utils.c

/*
 * Copyright (C) 2001-2003 Michael Niedermayer <michaelni@gmx.at>
 *
 * This file is part of Libav.
 *
 * Libav is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * Libav is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with Libav; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

#include "config.h"

#define _SVID_SOURCE // needed for MAP_ANONYMOUS
#include <assert.h>
#include <inttypes.h>
#include <math.h>
#include <stdio.h>
#include <string.h>
#if HAVE_SYS_MMAN_H
#include <sys/mman.h>
#if defined(MAP_ANON) && !defined(MAP_ANONYMOUS)
#define MAP_ANONYMOUS MAP_ANON
#endif
#endif
#if HAVE_VIRTUALALLOC
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#endif

#include "libavutil/attributes.h"
#include "libavutil/avutil.h"
#include "libavutil/bswap.h"
#include "libavutil/cpu.h"
#include "libavutil/intreadwrite.h"
#include "libavutil/mathematics.h"
#include "libavutil/opt.h"
#include "libavutil/pixdesc.h"
#include "libavutil/x86/asm.h"
#include "rgb2rgb.h"
#include "swscale.h"
#include "swscale_internal.h"

unsigned swscale_version(void)
{
    return LIBSWSCALE_VERSION_INT;
}

const char *swscale_configuration(void)
{
    return LIBAV_CONFIGURATION;
}

const char *swscale_license(void)
{
#define LICENSE_PREFIX "libswscale license: "
    return LICENSE_PREFIX LIBAV_LICENSE + sizeof(LICENSE_PREFIX) - 1;
}

#define RET 0xC3 // near return opcode for x86

typedef struct FormatEntry {
    int is_supported_in, is_supported_out;
} FormatEntry;

static const FormatEntry format_entries[PIX_FMT_NB] = {
    [PIX_FMT_YUV420P]     = { 1, 1 },
    [PIX_FMT_YUYV422]     = { 1, 1 },
    [PIX_FMT_RGB24]       = { 1, 1 },
    [PIX_FMT_BGR24]       = { 1, 1 },
    [PIX_FMT_YUV422P]     = { 1, 1 },
    [PIX_FMT_YUV444P]     = { 1, 1 },
    [PIX_FMT_YUV410P]     = { 1, 1 },
    [PIX_FMT_YUV411P]     = { 1, 1 },
    [PIX_FMT_GRAY8]       = { 1, 1 },
    [PIX_FMT_MONOWHITE]   = { 1, 1 },
    [PIX_FMT_MONOBLACK]   = { 1, 1 },
    [PIX_FMT_PAL8]        = { 1, 0 },
    [PIX_FMT_YUVJ420P]    = { 1, 1 },
    [PIX_FMT_YUVJ422P]    = { 1, 1 },
    [PIX_FMT_YUVJ444P]    = { 1, 1 },
    [PIX_FMT_UYVY422]     = { 1, 1 },
    [PIX_FMT_UYYVYY411]   = { 0, 0 },
    [PIX_FMT_BGR8]        = { 1, 1 },
    [PIX_FMT_BGR4]        = { 0, 1 },
    [PIX_FMT_BGR4_BYTE]   = { 1, 1 },
    [PIX_FMT_RGB8]        = { 1, 1 },
    [PIX_FMT_RGB4]        = { 0, 1 },
    [PIX_FMT_RGB4_BYTE]   = { 1, 1 },
    [PIX_FMT_NV12]        = { 1, 1 },
    [PIX_FMT_NV21]        = { 1, 1 },
    [PIX_FMT_ARGB]        = { 1, 1 },
    [PIX_FMT_RGBA]        = { 1, 1 },
    [PIX_FMT_ABGR]        = { 1, 1 },
    [PIX_FMT_BGRA]        = { 1, 1 },
    [PIX_FMT_GRAY16BE]    = { 1, 1 },
    [PIX_FMT_GRAY16LE]    = { 1, 1 },
    [PIX_FMT_YUV440P]     = { 1, 1 },
    [PIX_FMT_YUVJ440P]    = { 1, 1 },
    [PIX_FMT_YUVA420P]    = { 1, 1 },
    [PIX_FMT_RGB48BE]     = { 1, 1 },
    [PIX_FMT_RGB48LE]     = { 1, 1 },
    [PIX_FMT_RGB565BE]    = { 1, 1 },
    [PIX_FMT_RGB565LE]    = { 1, 1 },
    [PIX_FMT_RGB555BE]    = { 1, 1 },
    [PIX_FMT_RGB555LE]    = { 1, 1 },
    [PIX_FMT_BGR565BE]    = { 1, 1 },
    [PIX_FMT_BGR565LE]    = { 1, 1 },
    [PIX_FMT_BGR555BE]    = { 1, 1 },
    [PIX_FMT_BGR555LE]    = { 1, 1 },
    [PIX_FMT_YUV420P16LE] = { 1, 1 },
    [PIX_FMT_YUV420P16BE] = { 1, 1 },
    [PIX_FMT_YUV422P16LE] = { 1, 1 },
    [PIX_FMT_YUV422P16BE] = { 1, 1 },
    [PIX_FMT_YUV444P16LE] = { 1, 1 },
    [PIX_FMT_YUV444P16BE] = { 1, 1 },
    [PIX_FMT_RGB444LE]    = { 1, 1 },
    [PIX_FMT_RGB444BE]    = { 1, 1 },
    [PIX_FMT_BGR444LE]    = { 1, 1 },
    [PIX_FMT_BGR444BE]    = { 1, 1 },
    [PIX_FMT_Y400A]       = { 1, 0 },
    [PIX_FMT_BGR48BE]     = { 1, 1 },
    [PIX_FMT_BGR48LE]     = { 1, 1 },
    [PIX_FMT_YUV420P9BE]  = { 1, 1 },
    [PIX_FMT_YUV420P9LE]  = { 1, 1 },
    [PIX_FMT_YUV420P10BE] = { 1, 1 },
    [PIX_FMT_YUV420P10LE] = { 1, 1 },
    [PIX_FMT_YUV422P9BE]  = { 1, 1 },
    [PIX_FMT_YUV422P9LE]  = { 1, 1 },
    [PIX_FMT_YUV422P10BE] = { 1, 1 },
    [PIX_FMT_YUV422P10LE] = { 1, 1 },
    [PIX_FMT_YUV444P9BE]  = { 1, 1 },
    [PIX_FMT_YUV444P9LE]  = { 1, 1 },
    [PIX_FMT_YUV444P10BE] = { 1, 1 },
    [PIX_FMT_YUV444P10LE] = { 1, 1 },
    [PIX_FMT_GBRP]        = { 1, 0 },
    [PIX_FMT_GBRP9LE]     = { 1, 0 },
    [PIX_FMT_GBRP9BE]     = { 1, 0 },
    [PIX_FMT_GBRP10LE]    = { 1, 0 },
    [PIX_FMT_GBRP10BE]    = { 1, 0 },
    [PIX_FMT_GBRP16LE]    = { 1, 0 },
    [PIX_FMT_GBRP16BE]    = { 1, 0 },
};

