[phi]move reduce gpu impl funcs into pten/kernels/funcs (#39990)

* move reduce gpu impl funcs into pten/kernels/funcs

* change reduce header name and namespace

* fix spell word error

* change mutable_data to dev_ctx.Alloc

* modify place to devcontex

* format code style

* fix build error

* fix build error

* fix conflict

paddle/fluid/operators/fused/attn_bias_add.cu.h

@@ -191,9 +191,9 @@ void SetConfigForColumnReduce(const int max_threads, const int reduce_num,
 
   int num_block = (max_threads / left_num);
   if (num_block > 1 && reduce_num >= REDUCE_SPLIT_BOUNDARY) {
-    *blocking_size = phi::kernels::details::GetLastPow2(reduce_num / num_block);
+    *blocking_size = phi::funcs::details::GetLastPow2(reduce_num / num_block);
     if (*blocking_size <= 1) {
-      *blocking_size = phi::kernels::details::GetLastPow2(sqrt(reduce_num));
+      *blocking_size = phi::funcs::details::GetLastPow2(sqrt(reduce_num));
     } else if (*blocking_size * 2 < reduce_num) {
       *blocking_size *= 2;
     }

paddle/fluid/operators/reduce_ops/check_reduce_rank_test.cu

@@ -39,9 +39,9 @@ TEST(test_reduce_rank_check, all) {
       }
 
       if (is_valid) {
-        phi::kernels::details::CheckReduceRank(reduce_rank, rank);
+        phi::funcs::details::CheckReduceRank(reduce_rank, rank);
       } else {
-        ASSERT_THROW(phi::kernels::details::CheckReduceRank(reduce_rank, rank),
+        ASSERT_THROW(phi::funcs::details::CheckReduceRank(reduce_rank, rank),
                      paddle::platform::EnforceNotMet);
       }
     }

paddle/fluid/operators/reduce_ops/reduce_op.cu.h

@@ -23,8 +23,7 @@
 #include "paddle/fluid/framework/tensor.h"
 
 #include "paddle/phi/core/dense_tensor.h"
-#include "paddle/phi/kernels/gpu/reduce.h"
-
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 namespace paddle {
 namespace operators {
 
@@ -37,7 +36,7 @@ void TensorReduceImpl(const platform::CUDADeviceContext& dev_ctx,
                       gpuStream_t stream) {
   y->mutable_data<Ty>(x.place());
 
-  phi::kernels::TensorReduceImpl<Tx, Ty, ReduceOp, TransformOp>(
+  phi::funcs::TensorReduceImpl<Tx, Ty, ReduceOp, TransformOp>(
       static_cast<const phi::GPUContext&>(dev_ctx), x, y, transform,
       origin_reduce_dims, stream);
 }

paddle/phi/kernels/gpu/broadcast_tensors_grad_kernel.cu

@@ -20,7 +20,7 @@
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/enforce.h"
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/reduce.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/primitive/functor_primitives.h"
 
 namespace phi {
@@ -87,7 +87,7 @@ void BroadcastTensorsGradKernel(const Context& ctx,
           *input_tensor, ctx.GetPlace(), ctx, output_tensor);
     } else {
       // reduce_sum implementation on CUDA
-      kernels::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
           ctx,
           *input_tensor,
           output_tensor,

paddle/phi/kernels/gpu/compare_kernel.cu

@@ -80,7 +80,7 @@ inline void CompareAllKernelImpl(const Context& ctx,
   for (int i = 0; i < reduce_dims.size(); ++i) {
     reduce_dims[i] = i;
   }
-  kernels::TensorReduceImpl<bool, bool, BitwiseAdd, kps::IdentityFunctor<bool>>(
+  funcs::TensorReduceImpl<bool, bool, BitwiseAdd, kps::IdentityFunctor<bool>>(
       ctx, tmp, out, kps::IdentityFunctor<bool>(), reduce_dims, ctx.stream());
 }
 

paddle/phi/kernels/gpu/elementwise_grad.h

@@ -16,7 +16,7 @@ limitations under the License. */
 
 #include "paddle/phi/kernels/copy_kernel.h"
 #include "paddle/phi/kernels/funcs/elementwise_grad_base.h"
-#include "paddle/phi/kernels/gpu/reduce.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 
 namespace phi {
 
@@ -84,7 +84,7 @@ void DefaultElementwiseAddGrad(const GPUContext &ctx,
       std::vector<int> reduce_dims =
           funcs::GetReduceDim(x.dims(), out.dims(), axis);
       gpuStream_t stream = ctx.stream();
-      kernels::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
           ctx, dout, dx, kps::IdentityFunctor<T>(), reduce_dims, stream);
     }
   }
@@ -99,7 +99,7 @@ void DefaultElementwiseAddGrad(const GPUContext &ctx,
       std::vector<int> reduce_dims =
           funcs::GetReduceDim(y.dims(), out.dims(), axis);
       gpuStream_t stream = ctx.stream();
-      kernels::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
           ctx, dout, dy, kps::IdentityFunctor<T>(), reduce_dims, stream);
     }
   }
@@ -197,7 +197,7 @@ void default_elementwise_sub_grad(const GPUContext &ctx,
       std::vector<int> reduce_dims =
           funcs::GetReduceDim(x.dims(), out.dims(), axis);
       gpuStream_t stream = ctx.stream();
-      kernels::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
           ctx, dout, dx, kps::IdentityFunctor<T>(), reduce_dims, stream);
     }
   }
@@ -218,7 +218,7 @@ void default_elementwise_sub_grad(const GPUContext &ctx,
       std::vector<int> reduce_dims =
           funcs::GetReduceDim(y.dims(), out.dims(), axis);
       gpuStream_t stream = ctx.stream();
-      kernels::TensorReduceImpl<T, T, kps::AddFunctor, kps::InverseFunctor<T>>(
+      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::InverseFunctor<T>>(
           ctx, dout, dy, kps::InverseFunctor<T>(), reduce_dims, stream);
     }
   }

paddle/phi/kernels/gpu/trace_kernel.cu

@@ -17,7 +17,7 @@
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/funcs/diagonal.h"
-#include "paddle/phi/kernels/gpu/reduce.h"
+#include "paddle/phi/kernels/funcs/reduce_function.h"
 
 namespace phi {
 
@@ -34,7 +34,7 @@ void TraceKernel(const Context& ctx,
     auto stream = ctx.stream();
     std::vector<int> reduce_dims;
     reduce_dims.push_back(out->dims().size());
-    kernels::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+    funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
         ctx, diag, out, kps::IdentityFunctor<T>(), reduce_dims, stream);
   } else {
     phi::funcs::SetConstant<Context, T> functor;

paddle/phi/kernels/impl/matmul_grad_kernel_impl.h

@@ -60,7 +60,7 @@ struct ReduceSumForMatmulGrad<GPUContext, T> {
                   DenseTensor* output,
                   const std::vector<int>& reduce_dims) {
     auto stream = dev_ctx.stream();
-    kernels::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+    funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
         dev_ctx, input, output, kps::IdentityFunctor<T>(), reduce_dims, stream);
   }
 };