diff --git a/BUILD.bazel b/BUILD.bazel
index c3b141e..a069792 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -3123,13 +3123,11 @@
     "src/qs8-igemm/gen/4x8c4-minmax-rndnu-neondot.c",
     "src/qs8-igemm/gen/4x16c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/1x8c4-minmax-rndnu-neondot.c",
-    "src/qu8-gemm/gen/1x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/4x8c4-minmax-rndnu-neondot.c",
-    "src/qu8-gemm/gen/4x16c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/1x8c4-minmax-rndnu-neondot.c",
-    "src/qu8-igemm/gen/1x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/4x8c4-minmax-rndnu-neondot.c",
-    "src/qu8-igemm/gen/4x16c4-minmax-rndnu-neondot.c",
 ]
 
 ALL_NEONDOT_MICROKERNEL_SRCS = [
@@ -3171,16 +3169,28 @@
     "src/qs8-igemm/gen/8x16c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/1x8c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/1x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c",
+    "src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c",
+    "src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/4x8c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/4x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c",
+    "src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/6x8c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/6x16c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/8x8c4-minmax-rndnu-neondot.c",
     "src/qu8-gemm/gen/8x16c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/1x8c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/1x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c",
+    "src/qu8-igemm/gen/2x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-igemm/gen/3x8c4-minmax-rndnu-neondot.c",
+    "src/qu8-igemm/gen/3x16c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/4x8c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/4x16c4-minmax-rndnu-neondot.c",
+    "src/qu8-igemm/gen/5x8c4-minmax-rndnu-neondot.c",
+    "src/qu8-igemm/gen/5x16c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/6x8c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/6x16c4-minmax-rndnu-neondot.c",
     "src/qu8-igemm/gen/8x8c4-minmax-rndnu-neondot.c",
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9eff5d7..885b5b9 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -2311,13 +2311,11 @@
   src/qs8-igemm/gen/4x8c4-minmax-rndnu-neondot.c
   src/qs8-igemm/gen/4x16c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/1x8c4-minmax-rndnu-neondot.c
-  src/qu8-gemm/gen/1x16c4-minmax-rndnu-neondot.c
+  src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/4x8c4-minmax-rndnu-neondot.c
-  src/qu8-gemm/gen/4x16c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/1x8c4-minmax-rndnu-neondot.c
-  src/qu8-igemm/gen/1x16c4-minmax-rndnu-neondot.c
-  src/qu8-igemm/gen/4x8c4-minmax-rndnu-neondot.c
-  src/qu8-igemm/gen/4x16c4-minmax-rndnu-neondot.c)
+  src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c
+  src/qu8-igemm/gen/4x8c4-minmax-rndnu-neondot.c)
 
 SET(ALL_NEONDOT_MICROKERNEL_SRCS
   src/qc8-gemm/gen/1x8c4-minmax-fp32-neondot.c
@@ -2358,16 +2356,28 @@
   src/qs8-igemm/gen/8x16c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/1x8c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/1x16c4-minmax-rndnu-neondot.c
+  src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c
+  src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c
+  src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c
+  src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/4x8c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/4x16c4-minmax-rndnu-neondot.c
+  src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c
+  src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/6x8c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/6x16c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/8x8c4-minmax-rndnu-neondot.c
   src/qu8-gemm/gen/8x16c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/1x8c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/1x16c4-minmax-rndnu-neondot.c
+  src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c
+  src/qu8-igemm/gen/2x16c4-minmax-rndnu-neondot.c
+  src/qu8-igemm/gen/3x8c4-minmax-rndnu-neondot.c
+  src/qu8-igemm/gen/3x16c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/4x8c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/4x16c4-minmax-rndnu-neondot.c
+  src/qu8-igemm/gen/5x8c4-minmax-rndnu-neondot.c
+  src/qu8-igemm/gen/5x16c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/6x8c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/6x16c4-minmax-rndnu-neondot.c
   src/qu8-igemm/gen/8x8c4-minmax-rndnu-neondot.c
diff --git a/bench/qu8-gemm.cc b/bench/qu8-gemm.cc
index 3ee2200..3855d59 100644
--- a/bench/qu8-gemm.cc
+++ b/bench/qu8-gemm.cc
@@ -300,61 +300,131 @@
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
   static void qu8_gemm_1x8c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 1, 8, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      1, 8, 4, 1, benchmark::utils::CheckNEONDOT);
+  }
+  static void qu8_gemm_2x8c4__neondot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      2, 8, 4, 1, benchmark::utils::CheckNEONDOT);
+  }
+  static void qu8_gemm_3x8c4__neondot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      3, 8, 4, 1, benchmark::utils::CheckNEONDOT);
   }
   static void qu8_gemm_4x8c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 4, 8, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      4, 8, 4, 1, benchmark::utils::CheckNEONDOT);
+  }
+  static void qu8_gemm_5x8c4__neondot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      5, 8, 4, 1, benchmark::utils::CheckNEONDOT);
   }
   static void qu8_gemm_6x8c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 6, 8, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      6, 8, 4, 1, benchmark::utils::CheckNEONDOT);
   }
   static void qu8_gemm_8x8c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 8, 8, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      8, 8, 4, 1, benchmark::utils::CheckNEONDOT);
   }
   static void qu8_gemm_1x16c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 1, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      1, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+  }
+  static void qu8_gemm_2x16c4__neondot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      2, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+  }
+  static void qu8_gemm_3x16c4__neondot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      3, 16, 4, 1, benchmark::utils::CheckNEONDOT);
   }
   static void qu8_gemm_4x16c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 4, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      4, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+  }
+  static void qu8_gemm_5x16c4__neondot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      5, 16, 4, 1, benchmark::utils::CheckNEONDOT);
   }
   static void qu8_gemm_6x16c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 6, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      6, 16, 4, 1, benchmark::utils::CheckNEONDOT);
   }
   static void qu8_gemm_8x16c4__neondot(benchmark::State& state, const char* net) {
-    GEMMBenchmark(state, xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot,
-      xnn_init_qu8_conv_minmax_rndnu_neon_params, 8, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      8, 16, 4, 1, benchmark::utils::CheckNEONDOT);
+  }
+  static void qu8_gemm_1x8__neon_mlal_lane(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      1, 8, 1, 1, benchmark::utils::CheckNEON);
   }
   static void qu8_gemm_1x16__neon_mlal_lane(benchmark::State& state, const char* net) {
     GEMMBenchmark(state,
       xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane,
       xnn_init_qu8_conv_minmax_rndnu_neon_params,
-      1, 16, 1, 1,
-      benchmark::utils::CheckNEON);
+      1, 16, 1, 1, benchmark::utils::CheckNEON);
+  }
+  static void qu8_gemm_4x8__neon_mlal_lane(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane,
+      xnn_init_qu8_conv_minmax_rndnu_neon_params,
+      4, 8, 1, 1, benchmark::utils::CheckNEON);
   }
   static void qu8_gemm_4x16__neon_mlal_lane(benchmark::State& state, const char* net) {
     GEMMBenchmark(state,
       xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane,
       xnn_init_qu8_conv_minmax_rndnu_neon_params,
-      4, 16, 1, 1,
-      benchmark::utils::CheckNEON);
+      4, 16, 1, 1, benchmark::utils::CheckNEON);
   }
 
