Make FakeLowP tests work (#36525)

Summary:
Make the e2e FakeLowP python tests work with Glow lowering in OSS environment. Added a README.md as a guideline.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36525

Reviewed By: hyuen

Differential Revision: D21004706

Pulled By: yinghai

fbshipit-source-id: d182152e4a1a3368640bd7872cb9ea4d4bff4b02

caffe2/contrib/fakelowp/CMakeLists.txt

@@ -1,5 +1,5 @@
 if(USE_FAKELOWP)
-  message(STATUS "Including FAKELOWP operators")
+  message(STATUS "Including FakeLowP operators")
 
   # ---[ CPU files.
   file(GLOB_RECURSE tmp *.cc)
@@ -11,32 +11,22 @@ if(USE_FAKELOWP)
   # We will only build the perf kernel files if the compiler supports avx2
   # extensions.
   if(CAFFE2_COMPILER_SUPPORTS_AVX2_EXTENSIONS)
-    add_library(Caffe2_fakelowp_ops STATIC ${FAKELOWP_CPU_SRCS})  
-    add_dependencies(Caffe2_fakelowp_ops fbgemm Caffe2_PROTO c10)
-    target_include_directories(Caffe2_fakelowp_ops BEFORE
+    add_library(caffe2_fakelowp_ops OBJECT ${FAKELOWP_CPU_SRCS})  
+    add_dependencies(caffe2_fakelowp_ops fbgemm Caffe2_PROTO c10)
+    target_include_directories(caffe2_fakelowp_ops BEFORE
       PRIVATE $<BUILD_INTERFACE:${FBGEMM_SOURCE_DIR}/include>)
-    if(MSVC AND NOT "${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
-      target_compile_options(Caffe2_fakelowp_ops
-          PRIVATE "/arch:AVX2"
-          PRIVATE "/D__FMA__"
-          PRIVATE "/D__F16C__")
+    if(MSVC)
+      set_property(SOURCE ${FAKELOWP_CPU_SRCS}
+        APPEND_STRING PROPERTY COMPILE_FLAGS " /arch:AVX2 ")
     else()
-      target_compile_options(Caffe2_fakelowp_ops
-          PRIVATE "-mavx2"
-          PRIVATE "-mfma"
-          PRIVATE "-mavx"
-          PRIVATE "-mf16c")
+      set_property(SOURCE ${FAKELOWP_CPU_SRCS}
+        APPEND_STRING PROPERTY COMPILE_FLAGS " -mavx2 -mfma -mf16c -mxsave ")
     endif()
-    caffe2_interface_library(
-      Caffe2_fakelowp_ops Caffe2_fakelowp_ops_interface)
-    list(APPEND
-      Caffe2_DEPENDENCY_WHOLE_LINK_LIBS
-      "Caffe2_fakelowp_ops_interface")
+    set(Caffe2_CPU_SRCS ${Caffe2_CPU_SRCS}
+      $<TARGET_OBJECTS:caffe2_fakelowp_ops>)
   endif()
+  # ---[ Send the lists to the parent scope.
+  set(Caffe2_CPU_SRCS ${Caffe2_CPU_SRCS} PARENT_SCOPE)
 else()
   message(STATUS "Excluding FakeLowP operators")
 endif()
-
-set(Caffe2_DEPENDENCY_WHOLE_LINK_LIBS
-    ${Caffe2_DEPENDENCY_WHOLE_LINK_LIBS}
-    PARENT_SCOPE)

