removing parallel build for OSS (pytorch#509)

Summary:
Pull Request resolved: pytorch#509

1) Removing parallel build to see if we overcome g++ internal compiler error for oss Ubuntu builds
2) remove adding masm=intel to all files in bazel build. Add it only to inline asm files.

Reviewed By: shz0116

Differential Revision: D26255490

fbshipit-source-id: 2adff04b5d8f0bc393cbfce2e0ec02c09d137a86

.github/workflows/fbgemmci.yml

@@ -54,7 +54,7 @@ jobs:
         mkdir build_static
         cd build_static
         cmake -DUSE_SANITIZER=address -DFBGEMM_LIBRARY_TYPE=static ..
-        make -j
+        make
 
     - name: Test static FBGEMM lib
       if: contains(runner.os, 'linux')   # not run on macos-latest now due to supporting AVX2
@@ -69,7 +69,7 @@ jobs:
         mkdir build_shared
         cd build_shared
         cmake -DUSE_SANITIZER=address -DFBGEMM_LIBRARY_TYPE=shared ..
-        make -j
+        make
 
     - name: Test shared FBGEMM lib
       if: contains(runner.os, 'linux')   # not run on macos-latest now due to supporting AVX2

BUILD.bazel

@@ -5,7 +5,7 @@
 
 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_avx2_srcs", "get_fbgemm_inline_avx2_srcs", "get_fbgemm_avx512_srcs", "get_fbgemm_inline_avx512_srcs", "get_fbgemm_base_srcs", "get_fbgemm_generic_srcs", "get_fbgemm_public_headers", "get_fbgemm_tests")
 
 cc_library(
     name = "fbgemm_base",
@@ -32,7 +32,9 @@ cc_library(
     ],
     deps = [
         ":fbgemm_avx2",
+        ":fbgemm_inline_avx2",
         ":fbgemm_avx512",
+        ":fbgemm_inline_avx512",
         ":fbgemm_base",
         ":fbgemm_headers",
     ],
@@ -43,6 +45,23 @@ cc_library(
     name = "fbgemm_avx2",
     srcs = get_fbgemm_avx2_srcs(),
     hdrs = glob(["src/*.h"]),
+    copts = [
+        "-m64",
+        "-mavx2",
+        "-mfma",
+        "-mf16c",
+    ],
+    deps = [
+        ":fbgemm_base",
+        ":fbgemm_headers",
+    ],
+    linkstatic = 1,
+)
+
+cc_library(
+    name = "fbgemm_inline_avx2",
+    srcs = get_fbgemm_inline_avx2_srcs(),
+    hdrs = glob(["src/*.h"]),
     copts = [
         "-m64",
         "-mavx2",
@@ -61,6 +80,25 @@ cc_library(
     name = "fbgemm_avx512",
     srcs = get_fbgemm_avx512_srcs(),
     hdrs = glob(["src/*.h"]),
+    copts = [
+        "-m64",
+        "-mfma",
+        "-mavx512f",
+        "-mavx512bw",
+        "-mavx512dq",
+        "-mavx512vl",
+    ],
+    deps = [
+        ":fbgemm_base",
+        ":fbgemm_headers",
+    ],
+    linkstatic = 1,
+)
+
+cc_library(
+    name = "fbgemm_inline_avx512",
+    srcs = get_fbgemm_inline_avx512_srcs(),
+    hdrs = glob(["src/*.h"]),
     copts = [
         "-m64",
         "-mfma",

CMakeLists.txt

@@ -71,12 +71,17 @@ endif()
 # Define file lists
 get_filelist("get_fbgemm_generic_srcs(with_base=True)" FBGEMM_GENERIC_SRCS)
 get_filelist("get_fbgemm_avx2_srcs(msvc=${MSVC_BOOL})" FBGEMM_AVX2_SRCS)
+get_filelist("get_fbgemm_inline_avx2_srcs(msvc=${MSVC_BOOL})"
+  FBGEMM_AVX2_INLINE_SRCS)
 get_filelist("get_fbgemm_avx512_srcs(msvc=${MSVC_BOOL})" FBGEMM_AVX512_SRCS)
+get_filelist("get_fbgemm_inline_avx512_srcs(msvc=${MSVC_BOOL})"
+  FBGEMM_AVX512_INLINE_SRCS)
 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})
-add_library(fbgemm_avx512 OBJECT ${FBGEMM_AVX512_SRCS})
+add_library(fbgemm_avx2 OBJECT ${FBGEMM_AVX2_SRCS} ${FBGEMM_AVX2_INLINE_SRCS})
+add_library(fbgemm_avx512 OBJECT
+  ${FBGEMM_AVX512_SRCS} ${FBGEMM_AVX512_INLINE_SRCS})
 
 # Make libraries depend on defs.bzl
 add_custom_target(defs.bzl DEPENDS defs.bzl)

defs.bzl

@@ -83,6 +83,10 @@ def get_fbgemm_avx2_srcs(msvc = False):
         "src/PackDepthwiseConvMatrixAvx2.cc",
         "src/QuantUtilsAvx2.cc",
         "src/UtilsAvx2.cc",
+    ]
+
+def get_fbgemm_inline_avx2_srcs(msvc = False):
+    return [
         #FP16 kernels contain inline assembly and inline assembly syntax for MSVC is different.
         "src/FbgemmFP16UKernelsAvx2.cc" if not msvc else "src/FbgemmFP16UKernelsIntrinsicAvx2.cc",
     ]
@@ -94,6 +98,10 @@ def get_fbgemm_avx512_srcs(msvc = False):
         "src/FbgemmFloat16ConvertAvx512.cc",
         "src/QuantUtilsAvx512.cc",
         "src/UtilsAvx512.cc",
+    ]
+
+def get_fbgemm_inline_avx512_srcs(msvc = False):
+    return [
         #FP16 kernels contain inline assembly and inline assembly syntax for MSVC is different.
         "src/FbgemmFP16UKernelsAvx512.cc" if not msvc else "src/FbgemmFP16UKernelsIntrinsicAvx512.cc",
         "src/FbgemmFP16UKernelsAvx512_256.cc" if not msvc else "src/FbgemmFP16UKernelsIntrinsicAvx512_256.cc",