NASP/cuda-samples
35
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2024-11-24 18:49:18 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update globalToShmemAsyncCopy.cu

Browse Source 
Fix two obvious errors, the first one is that five tasks were submitted to pipeline at the same time and task 4 conflicts with task 0, the remaining two are copy errors
This commit is contained in:
aioprli 2024-05-06 22:03:16 +08:00 committed by GitHub
parent 5f97d7d0df
commit 6aac4717b8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
Samples/3_CUDA_Features/globalToShmemAsyncCopy/globalToShmemAsyncCopy.cu
View File

@ -128,7 +128,7 @@ __global__ void MatrixMulAsyncCopyMultiStageLargeChunk(
a <= aEnd; a += aStep, b += bStep, ++i) { a <= aEnd; a += aStep, b += bStep, ++i) {
// Load the matrices from device memory to shared memory; each thread loads // Load the matrices from device memory to shared memory; each thread loads
// one element of each matrix // one element of each matrix
for (; aStage <= a + aStep * maxPipelineStages; for (; aStage < a + aStep * maxPipelineStages;
aStage += aStep, bStage += bStep, ++iStage) { aStage += aStep, bStage += bStep, ++iStage) {
pipe.producer_acquire(); pipe.producer_acquire();
if (aStage <= aEnd && t4x < BLOCK_SIZE) { if (aStage <= aEnd && t4x < BLOCK_SIZE) {
@ -137,7 +137,7 @@ __global__ void MatrixMulAsyncCopyMultiStageLargeChunk(
cuda::memcpy_async(&As[j][threadIdx.y][t4x], cuda::memcpy_async(&As[j][threadIdx.y][t4x],
&A[aStage + wA * threadIdx.y + t4x], shape4, pipe); &A[aStage + wA * threadIdx.y + t4x], shape4, pipe);
cuda::memcpy_async(&Bs[j][threadIdx.y][t4x], cuda::memcpy_async(&Bs[j][threadIdx.y][t4x],
&B[aStage + wA * threadIdx.y + t4x], shape4, pipe); &B[bStage + wB * threadIdx.y + t4x], shape4, pipe);
} }
pipe.producer_commit(); pipe.producer_commit();
} }
@ -222,7 +222,7 @@ __global__ void MatrixMulAsyncCopyLargeChunk(float *__restrict__ C,
cuda::memcpy_async(&As[threadIdx.y][t4x], &A[a + wA * threadIdx.y + t4x], cuda::memcpy_async(&As[threadIdx.y][t4x], &A[a + wA * threadIdx.y + t4x],
shape4, pipe); shape4, pipe);
cuda::memcpy_async(&Bs[threadIdx.y][t4x], &B[a + wA * threadIdx.y + t4x], cuda::memcpy_async(&Bs[threadIdx.y][t4x], &B[b + wB * threadIdx.y + t4x],
shape4, pipe); shape4, pipe);
pipe.producer_commit(); pipe.producer_commit();