fbmmx.c

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

/*
 * Copyright © 2004 Red Hat, Inc.
 * Copyright © 2004 Nicholas Miell
 * Copyright © 2005 Trolltech AS
 *
 * Permission to use, copy, modify, distribute, and sell this software and its
 * documentation for any purpose is hereby granted without fee, provided that
 * the above copyright notice appear in all copies and that both that
 * copyright notice and this permission notice appear in supporting
 * documentation, and that the name of Red Hat not be used in advertising or
 * publicity pertaining to distribution of the software without specific,
 * written prior permission.  Red Hat makes no representations about the
 * suitability of this software for any purpose.  It is provided "as is"
 * without express or implied warranty.
 *
 * THE COPYRIGHT HOLDERS DISCLAIM ALL WARRANTIES WITH REGARD TO THIS
 * SOFTWARE, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND
 * FITNESS, IN NO EVENT SHALL THE COPYRIGHT HOLDERS BE LIABLE FOR ANY
 * SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN
 * AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING
 * OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS
 * SOFTWARE.
 *
 * Author:  Søren Sandmann ([e-mail])
 * Minor Improvements: Nicholas Miell ([e-mail])
 * MMX code paths for fbcompose.c by Lars Knoll ([e-mail])
 *
 * Based on work by Owen Taylor
 */

#ifdef HAVE_CONFIG_H
#include <config.h>
#endif

#include <assert.h>
#include "fbpict.h"
#include "pixman-xserver-compat.h"
#include "fbmmx.h"

#if defined(__amd64__) || defined(__x86_64__)
#define USE_SSE
#endif

#include <mmintrin.h>
#ifdef USE_SSE
#include <xmmintrin.h> /* for _mm_shuffle_pi16 and _MM_SHUFFLE */
#endif

#ifdef RENDER

#include "fbpict.h"

#define noVERBOSE

#ifdef VERBOSE
#define CHECKPOINT() ErrorF ("at %s %d\n", __FUNCTION__, __LINE__)
#else
#define CHECKPOINT()
#endif

/* Notes about writing mmx code
 *
 * give memory operands as the second operand. If you give it as the
 * first, gcc will first load it into a register, then use that
 * register
 *
 *   ie. use
 *
 *         _mm_mullo_pi16 (x, mmx_constant);
 *
 *   not
 *
 *         _mm_mullo_pi16 (mmx_constant, x);
 *
 * Also try to minimize dependencies. i.e. when you need a value, try
 * to calculate it from a value that was calculated as early as
 * possible.
 */

/* --------------- MMX primitivess ------------------------------------ */

typedef unsigned long long ullong;

typedef struct
{
    ullong mmx_4x00ff;
    ullong mmx_4x0080;
    ullong mmx_565_rgb;
    ullong mmx_565_unpack_multiplier;
    ullong mmx_565_r;
    ullong mmx_565_g;
    ullong mmx_565_b;
    ullong mmx_mask_0;
    ullong mmx_mask_1;
    ullong mmx_mask_2;
    ullong mmx_mask_3;
    ullong mmx_full_alpha;
    ullong mmx_ffff0000ffff0000;
    ullong mmx_0000ffff00000000;
    ullong mmx_000000000000ffff;
} MMXData;

static const MMXData c =
{
    .mmx_4x00ff =			0x00ff00ff00ff00ffULL,
    .mmx_4x0080 =			0x0080008000800080ULL,
    .mmx_565_rgb =			0x000001f0003f001fULL,
    .mmx_565_r =			0x000000f800000000ULL,
    .mmx_565_g =			0x0000000000fc0000ULL,
    .mmx_565_b =			0x00000000000000f8ULL,
    .mmx_mask_0 =			0xffffffffffff0000ULL,
    .mmx_mask_1 =			0xffffffff0000ffffULL,
    .mmx_mask_2 =			0xffff0000ffffffffULL,
    .mmx_mask_3 =			0x0000ffffffffffffULL,
    .mmx_full_alpha =			0x00ff000000000000ULL,
    .mmx_565_unpack_multiplier =	0x0000008404100840ULL,
    .mmx_ffff0000ffff0000 =		0xffff0000ffff0000ULL,
    .mmx_0000ffff00000000 =		0x0000ffff00000000ULL,
    .mmx_000000000000ffff =		0x000000000000ffffULL,
};

