fbmmx.c

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

    CHECKPOINT();

    fbComposeGetStart (pDst, xDst, yDst, CARD32, dstStride, dstLine, 1);
    fbComposeGetStart (pSrc, xSrc, ySrc, CARD32, srcStride, srcLine, 1);
    fbComposeGetStart (pMask, xMask, yMask, CARD8, maskStride, maskLine, 1);

    mask = *maskLine << 24 | *maskLine << 16 | *maskLine << 8 | *maskLine;
    vmask = load8888 (mask);
    srca = MC(4x00ff);

    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 (in_over (s, srca, vmask, d));

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

	while (w >= 16)
	{
	    __m64 vd0 = *(__m64 *)(dst + 0);
	    __m64 vd1 = *(__m64 *)(dst + 2);
	    __m64 vd2 = *(__m64 *)(dst + 4);
	    __m64 vd3 = *(__m64 *)(dst + 6);
	    __m64 vd4 = *(__m64 *)(dst + 8);
	    __m64 vd5 = *(__m64 *)(dst + 10);
	    __m64 vd6 = *(__m64 *)(dst + 12);
	    __m64 vd7 = *(__m64 *)(dst + 14);

	    __m64 vs0 = *(__m64 *)(src + 0);
	    __m64 vs1 = *(__m64 *)(src + 2);
	    __m64 vs2 = *(__m64 *)(src + 4);
	    __m64 vs3 = *(__m64 *)(src + 6);
	    __m64 vs4 = *(__m64 *)(src + 8);
	    __m64 vs5 = *(__m64 *)(src + 10);
	    __m64 vs6 = *(__m64 *)(src + 12);
	    __m64 vs7 = *(__m64 *)(src + 14);

	    vd0 = (__m64)pack8888 (
		in_over (expand8888 (vs0, 0), srca, vmask, expand8888 (vd0, 0)),
		in_over (expand8888 (vs0, 1), srca, vmask, expand8888 (vd0, 1)));

	    vd1 = (__m64)pack8888 (
		in_over (expand8888 (vs1, 0), srca, vmask, expand8888 (vd1, 0)),
		in_over (expand8888 (vs1, 1), srca, vmask, expand8888 (vd1, 1)));

	    vd2 = (__m64)pack8888 (
		in_over (expand8888 (vs2, 0), srca, vmask, expand8888 (vd2, 0)),
		in_over (expand8888 (vs2, 1), srca, vmask, expand8888 (vd2, 1)));

	    vd3 = (__m64)pack8888 (
		in_over (expand8888 (vs3, 0), srca, vmask, expand8888 (vd3, 0)),
		in_over (expand8888 (vs3, 1), srca, vmask, expand8888 (vd3, 1)));

	    vd4 = (__m64)pack8888 (
		in_over (expand8888 (vs4, 0), srca, vmask, expand8888 (vd4, 0)),
		in_over (expand8888 (vs4, 1), srca, vmask, expand8888 (vd4, 1)));

	    vd5 = (__m64)pack8888 (
		in_over (expand8888 (vs5, 0), srca, vmask, expand8888 (vd5, 0)),
		in_over (expand8888 (vs5, 1), srca, vmask, expand8888 (vd5, 1)));

	    vd6 = (__m64)pack8888 (
		in_over (expand8888 (vs6, 0), srca, vmask, expand8888 (vd6, 0)),
		in_over (expand8888 (vs6, 1), srca, vmask, expand8888 (vd6, 1)));

	    vd7 = (__m64)pack8888 (
		in_over (expand8888 (vs7, 0), srca, vmask, expand8888 (vd7, 0)),
		in_over (expand8888 (vs7, 1), srca, vmask, expand8888 (vd7, 1)));

	    *(__m64 *)(dst + 0) = vd0;
	    *(__m64 *)(dst + 2) = vd1;
	    *(__m64 *)(dst + 4) = vd2;
	    *(__m64 *)(dst + 6) = vd3;
	    *(__m64 *)(dst + 8) = vd4;
	    *(__m64 *)(dst + 10) = vd5;
	    *(__m64 *)(dst + 12) = vd6;
	    *(__m64 *)(dst + 14) = vd7;

	    w -= 16;
	    dst += 16;
	    src += 16;
	}

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

	    *dst = store8888 (in_over (s, srca, vmask, d));

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

    _mm_empty();
}

void
fbCompositeSrc_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, s;
    FbStride	dstStride, srcStride;
    CARD8	a;
    CARD16	w;

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

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

	while (w--)
	{
	    s = *src++;
	    a = s >> 24;
	    if (a == 0xff)
		*dst = s;
	    else if (a) {
		__m64 ms, sa;
		ms = load8888(s);
		sa = expand_alpha(ms);
		*dst = store8888(over(ms, sa, load8888(*dst)));
	    }
	    dst++;
	}
    }
    _mm_empty();
}

void
fbCompositeSolidMask_nx8x8888mmx (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;
    CARD32	*dstLine, *dst;
    CARD8	*maskLine, *mask;
    FbStride	dstStride, maskStride;
    CARD16	w;
    __m64	vsrc, vsrca;
    ullong	srcsrc;

