partially codegen adaptive_avgpool3d and backward (#3790)

* partially codegen adaptive_av_pool3d and backward

* Delete .torch_pin

Author: JackCaoG

torch_xla/csrc/aten_xla_type.cpp

@@ -21,6 +21,7 @@
 #include "torch_xla/csrc/aten_xla_bridge.h"
 #include "torch_xla/csrc/debug_util.h"
 #include "torch_xla/csrc/device.h"
+#include "torch_xla/csrc/generated/LazyIr.h"
 #include "torch_xla/csrc/generated/XLANativeFunctions.h"
 #include "torch_xla/csrc/helpers.h"
 #include "torch_xla/csrc/ops/as_strided.h"
@@ -330,8 +331,17 @@ at::Tensor XLANativeFunctions::_adaptive_avg_pool3d(
         &xla_cpu_fallback, ATEN_OP(_adaptive_avg_pool3d)>::call(self,
                                                                 output_size);
   }
-  return bridge::AtenFromXlaTensor(XLATensor::adaptive_avg_pool3d(
-      bridge::GetXlaTensor(self), output_size_list));
+  auto common_device = torch_xla::bridge::GetXlaDevice(self);
+  XLA_CHECK(common_device);
+  auto shapes =
+      torch::lazy::compute_shape__adaptive_avg_pool3d(self, output_size);
+  XLA_CHECK(shapes.size() == 1);
+  torch::lazy::NodePtr node = torch::lazy::MakeNode<AdaptiveAvgPool3d>(
+      bridge::GetXlaTensor(self)->GetIrValue(),
+      std::vector<int64_t>(output_size.begin(), output_size.end()),
+      std::move(shapes));
+  return torch_xla::bridge::AtenFromXlaTensor(
+      torch_xla::XLATensor::Create(std::move(node), *common_device));
 }
 
 at::Tensor XLANativeFunctions::_adaptive_avg_pool3d_backward(
@@ -347,8 +357,17 @@ at::Tensor XLANativeFunctions::_adaptive_avg_pool3d_backward(
         &xla_cpu_fallback,
         ATEN_OP(_adaptive_avg_pool3d_backward)>::call(grad_output, self);
   }
-  return bridge::AtenFromXlaTensor(XLATensor::adaptive_avg_pool3d_backward(
-      bridge::GetXlaTensor(grad_output), bridge::GetXlaTensor(self)));
+  auto common_device = torch_xla::bridge::GetXlaDevice(grad_output, self);
+  XLA_CHECK(common_device);
+  auto shapes = torch::lazy::compute_shape__adaptive_avg_pool3d_backward(
+      grad_output, self);
+  XLA_CHECK(shapes.size() == 1);
+  torch::lazy::NodePtr node = torch::lazy::MakeNode<AdaptiveAvgPool3dBackward>(
+      bridge::GetXlaTensor(grad_output)->GetIrValue(),
+      bridge::GetXlaTensor(self)->GetIrValue(), std::move(shapes));
+
+  return torch_xla::bridge::AtenFromXlaTensor(
+      torch_xla::XLATensor::Create(std::move(node), *common_device));
 }
 
 at::Tensor XLANativeFunctions::_adaptive_avg_pool2d(

torch_xla/csrc/ops/adaptive_avg_pool3d.cpp

@@ -355,32 +355,6 @@ torch::lazy::NodePtr AdaptiveMaxPool2dBackward(
                    std::move(lower_fn));
 }
 
