use mkl sparse matrix to improve performance (apache#14492)

* use mkl sparse matrix to improve performance

* fix build fail issue

* add 3rdparty/sparse matrix  in Makefile

* add macro for variable

* fix lib not find error

* fix gpu R test error

* fix Mac build error

* add lib/libsparse_matrix.so to CI

* fix indentation

* retrigger CI

3rdparty/sparse-matrix/Makefile

@@ -0,0 +1,21 @@
+CC = g++
+C = gcc
+MKLROOT = /opt/intel/mkl
+
+ifneq ($(USE_INTEL_PATH),)
+	MKLROOT = $(USE_INTEL_PATH)/mkl
+endif
+
+CFLAGS  = -fpic -O2 -I/opt/intel/mkl/include -c -Wall -Werror  -DMKL_ILP64 -m64  -std=c++11 
+LDFLAGS =  -Wl,--start-group -L${MKLROOT}/../compiler/lib/intel64 ${MKLROOT}/lib/intel64/libmkl_intel_ilp64.a ${MKLROOT}/lib/intel64/libmkl_intel_thread.a ${MKLROOT}/lib/intel64/libmkl_core.a -Wl,--end-group -liomp5 -lpthread -lm -ldl
+
+default: libsparse_matrix.so
+
+libsparse_matrix.so:  sparse_matrix.o
+	$(CC) -shared -o libsparse_matrix.so sparse_matrix.o $(LDFLAGS) 
+
+sparse_matrix.o:  sparse_matrix.cc sparse_matrix.h
+	$(CC) $(CFLAGS) sparse_matrix.cc 
+
+clean:
+	$(RM) libsparse_matrix.so *.o *~

3rdparty/sparse-matrix/sparse_matrix.cc

@@ -0,0 +1,45 @@
+#include <iostream>
+#include <string>
+#include <fstream>
+#include <mkl_spblas.h>
+#include "sparse_matrix.h"
+
+
+
+bool mkl_DotCsrDnsDns(SP_INT64* rows_start, SP_INT64* col_indx,
+	float* values, float* X, float* y,
+	int rows, int cols, int X_columns)
+{
+
+	sparse_index_base_t indexing = SPARSE_INDEX_BASE_ZERO;
+	sparse_status_t status;
+	sparse_matrix_t A = NULL;
+	sparse_layout_t layout = SPARSE_LAYOUT_ROW_MAJOR;
+	float one, zero;
+	one = (float)1.0;
+	zero = (float)0.0;
+
+	MKL_INT* rows_end = rows_start + 1;
+	status = mkl_sparse_s_create_csr(&A, indexing, rows, cols, rows_start, rows_end, col_indx, values);
+
+  if (status != SPARSE_STATUS_SUCCESS)
+  {
+    std::cout << "mkl_sparse_s_create_csr status :" << status << std::endl;
+    return false;
+  }
+	sparse_operation_t operation = SPARSE_OPERATION_NON_TRANSPOSE;
+	struct matrix_descr descrA;
+	descrA.type = SPARSE_MATRIX_TYPE_GENERAL;
+
+	status = mkl_sparse_s_mm(operation, one, A, descrA, layout, X, X_columns, X_columns, zero, y, X_columns);
+  if (status != SPARSE_STATUS_SUCCESS)
+  {
+    std::cout << "mkl_sparse_s_create_csr status :" << status << std::endl;
+    return false;
+  }
+
+	mkl_sparse_destroy(A);
+
+	return true;
+
+}

3rdparty/sparse-matrix/sparse_matrix.h

@@ -0,0 +1,48 @@
+#ifndef MXNET_OPERATOR_SPARSE_MATRIX_INL_H_
+#define MXNET_OPERATOR_SPARSE_MATRIX_INL_H_
+
+
+#if (!defined(__INTEL_COMPILER)) & defined(_MSC_VER)
+#define SP_INT64 __int64
+#define SP_UINT64 unsigned __int64
+#else
+#define SP_INT64 long long int
+#define SP_UINT64 unsigned long long int
+#endif
+
+
+#if defined _WIN32 || defined __CYGWIN__
+  #ifdef BUILDING_DLL
+    #ifdef __GNUC__
+      #define SPM_API_PUBLIC __attribute__ ((dllexport))
+    #else
+      #define SPM_API_PUBLIC __declspec(dllexport) // Note: actually gcc seems to also supports this syntax.
+    #endif
+  #else
+    #ifdef __GNUC__
+      #define SPM_API_PUBLIC __attribute__ ((dllimport))
+    #else
+      #define SPM_API_PUBLIC __declspec(dllimport) // Note: actually gcc seems to also supports this syntax.
+    #endif
+  #endif
+  #define SPM_API_LOCAL
+#else
+  #if __GNUC__ >= 4
+    #define SPM_API_PUBLIC __attribute__ ((visibility ("default")))
+    #define SPM_API_LOCAL  __attribute__ ((visibility ("hidden")))
+  #else
+    #define SPM_API_PUBLIC
+    #define SPM_API_LOCAL
+  #endif
+#endif
+
+
+
+extern "C"
+{
+	extern SPM_API_PUBLIC bool mkl_DotCsrDnsDns(SP_INT64* rows_start, SP_INT64* col_indx,
+		float* values, float* X, float* y, int rows, int cols, int X_columns);
+
+}
+
+#endif //MXNET_OPERATOR_SPARSE_MATRIX_INL_H_

