Move FakeFP16 back to internal to remove dependency on MKL (#36297)

Summary: Pull Request resolved: pytorch/pytorch#36297 Pull Request resolved: pytorch#343 We moved FakeFP16 back to close source and kept `RoundToFloat16` function in "fbgemm/FbgemmConvert.h". This is because FakeFP16 introduced dependency on MKL in the FBGEMM core. Also it doesn't seem to be needed for open source, as it is not used anywhere. Reviewed By: jspark1105 Differential Revision: D20937962 fbshipit-source-id: 9487a9fd2282b6df2f754c22bea36f2255a5c791
z-a-f · Apr 10, 2020 · 20b6cf1 · 20b6cf1
1 parent f1d3330
commit 20b6cf1
Show file tree

Hide file tree

Showing 10 changed files with 54 additions and 773 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -68,10 +68,10 @@ else()
 endif()
 
 # Define file lists
-get_filelist("get_fbgemm_generic_srcs(with_fp16=True)" FBGEMM_GENERIC_SRCS)
+get_filelist("get_fbgemm_generic_srcs()" FBGEMM_GENERIC_SRCS)
 get_filelist("get_fbgemm_avx2_srcs(msvc=${MSVC_BOOL})" FBGEMM_AVX2_SRCS)
 get_filelist("get_fbgemm_avx512_srcs(msvc=${MSVC_BOOL})" FBGEMM_AVX512_SRCS)
-get_filelist("get_fbgemm_public_headers(with_fp16=True)" FBGEMM_PUBLIC_HEADERS)
+get_filelist("get_fbgemm_public_headers()" FBGEMM_PUBLIC_HEADERS)
 
 add_library(fbgemm_generic OBJECT ${FBGEMM_GENERIC_SRCS})
 add_library(fbgemm_avx2 OBJECT ${FBGEMM_AVX2_SRCS})

diff --git a/bench/BenchUtils.cc b/bench/BenchUtils.cc
@@ -6,8 +6,8 @@
  */
 #include "./BenchUtils.h"
 
-#include <cstring>
 #include <algorithm>
+#include <cstring>
 #include <random>
 #include <type_traits>
 
@@ -43,8 +43,8 @@ randFill<uint8_t>(aligned_vector<uint8_t>& vec, uint8_t low, uint8_t high);
 template void
 randFill<int8_t>(aligned_vector<int8_t>& vec, int8_t low, int8_t high);
 template void randFill<int>(aligned_vector<int>& vec, int low, int high);
-//template void
-//randFill<int64_t>(aligned_vector<int64_t>& vec, int64_t low, int64_t high);
+// template void
+// randFill<int64_t>(aligned_vector<int64_t>& vec, int64_t low, int64_t high);
 template <>
 void randFill(aligned_vector<int64_t>& vec, int64_t low, int64_t high) {
   std::uniform_int_distribution<int64_t> dis(low, high);

diff --git a/bench/FakeFP16Benchmark.cc b/bench/FakeFP16Benchmark.cc
diff --git a/defs.bzl b/defs.bzl
@@ -1,4 +1,4 @@
-def get_fbgemm_generic_srcs(with_fp16 = False):
+def get_fbgemm_generic_srcs():
     return [
         "src/EmbeddingSpMDM.cc",
         "src/EmbeddingSpMDMNBit.cc",
@@ -32,9 +32,9 @@ def get_fbgemm_generic_srcs(with_fp16 = False):
         "src/RowWiseSparseAdagradFused.cc",
         "src/SparseAdagrad.cc",
         "src/Utils.cc",
-    ] + (["src/FbgemmFakeFP16.cc"] if with_fp16 else [])
+    ]
 
-def get_fbgemm_public_headers(with_fp16 = False):
+def get_fbgemm_public_headers():
     return [
         "include/fbgemm/Fbgemm.h",
         "include/fbgemm/FbgemmBuild.h",
@@ -52,7 +52,7 @@ def get_fbgemm_public_headers(with_fp16 = False):
         "include/fbgemm/ConvUtils.h",
         "include/fbgemm/Types.h",
         "include/fbgemm/FbgemmI8Spmdm.h",
-    ] + (["include/fbgemm/FbgemmFakeFP16.h"] if with_fp16 else [])
+    ]
 
 def get_fbgemm_avx2_srcs(msvc = False):
     return [

diff --git a/include/fbgemm/FbgemmConvert.h b/include/fbgemm/FbgemmConvert.h
@@ -136,4 +136,15 @@ FBGEMM_API void Float16ToFloat_avx2(const float16* src, float* dst, int size);
  */
 FBGEMM_API void Float16ToFloat_avx512(const float16* src, float* dst, int size);
 
+/**
+ * @brief Transform all entries in a matrix from fp32 to float16 and back to
+ * fp32.
+ */
+FBGEMM_API void RoundToFloat16(
+    const float* input,
+    float* output,
+    int len,
+    bool clamp = false,
+    bool clamp_denorms = false);
+
 }; // namespace fbgemm