[geometric]Add paddle.geometric.send_ue_recv API

paddle/fluid/operators/graph_send_ue_recv_op.cc

@@ -0,0 +1,150 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/infershape_utils.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/phi/core/infermeta_utils.h"
+#include "paddle/phi/infermeta/multiary.h"
+
+namespace paddle {
+namespace operators {
+
+class GraphSendUERecvOP : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"),
+        ctx.device_context());
+  }
+};
+
+class GraphSendUERecvGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    auto in_dims = ctx->GetInputDim("X");
+    ctx->SetOutputDim(framework::GradVarName("X"), in_dims);
+    auto y_dims = ctx->GetInputDim("Y");
+    ctx->SetOutputDim(framework::GradVarName("Y"), y_dims);
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
+                                       ctx, framework::GradVarName("Out")),
+                                   ctx.device_context());
+  }
+};
+
+class GraphSendUERecvOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "The input tensor with data type float32, float64, int32, int64.");
+    AddInput("Y",
+             "The input edge weight tensor, data type should be same with X");
+    AddInput("Src_index", "The source index tensor.");
+    AddInput("Dst_index", "The destination index tensor.");
+    AddInput("Out_size",
+             "(Tensor<int>, optional). The 0th dimension of the output."
+             "It has a higher priority than Attr(out_size).")
+        .AsDispensable();
+    AddOutput("Out", "Output tensor of graph_send_ue_recv op.");
+    AddOutput("Dst_count",
+              "Count tensor of Dst_index, mainly for MEAN pool_type.")
+        .AsIntermediate();
+    AddAttr<std::string>("compute_type",
+                         "(string, default 'ADD')"
+                         "Define differenct computation types between X and E.")
+        .SetDefault("ADD")
+        .InEnum({"ADD", "MUL"});
+    AddAttr<std::string>("pool_type",
+                         "(string, default 'SUM')"
+                         "Define different pool types to receive the result "
+                         "tensors of Dst_index.")
+        .SetDefault("SUM")
+        .InEnum({"SUM", "MEAN", "MIN", "MAX"});
+    AddAttr<std::vector<int64_t>>(
+        "out_size",
+        "(vector<int64_t>, default {0})"
+        "Define the first dimension of Output tensor."
+        "If set default {0}, then the shape of Out is the same with X.")
+        .SetDefault({0});
+    AddComment(R"DOC(
+Graph Learning Send_UE_Recv combine operator.
+
+$Out = Recv(Compute(Send(X, Src_index), Y, compute_type), Dst_index, pool_type)$
+
+This operator is mainly used in Graph Learning domain, and the main purpose is to reduce
+intermediate memory consumption in the process of message passing.
+
+Take `X` as the input tensor, we first use `src_index` to gather corresponding data.
+Then the gather data should compute with `Y` in different compute_types, like add, sub, mul, and div,
+and get the computation result. Then, use `dst_index` to update the corresponding position of output 
+tensor in different pooling types, like sum, mean, max, or min.
+
+)DOC");
+  }
+};
+
+template <typename T>
+class GraphSendUERecvGradOpMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("graph_send_ue_recv_grad");
+    op->SetInput("X", this->Input("X"));
+    op->SetInput("Y", this->Input("Y"));
+    op->SetInput("Src_index", this->Input("Src_index"));
+    op->SetInput("Dst_index", this->Input("Dst_index"));
+
+    if (PADDLE_GET_CONST(std::string, this->GetAttr("pool_type")) == "MEAN") {
+      op->SetInput("Dst_count", this->Output("Dst_count"));
+    }
+
+    if (PADDLE_GET_CONST(std::string, this->GetAttr("pool_type")) == "MIN" ||
+        PADDLE_GET_CONST(std::string, this->GetAttr("pool_type")) == "MAX") {
+      op->SetInput("Out", this->Output("Out"));
+    }
+
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    op->SetOutput(framework::GradVarName("Y"), this->InputGrad("Y"));
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+DECLARE_INFER_SHAPE_FUNCTOR(graph_send_ue_recv,
+                            GraphSendUERecvInferShapeFunctor,
+                            PD_INFER_META(phi::GraphSendUERecvInferMeta));
+REGISTER_OPERATOR(graph_send_ue_recv,
+                  ops::GraphSendUERecvOP,
+                  ops::GraphSendUERecvOpMaker,
+                  ops::GraphSendUERecvGradOpMaker<paddle::framework::OpDesc>,
+                  ops::GraphSendUERecvGradOpMaker<paddle::imperative::OpBase>,
+                  GraphSendUERecvInferShapeFunctor);
+REGISTER_OPERATOR(graph_send_ue_recv_grad, ops::GraphSendUERecvGradOp);

paddle/fluid/platform/device/gpu/gpu_primitives.h

@@ -419,6 +419,55 @@ CUDA_ATOMIC_WRAPPER(Max, double) {
   return __longlong_as_double(old);
 }
 