   BENCHMARK_GEMM(qu8_gemm_1x8c4__neondot)
+  BENCHMARK_GEMM(qu8_gemm_2x8c4__neondot)
+  BENCHMARK_GEMM(qu8_gemm_3x8c4__neondot)
   BENCHMARK_GEMM(qu8_gemm_4x8c4__neondot)
+  BENCHMARK_GEMM(qu8_gemm_5x8c4__neondot)
   BENCHMARK_GEMM(qu8_gemm_6x8c4__neondot)
   BENCHMARK_GEMM(qu8_gemm_8x8c4__neondot)
   BENCHMARK_GEMM(qu8_gemm_1x16c4__neondot)
+  BENCHMARK_GEMM(qu8_gemm_2x16c4__neondot)
+  BENCHMARK_GEMM(qu8_gemm_3x16c4__neondot)
   BENCHMARK_GEMM(qu8_gemm_4x16c4__neondot)
+  BENCHMARK_GEMM(qu8_gemm_5x16c4__neondot)
   BENCHMARK_GEMM(qu8_gemm_6x16c4__neondot)
   BENCHMARK_GEMM(qu8_gemm_8x16c4__neondot)
+  BENCHMARK_GEMM(qu8_gemm_1x8__neon_mlal_lane)
   BENCHMARK_GEMM(qu8_gemm_1x16__neon_mlal_lane)
+  BENCHMARK_GEMM(qu8_gemm_4x8__neon_mlal_lane)
   BENCHMARK_GEMM(qu8_gemm_4x16__neon_mlal_lane)
 #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
 
diff --git a/scripts/generate-qs8-gemm.sh b/scripts/generate-qs8-gemm.sh
index 3913bee..f0c99ce 100755
--- a/scripts/generate-qs8-gemm.sh
+++ b/scripts/generate-qs8-gemm.sh
@@ -269,14 +269,20 @@
 tools/xngen src/qs8-gemm/c4-neondot.c.in -D MR=6  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qs8-gemm/gen/6x16c4-minmax-rndnu-neondot.c
 tools/xngen src/qs8-gemm/c4-neondot.c.in -D MR=8  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qs8-gemm/gen/8x16c4-minmax-rndnu-neondot.c
 
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=1  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/1x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=4  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/4x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=6  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/6x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=8  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/8x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=1  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/1x16c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=4  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/4x16c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=6  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/6x16c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=8  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-gemm/gen/8x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=1  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/1x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=2  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=3  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=4  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/4x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=5  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=6  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/6x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=8  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/8x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=1  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/1x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=2  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=3  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=4  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/4x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=5  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=6  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/6x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-gemm/c4-neondot.c.in -D MR=8  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-gemm/gen/8x16c4-minmax-rndnu-neondot.c
 
 ############################### AArch64 assembly ##############################
 # Cortex A53 micro-kernel
diff --git a/scripts/generate-qs8-igemm.sh b/scripts/generate-qs8-igemm.sh
index 1d0ebe3..4b2cd2f 100755
--- a/scripts/generate-qs8-igemm.sh
+++ b/scripts/generate-qs8-igemm.sh
@@ -285,14 +285,20 @@
 tools/xngen src/qs8-igemm/c4-neondot.c.in -D MR=6  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qs8-igemm/gen/6x16c4-minmax-rndnu-neondot.c
 tools/xngen src/qs8-igemm/c4-neondot.c.in -D MR=8  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qs8-igemm/gen/8x16c4-minmax-rndnu-neondot.c
 
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=1  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/1x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=4  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/4x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=6  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/6x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=8  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/8x8c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=1  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/1x16c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=4  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/4x16c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=6  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/6x16c4-minmax-rndnu-neondot.c
-tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=8  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -D DATATYPE=QU8 -o src/qu8-igemm/gen/8x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=1  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/1x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=2  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=3  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/3x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=4  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/4x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=5  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/5x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=6  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/6x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=8  -D NR=8  -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/8x8c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=1  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/1x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=2  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/2x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=3  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/3x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=4  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/4x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=5  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/5x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=6  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/6x16c4-minmax-rndnu-neondot.c
+tools/xngen src/qu8-igemm/c4-neondot.c.in -D MR=8  -D NR=16 -D REQUANTIZATION=RNDNU    -D CHANNELWISE=0 -o src/qu8-igemm/gen/8x16c4-minmax-rndnu-neondot.c
 
 ############################### AArch64 assembly ##############################
 # Cortex A53 micro-kernel
diff --git a/src/init.c b/src/init.c
index 6e818b8..ac7eaa6 100644
--- a/src/init.c
+++ b/src/init.c
@@ -203,12 +203,12 @@
       init_flags |= XNN_INIT_FLAG_QU8;
 
       if (cpuinfo_has_arm_neon_dot()) {
-        xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot);
-        xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot);
+        xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot);
+        xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot);
         xnn_params.qu8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot);
         xnn_params.qu8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot);
         xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params;
-        xnn_params.qu8.gemm.mr = 4;
+        xnn_params.qu8.gemm.mr = 2;
         xnn_params.qu8.gemm.nr = 8;
         xnn_params.qu8.gemm.log2_kr = 2;
       } else {
diff --git a/src/qu8-gemm/c4-neondot.c.in b/src/qu8-gemm/c4-neondot.c.in
index ce81c94..c5b14b3 100644
--- a/src/qu8-gemm/c4-neondot.c.in
+++ b/src/qu8-gemm/c4-neondot.c.in
@@ -7,7 +7,6 @@
 $assert NR % 8 == 0
 $assert 8 <= NR <= 16
 $assert REQUANTIZATION == "RNDNU"
-$assert DATATYPE == "QU8"
 #include <assert.h>
 
