m53297601
/
mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

new add softplus and softplus grad gpu ops.

Browse Source
pull/7950/head
linqingke 4 years ago
parent f0e99e1099
commit 6dc3618758
8 changed files with 495 additions and 0 deletions
Show Stats Download Patch File Download Diff File

83
mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/softplus_impl.cu
Unescape View File

@ -0,0 +1,83 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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 "backend/kernel_compiler/gpu/cuda_impl/softplus_impl.cuh"
#include "runtime/device/gpu/cuda_common.h"
template <typename T>
__global__ void SoftplusKernel(const size_t size, const T *input_addr, T *output_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < size; pos += blockDim.x * gridDim.x) {
float x = input_addr[pos];
output_addr[pos] = logf(1. + exp(x));
}
}
template <>
__global__ void SoftplusKernel(const size_t size, const half *input_addr, half *output_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < size; pos += blockDim.x * gridDim.x) {
float x = __half2float(input_addr[pos]);
output_addr[pos] = __float2half(logf(1. + exp(x)));
}
}
template <typename T>
void Softplus(const size_t size, const T *input_addr, T *output_addr, cudaStream_t cuda_stream) {
SoftplusKernel<<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, input_addr, output_addr);
return;
}
template <>
void Softplus(const size_t size, const half *input_addr, half *output_addr, cudaStream_t cuda_stream) {
SoftplusKernel<half><<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, input_addr, output_addr);
return;
}
template <typename T>
__global__ void SoftplusGradKernel(const size_t size, const T *dy_addr, const T *x_addr, T *dx_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < (size); pos += blockDim.x * gridDim.x) {
T exp_x = exp(x_addr[pos]);
dx_addr[pos] = dy_addr[pos] * exp_x / (1. + exp_x);
}
}
template <typename T>
__global__ void SoftplusGradKernel(const size_t size, const half *dy_addr, const half *x_addr, half *dx_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < (size); pos += blockDim.x * gridDim.x) {
float x = __half2float(x_addr[pos]);
float dy = __half2float(dy_addr[pos]);
float exp_x = exp(x);
dx_addr[pos] = __float2half(dy * exp_x / (1. + exp_x));
}
}
template <typename T>
void SoftplusGrad(const size_t size, const T *dy_addr, const T *x_addr, T *dx_addr, cudaStream_t cuda_stream) {
SoftplusGradKernel<<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, dy_addr, x_addr, dx_addr);
return;
}
template <>
void SoftplusGrad(const size_t size, const half *dy_addr, const half *x_addr, half *dx_addr, cudaStream_t cuda_stream) {
SoftplusGradKernel<half><<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, dy_addr, x_addr, dx_addr);
return;
}
template void Softplus(const size_t size, const float *input_addr, float *output_addr, cudaStream_t cuda_stream);
template void Softplus(const size_t size, const half *input_addr, half *output_addr, cudaStream_t cuda_stream);
template void SoftplusGrad(const size_t size, const float *dy_addr, const float *x_addr, float *dx_addr,
cudaStream_t cuda_stream);
template void SoftplusGrad(const size_t size, const half *dy_addr, const half *x_addr, half *dx_addr,
cudaStream_t cuda_stream);

27
mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/softplus_impl.cuh
Unescape View File

@ -0,0 +1,27 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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.
*/
#ifndef MINDSPORE_CCSRC_KERNEL_GPU_CUDA_IMP_SOFTPLUS_H_
#define MINDSPORE_CCSRC_KERNEL_GPU_CUDA_IMP_SOFTPLUS_H_
#include "runtime/device/gpu/cuda_common.h"
template<typename T>
void Softplus(const size_t input_size, const T* input_addr, T* output_addr, cudaStream_t cuda_stream);
template<typename T>
void SoftplusGrad(const size_t size, const T* dy_addr, const T* x_addr, T* dx_addr, cudaStream_t cuda_stream);
#endif // MINDSPORE_CCSRC_KERNEL_GPU_CUDA_IMP_SOFTPLUS_H_

26
mindspore/ccsrc/backend/kernel_compiler/gpu/nn/softplus_gpu_kernel.cc
Unescape View File

@ -0,0 +1,26 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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 "backend/kernel_compiler/gpu/nn/softplus_gpu_kernel.h"
namespace mindspore {
namespace kernel {
MS_REG_GPU_KERNEL_ONE(Softplus, KernelAttr().AddInputAttr(kNumberTypeFloat32).AddOutputAttr(kNumberTypeFloat32),
SoftplusGpuKernel, float)
MS_REG_GPU_KERNEL_ONE(Softplus, KernelAttr().AddInputAttr(kNumberTypeFloat16).AddOutputAttr(kNumberTypeFloat16),
SoftplusGpuKernel, half)
} // namespace kernel
} // namespace mindspore

