快速图像降采样（AVX2、SSE）

最新推荐文章于 2023-04-13 09:08:55 发布

weixin_44591545

最新推荐文章于 2023-04-13 09:08:55 发布

阅读量210

点赞数

文章标签：算法 c++ 开发语言 opencv

本文链接：https://blog.csdn.net/weixin_44591545/article/details/128441181

版权

int DownSampling8X(unsigned char* src, unsigned char* dst, int width, int height)
{
    if (src == NULL || dst == NULL)
        return -1;
    short* temp = (short*)_mm_malloc(width * sizeof(short), 16);
    int block16 = width / 16;
    int block32 = width / 32;
    int block128 = width / 128;
    for (int y = 0; y < height - 7; y += 8)
    {
        int x = 0;
        // init with 1st line
        unsigned char* pLineSrc = src + y * width;
        for (; x < block32 * 32; x += 32)
        {
            short* pLineDst = temp + x;
            __m256i data = _mm256_loadu_si256((__m256i*)(pLineSrc + x));
            _mm256_storeu_si256((__m256i*)pLineDst, _mm256_cvtepu8_epi16(_mm256_extracti128_si256(data, 0)));
            _mm256_storeu_si256((__m256i*)(pLineDst + 16), _mm256_cvtepu8_epi16(_mm256_extracti128_si256(data, 1)));
        }
        for (; x < block16 * 16; x += 16)
        {
            short* pLineDst = temp + x;
            __m256i data = _mm256_cvtepu8_epi16(_mm_loadu_si128((__m128i*)(pLineSrc + x)));
            _mm256_storeu_si256((__m256i*)pLineDst, data);
        }
        for (; x < width; x++)
        {
            temp[x] = pLineSrc[x];
        }
        // calc sum of 8 lines
        for (int z = 1; z < 8; z++)
        {
            unsigned char* pLineSrc = src + (y + z) * width;
            int x = 0;
            for (; x < block32 * 32; x += 32)
            {
                short* pLineDst = temp + x;
                __m256i data = _mm256_loadu_si256((__m256i*)(pLineSrc + x));
                _mm256_storeu_si256((__m256i*)pLineDst, _mm256_add_epi16(_mm256_loadu_si256((__m256i*)pLineDst), _mm256_cvtepu8_epi16(_mm256_extracti128_si256(data, 0))));
                _mm256_storeu_si256((__m256i*)(pLineDst + 16), _mm256_add_epi16(_mm256_loadu_si256((__m256i*)(pLineDst + 16)), _mm256_cvtepu8_epi16(_mm256_extracti128_si256(data, 1))));
            }
            for (; x < block16 * 16; x += 16)
            {
                short* pLineDst = temp + x;
                __m256i data = _mm256_cvtepu8_epi16(_mm_loadu_si128((__m128i*)(pLineSrc + x)));
                _mm256_storeu_si256((__m256i*)pLineDst, _mm256_add_epi16(_mm256_loadu_si256((__m256i*)pLineDst), data));
            }
            for (; x < width; x++)
            {
                temp[x] = pLineSrc[x];
            }
        }
        // calc sum of 8 rows and get avg value
        x = 0;
        for (; x < block128 * 128; x += 128)
        {
            unsigned char* pLineDst = dst + (width / 8) * (y / 8) + x / 8;
            __m256i A = _mm256_loadu_si256((__m256i*)(temp + x));
            __m256i B = _mm256_loadu_si256((__m256i*)(temp + x + 16));
            __m256i C = _mm256_loadu_si256((__m256i*)(temp + x + 32));
            __m256i D = _mm256_loadu_si256((__m256i*)(temp + x + 48));
            __m256i E = _mm256_loadu_si256((__m256i*)(temp + x + 64));
            __m256i F = _mm256_loadu_si256((__m256i*)(temp + x + 80));
            __m256i G = _mm256_loadu_si256((__m256i*)(temp + x + 96));
            __m256i H = _mm256_loadu_si256((__m256i*)(temp + x + 112));
            __m256i AB = _mm256_permute4x64_epi64(_mm256_hadd_epi16(A, B), 0xD8);
            __m256i CD = _mm256_permute4x64_epi64(_mm256_hadd_epi16(C, D), 0xD8);
            __m256i EF = _mm256_permute4x64_epi64(_mm256_hadd_epi16(E, F), 0xD8);
            __m256i GH = _mm256_permute4x64_epi64(_mm256_hadd_epi16(G, H), 0xD8);
            __m256i ABCD = _mm256_permute4x64_epi64(_mm256_hadd_epi16(AB, CD), 0xD8);
            __m256i EFGH = _mm256_permute4x64_epi64(_mm256_hadd_epi16(EF, GH), 0xD8);
            __m256i ABCDEFGH = _mm256_permute4x64_epi64(_mm256_hadd_epi16(ABCD, EFGH), 0xD8);
            ABCDEFGH = _mm256_srai_epi16(ABCDEFGH, 6);
            __m128i ABCDEFGHX = _mm256_extracti128_si256(_mm256_packus_epi16(ABCDEFGH, _mm256_permute2x128_si256(ABCDEFGH, ABCDEFGH, 0x01)), 0);
            _mm_storeu_si128((__m128i*)pLineDst, ABCDEFGHX);
        }
        for (; x < block32 * 32; x += 32)
        {
            unsigned char* pLineDst = dst + (width / 8) * (y / 8) + x / 8;
            __m256i A = _mm256_loadu_si256((__m256i*)(temp + x));
            __m256i B = _mm256_loadu_si256((__m256i*)(temp + x + 16));
            __m256i AB = _mm256_permute4x64_epi64(_mm256_hadd_epi16(A, B), 0xD8);
            AB = _mm256_permute4x64_epi64(_mm256_hadd_epi16(AB, _mm256_setzero_si256()), 0xD8);
            AB = _mm256_permute4x64_epi64(_mm256_hadd_epi16(AB, _mm256_setzero_si256()), 0xD8);
            AB = _mm256_srai_epi16(AB, 6);
            __m128i ABx = _mm256_extracti128_si256(_mm256_packus_epi16(AB, _mm256_permute2x128_si256(AB, AB, 0x01)), 0);
            *((int32_t*)pLineDst) = _mm_cvtsi128_si32(ABx);
        }
        for (; x < width - 7; x += 8)
        {
            unsigned char* pLineDst = dst + (width / 8) * (y / 8) + x / 8;
            int sum = temp[x + 0] + temp[x + 1] + temp[x + 2] + temp[x + 3] + temp[x + 4] + temp[x + 5] + temp[x + 6] + temp[x + 7];
            *pLineDst = sum >> 6;
        }
    }
    _mm_free(temp);
    return 0;
}