CUDA小白 - NPP(2) - Arithmetic and Logical Operations(1)

cuda小白
原文链接 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。

如有问题，请指出，谢谢

Arithmetic Operations

当前模块主要是加减乘除，abs，平方，矩阵相乘，开根，ln，exp等。不同相同功能，以其中一个为例进行介绍。

AddC

针对图像中每一个像素添加一个常量值。与之类似的还有MulC，SubC，DivC，AbsDiffC，MulScale
大概接口样式：

// 两个结果的区别在于，有I的结果可以就原始的图像地址进行操作，无需进行拷贝
// Sfs的含义 表示可以对图像的数值范围进行约束操作。
NppStatus nppiAddC_[数据类型]_C[通道数]RSfs_[是否使用流]
NppStatus nppiAddC_[数据类型]_C[通道数]IRSfs_[是否使用流]

以为三通道的uint8_t的图像数据为例子：

NppStatus nppiAddC_8u_C3RSfs(const Npp8u * pSrc1,int nSrc1Step,const Npp8u aConstants[3],Npp8u *pDst,int nDstStep,NppiSize oSizeROI,int nScaleFactor);
NppStatus nppiAddC_8u_C3RSfs(const Npp8u aConstants[3],Npp8u *pDst,int nDstStep,NppiSize oSizeROI,int nScaleFactor);	

code

#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 = "../";// =============== load image ===============cv::Mat image = cv::imread(directory + "dog.png");if (image.empty()) {std::cout << "Load image error!" << std::endl;return -1;}int image_width = image.cols;int image_height = image.rows;int image_size = image_width * image_height * 3 * sizeof(uint8_t);std::cout << "Image info : image_width = " << image_width<< ", image_height = " << image_height << std::endl;// =============== malloc && cpy ===============uint8_t *in_ptr, *in_ptr2, *out_ptr, *roi_out_ptr;cudaMalloc((void**)&in_ptr, image_size);cudaMalloc((void**)&in_ptr2, image_size);cudaMalloc((void**)&out_ptr, image_size);cudaMalloc((void**)&roi_out_ptr, image_size);cudaMemcpy(in_ptr, image.data, image_size, cudaMemcpyHostToDevice);cudaMemcpy(in_ptr2, image.data, image_size, cudaMemcpyHostToDevice);uint8_t host_constant[3] = { (uint8_t)0, (uint8_t)20, (uint8_t)0 };NppiSize roi1, roi2;roi1.width = image_width;roi1.height = image_height;roi2.width = image_width / 2;roi2.height = image_height / 2;// nppiAddC_8u_C3RSfscv::Mat out_image = cv::Mat::zeros(image_height, image_width, CV_8UC3);NppStatus status;status = nppiAddC_8u_C3RSfs(in_ptr, image_width * 3, host_constant, out_ptr, image_width * 3, roi1, 0);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAddC_8u_C3RSfs failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr, image_size, cudaMemcpyDeviceToHost);cv::imwrite(directory + "add_constant.jpg", out_image);status = nppiAddC_8u_C3RSfs(in_ptr, image_width * 3, host_constant, out_ptr, image_width * 3, roi1, 1);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAddC_8u_C3RSfs failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr, image_size, cudaMemcpyDeviceToHost);cv::imwrite(directory + "add_constant_scale.jpg", out_image);status = nppiAddC_8u_C3RSfs(in_ptr, image_width * 3, host_constant, out_ptr, image_width * 3, roi2, 0);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAddC_8u_C3RSfs failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr, image_size, cudaMemcpyDeviceToHost);cv::imwrite(directory + "add_constant_roi.jpg", out_image);// freeCUDA_FREE(in_ptr)CUDA_FREE(in_ptr2)CUDA_FREE(out_ptr)CUDA_FREE(roi_out_ptr)
}

make

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}
)

result

注意点：

1. 对图像的进行值的范围进行一定的约束，最初的图像RGB的值都是[0, 255]，如果对应的scale设置为1的时候，相当于将数值的范围变为2的-nScaleFactor倍数，即[0, 128]，超过128的会变成128，因此整体的图像色调会变暗；同样的，如果设置为-1，则整体的会变为2倍，因为存图的限制只能是[0, 255]，所以整体看起来会变亮。
2. 如果指定的roi不是整张图的时候，由于roi的类型是NppiSize（width, height），因此输入的指针不是指向图像的起始位置，而是roi的起始位置。
3. MulScale相较于MulC来说，默认nScaleFactor为0；
4. AbsDiffC中主要功能就是得到整张图像与host_constant的绝对差值。
5. 每个结果都有一个对应的包含cudastream的版本，按需使用。

