[NPU] Support npu op pow and pow grad (#31247)

* [NPU] Support npu op: (1) pow (2) pow_grad

* Support fp16

paddle/fluid/memory/memcpy.cc

@@ -208,8 +208,16 @@ void Copy<platform::NPUPlace, platform::CPUPlace>(platform::NPUPlace dst_place,
   if (UNLIKELY(num == 0)) return;
 
   platform::SetNPUDeviceId(dst_place.device);
+
+  // NOTE(ascendrc): NPU memcpy async from host to device is a "real" async,
+  // which is different from CUDA. In Paddle, when async is called, "sync"
+  // is run actually, which means Paddle doesn't fully supported async.
+  // TODO(ascendrc): Support NPU memcpy async for better performance.
+  stream = nullptr;
+
   VLOG(4) << "memory::Copy " << num << " Bytes from " << src_place << " to "
           << dst_place << " by thream(" << stream << ")";
+
   if (stream) {
     platform::RecordEvent record_event("NpuMemcpyAsync:CPU->NPU");
     platform::NPUMemcpyAsync(dst, src, num, ACL_MEMCPY_HOST_TO_DEVICE, stream);
@@ -228,8 +236,16 @@ void Copy<platform::CPUPlace, platform::NPUPlace>(platform::CPUPlace dst_place,
   if (UNLIKELY(num == 0)) return;
 
   platform::SetNPUDeviceId(src_place.device);
+
+  // NOTE(ascendrc): NPU memcpy async from device to host is a "real" async,
+  // which is different from CUDA. In Paddle, when async is called, "sync"
+  // is run actually, which means Paddle doesn't fully supported async.
+  // TODO(ascendrc): Support NPU memcpy async for better performance.
+  stream = nullptr;
+
   VLOG(4) << "memory::Copy " << num << " Bytes from " << src_place << " to "
           << dst_place << " by thream(" << stream << ")";
+
   if (stream) {
     platform::RecordEvent record_event("NpuMemcpyAsync:NPU->CPU");
     platform::NPUMemcpyAsync(dst, src, num, ACL_MEMCPY_DEVICE_TO_HOST, stream);

paddle/fluid/operators/activation_op_npu.cc

@@ -0,0 +1,127 @@
+/* Copyright (c) 2021 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 Licnse. */
+
+#ifdef PADDLE_WITH_ASCEND_CL
+#include <memory>
+#include <string>
+
+#include "paddle/fluid/framework/ddim.h"
+#include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+class PowNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* out = ctx.Output<Tensor>("Out");
+    auto factor = ctx.Attr<float>("factor");
+
+    out->mutable_data<T>(ctx.GetPlace());
+
+    auto runner = NpuOpRunner("Power", {*x}, {*out},
+                              {{"power", factor},
+                               {"scale", static_cast<float>(1.0)},
+                               {"shift", static_cast<float>(0.0)}});
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class PowGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto factor = ctx.Attr<float>("factor");
+
+    auto x_dims = x->dims();
+
+    auto place = ctx.GetPlace();
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    // NOTE(liym27): dx = dout * factor * x.pow(factor-1)
+
+    // Step1: Compute x_pow = x.pow(factor-1)
+    Tensor x_pow(x->type());
+    x_pow.mutable_data<T>(x->dims(), place);
+    auto runner_pow = NpuOpRunner("Power", {*x}, {x_pow},
+                                  {{"power", factor - static_cast<float>(1)}});
+    runner_pow.Run(stream);
+
+    // Step 2: Construct a broadcast factor, which has the same shape with x.
+    // 2.1 Get the shape of x
+    Tensor x_shape(framework::proto::VarType::INT32);
+    x_shape.mutable_data<int32_t>({x_dims.size()}, place);
+    TensorFromVector(framework::vectorize<int32_t>(x_dims),
+                     ctx.device_context(), &x_shape);
+
+    // 2.2 Get a factor tensor with shape [1].
+    Tensor factor_tensor(framework::proto::VarType::FP32);
+    factor_tensor.mutable_data<float>({1}, place);
+    TensorFromVector(std::vector<float>{factor}, ctx.device_context(),
+                     &factor_tensor);
+
+    // 2.3 Get the factor which has the shape with x and the same value with
+    // factor.
+    Tensor factor_bc_tensor(framework::proto::VarType::FP32);
+    factor_bc_tensor.mutable_data<float>(x_dims, place);
+    auto runner_bc = NpuOpRunner("BroadcastTo", {factor_tensor, x_shape},
+                                 {factor_bc_tensor}, {});
+    runner_bc.Run(stream);
+
+    // Step 3: Compute x_power_mul_factor = factor * x.pow(factor-1)
+    Tensor x_power_mul_factor(x->type());
+    x_power_mul_factor.mutable_data<T>(x->dims(), place);
+    auto runner_mul_1 =
+        NpuOpRunner("Mul", {factor_bc_tensor, *x}, {x_power_mul_factor}, {});
+    runner_mul_1.Run(stream);
+
+    // Step 4: Compute dx = dout * factor * x.pow(factor-1)
+    dx->mutable_data<T>(place);
+    auto runner_mul_2 =
+        NpuOpRunner("Mul", {*dout, x_power_mul_factor}, {*dx}, {});
+    runner_mul_2.Run(stream);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_NPU_KERNEL(
+    pow, ops::PowNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::PowNPUKernel<paddle::platform::NPUDeviceContext,
+                      paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    pow_grad, ops::PowGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::PowGradNPUKernel<paddle::platform::NPUDeviceContext,
+                          paddle::platform::float16>);
+
+#endif

