[FFmpeg-devel] Yellow fever

[Måns Rullgård]

As you all have no doubt completely ignored, FATE turned much more
yellow yesterday.  The reason for this is that I disabled unaligned
memory access fixup on those machines which do not support it in
hardware (Alpha, ARMv5, MIPS).  Most of these are failing a single
test: lavfi-pixfmts_scale_le.

The problem here is btr32ToY() and related functions in libswscale:

#define BGR2Y(type, name, shr, shg, shb, maskr, maskg, maskb, RY, GY, BY, S)\
static inline void name(uint8_t *dst, const uint8_t *src, long width, uint32_t *unused)\
{\
    int i;\
    for (i=0; i<width; i++) {\
        int b= (((const type*)src)[i]>>shb)&maskb;\
        int g= (((const type*)src)[i]>>shg)&maskg;\
        int r= (((const type*)src)[i]>>shr)&maskr;\
\
        dst[i]= (((RY)*r + (GY)*g + (BY)*b + (33<<((S)-1)))>>(S));\
    }\
}

BGR2Y(uint32_t, bgr32ToY,16, 0, 0, 0x00FF, 0xFF00, 0x00FF, RY<< 8, GY   , BY<< 8, RGB2YUV_SHIFT+8)
BGR2Y(uint32_t, rgb32ToY, 0, 0,16, 0x00FF, 0xFF00, 0x00FF, RY<< 8, GY   , BY<< 8, RGB2YUV_SHIFT+8)
BGR2Y(uint16_t, bgr16ToY, 0, 0, 0, 0x001F, 0x07E0, 0xF800, RY<<11, GY<<5, BY    , RGB2YUV_SHIFT+8)
BGR2Y(uint16_t, bgr15ToY, 0, 0, 0, 0x001F, 0x03E0, 0x7C00, RY<<10, GY<<5, BY    , RGB2YUV_SHIFT+7)
BGR2Y(uint16_t, rgb16ToY, 0, 0, 0, 0xF800, 0x07E0, 0x001F, RY    , GY<<5, BY<<11, RGB2YUV_SHIFT+8)
BGR2Y(uint16_t, rgb15ToY, 0, 0, 0, 0x7C00, 0x03E0, 0x001F, RY    , GY<<5, BY<<10, RGB2YUV_SHIFT+7)

These functions are called from hyscale(), like this:

static inline void RENAME(hyscale)(SwsContext *c, uint16_t *dst, long dstWidth, const uint8_t *src, int srcW, int xInc,
                                   const int16_t *hLumFilter,
                                   const int16_t *hLumFilterPos, int hLumFilterSize,
                                   uint8_t *formatConvBuffer,
                                   uint32_t *pal, int isAlpha)
{
    void (*toYV12)(uint8_t *, const uint8_t *, long, uint32_t *) = isAlpha ? c->alpToYV12 : c->lumToYV12;
    void (*convertRange)(uint16_t *, int) = isAlpha ? NULL : c->lumConvertRange;

    src += isAlpha ? c->alpSrcOffset : c->lumSrcOffset;

    if (toYV12) {
        toYV12(formatConvBuffer, src, srcW, pal);
        src= formatConvBuffer;
    }
[...]
}

c->lumSrcOffset is either 1 or -1, meaning src will _always_ be
unaligned.  Clearly something must be done about this.  The trivial
fix would be to use AV_RN32() in the BGR2Y functions or change them to
use byte loads.  Better would be to make whatever changes are needed
to keep the pointers aligned, though I have no idea what that would
involve.

Comments?

-- 
M?ns Rullg?rd
mans at mansr.com