CUDA学习笔记2——CUDA程序基本框架

CUDA向量运算

CUDA程序的基本框架为：

头文件包含
常量定义/宏定义
C++ 自定义函数和CUDA核函数声明
int main(void)
{
分配主机与设备内存
初始化主机中的数据
将部分数据从主机拷贝至设备
调用核函数在设备中进行计算
将部分数据从设备拷贝至主机
释放主机与设备内存
}
c++ 自定义函数与CUDA核函数定义

例：

#include 
#include "cuda_runtime.h"
#include
#include  


//cuda内存  x[] + y[] = Z[]
//1.分配内存 2.内存拷贝 3.执行核函数 4.内存拷贝


__global__ void vecAdd(const double *x, const double *y,double *z, int count)
{
	const int index = blockDim.x * blockIdx.x + threadIdx.x;//使用索引让每个线程找到其要处理的数据
	//t00 t01 t02
	//t10 t11 t12			 [当前block之前有多少线程] +[当前线程中的排序]
	//t20 t21 t22    t21 (7) = blockDim(3)*blockIdx(2) + threadIdx(1)
	if (index < count)
	{
		z[index] = x[index] + y[index];
	}
}

void vecAdd_cpu(const double *x, const double *y, double *z, int count)
{
	for (int i = 0; i < count; i++)
	{
		z[i] = x[i] + y[i];
	}
}

int main()
{
	const int N = 1000;
	const int M = sizeof(double) * N;

	//cup内存分配
	double *h_x = (double*)malloc(M);
	double *h_y = (double*)malloc(M);
	double *h_z = (double*)malloc(M);
	double *result_cpu = (double*)malloc(M);
	//GPU内存分配
	double *d_x, *d_y, *d_z;
	cudaMalloc((void**)&d_x, M);
	cudaMalloc((void**)&d_y, M);
	cudaMalloc((void**)&d_z, M);

	for (int i=0;i<N;++i)
	{
		h_x[i] = 1;
		h_y[i] = 2; 
	}

	//cpu数据传输到GPU
	cudaMemcpy(d_x, h_x, M, cudaMemcpyHostToDevice);
	cudaMemcpy(d_y, h_y, M, cudaMemcpyHostToDevice);

	//调用核函数
	const int block_size = 128;
	const int gride_size = (N + block_size - 1) / block_size;
	vecAdd <<< gride_size, block_size>>> (d_x,d_y,d_z,N);

	//GPU数据传输到cpu
	cudaMemcpy(h_z, d_z, M, cudaMemcpyDeviceToHost);

	//cpu计算 
	vecAdd_cpu(h_x, h_y, result_cpu, N);
	bool error = false;
	for (int i = 0; i < N; i++)
	{
		if (fabs(result_cpu[i]-h_z[i])>(1.0e-10))
		{
			error = true;
		}
		printf("h_z[%d]: %f \n", i, h_z[i]);
	}
	printf("Result: %s\n", error ? "Errors" : "Pass");

	free(h_x);
	free(h_y);
	free(h_z);
	cudaFree(d_x);
	cudaFree(d_y);
	cudaFree(d_z);

}

cudaError_t cudaMalloc(void **address, size_t size); CUDA中设备内存动态分配

1. 第一个参数address 为待分配设备内存的指针，由于内存（地址）本身就是指针，因此待分配内存的指针即为指针的指针。
2. 第二个参数size 为待分配内存的字节数。
3. 返回值为错误代号，成功为cudaSuccess，失败为错误代号。

cudaError_t cudaFree(void* address) CUDA中释放内存
cudaError_t cudaMemcpy(void *dst, const void *src, size_t count, size_t count, enum cudaMemcpyKind kind); CUDA中主机与设备之间数据传递。

cuda矩阵运算 例：

#include 
#include "cuda_runtime.h"
#include
#include  
#include 

#define BLOCK_SIZE 16


//cuda矩阵运算  a[][] * b[][] = c[][]
//1.分配内存 2.内存拷贝 3.执行核函数 4.内存拷贝
__global__ void gpu_matrix_mult(int *a, int *b, int *c, const int size)
{
	int y = blockDim.y *blockIdx.y + threadIdx.y;
	int x = blockDim.x*blockIdx.x + threadIdx.x;
	int tmp = 0;

	if (x<size && y<size)
	{
		for (int step = 0; step < size; step++)
		{
			tmp += a[y*size + step] * b[step*size + x];
		}
		c[y*size + x] = tmp;
	}

}

void cpu_matrix_mult(int *a, int *b, int *c, const int size)
{
	for (int y = 0; y < size; y++)
	{
		for (int x = 0; x < size; x++)
		{
			int tmp = 0;
			for (int step = 0; step < size; step++)
			{
				tmp += a[y*size + step] * b[step * size + x];
			}
			c[y * size + x] = tmp;
		}
	}
}


int main()
{
	int matrix_size = 1000;
	int memsize = sizeof(int) * matrix_size * matrix_size;
	//cup上分配内存
	int *h_a, *h_b, *h_c, *h_cc;
	cudaMallocHost((void**)&h_a, memsize);
	cudaMallocHost((void**)&h_b, memsize);
	cudaMallocHost((void**)&h_c, memsize);
	cudaMallocHost((void**)&h_cc, memsize);

	for (int y = 0; y < matrix_size; y++)
	{
		for (int x = 0; x < matrix_size; x++)
		{
			h_a[y*matrix_size + x] = rand() % 1024;
			h_b[y*matrix_size + x] = rand() % 1024;
		}
	}
	//GPU上分配内存
	int *d_a, *d_b, *d_c;
	cudaMalloc((void**)&d_a, memsize);
	cudaMalloc((void**)&d_b, memsize);
	cudaMalloc((void**)&d_c, memsize);
	//将cpu数据拷贝到GPU
	cudaMemcpy(d_a, h_a, memsize, cudaMemcpyHostToDevice);
	cudaMemcpy(d_b, h_b, memsize, cudaMemcpyHostToDevice);

	unsigned int grid_rows = (matrix_size + BLOCK_SIZE - 1) / BLOCK_SIZE;
	unsigned int grid_cols = (matrix_size + BLOCK_SIZE - 1) / BLOCK_SIZE;

	dim3 dimGrid(grid_cols, grid_rows);
	//gpu warp 32个线程共享一个物理端 因此尽量为32的整数倍
	dim3 dimBlock(BLOCK_SIZE, BLOCK_SIZE);//  x*y*z<=1024   z未定义 默认为1

	gpu_matrix_mult << <dimGrid, dimBlock >> > (d_a, d_b, d_c, matrix_size);

	cudaMemcpy(h_c, d_c, memsize, cudaMemcpyDeviceToHost);

	cpu_matrix_mult(h_a, h_b, h_cc, matrix_size);

	bool errors = false;
	for (int y = 0; y < matrix_size; y++)
	{
		printf("%d \n", y);
		for (int x = 0; x < matrix_size; x++)
		{
			if (fabs(h_cc[y*matrix_size + x] - h_c[y*matrix_size + x]) >(1.0e-10))
			{
				errors = true;
			}
		}
	}
	printf("Result: %s\n", errors ? "Errors" : "Passed");

	cudaFreeHost(h_a);
	cudaFreeHost(h_b);
	cudaFreeHost(h_c);
	cudaFreeHost(h_cc);
	cudaFree(d_a);
	cudaFree(d_b);
	cudaFree(d_c);


}