diff options
Diffstat (limited to 'libtwin/twin_primitive.c')
| -rw-r--r-- | libtwin/twin_primitive.c | 137 |
1 files changed, 136 insertions, 1 deletions
diff --git a/libtwin/twin_primitive.c b/libtwin/twin_primitive.c index 362cbf1..28684cb 100644 --- a/libtwin/twin_primitive.c +++ b/libtwin/twin_primitive.c @@ -263,5 +263,140 @@ make_twin_op_srcs(op,argb32) \ make_twin_op_srcs(op,rgb16) \ make_twin_op_srcs(op,a8) -make_twin_op_dsts_srcs(over) +make_twin_op_dsts_srcs(over); make_twin_op_dsts_srcs(source) + +#ifdef HAVE_ALTIVEC + +#include <altivec.h> + +#define VUNALIGNED(p) (((unsigned long)(p)) & 0xf) + + +/* Altivec over function, some bits inspired by SDL */ +static inline vector unsigned int over_v (vector unsigned char dst, + vector unsigned char src + ) +{ + const vector unsigned char alphasplit = + vec_and (vec_lvsl (0, (int *)NULL), vec_splat_u8(0x0c)); + const vector unsigned char merge = + vec_add(vec_lvsl(0, (int *)NULL), + (vector unsigned char)vec_splat_u16(0x0f)); + vector unsigned char alpha, alphainv; + vector unsigned short dmule, dmulo; + const vector unsigned short v80 = vec_sl(vec_splat_u16(1), vec_splat_u16(7)); + const vector unsigned short v8= vec_splat_u8(8); + + /* get source alpha values all over the vector */ + alpha = vec_perm(src, src, alphasplit); + + /* invert alpha */ + alphainv = vec_nor(alpha, alpha); + + /* multiply destination values with inverse alpha into 2 u16 vectors */ + dmule = vec_mule(dst, alphainv); + dmulo = vec_mulo(dst, alphainv); + + /* round and merge back */ + dmule = vec_add(dmule, v80); + dmulo = vec_add(dmulo, v80); + dmule = vec_add(dmule, vec_sr(dmule, v8)); + dmulo = vec_add(dmulo, vec_sr(dmulo, v8)); + dst = vec_perm(dmule, dmulo, merge); + + /* return added value */ + return vec_adds(dst, src); +} + +void _twin_vec_argb32_over_argb32 (twin_pointer_t dst, + twin_source_u src, + int width) +{ + twin_argb32_t dst32; + twin_argb32_t src32; + vector unsigned char edgeperm; + vector unsigned char src0v, src1v, srcv, dstv; + + + /* Go scalar for small amounts as I can't be bothered */ + if (width < 8) { + _twin_argb32_over_argb32(dst, src, width); + return; + } + + /* first run scalar until destination is aligned */ + while (VUNALIGNED(dst.v) && width--) { + dst32 = dst_argb32_get; + src32 = src_argb32; + dst32 = over (dst32, src32); + dst_argb32_set (dst32); + } + + /* maybe we should have a special "aligned" version to avoid those + * permutations... + */ + edgeperm = vec_lvsl (0, src.p.argb32); + src0v = vec_ld (0, src.p.argb32); + while(width >= 4) { + dstv = vec_ld (0, dst.argb32); + src1v = vec_ld (16, src.p.argb32); + srcv = vec_perm (src0v, src1v, edgeperm); + dstv = over_v (dstv, srcv); + vec_st ((vector unsigned int)dstv, 0, dst.argb32); + src.p.argb32 += 4; + dst.argb32 += 4; + src0v = src1v; + width -= 4; + } + + /* then run scalar again for remaining bits */ + while (width--) { + dst32 = dst_argb32_get; + src32 = src_argb32; + dst32 = over (dst32, src32); + dst_argb32_set (dst32); + } +} + +void _twin_vec_argb32_source_argb32 (twin_pointer_t dst, + twin_source_u src, + int width) +{ + twin_argb32_t dst32; + twin_argb32_t src32; + vector unsigned char edgeperm; + vector unsigned char src0v, src1v, srcv; + + + /* first run scalar until destination is aligned */ + while (VUNALIGNED(dst.v) && width--) { + src32 = src_argb32; + dst32 = src32; + dst_argb32_set (dst32); + } + + /* maybe we should have a special "aligned" version to avoid those + * permutations... + */ + edgeperm = vec_lvsl (0, src.p.argb32); + src0v = vec_ld (0, src.p.argb32); + while(width >= 4) { + src1v = vec_ld (16, src.p.argb32); + srcv = vec_perm (src0v, src1v, edgeperm); + vec_st ((vector unsigned int)srcv, 0, dst.argb32); + src.p.argb32 += 4; + dst.argb32 += 4; + src0v = src1v; + width -= 4; + } + + /* then run scalar again for remaining bits */ + while (width--) { + src32 = src_argb32; + dst32 = src32; + dst_argb32_set (dst32); + } +} + +#endif /* HAVE_ALTIVEC */ |