mirror of
https://github.com/NVIDIA/cuda-samples.git
synced 2026-10-25 22:28:17 +08:00
Apply consistent code formatting across the repo. Add clang-format and pre-commit hooks.
This commit is contained in:
Rob Armstrong
@@ -30,19 +30,20 @@
|
||||
*/
|
||||
|
||||
// includes, system
|
||||
#include <stdlib.h>
|
||||
#include <stdio.h>
|
||||
#include <string.h>
|
||||
#include <math.h>
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include <string.h>
|
||||
|
||||
// includes, kernels
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
#include "stereoDisparity_kernel.cuh"
|
||||
|
||||
// includes, project
|
||||
#include <helper_functions.h> // helper for shared that are common to CUDA Samples
|
||||
#include <helper_cuda.h> // helper for checking cuda initialization and error checking
|
||||
#include <helper_string.h> // helper functions for string parsing
|
||||
#include <helper_cuda.h> // helper for checking cuda initialization and error checking
|
||||
#include <helper_functions.h> // helper for shared that are common to CUDA Samples
|
||||
#include <helper_string.h> // helper functions for string parsing
|
||||
|
||||
static const char *sSDKsample = "[stereoDisparity]\0";
|
||||
|
||||
@@ -55,228 +56,230 @@ void runTest(int argc, char **argv);
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// Program main
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
int main(int argc, char **argv) {
|
||||
printf("%s Starting...\n\n", sSDKsample);
|
||||
runTest(argc, argv);
|
||||
int main(int argc, char **argv)
|
||||
{
|
||||
printf("%s Starting...\n\n", sSDKsample);
|
||||
runTest(argc, argv);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
//! CUDA Sample for calculating depth maps
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
void runTest(int argc, char **argv) {
|
||||
cudaDeviceProp deviceProp;
|
||||
deviceProp.major = 0;
|
||||
deviceProp.minor = 0;
|
||||
int dev = 0;
|
||||
void runTest(int argc, char **argv)
|
||||
{
|
||||
cudaDeviceProp deviceProp;
|
||||
deviceProp.major = 0;
|
||||
deviceProp.minor = 0;
|
||||
int dev = 0;
|
||||
|
||||
// This will pick the best possible CUDA capable device
|
||||
dev = findCudaDevice(argc, (const char **)argv);
|
||||
// This will pick the best possible CUDA capable device
|
||||
dev = findCudaDevice(argc, (const char **)argv);
|
||||
|
||||
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
|
||||
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
|
||||
|
||||
// Statistics about the GPU device
|
||||
printf(
|
||||
"> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
|
||||
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
|
||||
// Statistics about the GPU device
|
||||
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
|
||||
deviceProp.multiProcessorCount,
|
||||
deviceProp.major,
|
||||
deviceProp.minor);
|
||||
|
||||
StopWatchInterface *timer;
|
||||
sdkCreateTimer(&timer);
|
||||
StopWatchInterface *timer;
|
||||
sdkCreateTimer(&timer);
|
||||
|
||||
// Search parameters
|
||||
int minDisp = -16;
|
||||
int maxDisp = 0;
|
||||
// Search parameters
|
||||
int minDisp = -16;
|
||||
int maxDisp = 0;
|
||||
|
||||
// Load image data
|
||||
// allocate mem for the images on host side
|
||||
// initialize pointers to NULL to request lib call to allocate as needed
|
||||
// PPM images are loaded into 4 byte/pixel memory (RGBX)
|
||||
unsigned char *h_img0 = NULL;
|
||||
unsigned char *h_img1 = NULL;
|
||||
unsigned int w, h;
|
||||
char *fname0 = sdkFindFilePath("stereo.im0.640x533.ppm", argv[0]);
|
||||
char *fname1 = sdkFindFilePath("stereo.im1.640x533.ppm", argv[0]);
|
||||