-torch::lazy::NodePtr AdaptiveAvgPool3dBackward(
-    const torch::lazy::Value& grad_output, const torch::lazy::Value& input) {
-  auto lower_fn = [](const XlaNode& node,
-                     LoweringContext* loctx) -> XlaOpVector {
-    xla::XlaOp grad_output = loctx->GetOutputOp(node.operand(0));
-    xla::XlaOp input = loctx->GetOutputOp(node.operand(1));
-    xla::XlaOp xla_output = BuildAdaptiveAvgPool3dBackward(
-        /*out_backprop=*/grad_output, /*input=*/input);
-    return node.ReturnOp(xla_output, loctx);
-  };
-  auto lower_for_shape_fn =
-      [](absl::Span<const xla::XlaOp> operands) -> xla::XlaOp {
-    XLA_CHECK_EQ(operands.size(), 2);
-    return BuildAdaptiveAvgPool3dBackward(/*out_backprop=*/operands[0],
-                                          /*input=*/operands[1]);
-  };
-  return GenericOp(torch::lazy::OpKind(at::aten::adaptive_avg_pool3d_backward),
-                   {grad_output, input},
-                   [&]() {
-                     return InferOutputShape(
-                         {GetXlaShape(grad_output), GetXlaShape(input)},
-                         lower_for_shape_fn);
-                   },
-                   std::move(lower_fn));
-}
-
 torch::lazy::NodePtr ComparisonOp(c10::Symbol kind,
                                   const torch::lazy::Value& input,
                                   const torch::lazy::Value& other) {

torch_xla/csrc/ops/adaptive_avg_pool3d.h

@@ -156,9 +156,6 @@ torch::lazy::NodePtr MatMul(const torch::lazy::Value& lhs,
 torch::lazy::NodePtr AdaptiveMaxPool2dBackward(
     const torch::lazy::Value& grad_output, const torch::lazy::Value& input);
 
-torch::lazy::NodePtr AdaptiveAvgPool3dBackward(
-    const torch::lazy::Value& grad_output, const torch::lazy::Value& input);
-
 torch::lazy::NodePtr ComparisonOp(c10::Symbol kind,
                                   const torch::lazy::Value& input,
                                   const torch::lazy::Value& other);

torch_xla/csrc/ops/ops.cpp

@@ -39,6 +39,20 @@ torch_xla::XlaOpVector AdaptiveAvgPool2dBackward::Lower(
                   loctx);
 }
 
+torch_xla::XlaOpVector AdaptiveAvgPool3d::Lower(LoweringContext* loctx) const {
+  xla::XlaOp input = loctx->GetOutputOp(operand(0));
+  return ReturnOp(BuildAdaptiveAvgPool3d(input, output_size), loctx);
+}
+
+torch_xla::XlaOpVector AdaptiveAvgPool3dBackward::Lower(
+    LoweringContext* loctx) const {
+  xla::XlaOp grad_output = loctx->GetOutputOp(operand(0));
+  xla::XlaOp input = loctx->GetOutputOp(operand(1));
+  xla::XlaOp xla_output = BuildAdaptiveAvgPool3dBackward(
+      /*out_backprop=*/grad_output, /*input=*/input);
+  return ReturnOp(xla_output, loctx);
+}
+
 torch_xla::XlaOpVector Asin::Lower(LoweringContext* loctx) const {
   xla::XlaOp xla_input = loctx->GetOutputOp(operand(0));
   return ReturnOp(xla::Asin(xla_input), loctx);

torch_xla/csrc/ops/ops.h

@@ -58,6 +58,28 @@ xla::Shape AdaptiveAvgPool2dBackwardOutputShape(
                           lower_for_shape_fn);
 }
 
+xla::Shape AdaptiveAvgPool3dOutputShape(const torch::lazy::Value& input,
+                                        absl::Span<const int64_t> output_size) {
+  auto lower_for_shape_fn =
+      [output_size](absl::Span<const xla::XlaOp> operands) -> xla::XlaOp {
+    XLA_CHECK_EQ(operands.size(), 1);
+    return BuildAdaptiveAvgPool3d(operands[0], output_size);
+  };
+  return InferOutputShape({GetXlaShape(input)}, lower_for_shape_fn);
+}
+
+xla::Shape AdaptiveAvgPool3dBackwardOutputShape(
+    const torch::lazy::Value& grad_output, const torch::lazy::Value& input) {
+  auto lower_for_shape_fn =
+      [](absl::Span<const xla::XlaOp> operands) -> xla::XlaOp {
+    XLA_CHECK_EQ(operands.size(), 2);
+    return BuildAdaptiveAvgPool3dBackward(/*out_backprop=*/operands[0],
+                                          /*input=*/operands[1]);
+  };
+  return InferOutputShape({GetXlaShape(grad_output), GetXlaShape(input)},
+                          lower_for_shape_fn);
+}
+
 xla::Shape AsinOutputShape(const torch::lazy::Value& input) {
   return GetXlaShape(input);
 }

