Revert D20956364: Move `transpose_simd` to TransposeUtils.cc

Differential Revision:
D20956364

Original commit changeset: 5ec1c8261a81

fbshipit-source-id: fa1875d9b9d0837c9115f6a624056984c92b2aeb

8 files changed, 98 insertions, 119 deletions

diff --git a/BUILD.bazel b/BUILD.bazel
index 4f75644..6173642 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -1,19 +1,13 @@
 load("@bazel_skylib//lib:paths.bzl", "paths")
 load("@rules_cc//cc:defs.bzl", "cc_library", "cc_test")
-load("defs.bzl", "get_fbgemm_avx2_srcs", "get_fbgemm_avx512_srcs", "get_fbgemm_base_srcs", "get_fbgemm_generic_srcs", "get_fbgemm_public_headers", "get_fbgemm_tests")
+load("defs.bzl", "get_fbgemm_generic_srcs", "get_fbgemm_avx2_srcs", "get_fbgemm_avx512_srcs", "get_fbgemm_public_headers")
 
 cc_library(
-    name = "fbgemm_base",
-    srcs = get_fbgemm_base_srcs()  + glob(["src/*.h"]),
-    includes = [
-        "src",
-    ],
-    deps = [
-        ":fbgemm_headers",
-        "@cpuinfo",
-        "@asmjit",
+    name = "fbgemm_src_headers",
+    hdrs = [
+        "src/RefImplementations.h",
     ],
-    linkstatic = 1,
+    include_prefix = "fbgemm",
 )
 
 cc_library(
@@ -26,7 +20,10 @@ cc_library(
     deps = [
         ":fbgemm_avx2",
         ":fbgemm_avx512",
-        ":fbgemm_base",
+        ":fbgemm_headers",
+        ":fbgemm_src_headers",
+        "@asmjit",
+        "@cpuinfo",
     ],
     linkstatic = 1,
 )
@@ -42,7 +39,8 @@ cc_library(
         "-masm=intel",
     ],
     deps = [
-        ":fbgemm_base",
+        ":fbgemm_headers",
+        "@asmjit",
     ],
     linkstatic = 1,
 )
@@ -60,7 +58,7 @@ cc_library(
         "-masm=intel",
     ],
     deps = [
-        ":fbgemm_base",
+        ":fbgemm_headers",
     ],
     linkstatic = 1,
 )
@@ -99,12 +97,12 @@ cc_library(
 [
   cc_test(
       name = paths.split_extension(paths.basename(filename))[0],
-      size = "medium",
+      size = "small",
       srcs = [
           filename,
       ],
       deps = [
           ":test_utils",
       ],
-  ) for filename in get_fbgemm_tests()
+  ) for filename in ["test/GConvTest.cc", "test/TransposeTest.cc", "test/QuantUtilsTest.cc", "test/I64Test.cc"]
 ]
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 946855c..ce27196 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -68,7 +68,7 @@ else()
 endif()
 
 # Define file lists
-get_filelist("get_fbgemm_generic_srcs(with_base=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()" FBGEMM_PUBLIC_HEADERS)
diff --git a/defs.bzl b/defs.bzl
index bf67a3b..f46b118 100644
--- a/defs.bzl
+++ b/defs.bzl
@@ -1,11 +1,4 @@
-def get_fbgemm_base_srcs():
-    return [
-        "src/GenerateI8Depthwise.cc",
-        "src/RefImplementations.cc",
-        "src/Utils.cc",
-    ]
-
-def get_fbgemm_generic_srcs(with_base = False):
+def get_fbgemm_generic_srcs():
     return [
         "src/EmbeddingSpMDM.cc",
         "src/EmbeddingSpMDMNBit.cc",
@@ -18,6 +11,7 @@ def get_fbgemm_generic_srcs(with_base = False):
         "src/FbgemmFloat16Convert.cc",
         "src/FbgemmI64.cc",
         "src/FbgemmI8Spmdm.cc",
+        "src/GenerateI8Depthwise.cc",
         "src/GenerateKernelU8S8S32ACC16.cc",
         "src/GenerateKernelU8S8S32ACC16Avx512.cc",  # Acc16 AVX512 JIT code gen
         "src/GenerateKernelU8S8S32ACC16Avx512VNNI.cc",
@@ -34,10 +28,11 @@ def get_fbgemm_generic_srcs(with_base = False):
         "src/PackWeightMatrixForGConv.cc",
         "src/PackWeightsForConv.cc",
         "src/QuantUtils.cc",
+        "src/RefImplementations.cc",
         "src/RowWiseSparseAdagradFused.cc",
         "src/SparseAdagrad.cc",
-        "src/TransposeUtils.cc",
-    ] + (get_fbgemm_base_srcs() if with_base else [])
+        "src/Utils.cc",
+    ]
 
