【slice】config default_kernel grid/block (#73109)

xiaoguoguo626807 · web-flow · commit 814308bba561 · 2025-06-06T16:17:39.000+08:00
* dpa3 fix

* modify tuple pop check

* add set_value.cu

* fix ci wrong case

* add no expand copy

* modify windows bug, input dim was change in win

* add one value copy

* modify bug

* temp vec optim

* add other optim path

* grid / block config set

* fix

* clear
diff --git a/paddle/phi/kernels/funcs/strided_copy_kernel.cu.h b/paddle/phi/kernels/funcs/strided_copy_kernel.cu.h
@@ -1,6 +1,6 @@
 /* Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
 Licensed under the Apache License, Version 2.0 (the "License");
-you may nint64_t use this file except in compliance with the License.
+you may not use this file except in compliance with the License.
 You may obtain a copy of the License at
     https://www.apache.org/licenses/LICENSE-2.0
 Unless required by applicable law or agreed to in writing, software
@@ -444,6 +444,7 @@ bool LaunchContiguous2StridedCaseZeroKernel(
   if (!VerifyStridedCopyThreadConfigurationParameters(block, grid)) {
     return false;
   }
+
   if (diff_dims) {
     switch (rank) {
 #define CASE_RANK(__Rk)                              \
@@ -547,8 +548,11 @@ void LaunchContiguous2StridedDefaultKernel(
     int rank,
     int64_t numel,
     bool diff_dims) {
-  int64_t block = 512;
-  int64_t grid = (numel + block - 1) / block;
+  constexpr int loop_count = 4;
+  auto config = phi::backends::gpu::GetGpuLaunchConfig1D(
+      dev_ctx, numel, VecSize * loop_count);
+  auto& grid = config.block_per_grid;
+  auto& block = config.thread_per_block;
 
   if (diff_dims) {
     if (VecSize == 4) {
@@ -728,44 +732,26 @@ void StrideCopyDiffDimKernel(
                                                                true)) {
   } else {
     switch (VecSize) {
-      case 1:
-        LaunchContiguous2StridedDefaultKernel<T, Context, 1>(dev_ctx,
-                                                             input_data,
-                                                             output_data,
-                                                             output_stride,
-                                                             output_dims,
-                                                             rank,
-                                                             numel,
-                                                             true);
-
-        break;
-      case 2:
-        LaunchContiguous2StridedDefaultKernel<T, Context, 2>(dev_ctx,
-                                                             input_data,
-                                                             output_data,
-                                                             output_stride,
-                                                             output_dims,
-                                                             rank,
-                                                             numel,
-                                                             true);
-        break;
-      case 4:
-        LaunchContiguous2StridedDefaultKernel<T, Context, 4>(dev_ctx,
-                                                             input_data,
-                                                             output_data,
-                                                             output_stride,
-                                                             output_dims,
-                                                             rank,
-                                                             numel,
-                                                             true);
-        break;
-
+#define CASE_VECSIZE(__Sz)                                                 \
+  case __Sz:                                                               \
+    LaunchContiguous2StridedDefaultKernel<T, Context, __Sz>(dev_ctx,       \
+                                                            input_data,    \
+                                                            output_data,   \
+                                                            output_stride, \
+                                                            output_dims,   \
+                                                            rank,          \
+                                                            numel,         \
+                                                            true);         \
+    break;
+      CASE_VECSIZE(1);
+      CASE_VECSIZE(2);
+      CASE_VECSIZE(4);
+#undef CASE_VECSIZE
       default:
         PADDLE_THROW(common::errors::InvalidArgument(
             "unsurport vecsize %d for StrideCopyDiffDimKernel", VecSize));
     }
   }
-  return;
 }
 
 }  // namespace phi
