Rename phi::func::TensorReduceImpl to phi::func::ReduceKernel. (#40183)

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

@@ -36,9 +36,9 @@ void TensorReduceImpl(const platform::CUDADeviceContext& dev_ctx,
                       gpuStream_t stream) {
   y->mutable_data<Ty>(x.place());
 
-  phi::funcs::TensorReduceImpl<Tx, Ty, ReduceOp, TransformOp>(
+  phi::funcs::ReduceKernel<Tx, Ty, ReduceOp, TransformOp>(
       static_cast<const phi::GPUContext&>(dev_ctx), x, y, transform,
-      origin_reduce_dims, stream);
+      origin_reduce_dims);
 }
 
 }  // namespace operators

paddle/phi/kernels/funcs/matrix_reduce.cu

@@ -45,13 +45,8 @@ class MatrixReduceSumFunctor<T, GPUContext> {
         out_reduce_dims.push_back(idx);
       }
     }
-    TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
-        dev_ctx,
-        in,
-        out,
-        kps::IdentityFunctor<T>(),
-        out_reduce_dims,
-        dev_ctx.stream());
+    ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+        dev_ctx, in, out, kps::IdentityFunctor<T>(), out_reduce_dims);
   }
 };
 

paddle/phi/kernels/funcs/reduce_function.h

@@ -1087,12 +1087,12 @@ template <typename Tx,
           typename Ty,
           template <typename> class ReduceOp,
           typename TransformOp>
-void TensorReduceImpl(const phi::GPUContext& dev_ctx,
+void ReduceKernel(const phi::GPUContext& dev_ctx,
                   const phi::DenseTensor& x,
                   phi::DenseTensor* y,
                   const TransformOp& transform,
-                      const std::vector<int>& origin_reduce_dims,
-                      KPStream stream) {
+                  const std::vector<int>& origin_reduce_dims) {
+  auto stream = dev_ctx.stream();
   dev_ctx.Alloc<Ty>(y);
 
   auto x_dim = phi::vectorize<int>(x.dims());

paddle/phi/kernels/gpu/broadcast_tensors_grad_kernel.cu

@@ -87,13 +87,12 @@ void BroadcastTensorsGradKernel(const Context& ctx,
           *input_tensor, ctx.GetPlace(), ctx, output_tensor);
     } else {
       // reduce_sum implementation on CUDA
-      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+      funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
           ctx,
           *input_tensor,
           output_tensor,
           kps::IdentityFunctor<T>(),
-          reduce_dims_vec,
-          ctx.stream());
+          reduce_dims_vec);
     }
   }
 }

paddle/phi/kernels/gpu/compare_kernel.cu

@@ -80,8 +80,8 @@ inline void CompareAllKernelImpl(const Context& ctx,
   for (int i = 0; i < reduce_dims.size(); ++i) {
     reduce_dims[i] = i;
   }
-  funcs::TensorReduceImpl<bool, bool, BitwiseAdd, kps::IdentityFunctor<bool>>(
-      ctx, tmp, out, kps::IdentityFunctor<bool>(), reduce_dims, ctx.stream());
+  funcs::ReduceKernel<bool, bool, BitwiseAdd, kps::IdentityFunctor<bool>>(
+      ctx, tmp, out, kps::IdentityFunctor<bool>(), reduce_dims);
 }
 
 }  // namespace phi

paddle/phi/kernels/gpu/elementwise_grad.h

@@ -29,13 +29,8 @@ void ReduceWrapper(const GPUContext &dev_ctx,
                    DenseTensor *dst) {
   std::vector<int> reduce_dims =
       funcs::GetReduceDim(dst->dims(), src->dims(), axis);
-  funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
-      dev_ctx,
-      *src,
-      dst,
-      kps::IdentityFunctor<T>(),
-      reduce_dims,
-      dev_ctx.stream());
+  funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+      dev_ctx, *src, dst, kps::IdentityFunctor<T>(), reduce_dims);
 }
 
