saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-02 03:55:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Make CentOS 6 CUDA build and run (#2159)

Browse source 
* Add manylinux1 source code changes

* Disable a python test
This commit is contained in:
Dmitri Smirnov 2019-10-19 15:33:31 -07:00 committed by Changming Sun
parent 96b33f4597
commit acec4b446f
18 changed files with 161 additions and 86 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
onnxruntime/core/graph/record.h
View file

@ -32,7 +32,7 @@ class Record {
values_ = other.values_;
}
Status GetName(int index, const std::string** pp_name) const {
Status GetName(size_t index, const std::string** pp_name) const {
if (nullptr == pp_name || index >= names_.size()) {
return Status(ONNXRUNTIME, common::INVALID_ARGUMENT);
}

4
onnxruntime/core/providers/cpu/math/matmul_helper.h
View file

@ -239,7 +239,7 @@ class MatMulComputeHelper {
template <typename T>
static void OffsetToArrays(T* p, const std::vector<size_t>& offsets, gsl::span<T*> arrays) {
auto len = offsets.size();
ORT_ENFORCE(arrays.size() == gsl::narrow_cast<ptrdiff_t>(len));
ORT_ENFORCE(arrays.size() == len);
for (size_t i = 0; i < len; i++) {
arrays[i] = p + offsets[i];
}
@ -248,7 +248,7 @@ class MatMulComputeHelper {
template <typename T>
static void OffsetToArrays(const T* p, const std::vector<size_t>& offsets, gsl::span<const T*> arrays) {
auto len = offsets.size();
ORT_ENFORCE(arrays.size() == gsl::narrow_cast<ptrdiff_t>(len));
ORT_ENFORCE(arrays.size() == len);
for (size_t i = 0; i < len; i++) {
arrays[i] = p + offsets[i];
}

4
onnxruntime/core/providers/cuda/nn/conv.cc
View file

@ -114,7 +114,7 @@ Status Conv<T>::ComputeInternal(OpKernelContext* context) const {
std::vector<int64_t> b_dims(2 + kernel_shape.size());
b_dims[0] = 1; // N
b_dims[1] = b_shape[0]; // C
for (int i = 0; i < kernel_shape.size(); i++)
for (size_t i = 0; i < kernel_shape.size(); i++)
b_dims[2 + i] = 1;
ORT_RETURN_IF_ERROR(s_.b_tensor.Set(b_dims, CudnnTensor::GetDataType<CudaT>()));
@ -212,7 +212,7 @@ Status CudnnConvolutionDescriptor::Set(
std::vector<int> pad_dims(rank);
std::vector<int> stride_dims(rank);
std::vector<int> dilation_dims(rank);
for (int i = 0; i < rank; i++) {
for (size_t i = 0; i < rank; i++) {
pad_dims[i] = gsl::narrow_cast<int>(pads[i]);
stride_dims[i] = gsl::narrow_cast<int>(strides[i]);
dilation_dims[i] = gsl::narrow_cast<int>(dilations[i]);

2
onnxruntime/core/providers/cuda/tensor/compress.cc
View file

@ -61,7 +61,7 @@ Status Compress::ComputeInternal(OpKernelContext* ctx) const {
int64_t axis_right_stride = 1;
if (has_axis_) {
for (int i = static_cast<int>(axis_ + 1); i < rank; ++i) {
for (auto i = static_cast<size_t>(axis_ + 1); i < rank; ++i) {
axis_right_stride *= input_dimensions[i];
}
}

2
onnxruntime/core/providers/cuda/tensor/concat.cc
View file

@ -40,7 +40,7 @@ Status Concat::ComputeInternal(OpKernelContext* ctx) const {
}
}
std::vector<int64_t> concat_sizes_range(concat_sizes);
for (int i = 1; i < concat_sizes_range.size(); ++i) {
for (size_t i = 1; i < concat_sizes_range.size(); ++i) {
concat_sizes_range[i] += concat_sizes_range[i - 1];
}

