提交 8fc116f9 authored 作者: Mathieu Germain's avatar Mathieu Germain 提交者: Marc-Alexandre Cote

Fix durp error

上级 3727eaeb
...@@ -142,7 +142,7 @@ class GpuCumsum(CumsumOp, GpuOp): ...@@ -142,7 +142,7 @@ class GpuCumsum(CumsumOp, GpuOp):
dim3 dimBlock(blockSize, 1, 1); dim3 dimBlock(blockSize, 1, 1);
dim3 dimGrid(dimGridX, 1, 1); dim3 dimGrid(dimGridX, 1, 1);
blockCumSum_1D_%(nodename)s<<<dimGrid, dimBlock>>> blockCumSum_1D_%(nodename)s<<<dimGrid, dimBlock, (2*blockSize) * sizeof(float)>>>
( (
CudaNdarray_DEV_DATA(%(x)s), CudaNdarray_DEV_DATA(%(x)s),
CudaNdarray_DEV_DATA(%(z)s), CudaNdarray_DEV_DATA(%(z)s),
...@@ -154,7 +154,7 @@ class GpuCumsum(CumsumOp, GpuOp): ...@@ -154,7 +154,7 @@ class GpuCumsum(CumsumOp, GpuOp):
cudaThreadSynchronize(); cudaThreadSynchronize();
dim3 dimGridBlockSum(1, 1, 1); dim3 dimGridBlockSum(1, 1, 1);
dim3 dimBlockBlockSum(dimGridX-1, 1, 1); dim3 dimBlockBlockSum(dimGridX-1, 1, 1);
blockCumSum_1D_%(nodename)s<<<dimGridBlockSum, dimBlockBlockSum, (2*blockSize) * sizeof(float)>>> blockCumSum_1D_%(nodename)s<<<dimGridBlockSum, dimBlockBlockSum, (2*(dimGridX-1)) * sizeof(float)>>>
( (
CudaNdarray_DEV_DATA(deviceBlockSum), CudaNdarray_DEV_DATA(deviceBlockSum),
CudaNdarray_DEV_DATA(deviceBlockSum), CudaNdarray_DEV_DATA(deviceBlockSum),
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论