int sws_isSupportedInput(enum PixelFormat pix_fmt)
{
    return (unsigned)pix_fmt < PIX_FMT_NB ?
           format_entries[pix_fmt].is_supported_in : 0;
}

int sws_isSupportedOutput(enum PixelFormat pix_fmt)
{
    return (unsigned)pix_fmt < PIX_FMT_NB ?
           format_entries[pix_fmt].is_supported_out : 0;
}

extern const int32_t ff_yuv2rgb_coeffs[8][4];

const char *sws_format_name(enum PixelFormat format)
{
    if ((unsigned)format < PIX_FMT_NB && av_pix_fmt_descriptors[format].name)
        return av_pix_fmt_descriptors[format].name;
    else
        return "Unknown format";
}

static double getSplineCoeff(double a, double b, double c, double d,
                             double dist)
{
    if (dist <= 1.0)
        return ((d * dist + c) * dist + b) * dist + a;
    else
        return getSplineCoeff(0.0,
                               b + 2.0 * c + 3.0 * d,
                               c + 3.0 * d,
                              -b - 3.0 * c - 6.0 * d,
                              dist - 1.0);
}

static int initFilter(int16_t **outFilter, int32_t **filterPos,
                      int *outFilterSize, int xInc, int srcW, int dstW,
                      int filterAlign, int one, int flags, int cpu_flags,
                      SwsVector *srcFilter, SwsVector *dstFilter,
                      double param[2], int is_horizontal)
{
    int i;
    int filterSize;
    int filter2Size;
    int minFilterSize;
    int64_t *filter    = NULL;
    int64_t *filter2   = NULL;
    const int64_t fone = 1LL << 54;
    int ret            = -1;

    emms_c(); // FIXME should not be required but IS (even for non-MMX versions)

    // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
    FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);

    if (FFABS(xInc - 0x10000) < 10) { // unscaled
        int i;
        filterSize = 1;
        FF_ALLOCZ_OR_GOTO(NULL, filter,
                          dstW * sizeof(*filter) * filterSize, fail);

        for (i = 0; i < dstW; i++) {
            filter[i * filterSize] = fone;
            (*filterPos)[i]        = i;
        }
    } else if (flags & SWS_POINT) { // lame looking point sampling mode
        int i;
        int xDstInSrc;
        filterSize = 1;
        FF_ALLOC_OR_GOTO(NULL, filter,
                         dstW * sizeof(*filter) * filterSize, fail);

        xDstInSrc = xInc / 2 - 0x8000;
        for (i = 0; i < dstW; i++) {
            int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;

            (*filterPos)[i] = xx;
            filter[i]       = fone;
            xDstInSrc      += xInc;
        }
    } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
               (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
        int i;
        int xDstInSrc;
        filterSize = 2;
        FF_ALLOC_OR_GOTO(NULL, filter,
                         dstW * sizeof(*filter) * filterSize, fail);

        xDstInSrc = xInc / 2 - 0x8000;
        for (i = 0; i < dstW; i++) {
            int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
            int j;

            (*filterPos)[i] = xx;
            // bilinear upscale / linear interpolate / area averaging
            for (j = 0; j < filterSize; j++) {
                int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
                                (fone >> 16);
                if (coeff < 0)
                    coeff = 0;
                filter[i * filterSize + j] = coeff;
                xx++;
            }
            xDstInSrc += xInc;
        }
    } else {
        int64_t xDstInSrc;
        int sizeFactor;

        if (flags & SWS_BICUBIC)
            sizeFactor = 4;
        else if (flags & SWS_X)
            sizeFactor = 8;
        else if (flags & SWS_AREA)
            sizeFactor = 1;     // downscale only, for upscale it is bilinear
        else if (flags & SWS_GAUSS)
            sizeFactor = 8;     // infinite ;)
        else if (flags & SWS_LANCZOS)
            sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
        else if (flags & SWS_SINC)
            sizeFactor = 20;    // infinite ;)
        else if (flags & SWS_SPLINE)
            sizeFactor = 20;    // infinite ;)
        else if (flags & SWS_BILINEAR)
            sizeFactor = 2;
        else {
            sizeFactor = 0;     // GCC warning killer
            assert(0);
        }

        if (xInc <= 1 << 16)
            filterSize = 1 + sizeFactor;    // upscale
        else
            filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;

        filterSize = FFMIN(filterSize, srcW - 2);
        filterSize = FFMAX(filterSize, 1);

        FF_ALLOC_OR_GOTO(NULL, filter,
                         dstW * sizeof(*filter) * filterSize, fail);

        xDstInSrc = xInc - 0x10000;
        for (i = 0; i < dstW; i++) {
            int xx = (xDstInSrc - ((filterSize - 2) << 16)) / (1 << 17);
            int j;
            (*filterPos)[i] = xx;
            for (j = 0; j < filterSize; j++) {
                int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
                double floatd;
                int64_t coeff;

                if (xInc > 1 << 16)
                    d = d * dstW / srcW;
                floatd = d * (1.0 / (1 << 30));

                if (flags & SWS_BICUBIC) {
                    int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] :   0) * (1 << 24);
                    int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);

                    if (d >= 1LL << 31) {
                        coeff = 0.0;
                    } else {
                        int64_t dd  = (d  * d) >> 30;
                        int64_t ddd = (dd * d) >> 30;

                        if (d < 1LL << 30)
                            coeff =  (12 * (1 << 24) -  9 * B - 6 * C) * ddd +
                                    (-18 * (1 << 24) + 12 * B + 6 * C) *  dd +
                                      (6 * (1 << 24) -  2 * B)         * (1 << 30);
                        else
                            coeff =      (-B -  6 * C) * ddd +
                                      (6 * B + 30 * C) * dd  +
                                    (-12 * B - 48 * C) * d   +
                                      (8 * B + 24 * C) * (1 << 30);
                    }
                    coeff *= fone >> (30 + 24);
                }
#if 0
                else if (flags & SWS_X) {
                    double p  = param ? param * 0.01 : 0.3;
                    coeff     = d ? sin(d * M_PI) / (d * M_PI) : 1.0;
                    coeff    *= pow(2.0, -p * d * d);
                }
#endif
                else if (flags & SWS_X) {
                    double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
                    double c;