+#ifdef PADDLE_CUDA_FP16
+inline static __device__ uint32_t max_to_low_half(uint32_t val, float x) {
+  float16 low_half;
+  // The float16 in lower 16bits
+  low_half.x = static_cast<uint16_t>(val & 0xFFFFu);
+  low_half = static_cast<float16>(max(static_cast<float>(low_half), x));
+  return (val & 0xFFFF0000u) | low_half.x;
+}
+
+inline static __device__ uint32_t max_to_high_half(uint32_t val, float x) {
+  float16 high_half;
+  // The float16 in higher 16bits
+  high_half.x = static_cast<uint16_t>(val >> 16);
+  high_half = static_cast<float16>(max(static_cast<float>(high_half), x));
+  return (val & 0xFFFFu) | (static_cast<uint32_t>(high_half.x) << 16);
+}
+
+CUDA_ATOMIC_WRAPPER(Max, float16) {
+  if (*address >= val) {
+    return *address;
+  }
+  uint32_t *address_as_ui = reinterpret_cast<uint32_t *>(
+      reinterpret_cast<char *>(address) -
+      (reinterpret_cast<uintptr_t>(address) & 0x02));
+  float val_f = static_cast<float>(val);
+  uint32_t old = *address_as_ui;
+  uint32_t assumed;
+  if (((uintptr_t)address & 0x02) == 0) {
+    // The float16 value stay at lower 16 bits of the address.
+    do {
+      assumed = old;
+      old = atomicCAS(address_as_ui, assumed, max_to_low_half(assumed, val_f));
+    } while (old != assumed);
+    float16 ret;
+    ret.x = old & 0xFFFFu;
+    return ret;
+  } else {
+    // The float16 value stay at higher 16 bits of the address.
+    do {
+      assumed = old;
+      old = atomicCAS(address_as_ui, assumed, max_to_high_half(assumed, val_f));
+    } while (old != assumed);
+    float16 ret;
+    ret.x = old >> 16;
+    return ret;
+  }
+}
+#endif
+
 // For atomicMin
 USE_CUDA_ATOMIC(Min, int);
 USE_CUDA_ATOMIC(Min, unsigned int);
@@ -503,5 +552,54 @@ CUDA_ATOMIC_WRAPPER(Min, double) {
   return __longlong_as_double(old);
 }
 
+#ifdef PADDLE_CUDA_FP16
+inline static __device__ uint32_t min_to_low_half(uint32_t val, float x) {
+  float16 low_half;
+  // The float16 in lower 16bits
+  low_half.x = static_cast<uint16_t>(val & 0xFFFFu);
+  low_half = static_cast<float16>(min(static_cast<float>(low_half), x));
+  return (val & 0xFFFF0000u) | low_half.x;
+}
+
+inline static __device__ uint32_t min_to_high_half(uint32_t val, float x) {
+  float16 high_half;
+  // The float16 in higher 16bits
+  high_half.x = static_cast<uint16_t>(val >> 16);
+  high_half = static_cast<float16>(min(static_cast<float>(high_half), x));
+  return (val & 0xFFFFu) | (static_cast<uint32_t>(high_half.x) << 16);
+}
+
+CUDA_ATOMIC_WRAPPER(Min, float16) {
+  if (*address <= val) {
+    return *address;
+  }
+  uint32_t *address_as_ui = reinterpret_cast<uint32_t *>(
+      reinterpret_cast<char *>(address) -
+      (reinterpret_cast<uintptr_t>(address) & 0x02));
+  float val_f = static_cast<float>(val);
+  uint32_t old = *address_as_ui;
+  uint32_t assumed;
+  if (((uintptr_t)address & 0x02) == 0) {
+    // The float16 value stay at lower 16 bits of the address.
+    do {
+      assumed = old;
+      old = atomicCAS(address_as_ui, assumed, min_to_low_half(assumed, val_f));
+    } while (old != assumed);
+    float16 ret;
+    ret.x = old & 0xFFFFu;
+    return ret;
+  } else {
+    // The float16 value stay at higher 16 bits of the address.
+    do {
+      assumed = old;
+      old = atomicCAS(address_as_ui, assumed, min_to_high_half(assumed, val_f));
+    } while (old != assumed);
+    float16 ret;
+    ret.x = old >> 16;
+    return ret;
+  }
+}
+#endif
+
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/pybind/op_function_generator.h

@@ -226,6 +226,7 @@ std::map<std::string, std::set<std::string>> op_ins_map = {
       "Mean3",
       "Var3"}},
     {"graph_send_recv", {"X", "Src_index", "Dst_index", "Out_size"}},
+    {"graph_send_ue_recv", {"X", "Y", "Src_index", "Dst_index", "Out_size"}},
 };
 
 // NOTE(zhiqiu): Like op_ins_map.

paddle/phi/api/yaml/legacy_api.yaml

@@ -1070,6 +1070,17 @@
   intermediate : dst_count
   backward : graph_send_recv_grad
 
+- api : graph_send_ue_recv
+  args : (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, str compute_type, str pool_type, IntArray out_size)
+  output : Tensor(out), Tensor(dst_count)
+  infer_meta :
+    func : GraphSendUERecvInferMeta
+  kernel :
+    func : graph_send_ue_recv
+    data_type : x
+  intermediate : dst_count
+  backward : graph_send_ue_recv_grad
+
 - api : greater_equal
   args : (Tensor x, Tensor y, int axis = -1)
   output : Tensor

paddle/phi/api/yaml/legacy_backward.yaml

@@ -952,6 +952,18 @@
     data_type : out_grad
   optional: out, dst_count
 
+- backward_api : graph_send_ue_recv_grad
+  forward : graph_send_ue_recv (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, str compute_type, str pool_type, IntArray out_size) -> Tensor(out), Tensor(dst_count)
+  args : (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, Tensor out, Tensor dst_count, Tensor out_grad, str compute_type, str pool_type)
+  output : Tensor(x_grad), Tensor(y_grad)
+  infer_meta :
+    func : GeneralBinaryGradInferMeta
+    param : [x, y]
+  kernel :
+    func : graph_send_ue_recv_grad
+    data_type : out_grad
+  optional: out, dst_count
+
 # grid sample
 - backward_api : grid_sample_grad
   forward : grid_sample (Tensor x, Tensor grid, str mode, str padding_mode, bool align_corners) -> Tensor(out)