以下CUDA sample是分别用C++和CUDA实现的模拟热传导生成的图像，并对其中使用到的CUDA函数进行了解说，code参考了《GPU高性能编程CUDA实战》一书的第七章，各个文件内容如下：
funset.cpp:
#include "funset.hpp"heat_conduction.cpp:
#include <random>
#include <iostream>
#include <vector>
#include <memory>
#include <string>
#include "common.hpp"
#include <opencv2/opencv.hpp>
int test_heat_conduction()
{
const int width{ 1024 }, height = width;
cv::Mat mat1(height, width, CV_8UC4), mat2(height, width, CV_8UC4);
const float speed{ 0.25f }, max_temp{ 1.f }, min_temp{0.0001f};
float elapsed_time1{ 0.f }, elapsed_time2{ 0.f }; // milliseconds
// intialize the constant data
std::unique_ptr<float[]> temp(new float[width * height]);
for (int i = 0; i < width*height; ++i) {
temp[i] = 0;
int x = i % width;
int y = i / height;
if ((x>300) && (x<600) && (y>310) && (y<601))
temp[i] = max_temp;
}
temp[width * 100 + 100] = (max_temp + min_temp) / 2;
temp[width * 700 + 100] = min_temp;
temp[width * 300 + 300] = min_temp;
temp[width * 200 + 700] = min_temp;
for (int y = 800; y < 900; ++y) {
for (int x = 400; x < 500; ++x) {
temp[x + y * width] = min_temp;
}
}
int ret = heat_conduction_cpu(mat1.data, width, height, temp.get(), speed, &elapsed_time1);
if (ret != 0) PRINT_ERROR_INFO(heat_conduction_cpu);
ret = heat_conduction_gpu(mat2.data, width, height, temp.get(), speed, &elapsed_time2);
if (ret != 0) PRINT_ERROR_INFO(heat_conduction_gpu);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
cv::Vec4b val1 = mat1.at<cv::Vec4b>(y, x);
cv::Vec4b val2 = mat2.at<cv::Vec4b>(y, x);
for (int i = 0; i < 4; ++i) {
if (val1[i] != val2[i]) {
fprintf(stderr, "their values are different at (%d, %d), i: %d, val1: %d, val2: %d\n",
x, y, i, val1[i], val2[i]);
//return -1;
}
}
}
}
std::string save_image_name{ "E:/GitCode/CUDA_Test/heat_conduction.jpg" };
cv::resize(mat2, mat2, cv::Size(width / 2, height / 2), 0.f, 0.f, 2);
cv::imwrite(save_image_name, mat2);
fprintf(stderr, "test heat conduction: cpu run time: %f ms, gpu run time: %f ms\n", elapsed_time1, elapsed_time2);
return 0;
}
#include "funset.hpp"#include <chrono>#include <memory>#include <vector>static void copy_const_kernel(float* iptr, const float* cptr, int width, int height){for (int y = 0; y < height; ++y) {for (int x = 0; x < width; ++x) {int offset = x + y * width;if (cptr[offset] != 0) iptr[offset] = cptr[offset];}}}static void blend_kernel(float* outSrc, const float* inSrc, int width, int height, float speed){for (int y = 0; y < height; ++y) {for (int x = 0; x < width; ++x) {int offset = x + y * width;int left = offset - 1;int right = offset + 1;if (x == 0) ++left;if (x == width - 1) --right;int top = offset - height;int bottom = offset + height;if (y == 0) top += height;if (y == height - 1) bottom -= height;outSrc[offset] = inSrc[offset] + speed * (inSrc[top] + inSrc[bottom] + inSrc[left] + inSrc[right] - inSrc[offset] * 4);}}}static unsigned char value(float n1, float n2, int hue){if (hue > 360) hue -= 360;else if (hue < 0) hue += 360;if (hue < 60)return (unsigned char)(255 * (n1 + (n2 - n1)*hue / 60));if (hue < 180)return (unsigned char)(255 * n2);if (hue < 240)return (unsigned char)(255 * (n1 + (n2 - n1)*(240 - hue) / 60));return (unsigned char)(255 * n1);}static void float_to_color(unsigned char *optr, const float *outSrc, int width, int height){for (int y = 0; y < height; ++y) {for (int x = 0; x < width; ++x) {int offset = x + y * width;float l = outSrc[offset];float s = 1;int h = (180 + (int)(360.0f * outSrc[offset])) % 360;float m1, m2;if (l <= 0.5f) m2 = l * (1 + s);else m2 = l + s - l * s;m1 = 2 * l - m2;optr[offset * 4 + 0] = value(m1, m2, h + 120);optr[offset * 4 + 1] = value(m1, m2, h);optr[offset * 4 + 2] = value(m1, m2, h - 120);optr[offset * 4 + 3] = 255;}}}int heat_conduction_cpu(unsigned char* ptr, int width, int height, const float* src, float speed, float* elapsed_time){auto start = std::chrono::steady_clock::now();std::vector<float> inSrc(width*height, 0.f);std::vector<float> outSrc(width*height, 0.f);for (int i = 0; i < 90; ++i) {copy_const_kernel(inSrc.data(), src, width, height);blend_kernel(outSrc.data(), inSrc.data(), width, height, speed);std::swap(inSrc, outSrc);}float_to_color(ptr, inSrc.data(), width, height);auto end = std::chrono::steady_clock::now();auto duration = std::chrono::duration_cast<std::chrono::nanoseconds>(end - start);*elapsed_time = duration.count() * 1.0e-6;return 0;}
heat_conduction.cu:
#include "funset.hpp"#include <iostream>#include <algorithm>#include <memory>#include <vector>#include <cuda_runtime.h> // For the CUDA runtime routines (prefixed with "cuda_")#include <device_launch_parameters.h>#include "common.hpp"/* __global__: 函数类型限定符;在设备上运行;在主机端调用,计算能力3.2及以上可以在设备端调用;声明的函数的返回值必须是void类型;对此类型函数的调用是异步的,即在设备完全完成它的运行之前就返回了;对此类型函数的调用必须指定执行配置,即用于在设备上执行函数时的grid和block的维度,以及相关的流(即插入<<< >>>运算符);a kernel,表示此函数为内核函数(运行在GPU上的CUDA并行计算函数称为kernel(内核函数),内核函数必须通过__global__函数类型限定符定义); */__global__ static void copy_const_kernel(float* iptr, const float* cptr){/* gridDim: 内置变量,用于描述线程网格的维度,对于所有线程块来说,这个变量是一个常数,用来保存线程格每一维的大小,即每个线程格中线程块的数量.一个grid最多只有二维,为dim3类型；blockDim: 内置变量,用于说明每个block的维度与尺寸.为dim3类型,包含了block在三个维度上的尺寸信息;对于所有线程块来说,这个变量是一个常数,保存的是线程块中每一维的线程数量;blockIdx: 内置变量,变量中包含的值就是当前执行设备代码的线程块的索引;用于说明当前thread所在的block在整个grid中的位置,blockIdx.x取值范围是[0,gridDim.x-1],blockIdx.y取值范围是[0, gridDim.y-1].