#define MC(x) ((__m64) c.mmx_##x)

static __inline__ __m64
shift (__m64 v, int s)
{
    if (s > 0)
	return _mm_slli_si64 (v, s);
    else if (s < 0)
	return _mm_srli_si64 (v, -s);
    else
	return v;
}

static __inline__ __m64
negate (__m64 mask)
{
    return _mm_xor_si64 (mask, MC(4x00ff));
}

static __inline__ __m64
pix_multiply (__m64 a, __m64 b)
{
    __m64 res;

    res = _mm_mullo_pi16 (a, b);
    res = _mm_adds_pu16 (res, MC(4x0080));
    res = _mm_adds_pu16 (res, _mm_srli_pi16 (res, 8));
    res = _mm_srli_pi16 (res, 8);

    return res;
}

static __inline__ __m64
pix_add (__m64 a, __m64 b)
{
    return  _mm_adds_pu8 (a, b);
}

#ifdef USE_SSE

static __inline__ __m64
expand_alpha (__m64 pixel)
{
    return _mm_shuffle_pi16 (pixel, _MM_SHUFFLE(3, 3, 3, 3));
}

static __inline__ __m64
expand_alpha_rev (__m64 pixel)
{
    return _mm_shuffle_pi16 (pixel, _MM_SHUFFLE(0, 0, 0, 0));
}

static __inline__ __m64
invert_colors (__m64 pixel)
{
    return _mm_shuffle_pi16 (pixel, _MM_SHUFFLE(3, 0, 1, 2));
}

#else

static __inline__ __m64
expand_alpha (__m64 pixel)
{
    __m64 t1, t2;

    t1 = shift (pixel, -48);
    t2 = shift (t1, 16);
    t1 = _mm_or_si64 (t1, t2);
    t2 = shift (t1, 32);
    t1 = _mm_or_si64 (t1, t2);

    return t1;
}

static __inline__ __m64
expand_alpha_rev (__m64 pixel)
{
    __m64 t1, t2;

    /* move alpha to low 16 bits and zero the rest */
    t1 = shift (pixel,  48);
    t1 = shift (t1, -48);

    t2 = shift (t1, 16);
    t1 = _mm_or_si64 (t1, t2);
    t2 = shift (t1, 32);
    t1 = _mm_or_si64 (t1, t2);

    return t1;
}

static __inline__ __m64
invert_colors (__m64 pixel)
{
    __m64 x, y, z;

    x = y = z = pixel;

    x = _mm_and_si64 (x, MC(ffff0000ffff0000));
    y = _mm_and_si64 (y, MC(000000000000ffff));
    z = _mm_and_si64 (z, MC(0000ffff00000000));

    y = shift (y, 32);
    z = shift (z, -32);

    x = _mm_or_si64 (x, y);
    x = _mm_or_si64 (x, z);

    return x;
}

#endif

static __inline__ __m64
over (__m64 src, __m64 srca, __m64 dest)
{
    return  _mm_adds_pu8 (src, pix_multiply(dest, negate(srca)));
}

static __inline__ __m64
over_rev_non_pre (__m64 src, __m64 dest)
{
    __m64 srca = expand_alpha (src);
    __m64 srcfaaa = _mm_or_si64 (srca, MC(full_alpha));

    return over(pix_multiply(invert_colors(src), srcfaaa), srca, dest);
}

static __inline__ __m64
in (__m64 src,
    __m64 mask)
{
    return pix_multiply (src, mask);
}

