snowdsp_mmx.c 38.6 KB
Newer Older
1 2 3 4
/*
 * MMX and SSE2 optimized snow DSP utils
 * Copyright (c) 2005-2006 Robert Edele <yartrebo@earthlink.net>
 *
5 6 7
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
8 9
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
10
 * version 2.1 of the License, or (at your option) any later version.
11
 *
12
 * FFmpeg is distributed in the hope that it will be useful,
13 14 15 16 17
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
18
 * License along with FFmpeg; if not, write to the Free Software
19 20 21
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

22 23 24
#include "libavutil/x86_cpu.h"
#include "libavcodec/avcodec.h"
#include "libavcodec/snow.h"
25

26
void ff_snow_horizontal_compose97i_sse2(IDWTELEM *b, int width){
27
    const int w2= (width+1)>>1;
28
    DECLARE_ALIGNED_16(IDWTELEM, temp[width>>1]);
29 30 31 32 33
    const int w_l= (width>>1);
    const int w_r= w2 - 1;
    int i;

    { // Lift 0
34 35
        IDWTELEM * const ref = b + w2 - 1;
        IDWTELEM b_0 = b[0]; //By allowing the first entry in b[0] to be calculated twice
36 37 38 39 40 41 42
        // (the first time erroneously), we allow the SSE2 code to run an extra pass.
        // The savings in code and time are well worth having to store this value and
        // calculate b[0] correctly afterwards.

        i = 0;
        asm volatile(
            "pcmpeqd   %%xmm7, %%xmm7         \n\t"
43 44 45
            "pcmpeqd   %%xmm3, %%xmm3         \n\t"
            "psllw         $1, %%xmm3         \n\t"
            "paddw     %%xmm7, %%xmm3         \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
46
            "psllw        $13, %%xmm3         \n\t"
47
        ::);
48
        for(; i<w_l-15; i+=16){
49 50 51
            asm volatile(
                "movdqu   (%1), %%xmm1        \n\t"
                "movdqu 16(%1), %%xmm5        \n\t"
52 53 54 55 56 57
                "movdqu  2(%1), %%xmm2        \n\t"
                "movdqu 18(%1), %%xmm6        \n\t"
                "paddw  %%xmm1, %%xmm2        \n\t"
                "paddw  %%xmm5, %%xmm6        \n\t"
                "paddw  %%xmm7, %%xmm2        \n\t"
                "paddw  %%xmm7, %%xmm6        \n\t"
58 59 60 61 62 63
                "pmulhw %%xmm3, %%xmm2        \n\t"
                "pmulhw %%xmm3, %%xmm6        \n\t"
                "paddw    (%0), %%xmm2        \n\t"
                "paddw  16(%0), %%xmm6        \n\t"
                "movdqa %%xmm2, (%0)          \n\t"
                "movdqa %%xmm6, 16(%0)        \n\t"
64 65 66 67 68 69 70 71 72
                :: "r"(&b[i]), "r"(&ref[i])
                : "memory"
            );
        }
        snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
        b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
    }

    { // Lift 1
73
        IDWTELEM * const dst = b+w2;
74 75

        i = 0;
76
        for(; (((x86_reg)&dst[i]) & 0x1F) && i<w_r; i++){
77 78
            dst[i] = dst[i] - (b[i] + b[i + 1]);
        }
79
        for(; i<w_r-15; i+=16){
80 81 82
            asm volatile(
                "movdqu   (%1), %%xmm1        \n\t"
                "movdqu 16(%1), %%xmm5        \n\t"
83 84 85 86
                "movdqu  2(%1), %%xmm2        \n\t"
                "movdqu 18(%1), %%xmm6        \n\t"
                "paddw  %%xmm1, %%xmm2        \n\t"
                "paddw  %%xmm5, %%xmm6        \n\t"
87 88
                "movdqa   (%0), %%xmm0        \n\t"
                "movdqa 16(%0), %%xmm4        \n\t"
89 90
                "psubw  %%xmm2, %%xmm0        \n\t"
                "psubw  %%xmm6, %%xmm4        \n\t"
91 92 93 94 95 96 97 98 99 100
                "movdqa %%xmm0, (%0)          \n\t"
                "movdqa %%xmm4, 16(%0)        \n\t"
                :: "r"(&dst[i]), "r"(&b[i])
                : "memory"
            );
        }
        snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
    }