                    if (floatd < 1.0)
                        c = cos(floatd * M_PI);
                    else
                        c = -1.0;
                    if (c < 0.0)
                        c = -pow(-c, A);
                    else
                        c = pow(c, A);
                    coeff = (c * 0.5 + 0.5) * fone;
                } else if (flags & SWS_AREA) {
                    int64_t d2 = d - (1 << 29);
                    if (d2 * xInc < -(1LL << (29 + 16)))
                        coeff = 1.0 * (1LL << (30 + 16));
                    else if (d2 * xInc < (1LL << (29 + 16)))
                        coeff = -d2 * xInc + (1LL << (29 + 16));
                    else
                        coeff = 0.0;
                    coeff *= fone >> (30 + 16);
                } else if (flags & SWS_GAUSS) {
                    double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
                    coeff = (pow(2.0, -p * floatd * floatd)) * fone;
                } else if (flags & SWS_SINC) {
                    coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
                } else if (flags & SWS_LANCZOS) {
                    double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
                    coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
                             (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
                    if (floatd > p)
                        coeff = 0;
                } else if (flags & SWS_BILINEAR) {
                    coeff = (1 << 30) - d;
                    if (coeff < 0)
                        coeff = 0;
                    coeff *= fone >> 30;
                } else if (flags & SWS_SPLINE) {
                    double p = -2.196152422706632;
                    coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
                } else {
                    coeff = 0.0; // GCC warning killer
                    assert(0);
                }

                filter[i * filterSize + j] = coeff;
                xx++;
            }
            xDstInSrc += 2 * xInc;
        }
    }

    /* apply src & dst Filter to filter -> filter2
     * av_free(filter);
     */
    assert(filterSize > 0);
    filter2Size = filterSize;
    if (srcFilter)
        filter2Size += srcFilter->length - 1;
    if (dstFilter)
        filter2Size += dstFilter->length - 1;
    assert(filter2Size > 0);
    FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);

    for (i = 0; i < dstW; i++) {
        int j, k;

        if (srcFilter) {
            for (k = 0; k < srcFilter->length; k++) {
                for (j = 0; j < filterSize; j++)
                    filter2[i * filter2Size + k + j] +=
                        srcFilter->coeff[k] * filter[i * filterSize + j];
            }
        } else {
            for (j = 0; j < filterSize; j++)
                filter2[i * filter2Size + j] = filter[i * filterSize + j];
        }
        // FIXME dstFilter

        (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
    }
    av_freep(&filter);

    /* try to reduce the filter-size (step1 find size and shift left) */
    // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
    minFilterSize = 0;
    for (i = dstW - 1; i >= 0; i--) {
        int min = filter2Size;
        int j;
        int64_t cutOff = 0.0;

        /* get rid of near zero elements on the left by shifting left */
        for (j = 0; j < filter2Size; j++) {
            int k;
            cutOff += FFABS(filter2[i * filter2Size]);

            if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
                break;

            /* preserve monotonicity because the core can't handle the
             * filter otherwise */
            if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
                break;

            // move filter coefficients left
            for (k = 1; k < filter2Size; k++)
                filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
            filter2[i * filter2Size + k - 1] = 0;
            (*filterPos)[i]++;
        }

        cutOff = 0;
        /* count near zeros on the right */
        for (j = filter2Size - 1; j > 0; j--) {
            cutOff += FFABS(filter2[i * filter2Size + j]);

            if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
                break;
            min--;
        }

        if (min > minFilterSize)
            minFilterSize = min;
    }

    if (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) {
        // we can handle the special case 4, so we don't want to go the full 8
        if (minFilterSize < 5)
            filterAlign = 4;

        /* We really don't want to waste our time doing useless computation, so
         * fall back on the scalar C code for very small filters.
         * Vectorizing is worth it only if you have a decent-sized vector. */
        if (minFilterSize < 3)
            filterAlign = 1;
    }

    if (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) {
        // special case for unscaled vertical filtering
        if (minFilterSize == 1 && filterAlign == 2)
            filterAlign = 1;
    }

    assert(minFilterSize > 0);
    filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
    assert(filterSize > 0);
    filter = av_malloc(filterSize * dstW * sizeof(*filter));
    if (filterSize >= MAX_FILTER_SIZE * 16 /
                      ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
        goto fail;
    *outFilterSize = filterSize;

    if (flags & SWS_PRINT_INFO)
        av_log(NULL, AV_LOG_VERBOSE,
               "SwScaler: reducing / aligning filtersize %d -> %d\n",
               filter2Size, filterSize);
    /* try to reduce the filter-size (step2 reduce it) */
    for (i = 0; i < dstW; i++) {
        int j;

        for (j = 0; j < filterSize; j++) {
            if (j >= filter2Size)
                filter[i * filterSize + j] = 0;
            else
                filter[i * filterSize + j] = filter2[i * filter2Size + j];
            if ((flags & SWS_BITEXACT) && j >= minFilterSize)
                filter[i * filterSize + j] = 0;
        }
    }

