Add aten::fractional_max_pool2d, fractional_max_pool3d (forward and backward) (#880)

- [x] fractional_max_pool2d
- [x] fractional_max_pool2d_backward
- [x] fractional_max_pool3d
- [x] fractional_max_pool3d_backward

---------

Co-authored-by: Yutao Xu <[email protected]>

Author: yucai-intel

src/ATen/native/xpu/FractionalMaxPool2d.cpp

@@ -0,0 +1,33 @@
+#include <ATen/core/Tensor.h>
+#include <ATen/core/op_registration/adaption.h>
+#include <ATen/native/cpu/mixed_data_type.h>
+#include <ATen/native/xpu/sycl/FractionalMaxPool2dKernels.h>
+
+#include <xpu/ATen/ops/fractional_max_pool2d_backward_native.h>
+#include <xpu/ATen/ops/fractional_max_pool2d_native.h>
+
+namespace at::native {
+
+TORCH_IMPL_FUNC(fractional_max_pool2d_out_xpu)
+(const Tensor& input,
+ IntArrayRef pool_size,
+ IntArrayRef output_size,
+ const Tensor& randomSamples,
+ const Tensor& output,
+ const Tensor& indices) {
+  xpu::fractional_max_pool2d_kernel(
+      input, pool_size, output_size, randomSamples, output, indices);
+}
+
+TORCH_IMPL_FUNC(fractional_max_pool2d_backward_xpu)
+(const Tensor& gradOutput,
+ const Tensor& input,
+ IntArrayRef pool_size /* unused */,
+ IntArrayRef output_size,
+ const Tensor& indices,
+ const Tensor& gradInput) {
+  xpu::fractional_max_pool2d_backward_kernel(
+      gradOutput, input, pool_size, output_size, indices, gradInput);
+}
+
+} // namespace at::native

src/ATen/native/xpu/FractionalMaxPool3d.cpp

@@ -0,0 +1,73 @@
+#include <ATen/core/Tensor.h>
+#include <ATen/core/op_registration/adaption.h>
+#include <ATen/native/cpu/mixed_data_type.h>
+#include <ATen/native/xpu/sycl/FractionalMaxPool3dKernels.h>
+#include <ATen/ops/empty.h>
+
+#include <xpu/ATen/ops/fractional_max_pool3d_backward_native.h>
+#include <xpu/ATen/ops/fractional_max_pool3d_native.h>
+
+namespace at::native {
+
+TORCH_IMPL_FUNC(fractional_max_pool3d_out_xpu)
+(const Tensor& input,
+ int64_t poolSizeT,
+ int64_t poolSizeH,
+ int64_t poolSizeW,
+ int64_t outputT,
+ int64_t outputH,
+ int64_t outputW,
+ const Tensor& randomSamples,
+ int64_t numBatch,
+ int64_t numPlanes,
+ int64_t inputT,
+ int64_t inputH,
+ int64_t inputW,
+ const Tensor& output,
+ const Tensor& indices) {
+  xpu::fractional_max_pool3d_kernel(
+      input,
+      poolSizeT,
+      poolSizeH,
+      poolSizeW,
+      outputT,
+      outputH,
+      outputW,
+      randomSamples,
+      numBatch,
+      numPlanes,
+      inputT,
+      inputH,
+      inputW,
+      output,
+      indices);
+}
+
+Tensor& fractional_max_pool3d_backward_out_xpu(
+    const Tensor& grad_output,
+    const Tensor& input,
+    IntArrayRef pool_size,
+    IntArrayRef output_size,
+    const Tensor& indices,
+    Tensor& grad_input) {
+  globalContext().alertNotDeterministic(
+      "fractional_max_pool3d_backward_out_xpu");
+  xpu::fractional_max_pool3d_backward_kernel(
+      grad_input, grad_output, input, output_size, indices);
+  return grad_input;
+}
+
+Tensor fractional_max_pool3d_backward_xpu(
+    const Tensor& grad_output,
+    const Tensor& input,
+    IntArrayRef pool_size,
+    IntArrayRef output_size,
+    const Tensor& indices) {
+  globalContext().alertNotDeterministic("fractional_max_pool3d_backward_xpu");
+  Tensor grad_input = at::empty({0}, input.options());
+  xpu::fractional_max_pool3d_backward_kernel(
+      grad_input, grad_output, input, output_size, indices);
+  return grad_input;
+}
+
+} // namespace at::native

src/ATen/native/xpu/XPUFallback.template

@@ -164,10 +164,6 @@ TORCH_LIBRARY_IMPL(aten, XPU, m) {
     "_fft_c2r",
     "_fft_r2c",
     "_flash_attention_forward",
-    "fractional_max_pool2d_backward.grad_input",
-    "fractional_max_pool2d.output",
-    "fractional_max_pool3d_backward",
-    "fractional_max_pool3d.output",
     "frexp.Tensor_out",
     "_fused_moving_avg_obs_fq_helper",
     "geqrf",