// Load image data
|
||||
// allocate mem for the images on host side
|
||||
// initialize pointers to NULL to request lib call to allocate as needed
|
||||
// PPM images are loaded into 4 byte/pixel memory (RGBX)
|
||||
unsigned char *h_img0 = NULL;
|
||||
unsigned char *h_img1 = NULL;
|
||||
unsigned int w, h;
|
||||
char *fname0 = sdkFindFilePath("stereo.im0.640x533.ppm", argv[0]);
|
||||
char *fname1 = sdkFindFilePath("stereo.im1.640x533.ppm", argv[0]);
|
||||
|
||||
printf("Loaded <%s> as image 0\n", fname0);
|
||||
printf("Loaded <%s> as image 0\n", fname0);
|
||||
|
||||
if (!sdkLoadPPM4ub(fname0, &h_img0, &w, &h)) {
|
||||
fprintf(stderr, "Failed to load <%s>\n", fname0);
|
||||
}
|
||||
if (!sdkLoadPPM4ub(fname0, &h_img0, &w, &h)) {
|
||||
fprintf(stderr, "Failed to load <%s>\n", fname0);
|
||||
}
|
||||
|
||||
printf("Loaded <%s> as image 1\n", fname1);
|
||||
printf("Loaded <%s> as image 1\n", fname1);
|
||||
|
||||
if (!sdkLoadPPM4ub(fname1, &h_img1, &w, &h)) {
|
||||
fprintf(stderr, "Failed to load <%s>\n", fname1);
|
||||
}
|
||||
if (!sdkLoadPPM4ub(fname1, &h_img1, &w, &h)) {
|
||||
fprintf(stderr, "Failed to load <%s>\n", fname1);
|
||||
}
|
||||
|
||||
dim3 numThreads = dim3(blockSize_x, blockSize_y, 1);
|
||||
dim3 numBlocks = dim3(iDivUp(w, numThreads.x), iDivUp(h, numThreads.y));
|
||||
unsigned int numData = w * h;
|
||||
unsigned int memSize = sizeof(int) * numData;
|
||||
dim3 numThreads = dim3(blockSize_x, blockSize_y, 1);
|
||||
dim3 numBlocks = dim3(iDivUp(w, numThreads.x), iDivUp(h, numThreads.y));
|
||||
unsigned int numData = w * h;
|
||||
unsigned int memSize = sizeof(int) * numData;
|
||||
|
||||
// allocate mem for the result on host side
|
||||
unsigned int *h_odata = (unsigned int *)malloc(memSize);
|
||||
// allocate mem for the result on host side
|
||||
unsigned int *h_odata = (unsigned int *)malloc(memSize);
|
||||
|
||||
// initialize the memory
|
||||
for (unsigned int i = 0; i < numData; i++) h_odata[i] = 0;
|
||||
// initialize the memory
|
||||
for (unsigned int i = 0; i < numData; i++)
|
||||
h_odata[i] = 0;
|
||||
|
||||
// allocate device memory for result
|
||||
unsigned int *d_odata, *d_img0, *d_img1;
|
||||
// allocate device memory for result
|
||||
unsigned int *d_odata, *d_img0, *d_img1;
|
||||
|
||||
checkCudaErrors(cudaMalloc((void **)&d_odata, memSize));
|
||||
checkCudaErrors(cudaMalloc((void **)&d_img0, memSize));
|
||||
checkCudaErrors(cudaMalloc((void **)&d_img1, memSize));
|
||||
checkCudaErrors(cudaMalloc((void **)&d_odata, memSize));
|
||||
checkCudaErrors(cudaMalloc((void **)&d_img0, memSize));
|
||||
checkCudaErrors(cudaMalloc((void **)&d_img1, memSize));
|
||||
|
||||
// copy host memory to device to initialize to zeros
|
||||
checkCudaErrors(cudaMemcpy(d_img0, h_img0, memSize, cudaMemcpyHostToDevice));
|
||||
checkCudaErrors(cudaMemcpy(d_img1, h_img1, memSize, cudaMemcpyHostToDevice));
|
||||
checkCudaErrors(
|
||||
cudaMemcpy(d_odata, h_odata, memSize, cudaMemcpyHostToDevice));
|
||||
// copy host memory to device to initialize to zeros
|
||||
checkCudaErrors(cudaMemcpy(d_img0, h_img0, memSize, cudaMemcpyHostToDevice));
|
||||
checkCudaErrors(cudaMemcpy(d_img1, h_img1, memSize, cudaMemcpyHostToDevice));
|
||||
checkCudaErrors(cudaMemcpy(d_odata, h_odata, memSize, cudaMemcpyHostToDevice));
|
||||
|
||||
cudaChannelFormatDesc ca_desc0 = cudaCreateChannelDesc<unsigned int>();
|
||||
cudaChannelFormatDesc ca_desc1 = cudaCreateChannelDesc<unsigned int>();
|
||||
cudaChannelFormatDesc ca_desc0 = cudaCreateChannelDesc<unsigned int>();
|
||||
cudaChannelFormatDesc ca_desc1 = cudaCreateChannelDesc<unsigned int>();
|