    // FIXME try to align filterPos if possible

    // fix borders
    if (is_horizontal) {
        for (i = 0; i < dstW; i++) {
            int j;
            if ((*filterPos)[i] < 0) {
                // move filter coefficients left to compensate for filterPos
                for (j = 1; j < filterSize; j++) {
                    int left = FFMAX(j + (*filterPos)[i], 0);
                    filter[i * filterSize + left] += filter[i * filterSize + j];
                    filter[i * filterSize + j]     = 0;
                }
                (*filterPos)[i] = 0;
            }

            if ((*filterPos)[i] + filterSize > srcW) {
                int shift = (*filterPos)[i] + filterSize - srcW;
                // move filter coefficients right to compensate for filterPos
                for (j = filterSize - 2; j >= 0; j--) {
                    int right = FFMIN(j + shift, filterSize - 1);
                    filter[i * filterSize + right] += filter[i * filterSize + j];
                    filter[i * filterSize + j]      = 0;
                }
                (*filterPos)[i] = srcW - filterSize;
            }
        }
    }

    // Note the +1 is for the MMX scaler which reads over the end
    /* align at 16 for AltiVec (needed by hScale_altivec_real) */
    FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
                      *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);

    /* normalize & store in outFilter */
    for (i = 0; i < dstW; i++) {
        int j;
        int64_t error = 0;
        int64_t sum   = 0;

        for (j = 0; j < filterSize; j++) {
            sum += filter[i * filterSize + j];
        }
        sum = (sum + one / 2) / one;
        for (j = 0; j < *outFilterSize; j++) {
            int64_t v = filter[i * filterSize + j] + error;
            int intV  = ROUNDED_DIV(v, sum);
            (*outFilter)[i * (*outFilterSize) + j] = intV;
            error                                  = v - intV * sum;
        }
    }

    (*filterPos)[dstW + 0] =
    (*filterPos)[dstW + 1] =
    (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
                                                      * read over the end */
    for (i = 0; i < *outFilterSize; i++) {
        int k = (dstW - 1) * (*outFilterSize) + i;
        (*outFilter)[k + 1 * (*outFilterSize)] =
        (*outFilter)[k + 2 * (*outFilterSize)] =
        (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
    }

    ret = 0;

fail:
    av_free(filter);
    av_free(filter2);
    return ret;
}

#if HAVE_MMXEXT && HAVE_INLINE_ASM
static int initMMX2HScaler(int dstW, int xInc, uint8_t *filterCode,
                           int16_t *filter, int32_t *filterPos, int numSplits)
{
    uint8_t *fragmentA;
    x86_reg imm8OfPShufW1A;
    x86_reg imm8OfPShufW2A;
    x86_reg fragmentLengthA;
    uint8_t *fragmentB;
    x86_reg imm8OfPShufW1B;
    x86_reg imm8OfPShufW2B;
    x86_reg fragmentLengthB;
    int fragmentPos;

    int xpos, i;

    // create an optimized horizontal scaling routine
    /* This scaler is made of runtime-generated MMX2 code using specially tuned
     * pshufw instructions. For every four output pixels, if four input pixels
     * are enough for the fast bilinear scaling, then a chunk of fragmentB is
     * used. If five input pixels are needed, then a chunk of fragmentA is used.
     */

    // code fragment

    __asm__ volatile (
        "jmp                         9f                 \n\t"
        // Begin
        "0:                                             \n\t"
        "movq    (%%"REG_d", %%"REG_a"), %%mm3          \n\t"
        "movd    (%%"REG_c", %%"REG_S"), %%mm0          \n\t"
        "movd   1(%%"REG_c", %%"REG_S"), %%mm1          \n\t"
        "punpcklbw                %%mm7, %%mm1          \n\t"
        "punpcklbw                %%mm7, %%mm0          \n\t"
        "pshufw                   $0xFF, %%mm1, %%mm1   \n\t"
        "1:                                             \n\t"
        "pshufw                   $0xFF, %%mm0, %%mm0   \n\t"
        "2:                                             \n\t"
        "psubw                    %%mm1, %%mm0          \n\t"
        "movl   8(%%"REG_b", %%"REG_a"), %%esi          \n\t"
        "pmullw                   %%mm3, %%mm0          \n\t"
        "psllw                       $7, %%mm1          \n\t"
        "paddw                    %%mm1, %%mm0          \n\t"

        "movq                     %%mm0, (%%"REG_D", %%"REG_a") \n\t"

