[PHI] add new supported datatype for tile and sigmoid_grad (#60119)

RuohengMa · web-flow · commit 4c975499456c · 2023-12-28T11:55:13.000+08:00
diff --git a/paddle/phi/backends/xpu/xpu2_op_list.cc b/paddle/phi/backends/xpu/xpu2_op_list.cc
@@ -835,7 +835,8 @@ XPUOpMap& get_kl2_ops() {
                      phi::DataType::BFLOAT16})},
       {"sigmoid",
        XPUKernelSet({phi::DataType::FLOAT32, phi::DataType::FLOAT16})},
-      {"sigmoid_grad", XPUKernelSet({phi::DataType::FLOAT32})},
+      {"sigmoid_grad",
+       XPUKernelSet({phi::DataType::FLOAT32, phi::DataType::FLOAT16})},
       {"sign", XPUKernelSet({phi::DataType::FLOAT32})},
       {"slice_grad",
        XPUKernelSet({phi::DataType::FLOAT32,
diff --git a/paddle/phi/kernels/xpu/activation_grad_kernel.cc b/paddle/phi/kernels/xpu/activation_grad_kernel.cc
@@ -734,14 +734,20 @@ PD_REGISTER_KERNEL(swish_grad,
                    phi::dtype::float16,
                    phi::dtype::bfloat16) {}
 
+PD_REGISTER_KERNEL(sigmoid_grad,
+                   XPU,
+                   ALL_LAYOUT,
+                   phi::SigmoidGradKernel,
+                   float,
+                   phi::dtype::float16) {}
+
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(exp_grad, ExpGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(log_grad, LogGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(leaky_relu_grad, LeakyReluGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(hardsigmoid_grad, HardSigmoidGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(hardswish_grad, HardSwishGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(reciprocal_grad, ReciprocalGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(relu6_grad, Relu6GradKernel)
-PD_REGISTER_ACTIVATION_GRAD_KERNEL(sigmoid_grad, SigmoidGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(sqrt_grad, SqrtGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(mish_grad, MishGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(softplus_grad, SoftplusGradKernel)
diff --git a/paddle/phi/kernels/xpu/tile_kernel.cc b/paddle/phi/kernels/xpu/tile_kernel.cc
@@ -29,7 +29,6 @@ void TileKernel(const Context& dev_ctx,
                 const DenseTensor& x,
                 const IntArray& repeat_times_arr,
                 DenseTensor* out) {
-  using XPUType = typename XPUTypeTrait<T>::Type;
   auto rank = x.dims().size();
   std::vector<int64_t> repeat_times = repeat_times_arr.GetData();
   int repeat_times_size = repeat_times.size();
@@ -123,24 +122,6 @@ void TileKernel(const Context& dev_ctx,
                                  vec_in_dims,
                                  vec_out_dims);
 
-  } else if (std::is_same<T, double>::value) {
-    float* x_t = RAII_GUARD.alloc_l3_or_gm<float>(x.numel());
-    float* y_t = RAII_GUARD.alloc_l3_or_gm<float>(out->numel());
-    int r =
-        xpu::cast<XPUType, float>(dev_ctx.x_context(),
-                                  reinterpret_cast<const XPUType*>(x.data<T>()),
-                                  x_t,
-                                  x.numel());
-    PADDLE_ENFORCE_XDNN_SUCCESS(r, "cast");
-    ret = xpu::broadcast<float>(
-        dev_ctx.x_context(), x_t, y_t, vec_in_dims, vec_out_dims);
-    PADDLE_ENFORCE_XDNN_SUCCESS(ret, "broadcast");
-    r = xpu::cast<float, XPUType>(dev_ctx.x_context(),
-                                  y_t,
-                                  reinterpret_cast<XPUType*>(out->data<T>()),
-                                  out->numel());
-    PADDLE_ENFORCE_XDNN_SUCCESS(r, "cast");
-
   } else {
     ret = xpu::broadcast<T>(dev_ctx.x_context(),
                             x.data<T>(),