 template <ElementwiseType ET, typename T, typename Functor>
@@ -172,9 +167,8 @@ void DefaultElementwiseAddGrad(const GPUContext &ctx,
       }
       std::vector<int> reduce_dims =
           funcs::GetReduceDim(x.dims(), out.dims(), axis);
-      gpuStream_t stream = ctx.stream();
-      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
-          ctx, dout, dx, kps::IdentityFunctor<T>(), reduce_dims, stream);
+      funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+          ctx, dout, dx, kps::IdentityFunctor<T>(), reduce_dims);
     }
   }
   // dy
@@ -187,9 +181,8 @@ void DefaultElementwiseAddGrad(const GPUContext &ctx,
     } else {
       std::vector<int> reduce_dims =
           funcs::GetReduceDim(y.dims(), out.dims(), axis);
-      gpuStream_t stream = ctx.stream();
-      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
-          ctx, dout, dy, kps::IdentityFunctor<T>(), reduce_dims, stream);
+      funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+          ctx, dout, dy, kps::IdentityFunctor<T>(), reduce_dims);
     }
   }
 }
@@ -285,9 +278,8 @@ 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();
-      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
-          ctx, dout, dx, kps::IdentityFunctor<T>(), reduce_dims, stream);
+      funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+          ctx, dout, dx, kps::IdentityFunctor<T>(), reduce_dims);
     }
   }
   // dy
@@ -306,9 +298,8 @@ void default_elementwise_sub_grad(const GPUContext &ctx,
     } else {
       std::vector<int> reduce_dims =
           funcs::GetReduceDim(y.dims(), out.dims(), axis);
-      gpuStream_t stream = ctx.stream();
-      funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::InverseFunctor<T>>(
-          ctx, dout, dy, kps::InverseFunctor<T>(), reduce_dims, stream);
+      funcs::ReduceKernel<T, T, kps::AddFunctor, kps::InverseFunctor<T>>(
+          ctx, dout, dy, kps::InverseFunctor<T>(), reduce_dims);
     }
   }
 }

paddle/phi/kernels/gpu/reduce.h

@@ -39,8 +39,6 @@ void Reduce(const KPDevice& dev_ctx,
     reduce_num *= (x.dims())[i];
   }
 
-  KPStream stream = dev_ctx.stream();
-
   if (out_dtype != phi::DataType::UNDEFINED && out_dtype != x.dtype()) {
     auto tmp_tensor = phi::Cast<T>(dev_ctx, x, out_dtype);
     PD_VISIT_BOOL_AND_FLOATING_AND_COMPLEX_AND_3_TYPES(
@@ -48,10 +46,10 @@ void Reduce(const KPDevice& dev_ctx,
         phi::DataType::INT64,
         phi::DataType::FLOAT16,
         out_dtype,
-        "TensorReduceImpl",
+        "ReduceKernel",
         ([&] {
           using MPType = typename kps::details::MPTypeTrait<data_t>::Type;
-          phi::funcs::TensorReduceImpl<data_t,
+          phi::funcs::ReduceKernel<data_t,
                                    data_t,
                                    ReduceOp,
                                    TransformOp<data_t, MPType>>(
@@ -59,18 +57,12 @@ void Reduce(const KPDevice& dev_ctx,
               tmp_tensor,
               out,
               TransformOp<data_t, MPType>(reduce_num),
-              reduce_dims,
-              stream);
+              reduce_dims);
         }));
   } else {
     using MPType = typename kps::details::MPTypeTrait<T>::Type;
-    phi::funcs::TensorReduceImpl<T, T, ReduceOp, TransformOp<T, MPType>>(
-        dev_ctx,
-        x,
-        out,
-        TransformOp<T, MPType>(reduce_num),
-        reduce_dims,
-        stream);
+    phi::funcs::ReduceKernel<T, T, ReduceOp, TransformOp<T, MPType>>(
+        dev_ctx, x, out, TransformOp<T, MPType>(reduce_num), reduce_dims);
   }
 }
 }  // namespace phi

paddle/phi/kernels/gpu/sigmoid_cross_entropy_with_logits_grad_kernel.cu

@@ -69,17 +69,12 @@ void SigmoidCrossEntropyWithLogitsGradKernel(const Context &dev_ctx,
   dev_ctx.template Alloc<T>(counts_tensor);
   counts_tensor->Resize(in_grad->dims());
 
-  int limit = in_grad->numel();
-  int blocks = NumBlocks(limit);
-  int threads = kNumCUDAThreads;
   std::vector<const DenseTensor *> ins = {&x, &label, &out_grad};
   std::vector<DenseTensor *> outs = {in_grad, counts_tensor};
   auto functor = SigmoidBwdFunctor<T>(ignore_index);
-  constexpr int Size = 2;
-  phi::funcs::ElementwiseKernel<T, decltype(functor), Size>(
+  phi::funcs::ElementwiseKernel<T, decltype(functor), 2>(
       dev_ctx, ins, &outs, functor);
   if (normalize) {
-    T *counts = dev_ctx.template Alloc<T>(counts_tensor);
     DenseTensor *norm_tensor = new DenseTensor();
     norm_tensor->Resize({sizeof(T)});
     dev_ctx.template Alloc<T>(norm_tensor);
@@ -89,13 +84,8 @@ void SigmoidCrossEntropyWithLogitsGradKernel(const Context &dev_ctx,
       reduce_dim.push_back(i);
     }
 
