fbmmx.c

按照官方的说法：Cairo is a vector graphics library with cross-device output support. 翻译过来

		vdest = pack565(over_rev_non_pre(load8888(s0), expand565(vdest, 0)), vdest, 0);
	        vdest = pack565(over_rev_non_pre(load8888(s1), expand565(vdest, 1)), vdest, 1);
		vdest = pack565(over_rev_non_pre(load8888(s2), expand565(vdest, 2)), vdest, 2);
		vdest = pack565(over_rev_non_pre(load8888(s3), expand565(vdest, 3)), vdest, 3);

		*(__m64 *)dst = vdest;
	    }

	    w -= 4;
	    dst += 4;
	    src += 4;
	}

	CHECKPOINT();

	while (w)
	{
	    __m64 vsrc = load8888 (*src);
	    ullong d = *dst;
	    __m64 vdest = expand565 ((__m64)d, 0);

	    vdest = pack565(over_rev_non_pre(vsrc, vdest), vdest, 0);

	    *dst = (ullong)vdest;

	    w--;
	    dst++;
	    src++;
	}
    }

    _mm_empty();
}

/* "8888RevNP" is GdkPixbuf's format: ABGR, non premultiplied */

void
fbCompositeSrc_8888RevNPx8888mmx (pixman_operator_t      op,
				  PicturePtr pSrc,
				  PicturePtr pMask,
				  PicturePtr pDst,
				  INT16      xSrc,
				  INT16      ySrc,
				  INT16      xMask,
				  INT16      yMask,
				  INT16      xDst,
				  INT16      yDst,
				  CARD16     width,
				  CARD16     height)
{
    CARD32	*dstLine, *dst;
    CARD32	*srcLine, *src;
    FbStride	dstStride, srcStride;
    CARD16	w;

    CHECKPOINT();

    fbComposeGetStart (pDst, xDst, yDst, CARD32, dstStride, dstLine, 1);
    fbComposeGetStart (pSrc, xSrc, ySrc, CARD32, srcStride, srcLine, 1);

    assert (pSrc->pDrawable == pMask->pDrawable);

    while (height--)
    {
	dst = dstLine;
	dstLine += dstStride;
	src = srcLine;
	srcLine += srcStride;
	w = width;

	while (w && (unsigned long)dst & 7)
	{
	    __m64 s = load8888 (*src);
	    __m64 d = load8888 (*dst);

	    *dst = store8888 (over_rev_non_pre (s, d));

	    w--;
	    dst++;
	    src++;
	}

	while (w >= 2)
	{
	    ullong s0, s1;
	    unsigned char a0, a1;
	    __m64 d0, d1;

	    s0 = *src;
	    s1 = *(src + 1);

	    a0 = (s0 >> 24);
	    a1 = (s1 >> 24);

	    if ((a0 & a1) == 0xFF)
	    {
		d0 = invert_colors(load8888(s0));
		d1 = invert_colors(load8888(s1));

		*(__m64 *)dst = pack8888 (d0, d1);
	    }
	    else if (a0 | a1)
	    {
		__m64 vdest = *(__m64 *)dst;

		d0 = over_rev_non_pre (load8888(s0), expand8888 (vdest, 0));
		d1 = over_rev_non_pre (load8888(s1), expand8888 (vdest, 1));

		*(__m64 *)dst = pack8888 (d0, d1);
	    }

	    w -= 2;
	    dst += 2;
	    src += 2;
	}

	while (w)
	{
	    __m64 s = load8888 (*src);
	    __m64 d = load8888 (*dst);

	    *dst = store8888 (over_rev_non_pre (s, d));

	    w--;
	    dst++;
	    src++;
	}
    }

    _mm_empty();
}

void
fbCompositeSolidMask_nx8888x0565Cmmx (pixman_operator_t      op,
				      PicturePtr pSrc,
				      PicturePtr pMask,
				      PicturePtr pDst,
				      INT16      xSrc,
				      INT16      ySrc,
				      INT16      xMask,
				      INT16      yMask,
				      INT16      xDst,
				      INT16      yDst,
				      CARD16     width,
				      CARD16     height)
{
    CARD32	src, srca;
    CARD16	*dstLine;
    CARD32	*maskLine;
    FbStride	dstStride, maskStride;
    __m64  vsrc, vsrca;

