Method onCopyBuffer

source/backend/vulkan/buffer/backend/VulkanBackend.cpp:441–534

}

void VulkanBackend::onCopyBuffer(const Tensor* srcTensor, const Tensor* dstTensor) const {
#ifdef MNN_VULKAN_DEBUG
    AUTOTIME;
    MNN_PRINT("Src: ");
    for (int i=0; i<srcTensor->dimensions(); ++i) {
        MNN_PRINT("%d , ", srcTensor->length(i));
    }
    MNN_PRINT("\n");
    MNN_PRINT("Dst: ");
    for (int i=0; i<dstTensor->dimensions(); ++i) {
        MNN_PRINT("%d , ", dstTensor->length(i));
    }
    MNN_PRINT("\n");
#endif

    auto calculateCpSize = [this] (const Tensor* tensor) -> size_t {
            size_t eleSize = (size_t) tensor->elementSize();
            return (tensor->getType().code == halide_type_float && this->mUseFP16) ?
                (eleSize * sizeof(uint16_t)) :
                (eleSize * sizeof(float));
        };

    std::shared_ptr<Tensor> tempTensor;
    if (srcTensor->host<float>() != nullptr) {
        _finish();
        auto format = TensorUtils::getDescribe(dstTensor)->dimensionFormat;
        auto buffer = reinterpret_cast<VulkanBuffer*>(dstTensor->deviceId());
        auto offset = TensorUtils::getDescribeOrigin(dstTensor)->offset;
        // host->gpu
        if(format != TensorUtils::getDescribe(srcTensor)->dimensionFormat) {
            tempTensor.reset(Tensor::create(dstTensor->shape(), dstTensor->getType(), nullptr, _convert(format)));
            MNNCPUCopyBuffer(srcTensor, tempTensor.get());
            srcTensor = tempTensor.get();
        }
        size_t cpSize = calculateCpSize(srcTensor);
        _requireHostBuffer(cpSize);
        _copyTensorToBuffer(srcTensor, mHostBuffer.get(), 0, srcTensor->getType().code == halide_type_float && mUseFP16);
        auto cmdbuffer = mCmdBufferForCopy;
        cmdbuffer->begin(0);
        VkBufferCopy bufferCopy;
        bufferCopy.size = cpSize;
        bufferCopy.dstOffset = offset;
        bufferCopy.srcOffset = 0;
        vkCmdCopyBuffer(cmdbuffer->get(), mHostBuffer->buffer(), buffer->buffer(),
                        1, &bufferCopy);
        cmdbuffer->end();
        pushCommand(cmdbuffer->get());
        _finish();
    } else if (dstTensor->host<float>() != nullptr) {
        // gpu->host
        _finish();
        auto format = TensorUtils::getDescribe(dstTensor)->dimensionFormat;
        if (format != TensorUtils::getDescribe(srcTensor)->dimensionFormat) {
            tempTensor.reset(Tensor::create(srcTensor->shape(), dstTensor->getType(), nullptr, _convert(TensorUtils::getDescribe(srcTensor)->dimensionFormat)), [dstTensor](void* t) {
                Tensor* temp = (Tensor*)t;
                MNNCPUCopyBuffer(temp, dstTensor);
                delete temp;
            });