为uint3类型,包含了一个block在grid中各个维度上的索引信息;threadIdx: 内置变量,变量中包含的值就是当前执行设备代码的线程索引;用于说明当前thread在block中的位置;如果线程是一维的可获取threadIdx.x,如果是二维的还可获取threadIdx.y,如果是三维的还可获取threadIdx.z;为uint3类型,包含了一个thread在block中各个维度的索引信息 */// map from threadIdx/BlockIdx to pixel positionint x = threadIdx.x + blockIdx.x * blockDim.x;int y = threadIdx.y + blockIdx.y * blockDim.y;int offset = x + y * blockDim.x * gridDim.x;if (cptr[offset] != 0) iptr[offset] = cptr[offset];}__global__ static void blend_kernel(float* outSrc, const float* inSrc, int width, int height, float speed){// map from threadIdx/BlockIdx to pixel positionint x = threadIdx.x + blockIdx.x * blockDim.x;int y = threadIdx.y + blockIdx.y * blockDim.y;int offset = x + y * blockDim.x * gridDim.x;int left = offset - 1;int right = offset + 1;if (x == 0) ++left;if (x == width - 1) --right;int top = offset - height;int bottom = offset + height;if (y == 0) top += height;if (y == height - 1) bottom -= height;outSrc[offset] = inSrc[offset] + speed * (inSrc[top] + inSrc[bottom] + inSrc[left] + inSrc[right] - inSrc[offset] * 4);}/* __device__: 函数类型限定符,表明被修饰的函数在设备上执行，只能从设备上调用，但只能在其它__device__函数或者__global__函数中调用；__device__函数不支持递归；__device__函数的函数体内不能声明静态变量；__device__函数的参数数目是不可变化的;不能对__device__函数取指针 */__device__ static unsigned char value(float n1, float n2, int hue){if (hue > 360) hue -= 360;else if (hue < 0) hue += 360;if (hue < 60)return (unsigned char)(255 * (n1 + (n2 - n1)*hue / 60));if (hue < 180)return (unsigned char)(255 * n2);if (hue < 240)return (unsigned char)(255 * (n1 + (n2 - n1)*(240 - hue) / 60));return (unsigned char)(255 * n1);}__global__ static void float_to_color(unsigned char *optr, const float *outSrc){// map from threadIdx/BlockIdx to pixel positionint x = threadIdx.x + blockIdx.x * blockDim.x;int y = threadIdx.y + blockIdx.y * blockDim.y;int offset = x + y * blockDim.x * gridDim.x;float l = outSrc[offset];float s = 1;int h = (180 + (int)(360.0f * outSrc[offset])) % 360;float m1, m2;if (l <= 0.5f) m2 = l * (1 + s);else m2 = l + s - l * s;m1 = 2 * l - m2;optr[offset * 4 + 0] = value(m1, m2, h + 120);optr[offset * 4 + 1] = value(m1, m2, h);optr[offset * 4 + 2] = value(m1, m2, h - 120);optr[offset * 4 + 3] = 255;}static int heat_conduction_gpu_1(unsigned char* ptr, int width, int height, const float* src, float speed, float* elapsed_time){/* cudaEvent_t: CUDA event types,结构体类型, CUDA事件,用于测量GPU在某个任务上花费的时间,CUDA中的事件本质上是一个GPU时间戳,由于CUDA事件是在GPU上实现的,因此它们不适于对同时包含设备代码和主机代码的混合代码计时 */cudaEvent_t start, stop;// cudaEventCreate: 创建一个事件对象,异步启动cudaEventCreate(&start);cudaEventCreate(&stop);// cudaEventRecord: 记录一个事件,异步启动,start记录起始时间cudaEventRecord(start, 0);float* dev_inSrc{ nullptr };float* dev_outSrc{ nullptr };float* dev_constSrc{ nullptr };unsigned char* dev_image{ nullptr };const size_t length1{ width * height * sizeof(float) };const size_t length2{ width * height * 4 * sizeof(unsigned char) };// cudaMalloc: 在设备端分配内存cudaMalloc(&dev_inSrc, length1);cudaMalloc(&dev_outSrc, length1);cudaMalloc(&dev_constSrc, length1);cudaMalloc(&dev_image, length2);/* cudaMemcpy: 