Method onEncode

source/backend/opencl/execution/image/RasterExecution.cpp:25–407

}

ErrorCode RasterExecution::onEncode(const std::vector<Tensor *> &____inputs, const std::vector<Tensor *> &outputs) {
#ifdef LOG_VERBOSE
    MNN_PRINT("start RasterExecution onResize !\n");
#endif
    mTempInput.clear();
    mTempOutput = nullptr;
    MNN_ASSERT(outputs.size() == 1);
    auto output = outputs[0];
    OpCommonUtils::rasterInputReset(____inputs, outputs[0]);

    auto des = TensorUtils::getDescribe(output);
    auto outputDes = TensorUtils::getDescribe(output);
    mNeedZero = !TensorUtils::regionIsFull(output);
    auto regionNum = des->regions.size();
    auto runtime = ((OpenCLBackend *)backend())->getOpenCLRuntime();
    mFast = false;
    if (outputDes->dimensionFormat == MNN_DATA_FORMAT_NC4HW4) {
        mFast = true;
        for (int i=0; i< des->regions.size(); ++i) {
            auto& slice = des->regions[i];
            if (TensorUtils::getDescribe(slice.origin)->dimensionFormat != MNN_DATA_FORMAT_NC4HW4) {
                mFast = false;
                break;
            }
            if (!OpCommonUtils::canBlitFast(slice, output)) {
                mFast = false;
                break;
            }
        }
    }

    if(mFast)
    {
        mUnits.resize(regionNum);
        int kernel_idx = 0;
        
        if(mNeedZero)
        {
            mUnits.resize(regionNum + 1);
            auto outputShape    = tensorShapeFormat(output);
            int region[] = {outputShape[0], UP_DIV(outputShape[3], 4), outputShape[1], outputShape[2]};//nhwc
            Unit &unit          = mUnits[kernel_idx++];
            unit.kernel         = runtime->buildKernel("raster", "image_set_zero", {}, mOpenCLBackend->getPrecision(), output, output);
            unit.localWorkSize  = {8, 8};
            unit.globalWorkSize = {(uint32_t)UP_DIV((region[1] * region[3]), 16)*16,
                                   (uint32_t)UP_DIV((region[0] * region[2]), 16)*16};

            int global_dim0 = region[1] * region[3];
            int global_dim1 = region[0] * region[2];

            uint32_t idx   = 0;
            cl_int ret = CL_SUCCESS;
            ret |= unit.kernel->get().setArg(idx++, global_dim0);
            ret |= unit.kernel->get().setArg(idx++, global_dim1);
            ret |= unit.kernel->get().setArg(idx++, openCLImage(output));
            if(ret != CL_SUCCESS)
            {
                MNN_PRINT("setArg err %d\n", (int)ret);