- html - 出于某种原因,IE8 对我的 Sass 文件中继承的 html5 CSS 不友好?
- JMeter 在响应断言中使用 span 标签的问题
- html - 在 :hover and :active? 上具有不同效果的 CSS 动画
- html - 相对于居中的 html 内容固定的 CSS 重复背景?
我刚刚开始学习 CUDA,并且一直在查看 NVIDIA 网站上的示例。具体来说,我已经实现了矩阵乘法的非共享版本(第一个示例是非共享版本,即使它在共享内存部分):
http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory
我在更改 block 大小时遇到输出问题。 NVIDIA 代码的默认 block 大小为 16,当我将两个矩阵相乘时,这为我提供了正确的输出。但是,如果我将 block 大小更改为大于 16 的任何值(同时仍然是 16 的倍数),我将得到矩阵中所有元素的零输出。我也在我的笔记本电脑上对此进行了测试,并注意到超过 32 而不是 16 的任何结果都是相同的。有人可以解释发生了什么吗?我在 SLI 中有两个 9800GTX+ 视频卡,所以我的最大块大小应该为 (512,512,1)。为什么我只能做16个?
此外,我注意到矩阵乘法的共享版本(也在 NVIDIA 页面上)中存在相同的行为。
我没有发布我的代码,因为如果我直接从 NVIDIA 网站复制代码,我会遇到同样的问题。
我非常感谢任何对此或资源的帮助,以了解有关此类 CUDA 详细信息的更多信息。
谢谢!
我已按要求附上代码:
#include "stdio.h"
#include <cuda.h>
#include <assert.h>
#include <time.h>
#include <math.h>
// This is an example CUDA program that compares the timings of a matrix multiplication.
// The comparisons are between the CPU, GPU, and the GPU with shared memory.
#define BLOCK_SIZE 32
typedef struct {
int width;
int height;
int stride;
float* elements;
} Matrix;
typedef void (*FuncPtr)(Matrix& A, Matrix& B, Matrix& C);
void multiplyMatrix(Matrix& A, Matrix& B, Matrix& C);
// Helper declarations
void initializeMatrix(Matrix& A, int rows, int cols, float val);
void copyMatrix(Matrix& dest, Matrix& src);
void freeMatrix(Matrix& A);
void printError(cudaError_t err);
void printMat(Matrix& A);
void setVal(Matrix& A, float val);
double applyMultFunc(FuncPtr func, Matrix& A, Matrix& B, Matrix& C, int numOfIters);
// CUDA declarations
__global__ void cudaMultMat(Matrix A, Matrix B, Matrix C);
int main() {
printf("Beginning Matrix Multiplication Comparison\n");
// Initialize matrix
Matrix A, B, C;
int rowsA = 32;
int colsA = 32;
int colsB = 32;
initializeMatrix(A, rowsA, colsA, 5.0f);
initializeMatrix(B, colsA, colsB, 2.0f);
initializeMatrix(C, rowsA, colsB, 0.0f);
// C = A * B using CPU, GPU, and GPU with shared memory
FuncPtr gpuMatMult = &multiplyMatrix;
int numOfIterations = 100;
double multTime = applyMultFunc(gpuMatMult, A, B, C, numOfIterations);
printMat(C);
// Update user
printf("Normal Mat Mult Time: %f\n", multTime);
// Cleanup
freeMatrix(A);
freeMatrix(B);
freeMatrix(C);
printf("\nPress Enter to continue...\n");
getchar();
return 0;
}
void multiplyMatrix(Matrix& A, Matrix& B, Matrix& C) {
// Initialize device matrices
Matrix deviceA, deviceB, deviceC;
copyMatrix(deviceA, A);
copyMatrix(deviceB, B);
copyMatrix(deviceC, C);
// Initialize number of blocks and threads
dim3 numOfThreadsPerBlock(BLOCK_SIZE, BLOCK_SIZE);
