fix(cuda): use cudnn8.0.4 as cu111 default libs

GitOrigin-RevId: 721ca73bae696b06872614fb095f5282dbb36662

dnn/test/cuda/conv_bias.cpp

@@ -216,7 +216,7 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_QS8) {
     }
 }
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST_F(CUDA, CONV_BIAS_NCHW_QS8) {
     //! not support NonlineMode::SIGMOID and NonlineMode::H_SWISH
     require_compute_capability(6, 1);

imperative/CMakeLists.txt

@@ -63,6 +63,7 @@ add_custom_command(
     TARGET ${MODULE_NAME} POST_BUILD
     COMMAND ${CMAKE_COMMAND} -E copy ${PROJECT_SOURCE_DIR}/LICENSE ${PROJECT_SOURCE_DIR}/ACKNOWLEDGMENTS ${PROJECT_BINARY_DIR}
     COMMAND ${CMAKE_COMMAND} -E remove -f ${CMAKE_CURRENT_SOURCE_DIR}/python/megengine/core/$<TARGET_FILE_NAME:${MODULE_NAME}> # clean develop
+    COMMAND ${CMAKE_COMMAND} -E remove -f ${CMAKE_CURRENT_SOURCE_DIR}/python/megengine/version.py # clean develop
     COMMAND ${CMAKE_COMMAND} -E copy_directory ${CMAKE_CURRENT_SOURCE_DIR}/python/megengine ${CMAKE_CURRENT_BINARY_DIR}/python/megengine
     COMMAND ${CMAKE_COMMAND} -E copy_directory ${CMAKE_CURRENT_SOURCE_DIR}/python/test ${CMAKE_CURRENT_BINARY_DIR}/python/test
     COMMAND ${CMAKE_COMMAND} -E copy ${CMAKE_CURRENT_SOURCE_DIR}/python/setup.py ${CMAKE_CURRENT_BINARY_DIR}/python/setup.py

scripts/whl/BUILD_PYTHON_WHL_README.md

@@ -74,7 +74,7 @@ ALL_PYTHON="36m" ./scripts/whl/manylinux2014/build_wheel_common.sh -sdk cu101
 
 * If you just want to build with cpu only version, you can set `BUILD_WHL_CPU_ONLY` environment 'ON'. eg:
 ```bash
-BUILD_WHL_CPU_ONLY="ON" ALL_PYTHON="36m" ./scripts/whl/manylinux2014/build_wheel_common.sh -sdk cu101
+ALL_PYTHON="36m" ./scripts/whl/manylinux2014/build_wheel_common.sh -sdk cpu
 ```
 
 ## Build for MacOS

scripts/whl/manylinux2014/build_wheel_common.sh

@@ -9,16 +9,16 @@ local_path=$(dirname $(readlink -f $0))
 CUDNN_LIB_DIR="/opt/cudnn/lib64/"
 CUDA_LIB_DIR="/usr/local/cuda/lib64/"
 
-CUDA_SDK="unknown"
+SDK_NAME="unknown"
 function usage() {
-    echo "use '-sdk cu111' to specify cuda toolkit config, also support cu101, cu112"
+    echo "use '-sdk cu111' to specify cuda toolkit config, also support cu101, cu112, cpu"
 }
 
 while [ "$1" != "" ]; do
     case $1 in
         -sdk)
            shift
-           CUDA_SDK=$1
+           SDK_NAME=$1
            shift
            ;;
         *)
@@ -27,17 +27,16 @@ while [ "$1" != "" ]; do
     esac
 done
 
-echo "Build with ${CUDA_SDK}"
+echo "Build with ${SDK_NAME}"
 
-if [ $CUDA_SDK == "cu101" ];then
+if [ $SDK_NAME == "cu101" ];then
     COPY_LIB_LIST="${CUDA_LIB_DIR}/libnvrtc.so.10.1"
     EXTRA_CMAKE_FLAG=" -DMGE_WITH_CUDNN_SHARED=OFF" 
-    OUT_DIR="cu101"  
     BUILD_GCC8="ON"  
     REQUIR_CUDA_VERSION="10010" 
     REQUIR_CUDNN_VERSION="7.6.3" 
     REQUIR_TENSORRT_VERSION="6.0.1.5" 
-elif [ $CUDA_SDK == "cu111" ];then
+elif [ $SDK_NAME == "cu111" ];then
     COPY_LIB_LIST="\
         ${CUDA_LIB_DIR}/libnvrtc.so.11.1:\
         ${CUDA_LIB_DIR}/libcublasLt.so.11:\
@@ -56,11 +55,10 @@ elif [ $CUDA_SDK == "cu111" ];then
         arch=compute_80,code=sm_80 \
         arch=compute_86,code=sm_86 \
         arch=compute_86,code=compute_86" 
-    OUT_DIR="cu111"  
     REQUIR_CUDA_VERSION="11010" 
-    REQUIR_CUDNN_VERSION="8.0.5" 
+    REQUIR_CUDNN_VERSION="8.0.4" 
     REQUIR_TENSORRT_VERSION="7.2.2.3" 
-elif [ $CUDA_SDK == "cu112" ];then
+elif [ $SDK_NAME == "cu112" ];then
     COPY_LIB_LIST="\
         ${CUDA_LIB_DIR}/libnvrtc.so.11.2:\
         ${CUDA_LIB_DIR}/libcublasLt.so.11:\
@@ -79,16 +77,17 @@ elif [ $CUDA_SDK == "cu112" ];then
         arch=compute_80,code=sm_80 \
         arch=compute_86,code=sm_86 \
         arch=compute_86,code=compute_86"  
-    OUT_DIR="cu112"  
     REQUIR_CUDA_VERSION="11020" 
-    REQUIR_CUDNN_VERSION="8.0.5" 
+    REQUIR_CUDNN_VERSION="8.0.4" 
     REQUIR_TENSORRT_VERSION="7.2.2.3" 
+elif [ $SDK_NAME == "cpu" ];then
+    echo "use $SDK_NAME without cuda support"
+    BUILD_WHL_CPU_ONLY="ON"
 else
-    echo "no support sdk ${CUDA_SDK}, please set by '-sdk cu111'"
+    echo "no support sdk ${SDK_NAME}, please set by '-sdk cu111'"
     exit -1
 fi
 