72
mindspore/ccsrc/backend/kernel_compiler/gpu/nn/softplus_gpu_kernel.h
Unescape View File

@ -0,0 +1,72 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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.
*/
#ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_SOFTPLUS_GPU_KERNEL_H_
#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_SOFTPLUS_GPU_KERNEL_H_
#include <vector>
#include "backend/kernel_compiler/gpu/gpu_kernel.h"
#include "backend/kernel_compiler/gpu/gpu_kernel_factory.h"
#include "backend/kernel_compiler/gpu/kernel_constants.h"
#include "backend/kernel_compiler/gpu/cuda_impl/softplus_impl.cuh"
namespace mindspore {
namespace kernel {
template <typename T>
class SoftplusGpuKernel : public GpuKernel {
public:
SoftplusGpuKernel() : input_size_(0) {}
~SoftplusGpuKernel() override = default;
const std::vector<size_t> &GetInputSizeList() const override { return input_size_list_; }
const std::vector<size_t> &GetOutputSizeList() const override { return output_size_list_; }
const std::vector<size_t> &GetWorkspaceSizeList() const override { return workspace_size_list_; }
bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
const std::vector<AddressPtr> &outputs, void *stream_ptr) override {
T *input_addr = GetDeviceAddress<T>(inputs, 0);
T *output_addr = GetDeviceAddress<T>(outputs, 0);
Softplus(input_size_ / sizeof(T), input_addr, output_addr, reinterpret_cast<cudaStream_t>(stream_ptr));
return true;
}
bool Init(const CNodePtr &kernel_node) override {
InitResource();
input_size_ = sizeof(T);
auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
for (auto dim : input_shape) {
input_size_ *= dim;
}
InitSizeLists();
return true;
}
protected:
void InitSizeLists() override {
input_size_list_.push_back(input_size_);
output_size_list_.push_back(input_size_);
}
private:
std::vector<size_t> input_size_list_;
std::vector<size_t> output_size_list_;
std::vector<size_t> workspace_size_list_;
size_t input_size_;
};
} // namespace kernel
} // namespace mindspore
#endif // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_SOFTPLUS_GPU_KERNEL_H_

30
mindspore/ccsrc/backend/kernel_compiler/gpu/nn/softplus_grad_gpu_kernel.cc
Unescape View File

@ -0,0 +1,30 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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 "backend/kernel_compiler/gpu/nn/softplus_grad_gpu_kernel.h"
namespace mindspore {
namespace kernel {
MS_REG_GPU_KERNEL_ONE(
SoftplusGrad,
KernelAttr().AddInputAttr(kNumberTypeFloat32).AddInputAttr(kNumberTypeFloat32).AddOutputAttr(kNumberTypeFloat32),
SoftplusGpuGradKernel, float)
MS_REG_GPU_KERNEL_ONE(
SoftplusGrad,
KernelAttr().AddInputAttr(kNumberTypeFloat16).AddInputAttr(kNumberTypeFloat16).AddOutputAttr(kNumberTypeFloat16),
SoftplusGpuGradKernel, half)
} // namespace kernel
} // namespace mindspore

74
mindspore/ccsrc/backend/kernel_compiler/gpu/nn/softplus_grad_gpu_kernel.h
Unescape View File

@ -0,0 +1,74 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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.
*/
#ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_SOFTPLUS_GRAD_KERNEL_H_
#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_SOFTPLUS_GRAD_KERNEL_H_
#include <vector>
#include "backend/kernel_compiler/gpu/gpu_kernel.h"
#include "backend/kernel_compiler/gpu/gpu_kernel_factory.h"
#include "backend/kernel_compiler/gpu/kernel_constants.h"
#include "backend/kernel_compiler/gpu/cuda_impl/softplus_impl.cuh"
namespace mindspore {
namespace kernel {
template <typename T>
class SoftplusGpuGradKernel : public GpuKernel {
public:
SoftplusGpuGradKernel() : input_size_(0) {}
~SoftplusGpuGradKernel() override = default;
const std::vector<size_t> &GetInputSizeList() const override { return input_size_list_; }
const std::vector<size_t> &GetOutputSizeList() const override { return output_size_list_; }
const std::vector<size_t> &GetWorkspaceSizeList() const override { return workspace_size_list_; }
bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
const std::vector<AddressPtr> &outputs, void *stream_ptr) override {
T *dy_addr = GetDeviceAddress<T>(inputs, 0);
T *x_addr = GetDeviceAddress<T>(inputs, 1);
T *dx_addr = GetDeviceAddress<T>(outputs, 0);
SoftplusGrad(input_size_ / sizeof(T), dy_addr, x_addr, dx_addr, reinterpret_cast<cudaStream_t>(stream_ptr));
return true;
}
bool Init(const CNodePtr &kernel_node) override {
InitResource();
input_size_ = sizeof(T);
auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
for (auto dim : input_shape) {
input_size_ *= dim;
}
InitSizeLists();
return true;
}
protected:
void InitSizeLists() override {
input_size_list_.push_back(input_size_);
input_size_list_.push_back(input_size_);
output_size_list_.push_back(input_size_);
}
private:
std::vector<size_t> input_size_list_;
std::vector<size_t> output_size_list_;
std::vector<size_t> workspace_size_list_;
size_t input_size_;
};
} // namespace kernel
} // namespace mindspore
#endif // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_SOFTPLUS_GRAD_KERNEL_H_