int xSize = (C.width + numOfThreadsPerBlock.x - 1) / numOfThreadsPerBlock.x;
int ySize = (C.height + numOfThreadsPerBlock.y - 1) / numOfThreadsPerBlock.y;
dim3 numOfBlocks(xSize, ySize);
// Call CUDA kernel
cudaMultMat<<<numOfBlocks, numOfThreadsPerBlock>>>(deviceA, deviceB, deviceC);
printError(cudaThreadSynchronize());
printError(cudaMemcpy(C.elements, deviceC.elements, C.height * C.width * sizeof(float), cudaMemcpyDeviceToHost));
// Free cuda memory
printError(cudaFree(deviceA.elements));
printError(cudaFree(deviceB.elements));
printError(cudaFree(deviceC.elements));
}
// CUDA definitions
// GPU matrix multiplication (non-shared memory)
__global__ void cudaMultMat(Matrix A, Matrix B, Matrix C) {
// If the matrices are of the wrong size then return
if(A.width != B.height) {
return;
}
// Initialize the indexes into the grid
int col = (blockDim.x * blockIdx.x) + threadIdx.x;
int row = (blockDim.y * blockIdx.y) + threadIdx.y;
// Initialize the result
float cVal = 0.0f;
// Find the result for the dot product of a row of A and a column of B
for(int i = 0; i < A.width; i++) {
cVal += A.elements[row * A.width + i] * B.elements[i * B.width + col];
}
// If we are in bounds then save the result
if(row < C.height && col < C.width) {
C.elements[row * C.width + col] = cVal;
}
}
// Helper functions
void initializeMatrix(Matrix& A, int rows, int cols, float val) {
A.width = cols;
A.height = rows;
A.stride = A.width;
int numOfElements = A.width * A.height;
A.elements = (float*) malloc(numOfElements * sizeof(float));
for(int i = 0; i < numOfElements; i++) {
A.elements[i] = val;
}
}
void copyMatrix(Matrix& dest, Matrix& src) {
dest.width = src.width;
dest.height = src.height;
dest.stride = src.stride;
int size = src.width * src.height * sizeof(float);
printError(cudaMalloc(&dest.elements, size));
printError(cudaMemcpy(dest.elements, src.elements, size, cudaMemcpyHostToDevice));
}
void freeMatrix(Matrix& A) {
free(A.elements);
}
void printError(cudaError_t err) {
if(err != 0) {
printf("CUDA ERROR: %s\n", cudaGetErrorString(err));
getchar();
}
}
void printMat(Matrix& A) {
printf("*********************************\n");
for(int i = 0; i < A.height; i++) {
for(int j = 0; j < A.width; j++) {
int index = i * A.width + j;
printf("%2.1f, ", A.elements[index]);
}
printf("\n");
}
}
void setVal(Matrix& A, float val) {
for(int i = 0; i < A.width * A.height; i++) {
A.elements[i] = val;
}
}
double applyMultFunc(FuncPtr func, Matrix& A, Matrix& B, Matrix& C, int numOfIters) {
clock_t startTime = clock();