4
onnxruntime/core/providers/cuda/tensor/expand.cc
View file

@ -28,7 +28,7 @@ Status Expand::ComputeInternal(OpKernelContext* ctx) const {
auto input_shape = input0.Shape().GetDims();
// pad input_dims with 1 to make ranks match
for (int i = 0; i < rank - input_shape.size(); i++) {
for (size_t i = 0; i < rank - input_shape.size(); i++) {
input_shape.insert(input_shape.begin(), 1);
}
@ -41,7 +41,7 @@ Status Expand::ComputeInternal(OpKernelContext* ctx) const {
auto out_span = fdm_output_dims.CpuSpan();
auto sdm_span = fdm_output_subdim_size.CpuSpan();
auto subdim_size = output_shape.Size();
for (auto i = 0; i < rank; i++) {
for (size_t i = 0; i < rank; i++) {
in_span[i] = fast_divmod(static_cast<int>(input_shape[i]));
out_span[i] = fast_divmod(static_cast<int>(output_shape[i]));
// output_shape[i] won't be 0 here, it's covered in (0 == output_shape.Size())

6
onnxruntime/core/providers/cuda/tensor/slice.cc
View file

@ -70,14 +70,14 @@ Status Slice<Tind, dynamic>::ComputeInternal(OpKernelContext* ctx) const {
}
CudaAsyncBuffer<int64_t> starts_buffer(this, dimension_count);
gsl::span<int64_t> starts_buffer_span = starts_buffer.CpuSpan();
for (int i = 0; i < dimension_count; ++i) {
for (size_t i = 0; i < dimension_count; ++i) {
starts_buffer_span[i] = starts[i];
}
starts_buffer.CopyToGpu();
CudaAsyncBuffer<int64_t> steps_buffer(this, dimension_count);
gsl::span<int64_t> steps_buffer_span = steps_buffer.CpuSpan();
for (int i = 0; i < dimension_count; ++i) {
for (size_t i = 0; i < dimension_count; ++i) {
steps_buffer_span[i] = steps[i];
}
steps_buffer.CopyToGpu();
@ -90,7 +90,7 @@ Status Slice<Tind, dynamic>::ComputeInternal(OpKernelContext* ctx) const {
CudaAsyncBuffer<fast_divmod> div_strides(this, dimension_count);
gsl::span<fast_divmod> div_strides_span = div_strides.CpuSpan();
for (int i = 0; i < dimension_count; ++i) {
for (size_t i = 0; i < dimension_count; ++i) {
div_strides_span[i] = fast_divmod(gsl::narrow_cast<int>(output_pitches[i]));
}
div_strides.CopyToGpu();

2
onnxruntime/core/providers/cuda/tensor/split.cc
View file

@ -63,7 +63,7 @@ Status Split::ComputeInternal(OpKernelContext* ctx) const {
split_sizes_gpu.CopyToGpu();
std::vector<int64_t> split_sizes_range(split_sizes);
for (int i = 1; i < split_sizes_range.size(); ++i) {
for (size_t i = 1; i < split_sizes_range.size(); ++i) {
split_sizes_range[i] += split_sizes_range[i - 1];
}
CudaAsyncBuffer<int64_t> split_sizes_range_gpu(this, split_sizes_range);

2
onnxruntime/core/providers/cuda/tensor/tile.cc
View file

@ -36,7 +36,7 @@ Status Tile<T>::ComputeInternal(OpKernelContext* ctx) const {
auto* repeats = repeats_tensor.template Data<int64_t>();
const auto& input_shape = input_tensor.Shape().GetDims();
std::vector<int64_t> output_dims(input_shape);
for (auto axis = 0; axis < rank; axis++)
for (size_t axis = 0; axis < rank; axis++)
output_dims[axis] *= repeats[axis];
TensorShape outputShape(output_dims);
auto& output_tensor = *ctx->Output(0, outputShape);

