About 2x faster than the c version.
Unified · Split
| @@ -166,6 +166,112 @@ align 16 · 165 unmodified lines | |||
| 166 | jl .loop | 166 | jl .loop |
| 167 | REP_RET | 167 | REP_RET |
| 168 | 168 | ||
| 169 | ;********************************************************** | ||
| 170 | ;void ps_hybrid_analysis_ileave_sse(float out[2][38][64], | ||
| 171 | ; float (*in)[32][2], | ||
| 172 | ; int i, int len) | ||
| 173 | ;********************************************************** | ||
| 174 | INIT_XMM sse | ||
| 175 | cglobal ps_hybrid_analysis_ileave, 3, 7, 5, out, in, i, len, in0, in1, tmp | ||
| 176 | movsxdifnidn iq, id | ||
| 177 | mov lend, 32 << 3 | ||
| 178 | lea inq, [inq+iq*4] | ||
| 179 | mov tmpd, id | ||
| 180 | shl tmpd, 8 | ||
| 181 | add outq, tmpq | ||
| 182 | mov tmpd, 64 | ||
| 183 | sub tmpd, id | ||
| 184 | mov id, tmpd | ||
| 185 | |||
| 186 | test id, 1 | ||
| 187 | jne .loop4 | ||
| 188 | test id, 2 | ||
| 189 | jne .loop8 | ||
| 190 | |||
| 191 | align 16 | ||
| 192 | .loop16: | ||
| 193 | mov in0q, inq | ||
| 194 | mov in1q, 38*64*4 | ||
| 195 | add in1q, in0q | ||
| 196 | mov tmpd, lend | ||
| 197 | |||
| 198 | .inner_loop16: | ||
| 199 | movaps m0, [in0q] | ||
| 200 | movaps m1, [in1q] | ||
| 201 | movaps m2, [in0q+lenq] | ||
| 202 | movaps m3, [in1q+lenq] | ||
| 203 | TRANSPOSE4x4PS 0, 1, 2, 3, 4 | ||
| 204 | movaps [outq], m0 | ||
| 205 | movaps [outq+lenq], m1 | ||
| 206 | movaps [outq+lenq*2], m2 | ||
| 207 | movaps [outq+3*32*2*4], m3 | ||
| 208 | lea in0q, [in0q+lenq*2] | ||
| 209 | lea in1q, [in1q+lenq*2] | ||
| 210 | add outq, mmsize | ||
| 211 | sub tmpd, mmsize | ||
| 212 | jg .inner_loop16 | ||
| 213 | add inq, 16 | ||
| 214 | add outq, 3*32*2*4 | ||
| 215 | sub id, 4 | ||
| 216 | jg .loop16 | ||
| 217 | RET | ||
| 218 | |||
| 219 | align 16 | ||
| 220 | .loop8: | ||
| 221 | mov in0q, inq | ||
| 222 | mov in1q, 38*64*4 | ||
| 223 | add in1q, in0q | ||
| 224 | mov tmpd, lend | ||
| 225 | |||
| 226 | .inner_loop8: | ||
| 227 | movlps m0, [in0q] | ||
| 228 | movlps m1, [in1q] | ||
| 229 | movhps m0, [in0q+lenq] | ||
| 230 | movhps m1, [in1q+lenq] | ||
| 231 | SBUTTERFLYPS 0, 1, 2 | ||
| 232 | SBUTTERFLYPD 0, 1, 2 | ||
| 233 | movaps [outq], m0 | ||
| 234 | movaps [outq+lenq], m1 | ||
| 235 | lea in0q, [in0q+lenq*2] | ||
| 236 | lea in1q, [in1q+lenq*2] | ||
| 237 | add outq, mmsize | ||
| 238 | sub tmpd, mmsize | ||
| 239 | jg .inner_loop8 | ||
| 240 | add inq, 8 | ||
| 241 | add outq, lenq | ||
| 242 | sub id, 2 | ||
| 243 | jg .loop16 | ||
| 244 | RET | ||
| 245 | |||
| 246 | align 16 | ||
| 247 | .loop4: | ||
| 248 | mov in0q, inq | ||
| 249 | mov in1q, 38*64*4 | ||
| 250 | add in1q, in0q | ||
| 251 | mov tmpd, lend | ||
| 252 | |||
| 253 | .inner_loop4: | ||
| 254 | movss m0, [in0q] | ||
| 255 | movss m1, [in1q] | ||
| 256 | movss m2, [in0q+lenq] | ||
| 257 | movss m3, [in1q+lenq] | ||
| 258 | movlhps m0, m1 | ||
| 259 | movlhps m2, m3 | ||
| 260 | shufps m0, m2, q2020 | ||
| 261 | movaps [outq], m0 | ||
| 262 | lea in0q, [in0q+lenq*2] | ||
| 263 | lea in1q, [in1q+lenq*2] | ||
| 264 | add outq, mmsize | ||
| 265 | sub tmpd, mmsize | ||
| 266 | jg .inner_loop4 | ||
| 267 | add inq, 4 | ||
| 268 | sub id, 1 | ||
| 269 | test id, 2 | ||
| 270 | jne .loop8 | ||
| 271 | cmp id, 4 | ||
| 272 | jge .loop16 | ||
| 273 | RET | ||
| 274 | |||
| 169 | ;*********************************************************** | 275 | ;*********************************************************** |
| 170 | ;void ps_hybrid_synthesis_deint_sse4(float out[2][38][64], | 276 | ;void ps_hybrid_synthesis_deint_sse4(float out[2][38][64], |
| 171 | ; float (*in)[32][2], | 277 | ; float (*in)[32][2], |
| @@ -44,6 +44,8 @@ void ff_ps_hybrid_synthesis_deint_sse(float out[2][38][64], float (*in)[32][2], · 43 unmodified lines | |||
| 44 | int i, int len); | 44 | int i, int len); |
| 45 | void ff_ps_hybrid_synthesis_deint_sse4(float out[2][38][64], float (*in)[32][2], | 45 | void ff_ps_hybrid_synthesis_deint_sse4(float out[2][38][64], float (*in)[32][2], |
| 46 | int i, int len); | 46 | int i, int len); |
| 47 | void ff_ps_hybrid_analysis_ileave_sse(float (*out)[32][2], float L[2][38][64], | ||
| 48 | int i, int len); | ||
| 47 | 49 | ||
| 48 | av_cold void ff_psdsp_init_x86(PSDSPContext *s) | 50 | av_cold void ff_psdsp_init_x86(PSDSPContext *s) |
| 49 | { | 51 | { |
| @@ -52,6 +54,7 @@ av_cold void ff_psdsp_init_x86(PSDSPContext *s) · 2 unmodified lines | |||
| 52 | if (EXTERNAL_SSE(cpu_flags)) { | 54 | if (EXTERNAL_SSE(cpu_flags)) { |
| 53 | s->add_squares = ff_ps_add_squares_sse; | 55 | s->add_squares = ff_ps_add_squares_sse; |
| 54 | s->mul_pair_single = ff_ps_mul_pair_single_sse; | 56 | s->mul_pair_single = ff_ps_mul_pair_single_sse; |
| 57 | s->hybrid_analysis_ileave = ff_ps_hybrid_analysis_ileave_sse; | ||
| 55 | s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse; | 58 | s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse; |
| 56 | s->hybrid_analysis = ff_ps_hybrid_analysis_sse; | 59 | s->hybrid_analysis = ff_ps_hybrid_analysis_sse; |
| 57 | } | 60 | } |