||||
|
||||
cudaTextureObject_t tex2Dleft, tex2Dright;
|
||||
cudaResourceDesc texRes;
|
||||
memset(&texRes, 0, sizeof(cudaResourceDesc));
|
||||
cudaTextureObject_t tex2Dleft, tex2Dright;
|
||||
cudaResourceDesc texRes;
|
||||
memset(&texRes, 0, sizeof(cudaResourceDesc));
|
||||
|
||||
texRes.resType = cudaResourceTypePitch2D;
|
||||
texRes.res.pitch2D.devPtr = d_img0;
|
||||
texRes.res.pitch2D.desc = ca_desc0;
|
||||
texRes.res.pitch2D.width = w;
|
||||
texRes.res.pitch2D.height = h;
|
||||
texRes.res.pitch2D.pitchInBytes = w * 4;
|
||||
texRes.resType = cudaResourceTypePitch2D;
|
||||
texRes.res.pitch2D.devPtr = d_img0;
|
||||
texRes.res.pitch2D.desc = ca_desc0;
|
||||
texRes.res.pitch2D.width = w;
|
||||
texRes.res.pitch2D.height = h;
|
||||
texRes.res.pitch2D.pitchInBytes = w * 4;
|
||||
|
||||
cudaTextureDesc texDescr;
|
||||
memset(&texDescr, 0, sizeof(cudaTextureDesc));
|
||||
cudaTextureDesc texDescr;
|
||||
memset(&texDescr, 0, sizeof(cudaTextureDesc));
|
||||
|
||||
texDescr.normalizedCoords = false;
|
||||
texDescr.filterMode = cudaFilterModePoint;
|
||||
texDescr.addressMode[0] = cudaAddressModeClamp;
|
||||
texDescr.addressMode[1] = cudaAddressModeClamp;
|
||||
texDescr.readMode = cudaReadModeElementType;
|
||||
texDescr.normalizedCoords = false;
|
||||
texDescr.filterMode = cudaFilterModePoint;
|
||||
texDescr.addressMode[0] = cudaAddressModeClamp;
|
||||
texDescr.addressMode[1] = cudaAddressModeClamp;
|
||||
texDescr.readMode = cudaReadModeElementType;
|
||||
|
||||
checkCudaErrors(
|
||||
cudaCreateTextureObject(&tex2Dleft, &texRes, &texDescr, NULL));
|
||||
checkCudaErrors(cudaCreateTextureObject(&tex2Dleft, &texRes, &texDescr, NULL));
|
||||
|
||||
memset(&texRes, 0, sizeof(cudaResourceDesc));
|
||||
memset(&texRes, 0, sizeof(cudaResourceDesc));
|
||||
|
||||
texRes.resType = cudaResourceTypePitch2D;
|
||||
texRes.res.pitch2D.devPtr = d_img1;
|
||||
texRes.res.pitch2D.desc = ca_desc1;
|
||||
texRes.res.pitch2D.width = w;
|
||||
texRes.res.pitch2D.height = h;
|
||||
texRes.res.pitch2D.pitchInBytes = w * 4;
|
||||
texRes.resType = cudaResourceTypePitch2D;
|
||||
texRes.res.pitch2D.devPtr = d_img1;
|
||||
texRes.res.pitch2D.desc = ca_desc1;
|
||||
texRes.res.pitch2D.width = w;
|
||||
texRes.res.pitch2D.height = h;
|
||||
texRes.res.pitch2D.pitchInBytes = w * 4;
|
||||
|
||||
memset(&texDescr, 0, sizeof(cudaTextureDesc));
|
||||
memset(&texDescr, 0, sizeof(cudaTextureDesc));
|
||||
|
||||
texDescr.normalizedCoords = false;
|
||||
texDescr.filterMode = cudaFilterModePoint;
|
||||
texDescr.addressMode[0] = cudaAddressModeClamp;
|
||||
texDescr.addressMode[1] = cudaAddressModeClamp;
|
||||
texDescr.readMode = cudaReadModeElementType;
|
||||
texDescr.normalizedCoords = false;
|
||||
texDescr.filterMode = cudaFilterModePoint;
|
||||
texDescr.addressMode[0] = cudaAddressModeClamp;
|
||||
texDescr.addressMode[1] = cudaAddressModeClamp;
|
||||
texDescr.readMode = cudaReadModeElementType;
|
||||
|
||||
checkCudaErrors(
|
||||
cudaCreateTextureObject(&tex2Dright, &texRes, &texDescr, NULL));
|
||||
checkCudaErrors(cudaCreateTextureObject(&tex2Dright, &texRes, &texDescr, NULL));
|
||||
|
||||
// First run the warmup kernel (which we'll use to get the GPU in the correct
|
||||
// max power state
|
||||
stereoDisparityKernel<<<numBlocks, numThreads>>>(
|
||||
d_img0, d_img1, d_odata, w, h, minDisp, maxDisp, tex2Dleft, tex2Dright);
|
||||
cudaDeviceSynchronize();
|
||||
// First run the warmup kernel (which we'll use to get the GPU in the correct
|
||||
// max power state
|
||||
stereoDisparityKernel<<<numBlocks, numThreads>>>(
|
||||
d_img0, d_img1, d_odata, w, h, minDisp, maxDisp, tex2Dleft, tex2Dright);
