mirror of https://github.com/alibaba/MNN.git
62 lines
1.7 KiB
C++
62 lines
1.7 KiB
C++
//
|
|
// VulkanRelu.hpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/01/31.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#ifndef VulkanRelu_hpp
|
|
#define VulkanRelu_hpp
|
|
|
|
#include <stdio.h>
|
|
|
|
#include "backend/vulkan/execution/VulkanBasicExecution.hpp"
|
|
|
|
namespace MNN {
|
|
|
|
class VulkanRelu : public VulkanBasicExecution {
|
|
public:
|
|
VulkanRelu(Backend* bn, float slope);
|
|
virtual ~VulkanRelu();
|
|
ErrorCode onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const VulkanCommandPool::Buffer* cmdBuffer) override;
|
|
|
|
private:
|
|
float mSlope;
|
|
std::shared_ptr<VulkanBuffer> mGpuReluParam;
|
|
const VulkanPipeline* mReluPipeline;
|
|
std::shared_ptr<VulkanPipeline::DescriptorSet> mDescriptorSet;
|
|
};
|
|
|
|
class VulkanRelu6 : public VulkanBasicExecution {
|
|
public:
|
|
VulkanRelu6(Backend* bn);
|
|
virtual ~VulkanRelu6();
|
|
ErrorCode onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const VulkanCommandPool::Buffer* cmdBuffer) override;
|
|
|
|
private:
|
|
std::shared_ptr<VulkanBuffer> mGpuRelu6Param;
|
|
const VulkanPipeline* mRelu6Pipeline;
|
|
std::shared_ptr<VulkanPipeline::DescriptorSet> mDescriptorSet;
|
|
};
|
|
|
|
class VulkanPrelu : public VulkanBasicExecution {
|
|
public:
|
|
VulkanPrelu(Backend* bn, const Op* op);
|
|
virtual ~VulkanPrelu();
|
|
ErrorCode onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const VulkanCommandPool::Buffer* cmdBuffer) override;
|
|
|
|
private:
|
|
std::shared_ptr<VulkanBuffer> mGpuPreluParam;
|
|
std::shared_ptr<VulkanImage> mSlope;
|
|
const VulkanPipeline* mPreluPipeline;
|
|
std::shared_ptr<VulkanPipeline::DescriptorSet> mDescriptorSet;
|
|
};
|
|
|
|
} // namespace MNN
|
|
|
|
#endif /* VulkanRelu_hpp */
|