 #include <arm_neon.h>
@@ -19,7 +18,7 @@
 
 
 $PARAMS_STRUCT = "fp32_neonv8" if REQUANTIZATION == "FP32" else REQUANTIZATION.lower() + "_neon"
-void xnn_${DATATYPE.lower()}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
+void xnn_qu8_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
     size_t mr,
     size_t nc,
     size_t kc,
diff --git a/src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..6881c90
--- /dev/null
+++ b/src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,261 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-gemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    const uint8_t* restrict a,
+    size_t a_stride,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 2);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(kc % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  const uint8_t* a0 = a;
+  uint8_t* c0 = c;
+  const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr != 2) {
+    a1 = a0;
+    c1 = c0;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  // Loop over groups of 16 columns.
+  do {
+    // Initialize accumulators with bias. 16 bias values are loaded from the
+    // weight matrix, at the start of the group of 16 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc1x89AB = vpacc0x89AB;
+    uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+
+    // Inner accumulation loop along the 16 columns.
+    size_t k = kc;
+    // 2x partial unrolled loop to load 8 bytes at a time.
+    while (k >= 8 * sizeof(uint8_t)) {
+      // Load a 2x8 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+
+      // Load a 8x16 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+
+      k -= 8 * sizeof(uint8_t);
+    }
+    // Handle up to 4 final positions of `k`
+    if XNN_UNLIKELY(k != 0) {
+      // Load a 2x4 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+
+      // Load a 4x16 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+    }
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+    int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+    int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+    vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+    vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+    vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+    vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+    vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+    vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+    uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+    uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+    vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+
+    vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+    vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+
+    if (nc >= 16) {
+      vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+      vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+
+      a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+      a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+
+      nc -= 16;
+    } else {
+      uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+      if (nc & 8) {
+        vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+        vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+        vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+      }
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..e6c759f
--- /dev/null
+++ b/src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,193 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-gemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    const uint8_t* restrict a,
+    size_t a_stride,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 2);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(kc % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  const uint8_t* a0 = a;
+  uint8_t* c0 = c;
+  const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr != 2) {
+    a1 = a0;
+    c1 = c0;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  // Loop over groups of 8 columns.
+  do {
+    // Initialize accumulators with bias. 8 bias values are loaded from the
+    // weight matrix, at the start of the group of 8 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+
+    // Inner accumulation loop along the 8 columns.
+    size_t k = kc;
+    // 2x partial unrolled loop to load 8 bytes at a time.
+    while (k >= 8 * sizeof(uint8_t)) {
+      // Load a 2x8 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+
+      // Load a 8x8 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+
+      k -= 8 * sizeof(uint8_t);
+    }
+    // Handle up to 4 final positions of `k`
+    if XNN_UNLIKELY(k != 0) {
+      // Load a 2x4 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+
+      // Load a 4x8 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+    }
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+
+    vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+
+    if (nc >= 8) {
+      vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+      vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+
+      a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+      a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+
+      nc -= 8;
+    } else {
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..8dafae1
--- /dev/null
+++ b/src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,336 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-gemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    const uint8_t* restrict a,
+    size_t a_stride,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 3);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(kc % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  const uint8_t* a0 = a;
+  uint8_t* c0 = c;
+  const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    a1 = a0;
+    c1 = c0;
+  }
+  const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    a2 = a1;
+    c2 = c1;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  // Loop over groups of 16 columns.
+  do {
+    // Initialize accumulators with bias. 16 bias values are loaded from the
+    // weight matrix, at the start of the group of 16 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc1x89AB = vpacc0x89AB;
+    uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x89AB = vpacc0x89AB;
+    uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc2xCDEF = vmovq_n_u32(0);
+
+    // Inner accumulation loop along the 16 columns.
+    size_t k = kc;
+    // 2x partial unrolled loop to load 8 bytes at a time.
+    while (k >= 8 * sizeof(uint8_t)) {
+      // Load a 3x8 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+
+      // Load a 8x16 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 3x8 * 8x16 --> 3x16.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+      vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+      vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+      vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+      vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
+      vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 1);
+      vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
+      vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 1);
+
+      k -= 8 * sizeof(uint8_t);
+    }
+    // Handle up to 4 final positions of `k`
+    if XNN_UNLIKELY(k != 0) {
+      // Load a 3x4 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+
+      // Load a 4x16 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+      vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+      vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+      vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+    }
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+    int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+    int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+    int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x89AB));
+    int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2xCDEF));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+    vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+    vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+    vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
+    vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+    vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+    vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+    vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
+    vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+    vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+    vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+    vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
+    vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+    uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+    uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+    uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+    uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+    vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+    vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
+
+    vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+    vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+    vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
+
+    if (nc >= 16) {
+      vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+      vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+      vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
+
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+
+      a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+      a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+      a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+
+      nc -= 16;
+    } else {
+      uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+      uint8x8_t vout2x01234567 = vget_low_u8(vout2x0123456789ABCDEF);
+      if (nc & 8) {
+        vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+        vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+        vst1_u8(c2, vout2x01234567); c2 += 8;
+        vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+        vout2x01234567 = vget_high_u8(vout2x0123456789ABCDEF);
+      }
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1_lane_u8(c2, vout2x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..7a46793
--- /dev/null
+++ b/src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,239 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-gemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    const uint8_t* restrict a,
+    size_t a_stride,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 3);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(kc % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  const uint8_t* a0 = a;
+  uint8_t* c0 = c;
+  const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    a1 = a0;
+    c1 = c0;
+  }
+  const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    a2 = a1;
+    c2 = c1;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  // Loop over groups of 8 columns.
+  do {
+    // Initialize accumulators with bias. 8 bias values are loaded from the
+    // weight matrix, at the start of the group of 8 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+
+    // Inner accumulation loop along the 8 columns.
+    size_t k = kc;
+    // 2x partial unrolled loop to load 8 bytes at a time.
+    while (k >= 8 * sizeof(uint8_t)) {
+      // Load a 3x8 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+
+      // Load a 8x8 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+
+      k -= 8 * sizeof(uint8_t);
+    }
+    // Handle up to 4 final positions of `k`
+    if XNN_UNLIKELY(k != 0) {
+      // Load a 3x4 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+
+      // Load a 4x8 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+    }
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+    uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+    uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+    vout2x01234567 = vmax_u8(vout2x01234567, vget_low_u8(voutput_min));
+
+    vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+    vout2x01234567 = vmin_u8(vout2x01234567, vget_low_u8(voutput_max));
+
+    if (nc >= 8) {
+      vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+      vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+      vst1_u8(c2 + 0, vout2x01234567);
+
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+
+      a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+      a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+      a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+
+      nc -= 8;
+    } else {
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1_lane_u8(c2, vout2x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..7381d8a
--- /dev/null
+++ b/src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,482 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-gemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    const uint8_t* restrict a,
+    size_t a_stride,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 5);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(kc % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  const uint8_t* a0 = a;
+  uint8_t* c0 = c;
+  const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    a1 = a0;
+    c1 = c0;
+  }
+  const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    a2 = a1;
+    c2 = c1;
+  }
+  const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
+  uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 4) {
+    a3 = a2;
+    c3 = c2;
+  }
+  const uint8_t* a4 = (const uint8_t*) ((uintptr_t) a3 + a_stride);
+  uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 4) {
+    a4 = a3;
+    c4 = c3;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  // Loop over groups of 16 columns.
+  do {
+    // Initialize accumulators with bias. 16 bias values are loaded from the
+    // weight matrix, at the start of the group of 16 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc1x89AB = vpacc0x89AB;
+    uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x89AB = vpacc0x89AB;
+    uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc3x0123 = vpacc0x0123;
+    uint32x4_t vpacc3x4567 = vpacc0x4567;
+    uint32x4_t vpacc3x89AB = vpacc0x89AB;
+    uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc4x0123 = vpacc0x0123;
+    uint32x4_t vpacc4x4567 = vpacc0x4567;
+    uint32x4_t vpacc4x89AB = vpacc0x89AB;
+    uint32x4_t vpacc4xCDEF = vpacc0xCDEF;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc2xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc3x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc3xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc4x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc4xCDEF = vmovq_n_u32(0);
+
+    // Inner accumulation loop along the 16 columns.
+    size_t k = kc;
+    // 2x partial unrolled loop to load 8 bytes at a time.
+    while (k >= 8 * sizeof(uint8_t)) {
+      // Load a 5x8 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+      const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
+      const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
+
+      // Load a 8x16 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+      vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+      vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+      vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+      vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+      vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+      vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+      vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+      vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
+      vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 0);
+      vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
+      vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 0);
+      vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+      vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+      vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+      vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+      vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
+      vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 0);
+      vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
+      vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 0);
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+      vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
+      vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 1);
+      vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
+      vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 1);
+      vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
+      vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 1);
+      vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
+      vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 1);
