Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

【NPU】Support npu op elementwise_div and elementwise_div_grad #31573

Merged
merged 6 commits into from
Mar 15, 2021
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
140 changes: 140 additions & 0 deletions paddle/fluid/operators/elementwise/elementwise_div_op_npu.cc
Original file line number Diff line number Diff line change
@@ -0,0 +1,140 @@
/* 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. */

#ifdef PADDLE_WITH_ASCEND_CL
#include <memory>
#include <string>

#include "paddle/fluid/operators/elementwise/elementwise_div_op.h"
#include "paddle/fluid/operators/npu_op_runner.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;

template <typename DeviceContext, typename T>
class ElementwiseDivNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");

auto* out = ctx.Output<Tensor>("Out");

auto place = ctx.GetPlace();

out->mutable_data<T>(place);

auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();

auto runner = NpuOpRunner("Div", {*x, *y}, {*out}, {});
runner.Run(stream);
}
};

template <typename DeviceContext, typename T>
class ElementwiseDivGradNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* out = ctx.Input<Tensor>("Out");
auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");

auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));

auto place = ctx.GetPlace();

auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();

Tensor y_power(y->type());
y_power.mutable_data<T>(y->dims(), place);
auto y_power_runner = NpuOpRunner("Power", {*y},
{y_power}, {{"power", static_cast<float>(-1)}});
y_power_runner.Run(stream);

if (dx) {
dx->mutable_data<T>(place);

Tensor tensor_zeros(x->type());
tensor_zeros.mutable_data<T>(x->dims(), place);
auto tensor_zeros_runner = NpuOpRunner("ZerosLike", {*x},
{tensor_zeros}, {});
tensor_zeros_runner.Run(stream);

Tensor x_zero(paddle::framework::proto::VarType::BOOL);
x_zero.mutable_data<bool>(x->dims(), place);
auto x_zero_runner = NpuOpRunner("Equal", {*x, tensor_zeros},
{x_zero}, {});
x_zero_runner.Run(stream);

Tensor x_nozero(paddle::framework::proto::VarType::BOOL);
x_nozero.mutable_data<bool>(x->dims(), place);
auto x_nozero_runner = NpuOpRunner("LogicalNot", {x_zero},
{x_nozero}, {});
x_nozero_runner.Run(stream);

Tensor x_nozero_f(x->type());
x_nozero_f.mutable_data<T>(x->dims(), place);
auto x_nozero_f_runner = NpuOpRunner("Cast", {x_nozero},
{x_nozero_f}, {{"dst_type", static_cast<int32_t>(0)}});
x_nozero_f_runner.Run(stream);

Tensor x_grad_w(x->type());
x_grad_w.mutable_data<T>(x->dims(), place);
auto x_grad_w_runner = NpuOpRunner("Mul", {x_nozero_f, y_power},
{x_grad_w}, {});
x_grad_w_runner.Run(stream);

auto x_grad_runner = NpuOpRunner("Mul", {x_grad_w, *dout}, {*dx}, {});
x_grad_runner.Run(stream);
}

if (dy) {
dy->mutable_data<T>(place);

Tensor y_grad_w(x->type());
y_grad_w.mutable_data<T>(y->dims(), place);
auto y_grad_w_runner = NpuOpRunner("Mul", {*out, y_power},
{y_grad_w}, {});
y_grad_w_runner.Run(stream);

auto y_grad_runner = NpuOpRunner("Mul", {y_grad_w, *dout}, {*dy}, {});
y_grad_runner.Run(stream);
}
}
};

} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;

REGISTER_OP_NPU_KERNEL(
elementwise_div,
ops::ElementwiseDivNPUKernel<paddle::platform::NPUDeviceContext, float>,
ops::ElementwiseDivNPUKernel<paddle::platform::NPUDeviceContext,
paddle::platform::float16>);

REGISTER_OP_NPU_KERNEL(
elementwise_div_grad,
ops::ElementwiseDivGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
ops::ElementwiseDivGradNPUKernel<paddle::platform::NPUDeviceContext,
paddle::platform::float16>);
#endif
171 changes: 171 additions & 0 deletions python/paddle/fluid/tests/unittests/npu/test_elementwise_div_op_npu.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,171 @@
# 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 TestElementwiseDiv(OpTest):
def setUp(self):
self.set_npu()
self.op_type = "elementwise_div"
self.place = paddle.NPUPlace(0)

self.init_dtype()
np.random.seed(SEED)
x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
y = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
out = np.divide(x, y)

self.inputs = {
'X': OpTest.np_dtype_to_fluid_dtype(x),
'Y': OpTest.np_dtype_to_fluid_dtype(y)
}
self.attrs = {}
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): Div 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 TestElementwiseDivFp16(OpTest):
def setUp(self):
self.set_npu()
self.op_type = "elementwise_div"
self.place = paddle.NPUPlace(0)

self.init_dtype()
np.random.seed(SEED)
x = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
y = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
out = np.divide(x, y)

self.inputs = {
'X': OpTest.np_dtype_to_fluid_dtype(x),
'Y': OpTest.np_dtype_to_fluid_dtype(y)
}
self.attrs = {}
self.outputs = {'Out': out}

def set_npu(self):
self.__class__.use_npu = True
self.__class__.no_need_check_grad = True

def init_dtype(self):
self.dtype = np.float16

def test_check_output(self):
self.check_output_with_place(self.place, check_dygraph=False, atol=1e-5)


@unittest.skipIf(not paddle.is_compiled_with_npu(),
"core is not compiled with NPU")
class TestElementwiseDivNet(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.uniform(1, 2, [32, 32]).astype('float32')
b_np = np.random.uniform(1, 2, [32, 32]).astype('float32')
c_np = np.random.uniform(1, 2, [32, 32]).astype('float32')
d_np = np.random.uniform(1, 2, [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')
c = paddle.static.data(name="c", shape=[32, 32], dtype='float32')
d = paddle.static.data(name="d", shape=[32, 32], dtype='float32')
label = paddle.static.data(
name="label", shape=[32, 1], dtype='int64')

e = paddle.multiply(a, b)
f = paddle.multiply(c, d)
f.stop_gradient = True
g = paddle.divide(e, f)

fc_1 = fluid.layers.fc(input=g, 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)

print("Start run on {}".format(place))
for epoch in range(100):

pred_res, loss_res = exe.run(main_prog,
feed={
"a": a_np,
"b": b_np,
"c": c_np,
"d": d_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()