caffe2/contrib/fakelowp/test/README.md

@@ -0,0 +1,41 @@
+# How to run FakeLowP vs Glow tests
+This was tested on Ubuntu 16.04 LTS but should work in general Linux system.
+
+## Install and Build Glow
+Follow https://github.com/pytorch/glow/blob/master/README.md to install the dependency of Glow. Then at glow root run
+```
+mkdir build && cd build
+cmake -G Ninja -DGLOW_BUILD_ONNXIFI_DYNLIB=ON ..
+ninja all
+```
+Note that here you probably want to add other flags like `-DGLOW_WITH_NNPI=1` to enable specific backend if you have the flow setup.
+Once built successfully, you will get an dynamic library at `build/lib/Onnxifi/libonnxifi.so`. We will use it later.
+
+## Install and Build PyTorch
+Follow https://github.com/pytorch/pytorch/blob/master/README.md to install the dependency of PyTorch. It might be easy to 
+setup a python virtualenv or conda. And please use Python > 3.5.2 because hypothesis library will expose a bug in Python which 
+is fixed after 3.5.2. Something like 3.7 mighr be good enough. Here I give a virtualenv flow:
+```
+python3.7 -m venv venv3
+source venv3/bin/active
+cd pytorch
+pip install -r requirements.txt
+```
+You probably need to install gflags-dev too with
+```
+sudo apt-get install libgflags-dev
+```
+
+Once you have all the dependency libs installed, build PyTorch with FakeLowP op support
+```
+USE_CUDA=0 USE_ROCM=0 USE_FAKELOWP=ON DEBUG=1 CMAKE_BUILD_TYPE=Debug USE_GFLAGS=1 USE_GLOG=1 USE_MKLDNN=0 python setup.py install
+```
+The key options here are `USE_FAKELOWP=ON` which enables building of FakeLowP operators and `USE_GFLAGS=1` which enables gflags as we 
+use gflags in Glow to pass options. Other flags are mostl for fast build time and debug purpose. 
+
+## Run the test
+You can now run the tests with command like the following  when you are inside the virtual python env:
+```
+OSS_ONNXIFI_LIB=${PATH_TO_GLOW}/build/lib/Onnxifi/libonnxifi.so python pytorch/caffe2/contrib/fakelowp/test/test_sls_nnpi_fp16.py
+```
+

caffe2/contrib/fakelowp/test/test_sls_nnpi_fp16.py

@@ -1,6 +1,7 @@
 from __future__ import absolute_import, division, print_function, unicode_literals
 
 import time
+import unittest
 
 # Must happen before importing caffe2.python.*
 import caffe2.python.fakelowp.init_shared_libs  # noqa
@@ -8,7 +9,6 @@ import numpy as np
 from caffe2.proto import caffe2_pb2
 from caffe2.python import core, workspace
 from caffe2.python.onnx.onnxifi import onnxifi_caffe2_net
-from caffe2.python.onnx.tests.test_utils import TestCase
 from caffe2.python.fakelowp.test_utils import print_test_debug_info
 
 
@@ -24,8 +24,8 @@ GLOW_MATMUL_ATOL = 1e-5
 GLOW_MATMUL_RTOL = 1e-3
 
 
-class SparseLengthsSumTest(TestCase):
-    def Test_SLS_NonQuantized_fp16(self):
+class SparseLengthsSumTest(unittest.TestCase):
+    def Skip_test_SLS_NonQuantized_fp16(self):
         N = 20000
         DIM = 64
         D = (4 * np.random.random_sample((N, DIM)) + 1).astype(np.float32)
@@ -614,3 +614,6 @@ class SparseLengthsSumTest(TestCase):
                 },
             )
             assert 0
+
+if __name__ == '__main__':
+    unittest.main()

caffe2/python/fakelowp/init_shared_libs.py

@@ -1 +1,10 @@
 from __future__ import absolute_import, division, print_function, unicode_literals
+
+import ctypes
+import os
+
+if 'OSS_ONNXIFI_LIB' in os.environ:
+    lib = os.environ['OSS_ONNXIFI_LIB']
+    print("Loading ONNXIFI lib: ".format(lib))
+    ctypes.CDLL(lib, ctypes.RTLD_GLOBAL)
+

cmake/Dependencies.cmake

@@ -1229,6 +1229,7 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
   if(ONNX_ML)
     add_definitions(-DONNX_ML=1)
   endif()
+  add_definitions(-DONNXIFI_ENABLE_EXT=1)
   # Add op schemas in "ai.onnx.pytorch" domain
   add_subdirectory("${CMAKE_CURRENT_LIST_DIR}/../caffe2/onnx/torch_ops")
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/../third_party/onnx EXCLUDE_FROM_ALL)