+      vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
+      vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 1);
+      vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
+      vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 1);
+      vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
+      vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 1);
+      vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
+      vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 1);
+      vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb4567x89AB, va4x01234567, 1);
+      vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 1);
+      vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
+      vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 1);
+
+      k -= 8 * sizeof(uint8_t);
+    }
+    // Handle up to 4 final positions of `k`
+    if XNN_UNLIKELY(k != 0) {
+      // Load a 5x4 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+      const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 4;
+      const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 4;
+
+      // Load a 4x16 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+      vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+      vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+      vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+      vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+      vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+      vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+      vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+      vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+      vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+      vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+      vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+      vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+      vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+      vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
+      vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 0);
+      vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
+      vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 0);
+      vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+      vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+      vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+      vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+      vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
+      vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 0);
+      vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
+      vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 0);
+    }
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+    int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+    int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+    int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x89AB));
+    int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2xCDEF));
+    int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
+    int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x4567));
+    int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x89AB));
+    int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3xCDEF));
+    int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
+    int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x4567));
+    int32x4_t vacc4x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc4x89AB, vnacc4x89AB));
+    int32x4_t vacc4xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc4xCDEF, vnacc4xCDEF));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+    vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+    vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+    vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
+    vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
+    vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
+    vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
+    vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
+    vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
+    vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
+    vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
+    vacc4x89AB = vshlq_s32(vacc4x89AB, vright_pre_shift);
+    vacc4xCDEF = vshlq_s32(vacc4xCDEF, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+    vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+    vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+    vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
+    vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
+    vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
+    vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
+    vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
+    vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
+    vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
+    vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
+    vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
+    vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+    vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+    vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+    vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
+    vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
+    vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
+    vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
+    vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
+    vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
+    vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
+    vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
+    vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
+    vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
+    const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
+    const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+    uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+    uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
+    uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
+    uint8x16_t vout4x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc4x89ABCDEF);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
+    const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
+    const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+    uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+    uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
+    uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
+    uint8x16_t vout4x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc4x89ABCDEF));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+    vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+    vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
+    vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
+    vout4x0123456789ABCDEF = vmaxq_u8(vout4x0123456789ABCDEF, voutput_min);
+
+    vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+    vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+    vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
+    vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
+    vout4x0123456789ABCDEF = vminq_u8(vout4x0123456789ABCDEF, voutput_max);
+
+    if (nc >= 16) {
+      vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+      vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+      vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
+      vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
+      vst1q_u8(c4 + 0, vout4x0123456789ABCDEF);
+
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+      c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
+      c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
+
+      a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+      a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+      a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+      a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
+      a4 = (const uint8_t*) ((uintptr_t) a4 - kc);
+
+      nc -= 16;
+    } else {
+      uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+      uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
+      uint8x8_t vout4x01234567 = vget_low_u8(vout4x0123456789ABCDEF);
+      if (nc & 8) {
+        vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+        vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+        vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
+        vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
+        vst1_u8(c4, vout4x01234567); c4 += 8;
+        vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+        vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
+        vout4x01234567 = vget_high_u8(vout4x0123456789ABCDEF);
+      }
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
+        vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
+        vst1_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
+        vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
+        vst1_lane_u8(c4, vout4x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..58a9a21
--- /dev/null
+++ b/src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,325 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-gemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    const uint8_t* restrict a,
+    size_t a_stride,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 5);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(kc % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  const uint8_t* a0 = a;
+  uint8_t* c0 = c;
+  const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    a1 = a0;
+    c1 = c0;
+  }
+  const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    a2 = a1;
+    c2 = c1;
+  }
+  const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
+  uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 4) {
+    a3 = a2;
+    c3 = c2;
+  }
+  const uint8_t* a4 = (const uint8_t*) ((uintptr_t) a3 + a_stride);
+  uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 4) {
+    a4 = a3;
+    c4 = c3;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  // Loop over groups of 8 columns.
+  do {
+    // Initialize accumulators with bias. 8 bias values are loaded from the
+    // weight matrix, at the start of the group of 8 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vpacc3x0123 = vpacc0x0123;
+    uint32x4_t vpacc3x4567 = vpacc0x4567;
+    uint32x4_t vpacc4x0123 = vpacc0x0123;
+    uint32x4_t vpacc4x4567 = vpacc0x4567;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x4567 = vmovq_n_u32(0);
+
+    // Inner accumulation loop along the 8 columns.
+    size_t k = kc;
+    // 2x partial unrolled loop to load 8 bytes at a time.
+    while (k >= 8 * sizeof(uint8_t)) {
+      // Load a 5x8 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+      const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
+      const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
+
+      // Load a 8x8 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 5x8 * 8x8 --> 5x8.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+      vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+      vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+      vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+      vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+      vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+      vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+      vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+      vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
+      vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 1);
+      vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
+      vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 1);
+      vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
+      vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 1);
+      vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
+      vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 1);
+
+      k -= 8 * sizeof(uint8_t);
+    }
+    // Handle up to 4 final positions of `k`
+    if XNN_UNLIKELY(k != 0) {
+      // Load a 5x4 block of activations.
+      const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+      const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+      const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+      const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 4;
+      const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 4;
+
+      // Load a 4x8 block of weights.
+      const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+      const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+      // Multiply-accumulate: 5x4 * 4x8 --> 5x8.
+      vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+      vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+      vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+      vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+      vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+      vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+      vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+      vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+      vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+      vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+      vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+      vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+      vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+      vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+      vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+      vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+      vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+      vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+    }
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+    int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
+    int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x4567));
+    int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
+    int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x4567));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+    vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
+    vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
+    vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
+    vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+    vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
+    vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
+    vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
+    vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+    vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
+    vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
+    vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
+    vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+    uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
+    uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+    uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
+    uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+    vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
+    vout4x01234567 = vmax_u8(vout4x01234567, vget_low_u8(voutput_min));
+
+    vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+    vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
+    vout4x01234567 = vmin_u8(vout4x01234567, vget_low_u8(voutput_max));
+
+    if (nc >= 8) {
+      vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+      vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+      vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
+      vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
+      vst1_u8(c4 + 0, vout4x01234567);
+
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+      c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
+      c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
+
+      a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+      a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+      a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+      a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
+      a4 = (const uint8_t*) ((uintptr_t) a4 - kc);
+
+      nc -= 8;
+    } else {
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
+        vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
+        vst1_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
+        vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
+        vst1_lane_u8(c4, vout4x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-igemm/c4-neondot.c.in b/src/qu8-igemm/c4-neondot.c.in
index 75e58a9..a011f66 100644
--- a/src/qu8-igemm/c4-neondot.c.in
+++ b/src/qu8-igemm/c4-neondot.c.in
@@ -7,7 +7,6 @@
 $assert NR % 8 == 0
 $assert 8 <= NR <= 16
 $assert REQUANTIZATION == "RNDNU"
-$assert DATATYPE == "QU8"
 #include <assert.h>
 