|
||||
cudaDeviceSynchronize();
|
||||
|
||||
// Allocate CUDA events that we'll use for timing
|
||||
cudaEvent_t start, stop;
|
||||
checkCudaErrors(cudaEventCreate(&start));
|
||||
checkCudaErrors(cudaEventCreate(&stop));
|
||||
// Allocate CUDA events that we'll use for timing
|
||||
cudaEvent_t start, stop;
|
||||
checkCudaErrors(cudaEventCreate(&start));
|
||||
checkCudaErrors(cudaEventCreate(&stop));
|
||||
|
||||
printf("Launching CUDA stereoDisparityKernel()\n");
|
||||
printf("Launching CUDA stereoDisparityKernel()\n");
|
||||
|
||||
// Record the start event
|
||||
checkCudaErrors(cudaEventRecord(start, NULL));
|
||||
// Record the start event
|
||||
checkCudaErrors(cudaEventRecord(start, NULL));
|
||||
|
||||
// launch the stereoDisparity kernel
|
||||
stereoDisparityKernel<<<numBlocks, numThreads>>>(
|
||||
d_img0, d_img1, d_odata, w, h, minDisp, maxDisp, tex2Dleft, tex2Dright);
|
||||
// launch the stereoDisparity kernel
|
||||
stereoDisparityKernel<<<numBlocks, numThreads>>>(
|
||||
d_img0, d_img1, d_odata, w, h, minDisp, maxDisp, tex2Dleft, tex2Dright);
|
||||
|
||||
// Record the stop event
|
||||
checkCudaErrors(cudaEventRecord(stop, NULL));
|
||||
// Record the stop event
|
||||
checkCudaErrors(cudaEventRecord(stop, NULL));
|
||||
|
||||
// Wait for the stop event to complete
|
||||
checkCudaErrors(cudaEventSynchronize(stop));
|
||||
// Wait for the stop event to complete
|
||||
checkCudaErrors(cudaEventSynchronize(stop));
|
||||
|
||||
// Check to make sure the kernel didn't fail
|
||||
getLastCudaError("Kernel execution failed");
|
||||
// Check to make sure the kernel didn't fail
|
||||
getLastCudaError("Kernel execution failed");
|
||||
|
||||
float msecTotal = 0.0f;
|
||||
checkCudaErrors(cudaEventElapsedTime(&msecTotal, start, stop));
|
||||
float msecTotal = 0.0f;
|
||||
checkCudaErrors(cudaEventElapsedTime(&msecTotal, start, stop));
|
||||
|
||||
// Copy result from device to host for verification
|
||||
checkCudaErrors(
|
||||
cudaMemcpy(h_odata, d_odata, memSize, cudaMemcpyDeviceToHost));
|
||||
// Copy result from device to host for verification
|
||||
checkCudaErrors(cudaMemcpy(h_odata, d_odata, memSize, cudaMemcpyDeviceToHost));
|
||||
|
||||
printf("Input Size [%dx%d], ", w, h);
|
||||
printf("Kernel size [%dx%d], ", (2 * RAD + 1), (2 * RAD + 1));
|
||||
printf("Disparities [%d:%d]\n", minDisp, maxDisp);
|
||||
printf("Input Size [%dx%d], ", w, h);
|
||||
printf("Kernel size [%dx%d], ", (2 * RAD + 1), (2 * RAD + 1));
|
||||
printf("Disparities [%d:%d]\n", minDisp, maxDisp);
|
||||
|
||||
printf("GPU processing time : %.4f (ms)\n", msecTotal);
|
||||
printf("Pixel throughput : %.3f Mpixels/sec\n",
|
||||
((float)(w * h * 1000.f) / msecTotal) / 1000000);
|
||||
printf("GPU processing time : %.4f (ms)\n", msecTotal);
|
||||
printf("Pixel throughput : %.3f Mpixels/sec\n", ((float)(w * h * 1000.f) / msecTotal) / 1000000);
|
||||
|
||||
// calculate sum of resultant GPU image
|
||||
unsigned int checkSum = 0;
|
||||
// calculate sum of resultant GPU image
|
||||
unsigned int checkSum = 0;
|
||||
|
||||
for (unsigned int i = 0; i < w * h; i++) {
|
||||
checkSum += h_odata[i];
|
||||
}
|
||||
for (unsigned int i = 0; i < w * h; i++) {
|
||||
checkSum += h_odata[i];
|
||||
}
|
||||
|
||||
printf("GPU Checksum = %u, ", checkSum);
|
||||
printf("GPU Checksum = %u, ", checkSum);
|
||||
|
||||
// write out the resulting disparity image.
|
||||
unsigned char *dispOut = (unsigned char *)malloc(numData);
|
||||
int mult = 20;
|
||||
const char *fnameOut = "output_GPU.pgm";
|
||||
// write out the resulting disparity image.
