Codegen addcdiv and addcmul (#3768)

* Codegen addcdiv and addcmul * pin * Use promoteAdd/Div/Mul * remove comment * Convert scalar to the right type * Delete .torch_pin
pytorch · Aug 16, 2022 · 6bfcd24 · 6bfcd24
1 parent 6639bcc
commit 6bfcd24
Show file tree

Hide file tree

Showing 9 changed files with 75 additions and 64 deletions.
diff --git a/scripts/gen_lazy_tensor.py b/scripts/gen_lazy_tensor.py
@@ -50,7 +50,7 @@ def node_base_ctor_call(self, schema: LazyIrSchema) -> str:
     base_ctor_value_args = ", ".join(base_ctor_value_args_list)
 
     shape_fn_inputs_list = [
-        f"{a.name}" for a in schema.positional_args
+        f"{a.name}" for a in (schema.positional_args + schema.keyword_args)
         if (a.is_lazy_value or isinstance(a.lazy_type, VectorCType) or
             is_boolean_dtype(a.lazy_type) or a.name == 'reduction')
     ]

diff --git a/torch_xla/csrc/aten_xla_type.cpp b/torch_xla/csrc/aten_xla_type.cpp
@@ -612,37 +612,6 @@ at::Tensor XLANativeFunctions::add(const at::Tensor& self,
                     });
 }
 
-at::Tensor XLANativeFunctions::addcdiv(const at::Tensor& self,
-                                       const at::Tensor& tensor1,
-                                       const at::Tensor& tensor2,
-                                       const at::Scalar& value) {
-  XLA_FN_COUNTER("xla::");
-  return bridge::AtenFromXlaTensor(XLATensor::addcdiv(
-      bridge::GetXlaTensor(self), value, bridge::GetXlaTensor(tensor1),
-      bridge::GetXlaTensor(tensor2)));
-}
-
-at::Tensor& XLANativeFunctions::addcdiv_(at::Tensor& self,
-                                         const at::Tensor& tensor1,
-                                         const at::Tensor& tensor2,
-                                         const at::Scalar& value) {
-  XLA_FN_COUNTER("xla::");
-  XLATensorPtr self_tensor = bridge::GetXlaTensor(self);
-  XLATensor::addcdiv_(self_tensor, value, bridge::GetXlaTensor(tensor1),
-                      bridge::GetXlaTensor(tensor2));
-  return self;
-}
-
-at::Tensor XLANativeFunctions::addcmul(const at::Tensor& self,
-                                       const at::Tensor& tensor1,
-                                       const at::Tensor& tensor2,
-                                       const at::Scalar& value) {
-  XLA_FN_COUNTER("xla::");
-  return bridge::AtenFromXlaTensor(XLATensor::addcmul(
-      bridge::GetXlaTensor(self), value, bridge::GetXlaTensor(tensor1),
-      bridge::GetXlaTensor(tensor2)));
-}
-
 at::Tensor XLANativeFunctions::addmm(const at::Tensor& self,
                                      const at::Tensor& mat1,
                                      const at::Tensor& mat2,

diff --git a/torch_xla/csrc/ops/ops_lower_fn.cpp b/torch_xla/csrc/ops/ops_lower_fn.cpp
@@ -7,6 +7,7 @@
 #include "torch_xla/csrc/matrix.h"
 #include "torch_xla/csrc/pooling.h"
 #include "torch_xla/csrc/reduction.h"
+#include "torch_xla/csrc/xla_lower_util.h"
 
 namespace torch_xla {
 torch_xla::XlaOpVector Abs::Lower(LoweringContext* loctx) const {
@@ -69,6 +70,22 @@ torch_xla::XlaOpVector Amin::Lower(LoweringContext* loctx) const {
   return ReturnOp(BuildMinInDims(input, dim, keepdim), loctx);
 }
 
