commit 4c032f34f870ee35a10ba4c5e203de6cc2565865
author Frank Barchard <fbarchard@google.com> Wed Feb 10 15:18:49 2021 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Wed Feb 10 15:19:29 2021 -0800
tree cefa00455e571ffa0ace4bdf7446b484ca99ec36
parent b8c9fa903518a91433ec389ea55c3f461c0c751e

E2E run benchmarks from fastest to slowest instruction set

PiperOrigin-RevId: 356838158

bench/f32-gemm-e2e.cc

diff --git a/bench/f32-gemm-e2e.cc b/bench/f32-gemm-e2e.cc
index 281b4d1..074c545 100644
--- a/bench/f32-gemm-e2e.cc
+++ b/bench/f32-gemm-e2e.cc
@@ -886,6 +886,37 @@
       benchmark::utils::CheckAVX512F);
   }
 
+  BENCHMARK_FP32_END2END(f32_gemm_4x16__avx512f_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_5x16__avx512f_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_6x16__avx512f_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_7x16__avx512f_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_8x16__avx512f_broadcast);
+
+  BENCHMARK_FP32_END2END(f32_gemm_3x16s4__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_4x16s4__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_5x16s4__fma3_broadcast);
+
+  BENCHMARK_FP32_END2END(f32_gemm_4x8__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_5x8__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_6x8__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_7x8__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_8x8__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_3x16__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_4x16__fma3_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_5x16__fma3_broadcast);
+
+  BENCHMARK_FP32_END2END(f32_gemm_4x8__avx_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_5x8__avx_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_6x8__avx_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_7x8__avx_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_3x16__avx_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_4x16__avx_broadcast);
+  BENCHMARK_FP32_END2END(f32_gemm_5x16__avx_broadcast);
+
+  BENCHMARK_FP32_END2END(f32_gemm_3x8__sse2_dup);
+  BENCHMARK_FP32_END2END(f32_gemm_4x8__sse2_dup);
+  BENCHMARK_FP32_END2END(f32_gemm_5x8__sse2_dup);
+
   BENCHMARK_FP32_END2END(f32_gemm_3x8__sse_load1);
   BENCHMARK_FP32_END2END(f32_gemm_4x8__sse_load1);
   BENCHMARK_FP32_END2END(f32_gemm_5x8__sse_load1);
@@ -897,37 +928,6 @@
   BENCHMARK_FP32_END2END(f32_gemm_3x8s4__sse);
   BENCHMARK_FP32_END2END(f32_gemm_4x8s4__sse);
   BENCHMARK_FP32_END2END(f32_gemm_5x8s4__sse);
-
-  BENCHMARK_FP32_END2END(f32_gemm_3x8__sse2_dup);
-  BENCHMARK_FP32_END2END(f32_gemm_4x8__sse2_dup);
-  BENCHMARK_FP32_END2END(f32_gemm_5x8__sse2_dup);
-
-  BENCHMARK_FP32_END2END(f32_gemm_4x8__avx_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_5x8__avx_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_6x8__avx_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_7x8__avx_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_3x16__avx_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_4x16__avx_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_5x16__avx_broadcast);
-
-  BENCHMARK_FP32_END2END(f32_gemm_4x8__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_5x8__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_6x8__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_7x8__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_8x8__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_3x16__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_4x16__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_5x16__fma3_broadcast);
-
-  BENCHMARK_FP32_END2END(f32_gemm_3x16s4__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_4x16s4__fma3_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_5x16s4__fma3_broadcast);
-
-  BENCHMARK_FP32_END2END(f32_gemm_4x16__avx512f_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_5x16__avx512f_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_6x16__avx512f_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_7x16__avx512f_broadcast);
-  BENCHMARK_FP32_END2END(f32_gemm_8x16__avx512f_broadcast);
 #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
 
 #if XNN_ARCH_WASMSIMD