alibaba--mnn/source/backend/vulkan/image/execution/VulkanUnary.cpp

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

#include "VulkanUnary.hpp"
#include "core/Macro.h"
#include "core/TensorUtils.hpp"

namespace MNN {

struct Param {
    ivec4 size;
};

VulkanUnary::VulkanUnary(const std::string& midType, Backend* bn, bool image) : VulkanBasicExecution(bn) {
    auto vkbackend = static_cast<VulkanBackend*>(bn);
    auto prefix = "glsl_unaryImage_";
    auto types = {
        VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
        VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
        VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
    };
    std::string posfix = "_comp";
    // get pipeline
    mUnaryPipeline = vkbackend->getPipeline(prefix + midType + posfix, types);
}

VulkanUnary::~VulkanUnary() {
}

static std::string _getMidType(const Op* op) {
    std::string midType = "";
    if (op->type() == OpType_TanH) {
        midType = "TANH";
    } else if (op->type() == OpType_Sigmoid) {
        midType = "SIGMOID";
    } else {
        // unary op
        auto unaryType = op->main_as_UnaryOp()->opType();
#define SETTYPE(type, name) if (unaryType == type) {midType = name; break;}
        do {
            SETTYPE(UnaryOpOperation_SIGMOID, "SIGMOID");
            SETTYPE(UnaryOpOperation_TANH, "TANH");
            SETTYPE(UnaryOpOperation_RSQRT, "RSQRT");
            SETTYPE(UnaryOpOperation_SIGN, "SIGN");
            SETTYPE(UnaryOpOperation_ABS, "ABS");
            SETTYPE(UnaryOpOperation_NEG, "NEG");
            SETTYPE(UnaryOpOperation_EXP, "EXP");
            SETTYPE(UnaryOpOperation_SQRT, "SQRT");
            SETTYPE(UnaryOpOperation_SQUARE, "SQUARE");
            SETTYPE(UnaryOpOperation_LOG, "LOG");
            SETTYPE(UnaryOpOperation_TAN, "TAN");
            SETTYPE(UnaryOpOperation_COS, "COS");
            SETTYPE(UnaryOpOperation_SIN, "SIN");
            SETTYPE(UnaryOpOperation_CEIL, "CEIL");
            SETTYPE(UnaryOpOperation_FLOOR, "FLOOR");
            SETTYPE(UnaryOpOperation_EXPM1, "EXPM1");
            SETTYPE(UnaryOpOperation_RECIPROCAL, "RECIPROCAL");

            SETTYPE(UnaryOpOperation_SINH, "SINH");
            SETTYPE(UnaryOpOperation_ASIN, "ASIN");
            SETTYPE(UnaryOpOperation_ASINH, "ASINH");
            SETTYPE(UnaryOpOperation_COSH, "COSH");
            SETTYPE(UnaryOpOperation_ACOS, "ACOS");
            SETTYPE(UnaryOpOperation_ACOSH, "ACOSH");
            SETTYPE(UnaryOpOperation_ATAN, "ATAN");
            SETTYPE(UnaryOpOperation_ATANH, "ATANH");
            SETTYPE(UnaryOpOperation_LOG1P, "LOG1P");
            
            SETTYPE(UnaryOpOperation_ROUND, "ROUND");
            SETTYPE(UnaryOpOperation_HARDSWISH, "HARDSWISH");
            SETTYPE(UnaryOpOperation_GELU, "GELU");
            // Since SPIR-V lacks a built-in erf (gauss error function) instruction and the existing shader implementation of GELU is essentially an approximation of erf, there is no need to add a new implementation of GELU_STANDARD.
            SETTYPE(UnaryOpOperation_GELU_STANDARD, "GELU");
            SETTYPE(UnaryOpOperation_SILU, "SILU");
        } while(false);
#undef SETTYPE
    }
    return midType;
}

bool VulkanUnary::encode(const Tensor* input, const Tensor* output, const VulkanCommandPool::Buffer* cmdBuffer, const Tensor::InsideDescribe::Region* region) {
    return true;
}
bool VulkanUnary::encoderSingle(const VulkanCommandPool::Buffer* cmdBuffer, const VulkanImage* dest, const VulkanImage* source,
                                const std::array<int, 3>& size
                   ) {
    auto vkbackend = static_cast<VulkanBackend*>(backend());
    auto param = std::make_shared<VulkanBuffer>(vkbackend->getMemoryPool(), false, sizeof(Param), nullptr,
                                            VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT);
    auto paramOrigin = (Param*)param->map();
    paramOrigin->size[0] = size[0] * size[1] * size[2];
    paramOrigin->size[1] = size[2]; // depth
    paramOrigin->size[2] = size[1]; // height
    paramOrigin->size[3] = size[0]; // width
    param->unmap();
    auto totalSize = size[0] * size[1] * size[2];
    std::shared_ptr<VulkanLayout::DescriptorSet> des(mUnaryPipeline->createSet());
    des->writeImage(dest->view(), vkbackend->getCommonSampler()->get(), VK_IMAGE_LAYOUT_GENERAL, 0);
    des->writeImage(source->view(), vkbackend->getCommonSampler()->get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 1);
    des->writeBuffer(param->buffer(), 2, sizeof(Param), 0);
    mUnaryPipeline->bind(cmdBuffer->get(), des->get());

    source->barrierRead(cmdBuffer->get());
    dest->barrierWrite(cmdBuffer->get());
    vkCmdDispatch(cmdBuffer->get(), UP_DIV(totalSize, 256), 1, 1);
    mDesSet.emplace_back(des);
    mParams.emplace_back(param);
    return true;
}

ErrorCode VulkanUnary::onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
                                const VulkanCommandPool::Buffer* cmdBuffer) {
    // set param
    auto vkbackend = static_cast<VulkanBackend*>(backend());
    auto inputTensor = (VulkanTensor*)(inputs[0]->deviceId());
    auto outputTensor = (VulkanTensor*)(outputs[0]->deviceId());
    mDesSet.clear();
    mParams.clear();
    for (int n=0; n<inputTensor->imageSize(); ++n) {
        auto inputT = inputTensor->image(n);
        auto outputT = outputTensor->image(n);
        encoderSingle(cmdBuffer, outputT, inputT, {outputT->width(), outputT->height(), 1});
    }
    return NO_ERROR;
}

class VulkanUnaryCreator : public VulkanBackend::Creator {
public:
    virtual VulkanBasicExecution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs, const MNN::Op* op, Backend* bn) const override {
        if (inputs[0]->buffer().type.code != halide_type_float) {
            return nullptr;
        }
        auto midType = _getMidType(op);
        if (midType.empty()) {
            return nullptr;
        }
        return new VulkanUnary(midType, bn, true);
    }
};

static bool gResistor = []() {
    VulkanBackend::addCreator(OpType_UnaryOp, new VulkanUnaryCreator);
    VulkanBackend::addCreator(OpType_TanH, new VulkanUnaryCreator);
    VulkanBackend::addCreator(OpType_Sigmoid, new VulkanUnaryCreator);
    return true;
}();

} // namespace MNN