Merge pull request #96 from rafbiels/dl-cifar-fix-host-overhead

[dl-cifar] Reduce the impact of host computations on benchmark results

Author: rmukhopa

dl-cifar/CUDA/CMakeLists.txt

@@ -34,7 +34,7 @@ option(USE_SM                  "Specifies which streaming multiprocessor archite
 option(DEVICE_TIMER            "Build using Device Timer" OFF)
 
 set(DEF_WL_CXX_FLAGS           " ")
-set(DEF_GENERAL_CXX_FLAGS      " -O2 ")
+set(DEF_GENERAL_CXX_FLAGS      " -O3 -ffast-math ")
 set(DEF_COMBINED_CXX_FLAGS     "${DEF_GENERAL_CXX_FLAGS} ${DEF_WL_CXX_FLAGS}")
 
 set(SOURCES 
@@ -114,4 +114,4 @@ set(CUDA_SEPARABLE_COMPILATION ON)
 message(STATUS "CXX  Compilation flags to: ${CMAKE_CXX_FLAGS}")
 link_libraries(stdc++fs cublas cudnn)
 cuda_add_executable(${PROJECT_NAME} ${SOURCES})
-target_link_libraries(${PROJECT_NAME} ${CUDA_LIBRARIES})
+target_link_libraries(${PROJECT_NAME} ${CUDA_LIBRARIES})

dl-cifar/HIP/CMakeLists.txt

@@ -33,7 +33,7 @@ option(DEVICE_TIMER            "Build using Device Timer" OFF)
 
 
 set(DEF_WL_CXX_FLAGS           " -D__HIP_PLATFORM_AMD__ ")
-set(DEF_GENERAL_CXX_FLAGS      " -Wall -O3 -Wextra ")
+set(DEF_GENERAL_CXX_FLAGS      " -Wall -O3 -ffast-math -Wextra ")
 set(DEF_COMBINED_CXX_FLAGS     "${DEF_GENERAL_CXX_FLAGS} ${DEF_WL_CXX_FLAGS}")
 
 

dl-cifar/common/image_processing.h

@@ -28,6 +28,7 @@
 #include <cassert>
 #include <vector>
 #include <exception>
+#include <memory>
 #include "tracing.h"
 #include"handle.h"
 #include "upsample.h"
@@ -102,10 +103,24 @@ namespace dl_cifar::common {
             static void initImage(float* image, int imageSize) {
                 Tracer::func_begin("ImageProcessor::initImage");
 
-                unsigned seed = 123456789;
-                for (int index = 0; index < imageSize; index++) {
+                static size_t cacheSize{0};
+                static std::unique_ptr<float[]> cacheImage{};
+                static unsigned seed = 123456789;
+
+                // grow the cache allocation to image size
+                if (imageSize > cacheSize) {
+                    auto newCacheImage = std::make_unique<float[]>(imageSize);
+                    std::memcpy(newCacheImage.get(), cacheImage.get(), cacheSize*sizeof(float));
+                    cacheImage.reset(newCacheImage.release());
+                }
+
+                // fill image with cached data and compute the remaining part
+                std::memcpy(image, cacheImage.get(), std::min(cacheSize,static_cast<size_t>(imageSize))*sizeof(float));
+                while (cacheSize < imageSize) {
+                    ++cacheSize;
                     seed         = (1103515245 * seed + 12345) & 0xffffffff;
-                    image[index] = float(seed) * 2.3283064e-10;  // 2^-32
+                    cacheImage[cacheSize-1] = float(seed) * 2.3283064e-10;  // 2^-32
+                    image[cacheSize-1] = cacheImage[cacheSize-1];
                 }
                 Tracer::func_end("ImageProcessor::initImage");
 

dl-cifar/common/vit/vit.h

@@ -181,10 +181,10 @@ namespace dl_cifar::common {
                     ImageProcessor::resize(langHandle, d_cifarRawImgs, d_resizedImgs, selectedVitParams.batchSize, 
                                             VitConfigs::cifarNoOfChannels, VitConfigs::cifarImgWidth, VitConfigs::cifarImgWidth, 
                                             selectedVitParams.imgWidth, selectedVitParams.imgHeight); 
-                    ImageProcessor::resizeInHost(langHandle, h_cifarRawImgs, h_resizedImgs, selectedVitParams.batchSize, 
-                                            VitConfigs::cifarNoOfChannels, VitConfigs::cifarImgWidth, VitConfigs::cifarImgWidth, 
-                                            selectedVitParams.imgWidth, selectedVitParams.imgHeight); 
-                    langHandle->memCpyH2D(d_resizedImgs, h_resizedImgs, sizeof(float) * resizedSize, true);
+                    // ImageProcessor::resizeInHost(langHandle, h_cifarRawImgs, h_resizedImgs, selectedVitParams.batchSize,
+                    //                         VitConfigs::cifarNoOfChannels, VitConfigs::cifarImgWidth, VitConfigs::cifarImgWidth,
+                    //                         selectedVitParams.imgWidth, selectedVitParams.imgHeight);
+                    // langHandle->memCpyH2D(d_resizedImgs, h_resizedImgs, sizeof(float) * resizedSize, true);