    { // Lift 2
101 102
        IDWTELEM * const ref = b+w2 - 1;
        IDWTELEM b_0 = b[0];
103 104 105

        i = 0;
        asm volatile(
106
            "psllw         $15, %%xmm7        \n\t"
107 108 109
            "pcmpeqw    %%xmm6, %%xmm6        \n\t"
            "psrlw         $13, %%xmm6        \n\t"
            "paddw      %%xmm7, %%xmm6        \n\t"
110
        ::);
111
        for(; i<w_l-15; i+=16){
112
            asm volatile(
113 114 115 116 117 118 119 120 121 122 123 124 125 126
                "movdqu   (%1), %%xmm0        \n\t"
                "movdqu 16(%1), %%xmm4        \n\t"
                "movdqu  2(%1), %%xmm1        \n\t"
                "movdqu 18(%1), %%xmm5        \n\t" //FIXME try aligned reads and shifts
                "paddw  %%xmm6, %%xmm0        \n\t"
                "paddw  %%xmm6, %%xmm4        \n\t"
                "paddw  %%xmm7, %%xmm1        \n\t"
                "paddw  %%xmm7, %%xmm5        \n\t"
                "pavgw  %%xmm1, %%xmm0        \n\t"
                "pavgw  %%xmm5, %%xmm4        \n\t"
                "psubw  %%xmm7, %%xmm0        \n\t"
                "psubw  %%xmm7, %%xmm4        \n\t"
                "psraw      $1, %%xmm0        \n\t"
                "psraw      $1, %%xmm4        \n\t"
127 128
                "movdqa   (%0), %%xmm1        \n\t"
                "movdqa 16(%0), %%xmm5        \n\t"
129 130 131 132 133 134
                "paddw  %%xmm1, %%xmm0        \n\t"
                "paddw  %%xmm5, %%xmm4        \n\t"
                "psraw      $2, %%xmm0        \n\t"
                "psraw      $2, %%xmm4        \n\t"
                "paddw  %%xmm1, %%xmm0        \n\t"
                "paddw  %%xmm5, %%xmm4        \n\t"
135 136 137 138 139 140 141
                "movdqa %%xmm0, (%0)          \n\t"
                "movdqa %%xmm4, 16(%0)        \n\t"
                :: "r"(&b[i]), "r"(&ref[i])
                : "memory"
            );
        }
        snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
142
        b[0] = b_0 + ((2 * ref[1] + W_BO-1 + 4 * b_0) >> W_BS);
143 144 145
    }

    { // Lift 3
146
        IDWTELEM * const src = b+w2;
147 148

        i = 0;
149
        for(; (((x86_reg)&temp[i]) & 0x1F) && i<w_r; i++){
150 151 152 153
            temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS);
        }
        for(; i<w_r-7; i+=8){
            asm volatile(
154 155 156 157
                "movdqu  2(%1), %%xmm2        \n\t"
                "movdqu 18(%1), %%xmm6        \n\t"
                "paddw    (%1), %%xmm2        \n\t"
                "paddw  16(%1), %%xmm6        \n\t"
158 159
                "movdqu   (%0), %%xmm0        \n\t"
                "movdqu 16(%0), %%xmm4        \n\t"
160 161 162 163 164 165
                "paddw  %%xmm2, %%xmm0        \n\t"
                "paddw  %%xmm6, %%xmm4        \n\t"
                "psraw      $1, %%xmm2        \n\t"
                "psraw      $1, %%xmm6        \n\t"
                "paddw  %%xmm0, %%xmm2        \n\t"
                "paddw  %%xmm4, %%xmm6        \n\t"
166 167 168 169 170 171
                "movdqa %%xmm2, (%2)          \n\t"
                "movdqa %%xmm6, 16(%2)        \n\t"
                :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
                 : "memory"
               );
        }
172
        snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
173 174 175 176 177
    }

    {
        snow_interleave_line_header(&i, width, b, temp);

