mmdeploy/csrc/preprocess/cuda/pad_impl.cpp
lzhangzz 640aa03538
Support Windows (#106)
* minor changes

* support windows

* fix GCC build

* fix lint

* reformat

* fix Windows build

* fix GCC build

* search backend ops for onnxruntime

* fix lint

* fix lint

* code clean-up

* code clean-up

* fix clang build

* fix trt support

* fix cmake for ncnn

* fix cmake for openvino

* fix SDK Python API

* handle ops for other backends (ncnn, trt)

* handle SDK Python API library location

* robustify linkage

* fix cuda

* minor fix for openvino & ncnn

* use CMAKE_CUDA_ARCHITECTURES if set

* fix cuda preprocessor

* fix misc

* fix pplnn & pplcv, drop support for pplcv<0.6.0

* robustify cmake

* update build.md (#2)

* build dynamic modules as module library & fix demo (partially)

* fix candidate path for mmdeploy_python

* move "enable CUDA" to cmake config for demo

* refine demo cmake

* add comment

* fix ubuntu build

* revert docs/en/build.md

* fix C API

* fix lint

* Windows build doc (#3)

* check in docs related to mmdeploy build on windows

* update build guide on windows platform

* update build guide on windows platform

* make path of thirdparty libraries consistent

* make path consistency

* correct build command for custom ops

* correct build command for sdk

* update sdk build instructions

* update doc

* correct build command

* fix lint

* correct build command and fix lint

Co-authored-by: lvhan <lvhan@pjlab.org>

* trailing whitespace (#4)

* minor fix

* fix sr sdk model

* fix type deduction

* fix cudaFree after driver shutting down

* update ppl.cv installation warning (#5)

* fix device allocator threshold & fix lint

* update doc (#6)

* update ppl.cv installation warning

* missing 'git clone'

Co-authored-by: chenxin <chenxin2@sensetime.com>
Co-authored-by: zhangli <zhangli@sensetime.com>
Co-authored-by: lvhan028 <lvhan_028@163.com>
Co-authored-by: lvhan <lvhan@pjlab.org>
2022-02-24 20:08:44 +08:00

117 lines
4.5 KiB
C++

// Copyright (c) OpenMMLab. All rights reserved.
#include "core/utils/device_utils.h"
#include "core/utils/formatter.h"
#include "ppl/cv/cuda/copymakeborder.h"
#include "preprocess/transform/pad.h"
using namespace std;
using namespace ppl::cv::cuda;
namespace mmdeploy {
namespace cuda {
class PadImpl : public ::mmdeploy::PadImpl {
public:
explicit PadImpl(const Value& args) : ::mmdeploy::PadImpl(args) {
#if PPLCV_VERSION_MAJOR >= 0 && PPLCV_VERSION_MINOR >= 6 && PPLCV_VERSION_PATCH >= 2
map<string, ppl::cv::BorderType> border_map{{"constant", ppl::cv::BORDER_CONSTANT},
{"edge", ppl::cv::BORDER_REPLICATE},
{"reflect", ppl::cv::BORDER_REFLECT_101},
{ "symmetric",
ppl::cv::BORDER_REFLECT }};
#else
map<string, ppl::cv::BorderType> border_map{{"constant", ppl::cv::BORDER_TYPE_CONSTANT},
{"edge", ppl::cv::BORDER_TYPE_REPLICATE},
{"reflect", ppl::cv::BORDER_TYPE_REFLECT_101},
{"symmetric", ppl::cv::BORDER_TYPE_REFLECT}};
#endif
if (border_map.find(arg_.padding_mode) == border_map.end()) {
MMDEPLOY_ERROR("unsupported padding_mode '{}'", arg_.padding_mode);
throw_exception(eNotSupported);
}
padding_mode_ = border_map[arg_.padding_mode];
}
protected:
Result<Tensor> PadImage(const Tensor& img, const array<int, 4>& padding) override {
OUTCOME_TRY(auto src_tensor, MakeAvailableOnDevice(img, device_, stream_));
auto desc = src_tensor.desc();
int height = desc.shape[1];
int width = desc.shape[2];
int c = desc.shape[3];
auto dst_height = height + padding[1] + padding[3];
auto dst_width = width + padding[0] + padding[2];
TensorShape dst_shape{1, dst_height, dst_width, c};
TensorDesc dst_desc{device_, desc.data_type, dst_shape, ""};
Tensor dst_tensor(dst_desc);
ppl::common::RetCode ret = 0;
cudaStream_t stream = ::mmdeploy::GetNative<cudaStream_t>(stream_);
if (desc.data_type == DataType::kFLOAT) {
auto src_buffer = src_tensor.data<float>();
auto dst_buffer = dst_tensor.data<float>();
if (3 == c) {
ret = CopyMakeBorder<float, 3>(stream, height, width, width * c, src_buffer, dst_width * c,
dst_buffer, padding[1], padding[3], padding[0], padding[2],
padding_mode_, arg_.pad_val);
} else if (1 == c) {
ret = CopyMakeBorder<float, 1>(stream, height, width, width * c, src_buffer, dst_width * c,
dst_buffer, padding[1], padding[3], padding[0], padding[2],
padding_mode_, arg_.pad_val);
} else {
MMDEPLOY_ERROR("unsupported channels {}", c);
assert(0);
return Status(eNotSupported);
}
} else if (desc.data_type == DataType::kINT8) {
auto src_buffer = src_tensor.data<uint8_t>();
auto dst_buffer = dst_tensor.data<uint8_t>();
if (3 == c) {
ret = CopyMakeBorder<ppl::cv::uchar, 3>(
stream, height, width, width * c, src_buffer, dst_width * c, dst_buffer, padding[1],
padding[3], padding[0], padding[2], padding_mode_, (ppl::cv::uchar)arg_.pad_val);
} else if (1 == c) {
ret = CopyMakeBorder<ppl::cv::uchar, 1>(
stream, height, width, width * c, src_buffer, dst_width * c, dst_buffer, padding[1],
padding[3], padding[0], padding[2], padding_mode_, (ppl::cv::uchar)arg_.pad_val);
} else {
MMDEPLOY_ERROR("unsupported channels {}", c);
assert(0);
return Status(eNotSupported);
}
} else {
MMDEPLOY_ERROR("unsupported data type {}", desc.data_type);
assert(0);
return Status(eNotSupported);
}
if (ret != 0) {
MMDEPLOY_ERROR("unexpected exception happened");
assert(0);
return Status(eNotSupported);
}
return dst_tensor;
}
private:
ppl::cv::BorderType padding_mode_;
};
class PadCreator : public Creator<::mmdeploy::PadImpl> {
public:
const char* GetName() const override { return "cuda"; }
int GetVersion() const override { return 1; }
ReturnType Create(const Value& args) override { return make_unique<PadImpl>(args); }
};
} // namespace cuda
} // namespace mmdeploy
using ::mmdeploy::PadImpl;
using mmdeploy::cuda::PadCreator;
REGISTER_MODULE(PadImpl, PadCreator);