    CHECKPOINT();

    fbComposeGetSolid(pSrc, pDst, src);

    srca = src >> 24;
    if (srca == 0)
	return;

    fbComposeGetStart (pDst, xDst, yDst, CARD16, dstStride, dstLine, 1);
    fbComposeGetStart (pMask, xMask, yMask, CARD32, maskStride, maskLine, 1);

    vsrc = load8888 (src);
    vsrca = expand_alpha (vsrc);

    while (height--)
    {
	int twidth = width;
	CARD32 *p = (CARD32 *)maskLine;
	CARD16 *q = (CARD16 *)dstLine;

	while (twidth && ((unsigned long)q & 7))
	{
	    CARD32 m = *(CARD32 *)p;

	    if (m)
	    {
		ullong d = *q;
		__m64 vdest = expand565 ((__m64)d, 0);
		vdest = pack565 (in_over (vsrc, vsrca, load8888 (m), vdest), vdest, 0);
		*q = (ullong)vdest;
	    }

	    twidth--;
	    p++;
	    q++;
	}

	while (twidth >= 4)
	{
	    CARD32 m0, m1, m2, m3;

	    m0 = *p;
	    m1 = *(p + 1);
	    m2 = *(p + 2);
	    m3 = *(p + 3);

	    if ((m0 | m1 | m2 | m3))
	    {
		__m64 vdest = *(__m64 *)q;

		vdest = pack565(in_over(vsrc, vsrca, load8888(m0), expand565(vdest, 0)), vdest, 0);
		vdest = pack565(in_over(vsrc, vsrca, load8888(m1), expand565(vdest, 1)), vdest, 1);
		vdest = pack565(in_over(vsrc, vsrca, load8888(m2), expand565(vdest, 2)), vdest, 2);
		vdest = pack565(in_over(vsrc, vsrca, load8888(m3), expand565(vdest, 3)), vdest, 3);

		*(__m64 *)q = vdest;
	    }
	    twidth -= 4;
	    p += 4;
	    q += 4;
	}

	while (twidth)
	{
	    CARD32 m;

	    m = *(CARD32 *)p;
	    if (m)
	    {
		ullong d = *q;
		__m64 vdest = expand565((__m64)d, 0);
		vdest = pack565 (in_over(vsrc, vsrca, load8888(m), vdest), vdest, 0);
		*q = (ullong)vdest;
	    }

	    twidth--;
	    p++;
	    q++;
	}

	maskLine += maskStride;
	dstLine += dstStride;
    }

    _mm_empty ();
}

void
fbCompositeSrcAdd_8000x8000mmx (pixman_operator_t	op,
				PicturePtr pSrc,
				PicturePtr pMask,
				PicturePtr pDst,
				INT16      xSrc,
				INT16      ySrc,
				INT16      xMask,
				INT16      yMask,
				INT16      xDst,
				INT16      yDst,
				CARD16     width,
				CARD16     height)
{
    CARD8	*dstLine, *dst;
    CARD8	*srcLine, *src;
    FbStride	dstStride, srcStride;
    CARD16	w;
    CARD8	s, d;
    CARD16	t;

    CHECKPOINT();

    fbComposeGetStart (pSrc, xSrc, ySrc, CARD8, srcStride, srcLine, 1);
    fbComposeGetStart (pDst, xDst, yDst, CARD8, dstStride, dstLine, 1);

    while (height--)
    {
	dst = dstLine;
	dstLine += dstStride;
	src = srcLine;
	srcLine += srcStride;
	w = width;

	while (w && (unsigned long)dst & 7)
	{
	    s = *src;
	    d = *dst;
	    t = d + s;
	    s = t | (0 - (t >> 8));
	    *dst = s;

	    dst++;
	    src++;
	    w--;
	}

	while (w >= 8)
	{
	    *(__m64*)dst = _mm_adds_pu8(*(__m64*)src, *(__m64*)dst);
	    dst += 8;
	    src += 8;
	    w -= 8;
	}

	while (w)
	{
	    s = *src;
	    d = *dst;
	    t = d + s;
	    s = t | (0 - (t >> 8));
	    *dst = s;

	    dst++;
	    src++;
	    w--;
	}
    }

    _mm_empty();
}

void
fbCompositeSrcAdd_8888x8888mmx (pixman_operator_t		op,
				PicturePtr	pSrc,
				PicturePtr	pMask,
				PicturePtr	 pDst,
				INT16		 xSrc,
				INT16      ySrc,
				INT16      xMask,
				INT16      yMask,
				INT16      xDst,
				INT16      yDst,
				CARD16     width,
				CARD16     height)
{
    CARD32	*dstLine, *dst;
    CARD32	*srcLine, *src;
    FbStride	dstStride, srcStride;
    CARD16	w;