178
        for (; (i & 0x3E) != 0x3E; i-=2){
179 180 181
            b[i+1] = temp[i>>1];
            b[i] = b[i>>1];
        }
182
        for (i-=62; i>=0; i-=64){
183 184 185 186 187 188 189 190 191
            asm volatile(
                "movdqa      (%1), %%xmm0       \n\t"
                "movdqa    16(%1), %%xmm2       \n\t"
                "movdqa    32(%1), %%xmm4       \n\t"
                "movdqa    48(%1), %%xmm6       \n\t"
                "movdqa      (%1), %%xmm1       \n\t"
                "movdqa    16(%1), %%xmm3       \n\t"
                "movdqa    32(%1), %%xmm5       \n\t"
                "movdqa    48(%1), %%xmm7       \n\t"
192 193 194 195
                "punpcklwd   (%2), %%xmm0       \n\t"
                "punpcklwd 16(%2), %%xmm2       \n\t"
                "punpcklwd 32(%2), %%xmm4       \n\t"
                "punpcklwd 48(%2), %%xmm6       \n\t"
196 197 198 199
                "movdqa    %%xmm0, (%0)         \n\t"
                "movdqa    %%xmm2, 32(%0)       \n\t"
                "movdqa    %%xmm4, 64(%0)       \n\t"
                "movdqa    %%xmm6, 96(%0)       \n\t"
200 201 202 203
                "punpckhwd   (%2), %%xmm1       \n\t"
                "punpckhwd 16(%2), %%xmm3       \n\t"
                "punpckhwd 32(%2), %%xmm5       \n\t"
                "punpckhwd 48(%2), %%xmm7       \n\t"
204 205 206 207 208 209 210 211 212 213 214
                "movdqa    %%xmm1, 16(%0)       \n\t"
                "movdqa    %%xmm3, 48(%0)       \n\t"
                "movdqa    %%xmm5, 80(%0)       \n\t"
                "movdqa    %%xmm7, 112(%0)      \n\t"
                :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1])
                 : "memory"
               );
        }
    }
}

215
void ff_snow_horizontal_compose97i_mmx(IDWTELEM *b, int width){
216
    const int w2= (width+1)>>1;
217
    IDWTELEM temp[width >> 1];
218 219 220 221 222
    const int w_l= (width>>1);
    const int w_r= w2 - 1;
    int i;

    { // Lift 0
223
        IDWTELEM * const ref = b + w2 - 1;
224 225 226 227

        i = 1;
        b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
        asm volatile(
228
            "pcmpeqw    %%mm7, %%mm7         \n\t"
229 230 231 232
            "pcmpeqw    %%mm3, %%mm3         \n\t"
            "psllw         $1, %%mm3         \n\t"
            "paddw      %%mm7, %%mm3         \n\t"
            "psllw        $13, %%mm3         \n\t"
233
           ::);
234
        for(; i<w_l-7; i+=8){
235 236 237
            asm volatile(
                "movq     (%1), %%mm2        \n\t"
                "movq    8(%1), %%mm6        \n\t"
238 239 240 241
                "paddw   2(%1), %%mm2        \n\t"
                "paddw  10(%1), %%mm6        \n\t"
                "paddw   %%mm7, %%mm2        \n\t"
                "paddw   %%mm7, %%mm6        \n\t"
242 243 244 245 246 247
                "pmulhw  %%mm3, %%mm2        \n\t"
                "pmulhw  %%mm3, %%mm6        \n\t"
                "paddw    (%0), %%mm2        \n\t"
                "paddw   8(%0), %%mm6        \n\t"
                "movq    %%mm2, (%0)         \n\t"
                "movq    %%mm6, 8(%0)        \n\t"
248 249 250 251 252 253 254 255
                :: "r"(&b[i]), "r"(&ref[i])
                 : "memory"
               );
        }
        snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
    }

    { // Lift 1
256
        IDWTELEM * const dst = b+w2;
257 258

        i = 0;
259
        for(; i<w_r-7; i+=8){
260 261 262
            asm volatile(
                "movq     (%1), %%mm2        \n\t"
                "movq    8(%1), %%mm6        \n\t"
263 264
                "paddw   2(%1), %%mm2        \n\t"
                "paddw  10(%1), %%mm6        \n\t"
265 266
                "movq     (%0), %%mm0        \n\t"
                "movq    8(%0), %%mm4        \n\t"
267 268
                "psubw   %%mm2, %%mm0        \n\t"
                "psubw   %%mm6, %%mm4        \n\t"
269 270 271 272 273 274 275 276 277 278
                "movq    %%mm0, (%0)         \n\t"
                "movq    %%mm4, 8(%0)        \n\t"
                :: "r"(&dst[i]), "r"(&b[i])
                 : "memory"
               );
        }
        snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
    }