        "add                         $8, %%"REG_a"      \n\t"
        // End
        "9:                                             \n\t"
        // "int $3                                         \n\t"
        "lea       " LOCAL_MANGLE(0b) ", %0             \n\t"
        "lea       " LOCAL_MANGLE(1b) ", %1             \n\t"
        "lea       " LOCAL_MANGLE(2b) ", %2             \n\t"
        "dec                         %1                 \n\t"
        "dec                         %2                 \n\t"
        "sub                         %0, %1             \n\t"
        "sub                         %0, %2             \n\t"
        "lea       " LOCAL_MANGLE(9b) ", %3             \n\t"
        "sub                         %0, %3             \n\t"


        : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
          "=r" (fragmentLengthA)
        );

    __asm__ volatile (
        "jmp                         9f                 \n\t"
        // Begin
        "0:                                             \n\t"
        "movq    (%%"REG_d", %%"REG_a"), %%mm3          \n\t"
        "movd    (%%"REG_c", %%"REG_S"), %%mm0          \n\t"
        "punpcklbw                %%mm7, %%mm0          \n\t"
        "pshufw                   $0xFF, %%mm0, %%mm1   \n\t"
        "1:                                             \n\t"
        "pshufw                   $0xFF, %%mm0, %%mm0   \n\t"
        "2:                                             \n\t"
        "psubw                    %%mm1, %%mm0          \n\t"
        "movl   8(%%"REG_b", %%"REG_a"), %%esi          \n\t"
        "pmullw                   %%mm3, %%mm0          \n\t"
        "psllw                       $7, %%mm1          \n\t"
        "paddw                    %%mm1, %%mm0          \n\t"

        "movq                     %%mm0, (%%"REG_D", %%"REG_a") \n\t"

        "add                         $8, %%"REG_a"      \n\t"
        // End
        "9:                                             \n\t"
        // "int                       $3                   \n\t"
        "lea       " LOCAL_MANGLE(0b) ", %0             \n\t"
        "lea       " LOCAL_MANGLE(1b) ", %1             \n\t"
        "lea       " LOCAL_MANGLE(2b) ", %2             \n\t"
        "dec                         %1                 \n\t"
        "dec                         %2                 \n\t"
        "sub                         %0, %1             \n\t"
        "sub                         %0, %2             \n\t"
        "lea       " LOCAL_MANGLE(9b) ", %3             \n\t"
        "sub                         %0, %3             \n\t"


        : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
          "=r" (fragmentLengthB)
        );

    xpos        = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
    fragmentPos = 0;

    for (i = 0; i < dstW / numSplits; i++) {
        int xx = xpos >> 16;

        if ((i & 3) == 0) {
            int a                  = 0;
            int b                  = ((xpos + xInc) >> 16) - xx;
            int c                  = ((xpos + xInc * 2) >> 16) - xx;
            int d                  = ((xpos + xInc * 3) >> 16) - xx;
            int inc                = (d + 1 < 4);
            uint8_t *fragment      = (d + 1 < 4) ? fragmentB : fragmentA;
            x86_reg imm8OfPShufW1  = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
            x86_reg imm8OfPShufW2  = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
            x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
            int maxShift           = 3 - (d + inc);
            int shift              = 0;

            if (filterCode) {
                filter[i]        = ((xpos              & 0xFFFF) ^ 0xFFFF) >> 9;
                filter[i + 1]    = (((xpos + xInc)     & 0xFFFF) ^ 0xFFFF) >> 9;
                filter[i + 2]    = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
                filter[i + 3]    = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
                filterPos[i / 2] = xx;

                memcpy(filterCode + fragmentPos, fragment, fragmentLength);

                filterCode[fragmentPos + imm8OfPShufW1] =  (a + inc)       |
                                                          ((b + inc) << 2) |
                                                          ((c + inc) << 4) |
                                                          ((d + inc) << 6);
                filterCode[fragmentPos + imm8OfPShufW2] =  a | (b << 2) |
                                                               (c << 4) |
                                                               (d << 6);

                if (i + 4 - inc >= dstW)
                    shift = maxShift;               // avoid overread
                else if ((filterPos[i / 2] & 3) <= maxShift)
                    shift = filterPos[i / 2] & 3;   // align

                if (shift && i >= shift) {
                    filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
                    filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
                    filterPos[i / 2]                        -= shift;
                }
            }

            fragmentPos += fragmentLength;

            if (filterCode)
                filterCode[fragmentPos] = RET;
        }
        xpos += xInc;
    }
    if (filterCode)
        filterPos[((i / 2) + 1) & (~1)] = xpos >> 16;  // needed to jump to the next part

