diff options
author | Christophe Gisquet <christophe.gisquet@gmail.com> | 2012-04-19 22:36:17 +0200 |
---|---|---|
committer | Diego Biurrun <diego@biurrun.de> | 2012-05-10 18:42:43 +0200 |
commit | 110d0cdc9d1ec414a658f841a3fbefbf6f796d61 (patch) | |
tree | d2f80a035204c7a75a6daa5c71357e61817ffd54 /libavcodec/x86/rv40dsp_init.c | |
parent | 706b998cdcea97c50fad2228f67488de0e06b2a2 (diff) | |
download | ffmpeg-110d0cdc9d1ec414a658f841a3fbefbf6f796d61.tar.gz |
rv40dsp x86: MMX/MMX2/3DNow/SSE2/SSSE3 implementations of MC
Code mostly inspired by vp8's MC, however:
- its MMX2 horizontal filter is worse because it can't take advantage of
the coefficient redundancy
- that same coefficient redundancy allows better code for non-SSSE3 versions
Benchmark (rounded to tens of unit):
V8x8 H8x8 2D8x8 V16x16 H16x16 2D16x16
C 445 358 985 1785 1559 3280
MMX* 219 271 478 714 929 1443
SSE2 131 158 294 425 515 892
SSSE3 120 122 248 387 390 763
End result is overall around a 15% speedup for SSSE3 version (on 6 sequences);
all loop filter functions now take around 55% of decoding time, while luma MC
dsp functions are around 6%, chroma ones are 1.3% and biweight around 2.3%.
Signed-off-by: Diego Biurrun <diego@biurrun.de>
Diffstat (limited to 'libavcodec/x86/rv40dsp_init.c')
-rw-r--r-- | libavcodec/x86/rv40dsp_init.c | 146 |
1 files changed, 146 insertions, 0 deletions
diff --git a/libavcodec/x86/rv40dsp_init.c b/libavcodec/x86/rv40dsp_init.c index df468aa9e5..3f42363e4e 100644 --- a/libavcodec/x86/rv40dsp_init.c +++ b/libavcodec/x86/rv40dsp_init.c @@ -22,8 +22,11 @@ /** * @file * RV40 decoder motion compensation functions x86-optimised + * 2,0 and 0,2 have h264 equivalents. + * 3,3 is bugged in the rv40 format and maps to _xy2 version */ +#include "libavcodec/x86/dsputil_mmx.h" #include "libavcodec/rv34dsp.h" void ff_put_rv40_chroma_mc8_mmx (uint8_t *dst, uint8_t *src, @@ -53,6 +56,132 @@ DECLARE_WEIGHT(mmx) DECLARE_WEIGHT(sse2) DECLARE_WEIGHT(ssse3) +/** @{ */ +/** + * Define one qpel function. + * LOOPSIZE must be already set to the number of pixels processed per + * iteration in the inner loop of the called functions. + * COFF(x) must be already defined so as to provide the offset into any + * array of coeffs used by the called function for the qpel position x. + */ +#define QPEL_FUNC_DECL(OP, SIZE, PH, PV, OPT) \ +static void OP ## rv40_qpel ##SIZE ##_mc ##PH ##PV ##OPT(uint8_t *dst, \ + uint8_t *src, \ + int stride) \ +{ \ + int i; \ + if (PH && PV) { \ + DECLARE_ALIGNED(16, uint8_t, tmp)[SIZE * (SIZE + 5)]; \ + uint8_t *tmpptr = tmp + SIZE * 2; \ + src -= stride * 2; \ + \ + for (i = 0; i < SIZE; i += LOOPSIZE) \ + ff_put_rv40_qpel_h ##OPT(tmp + i, SIZE, src + i, stride, \ + SIZE + 5, HCOFF(PH)); \ + for (i = 0; i < SIZE; i += LOOPSIZE) \ + ff_ ##OP ##rv40_qpel_v ##OPT(dst + i, stride, tmpptr + i, \ + SIZE, SIZE, VCOFF(PV)); \ + } else if (PV) { \ + for (i = 0; i < SIZE; i += LOOPSIZE) \ + ff_ ##OP ##rv40_qpel_v ## OPT(dst + i, stride, src + i, \ + stride, SIZE, VCOFF(PV)); \ + } else { \ + for (i = 0; i < SIZE; i += LOOPSIZE) \ + ff_ ##OP ##rv40_qpel_h ## OPT(dst + i, stride, src + i, \ + stride, SIZE, HCOFF(PH)); \ + } \ +}; + +/** Declare functions for sizes 8 and 16 and given operations + * and qpel position. */ +#define QPEL_FUNCS_DECL(OP, PH, PV, OPT) \ + QPEL_FUNC_DECL(OP, 8, PH, PV, OPT) \ + QPEL_FUNC_DECL(OP, 16, PH, PV, OPT) + +/** Declare all functions for all sizes and qpel positions */ +#define QPEL_MC_DECL(OP, OPT) \ +void ff_ ##OP ##rv40_qpel_h ##OPT(uint8_t *dst, ptrdiff_t dstStride, \ + const uint8_t *src, \ + ptrdiff_t srcStride, \ + int len, int m); \ +void ff_ ##OP ##rv40_qpel_v ##OPT(uint8_t *dst, ptrdiff_t dstStride, \ + const uint8_t *src, \ + ptrdiff_t srcStride, \ + int len, int m); \ +QPEL_FUNCS_DECL(OP, 0, 1, OPT) \ +QPEL_FUNCS_DECL(OP, 0, 3, OPT) \ +QPEL_FUNCS_DECL(OP, 1, 0, OPT) \ +QPEL_FUNCS_DECL(OP, 1, 1, OPT) \ +QPEL_FUNCS_DECL(OP, 1, 2, OPT) \ +QPEL_FUNCS_DECL(OP, 1, 3, OPT) \ +QPEL_FUNCS_DECL(OP, 2, 1, OPT) \ +QPEL_FUNCS_DECL(OP, 2, 2, OPT) \ +QPEL_FUNCS_DECL(OP, 2, 3, OPT) \ +QPEL_FUNCS_DECL(OP, 3, 0, OPT) \ +QPEL_FUNCS_DECL(OP, 3, 1, OPT) \ +QPEL_FUNCS_DECL(OP, 3, 2, OPT) +/** @} */ + +#define LOOPSIZE 8 +#define HCOFF(x) (32 * (x - 1)) +#define VCOFF(x) (32 * (x - 1)) +QPEL_MC_DECL(put_, _ssse3) +QPEL_MC_DECL(avg_, _ssse3) + +#undef LOOPSIZE +#undef HCOFF +#undef VCOFF +#define LOOPSIZE 8 +#define HCOFF(x) (64 * (x - 1)) +#define VCOFF(x) (64 * (x - 1)) +QPEL_MC_DECL(put_, _sse2) +QPEL_MC_DECL(avg_, _sse2) + +#if ARCH_X86_32 +#undef LOOPSIZE +#undef HCOFF +#undef VCOFF +#define LOOPSIZE 4 +#define HCOFF(x) (64 * (x - 1)) +#define VCOFF(x) (64 * (x - 1)) + +QPEL_MC_DECL(put_, _mmx) + +#define ff_put_rv40_qpel_h_mmx2 ff_put_rv40_qpel_h_mmx +#define ff_put_rv40_qpel_v_mmx2 ff_put_rv40_qpel_v_mmx +QPEL_MC_DECL(avg_, _mmx2) + +#define ff_put_rv40_qpel_h_3dnow ff_put_rv40_qpel_h_mmx +#define ff_put_rv40_qpel_v_3dnow ff_put_rv40_qpel_v_mmx +QPEL_MC_DECL(avg_, _3dnow) +#endif + +/** @{ */ +/** Set one function */ +#define QPEL_FUNC_SET(OP, SIZE, PH, PV, OPT) \ + c-> OP ## pixels_tab[2 - SIZE / 8][4 * PV + PH] = OP ## rv40_qpel ##SIZE ## _mc ##PH ##PV ##OPT; + +/** Set functions put and avg for sizes 8 and 16 and a given qpel position */ +#define QPEL_FUNCS_SET(OP, PH, PV, OPT) \ + QPEL_FUNC_SET(OP, 8, PH, PV, OPT) \ + QPEL_FUNC_SET(OP, 16, PH, PV, OPT) + +/** Set all functions for all sizes and qpel positions */ +#define QPEL_MC_SET(OP, OPT) \ +QPEL_FUNCS_SET (OP, 0, 1, OPT) \ +QPEL_FUNCS_SET (OP, 0, 3, OPT) \ +QPEL_FUNCS_SET (OP, 1, 0, OPT) \ +QPEL_FUNCS_SET (OP, 1, 1, OPT) \ +QPEL_FUNCS_SET (OP, 1, 2, OPT) \ +QPEL_FUNCS_SET (OP, 1, 3, OPT) \ +QPEL_FUNCS_SET (OP, 2, 1, OPT) \ +QPEL_FUNCS_SET (OP, 2, 2, OPT) \ +QPEL_FUNCS_SET (OP, 2, 3, OPT) \ +QPEL_FUNCS_SET (OP, 3, 0, OPT) \ +QPEL_FUNCS_SET (OP, 3, 1, OPT) \ +QPEL_FUNCS_SET (OP, 3, 2, OPT) +/** @} */ + void ff_rv40dsp_init_x86(RV34DSPContext *c, DSPContext *dsp) { #if HAVE_YASM @@ -65,25 +194,42 @@ void ff_rv40dsp_init_x86(RV34DSPContext *c, DSPContext *dsp) c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_mmx; c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_mmx; c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_mmx; + c->put_pixels_tab[0][15] = ff_put_rv40_qpel16_mc33_mmx; + c->put_pixels_tab[1][15] = ff_put_rv40_qpel8_mc33_mmx; + c->avg_pixels_tab[0][15] = ff_avg_rv40_qpel16_mc33_mmx; + c->avg_pixels_tab[1][15] = ff_avg_rv40_qpel8_mc33_mmx; +#if ARCH_X86_32 + QPEL_MC_SET(put_, _mmx) +#endif } if (mm_flags & AV_CPU_FLAG_MMX2) { c->avg_chroma_pixels_tab[0] = ff_avg_rv40_chroma_mc8_mmx2; c->avg_chroma_pixels_tab[1] = ff_avg_rv40_chroma_mc4_mmx2; +#if ARCH_X86_32 + QPEL_MC_SET(avg_, _mmx2) +#endif } else if (mm_flags & AV_CPU_FLAG_3DNOW) { c->avg_chroma_pixels_tab[0] = ff_avg_rv40_chroma_mc8_3dnow; c->avg_chroma_pixels_tab[1] = ff_avg_rv40_chroma_mc4_3dnow; +#if ARCH_X86_32 + QPEL_MC_SET(avg_, _3dnow) +#endif } if (mm_flags & AV_CPU_FLAG_SSE2) { c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_sse2; c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_sse2; c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_sse2; c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_sse2; + QPEL_MC_SET(put_, _sse2) + QPEL_MC_SET(avg_, _sse2) } if (mm_flags & AV_CPU_FLAG_SSSE3) { c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_ssse3; c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_ssse3; c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_ssse3; c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_ssse3; + QPEL_MC_SET(put_, _ssse3) + QPEL_MC_SET(avg_, _ssse3) } #endif } |