 #include <arm_neon.h>
@@ -19,7 +18,7 @@
 
 
 $PARAMS_STRUCT = "fp32_neonv8" if REQUANTIZATION == "FP32" else REQUANTIZATION.lower() + "_neon"
-void xnn_${DATATYPE.lower()}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
+void xnn_qu8_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
     size_t mr,
     size_t nc,
     size_t kc,
diff --git a/src/qu8-igemm/gen/2x16c4-minmax-rndnu-neondot.c b/src/qu8-igemm/gen/2x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..e213242
--- /dev/null
+++ b/src/qu8-igemm/gen/2x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,274 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-igemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    size_t ks,
+    const uint8_t** restrict a,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    size_t a_offset,
+    const uint8_t* zero,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 2);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(ks != 0);
+  assert(ks % (2 * sizeof(void*)) == 0);
+  assert(a_offset % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  uint8_t* c0 = c;
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr != 2) {
+    c1 = c0;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  do {
+    // Initialize accumulators with bias. 16 bias values are loaded from the
+    // weight matrix, at the start of the group of 16 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc1x89AB = vpacc0x89AB;
+    uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+
+    size_t p = ks;
+    do {
+      const uint8_t* restrict a0 = a[0];
+      if XNN_UNPREDICTABLE(a0 != zero) {
+        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
+      }
+      const uint8_t* restrict a1 = a[1];
+      if XNN_UNPREDICTABLE(a1 != zero) {
+        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
+      }
+      a += 2;
+
+      // Inner accumulation loop along the 16 columns.
+      size_t k = kc;
+      // 2x partial unrolled loop to load 8 bytes at a time.
+      while (k >= 8 * sizeof(uint8_t)) {
+        // Load a 2x8 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+        const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+
+        // Load a 8x16 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+
+        k -= 8 * sizeof(uint8_t);
+      }
+      // Handle up to 4 final positions of `k`
+      if XNN_UNLIKELY(k != 0) {
+        // Load a 2x4 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0);
+        const uint8x8_t va1x01234567 = vld1_u8(a1);
+
+        // Load a 4x16 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+      }
+      p -= 2 * sizeof(void*);
+    } while (p != 0);
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+    int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+    int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+    vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+    vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+    vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+    vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+    vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+    vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+    uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+    uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+    vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+
+    vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+    vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+
+    if (nc >= 16) {
+      vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+      vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+
+      a = (const uint8_t**restrict) ((uintptr_t) a - ks);
+
+      nc -= 16;
+    } else {
+      uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+      if (nc & 8) {
+        vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+        vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+        vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+      }
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c b/src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..d6c0a8b
--- /dev/null
+++ b/src/qu8-igemm/gen/2x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,206 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-igemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    size_t ks,
+    const uint8_t** restrict a,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    size_t a_offset,
+    const uint8_t* zero,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 2);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(ks != 0);
+  assert(ks % (2 * sizeof(void*)) == 0);
+  assert(a_offset % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  uint8_t* c0 = c;
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr != 2) {
+    c1 = c0;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  do {
+    // Initialize accumulators with bias. 8 bias values are loaded from the
+    // weight matrix, at the start of the group of 8 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+
+    size_t p = ks;
+    do {
+      const uint8_t* restrict a0 = a[0];
+      if XNN_UNPREDICTABLE(a0 != zero) {
+        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
+      }
+      const uint8_t* restrict a1 = a[1];
+      if XNN_UNPREDICTABLE(a1 != zero) {
+        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
+      }
+      a += 2;
+
+      // Inner accumulation loop along the 8 columns.
+      size_t k = kc;
+      // 2x partial unrolled loop to load 8 bytes at a time.
+      while (k >= 8 * sizeof(uint8_t)) {
+        // Load a 2x8 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+        const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+
+        // Load a 8x8 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+
+        k -= 8 * sizeof(uint8_t);
+      }
+      // Handle up to 4 final positions of `k`
+      if XNN_UNLIKELY(k != 0) {
+        // Load a 2x4 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0);
+        const uint8x8_t va1x01234567 = vld1_u8(a1);
+
+        // Load a 4x8 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+      }
+      p -= 2 * sizeof(void*);
+    } while (p != 0);
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+
+    vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+
+    if (nc >= 8) {
+      vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+      vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+
+      a = (const uint8_t**restrict) ((uintptr_t) a - ks);
+
+      nc -= 8;
+    } else {
+      if (nc & 4) {
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-igemm/gen/3x16c4-minmax-rndnu-neondot.c b/src/qu8-igemm/gen/3x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..a6668fb
--- /dev/null
+++ b/src/qu8-igemm/gen/3x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,350 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-igemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    size_t ks,
+    const uint8_t** restrict a,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    size_t a_offset,
+    const uint8_t* zero,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 3);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(ks != 0);
+  assert(ks % (3 * sizeof(void*)) == 0);
+  assert(a_offset % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  uint8_t* c0 = c;
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    c1 = c0;
+  }
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    c2 = c1;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  do {
+    // Initialize accumulators with bias. 16 bias values are loaded from the
+    // weight matrix, at the start of the group of 16 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc1x89AB = vpacc0x89AB;
+    uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x89AB = vpacc0x89AB;
+    uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc2xCDEF = vmovq_n_u32(0);
+
+    size_t p = ks;
+    do {
+      const uint8_t* restrict a0 = a[0];
+      if XNN_UNPREDICTABLE(a0 != zero) {
+        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
+      }
+      const uint8_t* restrict a1 = a[1];
+      if XNN_UNPREDICTABLE(a1 != zero) {
+        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
+      }
+      const uint8_t* restrict a2 = a[2];
+      if XNN_UNPREDICTABLE(a2 != zero) {
+        a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
+      }
+      a += 3;
+
+      // Inner accumulation loop along the 16 columns.
+      size_t k = kc;
+      // 2x partial unrolled loop to load 8 bytes at a time.
+      while (k >= 8 * sizeof(uint8_t)) {
+        // Load a 3x8 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+        const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+        const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+
+        // Load a 8x16 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 3x8 * 8x16 --> 3x16.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+        vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+        vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+        vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+        vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+        vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
+        vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 1);
+        vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
+        vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 1);
+
+        k -= 8 * sizeof(uint8_t);
+      }
+      // Handle up to 4 final positions of `k`
+      if XNN_UNLIKELY(k != 0) {
+        // Load a 3x4 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0);
+        const uint8x8_t va1x01234567 = vld1_u8(a1);
+        const uint8x8_t va2x01234567 = vld1_u8(a2);
+
+        // Load a 4x16 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+        vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+        vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+        vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+        vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+      }
+      p -= 3 * sizeof(void*);
+    } while (p != 0);
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+    int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+    int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+    int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x89AB));
+    int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2xCDEF));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+    vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+    vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+    vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
+    vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+    vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+    vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+    vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
+    vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+    vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+    vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+    vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
+    vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+    uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+    uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+    uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+    uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+    vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+    vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
+
+    vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+    vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+    vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
+
+    if (nc >= 16) {
+      vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
+      vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+      vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+
+      a = (const uint8_t**restrict) ((uintptr_t) a - ks);
+
+      nc -= 16;
+    } else {
+      uint8x8_t vout2x01234567 = vget_low_u8(vout2x0123456789ABCDEF);
+      uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+      if (nc & 8) {
+        vst1_u8(c2, vout2x01234567); c2 += 8;
+        vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+        vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+        vout2x01234567 = vget_high_u8(vout2x0123456789ABCDEF);
+        vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+      }
+      if (nc & 4) {
+        vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1_lane_u8(c2, vout2x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-igemm/gen/3x8c4-minmax-rndnu-neondot.c b/src/qu8-igemm/gen/3x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..949b984
--- /dev/null
+++ b/src/qu8-igemm/gen/3x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,253 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-igemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    size_t ks,
+    const uint8_t** restrict a,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    size_t a_offset,
+    const uint8_t* zero,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 3);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(ks != 0);
+  assert(ks % (3 * sizeof(void*)) == 0);
+  assert(a_offset % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  uint8_t* c0 = c;
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    c1 = c0;
+  }
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    c2 = c1;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  do {
+    // Initialize accumulators with bias. 8 bias values are loaded from the
+    // weight matrix, at the start of the group of 8 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+
+    size_t p = ks;
+    do {
+      const uint8_t* restrict a0 = a[0];
+      if XNN_UNPREDICTABLE(a0 != zero) {
+        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
+      }
+      const uint8_t* restrict a1 = a[1];
+      if XNN_UNPREDICTABLE(a1 != zero) {
