MNN/source/backend/vulkan/execution/VulkanPool.cpp

//
//  VulkanPool.cpp
//  MNN
//
//  Created by MNN on 2019/01/31.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#include "backend/vulkan/execution/VulkanPool.hpp"
#include "core/Macro.h"
namespace MNN {
struct ConstBuffer {
    ivec4 inputSize;
    ivec4 outputSize;
    ivec2 pad;
    ivec2 kernelSize;
    ivec2 stride;
};

VulkanPool::VulkanPool(const Op* op, Backend* bn) : VulkanBasicExecution(bn) {
    mCommon    = op->main_as_Pool();
    auto extra = static_cast<VulkanBackend*>(bn);

    std::vector<VkDescriptorType> types{VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
                                        VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER};
    switch (mCommon->type()) {
        case PoolType_MAXPOOL:
            mPoolPipeline =
                extra->getPipeline("glsl_maxpool_comp", /*glsl_maxpool_comp, glsl_maxpool_comp_len,*/ types);
            break;
        case PoolType_AVEPOOL:
            mPoolPipeline =
                extra->getPipeline("glsl_avgpool_comp", /*glsl_avgpool_comp, glsl_avgpool_comp_len,*/ types);
            break;
        default:
            break;
    }
    mConstBuffer.reset(new VulkanBuffer(extra->getMemoryPool(), false, sizeof(ConstBuffer), nullptr,
                                        VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT));
}
VulkanPool::~VulkanPool() {
}

ErrorCode VulkanPool::onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
                               const VulkanCommandPool::Buffer* cmdBuffer) {
    auto input  = inputs[0];
    auto output = outputs[0];
    int iw      = input->width();
    int ih      = input->height();
    int ow      = output->width();
    int oh      = output->height();
    int icDiv4  = UP_DIV(input->channel(), 4);
    int ocDiv4  = UP_DIV(output->channel(), 4);
    auto extra  = (VulkanBackend*)backend();
    // Set Const Buffer
    {
        auto pool = (ConstBuffer*)mConstBuffer->map();
        ::memset(pool, 0, sizeof(ConstBuffer));
        pool->inputSize[0]  = input->width();
        pool->inputSize[1]  = input->height();
        pool->inputSize[2]  = icDiv4 * input->batch();
        pool->outputSize[0] = ow;
        pool->outputSize[1] = oh;
        pool->outputSize[2] = ocDiv4 * output->batch();
        int padWidth     = mCommon->padX();
        int padHeight    = mCommon->padY();

        int strideWidth  = mCommon->strideX();
        int strideHeight = mCommon->strideY();

        // edit const if global
        int kernelWidth  = std::min(mCommon->kernelX(), iw);
        int kernelHeight = std::min(mCommon->kernelY(), ih);
        if (mCommon->isGlobal()) {
            kernelWidth  = iw;
            kernelHeight = ih;
            strideWidth  = iw;
            strideHeight = ih;
            padWidth     = 0;
            padHeight    = 0;
        }

        if(mCommon->padType() == PoolPadType_SAME){
            int padNeededWidth  = (output->width() - 1) * strideWidth + kernelWidth - input->width();
            int padNeededHeight = (output->height() - 1) * strideHeight + kernelHeight - input->height();
            padWidth            = padNeededWidth > 0 ? padNeededWidth / 2 : 0;
            padHeight           = padNeededHeight > 0 ? padNeededHeight / 2 : 0;
        } else if (mCommon->padType() == PoolPadType_VALID) {
            padWidth = padHeight = 0;
        }

        pool->pad[0]        = padWidth;
        pool->pad[1]        = padHeight;
        pool->stride[0]     = strideWidth;
        pool->stride[1]     = strideHeight;
        pool->kernelSize[0] = kernelWidth;
        pool->kernelSize[1] = kernelHeight;
        mConstBuffer->flush(true, 0, sizeof(ConstBuffer));
        mConstBuffer->unmap();
    }

    // Set Command Buffer
    {
        mDescriptorSet.reset(mPoolPipeline->createSet());
        mDescriptorSet->writeImage((VkImageView)output->deviceId(), extra->getCommonSampler()->get(),
                                   VK_IMAGE_LAYOUT_GENERAL, 0);
        mDescriptorSet->writeImage((VkImageView)input->deviceId(), extra->getCommonSampler()->get(),
                                   VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 1);
        mDescriptorSet->writeBuffer(mConstBuffer->buffer(), 2, mConstBuffer->size());
        mPoolPipeline->bind(cmdBuffer->get(), mDescriptorSet->get());
        vkCmdDispatch(cmdBuffer->get(), UP_DIV(ow, 8), UP_DIV(oh, 8), UP_DIV(ocDiv4 * output->batch(), 1));
    }
    return NO_ERROR;
}

class VulkanPoolCreator : public VulkanBackend::Creator {
public:
    virtual VulkanBasicExecution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs, const MNN::Op* op,
                                Backend* backend) const override {
        return new VulkanPool(op, backend);
    }
};

static bool gResistor = []() {
    VulkanBackend::addCreator(OpType_Pooling, new VulkanPoolCreator);
    return true;
}();

} // namespace MNN