|
||||
unsigned char *dispOut = (unsigned char *)malloc(numData);
|
||||
int mult = 20;
|
||||
const char *fnameOut = "output_GPU.pgm";
|
||||
|
||||
for (unsigned int i = 0; i < numData; i++) {
|
||||
dispOut[i] = (int)h_odata[i] * mult;
|
||||
}
|
||||
for (unsigned int i = 0; i < numData; i++) {
|
||||
dispOut[i] = (int)h_odata[i] * mult;
|
||||
}
|
||||
|
||||
printf("GPU image: <%s>\n", fnameOut);
|
||||
sdkSavePGM(fnameOut, dispOut, w, h);
|
||||
printf("GPU image: <%s>\n", fnameOut);
|
||||
sdkSavePGM(fnameOut, dispOut, w, h);
|
||||
|
||||
// compute reference solution
|
||||
printf("Computing CPU reference...\n");
|
||||
cpu_gold_stereo((unsigned int *)h_img0, (unsigned int *)h_img1,
|
||||
(unsigned int *)h_odata, w, h, minDisp, maxDisp);
|
||||
unsigned int cpuCheckSum = 0;
|
||||
// compute reference solution
|
||||
printf("Computing CPU reference...\n");
|
||||
cpu_gold_stereo((unsigned int *)h_img0, (unsigned int *)h_img1, (unsigned int *)h_odata, w, h, minDisp, maxDisp);
|
||||
unsigned int cpuCheckSum = 0;
|
||||
|
||||
for (unsigned int i = 0; i < w * h; i++) {
|
||||
cpuCheckSum += h_odata[i];
|
||||
}
|
||||
for (unsigned int i = 0; i < w * h; i++) {
|
||||
cpuCheckSum += h_odata[i];
|
||||
}
|
||||
|
||||
printf("CPU Checksum = %u, ", cpuCheckSum);
|
||||
const char *cpuFnameOut = "output_CPU.pgm";
|
||||
printf("CPU Checksum = %u, ", cpuCheckSum);
|
||||
const char *cpuFnameOut = "output_CPU.pgm";
|
||||
|
||||
for (unsigned int i = 0; i < numData; i++) {
|
||||
dispOut[i] = (int)h_odata[i] * mult;
|
||||
}
|
||||
for (unsigned int i = 0; i < numData; i++) {
|
||||
dispOut[i] = (int)h_odata[i] * mult;
|
||||
}
|
||||
|
||||
printf("CPU image: <%s>\n", cpuFnameOut);
|
||||
sdkSavePGM(cpuFnameOut, dispOut, w, h);
|
||||
printf("CPU image: <%s>\n", cpuFnameOut);
|
||||
sdkSavePGM(cpuFnameOut, dispOut, w, h);
|
||||
|
||||
// cleanup memory
|
||||
checkCudaErrors(cudaFree(d_odata));
|
||||
checkCudaErrors(cudaFree(d_img0));
|
||||
checkCudaErrors(cudaFree(d_img1));
|
||||
// cleanup memory
|
||||
checkCudaErrors(cudaFree(d_odata));
|
||||
checkCudaErrors(cudaFree(d_img0));
|
||||
checkCudaErrors(cudaFree(d_img1));
|
||||
|
||||
if (h_odata != NULL) free(h_odata);
|
||||
if (h_odata != NULL)
|
||||
free(h_odata);
|
||||
|
||||
if (h_img0 != NULL) free(h_img0);
|
||||
if (h_img0 != NULL)
|
||||
free(h_img0);
|
||||
|
||||
if (h_img1 != NULL) free(h_img1);
|
||||
if (h_img1 != NULL)
|
||||
free(h_img1);
|
||||
|
||||
if (dispOut != NULL) free(dispOut);
|
||||
if (dispOut != NULL)
|
||||
free(dispOut);
|
||||
|
||||
sdkDeleteTimer(&timer);
|
||||
sdkDeleteTimer(&timer);
|
||||
|
||||
exit((checkSum == cpuCheckSum) ? EXIT_SUCCESS : EXIT_FAILURE);
|
||||
exit((checkSum == cpuCheckSum) ? EXIT_SUCCESS : EXIT_FAILURE);
|
||||
}
|
||||
|
||||
@@ -54,18 +54,17 @@ namespace cg = cooperative_groups;
|
||||
// "ptx_isa_3.0K.pdf"
|
||||
// included in the NVIDIA GPU Computing Toolkit
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
__device__ unsigned int __usad4(unsigned int A, unsigned int B,
|
||||
unsigned int C = 0) {
|
||||
unsigned int result;
|
||||
__device__ unsigned int __usad4(unsigned int A, unsigned int B, unsigned int C = 0)
|
||||
{
|
||||
unsigned int result;
|
||||
|
||||
// Kepler (SM 3.x) and higher supports a 4 vector SAD SIMD
