diff options
Diffstat (limited to 'render/rastx86.c')
| -rw-r--r-- | render/rastx86.c | 89 |
1 files changed, 77 insertions, 12 deletions
diff --git a/render/rastx86.c b/render/rastx86.c index 79020fb1..9360b5e2 100644 --- a/render/rastx86.c +++ b/render/rastx86.c @@ -3,6 +3,8 @@ x86 specific render optims live here */ #include <fitz.h> +typedef unsigned char byte; + /* always surround cpu specific code with HAVE_XXX */ #ifdef HAVE_MMX @@ -10,15 +12,77 @@ x86 specific render optims live here shouldn't require anything */ #include <mmintrin.h> +static void duff_4i1o4mmx(byte *sp0, int sw, byte *mp0, int mw, byte *dp0, int dw, int w0, int h) +{ + /* + rendering all pages of + x11pdf ~/doc/OpenGL/Presentations/CEDEC2003_Venus_and_Vulcan.pdf + % cumulative self self total + time seconds seconds calls ms/call ms/call name + 30.50 20.04 20.04 261 76.76 76.76 duff_4i1o4 + 21.67 22.02 10.95 221 49.55 49.55 duff_4i1o4mmx + */ + __m64 mzero = _mm_setzero_si64(); + while (h--) + { + byte *sp = sp0; + byte *mp = mp0; + byte *dp = dp0; + + unsigned *s = (unsigned *)sp; + unsigned *d = (unsigned *)dp; + + int w = w0; + + /* TODO: unroll and process two pixels/iteration */ + while (w--) + { + int ts = *s++; + int ma = *mp++ + 1; + int sa = ((ts & 0xff) * ma) >> 8; + int ssa = 254 - sa; + + __m64 d0 = _mm_cvtsi32_si64(*d); + __m64 s0 = _mm_cvtsi32_si64(ts); + + /* 4 x 9 bit alpha value */ + __m64 mma = _mm_set1_pi16(ma); + __m64 mssa = _mm_set1_pi16(ssa); + + /* unpack 0000argb => a0r0g0b0 */ + __m64 d1 = _mm_unpacklo_pi8(d0, mzero); + __m64 s1 = _mm_unpacklo_pi8(s0, mzero); + + /* s1 * ma => a0r0g0b0 */ + __m64 msma = _mm_mullo_pi16(s1, mma); + /* d1 * mssa */ + __m64 mdssa = _mm_mullo_pi16(d1, mssa); + + __m64 res0 = _mm_add_pi16(msma, mdssa); + /* TODO: is it possible to get rid of the shift? */ + __m64 res1 = _mm_srli_pi16(res0, 8); + + /* pack */ + __m64 res2 = _mm_packs_pu16(res1, mzero); + + *d++ = _mm_cvtsi64_si32(res2); + } + + sp0 += sw; + mp0 += mw; + dp0 += dw; + } + + _mm_empty(); +} + static inline unsigned getargb(unsigned *s, int w, int h, int u, int v) { - if (u < 0 || u >= w) return 0; - if (v < 0 || v >= h) return 0; + if (u < 0 | u >= w | v < 0 | v >= h) return 0; return s[w * v + u]; } -/* this code has not been tested since refactoring */ static void img_4o4mmx(FZ_PSRC, FZ_PDST, FZ_PCTM) { /* since mmx does not have an unsigned multiply instruction we use @@ -33,12 +97,16 @@ static void img_4o4mmx(FZ_PSRC, FZ_PDST, FZ_PCTM) unsigned *d = (unsigned *)dst0; int u = u0; int v = v0; - int w = w0; + int w = w0; + + __m64 mzero = _mm_setzero_si64(); + __m64 m256 = _mm_set1_pi16(256); + __m64 malphamask = _mm_cvtsi32_si64(0xff); while (w--) { int iu = u >> 17; - int iv = u >> 17; + int iv = v >> 17; int fu = u & 0x7fff; int fv = v & 0x7fff; @@ -75,7 +143,6 @@ static void img_4o4mmx(FZ_PSRC, FZ_PDST, FZ_PCTM) } /* unpack src into 4x16bit vectors */ - __m64 mzero = _mm_setzero_si64(); __m64 ms0 = _mm_unpackhi_pi8(ms0s1, mzero); __m64 ms1 = _mm_unpacklo_pi8(ms0s1, mzero); __m64 ms2 = _mm_unpackhi_pi8(ms2s3, mzero); @@ -110,14 +177,12 @@ static void img_4o4mmx(FZ_PSRC, FZ_PDST, FZ_PCTM) __m64 d1 = _mm_unpacklo_pi8(d0, mzero); /* get src alpha */ - __m64 m256 = _mm_set1_pi16(256); - __m64 malphamask = _mm_cvtsi32_si64(0xff); - /* splat alpha TODO: better way? */ + /* splat alpha */ __m64 a0001 = _mm_and_si64(malphamask, t8); - __m64 a0010 = _mm_slli_si64(a0001, 16); - __m64 a0011 = _mm_or_si64(a0001, a0010); + __m64 a0011 = _mm_unpacklo_pi16(a0001, a0001); __m64 a1111 = _mm_unpacklo_pi16(a0011, a0011); + /* 255+1 - sa */ __m64 sna = _mm_sub_pi16(m256, a1111); @@ -152,9 +217,9 @@ fz_accelrastfuncs(fz_rastfuncs *tab) # ifdef HAVE_MMX if (fz_cpuflags & HAVE_MMX) { + tab->duff_4i1o4 = duff_4i1o4mmx; tab->img_4o4 = img_4o4mmx; } # endif } #endif - |