    return fragmentPos + 1;
}
#endif /* HAVE_MMXEXT && HAVE_INLINE_ASM */

static void getSubSampleFactors(int *h, int *v, enum PixelFormat format)
{
    *h = av_pix_fmt_descriptors[format].log2_chroma_w;
    *v = av_pix_fmt_descriptors[format].log2_chroma_h;
}

int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
                             int srcRange, const int table[4], int dstRange,
                             int brightness, int contrast, int saturation)
{
    memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
    memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);

    c->brightness = brightness;
    c->contrast   = contrast;
    c->saturation = saturation;
    c->srcRange   = srcRange;
    c->dstRange   = dstRange;
    if (isYUV(c->dstFormat) || isGray(c->dstFormat))
        return -1;

    c->dstFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[c->dstFormat]);
    c->srcFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[c->srcFormat]);

    ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
                             contrast, saturation);
    // FIXME factorize

    if (HAVE_ALTIVEC && av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)
        ff_yuv2rgb_init_tables_altivec(c, inv_table, brightness,
                                       contrast, saturation);
    return 0;
}

int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
                             int *srcRange, int **table, int *dstRange,
                             int *brightness, int *contrast, int *saturation)
{
    if (isYUV(c->dstFormat) || isGray(c->dstFormat))
        return -1;

    *inv_table  = c->srcColorspaceTable;
    *table      = c->dstColorspaceTable;
    *srcRange   = c->srcRange;
    *dstRange   = c->dstRange;
    *brightness = c->brightness;
    *contrast   = c->contrast;
    *saturation = c->saturation;

    return 0;
}

static int handle_jpeg(enum PixelFormat *format)
{
    switch (*format) {
    case PIX_FMT_YUVJ420P:
        *format = PIX_FMT_YUV420P;
        return 1;
    case PIX_FMT_YUVJ422P:
        *format = PIX_FMT_YUV422P;
        return 1;
    case PIX_FMT_YUVJ444P:
        *format = PIX_FMT_YUV444P;
        return 1;
    case PIX_FMT_YUVJ440P:
        *format = PIX_FMT_YUV440P;
        return 1;
    default:
        return 0;
    }
}

SwsContext *sws_alloc_context(void)
{
    SwsContext *c = av_mallocz(sizeof(SwsContext));

    c->av_class = &sws_context_class;
    av_opt_set_defaults(c);

    return c;
}

av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
                             SwsFilter *dstFilter)
{
    int i;
    int usesVFilter, usesHFilter;
    int unscaled;
    SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
    int srcW              = c->srcW;
    int srcH              = c->srcH;
    int dstW              = c->dstW;
    int dstH              = c->dstH;
    int dst_stride        = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
    int dst_stride_px     = dst_stride >> 1;
    int flags, cpu_flags;
    enum PixelFormat srcFormat = c->srcFormat;
    enum PixelFormat dstFormat = c->dstFormat;

    cpu_flags = av_get_cpu_flags();
    flags     = c->flags;
    emms_c();
    if (!rgb15to16)
        sws_rgb2rgb_init();

    unscaled = (srcW == dstW && srcH == dstH);

    if (!sws_isSupportedInput(srcFormat)) {
        av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
               sws_format_name(srcFormat));
        return AVERROR(EINVAL);
    }
    if (!sws_isSupportedOutput(dstFormat)) {
        av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
               sws_format_name(dstFormat));
        return AVERROR(EINVAL);
    }

    i = flags & (SWS_POINT         |
                 SWS_AREA          |
                 SWS_BILINEAR      |
                 SWS_FAST_BILINEAR |
                 SWS_BICUBIC       |
                 SWS_X             |
                 SWS_GAUSS         |
                 SWS_LANCZOS       |
                 SWS_SINC          |
                 SWS_SPLINE        |
                 SWS_BICUBLIN);
    if (!i || (i & (i - 1))) {
        av_log(c, AV_LOG_ERROR,
               "Exactly one scaler algorithm must be chosen\n");
        return AVERROR(EINVAL);
    }
    /* sanity check */
    if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
        /* FIXME check if these are enough and try to lower them after
         * fixing the relevant parts of the code */
        av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
               srcW, srcH, dstW, dstH);
        return AVERROR(EINVAL);
    }

    if (!dstFilter)
        dstFilter = &dummyFilter;
    if (!srcFilter)
        srcFilter = &dummyFilter;

    c->lumXInc      = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
    c->lumYInc      = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
    c->dstFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[dstFormat]);
    c->srcFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[srcFormat]);
    c->vRounder     = 4 * 0x0001000100010001ULL;

    usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
                  (srcFilter->chrV && srcFilter->chrV->length > 1) ||
                  (dstFilter->lumV && dstFilter->lumV->length > 1) ||
                  (dstFilter->chrV && dstFilter->chrV->length > 1);
    usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
                  (srcFilter->chrH && srcFilter->chrH->length > 1) ||
                  (dstFilter->lumH && dstFilter->lumH->length > 1) ||
                  (dstFilter->chrH && dstFilter->chrH->length > 1);

    getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
    getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);