for(int i = 0; i < numOfIters; i++) {
func(A, B, C);
}
clock_t endTime = clock();
return (double) (endTime - startTime) / CLOCKS_PER_SEC;
}
最佳答案
当您增加 block 大小时,您超出了 GPU 的每 block 线程数规范。
无论您如何创建 block ,9800GTX 的每个 block 都有 512 个线程的限制。 16*16 = 256 没问题。 32 x 32 = 1024 这是不对的。在这种情况下,内核无法运行,因此输出不正确。
你的笔记本电脑可能有一个更新的 GPU,它支持每 block 1024 个线程,所以 32 x 32 是可以的,但更大的就不行了。
如果添加proper cuda error checking您可以通过代码确认这一点。请注意,此代码似乎具有 cuda 错误检查,但在内核调用上实现的检查不完整。研究我提供的链接,您会看到不同之处。如果您修改带有完整错误检查的代码,您将看到该错误。
关于c++ - Cuda 矩阵示例 block 大小,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/20086047/
假设我有两个矩阵,每个矩阵有两列和不同的行数。我想检查并查看一个矩阵的哪些对在另一个矩阵中。如果这些是一维的,我通常只会做 a %in% x得到我的结果。 match似乎只适用于向量。 > a
关闭。这个问题是opinion-based .它目前不接受答案。 想要改进这个问题? 更新问题,以便 editing this post 可以用事实和引用来回答它. 关闭 9 个月前。 Improv
我只处理过 DirectX 矩阵 我读过一些文章,说不能将 DirectX 矩阵数学库用于 openGL 矩阵。 但我也读过,如果你的数学是一致的,你可以获得类似的结果。那只会让我更加困惑。 任何人都
我编写了一个C++代码来解决线性系统A.x = b,其中A是一个对称矩阵,方法是首先使用LAPACK(E)对角矩阵A = V.D.V^T(因为以后需要特征值),然后求解x = A^-1.b = V^T
我遇到了问题。我想创建二维数组 rows=3 cols=2我的代码如下 int **ptr; int row=3; int col=2; ptr=new int *[col]; for (int i=
我有一个 3d mxnxt 矩阵,我希望能够提取 t 2d nxm 矩阵。在我的例子中,我有一个 1024x1024x10 矩阵,我想要 10 张图像显示给我。 这不是 reshape ,我每次只需要
我在 MATLAB 中有一个 3d 矩阵 (n-by-m-by-t) 表示一段时间内网格中的 n-by-m 测量值.我想要一个二维矩阵,其中空间信息消失了,只剩下 n*m 随着时间 t 的测量值(即:
作为一个简化的示例,我有一个 3D numpy 矩阵,如下所示: a = np.array([[[1,2], [4,np.nan], [7,
作为一个简化的示例,我有一个 3D numpy 矩阵,如下所示: a = np.array([[[1,2], [4,np.nan], [7,
使用 eigen2 , 并给定一个矩阵 A a_0_0, a_0_1, a_0_2, ... a_1_0, a_1_0, a_1_2, ... ... 和一个矩阵B: b_0_0, b_0_1, b_
我想知道如何获得下面的布局。 在中型和大型设备上,我希望有 2 行和 2 列的布局(2 x 2 矩阵)。 在小型(和超小型)设备上或调整为小型设备时,我想要一个 4 行和 1 列的矩阵。 我将通过 a
有什么方法可以向量化以下内容: for i = 1:6 te = k(:,:,:,i).*(c(i)); end 我正在尝试将 4D 矩阵 k 乘以向量 c,方法是将其
如何从填充有 1 和 0 的矩阵中抽取 n 个随机点的样本? a=rep(0:1,5) b=rep(0,10) c=rep(1,10) dataset=matrix(cbind(a,b,c),nrow
我正在尝试创建一个包含 X 个 X 的矩阵。以下代码生成从左上角到右下角的 X 对 Angular 线,而不是从右上角到左下角的 X 对 Angular 线。我不确定从哪里开始。是否应该使用新变量创建
我想在 python 中创建一个每行三列的矩阵,并能够通过任何一行对它们进行索引。矩阵中的每个值都是唯一的。 据我所知,我可以设置如下矩阵: matrix = [["username", "name"
我有点迷茫 我创建了一个名为 person 的类,它具有 age 和 name 属性(以及 get set 方法)。然后在另一个类中,我想创建一个 persons 数组,其中每个人都有不同的年龄和姓名
我有 n 个类,它们要么堆叠,要么不堆叠。所有这些类都扩展了同一个类 (CellObject)。我知道更多类将添加到此列表中,我想创建一种易于在一个地方操纵“可堆叠性”的方法。 我正在考虑创建一个矩阵
我有一个包含 x 个字符串名称及其关联 ID 的文件。本质上是两列数据。 我想要的是一个格式为 x x x 的相关样式表(将相关数据同时作为 x 轴和 y 轴),但我想要 fuzzywuzzy 库的函
机器学习与传统编程的一个重要区别在于机器学习比传统编程涉及了更多的数学知识。不过,随着机器学习的飞速发展,各种框架应运而生,在数据分析等应用中使用机器学习时,使用现成的库和框架成为常态,似乎越来越不需
当我在 julia 中输入这个错误跳转但我不知道为什么,它应该工作。/ julia> A = [1 2 3 4; 5 6 7 8; 1 2 3 4; 5 6 7 8] 4×4 Array{Int64,
我是一名优秀的程序员,十分优秀!