“CUDA”版本间的差异
来自iCenter Wiki
| （相同用户的5个中间修订版本未显示） | |||
| 第12行： | 第12行： | ||
#include <cuda.h> | #include <cuda.h> | ||
__global__ void my_kernel() { | __global__ void my_kernel() { | ||
| − | + | printf(“Hello World from GPU!\n”); | |
} | } | ||
int main() | int main() | ||
{ | { | ||
| − | + | printf("Hello World from CPU!\n"); | |
| − | + | my_kernel<<<1,1>>>(); | |
| − | + | cudaDeviceSynchronize() | |
| − | + | return 0; | |
} | } | ||
| 第28行： | 第28行： | ||
Hello World from CPU! | Hello World from CPU! | ||
Hello World from GPU! | Hello World from GPU! | ||
| + | |||
| + | #include <stdio.h> | ||
| + | #include <cuda.h> | ||
| + | __global__ void my_kernel() { | ||
| + | printf(“GPU blk %i thread %i %i\n”, blockIdx.x, threadIdx.x, threadIdx.y); | ||
| + | } | ||
| + | int main() | ||
| + | { | ||
| + | const dim3 threadsPerBlock(2,2); | ||
| + | const dim3 blocksPerGrid(2); | ||
| + | my_kernel<<<threadsPerBlock, blocksPerGrid>>>(); | ||
| + | return 0; | ||
| + | } | ||
| + | |||
| + | |||
| + | $ nvcc hello_world.cu | ||
| + | $ ./a.out | ||
| + | GPU block 0 thread 0 0 | ||
| + | GPU block 0 thread 0 1 | ||
| + | GPU block 0 thread 1 0 | ||
| + | GPU block 0 thread 1 1 | ||
| + | GPU block 1 thread 0 0 | ||
| + | GPU block 1 thread 0 1 | ||
| + | GPU block 1 thread 1 0 | ||
| + | GPU block 1 thread 1 1 | ||
2017年4月20日 (四) 01:59的最后版本
CUDA（Compute Unified Device Architecture，统一计算架构）
异构平行计算系统（heterogeneous parallel computing systems）
latency devices (CPU cores)
throughput devices (GPU cores)
单进程多数据(single process and mutiple data, SPMD) 模型
#include <stdio.h>
#include <cuda.h>
__global__ void my_kernel() {
printf(“Hello World from GPU!\n”);
}
int main()
{
printf("Hello World from CPU!\n");
my_kernel<<<1,1>>>();
cudaDeviceSynchronize()
return 0;
}
编译输出
$ nvcc hello_world.cu $ ./a.out Hello World from CPU! Hello World from GPU!
#include <stdio.h>
#include <cuda.h>
__global__ void my_kernel() {
printf(“GPU blk %i thread %i %i\n”, blockIdx.x, threadIdx.x, threadIdx.y);
}
int main()
{
const dim3 threadsPerBlock(2,2);
const dim3 blocksPerGrid(2);
my_kernel<<<threadsPerBlock, blocksPerGrid>>>();
return 0;
}
$ nvcc hello_world.cu $ ./a.out GPU block 0 thread 0 0 GPU block 0 thread 0 1 GPU block 0 thread 1 0 GPU block 0 thread 1 1 GPU block 1 thread 0 0 GPU block 1 thread 0 1 GPU block 1 thread 1 0 GPU block 1 thread 1 1