4
onnxruntime/core/providers/cuda/tensor/upsample.cc
View file

@ -70,7 +70,7 @@ Status Upsample<T>::BaseCompute(OpKernelContext* context, const std::vector<floa
CudaAsyncBuffer<fast_divmod> output_div_pitches(this, rank);
gsl::span<fast_divmod> div_strides_span = output_div_pitches.CpuSpan();
for (int i = 0; i < rank; ++i) {
for (size_t i = 0; i < rank; ++i) {
input_stride_span[i] = input_pitches[i];
div_strides_span[i] = fast_divmod(gsl::narrow_cast<int>(output_pitches[i]));
}
@ -95,7 +95,7 @@ Status Upsample<T>::BaseCompute(OpKernelContext* context, const std::vector<floa
CudaAsyncBuffer<fast_divmod> scales_div(this, rank);
gsl::span<fast_divmod> scales_div_span = scales_div.CpuSpan();
for (int i = 0; i < rank; ++i) {
for (size_t i = 0; i < rank; ++i) {
scales_div_span[i] = fast_divmod(gsl::narrow_cast<int>(ceil(scales[i])));
}
scales_div.CopyToGpu();

76
onnxruntime/test/ir/graph_test.cc
View file

@ -114,16 +114,17 @@ TEST(GraphTraversalTest, ReverseDFS) {
Model model("graph_1");
auto& graph = model.MainGraph();
// Case 1: A normal graph.
// SouceNode
// / \
// node_1 (Variable) node_2 (Variable)
// \ /
// node_3 (Add)
// |
// node_4 (NoOp)
// |
// SinkNode
/* Case 1: A normal graph.
* SouceNode
* / \
* node_1 (Variable) node_2 (Variable)
* \ /
* node_3 (Add)
* |
* node_4 (NoOp)
* |
* SinkNode
*/
std::vector<NodeArg*> inputs;
std::vector<NodeArg*> outputs;
@ -267,16 +268,17 @@ TEST(ResolvingGraphTest, GraphConstruction_CheckIsAcyclic) {
Model model("graph_1");
auto& graph = model.MainGraph();
// A normal graph.
// SouceNode
// / \
// node_1 (Variable) node_2 (Variable)
// \ /
// node_3 (Add)
// |
// node_4 (NoOp)
// |
// SinkNode
/* A normal graph.
* SouceNode
* / \
* node_1 (Variable) node_2 (Variable)
* \ /
* node_3 (Add)
* |
* node_4 (NoOp)
* |
* SinkNode
*/
std::vector<NodeArg*> inputs;
std::vector<NodeArg*> outputs;
@ -445,14 +447,15 @@ TEST(ResolvingGraphTest, GraphConstruction_CheckGraphInputOutputOrderMaintained)
map.insert({std::to_string(i), i});
}
// | |
// b (Identity) a (Identity) values
// \ /
// c (Merge)
// |
// d (Split)
// / \
// 1 .. 10
/* | |
* b (Identity) a (Identity) values
* \ /
* c (Merge)
* |
* d (Split)
* / \
* 1 .. 10
*/
TypeProto tensor_int32;
tensor_int32.mutable_tensor_type()->set_elem_type(TensorProto_DataType_INT32);
tensor_int32.mutable_tensor_type()->mutable_shape()->add_dim()->set_dim_value(1);
@ -653,14 +656,15 @@ TEST(ResolvingGraphTest, GraphConstruction_TypeInference) {
Model model("graph_1");
auto& graph = model.MainGraph();
// Case 1: A normal graph.
// SourceNode
// / | \
// node_1 (Variable) node_2 (Variable) node_3 (Variable)
// \ | / (it's all 3 nodes above outputs to the one input of node_4)
// node_4 (Max)
// |
// SinkNode
/* Case 1: A normal graph.
* SourceNode
* / | \
* node_1 (Variable) node_2 (Variable) node_3 (Variable)
* \ | / (it's all 3 nodes above outputs to the one input of node_4)
* node_4 (Max)
* |
* SinkNode
*/
std::vector<NodeArg*> inputs;
std::vector<NodeArg*> outputs;

