Merge pull request PaddlePaddle#40 from mthreads/format_build_musa

[MTAI-484] feat(build): fix a bug in softmax.cu for MUSA
hanhaowen-mt · Aug 14, 2023 · 7bb11aa · 7bb11aa
2 parents 269609e + d485fb5
commit 7bb11aa
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 11 deletions.
diff --git a/paddle/phi/kernels/CMakeLists.txt b/paddle/phi/kernels/CMakeLists.txt
@@ -88,6 +88,8 @@ if(WITH_MUSA)
        "gpu/rnn_grad_kernel.cu.cc"
        "gpu/rnn_kernel.cu.cc"
        "gpu/slogdeterminant_grad_kernel.cu"
+       "gpu/softmax_grad_kernel.cu"
+       "gpu/softmax_kernel.cu"
        "gpu/solve_grad_kernel.cu"
        "gpu/solve_kernel.cu"
        "gpu/spectral_norm_grad_kernel.cu"

diff --git a/paddle/phi/kernels/funcs/CMakeLists.txt b/paddle/phi/kernels/funcs/CMakeLists.txt
@@ -20,7 +20,8 @@ endif()
 if(WITH_MUSA)
   list(REMOVE_ITEM func_cu_srcs
       "cross_entropy.cu"
-      "gru_compute.cu")
+      "gru_compute.cu"
+      "softmax.cu")
 endif()
 
 collect_srcs(kernels_srcs SRCS ${func_cc_srcs} ${func_cu_srcs})
diff --git a/paddle/phi/kernels/funcs/softmax.cu b/paddle/phi/kernels/funcs/softmax.cu
@@ -21,8 +21,6 @@ limitations under the License. */
 
 namespace phi {
 namespace funcs {
-// TODO(@caizhi): enable it
-#if 0
 using ScopedTensorDescriptor = phi::backends::gpu::ScopedTensorDescriptor;
 using DataLayout = phi::backends::gpu::DataLayout;
 template <typename T>
@@ -61,8 +59,6 @@ void SoftmaxCUDNNFunctor<T, DeviceContext>::operator()(
                                             context.template Alloc<T>(Y),
                                             MIOPEN_SOFTMAX_ACCURATE,
                                             MIOPEN_SOFTMAX_MODE_INSTANCE));
-#elif defined(PADDLE_WITH_MUSA)
-  // TODO
 #else
   cudnnTensorDescriptor_t cudnn_x_desc =
       xDesc.descriptor<T>(layout, cudnn_tensor_dims);
@@ -148,18 +144,16 @@ template class SoftmaxCUDNNFunctor<float, phi::GPUContext>;
 template class SoftmaxCUDNNFunctor<phi::dtype::float16, phi::GPUContext>;
 template class SoftmaxGradCUDNNFunctor<float, phi::GPUContext>;
 template class SoftmaxGradCUDNNFunctor<phi::dtype::float16, phi::GPUContext>;
-// TODO(@caizhi): enable it
-//#if CUDNN_VERSION_MIN(8, 1, 0)
-//template class SoftmaxCUDNNFunctor<phi::dtype::bfloat16, phi::GPUContext>;
-//template class SoftmaxGradCUDNNFunctor<phi::dtype::bfloat16, phi::GPUContext>;
-//#endif
+#if CUDNN_VERSION_MIN(8, 1, 0)
+template class SoftmaxCUDNNFunctor<phi::dtype::bfloat16, phi::GPUContext>;
+template class SoftmaxGradCUDNNFunctor<phi::dtype::bfloat16, phi::GPUContext>;
+#endif
 
 // MIOPEN do not support double
 #ifndef PADDLE_WITH_HIP
 template class SoftmaxCUDNNFunctor<double, phi::GPUContext>;
 template class SoftmaxGradCUDNNFunctor<double, phi::GPUContext>;
 #endif
-#endif 
 template class SoftmaxFunctor<phi::GPUContext, phi::dtype::float16>;
 template class SoftmaxFunctor<phi::GPUContext, phi::dtype::bfloat16>;
 template class SoftmaxFunctor<phi::GPUContext, float>;