torch_xla/csrc/ops/ops_lower_fn.cpp

@@ -15,6 +15,12 @@ xla::Shape AdaptiveAvgPool2dOutputShape(const torch::lazy::Value& input,
 xla::Shape AdaptiveAvgPool2dBackwardOutputShape(
     const torch::lazy::Value& grad_output, const torch::lazy::Value& input);
 
+xla::Shape AdaptiveAvgPool3dOutputShape(const torch::lazy::Value& input,
+                                        absl::Span<const int64_t> output_size);
+
+xla::Shape AdaptiveAvgPool3dBackwardOutputShape(
+    const torch::lazy::Value& grad_output, const torch::lazy::Value& input);
+
 xla::Shape AsinOutputShape(const torch::lazy::Value& input);
 
 xla::Shape AsinhOutputShape(const torch::lazy::Value& input);

torch_xla/csrc/ops/ops_xla_shape_fn.cpp

@@ -328,12 +328,6 @@ class XLATensor : public c10::intrusive_ptr_target {
   static XLATensorPtr adaptive_max_pool2d_backward(
       const XLATensorPtr& grad_output, const XLATensorPtr& input);
 
-  static XLATensorPtr adaptive_avg_pool3d(const XLATensorPtr& input,
-                                          std::vector<int64_t> output_size);
-
-  static XLATensorPtr adaptive_avg_pool3d_backward(
-      const XLATensorPtr& grad_output, const XLATensorPtr& input);
-
   static XLATensorPtr _adaptive_avg_pool2d(
       const XLATensorPtr& input, std::vector<int64_t> output_size,
       std::vector<torch::lazy::Shape>&& shapes);

torch_xla/csrc/ops/ops_xla_shape_fn.h

@@ -21,7 +21,6 @@
 #include "torch_xla/csrc/layout_manager.h"
 #include "torch_xla/csrc/lowering_context.h"
 #include "torch_xla/csrc/ops/adam_optimizer_step.h"
-#include "torch_xla/csrc/ops/adaptive_avg_pool3d.h"
 #include "torch_xla/csrc/ops/adaptive_max_pool2d.h"
 #include "torch_xla/csrc/ops/all.h"
 #include "torch_xla/csrc/ops/all_gather.h"
@@ -591,18 +590,6 @@ XLATensorPtr XLATensor::adaptive_max_pool2d_backward(
                                                      input->GetIrValue()));
 }
 
-XLATensorPtr XLATensor::adaptive_avg_pool3d(const XLATensorPtr& input,
-                                            std::vector<int64_t> output_size) {
-  return input->CreateFrom(torch::lazy::MakeNode<AdaptiveAvgPool3d>(
-      input->GetIrValue(), std::move(output_size)));
-}
-
-XLATensorPtr XLATensor::adaptive_avg_pool3d_backward(
-    const XLATensorPtr& grad_output, const XLATensorPtr& input) {
-  return input->CreateFrom(AdaptiveAvgPool3dBackward(grad_output->GetIrValue(),
-                                                     input->GetIrValue()));
-}
-
 XLATensorPtr XLATensor::_adaptive_avg_pool2d(
     const XLATensorPtr& input, std::vector<int64_t> output_size,
     std::vector<torch::lazy::Shape>&& shapes) {

torch_xla/csrc/tensor.h

@@ -53,6 +53,8 @@ full_codegen:
 ir_gen:
   - _adaptive_avg_pool2d
   - _adaptive_avg_pool2d_backward
+  - _adaptive_avg_pool3d
+  - _adaptive_avg_pool3d_backward
 supported:
   - __ilshift__.Scalar
   - __ilshift__.Tensor