PaddlePaddle · raindrops2sea · Apr 28, 2023 · Mar 24, 2023 · Mar 28, 2023 · Apr 7, 2023
diff --git a/paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h b/paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h
@@ -19,17 +19,13 @@
 namespace paddle {
 namespace operators {
 
-template <ElementwiseType ET,
-          typename InT,
-          typename OutT,
-          typename Functor,
-          int NumOuts = 1>
+template <typename OutT, typename Functor, int NumOuts = 1>
 void LaunchElementwiseCudaKernel(
     const KPDevice &ctx,
     const std::vector<const phi::DenseTensor *> &ins,
     std::vector<phi::DenseTensor *> *outs,
-    int axis,
-    Functor func) {
+    Functor func,
+    int axis = -1) {
   std::vector<const phi::DenseTensor *> pt_inputs;
   std::vector<phi::DenseTensor *> pt_outputs;
   // TODO(YuanRisheng) *_tmp for cache DenseTensor, because the temporary
@@ -53,8 +49,8 @@ void LaunchElementwiseCudaKernel(
   for (int i = 0; i < pt_outputs_tmp.size(); i++) {
     pt_outputs.push_back(pt_outputs_tmp[i].get());
   }
-  phi::funcs::BroadcastKernel<ET, InT, OutT, Functor, NumOuts>(
-      ctx, pt_inputs, &pt_outputs, axis, func);
+  phi::funcs::BroadcastKernel<OutT, Functor, NumOuts>(
+      ctx, pt_inputs, &pt_outputs, func, axis);
 }
 
 }  // namespace operators

diff --git a/paddle/fluid/operators/elementwise/elementwise_op_function.h b/paddle/fluid/operators/elementwise/elementwise_op_function.h
@@ -188,7 +188,7 @@ void ElementwiseComputeEx(const framework::ExecutionContext &ctx,
   z->mutable_data<OutType>(ctx.GetPlace());
   const auto &dev_ctx = ctx.template device_context<DeviceContext>();
   phi::funcs::ElementwiseCompute<Functor, T, OutType>(
-      dev_ctx, *x, *y, axis, func, z);
+      dev_ctx, *x, *y, func, z, axis);
 }
 
 // FusedElemwiseAndAct
@@ -1596,7 +1596,7 @@ static inline std::vector<int> GetReduceDim(const framework::DDim &in,
 
 #if defined(__NVCC__) || defined(__HIPCC__)
 
-template <ElementwiseType ET, typename T, typename Functor>
+template <typename T, typename Functor>
 void GetGradXAndYOut(const phi::GPUContext &dev_ctx,
                      const platform::Place &place,
                      int axis,
@@ -1605,20 +1605,19 @@ void GetGradXAndYOut(const phi::GPUContext &dev_ctx,
                      phi::DenseTensor *dx,
                      phi::DenseTensor *dy,
                      Functor func) {
-  phi::GetGradXAndYOut<ET, T, Functor>(
+  phi::GetGradXAndYOut<T, Functor>(
       dev_ctx, place, axis, ins, *dout, dx, dy, func);
 }
 
-template <ElementwiseType ET, typename T, typename Functor>
+template <typename T, typename Functor>
 void GetGradXOrYOut(const phi::GPUContext &dev_ctx,
                     const platform::Place &place,
                     int axis,
                     std::vector<const phi::DenseTensor *> ins,
                     const phi::DenseTensor *dout,
                     phi::DenseTensor *dxy,
                     Functor func) {
-  phi::GetGradXOrYOut<ET, T, Functor>(
-      dev_ctx, place, axis, ins, *dout, dxy, func);
+  phi::GetGradXOrYOut<T, Functor>(dev_ctx, place, axis, ins, *dout, dxy, func);
 }
 