-BUILD_WHL_CPU_ONLY=${BUILD_WHL_CPU_ONLY}
 if [[ -z ${BUILD_WHL_CPU_ONLY} ]]
 then
     BUILD_WHL_CPU_ONLY="OFF"
@@ -205,7 +204,7 @@ docker run --rm -it $TMPFS_ARGS \
     -e ALL_PYTHON="${ALL_PYTHON}" \
     -e EXTRA_CMAKE_FLAG="$EXTRA_CMAKE_FLAG" \
     -e COPY_LIB_LIST="$COPY_LIB_LIST"  \
-    -e OUT_DIR="$OUT_DIR"  \
+    -e SDK_NAME="$SDK_NAME"  \
     -v ${CUDA_ROOT_DIR}:/usr/local/cuda \
     -v ${CUDNN_ROOT_DIR}:/opt/cudnn \
     -v ${TENSORRT_ROOT_DIR}:/opt/tensorrt \

scripts/whl/manylinux2014/do_build_common.sh

@@ -119,13 +119,13 @@ do
     if [ ${USE_AUDITWHEEL} = "ON" ]; then
         LD_LIBRARY_PATH=${BUILD_DIR}/dnn/cuda-stub:$LD_LIBRARY_PATH auditwheel repair -L ${NEW_LIB_PATH} ${BUILD_DIR}/staging/dist/Meg*.whl
     else
-        mkdir -p ${SRC_DIR}/scripts/whl/manylinux2014/output/wheelhouse/${OUT_DIR}
+        mkdir -p ${SRC_DIR}/scripts/whl/manylinux2014/output/wheelhouse/${SDK_NAME}
         cd ${BUILD_DIR}/staging/dist/
         org_whl_name=`ls Meg*${ver}*.whl`
         compat_whl_name=`echo ${org_whl_name} | sed 's/linux/manylinux2014/'`
         echo "org whl name: ${org_whl_name}"
         echo "comapt whl name: ${compat_whl_name}"
-        mv ${org_whl_name} ${SRC_DIR}/scripts/whl/manylinux2014/output/wheelhouse/${OUT_DIR}/${compat_whl_name}
+        mv ${org_whl_name} ${SRC_DIR}/scripts/whl/manylinux2014/output/wheelhouse/${SDK_NAME}/${compat_whl_name}
         cd /home/output
     fi
     chown -R ${UID}.${UID} .

src/gopt/test/inference.cpp

@@ -1836,7 +1836,7 @@ TEST(TestEnableTensorCore, SmallInputShape) {
 }
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestEnableTensorCore, Nchw4Nchw) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -1936,7 +1936,7 @@ TEST(TestEnableTensorCore, Nchw4Nchw) {
 #endif
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestEnableTensorCore, ConvBiasWithZ) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -2002,7 +2002,7 @@ TEST(TestEnableTensorCore, ConvBiasWithZ) {
 #endif
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestEnableTensorCore, Pooling) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -2148,7 +2148,7 @@ TEST(TestGoptInference, EnableTensorCore) {
 }
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(FuseConvBiasZPass, BlockFuse) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -2273,7 +2273,7 @@ TEST(FuseConvBiasZPass, BlockFuse) {
 #endif
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestEnableTensorCore, ShuffleMerge) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -2450,7 +2450,7 @@ TEST(FuseConvBiasZPass, Basic) {
 
 #if MGB_CUDA
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestGoptInference, EnableCHWN4) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -2552,7 +2552,7 @@ TEST(TestGoptInference, EnableCHWN4) {
 #endif
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestGoptInference, EnableCHWN4WarpPespective) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -2731,7 +2731,7 @@ TEST(TestGoptInference, EnableCHWN4Pooling) {
 }
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestGoptInference, EnableCHWN4ShuffleRemove) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -2857,7 +2857,7 @@ TEST(TestGoptInference, EnableCHWN4ShuffleRemove) {
 #endif
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestGoptInference, ConvertFormatNCHW4GPU) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -3076,7 +3076,7 @@ TEST(TestGoptInference, ConvertFormatNCHW4) {
 }
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestGoptInference, ConvertFormatNCHW4Ic3) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");
@@ -3945,7 +3945,7 @@ TEST(TestGoptInference, FoldingConvDimshuffle) {
 }
 
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestGoptInference, FoldingConvDimshuffleNCHW4NCHW32) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");

src/tensorrt/test/opr_replace.cpp

@@ -1978,7 +1978,7 @@ TEST(TestTensorRTReplace, FuseConvAdd) {
     MGB_ASSERT_TENSOR_NEAR(outputs[1], outputs[3], 1e-3);
 }
 //! close for cu111 ci, reopen it when bug fixed
-#if 0
+#if CUDA_VERSION < 11000
 TEST(TestTensorRTReplace, FuseConvAddNchw2nchw4) {
     REQUIRE_GPU(1);
     auto cn = CompNode::load("gpu0");