Runtime feature detection (#13549)

* Prototype for runtime feature detection

* Includes from diamond to quotes

* Add CPU feature and BLAS flavour flags

* Add BLAS flavour and CPU SSE and AVX flags

* MXNET_USE_LAPACK

* Fix C++ linting errors

* Expose runtime feature detection in the public C API and in the Python API

* Refactor Storage -> FeatureSet

* Refine documentation

* Add failure case

* Fix pylint

* Address CR comments

CMakeLists.txt

@@ -423,12 +423,14 @@ if(USE_OPENMP)
     endif()
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
+    add_definitions(-DMXNET_USE_OPENMP=1)
   else()
     if(OPENMP_FOUND)
       set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
       set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
       set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
+      add_definitions(-DMXNET_USE_OPENMP=1)
     endif()
   endif()
 elseif(UNIX AND NOT ANDROID)

Makefile

@@ -212,6 +212,16 @@ ifeq ($(USE_CUDNN), 1)
 	LDFLAGS += -lcudnn
 endif
 
+ifeq ($(use_blas), open)
+	CFLAGS += -DMXNET_USE_BLAS_OPEN=1
+else ifeq ($(use_blas), atlas)
+	CFLAGS += -DMXNET_USE_BLAS_ATLAS=1
+else ifeq ($(use_blas), mkl)
+	CFLAGS += -DMXNET_USE_BLAS_MKL=1
+else ifeq ($(use_blas), apple)
+	CFLAGS += -DMXNET_USE_BLAS_APPLE=1
+endif
+
 # whether to use F16C instruction set extension for fast fp16 compute on CPU
 # if cross compiling you may want to explicitly turn it off if target system does not support it
 ifndef USE_F16C

cmake/ChooseBlas.cmake

@@ -37,22 +37,26 @@ if(BLAS STREQUAL "Atlas" OR BLAS STREQUAL "atlas")
   list(APPEND mshadow_LINKER_LIBS ${Atlas_LIBRARIES})
   add_definitions(-DMSHADOW_USE_CBLAS=1)
   add_definitions(-DMSHADOW_USE_MKL=0)
+  add_definitions(-DMXNET_USE_BLAS_ATLAS=1)
 elseif(BLAS STREQUAL "Open" OR BLAS STREQUAL "open")
   find_package(OpenBLAS REQUIRED)
   include_directories(SYSTEM ${OpenBLAS_INCLUDE_DIR})
   list(APPEND mshadow_LINKER_LIBS ${OpenBLAS_LIB})
   add_definitions(-DMSHADOW_USE_CBLAS=1)
   add_definitions(-DMSHADOW_USE_MKL=0)
+  add_definitions(-DMXNET_USE_BLAS_OPEN=1)
 elseif(BLAS STREQUAL "MKL" OR BLAS STREQUAL "mkl")
   find_package(MKL REQUIRED)
   include_directories(SYSTEM ${MKL_INCLUDE_DIR})
   list(APPEND mshadow_LINKER_LIBS ${MKL_LIBRARIES})
   add_definitions(-DMSHADOW_USE_CBLAS=0)
   add_definitions(-DMSHADOW_USE_MKL=1)
+  add_definitions(-DMXNET_USE_BLAS_MKL=1)
 elseif(BLAS STREQUAL "apple")
   find_package(Accelerate REQUIRED)
   include_directories(SYSTEM ${Accelerate_INCLUDE_DIR})
   list(APPEND mshadow_LINKER_LIBS ${Accelerate_LIBRARIES})
   add_definitions(-DMSHADOW_USE_MKL=0)
   add_definitions(-DMSHADOW_USE_CBLAS=1)
+  add_definitions(-DMXNET_USE_BLAS_APPLE=1)
 endif()

include/mxnet/base.h

@@ -25,47 +25,18 @@
 #ifndef MXNET_BASE_H_
 #define MXNET_BASE_H_
 
-#include <dmlc/base.h>
-#include <dmlc/io.h>
-#include <dmlc/type_traits.h>
-#include <dmlc/parameter.h>
-#include <mshadow/tensor.h>
-// nnvm headers for symbolic construction.
-#include <nnvm/op.h>
-#include <nnvm/tuple.h>
-#include <nnvm/symbolic.h>
+#include "dmlc/base.h"
 #include <string>
+#include "dmlc/io.h"
+#include "dmlc/type_traits.h"
+#include "dmlc/parameter.h"
+#include "mshadow/tensor.h"
+// nnvm headers for symbolic construction.
+#include "nnvm/op.h"
+#include "nnvm/tuple.h"
+#include "nnvm/symbolic.h"
+#include "mxfeatures.h"
 