python/paddle/fluid/tests/unittests/npu/test_pow_op_npu.py

@@ -0,0 +1,152 @@
+#  Copyright (c) 2021 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.
+
+from __future__ import print_function
+
+import numpy as np
+import unittest
+import sys
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestPow(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "pow"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
+        out = np.power(x, 3)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {'factor': 3.0}
+        self.outputs = {'Out': out}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+    # TODO(ascendrc): Add grad test
+    # def test_check_grad(self):
+    #     if self.dtype == np.float16:
+    #         return
+    #     self.check_grad(['X'], 'Out')
+    #
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestPowFp16(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "pow"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
+        out = np.power(x, 3)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {'factor': 3.0}
+        self.outputs = {'Out': out}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSubtractNet(unittest.TestCase):
+    def _test(self, run_npu=True):
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = SEED
+        startup_prog.random_seed = SEED
+        np.random.seed(SEED)
+
+        a_np = np.random.random(size=(32, 32)).astype('float32')
+        b_np = np.random.random(size=(32, 32)).astype('float32')
+        label_np = np.random.randint(2, size=(32, 1)).astype('int64')
+
+        with paddle.static.program_guard(main_prog, startup_prog):
+            a = paddle.static.data(name="a", shape=[32, 32], dtype='float32')
+            b = paddle.static.data(name="b", shape=[32, 32], dtype='float32')
+            label = paddle.static.data(
+                name="label", shape=[32, 1], dtype='int64')
+
+            sum = paddle.add(a, b)
+            z = paddle.pow(sum, 2.0)
+
+            fc_1 = fluid.layers.fc(input=z, size=128)
+            prediction = fluid.layers.fc(input=fc_1, size=2, act='softmax')
+
+            cost = fluid.layers.cross_entropy(input=prediction, label=label)
+            loss = fluid.layers.reduce_mean(cost)
+            sgd = fluid.optimizer.SGD(learning_rate=0.01)
+            sgd.minimize(loss)
+
+        if run_npu:
+            place = paddle.NPUPlace(0)
+        else:
+            place = paddle.CPUPlace()
+
+        exe = paddle.static.Executor(place)
+        exe.run(startup_prog)
+
+        for epoch in range(100):
+
+            pred_res, loss_res = exe.run(
+                main_prog,
+                feed={"a": a_np,
+                      "b": b_np,
+                      "label": label_np},
+                fetch_list=[prediction, loss])
+            if epoch % 10 == 0:
+                print("Epoch {} | Prediction[0]: {}, Loss: {}".format(
+                    epoch, pred_res[0], loss_res))
+
+        return pred_res, loss_res
+
+    def test_npu(self):
+        cpu_pred, cpu_loss = self._test(False)
+        npu_pred, npu_loss = self._test(True)
+
+        self.assertTrue(np.allclose(npu_pred, cpu_pred))
+        self.assertTrue(np.allclose(npu_loss, cpu_loss))
+
+
+if __name__ == '__main__':
+    unittest.main()