Remove profiling nvtx from develop branch.

This make trouble for continuous integration and is apparently not supported on all platforms.
Since it is a debug function, it's just as well to remove it from the mainstream tree.

Author: Carsten Griwodz

CMakeLists.txt

@@ -18,7 +18,6 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY "${PROJECT_BINARY_DIR}/${CMAKE_SYSTEM_NAME}-$
 
 option(PopSift_BUILD_EXAMPLES "Build PopSift applications."  ON)
 option(PopSift_BUILD_DOCS "Build PopSift documentation."  OFF)
-option(PopSift_USE_NVTX_PROFILING     "Use CUDA NVTX for profiling." OFF)
 option(PopSift_ERRCHK_AFTER_KERNEL     "Synchronize and check CUDA error after every kernel." OFF)
 option(PopSift_USE_POSITION_INDEPENDENT_CODE "Generate position independent code." ON)
 option(PopSift_USE_GRID_FILTER "Switch off grid filtering to massively reduce compile time while debugging other things." ON)
@@ -90,10 +89,6 @@ find_package(CUDAToolkit)
 message(STATUS "CUDA Version is ${CUDAToolkit_VERSION}")
 set(CUDA_VERSION ${CUDAToolkit_VERSION})
 
-if(PopSift_USE_NVTX_PROFILING)
-  message(STATUS "PROFILING CPU CODE: NVTX is in use")
-endif()
-
 if(PopSift_ERRCHK_AFTER_KERNEL)
   message(STATUS "Synchronizing and checking errors after every kernel call")
   list(APPEND CUDA_NVCC_FLAGS "-DERRCHK_AFTER_KERNEL")
@@ -150,13 +145,6 @@ else()
   set(DISABLE_GRID_FILTER   0)
 endif()
 
-if(PopSift_USE_NVTX_PROFILING)
-  # library required for NVTX profiling of the CPU
-  set(PopSift_USE_NVTX 1)
-else()
-  set(PopSift_USE_NVTX 0)
-endif()
-
 add_subdirectory(src)
 
 if(PopSift_BUILD_DOCS)
@@ -194,7 +182,6 @@ message(STATUS "Build Shared libs: " ${BUILD_SHARED_LIBS})
 message(STATUS "Build examples: " ${PopSift_BUILD_EXAMPLES})
 message(STATUS "Build documentation: " ${PopSift_BUILD_DOCS})
 message(STATUS "Generate position independent code: " ${CMAKE_POSITION_INDEPENDENT_CODE})
-message(STATUS "Use CUDA NVTX for profiling: " ${PopSift_USE_NVTX_PROFILING})
 message(STATUS "Synchronize and check CUDA error after every kernel: " ${PopSift_ERRCHK_AFTER_KERNEL})
 message(STATUS "Grid filtering: " ${PopSift_USE_GRID_FILTER})
 message(STATUS "Additional warning for CUDA nvcc: " ${PopSift_NVCC_WARNINGS})

appveyor.yml

@@ -44,7 +44,7 @@ install:
 before_build:
     - md build
     - cd build
-    - cmake -G "Visual Studio 17 2022" -A x64 -T v143,host=x64,cuda="%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5" -DBUILD_SHARED_LIBS:BOOL=ON -DPopSift_USE_NVTX_PROFILING:BOOL=OFF -DPopSift_USE_GRID_FILTER:BOOL=OFF -DPopSift_BUILD_DOCS:BOOL=OFF -DPopSift_USE_POSITION_INDEPENDENT_CODE:BOOL=ON -DPopSift_BUILD_EXAMPLES:BOOL=ON -DCMAKE_BUILD_TYPE=%configuration% -DCMAKE_TOOLCHAIN_FILE=c:/tools/vcpkg/scripts/buildsystems/vcpkg.cmake ..
+    - cmake -G "Visual Studio 17 2022" -A x64 -T v143,host=x64,cuda="%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5" -DBUILD_SHARED_LIBS:BOOL=ON -DPopSift_USE_GRID_FILTER:BOOL=OFF -DPopSift_BUILD_DOCS:BOOL=OFF -DPopSift_USE_POSITION_INDEPENDENT_CODE:BOOL=ON -DPopSift_BUILD_EXAMPLES:BOOL=ON -DCMAKE_BUILD_TYPE=%configuration% -DCMAKE_TOOLCHAIN_FILE=c:/tools/vcpkg/scripts/buildsystems/vcpkg.cmake ..
     - ls -l
 
 build:

cmake/sift_config.h.in

@@ -14,5 +14,4 @@
 #define POPSIFT_HAVE_SHFL_DOWN_SYNC()     @PopSift_HAVE_SHFL_DOWN_SYNC@
 #define POPSIFT_HAVE_NORMF()              @PopSift_HAVE_NORMF@
 #define POPSIFT_DISABLE_GRID_FILTER()     @DISABLE_GRID_FILTER@
-#define POPSIFT_USE_NVTX()                @PopSift_USE_NVTX@
 

cudaInstallAppveyor.cmd

@@ -4,14 +4,12 @@ echo Downloading CUDA toolkit 12 for Windows 10
 
 appveyor DownloadFile https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.5.82-archive.zip -Filename cuda_nvcc.zip
 appveyor DownloadFile https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.5.82-archive.zip -Filename cuda_cudart.zip
-appveyor DownloadFile https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.5.82-archive.zip -Filename cuda_nvtx.zip
 appveyor DownloadFile https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.5.82-archive.zip -Filename vs_integration.zip
 dir
 
 echo Unzipping CUDA toolkit 12
 tar -xf cuda_nvcc.zip
 tar -xf cuda_cudart.zip
-tar -xf cuda_nvtx.zip
 tar -xf vs_integration.zip
 dir
 
@@ -22,7 +20,6 @@ mkdir "%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5\extras"
 echo Copying toolkit files to install dir(s)
 xcopy cuda_cudart-windows-x86_64-12.5.82-archive "%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5" /s /e /i /y
 xcopy cuda_nvcc-windows-x86_64-12.5.82-archive "%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5" /s /e /i /y
-xcopy cuda_nvtx-windows-x86_64-12.5.82-archive "%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5" /s /e /i /y
 xcopy visual_studio_integration-windows-x86_64-12.5.82-archive "%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5\extras" /s /e /i /y
 
 

src/CMakeLists.txt

@@ -45,12 +45,6 @@ target_link_libraries(popsift
 		      CUDA::cudart
 		      Threads::Threads)
 
-if(PopSift_USE_NVTX_PROFILING)
-target_link_libraries(popsift
-                      PUBLIC
-		      CUDA::nvtx3)
-endif()
-
 set_target_properties(popsift PROPERTIES VERSION ${PROJECT_VERSION})
 set_target_properties(popsift PROPERTIES DEBUG_POSTFIX "d")
 set_target_properties(popsift PROPERTIES CUDA_SEPARABLE_COMPILATION ON)

src/popsift/popsift.cu

@@ -438,18 +438,12 @@ void SiftJob::setImg( popsift::ImageBase* img )
 
 popsift::ImageBase* SiftJob::getImg()
 {
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-    _nvtx_id = nvtxRangeStartA( "inserting image" );
-#endif
     return _img;
 }
 
 void SiftJob::setFeatures( popsift::FeaturesBase* f )
 {
     _p.set_value( f );
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-    nvtxRangeEnd( _nvtx_id );
-#endif
 }
 
 popsift::FeaturesHost* SiftJob::get()

src/popsift/popsift.h

@@ -23,13 +23,6 @@
 #include <thread>
 #include <vector>
 
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-#include <nvtx3/nvToolsExtCuda.h>
-#else
-#define nvtxRangeStartA(a)
-#define nvtxRangeEnd(a)
-#endif
-
 /* user parameters */
 namespace popsift
 {
@@ -50,9 +43,6 @@ class SiftJob
     unsigned char*      _imageData;
     popsift::ImageBase* _img;
     std::exception_ptr _err;
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-    nvtxRangeId_t       _nvtx_id;
-#endif
 
 public:
 

src/popsift/s_filtergrid.cu

@@ -9,13 +9,6 @@
 #include "sift_extremum.h"
 #include "sift_pyramid.h"
 
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-#include <nvtx3/nvToolsExtCuda.h>
-#else
-#define nvtxRangePushA(a)
-#define nvtxRangePop()
-#endif
-
 #if ! POPSIFT_IS_DEFINED(POPSIFT_DISABLE_GRID_FILTER)
 
 #include <thrust/copy.h>
@@ -317,9 +310,7 @@ int Pyramid::extrema_filter_grid( const Config& conf, int ext_total )
         }
     }
 
-    nvtxRangePushA( "writing back count" );
     writeDescCountersToDevice( );
-    nvtxRangePop( );
 
     return ret_ext_total;
 }

src/popsift/s_image.cu

@@ -15,13 +15,6 @@
 #include <fstream>
 #include <iostream>
 
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-#include <nvtx3/nvToolsExtCuda.h>
-#else
-#define nvtxRangePushA(a)
-#define nvtxRangePop()
-#endif
-
 using namespace std;
 
 namespace popsift {
@@ -98,8 +91,6 @@ void Image::resetDimensions( int w, int h )
         destroyTexture( );
         createTexture( );
     } else {
-        nvtxRangePushA( "reallocating host-side image memory" );
-
         _max_w = max( w, _max_w );
         _max_h = max( h, _max_h );
         _input_image_h.freeHost( popsift::CudaAllocated );
@@ -111,21 +102,15 @@ void Image::resetDimensions( int w, int h )
 