Makefile

@@ -144,6 +144,7 @@ ifeq ($(USE_MKLDNN), 1)
 	LDFLAGS += -L$(MKLDNNROOT)/lib -lmkldnn -Wl,-rpath,'$${ORIGIN}'
 endif
 
+
 # setup opencv
 ifeq ($(USE_OPENCV), 1)
 	CFLAGS += -DMXNET_USE_OPENCV=1
@@ -410,6 +411,14 @@ ifeq ($(USE_DIST_KVSTORE), 1)
 	LDFLAGS += $(PS_LDFLAGS_A)
 endif
 
+#sparse-matrix
+ifeq ($(USE_BLAS), mkl)
+	SPARSE_MATRIX_DIR =  $(ROOTDIR)/3rdparty/sparse-matrix
+	LIB_DEP += $(SPARSE_MATRIX_DIR)/libsparse_matrix.so
+	CFLAGS += -I$(SPARSE_MATRIX_DIR)
+	LDFLAGS += -L$(SPARSE_MATRIX_DIR) -lsparse_matrix -Wl,-rpath,'$${ORIGIN}'
+endif
+
 .PHONY: clean all extra-packages test lint docs clean_all rcpplint rcppexport roxygen\
 	cython2 cython3 cython cyclean
 
@@ -547,11 +556,30 @@ ifeq ($(UNAME_S), Darwin)
 endif
 endif
 
+ifeq ($(USE_BLAS), mkl)
+ifeq ($(UNAME_S), Darwin)
+	install_name_tool -change '@rpath/libsparse_matrix.dylib' '@loader_path/libsparse_matrix.dylib' $@
+endif
+endif
+
 $(PS_PATH)/build/libps.a: PSLITE
 
 PSLITE:
 	$(MAKE) CXX="$(CXX)" DEPS_PATH="$(DEPS_PATH)" -C $(PS_PATH) ps
 
+ifeq ($(USE_BLAS), mkl)
+$(SPARSE_MATRIX_DIR)/libsparse_matrix.so: SPARSE_MATRIX
+
+SPARSE_MATRIX:
+ifeq ($(USE_INTEL_PATH), NONE)
+	$(MAKE) -C $(SPARSE_MATRIX_DIR)
+else
+	$(MAKE) -C $(SPARSE_MATRIX_DIR) USE_INTEL_PATH=$(USE_INTEL_PATH)
+endif
+	mkdir -p $(ROOTDIR)/lib
+	cp $(SPARSE_MATRIX_DIR)/libsparse_matrix.so $(ROOTDIR)/lib/
+endif
+
 $(DMLC_CORE)/libdmlc.a: DMLCCORE
 
 DMLCCORE:
@@ -628,6 +656,10 @@ rpkg:
 		cp -rf lib/libmklml_intel.so R-package/inst/libs; \
 	fi
 