    { // Lift 2
279
        IDWTELEM * const ref = b+w2 - 1;
280 281

        i = 1;
282
        b[0] = b[0] + (((2 * ref[1] + W_BO) + 4 * b[0]) >> W_BS);
283
        asm volatile(
284
            "psllw         $15, %%mm7        \n\t"
285 286 287
            "pcmpeqw     %%mm6, %%mm6        \n\t"
            "psrlw         $13, %%mm6        \n\t"
            "paddw       %%mm7, %%mm6        \n\t"
288
           ::);
289
        for(; i<w_l-7; i+=8){
290 291 292
            asm volatile(
                "movq     (%1), %%mm0        \n\t"
                "movq    8(%1), %%mm4        \n\t"
293 294 295 296 297 298 299 300 301 302 303 304
                "movq    2(%1), %%mm1        \n\t"
                "movq   10(%1), %%mm5        \n\t"
                "paddw   %%mm6, %%mm0        \n\t"
                "paddw   %%mm6, %%mm4        \n\t"
                "paddw   %%mm7, %%mm1        \n\t"
                "paddw   %%mm7, %%mm5        \n\t"
                "pavgw   %%mm1, %%mm0        \n\t"
                "pavgw   %%mm5, %%mm4        \n\t"
                "psubw   %%mm7, %%mm0        \n\t"
                "psubw   %%mm7, %%mm4        \n\t"
                "psraw      $1, %%mm0        \n\t"
                "psraw      $1, %%mm4        \n\t"
305 306
                "movq     (%0), %%mm1        \n\t"
                "movq    8(%0), %%mm5        \n\t"
307 308 309 310 311 312
                "paddw   %%mm1, %%mm0        \n\t"
                "paddw   %%mm5, %%mm4        \n\t"
                "psraw      $2, %%mm0        \n\t"
                "psraw      $2, %%mm4        \n\t"
                "paddw   %%mm1, %%mm0        \n\t"
                "paddw   %%mm5, %%mm4        \n\t"
313 314 315 316 317 318 319 320 321 322
                "movq    %%mm0, (%0)         \n\t"
                "movq    %%mm4, 8(%0)        \n\t"
                :: "r"(&b[i]), "r"(&ref[i])
                 : "memory"
               );
        }
        snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
    }

    { // Lift 3
323
        IDWTELEM * const src = b+w2;
324 325
        i = 0;

326
        for(; i<w_r-7; i+=8){
327
            asm volatile(
328 329 330 331
                "movq    2(%1), %%mm2        \n\t"
                "movq   10(%1), %%mm6        \n\t"
                "paddw    (%1), %%mm2        \n\t"
                "paddw   8(%1), %%mm6        \n\t"
332 333
                "movq     (%0), %%mm0        \n\t"
                "movq    8(%0), %%mm4        \n\t"
334 335 336 337 338 339
                "paddw   %%mm2, %%mm0        \n\t"
                "paddw   %%mm6, %%mm4        \n\t"
                "psraw      $1, %%mm2        \n\t"
                "psraw      $1, %%mm6        \n\t"
                "paddw   %%mm0, %%mm2        \n\t"
                "paddw   %%mm4, %%mm6        \n\t"
340 341 342 343 344 345
                "movq    %%mm2, (%2)         \n\t"
                "movq    %%mm6, 8(%2)        \n\t"
                :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
                 : "memory"
               );
        }
346
        snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
347 348 349 350 351
    }

    {
        snow_interleave_line_header(&i, width, b, temp);

352
        for (; (i & 0x1E) != 0x1E; i-=2){
353 354 355
            b[i+1] = temp[i>>1];
            b[i] = b[i>>1];
        }
356
        for (i-=30; i>=0; i-=32){
357 358 359 360 361 362 363 364 365
            asm volatile(
                "movq        (%1), %%mm0       \n\t"
                "movq       8(%1), %%mm2       \n\t"
                "movq      16(%1), %%mm4       \n\t"
                "movq      24(%1), %%mm6       \n\t"
                "movq        (%1), %%mm1       \n\t"
                "movq       8(%1), %%mm3       \n\t"
                "movq      16(%1), %%mm5       \n\t"
                "movq      24(%1), %%mm7       \n\t"
366 367 368 369
                "punpcklwd   (%2), %%mm0       \n\t"
                "punpcklwd  8(%2), %%mm2       \n\t"
                "punpcklwd 16(%2), %%mm4       \n\t"
                "punpcklwd 24(%2), %%mm6       \n\t"
370 371 372 373
                "movq       %%mm0, (%0)        \n\t"
                "movq       %%mm2, 16(%0)      \n\t"
                "movq       %%mm4, 32(%0)      \n\t"
                "movq       %%mm6, 48(%0)      \n\t"
374 375 376 377
                "punpckhwd   (%2), %%mm1       \n\t"
                "punpckhwd  8(%2), %%mm3       \n\t"
                "punpckhwd 16(%2), %%mm5       \n\t"
                "punpckhwd 24(%2), %%mm7       \n\t"
378 379 380 381 382 383 384 385 386 387 388
                "movq       %%mm1, 8(%0)       \n\t"
                "movq       %%mm3, 24(%0)      \n\t"
                "movq       %%mm5, 40(%0)      \n\t"
                "movq       %%mm7, 56(%0)      \n\t"
                :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1])
                 : "memory"
               );
        }
    }
}

