Enable tvm_op for ci

CMakeLists.txt

@@ -761,7 +761,7 @@ if(USE_TVM_OP)
   add_custom_command(TARGET mxnet POST_BUILD
     COMMAND ${CMAKE_COMMAND} -E env
           PYTHONPATH="${CMAKE_CURRENT_SOURCE_DIR}/3rdparty/tvm/python:${CMAKE_CURRENT_SOURCE_DIR}/3rdparty/tvm/topi/python:${CMAKE_CURRENT_SOURCE_DIR}/contrib"
-          LD_LIBRARY_PATH="${CMAKE_CURRENT_BINARY_DIR}/3rdparty/tvm"
+          LD_LIBRARY_PATH=${CMAKE_CURRENT_BINARY_DIR}:${CMAKE_CURRENT_BINARY_DIR}/3rdparty/tvm:$ENV{LD_LIBRARY_PATH}
           ${Python3_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/contrib/tvmop/compile.py -o${CMAKE_CURRENT_BINARY_DIR}/libtvmop.so
     )
 endif()

Makefile

@@ -463,6 +463,20 @@ OBJ = $(patsubst %.cc, build/%.o, $(SRC))
 CUSRC = $(wildcard src/*/*/*/*.cu src/*/*/*.cu src/*/*.cu src/*.cu)
 CUOBJ = $(patsubst %.cu, build/%_gpu.o, $(CUSRC))
 
+ifeq ($(USE_TVM_OP), 1)
+LIB_DEP += lib/libtvm_runtime.so lib/libtvmop.so
+CFLAGS += -I$(TVM_PATH)/include -DMXNET_USE_TVM_OP=1
+LDFLAGS += -L$(ROOTDIR)/lib -ltvm_runtime -Wl,-rpath,'$${ORIGIN}'
+
+TVM_USE_CUDA := OFF
+ifeq ($(USE_CUDA), 1)
+	TVM_USE_CUDA := ON
+	ifneq ($(USE_CUDA_PATH), NONE)
+		TVM_USE_CUDA := $(USE_CUDA_PATH)
+	endif
+endif
+endif
+
 # extra operators
 ifneq ($(EXTRA_OPERATORS),)
 	EXTRA_SRC = $(wildcard $(patsubst %, %/*.cc, $(EXTRA_OPERATORS)) $(patsubst %, %/*/*.cc, $(EXTRA_OPERATORS)))
@@ -600,34 +614,23 @@ $(DMLC_CORE)/libdmlc.a: DMLCCORE
 DMLCCORE:
 	+ cd $(DMLC_CORE); $(MAKE) libdmlc.a USE_SSE=$(USE_SSE) config=$(ROOTDIR)/$(config); cd $(ROOTDIR)
 
-ifeq ($(USE_TVM_OP), 1)
-LIB_DEP += lib/libtvm_runtime.so lib/libtvmop.so
-CFLAGS += -I$(TVM_PATH)/include -DMXNET_USE_TVM_OP=1
-LDFLAGS += -L$(TVM_PATH)/build -ltvm_runtime
-
-TVM_USE_CUDA := OFF
-ifeq ($(USE_CUDA), 1)
-	TVM_USE_CUDA := ON
-	ifneq ($(USE_CUDA_PATH), NONE)
-		TVM_USE_CUDA := $(USE_CUDA_PATH)
-	endif
-endif
 lib/libtvm_runtime.so:
 	echo "Compile TVM"
 	[ -e $(LLVM_PATH)/bin/llvm-config ] || sh $(ROOTDIR)/contrib/tvmop/prepare_tvm.sh; \
 	cd $(TVM_PATH)/build; \
 	cmake -DUSE_LLVM="$(LLVM_PATH)/bin/llvm-config" \
 		  -DUSE_SORT=OFF -DUSE_CUDA=$(TVM_USE_CUDA) -DUSE_CUDNN=OFF ..; \
 	$(MAKE) VERBOSE=1; \
+	mkdir -p $(ROOTDIR)/lib; \
 	cp $(TVM_PATH)/build/libtvm_runtime.so $(ROOTDIR)/lib/libtvm_runtime.so; \
+	ls $(ROOTDIR)/lib; \
 	cd $(ROOTDIR)
 