+        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
+      }
+      const uint8_t* restrict a2 = a[2];
+      if XNN_UNPREDICTABLE(a2 != zero) {
+        a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
+      }
+      a += 3;
+
+      // Inner accumulation loop along the 8 columns.
+      size_t k = kc;
+      // 2x partial unrolled loop to load 8 bytes at a time.
+      while (k >= 8 * sizeof(uint8_t)) {
+        // Load a 3x8 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+        const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+        const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+
+        // Load a 8x8 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+
+        k -= 8 * sizeof(uint8_t);
+      }
+      // Handle up to 4 final positions of `k`
+      if XNN_UNLIKELY(k != 0) {
+        // Load a 3x4 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0);
+        const uint8x8_t va1x01234567 = vld1_u8(a1);
+        const uint8x8_t va2x01234567 = vld1_u8(a2);
+
+        // Load a 4x8 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+      }
+      p -= 3 * sizeof(void*);
+    } while (p != 0);
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+    uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+    uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+    vout2x01234567 = vmax_u8(vout2x01234567, vget_low_u8(voutput_min));
+
+    vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+    vout2x01234567 = vmin_u8(vout2x01234567, vget_low_u8(voutput_max));
+
+    if (nc >= 8) {
+      vst1_u8(c2 + 0, vout2x01234567);
+      vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+      vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+
+      a = (const uint8_t**restrict) ((uintptr_t) a - ks);
+
+      nc -= 8;
+    } else {
+      if (nc & 4) {
+        vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1_lane_u8(c2, vout2x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-igemm/gen/5x16c4-minmax-rndnu-neondot.c b/src/qu8-igemm/gen/5x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..ba906ba
--- /dev/null
+++ b/src/qu8-igemm/gen/5x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,498 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-igemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    size_t ks,
+    const uint8_t** restrict a,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    size_t a_offset,
+    const uint8_t* zero,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 5);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(ks != 0);
+  assert(ks % (5 * sizeof(void*)) == 0);
+  assert(a_offset % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  uint8_t* c0 = c;
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    c1 = c0;
+  }
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    c2 = c1;
+  }
+  uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 4) {
+    c3 = c2;
+  }
+  uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 4) {
+    c4 = c3;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  do {
+    // Initialize accumulators with bias. 16 bias values are loaded from the
+    // weight matrix, at the start of the group of 16 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc1x89AB = vpacc0x89AB;
+    uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x89AB = vpacc0x89AB;
+    uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc3x0123 = vpacc0x0123;
+    uint32x4_t vpacc3x4567 = vpacc0x4567;
+    uint32x4_t vpacc3x89AB = vpacc0x89AB;
+    uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
+    uint32x4_t vpacc4x0123 = vpacc0x0123;
+    uint32x4_t vpacc4x4567 = vpacc0x4567;
+    uint32x4_t vpacc4x89AB = vpacc0x89AB;
+    uint32x4_t vpacc4xCDEF = vpacc0xCDEF;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc2xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc3x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc3xCDEF = vmovq_n_u32(0);
+    uint32x4_t vnacc4x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x89AB = vmovq_n_u32(0);
+    uint32x4_t vnacc4xCDEF = vmovq_n_u32(0);
+
+    size_t p = ks;
+    do {
+      const uint8_t* restrict a0 = a[0];
+      if XNN_UNPREDICTABLE(a0 != zero) {
+        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
+      }
+      const uint8_t* restrict a1 = a[1];
+      if XNN_UNPREDICTABLE(a1 != zero) {
+        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
+      }
+      const uint8_t* restrict a2 = a[2];
+      if XNN_UNPREDICTABLE(a2 != zero) {
+        a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
+      }
+      const uint8_t* restrict a3 = a[3];
+      if XNN_UNPREDICTABLE(a3 != zero) {
+        a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
+      }
+      const uint8_t* restrict a4 = a[4];
+      if XNN_UNPREDICTABLE(a4 != zero) {
+        a4 = (const uint8_t*) ((uintptr_t) a4 + a_offset);
+      }
+      a += 5;
+
+      // Inner accumulation loop along the 16 columns.
+      size_t k = kc;
+      // 2x partial unrolled loop to load 8 bytes at a time.
+      while (k >= 8 * sizeof(uint8_t)) {
+        // Load a 5x8 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+        const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+        const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+        const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
+        const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
+
+        // Load a 8x16 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+        vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+        vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+        vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+        vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+        vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+        vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+        vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+        vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+        vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
+        vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 0);
+        vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
+        vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 0);
+        vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+        vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+        vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+        vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+        vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
+        vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 0);
+        vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
+        vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 0);
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+        vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
+        vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 1);
+        vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
+        vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 1);
+        vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
+        vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 1);
+        vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
+        vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 1);
+        vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
+        vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 1);
+        vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
+        vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 1);
+        vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
+        vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 1);
+        vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
+        vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 1);
+        vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb4567x89AB, va4x01234567, 1);
+        vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 1);
+        vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
+        vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 1);
+
+        k -= 8 * sizeof(uint8_t);
+      }
+      // Handle up to 4 final positions of `k`
+      if XNN_UNLIKELY(k != 0) {
+        // Load a 5x4 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0);
+        const uint8x8_t va1x01234567 = vld1_u8(a1);
+        const uint8x8_t va2x01234567 = vld1_u8(a2);
+        const uint8x8_t va3x01234567 = vld1_u8(a3);
+        const uint8x8_t va4x01234567 = vld1_u8(a4);
+
+        // Load a 4x16 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+        vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+        vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+        vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+        vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+        vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+        vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+        vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+        vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+        vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+        vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+        vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+        vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+        vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+        vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+        vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
+        vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 0);
+        vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
+        vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 0);
+        vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+        vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+        vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+        vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+        vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
+        vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 0);
+        vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
+        vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 0);
+      }
+      p -= 5 * sizeof(void*);
+    } while (p != 0);
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+    int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+    int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+    int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x89AB));
+    int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2xCDEF));
+    int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
+    int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x4567));
+    int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x89AB));
+    int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3xCDEF));
+    int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
+    int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x4567));
+    int32x4_t vacc4x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc4x89AB, vnacc4x89AB));
+    int32x4_t vacc4xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc4xCDEF, vnacc4xCDEF));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+    vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+    vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+    vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
+    vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
+    vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
+    vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
+    vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
+    vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
+    vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
+    vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
+    vacc4x89AB = vshlq_s32(vacc4x89AB, vright_pre_shift);
+    vacc4xCDEF = vshlq_s32(vacc4xCDEF, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+    vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+    vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+    vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
+    vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
+    vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
+    vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
+    vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
+    vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
+    vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
+    vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
+    vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
+    vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+    vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+    vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+    vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
+    vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
+    vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
+    vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
+    vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
+    vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
+    vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
+    vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
+    vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
+    vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
+    const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
+    const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+    uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+    uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
+    uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
+    uint8x16_t vout4x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc4x89ABCDEF);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+    const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
