[cherry-pick]Delete ElementwiseKernel in BroadcastKernel (#42779) (#43210)

Delete ElementwiseKernel in BroadcastKernel
减少所有Broadcast中重复功能调用，同时减少编译时间和问题体积

paddle/phi/kernels/funcs/broadcast_function.h

@@ -496,26 +496,16 @@ void BroadcastKernel(const KPDevice &ctx,
                      Functor func) {
   std::vector<int> dims_size;
   dims_size.reserve(ins.size());
-  bool no_broadcast_flag = true;
   for (auto *in : ins) {
-    no_broadcast_flag &= ins[0]->dims() == in->dims();
     dims_size.emplace_back(in->dims().size());
   }
 
-  if (ins.size() > 0 && outs->size() > 0) {
-    no_broadcast_flag &= outs->at(0)->dims() == ins[0]->dims();
-  }
-
-  if (no_broadcast_flag) {
-    phi::funcs::ElementwiseKernel<OutT, Functor, NumOuts>(ctx, ins, outs, func);
-  } else {
   axis = axis == -1
              ? *std::max_element(dims_size.begin(), dims_size.end()) -
                    *std::min_element(dims_size.begin(), dims_size.end())
              : axis;
   BroadcastKernelForDifferentVecSize<ET, InT, OutT, Functor, NumOuts>(
       ctx, ins, outs, axis, func);
-  }
 }
 
 template <typename Functor, typename T, typename OutType = T>

paddle/phi/kernels/gpu/bitwise_kernel.cu

@@ -46,9 +46,9 @@ void BitwiseNotKernel(const Context& dev_ctx,
   dev_ctx.template Alloc<T>(out);
   std::vector<const DenseTensor*> ins = {&x};
   std::vector<DenseTensor*> outs = {out};
-  funcs::BitwiseNotFunctor<T> func;
-  funcs::BroadcastKernel<ElementwiseType::kUnary, T, T>(
-      dev_ctx, ins, &outs, -1, func);
+  funcs::BitwiseNotFunctor<T> unary_func;
+  funcs::ElementwiseKernel<T, funcs::BitwiseNotFunctor<T>>(
+      dev_ctx, ins, &outs, unary_func);
 }
 
 }  // namespace phi

paddle/phi/kernels/gpu/gelu_grad_kernel.cu

@@ -81,11 +81,13 @@ void GeluGradKernel(const Context& dev_ctx,
       }
     }
 #endif
-    phi::funcs::BroadcastKernel<ElementwiseType::kBinary, T, T>(
-        dev_ctx, ins, &outs, 0, GeluWithApproximateGradFunctor<T>());
+    using Functor = GeluWithApproximateGradFunctor<T>;
+    phi::funcs::ElementwiseKernel<T, Functor, 1>(
+        dev_ctx, ins, &outs, Functor());
   } else {
-    phi::funcs::BroadcastKernel<ElementwiseType::kBinary, T, T>(
-        dev_ctx, ins, &outs, 0, GeluWithoutApproximateGradFunctor<T>());
+    using Functor = GeluWithoutApproximateGradFunctor<T>;
+    phi::funcs::ElementwiseKernel<T, Functor, 1>(
+        dev_ctx, ins, &outs, Functor());
   }
 }
 

paddle/phi/kernels/gpu/gelu_kernel.cu

@@ -71,11 +71,13 @@ void GeluKernel(const Context& dev_ctx,
       }
     }
 #endif
-    phi::funcs::BroadcastKernel<ElementwiseType::kBinary, T, T>(
-        dev_ctx, ins, &outs, 0, GeluWithApproximateFunctor<T>());
+    using Functor = GeluWithApproximateFunctor<T>;
+    phi::funcs::ElementwiseKernel<T, Functor, 1>(
+        dev_ctx, ins, &outs, Functor());
   } else {
-    phi::funcs::BroadcastKernel<ElementwiseType::kBinary, T, T>(
-        dev_ctx, ins, &outs, 0, GeluWithoutApproximateFunctor<T>());
+    using Functor = GeluWithoutApproximateFunctor<T>;
+    phi::funcs::ElementwiseKernel<T, Functor, 1>(
+        dev_ctx, ins, &outs, Functor());
   }
 }
 

paddle/phi/kernels/gpu/reduce_grad.h

@@ -43,22 +43,19 @@ void ReduceGrad(const GPUContext& dev_ctx,
       }));
 }
 
-template <typename T,
-          typename Context,
-          template <typename, typename> class TransformOp>
+template <typename T, typename OutT, typename Context, typename Functor>
 void ReduceGradKernel(const Context& dev_ctx,
                       const DenseTensor& x,
                       const DenseTensor& out_grad,
                       const std::vector<int64_t>& dims,
                       bool keep_dim,
                       bool reduce_all,
-                      DenseTensor* x_grad) {
+                      DenseTensor* x_grad,
+                      Functor functor) {
   auto* in_x = &x;
   auto* d_out = &out_grad;
   auto* d_x = x_grad;
 