2
onnxruntime/test/providers/provider_test_utils.cc
View file

@ -259,7 +259,7 @@ void Check<TensorSeq>(const OpTester::Data& expected_data, const TensorSeq& outp
// now check the contents of the tensors
auto null_deleter = [](void*) {};
for (int i = 0; i < output_num_tensors; ++i) {
for (size_t i = 0; i < output_num_tensors; ++i) {
OrtValue temp_value;
// Reason for null_deleter: we don't want the tensor destructor to be called as part of this OrtValue destructor
// as we're creating this OrtValue only to reuse the Check functionality

2
onnxruntime/test/providers/provider_test_utils.h
View file

@ -481,7 +481,7 @@ class OpTester {
ptr->dtype = DataTypeImpl::GetType<T>();
auto num_tensors = seq_tensors.tensors.size();
ptr->tensors.resize(num_tensors);
for (int i = 0; i < num_tensors; ++i) {
for (size_t i = 0; i < num_tensors; ++i) {
TensorShape shape{seq_tensors.tensors[i].shape};
auto values_count = static_cast<int64_t>(seq_tensors.tensors[i].data.size());
ORT_ENFORCE(shape.Size() == values_count, values_count,

26
onnxruntime/test/python/onnxruntime_test_python.py
View file

@ -302,32 +302,6 @@ class TestInferenceSession(unittest.TestCase):
['identity', 'test\x00\x00\x00\x00']], dtype=object)
np.testing.assert_equal(expr, res[0])
def testConvAutoPad(self):
sess = onnxrt.InferenceSession(self.get_name("conv_autopad.onnx"))
x = np.array(25 * [1.0], dtype=np.float32).reshape((1, 1, 5, 5))
x_name = sess.get_inputs()[0].name
self.assertEqual(x_name, "Input4")
x_shape = sess.get_inputs()[0].shape
self.assertEqual(x_shape, [1, 1, 5, 5])
x_type = sess.get_inputs()[0].type
self.assertEqual(x_type, 'tensor(float)')
output_name = sess.get_outputs()[0].name
self.assertEqual(output_name, "Convolution5_Output_0")
output_shape = sess.get_outputs()[0].shape
self.assertEqual(output_shape, [1, 1, 5, 5])
output_type = sess.get_outputs()[0].type
self.assertEqual(output_type, 'tensor(float)')
res = sess.run([output_name], {x_name: x})
output_expected = np.array([[[[24., 33., 33., 33., 20.],
[27., 36., 36., 36., 21.],
[27., 36., 36., 36., 21.],
[27., 36., 36., 36., 21.],
[12., 15., 15., 15., 8.]]]], dtype=np.float32)
np.testing.assert_allclose(output_expected, res[0])
def testZipMapStringFloat(self):
sess = onnxrt.InferenceSession(
self.get_name("zipmap_stringfloat.onnx"))

70
tools/ci_build/github/linux/docker/Dockerfile.centos6_gpu Normal file
View file