 #endif

diff --git a/paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h b/paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h
@@ -23,8 +23,6 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-using ElementwiseType = phi::ElementwiseType;
-
 template <typename OutT, typename Functor, int NumOuts = 1>
 void LaunchSameDimsElementwiseCudaKernel(
     const KPDevice &ctx,

diff --git a/paddle/fluid/operators/fused/attn_gemm.h b/paddle/fluid/operators/fused/attn_gemm.h
@@ -109,8 +109,8 @@ class AttnMatMul {
       // bias_out = output + bias
       std::vector<const phi::DenseTensor*> ins = {output, bias};
       std::vector<phi::DenseTensor*> outs = {bias_out};
-      phi::funcs::BroadcastKernel<phi::ElementwiseType::kBinary, T, T>(
-          dev_ctx_, ins, &outs, -1, phi::funcs::AddFunctor<T>());
+      phi::funcs::BroadcastKernel<T>(
+          dev_ctx_, ins, &outs, phi::funcs::AddFunctor<T>());
     }
   }
 

diff --git a/paddle/fluid/operators/fused/attn_gemm_int8.h b/paddle/fluid/operators/fused/attn_gemm_int8.h
@@ -85,8 +85,8 @@ class AttnMatmulINT8 {
       // bias_out = output + bias
       std::vector<const phi::DenseTensor*> ins = {output, bias};
       std::vector<phi::DenseTensor*> outs = {bias_out};
-      phi::funcs::BroadcastKernel<phi::ElementwiseType::kBinary, T, T>(
-          dev_ctx_, ins, &outs, -1, phi::funcs::AddFunctor<T>());
+      phi::funcs::BroadcastKernel<T>(
+          dev_ctx_, ins, &outs, phi::funcs::AddFunctor<T>());
       PADDLE_ENFORCE_EQ(cudaGetLastError(),
                         cudaSuccess,
                         platform::errors::Fatal(
@@ -139,8 +139,8 @@ class AttnMatmulINT8 {
       // bias_out = output + bias
       std::vector<const phi::DenseTensor*> ins = {output, bias};
       std::vector<phi::DenseTensor*> outs = {bias_out};
-      phi::funcs::BroadcastKernel<phi::ElementwiseType::kBinary, T, T>(
-          dev_ctx_, ins, &outs, -1, phi::funcs::AddFunctor<T>());
+      phi::funcs::BroadcastKernel<T>(
+          dev_ctx_, ins, &outs, phi::funcs::AddFunctor<T>());
       PADDLE_ENFORCE_EQ(cudaGetLastError(),
                         cudaSuccess,
                         platform::errors::Fatal(

diff --git a/paddle/fluid/operators/fused/fmha_ref.h b/paddle/fluid/operators/fused/fmha_ref.h
@@ -255,12 +255,11 @@ class FMHARef {
         ins.emplace_back(src_mask_tensor);
         outs.emplace_back(src_mask_out_tensor);
         int elewise_add_axis = -1;
-        phi::funcs::BroadcastKernel<phi::ElementwiseType::kBinary, T, T>(
-            dev_ctx_,
-            ins,
-            &outs,
-            elewise_add_axis,
-            phi::funcs::AddFunctor<T>());
+        phi::funcs::BroadcastKernel<T>(dev_ctx_,
+                                       ins,
+                                       &outs,
+                                       phi::funcs::AddFunctor<T>(),
+                                       elewise_add_axis);
 
         phi::SoftmaxForwardCUDAKernelDriver<T>(
             dev_ctx_, *src_mask_out_tensor, softmax_axis, softmax_out_tensor);
@@ -432,12 +431,11 @@ class FMHARef {
         ins.emplace_back(src_mask_tensor);
         outs.emplace_back(src_mask_out_tensor);
         int elewise_add_axis = -1;
-        phi::funcs::BroadcastKernel<phi::ElementwiseType::kBinary, T, T>(
-            dev_ctx_,
-            ins,
-            &outs,
-            elewise_add_axis,
-            phi::funcs::AddFunctor<T>());
+        phi::funcs::BroadcastKernel<T>(dev_ctx_,
+                                       ins,
+                                       &outs,
+                                       phi::funcs::AddFunctor<T>(),
+                                       elewise_add_axis);
 