    CHECKPOINT();

    fbComposeGetSolid(pSrc, pDst, src);

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

    srcsrc = (unsigned long long)src << 32 | src;

    fbComposeGetStart (pDst, xDst, yDst, CARD32, dstStride, dstLine, 1);
    fbComposeGetStart (pMask, xMask, yMask, CARD8, maskStride, maskLine, 1);

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

    while (height--)
    {
	dst = dstLine;
	dstLine += dstStride;
	mask = maskLine;
	maskLine += maskStride;
	w = width;

	CHECKPOINT();

	while (w && (unsigned long)dst & 7)
	{
	    ullong m = *mask;

	    if (m)
	    {
		__m64 vdest = in_over(vsrc, vsrca, expand_alpha_rev ((__m64)m), load8888(*dst));
		*dst = store8888(vdest);
	    }

	    w--;
	    mask++;
	    dst++;
	}

	CHECKPOINT();

	while (w >= 2)
	{
	    ullong m0, m1;
	    m0 = *mask;
	    m1 = *(mask + 1);

	    if (srca == 0xff && (m0 & m1) == 0xff)
	    {
		*(unsigned long long *)dst = srcsrc;
	    }
	    else if (m0 | m1)
	    {
		__m64 vdest;
		__m64 dest0, dest1;

		vdest = *(__m64 *)dst;

		dest0 = in_over(vsrc, vsrca, expand_alpha_rev ((__m64)m0), expand8888(vdest, 0));
		dest1 = in_over(vsrc, vsrca, expand_alpha_rev ((__m64)m1), expand8888(vdest, 1));

		*(__m64 *)dst = pack8888(dest0, dest1);
	    }

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

	CHECKPOINT();

	while (w)
	{
	    ullong m = *mask;

	    if (m)
	    {
		__m64 vdest = load8888(*dst);
		vdest = in_over(vsrc, vsrca, expand_alpha_rev ((__m64)m), vdest);
		*dst = store8888(vdest);
	    }

	    w--;
	    mask++;
	    dst++;
	}
    }

    _mm_empty();
}

void
fbCompositeSolidMaskSrc_nx8x8888mmx (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;
    CARD32	*dstLine, *dst;
    CARD8	*maskLine, *mask;
    FbStride	dstStride, maskStride;
    CARD16	w;
    __m64	vsrc, vsrca;
    ullong	srcsrc;

    CHECKPOINT();

    fbComposeGetSolid(pSrc, pDst, src);

    srca = src >> 24;
    if (srca == 0)
    {
	if (fbSolidFillmmx (pDst->pDrawable, xDst, yDst, width, height, 0))
	    return;
    }

    srcsrc = (unsigned long long)src << 32 | src;

    fbComposeGetStart (pDst, xDst, yDst, CARD32, dstStride, dstLine, 1);
    fbComposeGetStart (pMask, xMask, yMask, CARD8, maskStride, maskLine, 1);

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

    while (height--)
    {
	dst = dstLine;
	dstLine += dstStride;
	mask = maskLine;
	maskLine += maskStride;
	w = width;

	CHECKPOINT();

	while (w && (unsigned long)dst & 7)
	{
	    ullong m = *mask;

	    if (m)
	    {
		  __m64 vdest = in(vsrc, expand_alpha_rev ((__m64)m));
		  *dst = store8888(vdest);
	    }
	    else
	    {
		  *dst = 0;
	    }

	    w--;
	    mask++;
	    dst++;
	}

	CHECKPOINT();

	while (w >= 2)
	{
	    ullong m0, m1;
	    m0 = *mask;
	    m1 = *(mask + 1);

	    if (srca == 0xff && (m0 & m1) == 0xff)
	    {
		*(unsigned long long *)dst = srcsrc;
	    }
	    else if (m0 | m1)
	    {
		__m64 vdest;
		__m64 dest0, dest1;

		vdest = *(__m64 *)dst;

		dest0 = in(vsrc, expand_alpha_rev ((__m64)m0));
		dest1 = in(vsrc, expand_alpha_rev ((__m64)m1));

		*(__m64 *)dst = pack8888(dest0, dest1);
	    }
	    else
	    {
		*dst = 0;
	    }

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

	CHECKPOINT();

	while (w)
	{
	    ullong m = *mask;

	    if (m)
	    {
		__m64 vdest = load8888(*dst);
		vdest = in(vsrc, expand_alpha_rev ((__m64)m));
		*dst = store8888(vdest);
	    }
	    else
	    {
		*dst = 0;
	    }

	    w--;
	    mask++;
	    dst++;
	}
    }

    _mm_empty();
}

void
fbCompositeSolidMask_nx8x0565mmx (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, *dst;
    CARD8	*maskLine, *mask;
    FbStride	dstStride, maskStride;
    CARD16	w;
    __m64	vsrc, vsrca;
    unsigned long long srcsrcsrcsrc, src16;