 def get_fbgemm_public_headers():
     return [
diff --git a/src/FbgemmBfloat16Convert.cc b/src/FbgemmBfloat16Convert.cc
index b0b6591..1a528d3 100644
--- a/src/FbgemmBfloat16Convert.cc
+++ b/src/FbgemmBfloat16Convert.cc
@@ -42,6 +42,21 @@ using namespace std;
 
 namespace fbgemm {
 
+void FloatToBfloat16_ref(const float* src, bfloat16* dst, int size) {
+  for (int i = 0; i < size; i++) {
+    // Add 2^15 and right shift 16 to do round-nearest
+    dst[i] = (*reinterpret_cast<const uint32_t*>(src + i) + (1 << 15)) >> 16;
+  }
+}
+
+void Bfloat16ToFloat_ref(const bfloat16* src, float* dst, int size) {
+  for (int i = 0; i < size; i++) {
+    uint32_t val_fp32 =
+        static_cast<uint32_t>(reinterpret_cast<const uint16_t*>(src)[i]) << 16;
+    reinterpret_cast<uint32_t*>(dst)[i] = val_fp32;
+  }
+}
+
 void FloatToBfloat16_simd(const float* src, bfloat16* dst, int size) {
   // Run time CPU detection
   if (cpuinfo_initialize()) {
diff --git a/src/FbgemmFloat16Convert.cc b/src/FbgemmFloat16Convert.cc
index a714dff..79e40a4 100644
--- a/src/FbgemmFloat16Convert.cc
+++ b/src/FbgemmFloat16Convert.cc
@@ -32,6 +32,30 @@ using namespace std;
 
 namespace fbgemm {
 
+void FloatToFloat16_ref(
+    const float* src,
+    float16* dst,
+    int size,
+    bool do_clip) {
+  constexpr float FP16_MAX = 65504.f;
+  if (do_clip) {
+    for (int i = 0; i < size; i++) {
+      float cur_src = std::max(-FP16_MAX, std::min(src[i], FP16_MAX));
+      dst[i] = cpu_float2half_rn(cur_src);
+    }
+  } else {
+    for (int i = 0; i < size; i++) {
+      dst[i] = cpu_float2half_rn(src[i]);
+    }
+  }
+}
+
+void Float16ToFloat_ref(const float16* src, float* dst, int size) {
+  for (int i = 0; i < size; i++) {
+    dst[i] = cpu_half2float(src[i]);
+  }
+}
+
 void FloatToFloat16_simd(
     const float* src,
     float16* dst,
diff --git a/src/RefImplementations.cc b/src/RefImplementations.cc
index 86663ed..d0809db 100644
--- a/src/RefImplementations.cc
+++ b/src/RefImplementations.cc
@@ -8,7 +8,6 @@
 #include "./RefImplementations.h"
 
 #include "fbgemm/FbgemmBuild.h"
-#include "fbgemm/FbgemmConvert.h"
 #include "fbgemm/Types.h"
 
 #include <algorithm>
@@ -20,45 +19,6 @@ using namespace std;
 