+torch_xla::XlaOpVector Addcdiv::Lower(LoweringContext* loctx) const {
+  xla::XlaOp xla_input = loctx->GetOutputOp(operand(0));
+  xla::XlaOp xla_t1 = loctx->GetOutputOp(operand(1));
+  xla::XlaOp xla_t2 = loctx->GetOutputOp(operand(2));
+  xla::XlaOp xla_val = loctx->GetOutputOp(operand(3));
+  return ReturnOp(BuildAddcdiv(xla_input, xla_t1, xla_t2, xla_val), loctx);
+}
+
+torch_xla::XlaOpVector Addcmul::Lower(LoweringContext* loctx) const {
+  xla::XlaOp xla_input = loctx->GetOutputOp(operand(0));
+  xla::XlaOp xla_t1 = loctx->GetOutputOp(operand(1));
+  xla::XlaOp xla_t2 = loctx->GetOutputOp(operand(2));
+  xla::XlaOp xla_val = loctx->GetOutputOp(operand(3));
+  return ReturnOp(BuildAddcmul(xla_input, xla_t1, xla_t2, xla_val), loctx);
+}
+
 torch_xla::XlaOpVector Asin::Lower(LoweringContext* loctx) const {
   xla::XlaOp xla_input = loctx->GetOutputOp(operand(0));
   return ReturnOp(xla::Asin(xla_input), loctx);

diff --git a/torch_xla/csrc/ops/ops_xla_shape_fn.cpp b/torch_xla/csrc/ops/ops_xla_shape_fn.cpp
@@ -7,6 +7,7 @@
 #include "torch_xla/csrc/helpers.h"
 #include "torch_xla/csrc/pooling.h"
 #include "torch_xla/csrc/reduction.h"
+#include "torch_xla/csrc/xla_lower_util.h"
 
 namespace torch_xla {
 namespace {
@@ -109,6 +110,31 @@ xla::Shape AllOutputShape(const torch::lazy::Value& input) {
   return InferOutputShape({GetXlaShape(input)}, lower_for_shape_fn);
 }
 
+xla::Shape AddcdivOutputShape(const torch::lazy::Value& input,
+                              const torch::lazy::Value& t1,
+                              const torch::lazy::Value& t2,
+                              const torch::lazy::Value& value) {
+  auto shape_fn = [](absl::Span<const xla::XlaOp> operands) -> xla::XlaOp {
+    return BuildAddcdiv(operands[0], operands[1], operands[2], operands[3]);
+  };
+  return InferOutputShape({GetXlaShape(input), GetXlaShape(t1), GetXlaShape(t2),
+                           GetXlaShape(value)},
+                          shape_fn);
+}
+
+xla::Shape AddcmulOutputShape(const torch::lazy::Value& input,
+                              const torch::lazy::Value& t1,
+                              const torch::lazy::Value& t2,
+                              const torch::lazy::Value& value) {
+  auto shape_fn = [](absl::Span<const xla::XlaOp> operands) -> xla::XlaOp {
+    return BuildAddcmul(operands[0], operands[1], operands[2], operands[3]);
+  };
+
+  return InferOutputShape({GetXlaShape(input), GetXlaShape(t1), GetXlaShape(t2),
+                           GetXlaShape(value)},
+                          shape_fn);
+}
+
 xla::Shape AsinOutputShape(const torch::lazy::Value& input) {
   return GetXlaShape(input);
 }

diff --git a/torch_xla/csrc/ops/ops_xla_shape_fn.h b/torch_xla/csrc/ops/ops_xla_shape_fn.h
@@ -29,6 +29,16 @@ xla::Shape AminOutputShape(const torch::lazy::Value& input,
 
 xla::Shape AllOutputShape(const torch::lazy::Value& input);
 
+xla::Shape AddcdivOutputShape(const torch::lazy::Value& input,
+                              const torch::lazy::Value& t1,
+                              const torch::lazy::Value& t2,
+                              const torch::lazy::Value& value);
+
+xla::Shape AddcmulOutputShape(const torch::lazy::Value& input,
+                              const torch::lazy::Value& t1,
+                              const torch::lazy::Value& t2,
+                              const torch::lazy::Value& value);
+
 xla::Shape AsinOutputShape(const torch::lazy::Value& input);
 
 xla::Shape AsinhOutputShape(const torch::lazy::Value& input);

diff --git a/torch_xla/csrc/tensor_methods.cpp b/torch_xla/csrc/tensor_methods.cpp
@@ -660,35 +660,6 @@ XLATensorPtr XLATensor::add(
       logical_element_type);
 }
 
