上面的答案指出了块大小如何影响性能,并建议了基于占用率最大化的常见启发式选择。在不希望提供选择块大小的标准的情况下,值得一提的是CUDA 6.5(现在为Release Candidate版本)包括了几个新的运行时函数,以帮助进行占用率计算和启动配置,请参见
CUDA Pro提示:占用API简化了启动配置
有用的功能之一是cudaOccupancyMaxPotentialBlockSize
启发式地计算可达到最大占用率的块大小。然后,该功能提供的值可用作启动参数手动优化的起点。下面是一个小例子。
#include <stdio.h>
/************************/
/* TEST KERNEL FUNCTION */
/************************/
__global__ void MyKernel(int *a, int *b, int *c, int N)
{
int idx = threadIdx.x + blockIdx.x * blockDim.x;
if (idx < N) { c[idx] = a[idx] + b[idx]; }
}
/********/
/* MAIN */
/********/
void main()
{
const int N = 1000000;
int blockSize; // The launch configurator returned block size
int minGridSize; // The minimum grid size needed to achieve the maximum occupancy for a full device launch
int gridSize; // The actual grid size needed, based on input size
int* h_vec1 = (int*) malloc(N*sizeof(int));
int* h_vec2 = (int*) malloc(N*sizeof(int));
int* h_vec3 = (int*) malloc(N*sizeof(int));
int* h_vec4 = (int*) malloc(N*sizeof(int));
int* d_vec1; cudaMalloc((void**)&d_vec1, N*sizeof(int));
int* d_vec2; cudaMalloc((void**)&d_vec2, N*sizeof(int));
int* d_vec3; cudaMalloc((void**)&d_vec3, N*sizeof(int));
for (int i=0; i<N; i++) {
h_vec1[i] = 10;
h_vec2[i] = 20;
h_vec4[i] = h_vec1[i] + h_vec2[i];
}
cudaMemcpy(d_vec1, h_vec1, N*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_vec2, h_vec2, N*sizeof(int), cudaMemcpyHostToDevice);
float time;
cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start, 0);
cudaOccupancyMaxPotentialBlockSize(&minGridSize, &blockSize, MyKernel, 0, N);
// Round up according to array size
gridSize = (N + blockSize - 1) / blockSize;
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&time, start, stop);
printf("Occupancy calculator elapsed time: %3.3f ms \n", time);
cudaEventRecord(start, 0);
MyKernel<<<gridSize, blockSize>>>(d_vec1, d_vec2, d_vec3, N);
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&time, start, stop);
printf("Kernel elapsed time: %3.3f ms \n", time);
printf("Blocksize %i\n", blockSize);
cudaMemcpy(h_vec3, d_vec3, N*sizeof(int), cudaMemcpyDeviceToHost);
for (int i=0; i<N; i++) {
if (h_vec3[i] != h_vec4[i]) { printf("Error at i = %i! Host = %i; Device = %i\n", i, h_vec4[i], h_vec3[i]); return; };
}
printf("Test passed\n");
}
编辑
的cudaOccupancyMaxPotentialBlockSize
被中定义cuda_runtime.h
文件,并定义如下:
template<class T>
__inline__ __host__ CUDART_DEVICE cudaError_t cudaOccupancyMaxPotentialBlockSize(
int *minGridSize,
int *blockSize,
T func,
size_t dynamicSMemSize = 0,
int blockSizeLimit = 0)
{
return cudaOccupancyMaxPotentialBlockSizeVariableSMem(minGridSize, blockSize, func, __cudaOccupancyB2DHelper(dynamicSMemSize), blockSizeLimit);
}
参数的含义如下
minGridSize = Suggested min grid size to achieve a full machine launch.
blockSize = Suggested block size to achieve maximum occupancy.
func = Kernel function.
dynamicSMemSize = Size of dynamically allocated shared memory. Of course, it is known at runtime before any kernel launch. The size of the statically allocated shared memory is not needed as it is inferred by the properties of func.
blockSizeLimit = Maximum size for each block. In the case of 1D kernels, it can coincide with the number of input elements.
请注意,从CUDA 6.5开始,需要根据API建议的1D块大小来计算自己的2D / 3D块尺寸。
还要注意,CUDA驱动程序API包含功能上等效的用于占用率计算的API,因此可以按照cuOccupancyMaxPotentialBlockSize
上面示例中针对运行时API所示的相同方式在驱动程序API代码中使用。