389
#ifdef HAVE_7REGS
390
#define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\
391 392 393 394
        ""op" ("r",%%"REG_d"), %%"t0"      \n\t"\
        ""op" 16("r",%%"REG_d"), %%"t1"    \n\t"\
        ""op" 32("r",%%"REG_d"), %%"t2"    \n\t"\
        ""op" 48("r",%%"REG_d"), %%"t3"    \n\t"
395 396 397 398 399

#define snow_vertical_compose_sse2_load(r,t0,t1,t2,t3)\
        snow_vertical_compose_sse2_load_add("movdqa",r,t0,t1,t2,t3)

#define snow_vertical_compose_sse2_add(r,t0,t1,t2,t3)\
400
        snow_vertical_compose_sse2_load_add("paddw",r,t0,t1,t2,t3)
401

Michael Niedermayer's avatar
Michael Niedermayer committed
402
#define snow_vertical_compose_r2r_sub(s0,s1,s2,s3,t0,t1,t2,t3)\
403 404 405 406
        "psubw %%"s0", %%"t0" \n\t"\
        "psubw %%"s1", %%"t1" \n\t"\
        "psubw %%"s2", %%"t2" \n\t"\
        "psubw %%"s3", %%"t3" \n\t"
407 408

#define snow_vertical_compose_sse2_store(w,s0,s1,s2,s3)\
409 410 411 412
        "movdqa %%"s0", ("w",%%"REG_d")      \n\t"\
        "movdqa %%"s1", 16("w",%%"REG_d")    \n\t"\
        "movdqa %%"s2", 32("w",%%"REG_d")    \n\t"\
        "movdqa %%"s3", 48("w",%%"REG_d")    \n\t"
413

Michael Niedermayer's avatar
Michael Niedermayer committed
414
#define snow_vertical_compose_sra(n,t0,t1,t2,t3)\
415 416 417 418
        "psraw $"n", %%"t0" \n\t"\
        "psraw $"n", %%"t1" \n\t"\
        "psraw $"n", %%"t2" \n\t"\
        "psraw $"n", %%"t3" \n\t"
419

Michael Niedermayer's avatar
Michael Niedermayer committed
420
#define snow_vertical_compose_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\
421 422 423 424
        "paddw %%"s0", %%"t0" \n\t"\
        "paddw %%"s1", %%"t1" \n\t"\
        "paddw %%"s2", %%"t2" \n\t"\
        "paddw %%"s3", %%"t3" \n\t"
425

426 427 428 429 430 431
#define snow_vertical_compose_r2r_pmulhw(s0,s1,s2,s3,t0,t1,t2,t3)\
        "pmulhw %%"s0", %%"t0" \n\t"\
        "pmulhw %%"s1", %%"t1" \n\t"\
        "pmulhw %%"s2", %%"t2" \n\t"\
        "pmulhw %%"s3", %%"t3" \n\t"

432 433 434 435 436 437
#define snow_vertical_compose_sse2_move(s0,s1,s2,s3,t0,t1,t2,t3)\
        "movdqa %%"s0", %%"t0" \n\t"\
        "movdqa %%"s1", %%"t1" \n\t"\
        "movdqa %%"s2", %%"t2" \n\t"\
        "movdqa %%"s3", %%"t3" \n\t"

438
void ff_snow_vertical_compose97i_sse2(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
439
    x86_reg i = width;
440

441
    while(i & 0x1F)
442 443 444 445 446 447 448
    {
        i--;
        b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
        b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
        b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
        b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
    }
449
    i+=i;
450 451 452 453

         asm volatile (
        "jmp 2f                                      \n\t"
        "1:                                          \n\t"
454 455
        snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sse2_add("%6","xmm0","xmm2","xmm4","xmm6")
456 457