-XLATensorPtr XLATensor::addcdiv(const XLATensorPtr& input,
-                                const at::Scalar& value,
-                                const XLATensorPtr& tensor1,
-                                const XLATensorPtr& tensor2) {
-  torch::lazy::Value constant = GetIrValueForScalar(
-      value, tensor1->shape().get().element_type(), input->GetDevice());
-  torch::lazy::Value div = tensor1->GetIrValue() / tensor2->GetIrValue();
-  return input->CreateFrom(input->GetIrValue() + div * constant);
-}
-
-void XLATensor::addcdiv_(XLATensorPtr& input, const at::Scalar& value,
-                         const XLATensorPtr& tensor1,
-                         const XLATensorPtr& tensor2) {
-  torch::lazy::Value constant = GetIrValueForScalar(
-      value, tensor1->shape().get().element_type(), input->GetDevice());
-  torch::lazy::Value div = tensor1->GetIrValue() / tensor2->GetIrValue();
-  input->SetInPlaceIrValue(input->GetIrValue() + div * constant);
-}
-
-XLATensorPtr XLATensor::addcmul(const XLATensorPtr& input,
-                                const at::Scalar& value,
-                                const XLATensorPtr& tensor1,
-                                const XLATensorPtr& tensor2) {
-  torch::lazy::Value constant = GetIrValueForScalar(
-      value, tensor1->shape().get().element_type(), input->GetDevice());
-  torch::lazy::Value mul = tensor1->GetIrValue() * tensor2->GetIrValue();
-  return input->CreateFrom(input->GetIrValue() + mul * constant);
-}
-
 XLATensorPtr XLATensor::addmm(const XLATensorPtr& input,
                               const XLATensorPtr& weight,
                               const XLATensorPtr& bias) {

diff --git a/torch_xla/csrc/xla_lower_util.cpp b/torch_xla/csrc/xla_lower_util.cpp
@@ -1012,4 +1012,17 @@ xla::XlaOp BuildRoll(xla::XlaOp input, absl::Span<const int64_t> shifts,
   return need_flatten ? xla::Reshape(input, input_shape.dimensions()) : input;
 }
 
+xla::XlaOp BuildAddcdiv(xla::XlaOp input, xla::XlaOp t1, xla::XlaOp t2,
+                        xla::XlaOp val) {
+  return XlaHelpers::PromotedAdd(
+      input, XlaHelpers::PromotedMul(XlaHelpers::PromotedDiv(t1, t2), val));
+}
+
+xla::XlaOp BuildAddcmul(xla::XlaOp input, xla::XlaOp t1, xla::XlaOp t2,
+                        xla::XlaOp val) {
+  val = MaybeConvertTo(val, XlaHelpers::ShapeOfXlaOp(t1).element_type());
+  return XlaHelpers::PromotedAdd(
+      input, XlaHelpers::PromotedMul(XlaHelpers::PromotedMul(t1, t2), val));
+}
+
 }  // namespace torch_xla
diff --git a/torch_xla/csrc/xla_lower_util.h b/torch_xla/csrc/xla_lower_util.h
@@ -119,4 +119,10 @@ xla::XlaOp BuildXLogY(xla::XlaOp input, xla::XlaOp other);
 xla::XlaOp BuildRoll(xla::XlaOp input, absl::Span<const int64_t> shifts,
                      absl::Span<const int64_t> dims);
 
+xla::XlaOp BuildAddcdiv(xla::XlaOp input, xla::XlaOp t1, xla::XlaOp t2,
+                        xla::XlaOp val);
+
+xla::XlaOp BuildAddcmul(xla::XlaOp input, xla::XlaOp t1, xla::XlaOp t2,
+                        xla::XlaOp val);
+
 }  // namespace torch_xla
diff --git a/xla_native_functions.yaml b/xla_native_functions.yaml
@@ -7,6 +7,8 @@ full_codegen:
   - all
   - amax
   - amin
+  - addcdiv
+  - addcmul
   - asin
   - asinh
   - atan
@@ -95,9 +97,6 @@ supported:
   - adaptive_max_pool2d_backward
   - add.Scalar
   - add.Tensor
-  - addcdiv
-  - addcdiv_
-  - addcmul
   - addmm
   - alias
   - all.dim
-Original file line number
+Diff line change
@@ Expand Up / @@ -7,6 +7,8 @@ full_codegen: @@
       - all
       - amax
       - amin
+      - addcdiv
+      - addcmul
       - asin
       - asinh
       - atan
@@ Expand Down Expand Up / @@ -95,9 +97,6 @@ supported: @@
       - adaptive_max_pool2d_backward
       - add.Scalar
       - add.Tensor
-      - addcdiv
-      - addcdiv_
-      - addcmul
       - addmm
       - alias
       - all.dim
@@ Expand Down @@