commit 2f06150c8095b9674a5a48726802c073dc75a1b1
author Frank Barchard <fbarchard@google.com> Thu Mar 11 02:18:15 2021 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Thu Mar 11 02:18:55 2021 -0800
tree c47581fe7502b0bcc22fc101386815773f353bfc
parent 1dc9fef1d2c493a99b756733dbc61b717367b86b

xnn_qs8_gemm_minmax_ukernel_2x8c8__aarch64_neon_mlal_padal GEMM microkernel

Similar math to C16 but load order is different.

PiperOrigin-RevId: 362245222

bench/qs8-gemm.cc

diff --git a/bench/qs8-gemm.cc b/bench/qs8-gemm.cc
index 639d890..8a0d8f1 100644
--- a/bench/qs8-gemm.cc
+++ b/bench/qs8-gemm.cc
@@ -559,6 +559,12 @@
   static void qs8_gemm_4x16c4__aarch64_neondot_ld64(benchmark::State& state, const char* net) {
     GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_ld64, 4, 16, 4, 1, benchmark::utils::CheckNEONDOT);
   }
+  static void qs8_gemm_2x8c8__aarch64_neon_mull_padal(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x8c8__aarch64_neon_mull_padal, 2, 8, 8, 1, benchmark::utils::CheckNEON);
+  }
+  static void qs8_gemm_2x8c8__aarch64_neon_mlal_padal(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x8c8__aarch64_neon_mlal_padal, 2, 8, 8, 1, benchmark::utils::CheckNEON);
+  }
   static void qs8_gemm_2x8c16__aarch64_neon_mlal_padal(benchmark::State& state, const char* net) {
     GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_2x8c16__aarch64_neon_mlal_padal, 2, 8, 16, 1, benchmark::utils::CheckNEON);
   }
@@ -568,6 +574,8 @@
   BENCHMARK_GEMM(qs8_gemm_4x16c4__aarch64_neondot_ld32)
   BENCHMARK_GEMM(qs8_gemm_4x16c4__aarch64_neondot_ld64)
   BENCHMARK_GEMM(qs8_gemm_4x16c4__aarch64_neondot_cortex_a55)
+  BENCHMARK_GEMM(qs8_gemm_2x8c8__aarch64_neon_mull_padal)
+  BENCHMARK_GEMM(qs8_gemm_2x8c8__aarch64_neon_mlal_padal)
   BENCHMARK_GEMM(qs8_gemm_2x8c16__aarch64_neon_mlal_padal)
 #endif  // XNN_ARCH_ARM64