    CHECKPOINT();

    fbComposeGetStart (pSrc, xSrc, ySrc, CARD32, srcStride, srcLine, 1);
    fbComposeGetStart (pDst, xDst, yDst, CARD32, dstStride, dstLine, 1);

    while (height--)
    {
	dst = dstLine;
	dstLine += dstStride;
	src = srcLine;
	srcLine += srcStride;
	w = width;

	while (w && (unsigned long)dst & 7)
	{
	    *dst = _mm_cvtsi64_si32(_mm_adds_pu8(_mm_cvtsi32_si64(*src),
						 _mm_cvtsi32_si64(*dst)));
	    dst++;
	    src++;
	    w--;
	}

	while (w >= 2)
	{
	    *(ullong*)dst = (ullong) _mm_adds_pu8(*(__m64*)src, *(__m64*)dst);
	    dst += 2;
	    src += 2;
	    w -= 2;
	}

	if (w)
	{
	    *dst = _mm_cvtsi64_si32(_mm_adds_pu8(_mm_cvtsi32_si64(*src),
						 _mm_cvtsi32_si64(*dst)));

	}
    }

    _mm_empty();
}

Bool
fbSolidFillmmx (FbPixels	*pDraw,
		int		x,
		int		y,
		int		width,
		int		height,
		FbBits		xor)
{
    FbStride	stride;
    int		bpp;
    ullong	fill;
    __m64	vfill;
    CARD32	byte_width;
    CARD8	*byte_line;
    FbBits      *bits;
    int		xoff, yoff;

    CHECKPOINT();

    fbGetDrawable(pDraw, bits, stride, bpp, xoff, yoff);

    if (bpp == 16 && (xor >> 16 != (xor & 0xffff)))
	return FALSE;

    if (bpp != 16 && bpp != 32)
	return FALSE;

    if (bpp == 16)
    {
	stride = stride * sizeof (FbBits) / 2;
	byte_line = (CARD8 *)(((CARD16 *)bits) + stride * (y + yoff) + (x + xoff));
	byte_width = 2 * width;
	stride *= 2;
    }
    else
    {
	stride = stride * sizeof (FbBits) / 4;
	byte_line = (CARD8 *)(((CARD32 *)bits) + stride * (y + yoff) + (x + xoff));
	byte_width = 4 * width;
	stride *= 4;
    }

    fill = ((ullong)xor << 32) | xor;
    vfill = (__m64)fill;

    while (height--)
    {
	int w;
	CARD8 *d = byte_line;
	byte_line += stride;
	w = byte_width;

	while (w >= 2 && ((unsigned long)d & 3))
	{
	    *(CARD16 *)d = xor;
	    w -= 2;
	    d += 2;
	}

	while (w >= 4 && ((unsigned long)d & 7))
	{
	    *(CARD32 *)d = xor;

	    w -= 4;
	    d += 4;
	}

	while (w >= 64)
	{
	    *(__m64*) (d +  0) = vfill;
	    *(__m64*) (d +  8) = vfill;
	    *(__m64*) (d + 16) = vfill;
	    *(__m64*) (d + 24) = vfill;
	    *(__m64*) (d + 32) = vfill;
	    *(__m64*) (d + 40) = vfill;
	    *(__m64*) (d + 48) = vfill;
	    *(__m64*) (d + 56) = vfill;

	    w -= 64;
	    d += 64;
	}
	while (w >= 4)
	{
	    *(CARD32 *)d = xor;

	    w -= 4;
	    d += 4;
	}
	if (w >= 2)
	{
	    *(CARD16 *)d = xor;
	    w -= 2;
	    d += 2;
	}
    }

    _mm_empty();
    return TRUE;
}

Bool
fbCopyAreammx (FbPixels	*pSrc,
	       FbPixels	*pDst,
	       int		src_x,
	       int		src_y,
	       int		dst_x,
	       int		dst_y,
	       int		width,
	       int		height)
{
    FbBits *	src_bits;
    FbStride	src_stride;
    int		src_bpp;
    int		src_xoff;
    int		src_yoff;