    CHECKPOINT();

    fbComposeGetSolid(pSrc, pDst, src);

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

    fbComposeGetStart (pDst, xDst, yDst, CARD16, dstStride, dstLine, 1);
    fbComposeGetStart (pMask, xMask, yMask, CARD8, maskStride, maskLine, 1);

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

    src16 = (ullong)pack565(vsrc, _mm_setzero_si64(), 0);

    srcsrcsrcsrc = (ullong)src16 << 48 | (ullong)src16 << 32 |
	(ullong)src16 << 16 | (ullong)src16;

    while (height--)
    {
	dst = dstLine;
	dstLine += dstStride;
	mask = maskLine;
	maskLine += maskStride;
	w = width;

	CHECKPOINT();

	while (w && (unsigned long)dst & 7)
	{
	    ullong m = *mask;

	    if (m)
	    {
		ullong d = *dst;
		__m64 vd = (__m64)d;
		__m64 vdest = in_over(vsrc, vsrca, expand_alpha_rev ((__m64)m), expand565(vd, 0));
		*dst = (ullong)pack565(vdest, _mm_setzero_si64(), 0);
	    }

	    w--;
	    mask++;
	    dst++;
	}

	CHECKPOINT();

	while (w >= 4)
	{
	    ullong m0, m1, m2, m3;
	    m0 = *mask;
	    m1 = *(mask + 1);
	    m2 = *(mask + 2);
	    m3 = *(mask + 3);

	    if (srca == 0xff && (m0 & m1 & m2 & m3) == 0xff)
	    {
		*(unsigned long long *)dst = srcsrcsrcsrc;
	    }
	    else if (m0 | m1 | m2 | m3)
	    {
		__m64 vdest;
		__m64 vm0, vm1, vm2, vm3;

		vdest = *(__m64 *)dst;

		vm0 = (__m64)m0;
		vdest = pack565(in_over(vsrc, vsrca, expand_alpha_rev(vm0), expand565(vdest, 0)), vdest, 0);
		vm1 = (__m64)m1;
		vdest = pack565(in_over(vsrc, vsrca, expand_alpha_rev(vm1), expand565(vdest, 1)), vdest, 1);
		vm2 = (__m64)m2;
		vdest = pack565(in_over(vsrc, vsrca, expand_alpha_rev(vm2), expand565(vdest, 2)), vdest, 2);
		vm3 = (__m64)m3;
		vdest = pack565(in_over(vsrc, vsrca, expand_alpha_rev(vm3), expand565(vdest, 3)), vdest, 3);

		*(__m64 *)dst = vdest;
	    }

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

	CHECKPOINT();

	while (w)
	{
	    ullong m = *mask;

	    if (m)
	    {
		ullong d = *dst;
		__m64 vd = (__m64)d;
		__m64 vdest = in_over(vsrc, vsrca, expand_alpha_rev ((__m64)m), expand565(vd, 0));
		*dst = (ullong)pack565(vdest, _mm_setzero_si64(), 0);
	    }

	    w--;
	    mask++;
	    dst++;
	}
    }

    _mm_empty();
}

void
fbCompositeSrc_8888RevNPx0565mmx (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)
{
    CARD16	*dstLine, *dst;
    CARD32	*srcLine, *src;
    FbStride	dstStride, srcStride;
    CARD16	w;

    CHECKPOINT();

    fbComposeGetStart (pDst, xDst, yDst, CARD16, 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;

	CHECKPOINT();

	while (w && (unsigned long)dst & 7)
	{
	    __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++;
	}

	CHECKPOINT();

	while (w >= 4)
	{
	    CARD32 s0, s1, s2, s3;
	    unsigned char a0, a1, a2, a3;

	    s0 = *src;
	    s1 = *(src + 1);
	    s2 = *(src + 2);
	    s3 = *(src + 3);

	    a0 = (s0 >> 24);
	    a1 = (s1 >> 24);
	    a2 = (s2 >> 24);
	    a3 = (s3 >> 24);

	    if ((a0 & a1 & a2 & a3) == 0xFF)
	    {
		__m64 vdest;
		vdest = pack565(invert_colors(load8888(s0)), _mm_setzero_si64(), 0);
		vdest = pack565(invert_colors(load8888(s1)), vdest, 1);
		vdest = pack565(invert_colors(load8888(s2)), vdest, 2);
		vdest = pack565(invert_colors(load8888(s3)), vdest, 3);

		*(__m64 *)dst = vdest;
	    }
	    else if (a0 | a1 | a2 | a3)
	    {
		__m64 vdest = *(__m64 *)dst;