458 459 460 461 462 463 464 465
        "pcmpeqw    %%xmm0, %%xmm0                   \n\t"
        "pcmpeqw    %%xmm2, %%xmm2                   \n\t"
        "paddw      %%xmm2, %%xmm2                   \n\t"
        "paddw      %%xmm0, %%xmm2                   \n\t"
        "psllw         $13, %%xmm2                   \n\t"
        snow_vertical_compose_r2r_add("xmm0","xmm0","xmm0","xmm0","xmm1","xmm3","xmm5","xmm7")
        snow_vertical_compose_r2r_pmulhw("xmm2","xmm2","xmm2","xmm2","xmm1","xmm3","xmm5","xmm7")
        snow_vertical_compose_sse2_add("%5","xmm1","xmm3","xmm5","xmm7")
466 467 468
        snow_vertical_compose_sse2_store("%5","xmm1","xmm3","xmm5","xmm7")
        snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sse2_add("%3","xmm1","xmm3","xmm5","xmm7")
Michael Niedermayer's avatar
Michael Niedermayer committed
469
        snow_vertical_compose_r2r_sub("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
470
        snow_vertical_compose_sse2_store("%4","xmm0","xmm2","xmm4","xmm6")
471 472 473 474 475 476

        "pcmpeqw %%xmm7, %%xmm7                      \n\t"
        "pcmpeqw %%xmm5, %%xmm5                      \n\t"
        "psllw $15, %%xmm7                           \n\t"
        "psrlw $13, %%xmm5                           \n\t"
        "paddw %%xmm7, %%xmm5                        \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
477
        snow_vertical_compose_r2r_add("xmm5","xmm5","xmm5","xmm5","xmm0","xmm2","xmm4","xmm6")
478 479
        "movq   (%2,%%"REG_d"), %%xmm1        \n\t"
        "movq  8(%2,%%"REG_d"), %%xmm3        \n\t"
480 481 482 483
        "paddw %%xmm7, %%xmm1                        \n\t"
        "paddw %%xmm7, %%xmm3                        \n\t"
        "pavgw %%xmm1, %%xmm0                        \n\t"
        "pavgw %%xmm3, %%xmm2                        \n\t"
484 485
        "movq 16(%2,%%"REG_d"), %%xmm1        \n\t"
        "movq 24(%2,%%"REG_d"), %%xmm3        \n\t"
486 487 488 489
        "paddw %%xmm7, %%xmm1                        \n\t"
        "paddw %%xmm7, %%xmm3                        \n\t"
        "pavgw %%xmm1, %%xmm4                        \n\t"
        "pavgw %%xmm3, %%xmm6                        \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
490 491
        snow_vertical_compose_r2r_sub("xmm7","xmm7","xmm7","xmm7","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
492
        snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
493

Michael Niedermayer's avatar
Michael Niedermayer committed
494
        snow_vertical_compose_sra("2","xmm0","xmm2","xmm4","xmm6")
495 496 497
        snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sse2_store("%3","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sse2_add("%1","xmm0","xmm2","xmm4","xmm6")
498
        snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
Michael Niedermayer's avatar
Michael Niedermayer committed
499 500
        snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
501 502
        snow_vertical_compose_sse2_add("%2","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sse2_store("%2","xmm0","xmm2","xmm4","xmm6")
503 504

        "2:                                          \n\t"
505
        "sub $64, %%"REG_d"                          \n\t"
506 507
        "jge 1b                                      \n\t"
        :"+d"(i)
508
        :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
509 510 511
}

#define snow_vertical_compose_mmx_load_add(op,r,t0,t1,t2,t3)\
512 513 514 515
        ""op" ("r",%%"REG_d"), %%"t0"   \n\t"\
        ""op" 8("r",%%"REG_d"), %%"t1"  \n\t"\
        ""op" 16("r",%%"REG_d"), %%"t2" \n\t"\
        ""op" 24("r",%%"REG_d"), %%"t3" \n\t"
516 517 518 519 520

#define snow_vertical_compose_mmx_load(r,t0,t1,t2,t3)\
        snow_vertical_compose_mmx_load_add("movq",r,t0,t1,t2,t3)

#define snow_vertical_compose_mmx_add(r,t0,t1,t2,t3)\
521
        snow_vertical_compose_mmx_load_add("paddw",r,t0,t1,t2,t3)
522 523

#define snow_vertical_compose_mmx_store(w,s0,s1,s2,s3)\
524 525 526 527
        "movq %%"s0", ("w",%%"REG_d")   \n\t"\
        "movq %%"s1", 8("w",%%"REG_d")  \n\t"\
        "movq %%"s2", 16("w",%%"REG_d") \n\t"\
        "movq %%"s3", 24("w",%%"REG_d") \n\t"
528 529 530 531 532 533 534