|
||||
asm(
|
||||
"vabsdiff4.u32.u32.u32.add"
|
||||
" %0, %1, %2, %3;"
|
||||
: "=r"(result)
|
||||
: "r"(A), "r"(B), "r"(C));
|
||||
// Kepler (SM 3.x) and higher supports a 4 vector SAD SIMD
|
||||
asm("vabsdiff4.u32.u32.u32.add"
|
||||
" %0, %1, %2, %3;"
|
||||
: "=r"(result)
|
||||
: "r"(A), "r"(B), "r"(C));
|
||||
|
||||
return result;
|
||||
return result;
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -88,165 +87,178 @@ __device__ unsigned int __usad4(unsigned int A, unsigned int B,
|
||||
//! @param minDisparity leftmost search range
|
||||
//! @param maxDisparity rightmost search range
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
__global__ void stereoDisparityKernel(unsigned int *g_img0,
|
||||
unsigned int *g_img1,
|
||||
unsigned int *g_odata, int w, int h,
|
||||
int minDisparity, int maxDisparity,
|
||||
__global__ void stereoDisparityKernel(unsigned int *g_img0,
|
||||
unsigned int *g_img1,
|
||||
unsigned int *g_odata,
|
||||
int w,
|
||||
int h,
|
||||
int minDisparity,
|
||||
int maxDisparity,
|
||||
cudaTextureObject_t tex2Dleft,
|
||||
cudaTextureObject_t tex2Dright) {
|
||||
// Handle to thread block group
|
||||
cg::thread_block cta = cg::this_thread_block();
|
||||
// access thread id
|
||||
const int tidx = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
const int tidy = blockDim.y * blockIdx.y + threadIdx.y;
|
||||
const unsigned int sidx = threadIdx.x + RAD;
|
||||
const unsigned int sidy = threadIdx.y + RAD;
|
||||
cudaTextureObject_t tex2Dright)
|
||||
{
|
||||
// Handle to thread block group
|
||||
cg::thread_block cta = cg::this_thread_block();
|
||||
// access thread id
|
||||
const int tidx = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
const int tidy = blockDim.y * blockIdx.y + threadIdx.y;
|
||||
const unsigned int sidx = threadIdx.x + RAD;
|
||||
const unsigned int sidy = threadIdx.y + RAD;
|
||||
|
||||
unsigned int imLeft;
|
||||
unsigned int imRight;
|
||||
unsigned int cost;
|
||||
unsigned int bestCost = 9999999;
|
||||
unsigned int bestDisparity = 0;
|
||||
__shared__ unsigned int diff[blockSize_y + 2 * RAD][blockSize_x + 2 * RAD];
|
||||
unsigned int imLeft;
|
||||
unsigned int imRight;
|
||||
unsigned int cost;
|
||||
unsigned int bestCost = 9999999;
|
||||
unsigned int bestDisparity = 0;
|
||||
__shared__ unsigned int diff[blockSize_y + 2 * RAD][blockSize_x + 2 * RAD];
|
||||
|
||||
// store needed values for left image into registers (constant indexed local
|
||||
// vars)
|
||||
unsigned int imLeftA[STEPS];
|
||||
unsigned int imLeftB[STEPS];
|
||||
// store needed values for left image into registers (constant indexed local
|
||||
// vars)
|
||||
unsigned int imLeftA[STEPS];
|
||||
unsigned int imLeftB[STEPS];
|
||||
|
||||
for (int i = 0; i < STEPS; i++) {
|
||||
int offset = -RAD + i * RAD;
|
||||
imLeftA[i] = tex2D<unsigned int>(tex2Dleft, tidx - RAD, tidy + offset);
|
||||
imLeftB[i] =
|
||||
tex2D<unsigned int>(tex2Dleft, tidx - RAD + blockSize_x, tidy + offset);
|
||||
}
|
||||
for (int i = 0; i < STEPS; i++) {
|
||||
int offset = -RAD + i * RAD;
|
||||
imLeftA[i] = tex2D<unsigned int>(tex2Dleft, tidx - RAD, tidy + offset);
|
||||
imLeftB[i] = tex2D<unsigned int>(tex2Dleft, tidx - RAD + blockSize_x, tidy + offset);
|
||||
}
|
||||
|
||||