+    const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
+    const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
+
+    uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+    uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+    uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
+    uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
+    uint8x16_t vout4x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc4x89ABCDEF));
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+    vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+    vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
+    vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
+    vout4x0123456789ABCDEF = vmaxq_u8(vout4x0123456789ABCDEF, voutput_min);
+
+    vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+    vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+    vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
+    vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
+    vout4x0123456789ABCDEF = vminq_u8(vout4x0123456789ABCDEF, voutput_max);
+
+    if (nc >= 16) {
+      vst1q_u8(c4 + 0, vout4x0123456789ABCDEF);
+      vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
+      vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
+      vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+      vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+
+      c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
+      c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+
+      a = (const uint8_t**restrict) ((uintptr_t) a - ks);
+
+      nc -= 16;
+    } else {
+      uint8x8_t vout4x01234567 = vget_low_u8(vout4x0123456789ABCDEF);
+      uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
+      uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+      if (nc & 8) {
+        vst1_u8(c4, vout4x01234567); c4 += 8;
+        vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
+        vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
+        vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+        vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+        vout4x01234567 = vget_high_u8(vout4x0123456789ABCDEF);
+        vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
+        vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+      }
+      if (nc & 4) {
+        vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1_lane_u8(c4, vout4x01234567, 0);
+        vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
+        vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/qu8-igemm/gen/5x8c4-minmax-rndnu-neondot.c b/src/qu8-igemm/gen/5x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..f65a665
--- /dev/null
+++ b/src/qu8-igemm/gen/5x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,341 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qu8-igemm/c4-neondot.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/igemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot(
+    size_t mr,
+    size_t nc,
+    size_t kc,
+    size_t ks,
+    const uint8_t** restrict a,
+    const void* restrict w,
+    uint8_t* restrict c,
+    size_t cm_stride,
+    size_t cn_stride,
+    size_t a_offset,
+    const uint8_t* zero,
+    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  assert(mr != 0);
+  assert(mr <= 5);
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(ks != 0);
+  assert(ks % (5 * sizeof(void*)) == 0);
+  assert(a_offset % sizeof(uint8_t) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+
+  kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+  uint8_t* c0 = c;
+  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 2) {
+    c1 = c0;
+  }
+  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 2) {
+    c2 = c1;
+  }
+  uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
+  if XNN_UNPREDICTABLE(mr < 4) {
+    c3 = c2;
+  }
+  uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
+  if XNN_UNPREDICTABLE(mr <= 4) {
+    c4 = c3;
+  }
+
+  const uint8x16_t vb_zero_point = vld1q_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
+
+  do {
+    // Initialize accumulators with bias. 8 bias values are loaded from the
+    // weight matrix, at the start of the group of 8 columns.
+    uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+    uint32x4_t vpacc1x0123 = vpacc0x0123;
+    uint32x4_t vpacc1x4567 = vpacc0x4567;
+    uint32x4_t vpacc2x0123 = vpacc0x0123;
+    uint32x4_t vpacc2x4567 = vpacc0x4567;
+    uint32x4_t vpacc3x0123 = vpacc0x0123;
+    uint32x4_t vpacc3x4567 = vpacc0x4567;
+    uint32x4_t vpacc4x0123 = vpacc0x0123;
+    uint32x4_t vpacc4x4567 = vpacc0x4567;
+    uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc3x4567 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x0123 = vmovq_n_u32(0);
+    uint32x4_t vnacc4x4567 = vmovq_n_u32(0);
+
+    size_t p = ks;
+    do {
+      const uint8_t* restrict a0 = a[0];
+      if XNN_UNPREDICTABLE(a0 != zero) {
+        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
+      }
+      const uint8_t* restrict a1 = a[1];
+      if XNN_UNPREDICTABLE(a1 != zero) {
+        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
+      }
+      const uint8_t* restrict a2 = a[2];
+      if XNN_UNPREDICTABLE(a2 != zero) {
+        a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
+      }
+      const uint8_t* restrict a3 = a[3];
+      if XNN_UNPREDICTABLE(a3 != zero) {
+        a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
+      }
+      const uint8_t* restrict a4 = a[4];
+      if XNN_UNPREDICTABLE(a4 != zero) {
+        a4 = (const uint8_t*) ((uintptr_t) a4 + a_offset);
+      }
+      a += 5;
+
+      // Inner accumulation loop along the 8 columns.
+      size_t k = kc;
+      // 2x partial unrolled loop to load 8 bytes at a time.
+      while (k >= 8 * sizeof(uint8_t)) {
+        // Load a 5x8 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+        const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+        const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+        const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
+        const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
+
+        // Load a 8x8 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 5x8 * 8x8 --> 5x8.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+        vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+        vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+        vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+        vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+        vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+        vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+        vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+        vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+        vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
+        vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 1);
+        vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
+        vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 1);
+        vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
+        vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 1);
+        vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
+        vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 1);
+
+        k -= 8 * sizeof(uint8_t);
+      }
+      // Handle up to 4 final positions of `k`
+      if XNN_UNLIKELY(k != 0) {
+        // Load a 5x4 block of activations.
+        const uint8x8_t va0x01234567 = vld1_u8(a0);
+        const uint8x8_t va1x01234567 = vld1_u8(a1);
+        const uint8x8_t va2x01234567 = vld1_u8(a2);
+        const uint8x8_t va3x01234567 = vld1_u8(a3);
+        const uint8x8_t va4x01234567 = vld1_u8(a4);
+
+        // Load a 4x8 block of weights.
+        const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+        const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+        // Multiply-accumulate: 5x4 * 4x8 --> 5x8.
+        vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+        vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+        vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+        vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+        vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+        vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+        vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+        vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+        vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+        vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+        vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+        vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+        vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+        vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+        vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+        vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+        vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+        vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+        vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+        vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+      }
+      p -= 5 * sizeof(void*);
+    } while (p != 0);
+
+    // Subtract zero point accumulators with accumulators.
+    int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+    int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+    int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+    int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+    int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+    int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+    int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
+    int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x4567));
+    int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
+    int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x4567));
+
+    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
+    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
+    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
+
+    vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+    vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+    vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+    vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+    vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+    vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+    vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
+    vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
+    vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
+    vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
+
+    vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+    vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+    vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+    vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+    vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+    vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+    vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
+    vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
+    vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
+    vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
+
+    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+    vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
+    vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
+    vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
+    vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
+
+    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+    uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
+    uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
+#else
+    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+    const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
+    const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
+
+    uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+    uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
+    uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
+#endif
+    const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
+    const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
+
+    vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+    vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
+    vout4x01234567 = vmax_u8(vout4x01234567, vget_low_u8(voutput_min));
+
+    vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+    vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
+    vout4x01234567 = vmin_u8(vout4x01234567, vget_low_u8(voutput_max));
+
+    if (nc >= 8) {
+      vst1_u8(c4 + 0, vout4x01234567);
+      vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
+      vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
+      vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+      vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+
+      c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
+      c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
+      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+
+      a = (const uint8_t**restrict) ((uintptr_t) a - ks);
+
+      nc -= 8;
+    } else {
+      if (nc & 4) {
+        vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+      }
+      if (nc & 2) {
+        vst1_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
+        vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
+      }
+      if (nc & 1) {
+        vst1_lane_u8(c4, vout4x01234567, 0);
+        vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
+        vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
+        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+      }
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}
diff --git a/src/xnnpack/gemm.h b/src/xnnpack/gemm.h
index ef82976..ec53bde 100644
--- a/src/xnnpack/gemm.h
+++ b/src/xnnpack/gemm.h
@@ -517,11 +517,17 @@
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75)
 
