From 78d5cf7b4ec4baedf2186ac2502d36036c784f36 Mon Sep 17 00:00:00 2001
From: Aganlengzi <aganlengzi@gmail.com>
Date: Tue, 17 May 2022 10:24:36 +0800
Subject: [PATCH] [NPU] add reduce_max_grad op (#42672)

---
 .../operators/reduce_ops/reduce_max_op_npu.cc | 67 +++++++++++++++++++
 1 file changed, 67 insertions(+)

diff --git a/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc b/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc
index f99b72faba..04660fb501 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc
+++ b/paddle/fluid/operators/reduce_ops/reduce_max_op_npu.cc
@@ -105,6 +105,68 @@ class ReduceMaxNPUKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class ReduceMaxGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* out = context.Input<Tensor>("Out");
+    auto* out_grad = context.Input<Tensor>(framework::GradVarName("Out"));
+    int in_dtype = context.Attr<int>("in_dtype");
+
+    PADDLE_ENFORCE_EQ(
+        in_dtype == -1, true,
+        platform::errors::InvalidArgument(
+            "NPU only support in_dtype == -1 in reduce_max_grad op."));
+
+    auto* x_grad = context.Output<Tensor>(framework::GradVarName("X"));
+    x_grad->mutable_data<T>(context.GetPlace());
+
+    auto& dev_ctx =
+        context.template device_context<paddle::platform::NPUDeviceContext>();
+    auto place = context.GetPlace();
+    auto stream = dev_ctx.stream();
+
+    // broadcast
+    auto x_dims_vec = phi::vectorize(x->dims());
+    Tensor transformed_out(x->type());
+    transformed_out.Resize(phi::make_ddim(x_dims_vec));
+    transformed_out.mutable_data<T>(place);
+    NpuOpRunner r_brd_out;
+    r_brd_out.SetType("BroadcastTo")
+        .AddInput(*out)
+        .AddInput(std::move(x_dims_vec))
+        .AddOutput(transformed_out)
+        .Run(stream);
+    Tensor transformed_out_grad(x->type());
+    transformed_out_grad.Resize(phi::make_ddim(x_dims_vec));
+    transformed_out_grad.mutable_data<T>(place);
+    NpuOpRunner r_brd_out_grad;
+    r_brd_out_grad.SetType("BroadcastTo")
+        .AddInput(*out_grad)
+        .AddInput(std::move(x_dims_vec))
+        .AddOutput(transformed_out_grad)
+        .Run(stream);
+
+    // compare
+    Tensor equal_cond;
+    equal_cond.mutable_data<bool>(x_grad->dims(), place);
+    const auto& r_equal =
+        NpuOpRunner("Equal", {*x, transformed_out}, {equal_cond}, {});
+    r_equal.Run(stream);
+
+    // select
+    Tensor t_zero;
+    t_zero.mutable_data<T>(x_grad->dims(), place);
+    FillNpuTensorWithConstant(&t_zero, static_cast<T>(0));
+    t_zero.Resize(x_grad->dims());
+
+    const auto& r_sel = NpuOpRunner(
+        "SelectV2", {equal_cond, transformed_out_grad, t_zero}, {*x_grad}, {});
+    r_sel.Run(stream);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -115,3 +177,8 @@ REGISTER_OP_NPU_KERNEL(
     ops::ReduceMaxNPUKernel<plat::NPUDeviceContext, plat::float16>,
     ops::ReduceMaxNPUKernel<plat::NPUDeviceContext, int64_t>,
     ops::ReduceMaxNPUKernel<plat::NPUDeviceContext, int>);
+REGISTER_OP_NPU_KERNEL(
+    reduce_max_grad, ops::ReduceMaxGradNPUKernel<plat::NPUDeviceContext, float>,
+    ops::ReduceMaxGradNPUKernel<plat::NPUDeviceContext, plat::float16>,
+    ops::ReduceMaxGradNPUKernel<plat::NPUDeviceContext, int64_t>,
+    ops::ReduceMaxGradNPUKernel<plat::NPUDeviceContext, int>);
-- 
GitLab