-/*!
- *\brief whether to use opencv support
- */
-#ifndef MXNET_USE_OPENCV
-#define MXNET_USE_OPENCV 1
-#endif
-
-/*!
- *\brief whether to use cuda support
- */
-#ifndef MXNET_USE_CUDA
-#define MXNET_USE_CUDA MSHADOW_USE_CUDA
-#endif
-
-/*!
- *\brief whether to use cudnn library for convolution
- */
-#ifndef MXNET_USE_CUDNN
-#define MXNET_USE_CUDNN MSHADOW_USE_CUDNN
-#endif
-
-/*!
- *\brief whether to use cusolver library
- */
-#ifndef MXNET_USE_CUSOLVER
-#define MXNET_USE_CUSOLVER MSHADOW_USE_CUSOLVER
-#endif
-
-/*! \brief Error message for using gpu when MXNET_USE_CUDA==0 */
-#define MXNET_GPU_NOT_ENABLED_ERROR  "GPU is not enabled"
 
 /*!
  * \brief define compatible keywords in g++
@@ -412,6 +383,7 @@ inline std::ostream& operator<<(std::ostream &out, const Context &ctx) {
 #define MXNET_DESCRIBE(...) describe(__VA_ARGS__ "\n\nFrom:" __FILE__ ":" STRINGIZE(__LINE__))
 #define ADD_FILELINE "\n\nDefined in " __FILE__ ":L" STRINGIZE(__LINE__)
 
+
 #if MXNET_USE_MKLDNN == 1
 constexpr size_t kMKLDNNAlign = 64;
 #endif

include/mxnet/c_api.h

@@ -208,6 +208,15 @@ MXNET_DLL const char *MXGetLastError();
 //-------------------------------------
 // Part 0: Global State setups
 //-------------------------------------
+
+/*!
+ * \brief
+ * \param feature to check mxfeatures.h
+ * \param out set to true if the feature is enabled, false otherwise
+ * \return 0 when success, -1 when failure happens.
+ */
+MXNET_DLL int MXHasFeature(const mx_uint feature, bool* out);
+
 /*!
  * \brief Seed all global random number generators in mxnet.
  * \param seed the random number seed.
@@ -465,6 +474,7 @@ MXNET_DLL int MXGetGPUMemoryInformation64(int dev, uint64_t *free_mem, uint64_t
  */
 MXNET_DLL int MXGetVersion(int *out);
 
+
 //-------------------------------------
 // Part 1: NDArray creation and deletion
 //-------------------------------------

include/mxnet/io.h

@@ -25,12 +25,12 @@
 #ifndef MXNET_IO_H_
 #define MXNET_IO_H_
 
-#include <dmlc/data.h>
-#include <dmlc/registry.h>
 #include <vector>
 #include <string>
 #include <utility>
 #include <queue>
+#include "dmlc/data.h"
+#include "dmlc/registry.h"
 #include "./base.h"
 #include "./ndarray.h"
 

