cuda小白
原始API链接 NPP
GPU架构近些年也有不少的变化，具体的可以参考别的博主的介绍，都比较详细。还有一些cuda中的专有名词的含义，可以参考《详解CUDA的Context、Stream、Warp、SM、SP、Kernel、Block、Grid》
常见的NppStatus，可以看这里。
仿射变换
// 仿射变换
NppStatus nppiWarpAffine_8u_C3R(const Npp8u *pSrc,
NppiSize oSrcSize,
int nSrcStep,
NppiRect oSrcROI,
Npp8u *pDst,
int nDstStep,
NppiRect oDstROI,
const double aCoeffs[2][3],
int eInterpolation);
// 同样的2x3的矩阵的逆仿射变换
NppStatus nppiWarpAffineBack_8u_C3R(const Npp8u *pSrc,
NppiSize oSrcSize,
int nSrcStep,
NppiRect oSrcROI,
Npp8u *pDst,
int nDstStep,
NppiRect oDstROI,
const double aCoeffs[2][3],
int eInterpolation);
// 源图像的四边形 仿射变换到 目标的四边形，平时使用较少，后续不进行测试
NppStatus nppiWarpAffineQuad_8u_C3R(const Npp8u *pSrc,
NppiSize oSrcSize,
int nSrcStep,
NppiRect oSrcROI,
const double aSrcQuad[4][2],
Npp8u *pDst,
int nDstStep,
NppiRect oDstROI,
const double aDstQuad[4][2],
int eInterpolation);
#include <iostream>
#include <cuda_runtime.h>
#include <npp.h>
#include <opencv2/opencv.hpp>
#define CUDA_FREE(ptr) { if (ptr != nullptr) { cudaFree(ptr); ptr = nullptr; } }
int main() {
std::string directory = "../";
cv::Mat image_dog = cv::imread(directory + "dog.png");
int image_width = image_dog.cols;
int image_height = image_dog.rows;
int image_size = image_width * image_height;
// =============== device memory ===============
// input
uint8_t *in_image;
cudaMalloc((void**)&in_image, image_size * 3 * sizeof(uint8_t));
cudaMemcpy(in_image, image_dog.data, image_size * 3 * sizeof(uint8_t), cudaMemcpyHostToDevice);
// output
uint8_t *out_ptr1, *out_ptr2;
cudaMalloc((void**)&out_ptr1, image_size * 3 * sizeof(uint8_t)); // 三通道
cudaMalloc((void**)&out_ptr2, image_size * 3 * sizeof(uint8_t)); // 三通道
double angle = 30.0;
double scale = 0.6;
cv::Point center = cv::Point(image_width / 2, image_height / 2);
cv::Mat rot_mat = cv::getRotationMatrix2D(center, angle, scale);
double coeffs[2][3] = { rot_mat.at<double>(0, 0),
rot_mat.at<double>(0, 1),
rot_mat.at<double>(0, 2),
rot_mat.at<double>(1, 0),
rot_mat.at<double>(1, 1),
rot_mat.at<double>(1, 2)};
NppiSize in_size;
in_size.width = image_width;
in_size.height = image_height;
NppiRect rc;
rc.x = 0;
rc.y = 0;
rc.width = image_width;
rc.height = image_height;
cv::Mat out_image = cv::Mat::zeros(image_height, image_width, CV_8UC3);
NppStatus status;
// =============== nppiWarpAffine_8u_C3R ===============
status = nppiWarpAffine_8u_C3R(in_image, in_size, image_width * 3, rc, out_ptr1, image_width * 3,
rc, coeffs, NPPI_INTER_LINEAR);
if (status != NPP_SUCCESS) {
std::cout << "[GPU] ERROR nppiWarpAffine_8u_C3R failed, status = " << status << std::endl;
return false;
}
cudaMemcpy(out_image.data, out_ptr1, image_size * 3, cudaMemcpyDeviceToHost);
cv::imwrite(directory + "affine.jpg", out_image);
// =============== nppiWarpAffineBack_8u_C3R ===============
status = nppiWarpAffineBack_8u_C3R(out_ptr1, in_size, image_width * 3, rc, out_ptr2, image_width * 3,
rc, coeffs, NPPI_INTER_LINEAR);
if (status != NPP_SUCCESS) {
std::cout << "[GPU] ERROR nppiWarpAffineBack_8u_C3R failed, status = " << status << std::endl;
return false;
}
cudaMemcpy(out_image.data, out_ptr2, image_size * 3, cudaMemcpyDeviceToHost);
cv::imwrite(directory + "affine_back.jpg", out_image);
// free
CUDA_FREE(in_image)
CUDA_FREE(out_ptr1)
CUDA_FREE(out_ptr2)
}
cmake_minimum_required(VERSION 3.20)
project(test)
find_package(OpenCV REQUIRED)
include_directories(${OpenCV_INCLUDE_DIRS})
find_package(CUDA REQUIRED)
include_directories(${CUDA_INCLUDE_DIRS})
file(GLOB CUDA_LIBS "/usr/local/cuda/lib64/*.so")
add_executable(test test.cpp)
target_link_libraries(test
${OpenCV_LIBS}
${CUDA_LIBS}
)
透视变换
NppStatus nppiWarpPerspective_8u_C3R(const Npp8u *pSrc,
