提交 009733e0 authored 作者: abergeron's avatar abergeron

Merge pull request #19 from nouiz/abergeron-cudnn_r2_alg

Fix cudnn crash with v1
......@@ -73,6 +73,20 @@ cudnnGetConvolutionForwardAlgorithm(
return CUDNN_STATUS_SUCCESS;
}
static inline cudnnStatus_t
cudnnGetConvolutionForwardWorkspaceSize(
cudnnHandle_t handle,
const cudnnTensorDescriptor_t srcDesc,
const cudnnFilterDescriptor_t filterDesc,
const cudnnConvolutionDescriptor_t convDesc,
const cudnnTensor4dDescriptor_t destDesc,
cudnnConvolutionFwdAlgo_t algo,
size_t *sizeInBytes) {
*sizeInBytes = 0;
return CUDNN_STATUS_SUCCESS;
}
static inline cudnnStatus_t
cudnnConvolutionForward_v2(
cudnnHandle_t handle,
......
......@@ -362,6 +362,8 @@ class GpuDnnConv(DnnBase, COp):
self.workmem = 'small'
def get_op_params(self):
if version() == -1:
return [('CONV_ALGO', "0")]
if self.workmem == 'none':
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM'
elif self.workmem == 'small':
......
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