ncnn/src/layer/cast.cpp

// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#include "cast.h"

namespace ncnn {

Cast::Cast()
{
    one_blob_only = true;
    support_inplace = false;
    support_packing = true;
}

int Cast::load_param(const ParamDict& pd)
{
    type_from = pd.get(0, 0);
    type_to = pd.get(1, 0);

    return 0;
}

// round to nearest
signed char float32_to_int8(float value)
{
    float tmp;
    if (value >= 0.f)
        tmp = value + 0.5f;
    else
        tmp = value - 0.5f;

    if (tmp > 127)
        return 127;
    if (tmp < -128)
        return -128;

    return static_cast<signed char>(tmp);
}

int Cast::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
{
    if (type_from == type_to)
    {
        top_blob = bottom_blob;
        return 0;
    }

    int w = bottom_blob.w;
    int h = bottom_blob.h;
    int channels = bottom_blob.c;
    int dims = bottom_blob.dims;
    size_t elemsize = bottom_blob.elemsize;
    int elempack = bottom_blob.elempack;

    size_t out_elemsize = elemsize;
    if (type_to == 1)
    {
        // float32
        out_elemsize = 4 * elempack;
    }
    else if (type_to == 2)
    {
        // float16
        out_elemsize = 2 * elempack;
    }
    else if (type_to == 3)
    {
        // int8
        out_elemsize = elempack;
    }
    else if (type_to == 4)
    {
        // bfloat16
        out_elemsize = 2 * elempack;
    }

    if (dims == 1)
    {
        top_blob.create(w, out_elemsize, elempack, opt.blob_allocator);
    }
    else if (dims == 2)
    {
        top_blob.create(w, h, out_elemsize, elempack, opt.blob_allocator);
    }
    else if (dims == 3)
    {
        top_blob.create(w, h, channels, out_elemsize, elempack, opt.blob_allocator);
    }
    if (top_blob.empty())
        return -100;

    int size = w * h * elempack;

    if (type_from == 1 && type_to == 2)
    {
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const float* ptr = bottom_blob.channel(q);
            unsigned short* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = float32_to_float16(ptr[i]);
            }
        }
    }

    if (type_from == 2 && type_to == 1)
    {
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const unsigned short* ptr = bottom_blob.channel(q);
            float* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = float16_to_float32(ptr[i]);
            }
        }
    }

    if (type_from == 3 && type_to == 1)
    {
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const signed char* ptr = bottom_blob.channel(q);
            float* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = (float)ptr[i];
            }
        }
    }

    if (type_from == 1 && type_to == 4)
    {
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const float* ptr = bottom_blob.channel(q);
            unsigned short* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = float32_to_bfloat16(ptr[i]);
            }
        }
    }

    if (type_from == 4 && type_to == 1)
    {
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const unsigned short* ptr = bottom_blob.channel(q);
            float* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = bfloat16_to_float32(ptr[i]);
            }
        }
    }

    // TODO more cast type

    return 0;
}

} // namespace ncnn