         destroyTexture( );
         createTexture( );
-
-        nvtxRangePop(); // "reallocating host-side image memory"
     }
 }
 
 void Image::allocate( int w, int h )
 {
-    nvtxRangePushA( "allocating host-side image memory" );
-
     _input_image_h.allocHost( w, h, popsift::CudaAllocated );
     _input_image_d.allocDev( w, h );
 
     createTexture( );
-
-    nvtxRangePop(); // "allocating host-side image memory"
 }
 
 void Image::destroyTexture( )
@@ -222,8 +207,6 @@ void ImageFloat::resetDimensions( int w, int h )
         destroyTexture( );
         createTexture( );
     } else {
-        nvtxRangePushA( "reallocating host-side image memory" );
-
         _max_w = max( w, _max_w );
         _max_h = max( h, _max_h );
         _input_image_h.freeHost( popsift::CudaAllocated );
@@ -235,21 +218,15 @@ void ImageFloat::resetDimensions( int w, int h )
 
         destroyTexture( );
         createTexture( );
-
-        nvtxRangePop(); // "reallocating host-side image memory"
     }
 }
 
 void ImageFloat::allocate( int w, int h )
 {
-    nvtxRangePushA( "allocating host-side image memory" );
-
     _input_image_h.allocHost( w, h, popsift::CudaAllocated );
     _input_image_d.allocDev( w, h );
 
     createTexture( );
-
-    nvtxRangePop(); // "allocating host-side image memory"
 }
 
 void ImageFloat::destroyTexture( )

src/popsift/s_orientation.cu

@@ -18,13 +18,6 @@
 #include <cmath>
 #include <cstdio>
 
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-#include <nvtx3/nvToolsExtCuda.h>
-#else
-#define nvtxRangePushA(a)
-#define nvtxRangePop()
-#endif
-
 using namespace popsift;
 using namespace std;
 

src/popsift/sift_desc.cu

@@ -21,13 +21,6 @@
 #include <cstdio>
 #include <iostream>
 
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-#include <nvtx3/nvToolsExtCuda.h>
-#else
-#define nvtxRangePushA(a)
-#define nvtxRangePop()
-#endif
-
 using namespace popsift;
 using namespace std;
 
@@ -55,11 +48,8 @@ using namespace std;
 __host__
 void Pyramid::descriptors( const Config& conf )
 {
-   nvtxRangePushA("Reading orientation count");
-
    readDescCountersFromDevice( _octaves[0].getStream() );
    cudaStreamSynchronize( _octaves[0].getStream() );
-   nvtxRangePop( );
 
     for( int octave=_num_octaves-1; octave>=0; octave-- )
     // for( int octave=0; octave<_num_octaves; octave++ )

src/popsift/sift_pyramid.cu

@@ -25,13 +25,6 @@
 #define mkdir(path, perm) _mkdir(path)
 #endif
 
-#if POPSIFT_IS_DEFINED(POPSIFT_USE_NVTX)
-#include <nvtx3/nvToolsExtCuda.h>
-#else
-#define nvtxRangePushA(a)
-#define nvtxRangePop()
-#endif
-
 #define PYRAMID_PRINT_DEBUG 0
 
 using namespace std;
@@ -285,22 +278,18 @@ FeaturesHost* Pyramid::get_descriptors( const Config& conf )
 
     readDescCountersFromDevice();
 
-    nvtxRangePushA( "download descriptors" );
     FeaturesHost* features = new FeaturesHost( hct.ext_total, hct.ori_total );
 
     if( hct.ext_total == 0 || hct.ori_total == 0 )
     {
-        nvtxRangePop();
         return features;
     }
 
     dim3 grid( grid_divide( hct.ext_total, 32 ) );
     prep_features<<<grid,32,0,_download_stream>>>( features->getDescriptors(), up_fac );
     POP_SYNC_CHK;
 
-    nvtxRangePushA( "register host memory" );
     features->pin( );
-    nvtxRangePop();
     popcuda_memcpy_async( features->getFeatures(),
                           dobuf_shadow.features,
                           hct.ext_total * sizeof(Feature),
@@ -313,10 +302,7 @@ FeaturesHost* Pyramid::get_descriptors( const Config& conf )
                           cudaMemcpyDeviceToHost,
                           _download_stream );
     cudaStreamSynchronize( _download_stream );
-    nvtxRangePushA( "unregister host memory" );
     features->unpin( );
-    nvtxRangePop();
-    nvtxRangePop();
 
     return features;
 }