在主机端和设备端拷贝数据,此函数第四个参数仅能是下面之一:(1). cudaMemcpyHostToHost: 拷贝数据从主机端到主机端(2). cudaMemcpyHostToDevice: 拷贝数据从主机端到设备端(3). cudaMemcpyDeviceToHost: 拷贝数据从设备端到主机端(4). cudaMemcpyDeviceToDevice: 拷贝数据从设备端到设备端(5). cudaMemcpyDefault: 从指针值自动推断拷贝数据方向,需要支持统一虚拟寻址(CUDA6.0及以上版本)cudaMemcpy函数对于主机是同步的 */cudaMemcpy(dev_constSrc, src, length1, cudaMemcpyHostToDevice);const int threads_block{ 16 };/* dim3: 基于uint3定义的内置矢量类型，相当于由3个unsigned int类型组成的结构体，可表示一个三维数组，在定义dim3类型变量时，凡是没有赋值的元素都会被赋予默认值1 */dim3 blocks(width / threads_block, height / threads_block);dim3 threads(threads_block, threads_block);for (int i = 0; i < 90; ++i) {copy_const_kernel << <blocks, threads >> >(dev_inSrc, dev_constSrc);blend_kernel << <blocks, threads >> >(dev_outSrc, dev_inSrc, width, height, speed);std::swap(dev_inSrc, dev_outSrc);}/* <<< >>>: 为CUDA引入的运算符,指定线程网格和线程块维度等,传递执行参数给CUDA编译器和运行时系统,用于说明内核函数中的线程数量,以及线程是如何组织的;尖括号中这些参数并不是传递给设备代码的参数,而是告诉运行时如何启动设备代码,传递给设备代码本身的参数是放在圆括号中传递的,就像标准的函数调用一样;不同计算能力的设备对线程的总数和组织方式有不同的约束;必须先为kernel中用到的数组或变量分配好足够的空间,再调用kernel函数,否则在GPU计算时会发生错误,例如越界等;使用运行时API时,需要在调用的内核函数名与参数列表直接以<<<Dg,Db,Ns,S>>>的形式设置执行配置,其中：Dg是一个dim3型变量,用于设置grid的维度和各个维度上的尺寸.设置好Dg后,grid中将有Dg.x*Dg.y个block,Dg.z必须为1;Db是一个dim3型变量,用于设置block的维度和各个维度上的尺寸.设置好Db后,每个block中将有Db.x*Db.y*Db.z个thread;Ns是一个size_t型变量,指定各块为此调用动态分配的共享存储器大小,这些动态分配的存储器可供声明为外部数组(extern __shared__)的其他任何变量使用;Ns是一个可选参数,默认值为0;S为cudaStream_t类型,用于设置与内核函数关联的流.S是一个可选参数,默认值0. */float_to_color << <blocks, threads >> >(dev_image, dev_inSrc);cudaMemcpy(ptr, dev_image, length2, cudaMemcpyDeviceToHost);// cudaFree: 释放设备上由cudaMalloc函数分配的内存cudaFree(dev_inSrc);cudaFree(dev_outSrc);cudaFree(dev_constSrc);cudaFree(dev_image);// cudaEventRecord: 记录一个事件,异步启动,stop记录结束时间cudaEventRecord(stop, 0);// cudaEventSynchronize: 事件同步,等待一个事件完成,异步启动cudaEventSynchronize(stop);// cudaEventElapseTime: 计算两个事件之间经历的时间,单位为毫秒,异步启动cudaEventElapsedTime(elapsed_time, start, stop);// cudaEventDestroy: 销毁事件对象,异步启动cudaEventDestroy(start);cudaEventDestroy(stop);return 0;}static int heat_conduction_gpu_2(unsigned char* ptr, int width, int height, const float* src, float speed, float* elapsed_time){return 0;}static int heat_conduction_gpu_3(unsigned char* ptr, int width, int height, const float* src, float speed, float* elapsed_time){return 0;}int heat_conduction_gpu(unsigned char* ptr, int width, int height, const float* src, float speed, float* elapsed_time){int ret{ 0 };ret = heat_conduction_gpu_1(ptr, width, height, src, speed, elapsed_time); // 没有采用纹理内存//ret = heat_conduction_gpu_2(ptr, width, height, src, speed, elapsed_time); // 采用一维纹理内存//ret = heat_conduction_gpu_3(ptr, width, height, src, speed, elapsed_time); // 采用二维纹理内存return ret;}
生成的结果图像如下：
执行结果如下：可见使用C++和CUDA实现的结果是完全一致的：
