commit 4ecae2e7b73de22f0c1f31c16b747de70177c91b
author Frank Barchard <fbarchard@google.com> Thu Jan 14 15:52:45 2021 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Thu Jan 14 15:53:29 2021 -0800
tree 60a87651a67db08c8821bc9e6c1f2705761bd877
parent b94e34b238491ab83a8ad94e7aa147db2277dbec

QS8 Neon GEMM microkernel with 8 bit multiply

PiperOrigin-RevId: 351893800

bench/qs8-gemm.cc

diff --git a/bench/qs8-gemm.cc b/bench/qs8-gemm.cc
index a231fce..63e09f3 100644
--- a/bench/qs8-gemm.cc
+++ b/bench/qs8-gemm.cc
@@ -296,6 +296,18 @@
   static void qs8_gemm_4x16__neon_mlal_lane(benchmark::State& state, const char* net) {
     GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane, 4, 16, 1, 1, benchmark::utils::CheckNEON);
   }
+  static void qs8_gemm_2x8__neon_mull_addw_dup(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup, 2, 8, 1, 1, benchmark::utils::CheckNEON);
+  }
+  static void qs8_gemm_4x8__neon_mull_addw_dup(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup, 4, 8, 1, 1, benchmark::utils::CheckNEON);
+  }
+  static void qs8_gemm_2x16__neon_mull_addw_dup(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup, 2, 16, 1, 1, benchmark::utils::CheckNEON);
+  }
+  static void qs8_gemm_4x16__neon_mull_addw_dup(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup, 4, 16, 1, 1, benchmark::utils::CheckNEON);
+  }
   static void qs8_gemm_1x8c4__neondot(benchmark::State& state, const char* net) {
     GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot, 1, 8, 4, 1, benchmark::utils::CheckNEONDOT);
   }
@@ -328,6 +340,10 @@
   BENCHMARK_GEMM(qs8_gemm_4x8__neon_mlal_lane)
   BENCHMARK_GEMM(qs8_gemm_2x16__neon_mlal_lane)
   BENCHMARK_GEMM(qs8_gemm_4x16__neon_mlal_lane)
+  BENCHMARK_GEMM(qs8_gemm_2x8__neon_mull_addw_dup)
+  BENCHMARK_GEMM(qs8_gemm_4x8__neon_mull_addw_dup)
+  BENCHMARK_GEMM(qs8_gemm_2x16__neon_mull_addw_dup)
+  BENCHMARK_GEMM(qs8_gemm_4x16__neon_mull_addw_dup)
   BENCHMARK_GEMM(qs8_gemm_1x8c4__neondot)
   BENCHMARK_GEMM(qs8_gemm_4x8c4__neondot)
   BENCHMARK_GEMM(qs8_gemm_6x8c4__neondot)