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot)
+DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot)
+DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot)
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot)
+DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot)
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot)
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot)
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot)
+DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot)
+DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot)
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot)
+DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot)
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot)
 DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot)
 
diff --git a/src/xnnpack/igemm.h b/src/xnnpack/igemm.h
index 0f5aaee..86e9e61 100644
--- a/src/xnnpack/igemm.h
+++ b/src/xnnpack/igemm.h
@@ -326,11 +326,17 @@
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75)
 
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot)
+DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot)
+DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot)
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot)
+DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot)
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot)
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot)
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot)
+DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot)
+DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot)
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot)
+DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot)
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot)
 DECLARE_QU8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_igemm_minmax_rndnu_ukernel_8x16c4__neondot)
 
diff --git a/test/qu8-gemm-minmax-rndnu.cc b/test/qu8-gemm-minmax-rndnu.cc
index f1d3329..220bb6e 100644
--- a/test/qu8-gemm-minmax-rndnu.cc
+++ b/test/qu8-gemm-minmax-rndnu.cc
@@ -528,6 +528,1016 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .cn_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .a_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      for (uint32_t n = 1; n <= 8; n++) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(8)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 8; n++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_stride(11)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_stride(19)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_stride(83)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(11)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .cm_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .cn_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .a_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      for (uint32_t n = 1; n <= 8; n++) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(8)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 8; n++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_stride(11)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_stride(19)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_stride(83)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(11)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .cm_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
@@ -1033,6 +2043,511 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .cn_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .a_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      for (uint32_t n = 1; n <= 8; n++) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(8)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 8; n++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .a_stride(11)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .a_stride(19)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .a_stride(83)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(11)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .cm_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_GEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
@@ -2548,6 +4063,1016 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .cn_stride(19)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .a_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      for (uint32_t n = 1; n <= 16; n++) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(16)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 16; n++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_lt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_stride(11)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_stride(19)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_stride(83)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(19)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .cm_stride(19)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_2X16C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .cn_stride(19)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .a_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      for (uint32_t n = 1; n <= 16; n++) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(16)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 16; n++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_lt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_stride(11)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_stride(19)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_stride(83)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(19)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .cm_stride(19)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_3X16C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
@@ -3053,6 +5578,511 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .cn_stride(19)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .a_stride(11)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      for (uint32_t n = 1; n <= 16; n++) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(16)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 16; n++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_lt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .a_stride(11)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_gt_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .a_stride(19)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_div_8_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .a_stride(83)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .a_stride(43)
+          .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(19)
+            .iterations(1)
+            .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .cm_stride(19)
+      .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_GEMM_MINMAX_RNDNU_5X16C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
diff --git a/test/qu8-gemm-minmax-rndnu.yaml b/test/qu8-gemm-minmax-rndnu.yaml
index c42833d..7cf2f78 100644
--- a/test/qu8-gemm-minmax-rndnu.yaml
+++ b/test/qu8-gemm-minmax-rndnu.yaml
@@ -6,9 +6,18 @@
 - name: xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
+- name: xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
@@ -18,9 +27,18 @@
 - name: xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
+- name: xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
diff --git a/test/qu8-igemm-minmax-rndnu.cc b/test/qu8-igemm-minmax-rndnu.cc
index 24438a7..f287c46 100644
--- a/test/qu8-igemm-minmax-rndnu.cc
+++ b/test/qu8-igemm-minmax-rndnu.cc
@@ -540,6 +540,1040 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .cn_stride(11)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      for (uint32_t n = 1; n <= 8; n++) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(8)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 8; n++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .ks(3)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, small_kernel_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .ks(3)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(11)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, a_offset) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .ks(3)
+        .a_offset(83)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, zero) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t mz = 0; mz < 2; mz++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .a_offset(83)
+          .zero_index(mz)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(8)
+      .k(8)
+      .cm_stride(11)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .cn_stride(11)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      for (uint32_t n = 1; n <= 8; n++) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(8)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 8; n++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .ks(3)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, small_kernel_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .ks(3)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(11)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, a_offset) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .ks(3)
+        .a_offset(127)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, zero) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t mz = 0; mz < 3; mz++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .a_offset(127)
+          .zero_index(mz)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(8)
+      .k(8)
+      .cm_stride(11)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_IGEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
@@ -1057,6 +2091,523 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .cn_stride(11)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      for (uint32_t n = 1; n <= 8; n++) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(8)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 8; n++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .cn_stride(11)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .ks(3)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, small_kernel_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .ks(3)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 9; n < 16; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 16; n <= 24; n += 8) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(8)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(11)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, a_offset) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .ks(3)
+        .a_offset(211)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, zero) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t mz = 0; mz < 5; mz++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(8)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(8)
+          .k(k)
+          .ks(3)
+          .a_offset(211)
+          .zero_index(mz)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(8)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(8)
+      .k(8)
+      .cm_stride(11)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(8)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(8)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
@@ -2608,6 +4159,1040 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .cn_stride(19)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      for (uint32_t n = 1; n <= 16; n++) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 2; m++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(16)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 16; n++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(n)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 2; m++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .ks(3)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, small_kernel_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .ks(3)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_gt_16_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, n_div_16_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(19)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, a_offset) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .ks(3)
+        .a_offset(83)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, zero) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t mz = 0; mz < 2; mz++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(2)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(2)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .a_offset(83)
+          .zero_index(mz)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(2)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(2)
+      .n(16)
+      .k(8)
+      .cm_stride(19)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_2X16C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .cn_stride(19)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      for (uint32_t n = 1; n <= 16; n++) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 3; m++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(16)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 16; n++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(n)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 3; m++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .ks(3)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, small_kernel_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .ks(3)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_gt_16_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, n_div_16_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(19)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, a_offset) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .ks(3)
+        .a_offset(127)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, zero) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t mz = 0; mz < 3; mz++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(3)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(3)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .a_offset(127)
+          .zero_index(mz)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(3)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(3)
+      .n(16)
+      .k(8)
+      .cm_stride(19)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_3X16C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
@@ -3125,6 +5710,523 @@
 
 
 #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .cn_stride(19)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      for (uint32_t n = 1; n <= 16; n++) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(m)
+          .n(n)
+          .k(8)
+          .iterations(1)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8_subtile_m) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t m = 1; m <= 5; m++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(m)
+        .n(16)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_eq_8_subtile_n) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 1; n <= 16; n++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(n)
+        .k(8)
+        .iterations(1)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_lt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_lt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k < 8; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_gt_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_gt_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 9; k < 16; k++) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_div_8) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, k_div_8_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 16; k <= 80; k += 8) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16_strided_cn) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(n)
+          .k(k)
+          .cn_stride(19)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        for (uint32_t m = 1; m <= 5; m++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .ks(3)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, small_kernel_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .ks(3)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_gt_16_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 17; n < 32; n++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, n_div_16_small_kernel) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t n = 32; n <= 48; n += 16) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, strided_cm_subtile) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      for (uint32_t m = 1; m <= 5; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(5)
+            .nr(16)
+            .kr(4)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .cm_stride(19)
+            .iterations(1)
+            .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+        }
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, a_offset) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .ks(3)
+        .a_offset(211)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, zero) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (uint32_t mz = 0; mz < 5; mz++) {
+      for (size_t k = 1; k <= 40; k += 9) {
+        GemmMicrokernelTester()
+          .mr(5)
+          .nr(16)
+          .kr(4)
+          .sr(1)
+          .m(5)
+          .n(16)
+          .k(k)
+          .ks(3)
+          .a_offset(211)
+          .zero_index(mz)
+          .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+      }
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, qmin) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .qmin(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, qmax) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .qmax(128)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, strided_cm) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    GemmMicrokernelTester()
+      .mr(5)
+      .nr(16)
+      .kr(4)
+      .sr(1)
+      .m(5)
+      .n(16)
+      .k(8)
+      .cm_stride(19)
+      .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, no_a_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, no_b_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+
+  TEST(QU8_IGEMM_MINMAX_RNDNU_5X16C4__NEONDOT, no_zero_point) {
+    TEST_REQUIRES_ARM_NEON_DOT;
+    for (size_t k = 1; k <= 40; k += 9) {
+      GemmMicrokernelTester()
+        .mr(5)
+        .nr(16)
+        .kr(4)
+        .sr(1)
+        .m(5)
+        .n(16)
+        .k(k)
+        .a_zero_point(0)
+        .b_zero_point(0)
+        .Test(xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_init_qu8_requantization_rndnu_params, xnn_qu8_requantize_rndnu);
+    }
+  }
+#endif  // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
   TEST(QU8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8) {
     TEST_REQUIRES_ARM_NEON_DOT;
     GemmMicrokernelTester()
diff --git a/test/qu8-igemm-minmax-rndnu.yaml b/test/qu8-igemm-minmax-rndnu.yaml
index 1b75977..5d39697 100644
--- a/test/qu8-igemm-minmax-rndnu.yaml
+++ b/test/qu8-igemm-minmax-rndnu.yaml
@@ -6,9 +6,18 @@
 - name: xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
+- name: xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
@@ -18,9 +27,18 @@
 - name: xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
+- name: xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8
+- name: xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot
+  init: xnn_init_qu8_conv_minmax_rndnu_neon_params
+  k-block: 8
 - name: xnn_qu8_igemm_minmax_rndnu_ukernel_6x16c4__neondot
   init: xnn_init_qu8_conv_minmax_rndnu_neon_params
   k-block: 8