-    funcs::TensorReduceImpl<T, T, kps::AddFunctor, NonzeroFunctor<T>>(
-        dev_ctx,
-        *counts_tensor,
-        norm_tensor,
-        NonzeroFunctor<T>(),
-        reduce_dim,
-        dev_ctx.stream());
+    funcs::ReduceKernel<T, T, kps::AddFunctor, NonzeroFunctor<T>>(
+        dev_ctx, *counts_tensor, norm_tensor, NonzeroFunctor<T>(), reduce_dim);
     T *norm = dev_ctx.template Alloc<T>(norm_tensor);
     auto norm_cpu_mem = paddle::memory::Alloc(phi::CPUPlace(), sizeof(T));
     T *norm_cpu_ptr = reinterpret_cast<T *>(norm_cpu_mem->ptr());
@@ -114,6 +104,7 @@ void SigmoidCrossEntropyWithLogitsGradKernel(const Context &dev_ctx,
     phi::funcs::ElementwiseKernel<T>(dev_ctx, div_ins, &div_outs, div_functor);
     delete norm_tensor;
   }
+  delete counts_tensor;
 }
 
 }  // namespace phi

paddle/phi/kernels/gpu/sigmoid_cross_entropy_with_logits_kernel.cu

@@ -69,17 +69,12 @@ void SigmoidCrossEntropyWithLogitsKernel(const Context &dev_ctx,
   dev_ctx.template Alloc<T>(counts_tensor);
   counts_tensor->Resize(out->dims());
 
-  int limit = out->numel();
-  int blocks = NumBlocks(limit);
-  int threads = kNumCUDAThreads;
   std::vector<const DenseTensor *> ins = {&x, &label};
   std::vector<DenseTensor *> outs = {out, counts_tensor};
   auto functor = SigmoidFwdFunctor<T>(ignore_index);
-  constexpr int Size = 2;
-  phi::funcs::ElementwiseKernel<T, decltype(functor), Size>(
+  phi::funcs::ElementwiseKernel<T, decltype(functor), 2>(
       dev_ctx, ins, &outs, functor);
   if (normalize) {
-    T *counts = dev_ctx.template Alloc<T>(counts_tensor);
     DenseTensor *norm_tensor = new DenseTensor();
     norm_tensor->Resize({sizeof(T)});
     dev_ctx.template Alloc<T>(norm_tensor);
@@ -89,13 +84,8 @@ void SigmoidCrossEntropyWithLogitsKernel(const Context &dev_ctx,
       reduce_dim.push_back(i);
     }
 
-    funcs::TensorReduceImpl<T, T, kps::AddFunctor, NonzeroFunctor<T>>(
-        dev_ctx,
-        *counts_tensor,
-        norm_tensor,
-        NonzeroFunctor<T>(),
-        reduce_dim,
-        dev_ctx.stream());
+    funcs::ReduceKernel<T, T, kps::AddFunctor, NonzeroFunctor<T>>(
+        dev_ctx, *counts_tensor, norm_tensor, NonzeroFunctor<T>(), reduce_dim);
     T *norm = dev_ctx.template Alloc<T>(norm_tensor);
     auto norm_cpu_mem = paddle::memory::Alloc(phi::CPUPlace(), sizeof(T));
     T *norm_cpu_ptr = reinterpret_cast<T *>(norm_cpu_mem->ptr());
@@ -114,8 +104,8 @@ void SigmoidCrossEntropyWithLogitsKernel(const Context &dev_ctx,
     phi::funcs::ElementwiseKernel<T>(dev_ctx, div_ins, &div_outs, div_functor);
 
     delete norm_tensor;
-    delete counts_tensor;
   }
+  delete counts_tensor;
 }
 
 }  // namespace phi

paddle/phi/kernels/gpu/trace_kernel.cu

@@ -31,11 +31,10 @@ void TraceKernel(const Context& ctx,
   T* out_data = ctx.template Alloc<T>(out);
   auto diag = funcs::Diagonal<T, Context>(ctx, &x, offset, axis1, axis2);
   if (diag.numel() > 0) {
-    auto stream = ctx.stream();
     std::vector<int> reduce_dims;
     reduce_dims.push_back(out->dims().size());
-    funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
-        ctx, diag, out, kps::IdentityFunctor<T>(), reduce_dims, stream);
+    funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+        ctx, diag, out, kps::IdentityFunctor<T>(), reduce_dims);
   } else {
     phi::funcs::SetConstant<Context, T> functor;
     functor(ctx, out, static_cast<T>(0));

paddle/phi/kernels/impl/matmul_grad_kernel_impl.h

@@ -59,9 +59,8 @@ struct ReduceSumForMatmulGrad<GPUContext, T> {
                   const DenseTensor& input,
                   DenseTensor* output,
                   const std::vector<int>& reduce_dims) {
-    auto stream = dev_ctx.stream();
-    funcs::TensorReduceImpl<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
-        dev_ctx, input, output, kps::IdentityFunctor<T>(), reduce_dims, stream);
+    funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<T>>(
+        dev_ctx, input, output, kps::IdentityFunctor<T>(), reduce_dims);
   }
 };
 #endif