#define snow_vertical_compose_mmx_move(s0,s1,s2,s3,t0,t1,t2,t3)\
        "movq %%"s0", %%"t0" \n\t"\
        "movq %%"s1", %%"t1" \n\t"\
        "movq %%"s2", %%"t2" \n\t"\
        "movq %%"s3", %%"t3" \n\t"

535

536
void ff_snow_vertical_compose97i_mmx(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
537
    x86_reg i = width;
538
    while(i & 15)
539 540 541 542 543 544 545
    {
        i--;
        b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS;
        b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS;
        b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS;
        b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS;
    }
546
    i+=i;
547 548 549 550
    asm volatile(
        "jmp 2f                                      \n\t"
        "1:                                          \n\t"

551 552 553 554 555 556 557 558 559 560
        snow_vertical_compose_mmx_load("%4","mm1","mm3","mm5","mm7")
        snow_vertical_compose_mmx_add("%6","mm1","mm3","mm5","mm7")
        "pcmpeqw    %%mm0, %%mm0                     \n\t"
        "pcmpeqw    %%mm2, %%mm2                     \n\t"
        "paddw      %%mm2, %%mm2                     \n\t"
        "paddw      %%mm0, %%mm2                     \n\t"
        "psllw        $13, %%mm2                     \n\t"
        snow_vertical_compose_r2r_add("mm0","mm0","mm0","mm0","mm1","mm3","mm5","mm7")
        snow_vertical_compose_r2r_pmulhw("mm2","mm2","mm2","mm2","mm1","mm3","mm5","mm7")
        snow_vertical_compose_mmx_add("%5","mm1","mm3","mm5","mm7")
561 562 563
        snow_vertical_compose_mmx_store("%5","mm1","mm3","mm5","mm7")
        snow_vertical_compose_mmx_load("%4","mm0","mm2","mm4","mm6")
        snow_vertical_compose_mmx_add("%3","mm1","mm3","mm5","mm7")
Michael Niedermayer's avatar
Michael Niedermayer committed
564
        snow_vertical_compose_r2r_sub("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
565
        snow_vertical_compose_mmx_store("%4","mm0","mm2","mm4","mm6")
566 567 568 569 570
        "pcmpeqw %%mm7, %%mm7                        \n\t"
        "pcmpeqw %%mm5, %%mm5                        \n\t"
        "psllw $15, %%mm7                            \n\t"
        "psrlw $13, %%mm5                            \n\t"
        "paddw %%mm7, %%mm5                          \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
571
        snow_vertical_compose_r2r_add("mm5","mm5","mm5","mm5","mm0","mm2","mm4","mm6")
572 573
        "movq   (%2,%%"REG_d"), %%mm1         \n\t"
        "movq  8(%2,%%"REG_d"), %%mm3         \n\t"
574 575 576 577
        "paddw %%mm7, %%mm1                          \n\t"
        "paddw %%mm7, %%mm3                          \n\t"
        "pavgw %%mm1, %%mm0                          \n\t"
        "pavgw %%mm3, %%mm2                          \n\t"
578 579
        "movq 16(%2,%%"REG_d"), %%mm1         \n\t"
        "movq 24(%2,%%"REG_d"), %%mm3         \n\t"
580 581 582 583
        "paddw %%mm7, %%mm1                          \n\t"
        "paddw %%mm7, %%mm3                          \n\t"
        "pavgw %%mm1, %%mm4                          \n\t"
        "pavgw %%mm3, %%mm6                          \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
584 585
        snow_vertical_compose_r2r_sub("mm7","mm7","mm7","mm7","mm0","mm2","mm4","mm6")
        snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
586
        snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
587

Michael Niedermayer's avatar
Michael Niedermayer committed
588
        snow_vertical_compose_sra("2","mm0","mm2","mm4","mm6")
589 590 591
        snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
        snow_vertical_compose_mmx_store("%3","mm0","mm2","mm4","mm6")
        snow_vertical_compose_mmx_add("%1","mm0","mm2","mm4","mm6")
592
        snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
Michael Niedermayer's avatar
Michael Niedermayer committed
593 594
        snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
        snow_vertical_compose_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
595 596
        snow_vertical_compose_mmx_add("%2","mm0","mm2","mm4","mm6")
        snow_vertical_compose_mmx_store("%2","mm0","mm2","mm4","mm6")
597 598