 lib/libtvmop.so: lib/libtvm_runtime.so $(wildcard contrib/tvmop/*/*.py contrib/tvmop/*.py)
 	echo "Compile TVM operators"
-	PYTHONPATH=$(TVM_PATH)/python:$(TVM_PATH)/topi/python:$(ROOTDIR)/contrib:$PYTHONPATH \
-		LD_LIBRARY_PATH=lib \
+	PYTHONPATH=$(TVM_PATH)/python:$(TVM_PATH)/topi/python:$(ROOTDIR)/contrib \
+		LD_LIBRARY_PATH=$(ROOTDIR)/lib \
 	    python3 $(ROOTDIR)/contrib/tvmop/compile.py -o $(ROOTDIR)/lib/libtvmop.so
-endif
 
 NNVM_INC = $(wildcard $(NNVM_PATH)/include/*/*.h)
 NNVM_SRC = $(wildcard $(NNVM_PATH)/src/*/*/*.cc $(NNVM_PATH)/src/*/*.cc $(NNVM_PATH)/src/*.cc)
@@ -703,6 +706,10 @@ rpkg:
 		cp -rf lib/libmklml_intel.so R-package/inst/libs; \
 	fi
 
+	if [ -e "lib/libtvm_runtime.so" ]; then \
+		cp -rf lib/libtvm_runtime.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/')}"

ci/docker/Dockerfile.build.ubuntu_gpu_cu101

@@ -80,3 +80,4 @@ RUN /work/ubuntu_adduser.sh
 COPY runtime_functions.sh /work/
 
 WORKDIR /work/mxnet
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/compat

ci/docker/install/ubuntu_python.sh

@@ -31,4 +31,4 @@ python3 get-pip.py
 python2 get-pip.py
 
 pip2 install nose cpplint==1.3.0 'numpy>1.16.0,<2.0.0' nose-timer 'requests<2.19.0,>=2.18.4' h5py==2.8.0rc1 scipy==1.0.1 boto3 Cython==0.29.7
-pip3 install nose cpplint==1.3.0 pylint==2.3.1 'numpy>1.16.0,<2.0.0' nose-timer 'requests<2.19.0,>=2.18.4' h5py==2.8.0rc1 scipy==1.0.1 boto3 Cython==0.29.7
+pip3 install nose cpplint==1.3.0 pylint==2.3.1 'numpy>1.16.0,<2.0.0' nose-timer 'requests<2.19.0,>=2.18.4' h5py==2.8.0rc1 scipy==1.0.1 boto3 Cython==0.29.7 decorator

