Cuda 4 GPU 间同步

发布于 2024-10-29 23:04:38 字数 491 浏览 0 评论 0原文

有人在 Cuda4 上成功测试过这个功能吗？看来我无法让它正常工作，不确定这是我的代码中的错误还是该功能尚未完全实现。

例如，我想在将内存从 1 个 GPU 复制到另一个 GPU 后执行内核调用。流0，mem0在gpu0上创建流1，mem1在gpu1上创建。两者都是 Fermi Tesla (C2070)

cudaMemcpyAsync( mem1, mem0, size, cudaMemcpyDefault, stream0 );
cudaEventRecord(P2Pevent, stream0);

cudaStreamWaitEvent(stream1, P2Pevent, 0);
cudaKernel<<<block,thread,0,stream1>>>(mem1);

有时内存似乎还没有为内核准备好，因此结果不正确。如果我在 Gpu0 上添加 cudaDeviceSynchronize 那么它工作正常。

谢谢

原文

Did anyone successfully test this feature on Cuda4 ? It seems I can't have it worked correctly, not sure if it's a bug in my code or the feature is not fully implemented yet.

For example, I want to do a kernel call after copying a memory from 1 gpu to another. Stream 0, mem0 is created on gpu0 stream 1,mem1 is created on gpu1. Both are Fermi Tesla (C2070)

cudaMemcpyAsync( mem1, mem0, size, cudaMemcpyDefault, stream0 );
cudaEventRecord(P2Pevent, stream0);

cudaStreamWaitEvent(stream1, P2Pevent, 0);
cudaKernel<<<block,thread,0,stream1>>>(mem1);

Sometime it seems the memory is not ready for the kernel yet so the result is incorrect. If I add an cudaDeviceSynchronize on Gpu0 then it works fine.

Thanks

分享到QQ

分享到微博