CUDA Kernel on Custom Device - FIX (#72867)

datutu-L · YqGe585 · web-flow · commit 700b5f425af6 · 2025-05-28T16:19:17.000+08:00
* seperate blas

fix GetDeviceProperties

small fix

fix dcu error in gpu_types.h

modified PD_REGISTER_KERNEL_FOR_ALL_BACKEND_DTYPE 250513

* fix stack_kernel compile error and modified 3 custom kernel

* Modified PD_REGISTER_KERNEL_FOR_ALL_DTYPE

* Modified PD_REGISTER_KERNEL_FOR_ALL_DTYPE

* 删除一些未使用的blas.h头文件

* 还原多余的删除

* 修复了CPU 编译不过

---------

Co-authored-by: geyuqiang &lt;geyuqiang@baidu.com&gt;
Co-authored-by: Yuqiang Ge &lt;143453447+YqGe585@users.noreply.github.com&gt;
diff --git a/paddle/phi/core/kernel_registry.h b/paddle/phi/core/kernel_registry.h
@@ -1247,10 +1247,20 @@ struct KernelRegistrar {
  * with one template argument.
  */
 
+#if (defined(PADDLE_WITH_CUSTOM_DEVICE) && defined(PADDLE_WITH_CUDA))
+#define PD_REGISTER_KERNEL_FOR_ALL_DTYPE(                                \
+    kernel_name, backend, layout, kernel_fn)                             \
+  template decltype(kernel_fn) kernel_fn;                                \
+  static void                                                            \
+      __FAKE_PD_KERNEL_args_def_FN_##kernel_name##_##backend##_##layout( \
+          const ::phi::KernelKey& kernel_key UNUSED,                     \
+          ::phi::Kernel* kernel UNUSED)
+#else
 #define PD_REGISTER_KERNEL_FOR_ALL_DTYPE(    \
     kernel_name, backend, layout, kernel_fn) \
   _PD_REGISTER_KERNEL_FOR_ALL_DTYPE(         \
       ::phi::RegType::INNER, kernel_name, backend, layout, kernel_fn)
+#endif
 
 #define _PD_REGISTER_KERNEL_FOR_ALL_DTYPE(                                   \
     reg_type, kernel_name, backend, layout, kernel_fn)                       \
diff --git a/paddle/phi/kernels/funcs/emb_eltwise_layer_norm_functor.cu b/paddle/phi/kernels/funcs/emb_eltwise_layer_norm_functor.cu
@@ -28,7 +28,6 @@ namespace cub = hipcub;
 #include "paddle/phi/kernels/funcs/emb_eltwise_layer_norm_functor.h"
 
 #include "paddle/phi/common/float16.h"
-#include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/math_cuda_utils.h"
 
 namespace phi {
diff --git a/paddle/phi/kernels/funcs/fake_dequantize_functor.h b/paddle/phi/kernels/funcs/fake_dequantize_functor.h
@@ -18,7 +18,6 @@ limitations under the License. */
 #include "paddle/phi/backends/all_context.h"
 #include "paddle/phi/common/scalar.h"
 #include "paddle/phi/core/dense_tensor.h"
-#include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/eigen/common.h"
 
 namespace phi {
diff --git a/paddle/phi/kernels/funcs/fake_quantize_functor.h b/paddle/phi/kernels/funcs/fake_quantize_functor.h
@@ -19,7 +19,6 @@ limitations under the License. */
 #include "paddle/phi/common/scalar.h"
 #include "paddle/phi/common/transform.h"
 #include "paddle/phi/core/dense_tensor.h"
-#include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/eigen/common.h"
 #include "paddle/phi/kernels/impl/clip_kernel_impl.h"
 
diff --git a/paddle/phi/kernels/funcs/math/bert_encoder_functor.cu b/paddle/phi/kernels/funcs/math/bert_encoder_functor.cu
@@ -18,7 +18,6 @@
 
 #include "paddle/phi/core/enforce.h"
 #include "paddle/phi/core/tensor_utils.h"
-#include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/math_cuda_utils.h"
 
 namespace phi {