ci/docker/runtime_functions.sh

@@ -370,6 +370,7 @@ build_ubuntu_cpu_openblas() {
     build_ccache_wrappers
     make \
         DEV=1                         \
+        USE_TVM_OP=1                  \
         USE_CPP_PACKAGE=1             \
         USE_BLAS=openblas             \
         USE_MKLDNN=0                  \
@@ -389,6 +390,7 @@ build_ubuntu_cpu_mkl() {
         DEV=1                         \
         USE_CPP_PACKAGE=1             \
         USE_BLAS=mkl                  \
+        USE_TVM_OP=1                  \
         USE_MKLDNN=0                  \
         USE_INTEL_PATH=/opt/intel     \
         USE_DIST_KVSTORE=1            \
@@ -405,6 +407,8 @@ build_ubuntu_cpu_cmake_debug() {
         -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
         -DCMAKE_C_COMPILER_LAUNCHER=ccache \
         -DUSE_CUDA=OFF \
+        -DUSE_TVM_OP=ON \
+        -DPython3_EXECUTABLE=/usr/bin/python3 \
         -DUSE_MKL_IF_AVAILABLE=OFF \
         -DUSE_OPENMP=OFF \
         -DUSE_OPENCV=ON \
@@ -547,6 +551,7 @@ build_ubuntu_cpu_mkldnn() {
     make  \
         DEV=1                         \
         USE_CPP_PACKAGE=1             \
+        USE_TVM_OP=1                  \
         USE_BLAS=openblas             \
         USE_SIGNAL_HANDLER=1          \
         -j$(nproc)
@@ -560,6 +565,7 @@ build_ubuntu_cpu_mkldnn_mkl() {
     make  \
         DEV=1                         \
         USE_CPP_PACKAGE=1             \
+        USE_TVM_OP=1                  \
         USE_BLAS=mkl                  \
         USE_SIGNAL_HANDLER=1          \
         -j$(nproc)
@@ -642,6 +648,7 @@ build_ubuntu_gpu_mkldnn() {
         USE_CUDA=1                                \
         USE_CUDA_PATH=/usr/local/cuda             \
         USE_CUDNN=1                               \
+        USE_TVM_OP=1                              \
         CUDA_ARCH="$CI_CUDA_COMPUTE_CAPABILITIES" \
         USE_SIGNAL_HANDLER=1                      \
         -j$(nproc)
@@ -658,6 +665,7 @@ build_ubuntu_gpu_mkldnn_nocudnn() {
         USE_CUDA=1                                \
         USE_CUDA_PATH=/usr/local/cuda             \
         USE_CUDNN=0                               \
+        USE_TVM_OP=1                              \
         CUDA_ARCH="$CI_CUDA_COMPUTE_CAPABILITIES" \
         USE_SIGNAL_HANDLER=1                      \
         -j$(nproc)
@@ -673,6 +681,7 @@ build_ubuntu_gpu_cuda101_cudnn7() {
         USE_CUDA=1                                \
         USE_CUDA_PATH=/usr/local/cuda             \
         USE_CUDNN=1                               \
+        USE_TVM_OP=1                              \
         USE_CPP_PACKAGE=1                         \
         USE_DIST_KVSTORE=1                        \
         CUDA_ARCH="$CI_CUDA_COMPUTE_CAPABILITIES" \
@@ -713,6 +722,8 @@ build_ubuntu_gpu_cmake_mkldnn() {
         -DUSE_SIGNAL_HANDLER=ON                 \
         -DUSE_CUDA=1                            \
         -DUSE_CUDNN=1                           \
+        -DUSE_TVM_OP=1                          \
+        -DPython3_EXECUTABLE=/usr/bin/python3   \
         -DUSE_MKLML_MKL=1                       \
         -DCMAKE_BUILD_TYPE=Release              \
         -DCUDA_ARCH_NAME=Manual                 \
@@ -737,6 +748,8 @@ build_ubuntu_gpu_cmake() {
         -DUSE_SIGNAL_HANDLER=ON                 \
         -DUSE_CUDA=ON                           \
         -DUSE_CUDNN=ON                          \
+        -DUSE_TVM_OP=ON                         \
+        -DPython3_EXECUTABLE=/usr/bin/python3   \
         -DUSE_MKL_IF_AVAILABLE=OFF              \
         -DUSE_MKLML_MKL=OFF                     \
         -DUSE_MKLDNN=OFF                        \
@@ -782,6 +795,8 @@ build_ubuntu_gpu_large_tensor() {
         -DUSE_SIGNAL_HANDLER=ON                 \
         -DUSE_CUDA=ON                           \
         -DUSE_CUDNN=ON                          \
+        -DUSE_TVM_OP=ON                         \
+        -DPython3_EXECUTABLE=/usr/bin/python3   \
         -DUSE_MKL_IF_AVAILABLE=OFF              \
         -DUSE_MKLML_MKL=OFF                     \
         -DUSE_MKLDNN=OFF                        \

ci/jenkins/Jenkins_steps.groovy

@@ -23,22 +23,22 @@
 utils = load('ci/Jenkinsfile_utils.groovy')
 
 // mxnet libraries
-mx_lib = 'lib/libmxnet.so, lib/libmxnet.a, libsample_lib.so, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a'
-mx_lib_cython = 'lib/libmxnet.so, lib/libmxnet.a, libsample_lib.so, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a, python/mxnet/_cy2/*.so, python/mxnet/_cy3/*.so'
+mx_lib = 'lib/libmxnet.so, lib/libmxnet.a, lib/libtvm_runtime.so, lib/libtvmop.so, libsample_lib.so, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a'
+mx_lib_cython = 'lib/libmxnet.so, lib/libmxnet.a, lib/libtvm_runtime.so, lib/libtvmop.so, libsample_lib.so, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a, python/mxnet/_cy2/*.so, python/mxnet/_cy3/*.so'
 
 // Python wheels
 mx_pip = 'build/*.whl'
 