static __inline__ __m64
in_over (__m64 src,
	 __m64 srca,
	 __m64 mask,
	 __m64 dest)
{
    return over(in(src, mask), pix_multiply(srca, mask), dest);
}

static __inline__ __m64
load8888 (CARD32 v)
{
    return _mm_unpacklo_pi8 (_mm_cvtsi32_si64 (v), _mm_setzero_si64());
}

static __inline__ __m64
pack8888 (__m64 lo, __m64 hi)
{
    return _mm_packs_pu16 (lo, hi);
}

static __inline__ CARD32
store8888 (__m64 v)
{
    return _mm_cvtsi64_si32(pack8888(v, _mm_setzero_si64()));
}

/* Expand 16 bits positioned at @pos (0-3) of a mmx register into
 *
 *    00RR00GG00BB
 *
 * --- Expanding 565 in the low word ---
 *
 * m = (m << (32 - 3)) | (m << (16 - 5)) | m;
 * m = m & (01f0003f001f);
 * m = m * (008404100840);
 * m = m >> 8;
 *
 * Note the trick here - the top word is shifted by another nibble to
 * avoid it bumping into the middle word
 */
static __inline__ __m64
expand565 (__m64 pixel, int pos)
{
    __m64 p = pixel;
    __m64 t1, t2;

    /* move pixel to low 16 bit and zero the rest */
    p = shift (shift (p, (3 - pos) * 16), -48);

    t1 = shift (p, 36 - 11);
    t2 = shift (p, 16 - 5);

    p = _mm_or_si64 (t1, p);
    p = _mm_or_si64 (t2, p);
    p = _mm_and_si64 (p, MC(565_rgb));

    pixel = _mm_mullo_pi16 (p, MC(565_unpack_multiplier));
    return _mm_srli_pi16 (pixel, 8);
}

static __inline__ __m64
expand8888 (__m64 in, int pos)
{
    if (pos == 0)
	return _mm_unpacklo_pi8 (in, _mm_setzero_si64());
    else
	return _mm_unpackhi_pi8 (in, _mm_setzero_si64());
}

static __inline__ __m64
pack565 (__m64 pixel, __m64 target, int pos)
{
    __m64 p = pixel;
    __m64 t = target;
    __m64 r, g, b;

    r = _mm_and_si64 (p, MC(565_r));
    g = _mm_and_si64 (p, MC(565_g));
    b = _mm_and_si64 (p, MC(565_b));

    r = shift (r, - (32 - 8) + pos * 16);
    g = shift (g, - (16 - 3) + pos * 16);
    b = shift (b, - (0  + 3) + pos * 16);

    if (pos == 0)
	t = _mm_and_si64 (t, MC(mask_0));
    else if (pos == 1)
	t = _mm_and_si64 (t, MC(mask_1));
    else if (pos == 2)
	t = _mm_and_si64 (t, MC(mask_2));
    else if (pos == 3)
	t = _mm_and_si64 (t, MC(mask_3));

    p = _mm_or_si64 (r, t);
    p = _mm_or_si64 (g, p);

    return _mm_or_si64 (b, p);
}

static __inline__ __m64
pix_add_mul (__m64 x, __m64 a, __m64 y, __m64 b)
{
    x = _mm_mullo_pi16 (x, a);
    y = _mm_mullo_pi16 (y, b);
    x = _mm_adds_pu16 (x, MC(4x0080));
    x = _mm_adds_pu16 (x, y);
    x = _mm_adds_pu16 (x, _mm_srli_pi16 (x, 8));
    x = _mm_srli_pi16 (x, 8);

