Mercurial > libavcodec.hg
changeset 1565:1a9a63f59849 libavcodec
minor mmx2 optimization if the dct
| author | michael |
|---|---|
| date | Thu, 23 Oct 2003 13:22:20 +0000 |
| parents | b6b7d080f1a1 |
| children | 396e8526e82c |
| files | dsputil.h i386/dsputil_mmx.c i386/fdct_mmx.c i386/mpegvideo_mmx.c i386/mpegvideo_mmx_template.c |
| diffstat | 5 files changed, 41 insertions(+), 5 deletions(-) [+] |
line wrap: on
line diff
--- a/dsputil.h Thu Oct 23 12:59:21 2003 +0000 +++ b/dsputil.h Thu Oct 23 13:22:20 2003 +0000 @@ -41,6 +41,7 @@ void j_rev_dct (DCTELEM *data); void ff_fdct_mmx(DCTELEM *block); +void ff_fdct_mmx2(DCTELEM *block); /* encoding scans */ extern const uint8_t ff_alternate_horizontal_scan[64];
--- a/i386/dsputil_mmx.c Thu Oct 23 12:59:21 2003 +0000 +++ b/i386/dsputil_mmx.c Thu Oct 23 13:22:20 2003 +0000 @@ -1603,8 +1603,13 @@ const int idct_algo= avctx->idct_algo; #ifdef CONFIG_ENCODERS - if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX) - c->fdct = ff_fdct_mmx; + if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){ + if(mm_flags & MM_MMXEXT){ + c->fdct = ff_fdct_mmx2; + }else{ + c->fdct = ff_fdct_mmx; + } + } #endif //CONFIG_ENCODERS if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){
--- a/i386/fdct_mmx.c Thu Oct 23 12:59:21 2003 +0000 +++ b/i386/fdct_mmx.c Thu Oct 23 13:22:20 2003 +0000 @@ -210,14 +210,19 @@ movq_r2m(mm3, *(out + offset + 7 * 8)); } -static always_inline void fdct_row(const int16_t *in, int16_t *out, const int16_t *table) +static always_inline void fdct_row(const int16_t *in, int16_t *out, const int16_t *table, int mmx2) { + if(mmx2){ + pshufw_m2r(*(in + 4), mm5, 0x1B); + movq_m2r(*(in + 0), mm0); + }else{ movd_m2r(*(in + 6), mm5); punpcklwd_m2r(*(in + 4), mm5); movq_r2r(mm5, mm2); psrlq_i2r(0x20, mm5); movq_m2r(*(in + 0), mm0); punpcklwd_r2r(mm2, mm5); + } movq_r2r(mm0, mm1); paddsw_r2r(mm5, mm0); psubsw_r2r(mm5, mm1); @@ -283,9 +288,32 @@ table = tab_frw_01234567; out = block; for(i=8;i>0;i--) { - fdct_row(block1, out, table); + fdct_row(block1, out, table, 0); block1 += 8; table += 32; out += 8; } } + +void ff_fdct_mmx2(int16_t *block) +{ + int64_t align_tmp[16] ATTR_ALIGN(8); + int16_t * const block_tmp= (int16_t*)align_tmp; + int16_t *block1, *out; + const int16_t *table; + int i; + + block1 = block_tmp; + fdct_col(block, block1, 0); + fdct_col(block, block1, 4); + + block1 = block_tmp; + table = tab_frw_01234567; + out = block; + for(i=8;i>0;i--) { + fdct_row(block1, out, table, 1); + block1 += 8; + table += 32; + out += 8; + } +}
--- a/i386/mpegvideo_mmx.c Thu Oct 23 12:59:21 2003 +0000 +++ b/i386/mpegvideo_mmx.c Thu Oct 23 13:22:20 2003 +0000 @@ -490,11 +490,13 @@ #undef HAVE_MMX2 #define RENAME(a) a ## _MMX +#define RENAMEl(a) a ## _mmx #include "mpegvideo_mmx_template.c" #define HAVE_MMX2 #undef RENAME #define RENAME(a) a ## _MMX2 +#define RENAMEl(a) a ## _mmx2 #include "mpegvideo_mmx_template.c" void MPV_common_init_mmx(MpegEncContext *s)
--- a/i386/mpegvideo_mmx_template.c Thu Oct 23 12:59:21 2003 +0000 +++ b/i386/mpegvideo_mmx_template.c Thu Oct 23 13:22:20 2003 +0000 @@ -43,7 +43,7 @@ assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly? //s->fdct (block); - ff_fdct_mmx (block); //cant be anything else ... + RENAMEl(ff_fdct) (block); //cant be anything else ... if (s->mb_intra) { int dummy;