This patch doesn't handle cases where alpha is zero but components are not. Not sure<div>if this is either possible or frequent.</div><div> </div><div> <div class="gmail_quote">On Mon, Dec 20, 2010 at 8:50 PM, Søren Sandmann <<a href="mailto:sandmann@daimi.au.dk">sandmann@daimi.au.dk</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">From: Søren Sandmann Pedersen <<a href="mailto:ssp@redhat.com">ssp@redhat.com</a>> This is a big speed-up in the SVG helicopter game: <a href="http://ie.microsoft.com/testdrive/Performance/Helicopter/Default.xhtml" target="_blank">http://ie.microsoft.com/testdrive/Performance/Helicopter/Default.xhtml</a> when rendered by Firefox 4 since it is compositing big images consisting almost entirely of zeros. --- pixman/pixman-sse2.c | 75 +++++++++++++++++++++++++++++-------------------- 1 files changed, 44 insertions(+), 31 deletions(-) diff --git a/pixman/pixman-sse2.c b/pixman/pixman-sse2.c index 5907de0..032f13b 100644 --- a/pixman/pixman-sse2.c +++ b/pixman/pixman-sse2.c @@ -3051,37 +3051,45 @@ sse2_composite_over_8888_n_8888 (pixman_implementation_t *imp, while (w && (unsigned long)dst & 15) { uint32_t s = *src++; - uint32_t d = *dst; - - __m64 ms = unpack_32_1x64 (s); - __m64 alpha = expand_alpha_1x64 (ms); - __m64 dest = _mm_movepi64_pi64 (xmm_mask); - __m64 alpha_dst = unpack_32_1x64 (d); - - *dst++ = pack_1x64_32 ( - in_over_1x64 (&ms, &alpha, &dest, &alpha_dst)); + if (s) + { + uint32_t d = *dst; + + __m64 ms = unpack_32_1x64 (s); + __m64 alpha = expand_alpha_1x64 (ms); + __m64 dest = _mm_movepi64_pi64 (xmm_mask); + __m64 alpha_dst = unpack_32_1x64 (d); + + *dst = pack_1x64_32 ( + in_over_1x64 (&ms, &alpha, &dest, &alpha_dst)); + } + dst++; w--; } while (w >= 4) { xmm_src = load_128_unaligned ((__m128i*)src); - xmm_dst = load_128_aligned ((__m128i*)dst); - - unpack_128_2x128 (xmm_src, &xmm_src_lo, &xmm_src_hi); - unpack_128_2x128 (xmm_dst, &xmm_dst_lo, &xmm_dst_hi); - expand_alpha_2x128 (xmm_src_lo, xmm_src_hi, - &xmm_alpha_lo, &xmm_alpha_hi); - - in_over_2x128 (&xmm_src_lo, &xmm_src_hi, - &xmm_alpha_lo, &xmm_alpha_hi, - &xmm_mask, &xmm_mask, - &xmm_dst_lo, &xmm_dst_hi); - - save_128_aligned ( - (__m128i*)dst, pack_2x128_128 (xmm_dst_lo, xmm_dst_hi)); + if (!is_zero (xmm_src)) + { + xmm_dst = load_128_aligned ((__m128i*)dst); + + unpack_128_2x128 (xmm_src, &xmm_src_lo, &xmm_src_hi); + unpack_128_2x128 (xmm_dst, &xmm_dst_lo, &xmm_dst_hi); + expand_alpha_2x128 (xmm_src_lo, xmm_src_hi, + &xmm_alpha_lo, &xmm_alpha_hi); + + in_over_2x128 (&xmm_src_lo, &xmm_src_hi, + &xmm_alpha_lo, &xmm_alpha_hi, + &xmm_mask, &xmm_mask, + &xmm_dst_lo, &xmm_dst_hi); + + save_128_aligned ( + (__m128i*)dst, pack_2x128_128 (xmm_dst_lo, xmm_dst_hi)); + } + dst += 4; src += 4; w -= 4; @@ -3090,16 +3098,21 @@ sse2_composite_over_8888_n_8888 (pixman_implementation_t *imp, while (w) { uint32_t s = *src++; - uint32_t d = *dst; - __m64 ms = unpack_32_1x64 (s); - __m64 alpha = expand_alpha_1x64 (ms); - __m64 mask = _mm_movepi64_pi64 (xmm_mask); - __m64 dest = unpack_32_1x64 (d); - - *dst++ = pack_1x64_32 ( - in_over_1x64 (&ms, &alpha, &mask, &dest)); + if (s) + { + uint32_t d = *dst; + + __m64 ms = unpack_32_1x64 (s); + __m64 alpha = expand_alpha_1x64 (ms); + __m64 mask = _mm_movepi64_pi64 (xmm_mask); + __m64 dest = unpack_32_1x64 (d); + + *dst = pack_1x64_32 ( + in_over_1x64 (&ms, &alpha, &mask, &dest)); + } + dst++; w--; } } -- 1.7.3.1 _______________________________________________ Pixman mailing list <a href="mailto:Pixman@lists.freedesktop.org">Pixman@lists.freedesktop.org</a> <a href="http://lists.freedesktop.org/mailman/listinfo/pixman" target="_blank">http://lists.freedesktop.org/mailman/listinfo/pixman</a> </blockquote></div> </div>