diff options
Diffstat (limited to 'fb')
| -rw-r--r-- | fb/fbmmx.c | 80 | ||||
| -rw-r--r-- | fb/fbmmx.h | 12 | ||||
| -rw-r--r-- | fb/fbpict.c | 14 | ||||
| -rw-r--r-- | fb/fbpict.h | 13 |
4 files changed, 117 insertions, 2 deletions
diff --git a/fb/fbmmx.c b/fb/fbmmx.c index 485fa4a7c..ead28405e 100644 --- a/fb/fbmmx.c +++ b/fb/fbmmx.c @@ -732,6 +732,86 @@ fbCompositeSrc_8888x8x8888mmx (CARD8 op, } void +fbCompositeSrc_8888x8888mmx (CARD8 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; + __m64 srca; + + CHECKPOINT(); + + fbComposeGetStart (pDst, xDst, yDst, CARD32, dstStride, dstLine, 1); + fbComposeGetStart (pSrc, xSrc, ySrc, CARD32, srcStride, srcLine, 1); + + 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 = (ullong)pack8888 (over (s, expand_alpha (s), d), (__m64)_mm_setzero_si64()); + + w--; + dst++; + src++; + } + + while (w >= 2) + { + __m64 vd = *(__m64 *)(dst + 0); + __m64 vs = *(__m64 *)(src + 0); + __m64 vs0 = expand8888 (vs, 0); + __m64 vs1 = expand8888 (vs, 1); + + *(__m64 *)dst = (__m64)pack8888 ( + over (vs0, expand_alpha (vs0), expand8888 (vd, 0)), + over (vs1, expand_alpha (vs1), expand8888 (vd, 1))); + + w -= 2; + dst += 2; + src += 2; + } + + while (w) + { + __m64 s = load8888 (*src); + __m64 d = load8888 (*dst); + + *dst = (ullong)pack8888 (over (s, expand_alpha (s), d), + (__m64)_mm_setzero_si64()); + + w--; + dst++; + src++; + } + } + + _mm_empty(); +} + +void fbCompositeSolidMask_nx8x8888mmx (CARD8 op, PicturePtr pSrc, PicturePtr pMask, diff --git a/fb/fbmmx.h b/fb/fbmmx.h index c0333a344..39e8d1e9c 100644 --- a/fb/fbmmx.h +++ b/fb/fbmmx.h @@ -60,6 +60,18 @@ void fbCompositeSrcAdd_8888x8888mmx (CARD8 op, INT16 yDst, CARD16 width, CARD16 height); +void fbCompositeSrc_8888x8888mmx (CARD8 op, + PicturePtr pSrc, + PicturePtr pMask, + PicturePtr pDst, + INT16 xSrc, + INT16 ySrc, + INT16 xMask, + INT16 yMask, + INT16 xDst, + INT16 yDst, + CARD16 width, + CARD16 height); void fbCompositeSolidMask_nx8888x8888Cmmx (CARD8 op, PicturePtr pSrc, PicturePtr pMask, diff --git a/fb/fbpict.c b/fb/fbpict.c index 4843eb2a1..67db2e1c1 100644 --- a/fb/fbpict.c +++ b/fb/fbpict.c @@ -1088,7 +1088,12 @@ fbComposite (CARD8 op, switch (pDst->format) { case PICT_a8r8g8b8: case PICT_x8r8g8b8: - func = fbCompositeSrc_8888x8888; +#ifdef USE_MMX + if (fbHaveMMX()) + func = fbCompositeSrc_8888x8888mmx; + else +#endif + func = fbCompositeSrc_8888x8888; break; case PICT_r8g8b8: func = fbCompositeSrc_8888x0888; @@ -1123,7 +1128,12 @@ fbComposite (CARD8 op, switch (pDst->format) { case PICT_a8b8g8r8: case PICT_x8b8g8r8: - func = fbCompositeSrc_8888x8888; +#ifdef USE_MMX + if (fbHaveMMX()) + func = fbCompositeSrc_8888x8888mmx; + else +#endif + func = fbCompositeSrc_8888x8888; break; case PICT_b8g8r8: func = fbCompositeSrc_8888x0888; diff --git a/fb/fbpict.h b/fb/fbpict.h index 3b88bfa37..4e8c95c30 100644 --- a/fb/fbpict.h +++ b/fb/fbpict.h @@ -1029,6 +1029,19 @@ fbCompositeSolidMask_nx1xn (CARD8 op, INT16 yDst, CARD16 width, CARD16 height); +void +fbCompositeSrc_8888x8888mmx (CARD8 op, + PicturePtr pSrc, + PicturePtr pMask, + PicturePtr pDst, + INT16 xSrc, + INT16 ySrc, + INT16 xMask, + INT16 yMask, + INT16 xDst, + INT16 yDst, + CARD16 width, + CARD16 height); void fbComposite (CARD8 op, |