// for a fixed camera system this could be hardcoded and loop unrolled
|
||||
for (int d = minDisparity; d <= maxDisparity; d++) {
|
||||
// for a fixed camera system this could be hardcoded and loop unrolled
|
||||
for (int d = minDisparity; d <= maxDisparity; d++) {
|
||||
// LEFT
|
||||
#pragma unroll
|
||||
for (int i = 0; i < STEPS; i++) {
|
||||
int offset = -RAD + i * RAD;
|
||||
// imLeft = tex2D( tex2Dleft, tidx-RAD, tidy+offset );
|
||||
imLeft = imLeftA[i];
|
||||
imRight = tex2D<unsigned int>(tex2Dright, tidx - RAD + d, tidy + offset);
|
||||
cost = __usad4(imLeft, imRight);
|
||||
diff[sidy + offset][sidx - RAD] = cost;
|
||||
}
|
||||
for (int i = 0; i < STEPS; i++) {
|
||||
int offset = -RAD + i * RAD;
|
||||
// imLeft = tex2D( tex2Dleft, tidx-RAD, tidy+offset );
|
||||
imLeft = imLeftA[i];
|
||||
imRight = tex2D<unsigned int>(tex2Dright, tidx - RAD + d, tidy + offset);
|
||||
cost = __usad4(imLeft, imRight);
|
||||
diff[sidy + offset][sidx - RAD] = cost;
|
||||
}
|
||||
|
||||
// RIGHT
|
||||
#pragma unroll
|
||||
|
||||
for (int i = 0; i < STEPS; i++) {
|
||||
int offset = -RAD + i * RAD;
|
||||
for (int i = 0; i < STEPS; i++) {
|
||||
int offset = -RAD + i * RAD;
|
||||
|
||||
if (threadIdx.x < 2 * RAD) {
|
||||
// imLeft = tex2D( tex2Dleft, tidx-RAD+blockSize_x, tidy+offset );
|
||||
imLeft = imLeftB[i];
|
||||
imRight = tex2D<unsigned int>(tex2Dright, tidx - RAD + blockSize_x + d,
|
||||
tidy + offset);
|
||||
cost = __usad4(imLeft, imRight);
|
||||
diff[sidy + offset][sidx - RAD + blockSize_x] = cost;
|
||||
}
|
||||
}
|
||||
if (threadIdx.x < 2 * RAD) {
|
||||
// imLeft = tex2D( tex2Dleft, tidx-RAD+blockSize_x, tidy+offset );
|
||||
imLeft = imLeftB[i];
|
||||
imRight = tex2D<unsigned int>(tex2Dright, tidx - RAD + blockSize_x + d, tidy + offset);
|
||||
cost = __usad4(imLeft, imRight);
|
||||
diff[sidy + offset][sidx - RAD + blockSize_x] = cost;
|
||||
}
|
||||
}
|
||||
|
||||
cg::sync(cta);
|
||||
cg::sync(cta);
|
||||
|
||||
// sum cost horizontally
|
||||
#pragma unroll
|
||||
|
||||
for (int j = 0; j < STEPS; j++) {
|
||||
int offset = -RAD + j * RAD;
|
||||
cost = 0;
|
||||
for (int j = 0; j < STEPS; j++) {
|
||||
int offset = -RAD + j * RAD;
|
||||
cost = 0;
|
||||
#pragma unroll
|
||||
|
||||
for (int i = -RAD; i <= RAD; i++) {
|
||||
cost += diff[sidy + offset][sidx + i];
|
||||
}
|
||||
|
||||
cg::sync(cta);
|
||||
diff[sidy + offset][sidx] = cost;
|
||||
cg::sync(cta);
|
||||
}
|
||||
|
||||
// sum cost vertically
|
||||
cost = 0;
|
||||
#pragma unroll
|
||||
|
||||
for (int i = -RAD; i <= RAD; i++) {
|
||||
cost += diff[sidy + i][sidx];
|
||||
}
|
||||
|
||||
// see if it is better or not
|
||||
if (cost < bestCost) {
|
||||
bestCost = cost;
|
||||
bestDisparity = d + 8;
|
||||
}
|
||||
|
||||
cg::sync(cta);
|
||||
}
|
||||
|
||||
if (tidy < h && tidx < w) {
|
||||
g_odata[tidy * w + tidx] = bestDisparity;
|
||||
}
|
||||
}
|
||||
|
||||
void cpu_gold_stereo(unsigned int *img0, unsigned int *img1,
|
||||
unsigned int *odata, int w, int h, int minDisparity,
|
||||
int maxDisparity) {
|
||||
for (int y = 0; y < h; y++) {
|
||||
for (int x = 0; x < w; x++) {
|
||||
unsigned int bestCost = 9999999;
|
||||
unsigned int bestDisparity = 0;
|
||||
|
||||
for (int d = minDisparity; d <= maxDisparity; d++) {
|
||||
unsigned int cost = 0;
|
||||
|
