The add_dirac_obmc8_mmx function was the only MMX function left. This patch migrates it to SSE2. Here are the checkasm benchmark results: diracdsp.add_dirac_obmc_8_c: 2299.1 ( 1.00x) diracdsp.add_dirac_obmc_8_mmx: 237.6 ( 9.68x) diracdsp.add_dirac_obmc_8_sse2: 109.1 (21.07x)
Unified · Split
| @@ -228,7 +228,7 @@ cglobal add_dirac_obmc%1_%2, 5,5,5, dst, src, stride, obmc, yblen · 227 unmodified lines | ||
| 228 | 228 | punpckhbw m1, m4 |
| 229 | 229 | mova m2, [obmcq+i] |
| 230 | 230 | mova m3, m2 |
| 231 | − punpcklbw m2, m4 | |
| 231 | + punpcklbw m2, m4 | |
| 232 | 232 | punpckhbw m3, m4 |
| 233 | 233 | pmullw m0, m2 |
| 234 | 234 | pmullw m1, m3 |
| @@ -248,9 +248,6 @@ cglobal add_dirac_obmc%1_%2, 5,5,5, dst, src, stride, obmc, yblen · 13 unmodified lines | ||
| 248 | 248 | RET |
| 249 | 249 | %endm |
| 250 | 250 | |
| 251 | −INIT_MMX | |
| 252 | −ADD_OBMC 8, mmx | |
| 253 | − | |
| 254 | 251 | INIT_XMM |
| 255 | 252 | PUT_RECT sse2 |
| 256 | 253 | ADD_RECT sse2 |
| @@ -259,6 +256,25 @@ HPEL_FILTER sse2 · 2 unmodified lines | ||
| 259 | 256 | ADD_OBMC 32, sse2 |
| 260 | 257 | ADD_OBMC 16, sse2 |
| 261 | 258 | |
| 259 | +cglobal add_dirac_obmc8_sse2, 5,5,4, dst, src, stride, obmc, yblen | |
| 260 | + pxor m3, m3 | |
| 261 | + movsxdifnidn strideq, strided | |
| 262 | +.loop: | |
| 263 | + movh m0, [srcq] | |
| 264 | + punpcklbw m0, m3 | |
| 265 | + movh m1, [obmcq] | |
| 266 | + punpcklbw m1, m3 | |
| 267 | + pmullw m0, m1 | |
| 268 | + movu m1, [dstq] | |
| 269 | + paddw m0, m1 | |
| 270 | + movu [dstq], m0 | |
| 271 | + lea srcq, [srcq+strideq] | |
| 272 | + lea dstq, [dstq+2*strideq] | |
| 273 | + add obmcq, 32 | |
| 274 | + sub yblend, 1 | |
| 275 | + jg .loop | |
| 276 | + RET | |
| 277 | + | |
| 262 | 278 | INIT_XMM sse4 |
| 263 | 279 | |
| 264 | 280 | ; void dequant_subband_32(uint8_t *src, uint8_t *dst, ptrdiff_t stride, const int qf, const int qs, int tot_v, int tot_h) |
| @@ -24,8 +24,7 @@ · 23 unmodified lines | ||
| 24 | 24 | |
| 25 | 25 | void ff_add_rect_clamped_sse2(uint8_t *, const uint16_t *, int, const int16_t *, int, int, int); |
| 26 | 26 | |
| 27 | −void ff_add_dirac_obmc8_mmx(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen); | |
| 28 | − | |
| 27 | +void ff_add_dirac_obmc8_sse2(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen); | |
| 29 | 28 | void ff_add_dirac_obmc16_sse2(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen); |
| 30 | 29 | void ff_add_dirac_obmc32_sse2(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen); |
| 31 | 30 | |
| @@ -94,15 +93,12 @@ void ff_diracdsp_init_x86(DiracDSPContext* c) · 62 unmodified lines | ||
| 94 | 93 | #if HAVE_X86ASM |
| 95 | 94 | int mm_flags = av_get_cpu_flags(); |
| 96 | 95 | |
| 97 | − if (EXTERNAL_MMX(mm_flags)) { | |
| 98 | − c->add_dirac_obmc[0] = ff_add_dirac_obmc8_mmx; | |
| 99 | − } | |
| 100 | − | |
| 101 | 96 | if (EXTERNAL_SSE2(mm_flags)) { |
| 102 | 97 | c->dirac_hpel_filter = dirac_hpel_filter_sse2; |
| 103 | 98 | c->add_rect_clamped = ff_add_rect_clamped_sse2; |
| 104 | 99 | c->put_signed_rect_clamped[0] = (void *)ff_put_signed_rect_clamped_sse2; |
| 105 | 100 | |
| 101 | + c->add_dirac_obmc[0] = ff_add_dirac_obmc8_sse2; | |
| 106 | 102 | c->add_dirac_obmc[1] = ff_add_dirac_obmc16_sse2; |
| 107 | 103 | c->add_dirac_obmc[2] = ff_add_dirac_obmc32_sse2; |
| 108 | 104 | |
| @@ -116,5 +112,5 @@ void ff_diracdsp_init_x86(DiracDSPContext* c) · 7 unmodified lines | ||
| 116 | 112 | c->dequant_subband[1] = ff_dequant_subband_32_sse4; |
| 117 | 113 | c->put_signed_rect_clamped[1] = ff_put_signed_rect_clamped_10_sse4; |
| 118 | 114 | } |
| 119 | −#endif | |
| 115 | +#endif // HAVE_X86ASM | |
| 120 | 116 | } |