77
tests/st/ops/gpu/test_softplus_grad_op.py
Unescape View File

@ -0,0 +1,77 @@
# Copyright 2020 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
import numpy as np
import pytest
import mindspore.context as context
import mindspore.nn as nn
from mindspore import Tensor
from mindspore.ops import composite as C
from mindspore.ops import operations as P
context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
class SoftplusNet(nn.Cell):
def __init__(self):
super(SoftplusNet, self).__init__()
self.softplus = P.Softplus()
def construct(self, x):
return self.softplus(x)
class Grad(nn.Cell):
def __init__(self, network):
super(Grad, self).__init__()
self.grad = C.GradOperation(get_all=True, sens_param=True)
self.network = network
def construct(self, input_data, sens):
gout = self.grad(self.network)(input_data, sens)
return gout
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_softplusgrad():
x = np.array([0.58401114, 0.68800163, 0.9760397, 0.14702141, 0.46563736, 0.9607501,
0.14567593, 0.12261796, 0.37054458, 0.46421242]).astype(np.float32)
dy = np.array([0.5559598, 0.96994054, 0.24770357, 0.34646875, 0.2984393, 0.03287048,
0.55681044, 0.966908, 0.06015943, 0.6099489]).astype(np.float32)
x_ms = Tensor(x)
dy_ms = Tensor(dy)
net = SoftplusNet()
grad = Grad(net)
output = grad(x_ms, dy_ms)
expect = dy * np.exp(x) / (1 + np.exp(x))
assert np.allclose(output[0].asnumpy(), expect, rtol=1e-3)
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_softplusgrad_fp16():
np.random.seed(42)
x_np = np.random.randn(5, 3, 6).astype(np.float16)
dy_np = np.random.randn(5, 3, 6).astype(np.float16)
net = SoftplusNet()
grad = Grad(net)
output = grad(Tensor(x_np), Tensor(dy_np))
expect = dy_np * np.exp(x_np) / (1 + np.exp(x_np))
assert np.allclose(output[0].asnumpy(), expect, rtol=1e-2)

106
tests/st/ops/gpu/test_softplus_op.py
Unescape View File

@ -0,0 +1,106 @@
# Copyright 2020 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
import numpy as np
import pytest
import mindspore.context as context
import mindspore.nn as nn
from mindspore import Tensor
from mindspore.ops import operations as P
context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
class SoftplusNet(nn.Cell):
def __init__(self):
super(SoftplusNet, self).__init__()
self.softplus = P.Softplus()
def construct(self, x):
return self.softplus(x)
def SoftplusCompute(x):
return np.log(1 + np.exp(x))
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_softplus_1d():
x_np = np.random.random((50,)).astype(np.float32)
y_np = SoftplusCompute(x_np)
x_ms = Tensor(x_np)
net = SoftplusNet()
y_ms = net(x_ms)
assert np.allclose(y_np, y_ms.asnumpy())
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_softplus_2d():
x_np = np.random.random((50, 40)).astype(np.float32)
y_np = SoftplusCompute(x_np)
x_ms = Tensor(x_np)
net = SoftplusNet()
y_ms = net(x_ms)
assert np.allclose(y_np, y_ms.asnumpy())
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_softplus_4d():
x_np = np.random.random((32, 3, 224, 224)).astype(np.float32)
y_np = SoftplusCompute(x_np)
x_ms = Tensor(x_np)
net = SoftplusNet()
y_ms = net(x_ms)
assert np.allclose(y_np, y_ms.asnumpy())
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_softplus_neg():
x_np = np.random.random((32, 3, 224, 224)).astype(np.float32) * -1
y_np = SoftplusCompute(x_np)
x_ms = Tensor(x_np)
net = SoftplusNet()
y_ms = net(x_ms)
assert np.allclose(y_np, y_ms.asnumpy())
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_softplus_4d_fp16():
x_np = np.random.random((32, 3, 224, 224)).astype(np.float16)
y_np = SoftplusCompute(x_np)
x_ms = Tensor(x_np)
net = SoftplusNet()
y_ms = net(x_ms)
assert np.allclose(y_np, y_ms.asnumpy(), rtol=5e-3)
Write Preview
Loading…
Cancel
Save