         phi::SoftmaxForwardCUDAKernelDriver<T>(
             dev_ctx_, *src_mask_out_tensor, softmax_axis, softmax_out_tensor);

diff --git a/paddle/fluid/operators/fused/fused_gate_attention.h b/paddle/fluid/operators/fused/fused_gate_attention.h
@@ -689,13 +689,13 @@ class FMHAGateRef {
       std::vector<const phi::DenseTensor*> ins = {
           qk_out, src_mask, nonbatched_bias};
       std::vector<phi::DenseTensor*> outs = {qk_out};
-      phi::funcs::BroadcastKernel<phi::ElementwiseType::kTernary, T, T>(
-          dev_ctx_, ins, &outs, -1, TernaryAddFunctor<T>());
+      phi::funcs::BroadcastKernel<T>(
+          dev_ctx_, ins, &outs, TernaryAddFunctor<T>());
     } else {
       std::vector<const phi::DenseTensor*> ins = {qk_out, src_mask};
       std::vector<phi::DenseTensor*> outs = {qk_out};
-      phi::funcs::BroadcastKernel<phi::ElementwiseType::kBinary, T, T>(
-          dev_ctx_, ins, &outs, -1, phi::funcs::AddFunctor<T>());
+      phi::funcs::BroadcastKernel<T>(
+          dev_ctx_, ins, &outs, phi::funcs::AddFunctor<T>());
     }
     phi::SoftmaxForwardCUDAKernelDriver<T>(dev_ctx_, *qk_out, -1, softmax_out);
   }

diff --git a/paddle/fluid/operators/fused_token_prune_op.cu b/paddle/fluid/operators/fused_token_prune_op.cu
@@ -141,8 +141,7 @@ class FusedTokenPruneOpCUDAKernel : public framework::OpKernel<T> {
     ins.emplace_back(attn);
     ins.emplace_back(mask);
     outs.emplace_back(&attn_tmp);
-    LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
-        dev_ctx, ins, &outs, -1, AttnMaskFunctor<T>());
+    LaunchElementwiseCudaKernel<T>(dev_ctx, ins, &outs, AttnMaskFunctor<T>());
 
     // 2. Reduce sum
     const std::vector<int64_t> reduce_dims{1, 2};

diff --git a/paddle/fluid/operators/reduce_ops/reduce_op.h b/paddle/fluid/operators/reduce_ops/reduce_op.h
@@ -834,12 +834,11 @@ class ReduceCudaGradKernel : public framework::OpKernel<T> {
     }
 
     using MPType = typename kps::details::MPTypeTrait<T>::Type;
-    phi::ReduceGrad<T, TransformOp<T, MPType>>(
-        dev_ctx,
-        pt_d_out.get(),
-        pt_d_x.get(),
-        pt_out_dtype,
-        TransformOp<T, MPType>(reduce_num));
+    phi::ReduceGrad<TransformOp<T, MPType>>(dev_ctx,
+                                            pt_d_out.get(),
+                                            pt_d_x.get(),
+                                            pt_out_dtype,
+                                            TransformOp<T, MPType>(reduce_num));
   }
 };
 

diff --git a/paddle/phi/kernels/cpu/bitwise_kernel.cc b/paddle/phi/kernels/cpu/bitwise_kernel.cc
@@ -24,15 +24,15 @@ limitations under the License. */
 
 namespace phi {
 
-#define DEFINE_BITWISE_KERNEL(op_type)                                    \
-  template <typename T, typename Context>                                 \
-  void Bitwise##op_type##Kernel(const Context& dev_ctx,                   \
-                                const DenseTensor& x,                     \
-                                const DenseTensor& y,                     \
-                                DenseTensor* out) {                       \
-    funcs::Bitwise##op_type##Functor<T> func;                             \
-    funcs::ElementwiseCompute<funcs::Bitwise##op_type##Functor<T>, T, T>( \
-        dev_ctx, x, y, -1, func, out);                                    \
+#define DEFINE_BITWISE_KERNEL(op_type)                                 \
+  template <typename T, typename Context>                              \
+  void Bitwise##op_type##Kernel(const Context& dev_ctx,                \
+                                const DenseTensor& x,                  \
+                                const DenseTensor& y,                  \
+                                DenseTensor* out) {                    \
+    funcs::Bitwise##op_type##Functor<T> func;                          \
+    funcs::ElementwiseCompute<funcs::Bitwise##op_type##Functor<T>, T>( \
+        dev_ctx, x, y, func, out);                                     \
   }
 