+	if [ -e "lib/libsparse_matrix.so" ]; then \
+		cp -rf lib/libsparse_matrix.so R-package/inst/libs; \
+	fi
+
 	mkdir -p R-package/inst/include
 	cp -rl include/* R-package/inst/include
 	Rscript -e "if(!require(devtools)){install.packages('devtools', repo = 'https://cloud.r-project.org/')}"
@@ -673,6 +705,7 @@ clean: rclean cyclean $(EXTRA_PACKAGES_CLEAN)
 	(cd scala-package && mvn clean) || true
 	cd $(DMLC_CORE); $(MAKE) clean; cd -
 	cd $(PS_PATH); $(MAKE) clean; cd -
+	cd $(SPARSE_MATRIX_DIR); $(MAKE) clean; cd -
 	cd $(NNVM_PATH); $(MAKE) clean; cd -
 	cd $(AMALGAMATION_PATH); $(MAKE) clean; cd -
 	$(RM) -r  $(patsubst %, %/*.d, $(EXTRA_OPERATORS)) $(patsubst %, %/*/*.d, $(EXTRA_OPERATORS))
@@ -683,6 +716,7 @@ clean: rclean mkldnn_clean cyclean testclean $(EXTRA_PACKAGES_CLEAN)
 	(cd scala-package && mvn clean) || true
 	cd $(DMLC_CORE); $(MAKE) clean; cd -
 	cd $(PS_PATH); $(MAKE) clean; cd -
+	cd $(SPARSE_MATRIX_DIR); $(MAKE) clean; cd -
 	cd $(NNVM_PATH); $(MAKE) clean; cd -
 	cd $(AMALGAMATION_PATH); $(MAKE) clean; cd -
 endif

ci/jenkins/Jenkins_steps.groovy

@@ -33,7 +33,7 @@ mx_cmake_lib = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/dmlc-core/li
 // mxnet cmake libraries, in cmake builds we do not produce a libnvvm static library by default.
 mx_cmake_lib_debug = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests'
 mx_cmake_mkldnn_lib = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests, build/3rdparty/openmp/runtime/src/libomp.so, build/3rdparty/mkldnn/src/libmkldnn.so.0'
-mx_mkldnn_lib = 'lib/libmxnet.so, lib/libmxnet.a, lib/libiomp5.so, lib/libmkldnn.so.0, lib/libmklml_intel.so, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a'
+mx_mkldnn_lib = 'lib/libmxnet.so, lib/libmxnet.a, lib/libiomp5.so, lib/libmkldnn.so.0, lib/libmklml_intel.so, lib/libsparse_matrix.so, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a'
 mx_tensorrt_lib = 'build/libmxnet.so, lib/libnvonnxparser_runtime.so.0, lib/libnvonnxparser.so.0, lib/libonnx_proto.so, lib/libonnx.so'
 mx_lib_cpp_examples = 'lib/libmxnet.so, lib/libmxnet.a, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a, 3rdparty/ps-lite/build/libps.a, deps/lib/libprotobuf-lite.a, deps/lib/libzmq.a, build/cpp-package/example/*'
 mx_lib_cpp_examples_cpu = 'build/libmxnet.so, build/cpp-package/example/*'

src/operator/tensor/dot-inl.h

@@ -38,7 +38,9 @@
 #ifdef __CUDACC__
 #include "./dot-inl.cuh"
 #endif  // __CUDACC__
-
+#if (MSHADOW_USE_MKL == 1)
+#include "sparse_matrix.h"
+#endif
 namespace mxnet {
 namespace op {
 
@@ -775,13 +777,35 @@ inline void DotCsrDnsDnsImpl(const OpContext& ctx,
   }
 
   using nnvm::dim_t;
-
+#if (MSHADOW_USE_MKL == 1)
+  TShape lhs_shape = lhs.shape();
+  TShape rhs_shape = rhs.shape_;
+#endif
   const TBlob data_l = lhs.data();
   const TBlob indptr_l = lhs.aux_data(csr::kIndPtr);
   const TBlob col_idx_l = lhs.aux_data(csr::kIdx);
   const TBlob& data_r = rhs;
   const TBlob data_out = *ret;
 
+#if (MSHADOW_USE_MKL == 1)
+  if (data_l.type_flag_ == mshadow::kFloat32
+    && indptr_l.type_flag_ == mshadow::kInt64
+    && col_idx_l.type_flag_ == mshadow::kInt64
+    && !trans_lhs) {
+    bool ret = mkl_DotCsrDnsDns(static_cast<SP_INT64*>(indptr_l.dptr_),
+                                static_cast<SP_INT64*>(col_idx_l.dptr_),
+                                data_l.dptr<float>(),
+                                data_r.dptr<float>(),
+                                data_out.dptr<float>(),
+                                lhs_shape[0],
+                                lhs_shape[1],
+                                rhs_shape[1]);
+    if (ret) {
+      return;
+    }
+  }
+#endif
+
   MSHADOW_SGL_DBL_TYPE_SWITCH(data_l.type_flag_, DType, {  // data type
     MSHADOW_IDX_TYPE_SWITCH(indptr_l.type_flag_, IType, {  // indptr type
       MSHADOW_IDX_TYPE_SWITCH(col_idx_l.type_flag_, CType, {  // col idx type