From f8b8811dcb7a0775443a39073966d626287d1029 Mon Sep 17 00:00:00 2001
From: lvmengsi <ceci3@users.noreply.github.com>
Date: Sun, 20 Oct 2019 19:06:44 +0800
Subject: [PATCH] fix_depthwise_conv_cudnn, test=develop (#20712) (#20727)

---
 paddle/fluid/operators/conv_cudnn_op.cu | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/paddle/fluid/operators/conv_cudnn_op.cu b/paddle/fluid/operators/conv_cudnn_op.cu
index 274da9abf08..c6af6da582b 100644
--- a/paddle/fluid/operators/conv_cudnn_op.cu
+++ b/paddle/fluid/operators/conv_cudnn_op.cu
@@ -265,6 +265,16 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
     algo = search::Find<T>(args, exhaustive_search, false, 0, ctx);
     workspace_size = search::GetWorkspaceSize(args, algo);
 
+#if CUDNN_VERSION_MIN(7, 0, 1)
+    // when groups > 1, SearchAlgorithm find algo is CUDNN_CONVOLUTION_\
+    // FWD_ALGO_WINOGRAD_NONFUSED, but this kind of algorithm is unstable
+    // in forward computation, so change the algorithm to CUDNN_CONVOLUTION_\
+    // FWD_ALGO_IMPLICIT_GEMM manually.
+    if (ctx.Attr<int>("groups") > 1) {
+      algo = static_cast<cudnnConvolutionFwdAlgo_t>(0);
+    }
+#endif
+
     // ------------------- cudnn conv forward ---------------------
     ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
     for (int i = 0; i < groups; i++) {
@@ -805,6 +815,7 @@ class CUDNNConvDoubleGradOpKernel : public framework::OpKernel<T> {
 #if CUDNN_VERSION_MIN(7, 0, 1)
     iwo_group = 1;
     c_group = groups;
+    groups = 1;
 #endif
     auto dtype = platform::CudnnDataType<T>::type;
 
-- 
GitLab