 DEFINE_BITWISE_KERNEL(And)

diff --git a/paddle/phi/kernels/cpu/compare_kernel.cc b/paddle/phi/kernels/cpu/compare_kernel.cc
@@ -33,10 +33,10 @@ inline void CompareKernelImpl(const Context& ctx,
   ctx.template Alloc<bool>(out);
   if (x.dims().size() >= y.dims().size()) {
     funcs::ElementwiseCompute<Functor, T, bool>(
-        ctx, x, y, axis, Functor(), out);
+        ctx, x, y, Functor(), out, axis);
   } else {
     funcs::ElementwiseCompute<InverseFunctor, T, bool>(
-        ctx, x, y, axis, InverseFunctor(), out);
+        ctx, x, y, InverseFunctor(), out, axis);
   }
 }
 
@@ -59,7 +59,7 @@ inline void CompareAllKernelImpl(const Context& ctx,
       tmp_data[0] = Functor()(x.data<T>()[0], y.data<T>()[0]);
     } else {
       funcs::ElementwiseCompute<Functor, T, bool>(
-          ctx, x, y, 0, Functor(), &tmp);
+          ctx, x, y, Functor(), &tmp, 0);
     }
     auto tmp_flat = EigenVector<bool>::Flatten(tmp);
     auto out_es = EigenScalar<bool>::From(*out);

diff --git a/paddle/phi/kernels/cpu/dirichlet_kernel.cc b/paddle/phi/kernels/cpu/dirichlet_kernel.cc
@@ -91,8 +91,8 @@ struct DirichletSampler<CPUContext, T> {
         true,
         false);
 
-    funcs::ElementwiseCompute<funcs::DivideFunctor<T>, T, T>(
-        dev_ctx, gamma_samples, gamma_sum, -1, funcs::DivideFunctor<T>(), out);
+    funcs::ElementwiseCompute<funcs::DivideFunctor<T>, T>(
+        dev_ctx, gamma_samples, gamma_sum, funcs::DivideFunctor<T>(), out);
   }
 };
 

diff --git a/paddle/phi/kernels/cpu/elementwise_divide_kernel.cc b/paddle/phi/kernels/cpu/elementwise_divide_kernel.cc
@@ -38,10 +38,10 @@ void DivideRawKernel(const Context& dev_ctx,
     auto y_dims = y.dims();
     if (x_dims.size() >= y_dims.size()) {
       funcs::ElementwiseCompute<funcs::DivideFunctor<T>, T>(
-          dev_ctx, x, y, axis, funcs::DivideFunctor<T>(), out);
+          dev_ctx, x, y, funcs::DivideFunctor<T>(), out, axis);
     } else {
       funcs::ElementwiseCompute<funcs::InverseDivideFunctor<T>, T>(
-          dev_ctx, x, y, axis, funcs::InverseDivideFunctor<T>(), out);
+          dev_ctx, x, y, funcs::InverseDivideFunctor<T>(), out, axis);
     }
   }
 }

diff --git a/paddle/phi/kernels/cpu/elementwise_kernel.cc b/paddle/phi/kernels/cpu/elementwise_kernel.cc
@@ -75,7 +75,7 @@ void HeavisideKernel(const Context& dev_ctx,
   // allocate memory for out
   dev_ctx.template Alloc<T>(out);
   funcs::ElementwiseCompute<funcs::ElementwiseHeavisideFunctor<T>, T>(
-      dev_ctx, x, y, -1, funcs::ElementwiseHeavisideFunctor<T>(), out);
+      dev_ctx, x, y, funcs::ElementwiseHeavisideFunctor<T>(), out);
 }
 
 }  // namespace phi