Adding NVTX support

Author: rarzumanyan

PyNvCodec/TC/CMakeLists.txt

@@ -20,6 +20,11 @@ project(TC)
 enable_language(CUDA)
 add_subdirectory(TC_CORE)
 
+set(USE_NVTX FALSE CACHE BOOL "Use NVTX for profiling")
+if(USE_NVTX)
+	add_definitions(-DUSE_NVTX)
+endif()
+
 set(VIDEO_CODEC_SDK_DIR "" CACHE PATH "Path to Nvidia Video Codec SDK")
 set(FFMPEG_DIR "" CACHE PATH "Path to FFMpeg")
 

PyNvCodec/TC/inc/Tasks.hpp

@@ -22,10 +22,31 @@ extern "C" {
   #include <libavutil/frame.h>
 }
 
+#ifdef USE_NVTX
+#include <nvtx3/nvToolsExt.h>
+#define NVTX_PUSH(FNAME)                                                       \
+  do {                                                                         \
+    nvtxRangePush(FNAME);                                                      \
+  } while (0);
+#define NVTX_POP                                                               \
+  do {                                                                         \
+    nvtxRangePop();                                                            \
+  } while (0);
+#else
+#define NVTX_PUSH(FNAME)
+#define NVTX_POP
+#endif
+
 using namespace VPF;
 
 // VPF stands for Video Processing Framework;
 namespace VPF {
+class DllExport NvtxMark {
+public:
+  NvtxMark(const char *fname) { NVTX_PUSH(fname) }
+  ~NvtxMark() { NVTX_POP }
+};
+
 class DllExport NvencEncodeFrame final : public Task {
 public:
   NvencEncodeFrame() = delete;

PyNvCodec/TC/src/Tasks.cpp

@@ -137,6 +137,7 @@ NvencEncodeFrame::NvencEncodeFrame(CUstream cuStream, CUcontext cuContext,
 NvencEncodeFrame::~NvencEncodeFrame() { delete pImpl; };
 
 TaskExecStatus NvencEncodeFrame::Execute() {
+  NvtxMark tick(__FUNCTION__);
   SetOutput(nullptr, 0U);
 
   try {
@@ -279,6 +280,7 @@ NvdecDecodeFrame::~NvdecDecodeFrame() {
 }
 
 TaskExecStatus NvdecDecodeFrame::Execute() {
+  NvtxMark tick(__FUNCTION__);
   ClearOutputs();
 
   auto &decoder = pImpl->nvDecoder;
@@ -414,6 +416,7 @@ CudaUploadFrame::CudaUploadFrame(CUstream cuStream, CUcontext cuContext,
 CudaUploadFrame::~CudaUploadFrame() { delete pImpl; }
 
 TaskExecStatus CudaUploadFrame::Execute() {
+  NvtxMark tick(__FUNCTION__);
   if (!GetInput()) {
     return TASK_EXEC_FAIL;
   }
@@ -513,6 +516,7 @@ CudaDownloadSurface::CudaDownloadSurface(CUstream cuStream, CUcontext cuContext,
 CudaDownloadSurface::~CudaDownloadSurface() { delete pImpl; }
 
 TaskExecStatus CudaDownloadSurface::Execute() {
+  NvtxMark tick(__FUNCTION__);
 
   if (!GetInput()) {
     return TASK_EXEC_FAIL;
@@ -612,6 +616,7 @@ DemuxFrame::~DemuxFrame() { delete pImpl; }
 void DemuxFrame::Flush() { pImpl->demuxer.Flush(); }
 
 TaskExecStatus DemuxFrame::Execute() {
+  NvtxMark tick(__FUNCTION__);
   ClearOutputs();
 
   uint8_t *pVideo = nullptr;
@@ -800,6 +805,7 @@ MuxFrame::~MuxFrame() {
 }
 
 TaskExecStatus MuxFrame::Execute() {
+  NvtxMark tick(__FUNCTION__);
   auto elementaryVideo = (Buffer *)GetInput(0U);
   auto muxingParamsBuffer = (Buffer *)GetInput(1U);
 
@@ -894,6 +900,7 @@ struct NppResizeSurfacePacked3C_Impl final : ResizeSurface_Impl {
   ~NppResizeSurfacePacked3C_Impl() { delete pSurface; }
 