 namespace fbgemm {
 
-void FloatToFloat16_ref(
-    const float* src,
-    float16* dst,
-    int size,
-    bool do_clip) {
-  constexpr float FP16_MAX = 65504.f;
-  if (do_clip) {
-    for (int i = 0; i < size; i++) {
-      float cur_src = std::max(-FP16_MAX, std::min(src[i], FP16_MAX));
-      dst[i] = cpu_float2half_rn(cur_src);
-    }
-  } else {
-    for (int i = 0; i < size; i++) {
-      dst[i] = cpu_float2half_rn(src[i]);
-    }
-  }
-}
-
-void Float16ToFloat_ref(const float16* src, float* dst, int size) {
-  for (int i = 0; i < size; i++) {
-    dst[i] = cpu_half2float(src[i]);
-  }
-}
-
-void FloatToBfloat16_ref(const float* src, bfloat16* dst, int size) {
-  for (int i = 0; i < size; i++) {
-    // Add 2^15 and right shift 16 to do round-nearest
-    dst[i] = (*reinterpret_cast<const uint32_t*>(src + i) + (1 << 15)) >> 16;
-  }
-}
-
-void Bfloat16ToFloat_ref(const bfloat16* src, float* dst, int size) {
-  for (int i = 0; i < size; i++) {
-    uint32_t val_fp32 =
-        static_cast<uint32_t>(reinterpret_cast<const uint16_t*>(src)[i]) << 16;
-    reinterpret_cast<uint32_t*>(dst)[i] = val_fp32;
-  }
-}
-
 void requantize_u8acc32_ref(
     int M,
     int N,
diff --git a/src/TransposeUtils.cc b/src/TransposeUtils.cc
deleted file mode 100644
index 27afa62..0000000
--- a/src/TransposeUtils.cc
+++ /dev/null
@@ -1,52 +0,0 @@
-/*
- * Copyright (c) Facebook, Inc. and its affiliates.
- * All rights reserved.
- * This source code is licensed under the BSD-style license found in the
- * LICENSE file in the root directory of this source tree.
- */
-#define FBGEMM_EXPORTS
-#include "./TransposeUtils.h"
-#include "fbgemm/Utils.h"
-#include <cstring>
-
-namespace fbgemm {
-
-void transpose_ref(
-    int M,
-    int N,
-    const float* src,
-    int ld_src,
-    float* dst,
-    int ld_dst) {
-  for (int j = 0; j < N; j++) {
-    for (int i = 0; i < M; i++) {
-      dst[i + j * ld_dst] = src[i * ld_src + j];
-    }
-  } // for each output row
-}
-
-void transpose_simd(
-    int M,
-    int N,
-    const float* src,
-    int ld_src,
-    float* dst,
-    int ld_dst) {
-  if ((M == 1 && ld_dst == 1) || (N == 1 && ld_src == 1)) {
-    if (dst != src) {
-      memcpy(dst, src, M * N * sizeof(float));
-    }
-    return;
-  }
-  static const auto iset = fbgemmInstructionSet();
-  // Run time CPU detection
-  if (isZmm(iset)) {
-    internal::transpose_avx512(M, N, src, ld_src, dst, ld_dst);
-  } else if (isYmm(iset)) {
-    internal::transpose_avx2(M, N, src, ld_src, dst, ld_dst);
-  } else {
-    transpose_ref(M, N, src, ld_src, dst, ld_dst);
-  }
-}
-
-} // namespace fbgemm
diff --git a/src/Utils.cc b/src/Utils.cc
index 51fa2a8..266da6f 100644
--- a/src/Utils.cc
+++ b/src/Utils.cc
@@ -18,6 +18,7 @@
 #include <stdexcept>
 #include <unordered_map>
 #include <unordered_set>
+#include "./TransposeUtils.h"
 
 namespace fbgemm {
 
@@ -169,6 +170,44 @@ template void printMatrix<int32_t>(
     size_t ld,
     std::string name);
 
+void transpose_ref(
+    int M,
+    int N,
+    const float* src,
+    int ld_src,
+    float* dst,
+    int ld_dst) {
+  for (int j = 0; j < N; j++) {
+    for (int i = 0; i < M; i++) {
+      dst[i + j * ld_dst] = src[i * ld_src + j];
+    }
+  } // for each output row
+}
+
+void transpose_simd(
+    int M,
+    int N,
+    const float* src,
+    int ld_src,
+    float* dst,
+    int ld_dst) {
+  if ((M == 1 && ld_dst == 1) || (N == 1 && ld_src == 1)) {
+    if (dst != src) {
+      memcpy(dst, src, M * N * sizeof(float));
+    }
+    return;
+  }
+  static const auto iset = fbgemmInstructionSet();
+  // Run time CPU detection
+  if (isZmm(iset)) {
+    internal::transpose_avx512(M, N, src, ld_src, dst, ld_dst);
+  } else if (isYmm(iset)) {
+    internal::transpose_avx2(M, N, src, ld_src, dst, ld_dst);
+  } else {
+    transpose_ref(M, N, src, ld_src, dst, ld_dst);
+  }
+}
+
 namespace {
 inst_set_t g_forced_isa = inst_set_t::anyarch;
 bool g_Avx512_Ymm_enabled = false;