 // mxnet cmake libraries, in cmake builds we do not produce a libnvvm static library by default.
-mx_cmake_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'
-mx_cmake_lib_cython = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests, build/3rdparty/openmp/runtime/src/libomp.so, python/mxnet/_cy2/*.so, python/mxnet/_cy3/*.so'
+mx_cmake_lib = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/tvm/libtvm_runtime.so, build/libtvmop.so, build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests, build/3rdparty/openmp/runtime/src/libomp.so'
+mx_cmake_lib_cython = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/tvm/libtvm_runtime.so, build/libtvmop.so, build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests, build/3rdparty/openmp/runtime/src/libomp.so, python/mxnet/_cy2/*.so, python/mxnet/_cy3/*.so'
 // 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/libsample_lib.so, 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, libsample_lib.so, lib/libiomp5.so, lib/libmkldnn.so.0, lib/libmklml_intel.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, libsample_lib.so, 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/*, python/mxnet/_cy2/*.so, python/mxnet/_cy3/*.so'
-mx_lib_cpp_examples_cpu = 'build/libmxnet.so, build/cpp-package/example/*'
+mx_cmake_lib_debug = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/tvm/libtvm_runtime.so, build/libtvmop.so, build/libsample_lib.so, build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests'
+mx_cmake_mkldnn_lib = 'build/libmxnet.so, build/libmxnet.a, build/3rdparty/tvm/libtvm_runtime.so, build/libtvmop.so, 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/libtvm_runtime.so, lib/libtvmop.so, libsample_lib.so, lib/libiomp5.so, lib/libmkldnn.so.0, lib/libmklml_intel.so, 3rdparty/dmlc-core/libdmlc.a, 3rdparty/tvm/nnvm/lib/libnnvm.a'
+mx_tensorrt_lib = 'build/libmxnet.so, build/3rdparty/tvm/libtvm_runtime.so, build/libtvmop.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, lib/libtvm_runtime.so, lib/libtvmop.so, libsample_lib.so, 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/*, python/mxnet/_cy2/*.so, python/mxnet/_cy3/*.so'
+mx_lib_cpp_examples_cpu = 'build/libmxnet.so, build/3rdparty/tvm/libtvm_runtime.so, build/libtvmop.so, build/cpp-package/example/*'
 
 // Python unittest for CPU
 // Python 2

scala-package/assembly/src/main/assembly/assembly.xml

@@ -54,6 +54,7 @@
       <directory>${MXNET_DIR}/lib</directory>
       <includes>
         <include>libmxnet.so</include>
+        <include>libtvm_runtime.so</include>
         <include>libgfortran.so.3</include>
         <include>libquadmath.so.0</include>
         <include>libiomp5.so</include>

scala-package/core/src/main/scala/org/apache/mxnet/util/NativeLibraryLoader.scala

@@ -86,6 +86,7 @@ private[mxnet] object NativeLibraryLoader {
     logger.debug(s"Attempting to load $loadLibname")
     val libFileInJar = libPathInJar + loadLibname
     saveLibraryToTemp("libmxnet.so", "/lib/native/libmxnet.so", true)
+    saveLibraryToTemp("libtvm_runtime.so", "/lib/native/libtvm_runtime.so", false)
     saveLibraryToTemp("libgfortran.so.3", "/lib/native/libgfortran.so.3", false)
     saveLibraryToTemp("libquadmath.so.0", "/lib/native/libquadmath.so.0", false)
     saveLibraryToTemp("libiomp5.so", "/lib/native/libiomp5.so", false)

src/operator/contrib/tvmop/ufunc.cc

@@ -96,7 +96,7 @@ void TVMBinaryBackwardComputeUseNone(const nnvm::NodeAttrs& attrs,
     for (int i = ov.size(); i < max_dim; ++i) {
       ov.push_back(1);
     }
-    for (int i = flag; i < ov.size(); i += 2) {
+    for (uint32_t i = flag; i < ov.size(); i += 2) {
       iv.push_back(ov[i]);
     }
     TShape oshape(ov.begin(), ov.end()), ishape(iv.begin(), iv.end());

tests/python/unittest/test_tvm_op.py

@@ -67,7 +67,6 @@ def test_tvm_broadcast_add():
             assert same(a.grad.asnumpy(), expected_grad_a)
             assert same(b.grad.asnumpy(), expected_grad_b)
 
-
 if __name__ == '__main__':
     import nose
     nose.runmodule()