mmdeploy/csrc/preprocess/cuda/normalize_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

82 lines
2.8 KiB
C++

// Copyright (c) OpenMMLab. All rights reserved.
#include <cuda_runtime.h>
#include "core/utils/device_utils.h"
#include "core/utils/formatter.h"
#include "preprocess/transform/normalize.h"
using namespace std;
namespace mmdeploy::cuda {
template <typename T, int channels>
void Normalize(const T* src, int height, int width, int stride, float* output, const float* mean,
const float* std, bool to_rgb, cudaStream_t stream);
class NormalizeImpl : public ::mmdeploy::NormalizeImpl {
public:
explicit NormalizeImpl(const Value& args) : ::mmdeploy::NormalizeImpl(args) {}
protected:
Result<Tensor> NormalizeImage(const Tensor& tensor) override {
OUTCOME_TRY(auto src_tensor, MakeAvailableOnDevice(tensor, device_, stream_));
auto src_desc = src_tensor.desc();
int h = (int)src_desc.shape[1];
int w = (int)src_desc.shape[2];
int c = (int)src_desc.shape[3];
int stride = w * c;
TensorDesc dst_desc{device_, DataType::kFLOAT, src_desc.shape, src_desc.name};
Tensor dst_tensor{dst_desc};
auto output = dst_tensor.data<float>();
auto stream = ::mmdeploy::GetNative<cudaStream_t>(stream_);
if (DataType::kINT8 == src_desc.data_type) {
auto input = src_tensor.data<uint8_t>();
if (3 == c) {
Normalize<uint8_t, 3>(input, h, w, stride, output, arg_.mean.data(), arg_.std.data(),
arg_.to_rgb, stream);
} else if (1 == c) {
Normalize<uint8_t, 1>(input, h, w, stride, output, arg_.mean.data(), arg_.std.data(),
arg_.to_rgb, stream);
} else {
MMDEPLOY_ERROR("unsupported channels {}", c);
return Status(eNotSupported);
}
} else if (DataType::kFLOAT == src_desc.data_type) {
auto input = src_tensor.data<float>();
if (3 == c) {
Normalize<float, 3>(input, h, w, stride, output, arg_.mean.data(), arg_.std.data(),
arg_.to_rgb, stream);
} else if (1 == c) {
Normalize<float, 1>(input, h, w, stride, output, arg_.mean.data(), arg_.std.data(),
arg_.to_rgb, stream);
} else {
MMDEPLOY_ERROR("unsupported channels {}", c);
return Status(eNotSupported);
}
} else {
MMDEPLOY_ERROR("unsupported data type {}", src_desc.data_type);
assert(0);
return Status(eNotSupported);
}
return dst_tensor;
}
};
class NormalizeImplCreator : public Creator<::mmdeploy::NormalizeImpl> {
public:
const char* GetName() const override { return "cuda"; }
int GetVersion() const override { return 1; }
std::unique_ptr<::mmdeploy::NormalizeImpl> Create(const Value& args) override {
return make_unique<NormalizeImpl>(args);
}
};
} // namespace mmdeploy::cuda
using mmdeploy::NormalizeImpl;
using mmdeploy::cuda::NormalizeImplCreator;
REGISTER_MODULE(NormalizeImpl, NormalizeImplCreator);