    return x;
}

/* --------------- MMX code patch for fbcompose.c --------------------- */

static FASTCALL void
mmxCombineMaskU (CARD32 *src, const CARD32 *mask, int width)
{
    const CARD32 *end = mask + width;
    while (mask < end) {
	CARD32 mmask = *mask;
	CARD32 maska = mmask >> 24;
	if (maska == 0) {
	    *src = 0;
	} else if (maska != 0xff) {
	    __m64 a = load8888(mmask);
	    __m64 s = load8888(*src);
	    a = expand_alpha(a);
	    s = pix_multiply(s, a);
	    *src = store8888(s);
	}
        ++src;
        ++mask;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineOverU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
	CARD32 ssrc = *src;
	CARD32 a = ssrc >> 24;
	if (a == 0xff) {
	    *dest = ssrc;
	} else if (a) {
	    __m64 s, sa;
	    s = load8888(ssrc);
	    sa = expand_alpha(s);
	    *dest = store8888(over(s, sa, load8888(*dest)));
	}
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineOverReverseU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
	__m64 d, da;
	d = load8888(*dest);
	da = expand_alpha(d);
	*dest = store8888(over (d, da, load8888(*src)));
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineInU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
        __m64 x, a;
        x = load8888(*src);
        a = load8888(*dest);
        a = expand_alpha(a);
        x = pix_multiply(x, a);
        *dest = store8888(x);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineInReverseU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
        __m64 x, a;
        x = load8888(*dest);
        a = load8888(*src);
        a = expand_alpha(a);
        x = pix_multiply(x, a);
        *dest = store8888(x);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineOutU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
        __m64 x, a;
        x = load8888(*src);
        a = load8888(*dest);
        a = expand_alpha(a);
        a = negate(a);
        x = pix_multiply(x, a);
        *dest = store8888(x);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineOutReverseU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
        __m64 x, a;
        x = load8888(*dest);
        a = load8888(*src);
        a = expand_alpha(a);
        a = negate(a);
        x = pix_multiply(x, a);
        *dest = store8888(x);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineAtopU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
        __m64 s, da, d, sia;
        s = load8888(*src);
        d = load8888(*dest);
        sia = expand_alpha(s);
        sia = negate(sia);
        da = expand_alpha(d);
        s = pix_add_mul (s, da, d, sia);
        *dest = store8888(s);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineAtopReverseU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end;

    end = dest + width;

    while (dest < end) {
        __m64 s, dia, d, sa;
        s = load8888(*src);
        d = load8888(*dest);
        sa = expand_alpha(s);
        dia = expand_alpha(d);
        dia = negate(dia);
	s = pix_add_mul (s, dia, d, sa);
        *dest = store8888(s);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineXorU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;

    while (dest < end) {
        __m64 s, dia, d, sia;
        s = load8888(*src);
        d = load8888(*dest);
        sia = expand_alpha(s);
        dia = expand_alpha(d);
        sia = negate(sia);
        dia = negate(dia);
	s = pix_add_mul (s, dia, d, sia);
        *dest = store8888(s);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineAddU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;
    while (dest < end) {
        __m64 s, d;
	s = load8888(*src);
	d = load8888(*dest);
	s = pix_add(s, d);
	*dest = store8888(s);
        ++dest;
        ++src;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineSaturateU (CARD32 *dest, const CARD32 *src, int width)
{
    const CARD32 *end = dest + width;
    while (dest < end) {
        CARD32 s = *src;
        CARD32 d = *dest;
        __m64 ms = load8888(s);
        __m64 md = load8888(d);
        CARD32 sa = s >> 24;
        CARD32 da = ~d >> 24;

        if (sa > da) {
            __m64 msa = load8888(FbIntDiv(da, sa)<<24);
            msa = expand_alpha(msa);
            ms = pix_multiply(ms, msa);
        }
        md = pix_add(md, ms);
        *dest = store8888(md);
        ++src;
        ++dest;
    }
    _mm_empty();
}

static FASTCALL void
mmxCombineSrcC (CARD32 *dest, CARD32 *src, CARD32 *mask, int width)
{
    const CARD32 *end = src + width;
    while (src < end) {
        __m64 a = load8888(*mask);
        __m64 s = load8888(*src);
        s = pix_multiply(s, a);