@ -0,0 +1,70 @@
# FROM mcr.microsoft.com/dotnet-buildtools/prereqs:centos-7-50f0d02-20190918214028
FROM centos:6
ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
ARG PYTHON_VERSION
ADD scripts /tmp/scripts
RUN cd /tmp/scripts && /tmp/scripts/install_centos.sh
ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:/bin:/usr/bin:/usr/local/bin:/opt/rh/devtoolset-7/root/usr/bin:{PATH}
RUN /tmp/scripts/install_deps.sh -p $PYTHON_VERSION && rm -rf /tmp/scripts
#Below are copied from https://gitlab.com/nvidia/container-images/cuda/tree/master/dist/centos6
RUN NVIDIA_GPGKEY_SUM=d1be581509378368edeec8c1eb2958702feedf3bc3d17011adbf24efacce4ab5 && \
curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64/7fa2af80.pub | sed '/^Version/d' > /etc/pki/rpm-gpg/RPM-GPG-KEY-NVIDIA && \
echo "$NVIDIA_GPGKEY_SUM /etc/pki/rpm-gpg/RPM-GPG-KEY-NVIDIA" | sha256sum -c -
COPY cuda_manylinux2010.repo /etc/yum.repos.d/cuda.repo
ENV CUDA_VERSION 10.0.130
ENV CUDA_PKG_VERSION 10-0-$CUDA_VERSION-1
RUN yum install -y \
cuda-cudart-$CUDA_PKG_VERSION \
cuda-libraries-$CUDA_PKG_VERSION \
cuda-nvtx-$CUDA_PKG_VERSION \
cuda-nvml-dev-$CUDA_PKG_VERSION \
cuda-command-line-tools-$CUDA_PKG_VERSION \
cuda-libraries-dev-$CUDA_PKG_VERSION \
cuda-minimal-build-$CUDA_PKG_VERSION \
&& \
ln -s cuda-10.0 /usr/local/cuda && \
rpm -e --nodeps gcc gcc-c++ && \
rm -rf /var/cache/yum/*
# nvidia-docker 1.0
RUN echo "/usr/local/nvidia/lib" >> /etc/ld.so.conf.d/nvidia.conf && \
echo "/usr/local/nvidia/lib64" >> /etc/ld.so.conf.d/nvidia.conf
ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:${PATH}
ENV LD_LIBRARY_PATH /usr/local/nvidia/lib:/usr/local/nvidia/lib64:${LD_LIBRARY_PATH}
# nvidia-container-runtime
ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
ENV NVIDIA_REQUIRE_CUDA "cuda>=10.0 brand=tesla,driver>=384,driver<385 brand=tesla,driver>=410,driver<411"
ENV LIBRARY_PATH /usr/local/cuda/lib64/stubs
ENV CUDNN_VERSION 7.6.4.38
LABEL com.nvidia.cudnn.version="${CUDNN_VERSION}"
# cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
RUN CUDNN_DOWNLOAD_SUM=417bb5daf51377037eb2f5c87649000ca1b9cec0acb16cfe07cb1d3e9a961dbf && \
curl -fsSL http://developer.download.nvidia.com/compute/redist/cudnn/v7.6.4/cudnn-10.0-linux-x64-v7.6.4.38.tgz -O && \
echo "$CUDNN_DOWNLOAD_SUM cudnn-10.0-linux-x64-v7.6.4.38.tgz" | sha256sum -c - && \
tar --no-same-owner -xzf cudnn-10.0-linux-x64-v7.6.4.38.tgz -C /usr/local && \
rm cudnn-10.0-linux-x64-v7.6.4.38.tgz && \
/sbin/ldconfig
ARG BUILD_UID=1000
ARG BUILD_USER=onnxruntimedev
RUN /usr/sbin/adduser --uid $BUILD_UID $BUILD_USER
WORKDIR /home/$BUILD_USER
USER $BUILD_USER