include/mxnet/mxfeatures.h

@@ -0,0 +1,188 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2018 by Contributors
+ * \file mxfeatures.h
+ * \brief check MXNet features including compile time support
+ */
+
+#pragma once
+
+#include "dmlc/base.h"
+#include "mshadow/base.h"
+
+/*!
+ *\brief whether to use opencv support
+ */
+#ifndef MXNET_USE_OPENCV
+#define MXNET_USE_OPENCV 1
+#endif
+
+/*!
+ *\brief whether to use cuda support
+ */
+#ifndef MXNET_USE_CUDA
+#define MXNET_USE_CUDA MSHADOW_USE_CUDA
+#endif
+
+/*!
+ *\brief whether to use cudnn library for convolution
+ */
+#ifndef MXNET_USE_CUDNN
+#define MXNET_USE_CUDNN MSHADOW_USE_CUDNN
+#endif
+
+#ifndef MXNET_USE_NCCL
+#define MXNET_USE_NCCL 0
+#endif
+
+/*!
+ *\brief whether to use cusolver library
+ */
+#ifndef MXNET_USE_CUSOLVER
+#define MXNET_USE_CUSOLVER MSHADOW_USE_CUSOLVER
+#endif
+
+#ifndef MXNET_ENABLE_CUDA_RTC
+#define MXNET_ENABLE_CUDA_RTC 0
+#endif
+
+/*! \brief Error message for using gpu when MXNET_USE_CUDA==0 */
+#define MXNET_GPU_NOT_ENABLED_ERROR  "GPU is not enabled"
+
+
+#ifndef MXNET_USE_TENSORRT
+#define MXNET_USE_TENSORRT 0
+#endif
+
+
+#ifndef MXNET_USE_BLAS_ATLAS
+#define MXNET_USE_BLAS_ATLAS 0
+#endif
+
+#ifndef MXNET_USE_BLAS_OPEN
+#define MXNET_USE_BLAS_OPEN 0
+#endif
+
+#ifndef MXNET_USE_BLAS_MKL
+#define MXNET_USE_BLAS_MKL 0
+#endif
+
+#ifndef MXNET_USE_BLAS_APPLE
+#define MXNET_USE_BLAS_APPLE 0
+#endif
+
+#ifndef MXNET_USE_LAPACK
+#define MXNET_USE_LAPACK 0
+#endif
+
+#ifndef MXNET_USE_MKLDNN
+#define MXNET_USE_MKLDNN 0
+#endif
+
+#ifndef MXNET_USE_OPENMP
+#define MXNET_USE_OPENMP 0
+#endif
+
+#ifndef MXNET_USE_F16C
+#define MXNET_USE_F16C MSHADOW_USE_F16C
+#endif
+
+#ifndef MXNET_USE_CAFFE
+#define MXNET_USE_CAFFE 0
+#endif
+
+#ifndef MXNET_USE_DIST_KVSTORE
+#define MXNET_USE_DIST_KVSTORE 0
+#endif
+
+#ifndef MXNET_USE_SIGNAL_HANDLER
+#define MXNET_USE_SIGNAL_HANDLER 0
+#endif
+
+
+
+namespace mxnet {
+namespace features {
+// Check compile flags such as CMakeLists.txt
+
+/// Compile time features
+enum : uint32_t {
+  // NVIDIA, CUDA
+  CUDA = 0,
+  CUDNN,
+  NCCL,
+  CUDA_RTC,
+  TENSORRT,
+
+  // CPU Features / optimizations
+  CPU_SSE,
+  CPU_SSE2,
+  CPU_SSE3,
+  CPU_SSE4_1,
+  CPU_SSE4_2,
+  CPU_SSE4A,  // AMD extensions to SSE4
+  CPU_AVX,
+  CPU_AVX2,
+
+
+  // Multiprocessing / CPU / System
+  OPENMP,
+  SSE,
+  F16C,
+  JEMALLOC,
+
+  // Math libraries & BLAS
+  // Flavour of BLAS
+  BLAS_OPEN,
+  BLAS_ATLAS,
+  // Intel(R) Math Kernel Library
+  BLAS_MKL,
+  BLAS_APPLE,
+  // Other math libraries:
+  // Linear Algebra PACKage
+  LAPACK,
+  // Intel(R) Math Kernel Library for Deep Neural Networks
+  MKLDNN,
+
+  // Image processing
+  OPENCV,
+
+  // Misc
+  CAFFE,
+  PROFILER,
+  DIST_KVSTORE,
+  CXX14,
+  // Signal handler to print stack traces on exceptions
+  SIGNAL_HANDLER,
+  DEBUG,
+
+  // size indicator
+  MAX_FEATURES
+};
+
+
+/*!
+ * \return true if the given feature is supported
+ */
+bool is_enabled(uint32_t feat);
+
+}  // namespace features
+}  // namespace mxnet

python/mxnet/gluon/utils.py

@@ -222,7 +222,7 @@ def _replace_atomic(src, dst):
     _MOVEFILE_WRITE_THROUGH = 0x8
     _windows_default_flags = _MOVEFILE_WRITE_THROUGH
 
-    text_type = unicode if sys.version_info[0] == 2 else str  # noqa
+    text_type = unicode if sys.version_info[0] == 2 else str  # pylint: disable=undefined-variable
 
     def _str_to_unicode(x):
         """Handle text decoding. Internal use only"""