    FbBits *	dst_bits;
    FbStride	dst_stride;
    int		dst_bpp;
    int		dst_xoff;
    int		dst_yoff;

    CARD8 *	src_bytes;
    CARD8 *	dst_bytes;
    int		byte_width;

    fbGetDrawable(pSrc, src_bits, src_stride, src_bpp, src_xoff, src_yoff);
    fbGetDrawable(pDst, dst_bits, dst_stride, dst_bpp, dst_xoff, dst_yoff);

    if (src_bpp != 16 && src_bpp != 32)
	return FALSE;

    if (dst_bpp != 16 && dst_bpp != 32)
	return FALSE;

    if (src_bpp != dst_bpp)
    {
	return FALSE;
    }

    if (src_bpp == 16)
    {
	src_stride = src_stride * sizeof (FbBits) / 2;
	dst_stride = dst_stride * sizeof (FbBits) / 2;
	src_bytes = (CARD8 *)(((CARD16 *)src_bits) + src_stride * (src_y + src_yoff) + (src_x + src_xoff));
	dst_bytes = (CARD8 *)(((CARD16 *)dst_bits) + dst_stride * (dst_y + dst_yoff) + (dst_x + dst_xoff));
	byte_width = 2 * width;
	src_stride *= 2;
	dst_stride *= 2;
    }
    else
    {
	src_stride = src_stride * sizeof (FbBits) / 4;
	dst_stride = dst_stride * sizeof (FbBits) / 4;
	src_bytes = (CARD8 *)(((CARD32 *)src_bits) + src_stride * (src_y + src_yoff) + (src_x + src_xoff));
	dst_bytes = (CARD8 *)(((CARD32 *)dst_bits) + dst_stride * (dst_y + dst_yoff) + (dst_x + dst_xoff));
	byte_width = 4 * width;
	src_stride *= 4;
	dst_stride *= 4;
    }

    while (height--)
    {
	int w;
	CARD8 *s = src_bytes;
	CARD8 *d = dst_bytes;
	src_bytes += src_stride;
	dst_bytes += dst_stride;
	w = byte_width;

	while (w >= 2 && ((unsigned long)d & 3))
	{
	    *(CARD16 *)d = *(CARD16 *)s;
	    w -= 2;
	    s += 2;
	    d += 2;
	}

	while (w >= 4 && ((unsigned int)d & 7))
	{
	    *(CARD32 *)d = *(CARD32 *)s;

	    w -= 4;
	    s += 4;
	    d += 4;
	}

	while (w >= 64)
	{
	    *(__m64 *)(d + 0)  = *(__m64 *)(s + 0);
	    *(__m64 *)(d + 8)  = *(__m64 *)(s + 8);
	    *(__m64 *)(d + 16) = *(__m64 *)(s + 16);
	    *(__m64 *)(d + 24) = *(__m64 *)(s + 24);
	    *(__m64 *)(d + 32) = *(__m64 *)(s + 32);
	    *(__m64 *)(d + 40) = *(__m64 *)(s + 40);
	    *(__m64 *)(d + 48) = *(__m64 *)(s + 48);
	    *(__m64 *)(d + 56) = *(__m64 *)(s + 56);
	    w -= 64;
	    s += 64;
	    d += 64;
	}
	while (w >= 4)
	{
	    *(CARD32 *)d = *(CARD32 *)s;

	    w -= 4;
	    s += 4;
	    d += 4;
	}
	if (w >= 2)
	{
	    *(CARD16 *)d = *(CARD16 *)s;
	    w -= 2;
	    s += 2;
	    d += 2;
	}
    }

    _mm_empty();
    return TRUE;
}

void
fbCompositeCopyAreammx (pixman_operator_t		op,
			PicturePtr	pSrc,
			PicturePtr	pMask,
			PicturePtr	pDst,
			INT16		xSrc,
			INT16		ySrc,
			INT16		xMask,
			INT16		yMask,
			INT16		xDst,
			INT16		yDst,
			CARD16		width,
			CARD16		height)
{
    fbCopyAreammx (pSrc->pDrawable,
		   pDst->pDrawable,
		   xSrc, ySrc,
		   xDst, yDst,
		   width, height);
}

#endif /* RENDER */