-  auto pt_out_dtype = x.dtype();
-
   // get reduce_dim and reduce_num for reduce_mean_grad
   int dim_size = in_x->dims().size();
   std::vector<int> reduce_dims =
@@ -79,14 +76,10 @@ void ReduceGradKernel(const Context& dev_ctx,
 
   auto pt_d_out = new_d_out;
   auto pt_d_x = *d_x;
-  using MPType = typename kps::details::MPTypeTrait<T>::Type;
-
-  phi::ReduceGrad<T, TransformOp<T, MPType>>(
-      dev_ctx,
-      &pt_d_out,
-      &pt_d_x,
-      pt_out_dtype,
-      TransformOp<T, MPType>(reduce_num));
+  std::vector<const DenseTensor*> inputs = {&pt_d_out};
+  std::vector<DenseTensor*> outputs = {&pt_d_x};
+  funcs::BroadcastKernel<phi::ElementwiseType::kUnary, T, OutT>(
+      dev_ctx, inputs, &outputs, 0, functor);
 }
 
 }  // namespace phi

paddle/phi/kernels/gpu/reduce_mean_grad_kernel.cu

@@ -29,8 +29,23 @@ void ReduceMeanGradKernel(const Context& dev_ctx,
                           bool keep_dim,
                           bool reduce_all,
                           DenseTensor* x_grad) {
-  ReduceGradKernel<T, Context, kps::DivideFunctor>(
-      dev_ctx, x, out_grad, dims, keep_dim, reduce_all, x_grad);
+  int dim_size = x.dims().size();
+  std::vector<int> reduce_dims =
+      funcs::details::GetReduceDim(dims, dim_size, reduce_all);
+  int reduce_num = 1;
+  for (auto i : reduce_dims) {
+    reduce_num *= (x.dims())[i];
+  }
+  using MPType = typename kps::details::MPTypeTrait<T>::Type;
+  ReduceGradKernel<T, T, Context, kps::DivideFunctor<T, MPType>>(
+      dev_ctx,
+      x,
+      out_grad,
+      dims,
+      keep_dim,
+      reduce_all,
+      x_grad,
+      kps::DivideFunctor<T, MPType>(reduce_num));
 }
 
 }  // namespace phi

paddle/phi/kernels/gpu/reduce_sum_grad_kernel.cu

@@ -29,8 +29,40 @@ void ReduceSumGradKernel(const Context& dev_ctx,
                          bool keep_dim,
                          bool reduce_all,
                          DenseTensor* x_grad) {
-  ReduceGradKernel<T, Context, kps::IdentityFunctor>(
-      dev_ctx, x, out_grad, dims, keep_dim, reduce_all, x_grad);
+  using MPType = typename kps::details::MPTypeTrait<T>::Type;
+  auto out_dtype = x.dtype();
+  auto* in_x = &x;
+  auto* d_out = &out_grad;
+  auto* d_x = x_grad;
+
+  // get reduce_dim and reduce_num for reduce_mean_grad
+  int dim_size = in_x->dims().size();
+  std::vector<int> reduce_dims =
+      funcs::details::GetReduceDim(dims, dim_size, reduce_all);
+
+  auto update_dims = vectorize(d_x->dims());
+  int reduce_num = 1;
+  for (auto i : reduce_dims) {
+    reduce_num *= (in_x->dims())[i];
+    update_dims[i] = 1;
+  }
+  // make new tensor
+  DenseTensor new_d_out(d_out->dtype());
+  new_d_out.ShareDataWith(*d_out);
+  new_d_out.Resize(phi::make_ddim(update_dims));
+
+  dev_ctx.Alloc(d_x, x.dtype());
+  auto pt_out_dtype = x.dtype();
+  auto pt_d_out = new_d_out;
+  auto pt_d_x = *d_x;
+  std::vector<const DenseTensor*> inputs = {&pt_d_out};
+  std::vector<DenseTensor*> outputs = {&pt_d_x};
+  phi::ReduceGrad<T, kps::IdentityFunctor<T, MPType>>(
+      dev_ctx,
+      &pt_d_out,
+      &pt_d_x,
+      pt_out_dtype,
+      kps::IdentityFunctor<T, MPType>());
 }
 
 }  // namespace phi
@@ -48,4 +80,3 @@ PD_REGISTER_KERNEL(sum_grad,
                    int64_t,
                    phi::dtype::complex<float>,
                    phi::dtype::complex<double>) {}
-

paddle/phi/kernels/gpu/where_kernel.cu

@@ -40,8 +40,7 @@ void WhereKernel(const Context& ctx,
   ctx.template Alloc<T>(out);
 
   CondFunctor<T> func;
-  funcs::BroadcastKernel<ElementwiseType::kTernary, T, T>(
-      ctx, ins, &outs, -1, func);
+  funcs::ElementwiseKernel<T, CondFunctor<T>, 1>(ctx, ins, &outs, func);
 }
 
 }  // namespace phi