Xpu ernie3: support fp16 for xpu kernels: full_like/stack/where (#51271)

* [xpu-ernie3] support fp16 for full_like/stack/where xpu kernels

* [xpu-ernie3] support fp16 for full_like/stack/where xpu kernels

paddle/phi/backends/xpu/xpu2_op_list.cc

@@ -319,6 +319,16 @@ XPUOpMap& get_kl2_ops() {
                      phi::DataType::INT32,
                      phi::DataType::INT8,
                      phi::DataType::FLOAT32})},
+      {"full_batch_size_like",
+       XPUKernelSet({phi::DataType::INT64,
+                     phi::DataType::INT32,
+                     phi::DataType::FLOAT32,
+                     phi::DataType::FLOAT16})},
+      {"fill_constant_batch_size_like",
+       XPUKernelSet({phi::DataType::INT64,
+                     phi::DataType::INT32,
+                     phi::DataType::FLOAT32,
+                     phi::DataType::FLOAT16})},
       {"unfold",
        XPUKernelSet({phi::DataType::FLOAT32, phi::DataType::FLOAT16})},
       {"unfold_grad",
@@ -595,6 +605,7 @@ XPUOpMap& get_kl2_ops() {
       {"shape",
        XPUKernelSet({phi::DataType::FLOAT32,
                      phi::DataType::INT64,
+                     phi::DataType::INT32,
                      phi::DataType::FLOAT16})},
       {"sigmoid", XPUKernelSet({phi::DataType::FLOAT32})},
       {"sigmoid_grad", XPUKernelSet({phi::DataType::FLOAT32})},
@@ -668,7 +679,8 @@ XPUOpMap& get_kl2_ops() {
       {"stack",
        XPUKernelSet({phi::DataType::FLOAT32,
                      phi::DataType::INT64,
-                     phi::DataType::INT32})},
+                     phi::DataType::INT32,
+                     phi::DataType::FLOAT16})},
       {"stack_grad",
        XPUKernelSet({phi::DataType::FLOAT32, phi::DataType::INT32})},
       {"strided_slice",
@@ -810,7 +822,8 @@ XPUOpMap& get_kl2_ops() {
       {"where",
        XPUKernelSet({phi::DataType::INT32,
                      phi::DataType::INT64,
-                     phi::DataType::FLOAT32})},
+                     phi::DataType::FLOAT32,
+                     phi::DataType::FLOAT16})},
       {"sin", XPUKernelSet({phi::DataType::FLOAT32})},
       {"cos", XPUKernelSet({phi::DataType::FLOAT32})},
       {"linspace",

paddle/phi/kernels/xpu/full_kernel.cc

@@ -88,13 +88,32 @@ void FullLikeKernel(const Context& dev_ctx,
                     phi::errors::InvalidArgument("The filled value is Inf."));
 
   auto out_data = reinterpret_cast<XPUInTDType*>(out->data<T>());
-  int r = xpu::constant(dev_ctx.x_context(),
-                        out_data,
-                        out->numel(),
-                        static_cast<XPUInTDType>(value));
-  PADDLE_ENFORCE_XDNN_SUCCESS(r, "constant");
+  if (out->numel() > 0) {
+    int r = xpu::constant(dev_ctx.x_context(),
+                          out_data,
+                          out->numel(),
+                          static_cast<XPUInTDType>(value));
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "constant");
+  }
 }
 
+template <typename T, typename Context>
+void FullBatchSizeLikeKernel(const Context& dev_ctx,
+                             const DenseTensor& x,
+                             const std::vector<int>& shape,
+                             const Scalar& val,
+                             DataType dtype,
+                             int x_batch_size_dim,
+                             int out_batch_size_dim,
+                             DenseTensor* out) {
+  if (x.lod().size() && x_batch_size_dim == 0) {
+    // set the correct batch size for the LoDTensor.
+    auto odims = out->dims();
+    odims[out_batch_size_dim] = static_cast<int>(x.lod().back().size()) - 1;
+    FullKernel<T, Context>(dev_ctx, phi::vectorize(odims), val, dtype, out);
+  }
+  FullLikeKernel<T, Context>(dev_ctx, x, val, dtype, out);
+}
 }  // namespace phi
 
 PD_REGISTER_KERNEL(full,
@@ -122,3 +141,15 @@ PD_REGISTER_KERNEL(full_like,
                    phi::dtype::float16) {
   kernel->InputAt(0).SetBackend(phi::Backend::ALL_BACKEND);
 }
+
+PD_REGISTER_KERNEL(full_batch_size_like,
+                   XPU,
+                   ALL_LAYOUT,
+                   phi::FullBatchSizeLikeKernel,
+                   float,
+                   int,
+                   int64_t,
+                   bool,
+                   phi::dtype::float16) {
+  kernel->InputAt(0).SetBackend(phi::Backend::ALL_BACKEND);
+}

paddle/phi/kernels/xpu/stack_kernel.cc

@@ -55,5 +55,11 @@ void StackKernel(const Context& dev_ctx,
 }
 }  // namespace phi
 
-PD_REGISTER_KERNEL(
-    stack, XPU, ALL_LAYOUT, phi::StackKernel, float, int, int64_t) {}
+PD_REGISTER_KERNEL(stack,
+                   XPU,
+                   ALL_LAYOUT,
+                   phi::StackKernel,
+                   float,
+                   int,
+                   int64_t,
+                   phi::dtype::float16) {}

paddle/phi/kernels/xpu/where_kernel.cc

@@ -25,10 +25,11 @@ void WhereKernel(const Context& ctx,
                  const DenseTensor& x,
                  const DenseTensor& y,
                  DenseTensor* out) {
+  using XPUType = typename XPUTypeTrait<T>::Type;
   const bool* cond_data = condition.data<bool>();
-  const T* x_data = x.data<T>();
-  const T* y_data = y.data<T>();
-  T* out_data = ctx.template Alloc<T>(out);
+  const XPUType* x_data = reinterpret_cast<const XPUType*>(x.data<T>());
+  const XPUType* y_data = reinterpret_cast<const XPUType*>(y.data<T>());
+  XPUType* out_data = reinterpret_cast<XPUType*>(ctx.template Alloc<T>(out));
 
   auto cond_dims = phi::vectorize<int>(condition.dims());
   auto x_dims = phi::vectorize<int>(x.dims());
@@ -44,10 +45,16 @@ void WhereKernel(const Context& ctx,
   int ret = xpu::select(
       ctx.x_context(), cond_data, x_data, y_data, out_data, cond_dims, x_dims);
 
-  PADDLE_ENFORCE_XDNN_SUCCESS(ret, "select");
+  PADDLE_ENFORCE_XDNN_SUCCESS(ret, "xpu::select");
 }
 
 }  // namespace phi
 
-PD_REGISTER_KERNEL(
-    where, XPU, ALL_LAYOUT, phi::WhereKernel, float, int, int64_t) {}
+PD_REGISTER_KERNEL(where,
+                   XPU,
+                   ALL_LAYOUT,
+                   phi::WhereKernel,
+                   float,
+                   int,
+                   int64_t,
+                   phi::dtype::float16) {}