Add

与AddC不同的是，Add输入的是两张图像。同样的还有Mul，MulScale，Sub，Div，Div_round，Abs，AbsDiff，Sqr，Sqrt，Ln，Exp。（由于Abs，AbsDiff，Sqr，Sqrt，Ln，Exp在图像出列方面使用的不是很多，就不细述）。
以uint8_t的三通道图像为例：

// 命名规则与nppiAddC*类似
NppStatus nppiAdd_8u_C3RSfs(const Npp8u * pSrc1,int nSrc1Step,const Npp8u *pSrc2,int nSrc2Step,Npp8u * pDst,int nDstStep,NppiSize oSizeROI,int nScaleFactor);
NppStatus nppiAdd_8u_C3IRSfs(const Npp8u *pSrc,int nSrcStep,Npp8u *pSrcDst,int nSrcDstStep,NppiSize oSizeROI,int nScaleFactor);

code

#include <iostream>
#include <cuda_runtime.h>
#include <npp.h>
#include <opencv2/opencv.hpp>#define PRINT_VALUE(value) {  \std::cout << "[GPU] " << #value << " = " << value << std::endl; }#define CUDA_FREE(ptr) { if (ptr != nullptr) { cudaFree(ptr); ptr = nullptr; } }int main() {std::string directory = "../";// =============== load image ===============cv::Mat image = cv::imread(directory + "dog.png");if (image.empty()) {std::cout << "Load image error!" << std::endl;return -1;}int image_width = image.cols;int image_height = image.rows;int image_size = image_width * image_height * 3 * sizeof(uint8_t);std::cout << "Image info : image_width = " << image_width<< ", image_height = " << image_height << std::endl;// =============== malloc && cpy ===============uint8_t *in_ptr, *in_ptr2, *out_ptr, *roi_out_ptr;cudaMalloc((void**)&in_ptr, image_size);cudaMalloc((void**)&in_ptr2, image_size);cudaMalloc((void**)&out_ptr, image_size);cudaMalloc((void**)&roi_out_ptr, image_size);cudaMemcpy(in_ptr, image.data, image_size, cudaMemcpyHostToDevice);cudaMemcpy(in_ptr2, image.data, image_size, cudaMemcpyHostToDevice);NppiSize roi1, roi2;roi1.width = image_width;roi1.height = image_height;roi2.width = image_width / 2;roi2.height = image_height / 2;// nppiAdd_8u_C3RSfscv::Mat out_image = cv::Mat::zeros(image_height, image_width, CV_8UC3);NppStatus status;status = nppiAdd_8u_C3RSfs(in_ptr, image_width * 3, in_ptr2, image_width * 3, out_ptr, image_width * 3, roi1, 0);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAdd_8u_C3RSfs failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr, image_size, cudaMemcpyDeviceToHost);cv::imwrite(directory + "add.jpg", out_image);status = nppiAdd_8u_C3RSfs(in_ptr, image_width * 3, in_ptr2, image_width * 3, out_ptr, image_width * 3, roi1, 1);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAdd_8u_C3RSfs failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr, image_size, cudaMemcpyDeviceToHost);cv::imwrite(directory + "add_scale.jpg", out_image);status = nppiAdd_8u_C3RSfs(in_ptr, image_width * 3, in_ptr2, image_width * 3, out_ptr, image_width * 3, roi2, 0);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAdd_8u_C3RSfs failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr, image_size, cudaMemcpyDeviceToHost);cv::imwrite(directory + "add_roi.jpg", out_image);// freeCUDA_FREE(in_ptr)CUDA_FREE(in_ptr2)CUDA_FREE(out_ptr)CUDA_FREE(roi_out_ptr)
}

make

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}
)

result

注意点：

1. nScaleFactor与AddC中的功能一致。roi的操作也与AddC中的一致。
2. 由于使用的是两个相同的图片进行相加，因此在nScaleFactor为1的时候，所有的数值都变成原来的值，因此保存的图像与原图一致。
3. 同样有cudastream版本，按需使用。

