Method onExecute

source/backend/cpu/CPUDynamicQuant.cpp:30–56

}

ErrorCode CPUDynamicQuant::onExecute(const std::vector<Tensor*> &inputs,
                                  const std::vector<Tensor*> &outputs) {
    auto core = static_cast<CPUBackend*>(backend())->functions();
    auto int8core = static_cast<CPUBackend*>(backend())->int8Functions();
    float *inputPtr = inputs[0]->host<float>();
    int8_t *outputPtr = outputs[0]->host<int8_t>();
    int size = static_cast<CPUBackend*>(backend())->getTensorSize(inputs[0]);
    float quantScale = 0.f, dequantScale = 0.f, zeroPoint = 0.f;
    float maxVal = 0.f, minVal = 0.f;
    core->MNNCountMaxMinValue(inputPtr, &minVal, &maxVal, size);
    // Compute scale and zero
    float range = maxVal - minVal;
    MNN_ASSERT(range != 0);
    quantScale = 255.0f / range;
    dequantScale = range / 255.0f;
    zeroPoint = roundf(-(minVal * 255.f) / range) - 128.0f;
    int pack = core->pack;
    std::vector<float> qsVec(pack, quantScale);
    int sizeDiv = UP_DIV(size, pack);
    int8core->MNNFloat2Int8(inputPtr, outputPtr, sizeDiv, &quantScale, -128, 127, &zeroPoint, 0);
    float* scale = outputs[1]->host<float>();
    float* zeros = outputs[2]->host<float>();
    *scale = dequantScale;
    *zeros = zeroPoint;
    
    return NO_ERROR;
}