   TaskExecStatus Execute(Surface &source) {
+    NvtxMark tick(__FUNCTION__);
 
     if (pSurface->PixelFormat() != source.PixelFormat()) {
       return TaskExecStatus::TASK_EXEC_FAIL;
@@ -950,6 +957,7 @@ struct NppResizeSurfacePlanar420_Impl final : ResizeSurface_Impl {
   ~NppResizeSurfacePlanar420_Impl() { delete pSurface; }
 
   TaskExecStatus Execute(Surface &source) {
+    NvtxMark tick(__FUNCTION__);
 
     if (pSurface->PixelFormat() != source.PixelFormat()) {
       cerr << "Actual pixel format is " << source.PixelFormat() << endl;
@@ -1016,6 +1024,7 @@ ResizeSurface::ResizeSurface(uint32_t width, uint32_t height,
 ResizeSurface::~ResizeSurface() { delete pImpl; }
 
 TaskExecStatus ResizeSurface::Execute() {
+  NvtxMark tick(__FUNCTION__);
   ClearOutputs();
 
   auto pInputSurface = (Surface *)GetInput();

PyNvCodec/TC/src/TasksColorCvt.cpp

@@ -50,6 +50,7 @@ struct nv12_bgr final : public NppConvertSurface_Impl {
   ~nv12_bgr() { delete pSurface; }
 
   Token *Execute(Token *pInputNV12) override {
+    NvtxMark tick(__FUNCTION__);
     if (!pInputNV12) {
       return nullptr;
     }
@@ -85,6 +86,7 @@ struct nv12_rgb final : public NppConvertSurface_Impl {
   ~nv12_rgb() { delete pSurface; }
 
   Token *Execute(Token *pInputNV12) override {
+    NvtxMark tick(__FUNCTION__);
     if (!pInputNV12) {
       return nullptr;
     }
@@ -121,6 +123,7 @@ struct nv12_yuv420 final : public NppConvertSurface_Impl {
   ~nv12_yuv420() { delete pSurface; }
 
   Token *Execute(Token *pInputNV12) override {
+    NvtxMark tick(__FUNCTION__);
     if (!pInputNV12) {
       return nullptr;
     }
@@ -163,6 +166,7 @@ struct yuv420_rgb final : public NppConvertSurface_Impl {
   ~yuv420_rgb() { delete pSurface; }
 
   Token *Execute(Token *pInputYUV420) override {
+    NvtxMark tick(__FUNCTION__);
     if (!pInputYUV420) {
       return nullptr;
     }
@@ -202,6 +206,7 @@ struct bgr_ycbcr final : public NppConvertSurface_Impl {
   ~bgr_ycbcr() { delete pSurface; }
 
   Token *Execute(Token *pInput) override {
+    NvtxMark tick(__FUNCTION__);
     auto pInputBGR = (SurfaceRGB *)pInput;
 
     if (BGR != pInputBGR->PixelFormat()) {
@@ -248,6 +253,7 @@ struct rgb_yuv420 final : public NppConvertSurface_Impl {
   ~rgb_yuv420() { delete pSurface; }
 
   Token *Execute(Token *pInput) override {
+    NvtxMark tick(__FUNCTION__);
     auto pInputRGB8 = (SurfaceRGB *)pInput;
 
     if (RGB != pInputRGB8->PixelFormat()) {
@@ -288,6 +294,7 @@ struct yuv420_nv12 final : public NppConvertSurface_Impl {
   ~yuv420_nv12() { delete pSurface; }
 
   Token *Execute(Token *pInputYUV420) override {
+    NvtxMark tick(__FUNCTION__);
     if (!pInputYUV420) {
       return nullptr;
     }
@@ -331,6 +338,7 @@ struct rgb8_deinterleave final : public NppConvertSurface_Impl {
   ~rgb8_deinterleave() { delete pSurface; }
 
   Token *Execute(Token *pInput) override {
+    NvtxMark tick(__FUNCTION__);
     auto pInputRGB8 = (SurfaceRGB *)pInput;
 
     if (RGB != pInputRGB8->PixelFormat()) {
@@ -374,6 +382,7 @@ struct rbg8_swapchannel final : public NppConvertSurface_Impl {
   ~rbg8_swapchannel() { delete pSurface; }
 
   Token *Execute(Token *pInput) override {
+    NvtxMark tick(__FUNCTION__);
     if (!pInput) {
       return nullptr;
     }