    /* reuse chroma for 2 pixels RGB/BGR unless user wants full
     * chroma interpolation */
    if (flags & SWS_FULL_CHR_H_INT &&
        isAnyRGB(dstFormat)        &&
        dstFormat != PIX_FMT_RGBA  &&
        dstFormat != PIX_FMT_ARGB  &&
        dstFormat != PIX_FMT_BGRA  &&
        dstFormat != PIX_FMT_ABGR  &&
        dstFormat != PIX_FMT_RGB24 &&
        dstFormat != PIX_FMT_BGR24) {
        av_log(c, AV_LOG_ERROR,
               "full chroma interpolation for destination format '%s' not yet implemented\n",
               sws_format_name(dstFormat));
        flags   &= ~SWS_FULL_CHR_H_INT;
        c->flags = flags;
    }
    if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
        c->chrDstHSubSample = 1;

    // drop some chroma lines if the user wants it
    c->vChrDrop          = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
                           SWS_SRC_V_CHR_DROP_SHIFT;
    c->chrSrcVSubSample += c->vChrDrop;

    /* drop every other pixel for chroma calculation unless user
     * wants full chroma */
    if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP)   &&
        srcFormat != PIX_FMT_RGB8 && srcFormat != PIX_FMT_BGR8 &&
        srcFormat != PIX_FMT_RGB4 && srcFormat != PIX_FMT_BGR4 &&
        srcFormat != PIX_FMT_RGB4_BYTE && srcFormat != PIX_FMT_BGR4_BYTE &&
        ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
         (flags & SWS_FAST_BILINEAR)))
        c->chrSrcHSubSample = 1;

    // Note the -((-x)>>y) is so that we always round toward +inf.
    c->chrSrcW = -((-srcW) >> c->chrSrcHSubSample);
    c->chrSrcH = -((-srcH) >> c->chrSrcVSubSample);
    c->chrDstW = -((-dstW) >> c->chrDstHSubSample);
    c->chrDstH = -((-dstH) >> c->chrDstVSubSample);

    /* unscaled special cases */
    if (unscaled && !usesHFilter && !usesVFilter &&
        (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
        ff_get_unscaled_swscale(c);

        if (c->swScale) {
            if (flags & SWS_PRINT_INFO)
                av_log(c, AV_LOG_INFO,
                       "using unscaled %s -> %s special converter\n",
                       sws_format_name(srcFormat), sws_format_name(dstFormat));
            return 0;
        }
    }

    c->srcBpc = 1 + av_pix_fmt_descriptors[srcFormat].comp[0].depth_minus1;
    if (c->srcBpc < 8)
        c->srcBpc = 8;
    c->dstBpc = 1 + av_pix_fmt_descriptors[dstFormat].comp[0].depth_minus1;
    if (c->dstBpc < 8)
        c->dstBpc = 8;
    if (c->dstBpc == 16)
        dst_stride <<= 1;
    FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
                     (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
                     fail);
    if (HAVE_MMXEXT && HAVE_INLINE_ASM && cpu_flags & AV_CPU_FLAG_MMXEXT &&
        c->srcBpc == 8 && c->dstBpc <= 10) {
        c->canMMX2BeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
                            (srcW & 15) == 0) ? 1 : 0;
        if (!c->canMMX2BeUsed && dstW >= srcW && (srcW & 15) == 0
            && (flags & SWS_FAST_BILINEAR)) {
            if (flags & SWS_PRINT_INFO)
                av_log(c, AV_LOG_INFO,
                       "output width is not a multiple of 32 -> no MMX2 scaler\n");
        }
        if (usesHFilter)
            c->canMMX2BeUsed = 0;
    } else
        c->canMMX2BeUsed = 0;

    c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
    c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;

    /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
     * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
     * correct scaling.
     * n-2 is the last chrominance sample available.
     * This is not perfect, but no one should notice the difference, the more
     * correct variant would be like the vertical one, but that would require
     * some special code for the first and last pixel */