NppiSize oSrcSize,
int nSrcStep,
NppiRect oSrcROI,
Npp8u *pDst,
int nDstStep,
NppiRect oDstROI,
const double aCoeffs[3][3],
int eInterpolation);
NppStatus nppiWarpPerspectiveBack_8u_C3R(const Npp8u *pSrc,
NppiSize oSrcSize,
int nSrcStep,
NppiRect oSrcROI,
Npp8u *pDst,
int nDstStep,
NppiRect oDstROI,
const double aCoeffs[3][3],
int eInterpolation);
#include <iostream>
#include <cuda_runtime.h>
#include <npp.h>
#include <opencv2/opencv.hpp>
#define CUDA_FREE(ptr) { if (ptr != nullptr) { cudaFree(ptr); ptr = nullptr; } }
int main() {
std::string directory = "../";
cv::Mat image_dog = cv::imread(directory + "dog.png");
int image_width = image_dog.cols;
int image_height = image_dog.rows;
int image_size = image_width * image_height;
// =============== device memory ===============
// input
uint8_t *in_image;
cudaMalloc((void**)&in_image, image_size * 3 * sizeof(uint8_t));
cudaMemcpy(in_image, image_dog.data, image_size * 3 * sizeof(uint8_t), cudaMemcpyHostToDevice);
// output
uint8_t *out_ptr1, *out_ptr2;
cudaMalloc((void**)&out_ptr1, image_size * 3 * sizeof(uint8_t)); // 三通道
cudaMalloc((void**)&out_ptr2, image_size * 3 * sizeof(uint8_t)); // 三通道
cv::Point2f src_pts[4], dst_pts[4];
src_pts[0].x = 0.0;
src_pts[0].y = 0.0;
src_pts[1].x = image_width - 1.0;
src_pts[1].y = 0.0;
src_pts[2].x = 0.0;
src_pts[2].y = image_height - 1.0;
src_pts[3].x = image_width - 1.0;
src_pts[3].y = image_height - 1.0;
dst_pts[0].x = image_width * 0.05;
dst_pts[0].y = image_height * 0.05;
dst_pts[1].x = image_width * 0.9;
dst_pts[1].y = image_height * 0.1;
dst_pts[2].x = image_width * 0.2;
dst_pts[2].y = image_height * 0.8;
dst_pts[3].x = image_width * 0.85;
dst_pts[3].y = image_height * 0.85;
cv::Mat warp_mat = cv::getPerspectiveTransform(src_pts, dst_pts);
double coeffs[3][3] = { warp_mat.at<double>(0, 0),
warp_mat.at<double>(0, 1),
warp_mat.at<double>(0, 2),
warp_mat.at<double>(1, 0),
warp_mat.at<double>(1, 1),
warp_mat.at<double>(1, 2),
warp_mat.at<double>(2, 0),
warp_mat.at<double>(2, 1),
warp_mat.at<double>(2, 2) };
NppiSize in_size;
in_size.width = image_width;
in_size.height = image_height;
NppiRect rc;
rc.x = 0;
rc.y = 0;
rc.width = image_width;
rc.height = image_height;
cv::Mat out_image = cv::Mat::zeros(image_height, image_width, CV_8UC3);
NppStatus status;
// =============== nppiWarpPerspective_8u_C3R ===============
status = nppiWarpPerspective_8u_C3R(in_image, in_size, image_width * 3, rc, out_ptr1, image_width * 3,
rc, coeffs, NPPI_INTER_LINEAR);
if (status != NPP_SUCCESS) {
std::cout << "[GPU] ERROR nppiWarpPerspective_8u_C3R failed, status = " << status << std::endl;
return false;
}
cudaMemcpy(out_image.data, out_ptr1, image_size * 3, cudaMemcpyDeviceToHost);
cv::imwrite(directory + "perspective.jpg", out_image);
// =============== nppiWarpPerspectiveBack_8u_C3R ===============
status = nppiWarpPerspectiveBack_8u_C3R(out_ptr1, in_size, image_width * 3, rc, out_ptr2, image_width * 3,
rc, coeffs, NPPI_INTER_LINEAR);
if (status != NPP_SUCCESS) {
std::cout << "[GPU] ERROR nppiWarpPerspectiveBack_8u_C3R failed, status = " << status << std::endl;
return false;
}
cudaMemcpy(out_image.data, out_ptr2, image_size * 3, cudaMemcpyDeviceToHost);
cv::imwrite(directory + "perspective_back.jpg", out_image);
// free
CUDA_FREE(in_image)
CUDA_FREE(out_ptr1)
CUDA_FREE(out_ptr2)
}
cmake_minimum_required(VERSION 3.20)
project(test)
find_package(OpenCV REQUIRED)
include_directories(${OpenCV_INCLUDE_DIRS})
find_package(CUDA REQUIRED)
include_directories(${CUDA_INCLUDE_DIRS})
file(GLOB CUDA_LIBS "/usr/local/cuda/lib64/*.so")
add_executable(test test.cpp)
target_link_libraries(test
${OpenCV_LIBS}
${CUDA_LIBS}
)
