add dummy compute to cache evict func (#337)

Summary:
Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/337

https://github.com/pytorch/FBGEMM/issues/333

Reviewed By: jianyuh

Differential Revision: D20799659

fbshipit-source-id: 6b38301b39423806cb4a5087c0925116c336854f

1 files changed, 13 insertions, 1 deletions

diff --git a/bench/BenchUtils.h b/bench/BenchUtils.h
index 2b63068..64dafd3 100644
--- a/bench/BenchUtils.h
+++ b/bench/BenchUtils.h
@@ -15,6 +15,7 @@
 #include <omp.h>
 #endif
 #include "./AlignedVec.h"
+#include "fbgemm/FbgemmBuild.h"
 
 namespace fbgemm {
 
@@ -31,16 +32,27 @@ int fbgemm_get_num_threads();
 int fbgemm_get_thread_num();
 
 template <typename T>
-void cache_evict(const T& vec) {
+NOINLINE
+float cache_evict(const T& vec) {
   auto const size = vec.size();
   auto const elemSize = sizeof(typename T::value_type);
   auto const dataSize = size * elemSize;
 
   const char* data = reinterpret_cast<const char*>(vec.data());
   constexpr int CACHE_LINE_SIZE = 64;
+  // Not having this dummy computation significantly slows down the computation
+  // that follows.
+  float dummy = 0.0f;
   for (std::size_t i = 0; i < dataSize; i += CACHE_LINE_SIZE) {
+    dummy += data[i] * 1.0f;
+    _mm_mfence();
+#ifndef _MSC_VER
+    asm volatile("" ::: "memory");
+#endif
     _mm_clflush(&data[i]);
   }
+
+  return dummy;
 }
 
 /**