AddWeighted

将特定区域的图像进行填充weight

NppStatus nppiAddWeighted_8u32f_C1IR(const Npp8u *pSrc,int nSrcStep,Npp32f * pSrcDst,int nSrcDstStep,NppiSize oSizeROI,Npp32f nAlpha);
NppStatus nppiAddWeighted_8u32f_C1IMR(const Npp8u *pSrc,int nSrcStep,const Npp8u *pMask,int nMaskStep,Npp32f * pSrcDst,int nSrcDstStep,NppiSize oSizeROI,Npp32f nAlpha);	

code

#include <iostream>
#include <cuda_runtime.h>
#include <npp.h>
#include <opencv2/opencv.hpp>#define PRINT_VALUE(value) {  \std::cout << "[GPU] " << #value << " = " << value << std::endl; }#define CUDA_FREE(ptr) { if (ptr != nullptr) { cudaFree(ptr); ptr = nullptr; } }int main() {std::string directory = "../";// =============== load image ===============cv::Mat image = cv::imread(directory + "dog.png");if (image.empty()) {std::cout << "Load image error!" << std::endl;return -1;}cv::Mat gray;cv::cvtColor(image, gray, CV_BGR2GRAY);cv::imwrite(directory + "gray.jpg", gray);int image_width = gray.cols;int image_height = gray.rows;int image_size = image_width * image_height;std::cout << "Image info : image_width = " << image_width<< ", image_height = " << image_height << std::endl;cv::Mat mat_mask = cv::Mat::ones(image_height, image_width, CV_8UC1);cv::Rect rc_center = cv::Rect(image_width / 4, image_height / 4, image_width / 2, image_height / 2);mat_mask(rc_center) = cv::Mat::ones(image_height / 2, image_width / 2, CV_8UC1) * 255;cv::imwrite(directory + "mask.jpg", mat_mask);// =============== malloc && cpy ===============uint8_t *in_ptr, *mask;cudaMalloc((void**)&in_ptr, image_size * sizeof(uint8_t));cudaMalloc((void**)&mask, image_size * sizeof(uint8_t));cudaMemcpy(in_ptr, gray.data, image_size, cudaMemcpyHostToDevice);cudaMemcpy(mask, mat_mask.data, image_size, cudaMemcpyHostToDevice);float *out_ptr, *out_ptr1, *out_ptr2;cudaMalloc((void**)&out_ptr, image_size * sizeof(float));cudaMalloc((void**)&out_ptr1, image_size * sizeof(float));cudaMalloc((void**)&out_ptr2, image_size * sizeof(float));NppiSize roi1, roi2;roi1.width = image_width;roi1.height = image_height;roi2.width = image_width / 2;roi2.height = image_height / 2;// nppiAdd_8u_C3RSfscv::Mat out_image = cv::Mat::zeros(image_height, image_width, CV_32FC1);NppStatus status;status = nppiAddWeighted_8u32f_C1IMR(in_ptr, image_width * sizeof(uint8_t), mask, image_width * sizeof(uint8_t), out_ptr, image_width * sizeof(float), roi1, 1.0);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAddWeighted_8u32f_C1IMR failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr, image_size * sizeof(float), cudaMemcpyDeviceToHost);cv::imwrite(directory + "addweight.jpg", out_image);status = nppiAddWeighted_8u32f_C1IMR(in_ptr, image_width * sizeof(uint8_t), mask, image_width * sizeof(uint8_t), out_ptr1, image_width * sizeof(float), roi1, 0.5);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAddWeighted_8u32f_C1IMR failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr1, image_size * sizeof(float), cudaMemcpyDeviceToHost);cv::imwrite(directory + "addweight_scale.jpg", out_image);status = nppiAddWeighted_8u32f_C1IMR(in_ptr, image_width * sizeof(uint8_t), mask, image_width * sizeof(uint8_t), out_ptr2, image_width * sizeof(float), roi2, 0.5);if (status != NPP_SUCCESS) {std::cout << "[GPU] ERROR nppiAddWeighted_8u32f_C1IMR failed, status = " << status << std::endl;return false;}cudaMemcpy(out_image.data, out_ptr2, image_size * sizeof(float), cudaMemcpyDeviceToHost);cv::imwrite(directory + "addweight_roi_scale.jpg", out_image);// freeCUDA_FREE(in_ptr)CUDA_FREE(mask)CUDA_FREE(out_ptr)CUDA_FREE(out_ptr1)CUDA_FREE(out_ptr2)
}

make

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}
)

result

注意点：
1.nAlpha是针对原图中的每一个像素的值需要添加的权重，mask仅影响目标位置中那些部分需要输出。
2. roi表示输入的区域约束。