27
tools/ci_build/github/linux/docker/scripts/install_centos.sh
View file

@ -7,16 +7,33 @@ os_major_version=$(cat /etc/redhat-release | tr -dc '0-9.'|cut -d \. -f1)
if ! rpm -q --quiet epel-release ; then
yum install -y https://dl.fedoraproject.org/pub/epel/epel-release-latest-$os_major_version.noarch.rpm
fi
echo "installing for os major version : $os_major_version"
if [ "$os_major_version" == "5" ]; then
#Be careful, don't pull gcc into the base system, because we already have one in /opt/rh/devtoolset-2/root/usr/bin
yum install -y redhat-lsb expat-devel libcurl-devel tar unzip curl zlib-devel make python2-devel icu rsync bzip2 git bzip2-devel
#Be careful, don't pull gcc into the base system, because we already have one in /opt/rh/devtoolset-2/root/usr/bin
yum install -y redhat-lsb expat-devel libcurl-devel tar unzip curl zlib-devel make python2-devel icu rsync bzip2 git bzip2-devel
elif [ "$os_major_version" == "6" ] && [ ! -d "/opt/python/cp35-cp35m" ]; then
yum install -y centos-release-scl
yum repolist
yum install -y redhat-lsb-core expat-devel libcurl-devel tar unzip curl zlib-devel make libunwind icu aria2 rsync bzip2 git bzip2-devel
yum upgrade -y
yum install -y \
ccache \
devtoolset-7-binutils \
devtoolset-7-gcc \
devtoolset-7-gcc-c++ \
devtoolset-7-gcc-gfortran
# The way to get python 3.6.8
yum install -y https://centos6.iuscommunity.org/ius-release.rpm
yum --enablerepo=ius install -y python36u python36u-devel python36u-pip python36u-numpy python36u-setuptools python36u-wheel protobuf
/usr/bin/python3.6 -m pip install --upgrade pip
else
yum install -y redhat-lsb-core expat-devel libcurl-devel tar unzip curl zlib-devel make python2-devel libunwind icu aria2 rsync bzip2 git bzip2-devel
yum install -y redhat-lsb-core expat-devel libcurl-devel tar unzip curl zlib-devel make python2-devel libunwind icu aria2 rsync bzip2 git bzip2-devel
fi
#If the /opt/python folder exists, we assume this is the manylinux docker image
if [ ! -d "/opt/python/cp35-cp35m" ]
if [ "$os_major_version" != "6" ] && [ ! -d "/opt/python/cp35-cp35m" ]
then
yum install -y ccache gcc gcc-c++ python3-devel python3-pip python3-numpy python3-setuptools python3-wheel
yum install -y ccache gcc gcc-c++ python3 python3-devel python3-pip python3-numpy python3-setuptools python3-wheel
fi

1
tools/ci_build/github/linux/docker/scripts/install_deps.sh
View file

@ -1,6 +1,7 @@
#!/bin/bash
set -e
while getopts p:d: parameter_Option
do case "${parameter_Option}"
in

11
tools/ci_build/github/linux/docker/scripts/install_onnx.sh
View file

@ -25,6 +25,8 @@ else
PYTHON_EXE="/usr/bin/python${PYTHON_VER}"
fi
${PYTHON_EXE} -m pip install protobuf
version2tag=(5af210ca8a1c73aa6bae8754c9346ec54d0a756e-onnx123
bae6333e149a59a3faa9c4d9c44974373dcf5256-onnx130
9e55ace55aad1ada27516038dfbdc66a8a0763db-onnx141
@ -46,7 +48,14 @@ for v2t in ${version2tag[*]}; do
if [ ! -d "third_party/pybind11/pybind11" ]; then
git clone https://github.com/pybind/pybind11.git third_party/pybind11
fi
${PYTHON_EXE} -m pip install .
# We need to make the adjustment only for CentOS6 OR we substitue this only for
# ${PYTHON_EXE} where we'd need to escape slashes
# Make sure we do not hit pyhon2 as on CentOS 6 it does not work
ESCAPED_PY=$(echo "${PYTHON_EXE}" | sed 's/\//\\\//g')
sed "1,1 s/\/usr\/bin\/env python/$ESCAPED_PY/" /tmp/src/onnx-$onnx_version/tools/protoc-gen-mypy.py > /tmp/src/onnx-$onnx_version/tools/repl_protoc-gen-mypy.py
chmod a+w /tmp/src/onnx-$onnx_version/tools/protoc-gen-mypy.py
mv /tmp/src/onnx-$onnx_version/tools/repl_protoc-gen-mypy.py /tmp/src/onnx-$onnx_version/tools/protoc-gen-mypy.py
mkdir -p /data/onnx/${onnx_tag}
${PYTHON_EXE} -m pip install .
backend-test-tools generate-data -o /data/onnx/$onnx_tag
done