||||
for (int i = -RAD; i <= RAD; i++) {
|
||||
for (int j = -RAD; j <= RAD; j++) {
|
||||
// border clamping
|
||||
int yy, xx, xxd;
|
||||
yy = y + i;
|
||||
|
||||
if (yy < 0) yy = 0;
|
||||
|
||||
if (yy >= h) yy = h - 1;
|
||||
|
||||
xx = x + j;
|
||||
|
||||
if (xx < 0) xx = 0;
|
||||
|
||||
if (xx >= w) xx = w - 1;
|
||||
|
||||
xxd = x + j + d;
|
||||
|
||||
if (xxd < 0) xxd = 0;
|
||||
|
||||
if (xxd >= w) xxd = w - 1;
|
||||
|
||||
// sum abs diff across components
|
||||
unsigned char *A = (unsigned char *)&img0[yy * w + xx];
|
||||
unsigned char *B = (unsigned char *)&img1[yy * w + xxd];
|
||||
unsigned int absdiff = 0;
|
||||
|
||||
for (int k = 0; k < 4; k++) {
|
||||
absdiff += abs((int)(A[k] - B[k]));
|
||||
for (int i = -RAD; i <= RAD; i++) {
|
||||
cost += diff[sidy + offset][sidx + i];
|
||||
}
|
||||
|
||||
cost += absdiff;
|
||||
}
|
||||
cg::sync(cta);
|
||||
diff[sidy + offset][sidx] = cost;
|
||||
cg::sync(cta);
|
||||
}
|
||||
|
||||
// sum cost vertically
|
||||
cost = 0;
|
||||
#pragma unroll
|
||||
|
||||
for (int i = -RAD; i <= RAD; i++) {
|
||||
cost += diff[sidy + i][sidx];
|
||||
}
|
||||
|
||||
// see if it is better or not
|
||||
if (cost < bestCost) {
|
||||
bestCost = cost;
|
||||
bestDisparity = d + 8;
|
||||
bestCost = cost;
|
||||
bestDisparity = d + 8;
|
||||
}
|
||||
|
||||
} // end for disparities
|
||||
|
||||
// store to best disparity
|
||||
odata[y * w + x] = bestDisparity;
|
||||
cg::sync(cta);
|
||||
}
|
||||
|
||||
if (tidy < h && tidx < w) {
|
||||
g_odata[tidy * w + tidx] = bestDisparity;
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif // #ifndef _STEREODISPARITY_KERNEL_H_
|
||||
|
||||
void cpu_gold_stereo(unsigned int *img0,
|
||||
unsigned int *img1,
|
||||
unsigned int *odata,
|
||||
int w,
|
||||
int h,
|
||||
int minDisparity,
|
||||
int maxDisparity)
|
||||
{
|
||||
for (int y = 0; y < h; y++) {
|
||||
for (int x = 0; x < w; x++) {
|
||||
unsigned int bestCost = 9999999;
|
||||
unsigned int bestDisparity = 0;
|
||||
|
||||
for (int d = minDisparity; d <= maxDisparity; d++) {
|
||||
unsigned int cost = 0;
|
||||
|
||||
for (int i = -RAD; i <= RAD; i++) {
|
||||
for (int j = -RAD; j <= RAD; j++) {
|
||||
// border clamping
|
||||
int yy, xx, xxd;
|
||||
yy = y + i;
|
||||
|
||||
if (yy < 0)
|
||||
yy = 0;
|
||||
|
||||
if (yy >= h)
|
||||
yy = h - 1;
|
||||
|
||||
xx = x + j;
|
||||
|
||||
if (xx < 0)
|
||||
xx = 0;
|
||||
|
||||
if (xx >= w)
|
||||
xx = w - 1;
|
||||
|
||||
xxd = x + j + d;
|
||||
|
||||
if (xxd < 0)
|
||||
xxd = 0;
|
||||
|
||||
if (xxd >= w)
|
||||
xxd = w - 1;
|
||||
|
||||
// sum abs diff across components
|
||||
unsigned char *A = (unsigned char *)&img0[yy * w + xx];
|
||||
unsigned char *B = (unsigned char *)&img1[yy * w + xxd];
|
||||
unsigned int absdiff = 0;
|
||||
|
||||
for (int k = 0; k < 4; k++) {
|
||||
absdiff += abs((int)(A[k] - B[k]));
|
||||
}
|
||||
|
||||
cost += absdiff;
|
||||
}
|
||||
}
|
||||
|
||||
if (cost < bestCost) {
|
||||
bestCost = cost;
|
||||
bestDisparity = d + 8;
|
||||
}
|
||||
|
||||
} // end for disparities
|
||||
|
||||
// store to best disparity
|
||||
odata[y * w + x] = bestDisparity;
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif // #ifndef _STEREODISPARITY_KERNEL_H_
|
||||
|
||||
Reference in New Issue
Block a user