fix conv_fusion in multi thread. (#55374)

paddle/phi/kernels/fusion/gpu/conv_fusion_kernel.cu

@@ -130,7 +130,6 @@ class CudnnConvDescManager {
     XXH64_hash_t hash_key = XXH64_digest(state);
     XXH64_freeState(state);
 
-    if (!cudnn_conv_cache_.count(hash_key)) {
     std::lock_guard<std::mutex> lock(cache_mutex_);
     if (!cudnn_conv_cache_.count(hash_key)) {
       cudnn_conv_cache_[hash_key] = CudnnCacheInfo();
@@ -158,7 +157,6 @@ class CudnnConvDescManager {
       cudnn_conv_cache_[hash_key].workspace_size = workspace_size;
       cudnn_conv_cache_[hash_key].algo = algo;
     }
-    }
 
     return &cudnn_conv_cache_.at(hash_key);
   }
@@ -199,7 +197,6 @@ class CudnnConvDescManager {
     XXH64_hash_t hash_key = XXH64_digest(state);
     XXH64_freeState(state);
 
-    if (!conv_attr_cache_.count(hash_key)) {
     std::lock_guard<std::mutex> lock(attr_mutex_);
     if (!conv_attr_cache_.count(hash_key)) {
       ConvAttrCacheInfo cache;
@@ -254,12 +251,10 @@ class CudnnConvDescManager {
           }
           if (format == CUDNN_TENSOR_NCHW) {
             input_pad[2 * i + 4] = paddings[2 * i] - padding_common[i];
-              input_pad[2 * i + 4 + 1] =
-                  paddings[2 * i + 1] - padding_common[i];
+            input_pad[2 * i + 4 + 1] = paddings[2 * i + 1] - padding_common[i];
           } else {
             input_pad[2 * i + 2] = paddings[2 * i] - padding_common[i];
-              input_pad[2 * i + 2 + 1] =
-                  paddings[2 * i + 1] - padding_common[i];
+            input_pad[2 * i + 2 + 1] = paddings[2 * i + 1] - padding_common[i];
           }
         }
 
@@ -283,7 +278,6 @@ class CudnnConvDescManager {
       cache.paddings = padding_common;
       conv_attr_cache_[hash_key] = cache;
     }
-    }
 
     return &conv_attr_cache_.at(hash_key);
   }