From e65720d49bbaf4c333cb079cf529c4eac88c32a1 Mon Sep 17 00:00:00 2001 From: cmhu <2657262686@qq.com> Date: Sat, 8 Oct 2022 17:23:18 +0800 Subject: [PATCH] 优化demo --- .vscode/settings.json | 6 +++++- src/NvJpegEncoder.cpp | 22 +++++++++++----------- src/NvJpegEncoder.h | 2 +- src/main.cpp | 120 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-------------------------------- 4 files changed, 105 insertions(+), 45 deletions(-) diff --git a/.vscode/settings.json b/.vscode/settings.json index 9590c79..c2847e1 100644 --- a/.vscode/settings.json +++ b/.vscode/settings.json @@ -5,6 +5,10 @@ "array": "cpp", "string": "cpp", "string_view": "cpp", - "iosfwd": "cpp" + "iosfwd": "cpp", + "typeinfo": "cpp", + "iostream": "cpp", + "thread": "cpp", + "nvdec.h": "c" } } \ No newline at end of file diff --git a/src/NvJpegEncoder.cpp b/src/NvJpegEncoder.cpp index 8cee91f..7ee0727 100644 --- a/src/NvJpegEncoder.cpp +++ b/src/NvJpegEncoder.cpp @@ -11,7 +11,7 @@ } while (false) -int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int height) +int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int height, cudaStream_t stream) { nvjpegHandle_t nvjpeg_handle; nvjpegEncoderState_t encoder_state; @@ -41,17 +41,17 @@ int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int heig CHECK_NVJPEG(nvjpegCreate(backend, nullptr, &nvjpeg_handle)); - CHECK_NVJPEG(nvjpegEncoderParamsCreate(nvjpeg_handle, &encoder_params, NULL)); - CHECK_NVJPEG(nvjpegEncoderStateCreate(nvjpeg_handle, &encoder_state, NULL)); + CHECK_NVJPEG(nvjpegEncoderParamsCreate(nvjpeg_handle, &encoder_params, stream)); + CHECK_NVJPEG(nvjpegEncoderStateCreate(nvjpeg_handle, &encoder_state, stream)); // set params - CHECK_NVJPEG(nvjpegEncoderParamsSetEncoding(encoder_params, nvjpegJpegEncoding_t::NVJPEG_ENCODING_PROGRESSIVE_DCT_HUFFMAN, NULL)); - CHECK_NVJPEG(nvjpegEncoderParamsSetOptimizedHuffman(encoder_params, 1, NULL)); - CHECK_NVJPEG(nvjpegEncoderParamsSetQuality(encoder_params, 70, NULL)); - CHECK_NVJPEG(nvjpegEncoderParamsSetSamplingFactors(encoder_params, nvjpegChromaSubsampling_t::NVJPEG_CSS_420, NULL)); + CHECK_NVJPEG(nvjpegEncoderParamsSetEncoding(encoder_params, nvjpegJpegEncoding_t::NVJPEG_ENCODING_PROGRESSIVE_DCT_HUFFMAN, stream)); + CHECK_NVJPEG(nvjpegEncoderParamsSetOptimizedHuffman(encoder_params, 1, stream)); + CHECK_NVJPEG(nvjpegEncoderParamsSetQuality(encoder_params, 70, stream)); + CHECK_NVJPEG(nvjpegEncoderParamsSetSamplingFactors(encoder_params, nvjpegChromaSubsampling_t::NVJPEG_CSS_420, stream)); cudaEventRecord(ev_start); - CHECK_NVJPEG(nvjpegEncodeImage(nvjpeg_handle, encoder_state, encoder_params, &input, input_format, image_width, image_height, NULL)); + CHECK_NVJPEG(nvjpegEncodeImage(nvjpeg_handle, encoder_state, encoder_params, &input, input_format, image_width, image_height, stream)); cudaEventRecord(ev_end); std::vector obuffer; @@ -61,7 +61,7 @@ int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int heig encoder_state, NULL, &length, - NULL)); + stream)); obuffer.resize(length); CHECK_NVJPEG(nvjpegEncodeRetrieveBitstream( @@ -69,7 +69,7 @@ int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int heig encoder_state, obuffer.data(), &length, - NULL)); + stream)); cudaEventSynchronize(ev_end); @@ -80,7 +80,7 @@ int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int heig float ms; cudaEventElapsedTime(&ms, ev_start, ev_end); - std::cout << "time spend " << ms << " ms" << std::endl; + // std::cout << "time spend " << ms << " ms" << std::endl; std::ofstream outputFile(filepath, std::ios::out | std::ios::binary); outputFile.write(reinterpret_cast(obuffer.data()), static_cast(length)); diff --git a/src/NvJpegEncoder.h b/src/NvJpegEncoder.h index 8d63d97..3c27ba8 100644 --- a/src/NvJpegEncoder.h +++ b/src/NvJpegEncoder.h @@ -1,3 +1,3 @@ #include -int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int height); \ No newline at end of file +int saveJpeg(const char * filepath, unsigned char* d_srcBGR, int width, int height, cudaStream_t stream); \ No newline at end of file diff --git a/src/main.cpp b/src/main.cpp index 2ecd8d1..e0fbba6 100644 --- a/src/main.cpp +++ b/src/main.cpp @@ -10,7 +10,12 @@ #include -unsigned char *pHwRgb = nullptr; +unsigned char *pHwRgb[2] = {nullptr, nullptr}; + +int sum1 = 0; +int sum2 = 0; + +cudaStream_t stream[2]; /** * 注意： gpuFrame 在解码器设置的显卡上，后续操作要十分注意这一点，尤其是多线程情况 @@ -21,35 +26,63 @@ void postDecoded(const void * userPtr, AVFrame * gpuFrame){ { // cout << "decode name: " << decoder->getName() << endl; - if (decoder->getName() == "dec1") - { - /* code */ - } - - // const char* gpu_pixfmt = av_get_pix_fmt_name((AVPixelFormat)gpuFrame->format); - // cout << "pixfmt: " << gpu_pixfmt << endl; - // cout << "keyframe: " << gpuFrame->key_frame << " width: " << gpuFrame->width << " height: "<< gpuFrame->height << endl; - // cout << "decode successed ✿✿ヽ(°▽°)ノ✿ " << endl; - - if (gpuFrame->format == AV_PIX_FMT_CUDA) - { - // cout << "gpuid = " << atoi(decoder->m_cfg.gpuid.c_str()) << endl; - // cudaSetDevice(atoi(decoder->m_cfg.gpuid.c_str())); - // cudaError_t cudaStatus; - // if(pHwRgb == nullptr){ - // cuda_common::setColorSpace2( ITU709, 0 ); - // cudaStatus = cudaMalloc((void **)&pHwRgb, 3 * gpuFrame->width * gpuFrame->height * sizeof(unsigned char)); - // } - // cudaStatus = cuda_common::CUDAToBGR((CUdeviceptr)gpuFrame->data[0],(CUdeviceptr)gpuFrame->data[1], gpuFrame->linesize[0], gpuFrame->linesize[1], pHwRgb, gpuFrame->width, gpuFrame->height); - // cudaDeviceSynchronize(); - // if (cudaStatus != cudaSuccess) { - // cout << "CUDAToBGR failed !!!" << endl; - // return; - // } - - // string path = "/home/cmhu/FFNvDecoder/" + decoder->m_cfg.gpuid + ".jpg"; - // saveJpeg(path.c_str(), pHwRgb, gpuFrame->width, gpuFrame->height); // 验证 CUDAToRGB - } + // const char* gpu_pixfmt = av_get_pix_fmt_name((AVPixelFormat)gpuFrame->format); + // cout << "pixfmt: " << gpu_pixfmt << endl; + // cout << "keyframe: " << gpuFrame->key_frame << " width: " << gpuFrame->width << " height: "<< gpuFrame->height << endl; + // cout << "decode successed ✿✿ヽ(°▽°)ノ✿ " << endl; + + int sum = sum1; + if (decoder->getName() == "dec1") + { + sum1 ++ ; + sum = sum1; + + if (gpuFrame->format == AV_PIX_FMT_CUDA) + { + cout << "gpuid = " << atoi(decoder->m_cfg.gpuid.c_str()) << endl; + cudaSetDevice(atoi(decoder->m_cfg.gpuid.c_str())); + cudaError_t cudaStatus; + if(pHwRgb[0] == nullptr){ + // cudaStreamCreate(&stream[0]); + cuda_common::setColorSpace2( ITU709, 0 ); + cudaStatus = cudaMalloc((void **)&pHwRgb[0], 3 * gpuFrame->width * gpuFrame->height * sizeof(unsigned char)); + } + cudaStatus = cuda_common::CUDAToBGR((CUdeviceptr)gpuFrame->data[0],(CUdeviceptr)gpuFrame->data[1], gpuFrame->linesize[0], gpuFrame->linesize[1], pHwRgb[0], gpuFrame->width, gpuFrame->height); + cudaDeviceSynchronize(); + if (cudaStatus != cudaSuccess) { + cout << "CUDAToBGR failed !!!" << endl; + return; + } + + string path = "/home/cmhu/data/" + decoder->getName() + "/" + to_string(sum) + ".jpg"; + saveJpeg(path.c_str(), pHwRgb[0], gpuFrame->width, gpuFrame->height, stream[0]); // 验证 CUDAToRGB + } + } else if (decoder->getName() == "dec2") + { + sum2 ++ ; + sum = sum2; + + if (gpuFrame->format == AV_PIX_FMT_CUDA) + { + cout << "gpuid = " << atoi(decoder->m_cfg.gpuid.c_str()) << endl; + cudaSetDevice(atoi(decoder->m_cfg.gpuid.c_str())); + cudaError_t cudaStatus; + if(pHwRgb[1] == nullptr){ + // cudaStreamCreate(&stream[1]); + cuda_common::setColorSpace2( ITU709, 0 ); + cudaStatus = cudaMalloc((void **)&pHwRgb[1], 3 * gpuFrame->width * gpuFrame->height * sizeof(unsigned char)); + } + cudaStatus = cuda_common::CUDAToBGR((CUdeviceptr)gpuFrame->data[0],(CUdeviceptr)gpuFrame->data[1], gpuFrame->linesize[0], gpuFrame->linesize[1], pHwRgb[1], gpuFrame->width, gpuFrame->height); + cudaDeviceSynchronize(); + if (cudaStatus != cudaSuccess) { + cout << "CUDAToBGR failed !!!" << endl; + return; + } + + string path = "/home/cmhu/data/" + decoder->getName() + "/" + to_string(sum) + ".jpg"; + saveJpeg(path.c_str(), pHwRgb[1], gpuFrame->width, gpuFrame->height, stream[1]); // 验证 CUDAToRGB + } + } } } @@ -72,6 +105,8 @@ long long get_cur_time(){ } int sum = 0; +unsigned char *pHwData = nullptr; + void postDecoded0(const void * userPtr, AVFrame * gpuFrame){ FFNvDecoder* decoder = (FFNvDecoder*)userPtr; if (decoder!= nullptr) @@ -98,12 +133,33 @@ void postDecoded0(const void * userPtr, AVFrame * gpuFrame){ count_flag = false; } + + if (gpuFrame->format == AV_PIX_FMT_CUDA) + { + cudaSetDevice(atoi(decoder->m_cfg.gpuid.c_str())); + cout << "gpu id : " << decoder->m_cfg.gpuid.c_str() << endl; + cudaError_t cudaStatus; + if(pHwData == nullptr){ + cuda_common::setColorSpace2( ITU709, 0 ); + cudaStatus = cudaMalloc((void **)&pHwData, 3 * gpuFrame->width * gpuFrame->height * sizeof(unsigned char)); + } + cudaStatus = cuda_common::CUDAToBGR((CUdeviceptr)gpuFrame->data[0],(CUdeviceptr)gpuFrame->data[1], gpuFrame->linesize[0], gpuFrame->linesize[1], pHwData, gpuFrame->width, gpuFrame->height); + cudaDeviceSynchronize(); + if (cudaStatus != cudaSuccess) { + cout << "CUDAToBGR failed !!!" << endl; + return; + } + + string path = "/home/cmhu/data/test/" + to_string(sum) + ".jpg"; + saveJpeg(path.c_str(), pHwData, gpuFrame->width, gpuFrame->height, nullptr); // 验证 CUDAToRGB + } } } } // string test_uri = "rtmp://192.168.10.56:1935/objecteye/1"; -string test_uri = "/home/cmhu/data/test.mp4"; +// string test_uri = "/home/cmhu/data/output_800x480.mp4"; +string test_uri = "/home/cmhu/data/output_1920x1080.mp4"; void createDecode(int index){ FFNvDecoderManager* pDecManager = FFNvDecoderManager::getInstance(); @@ -170,7 +226,7 @@ int main(){ FFNvDecoderManager* pDecManager = FFNvDecoderManager::getInstance(); - int count = 105; + int count = 99; for (size_t i = 0; i < count ; i++) { createDecode(i); -- libgit2 0.21.4