        "2:                                          \n\t"
599
        "sub $32, %%"REG_d"                          \n\t"
600 601
        "jge 1b                                      \n\t"
        :"+d"(i)
602
        :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
603
}
604
#endif //HAVE_7REGS
605 606

#define snow_inner_add_yblock_sse2_header \
607
    IDWTELEM * * dst_array = sb->line + src_y;\
608
    x86_reg tmp;\
609
    asm volatile(\
610 611 612
             "mov  %7, %%"REG_c"             \n\t"\
             "mov  %6, %2                    \n\t"\
             "mov  %4, %%"REG_S"             \n\t"\
613 614
             "pxor %%xmm7, %%xmm7            \n\t" /* 0 */\
             "pcmpeqd %%xmm3, %%xmm3         \n\t"\
615 616
             "psllw $15, %%xmm3              \n\t"\
             "psrlw $12, %%xmm3              \n\t" /* FRAC_BITS >> 1 */\
617 618 619
             "1:                             \n\t"\
             "mov %1, %%"REG_D"              \n\t"\
             "mov (%%"REG_D"), %%"REG_D"     \n\t"\
620
             "add %3, %%"REG_D"              \n\t"
621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667

#define snow_inner_add_yblock_sse2_start_8(out_reg1, out_reg2, ptr_offset, s_offset)\
             "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
             "movq (%%"REG_d"), %%"out_reg1" \n\t"\
             "movq (%%"REG_d", %%"REG_c"), %%"out_reg2" \n\t"\
             "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
             "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
             "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
             "movq "s_offset"+16(%%"REG_S"), %%xmm4 \n\t"\
             "punpcklbw %%xmm7, %%xmm0       \n\t"\
             "punpcklbw %%xmm7, %%xmm4       \n\t"\
             "pmullw %%xmm0, %%"out_reg1"    \n\t"\
             "pmullw %%xmm4, %%"out_reg2"    \n\t"

#define snow_inner_add_yblock_sse2_start_16(out_reg1, out_reg2, ptr_offset, s_offset)\
             "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
             "movq (%%"REG_d"), %%"out_reg1" \n\t"\
             "movq 8(%%"REG_d"), %%"out_reg2" \n\t"\
             "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
             "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
             "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
             "movq "s_offset"+8(%%"REG_S"), %%xmm4 \n\t"\
             "punpcklbw %%xmm7, %%xmm0       \n\t"\
             "punpcklbw %%xmm7, %%xmm4       \n\t"\
             "pmullw %%xmm0, %%"out_reg1"    \n\t"\
             "pmullw %%xmm4, %%"out_reg2"    \n\t"

#define snow_inner_add_yblock_sse2_accum_8(ptr_offset, s_offset) \
             snow_inner_add_yblock_sse2_start_8("xmm2", "xmm6", ptr_offset, s_offset)\
             "paddusw %%xmm2, %%xmm1         \n\t"\
             "paddusw %%xmm6, %%xmm5         \n\t"

#define snow_inner_add_yblock_sse2_accum_16(ptr_offset, s_offset) \
             snow_inner_add_yblock_sse2_start_16("xmm2", "xmm6", ptr_offset, s_offset)\
             "paddusw %%xmm2, %%xmm1         \n\t"\
             "paddusw %%xmm6, %%xmm5         \n\t"

#define snow_inner_add_yblock_sse2_end_common1\
             "add $32, %%"REG_S"             \n\t"\
             "add %%"REG_c", %0              \n\t"\
             "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
             "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
             "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
             "add %%"REG_c", (%%"REG_a")     \n\t"

#define snow_inner_add_yblock_sse2_end_common2\
             "jnz 1b                         \n\t"\
668
             :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
669
             :\
670
             "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"((x86_reg)b_h),"m"((x86_reg)src_stride):\
671
             "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
672 673 674 675 676 677

#define snow_inner_add_yblock_sse2_end_8\
             "sal $1, %%"REG_c"              \n\t"\
             "add $"PTR_SIZE"*2, %1          \n\t"\
             snow_inner_add_yblock_sse2_end_common1\
             "sar $1, %%"REG_c"              \n\t"\
678
             "sub $2, %2                     \n\t"\
679 680 681 682 683
             snow_inner_add_yblock_sse2_end_common2

#define snow_inner_add_yblock_sse2_end_16\
             "add $"PTR_SIZE"*1, %1          \n\t"\
             snow_inner_add_yblock_sse2_end_common1\
684
             "dec %2                         \n\t"\
685 686
             snow_inner_add_yblock_sse2_end_common2

687 688
static void inner_add_yblock_bw_8_obmc_16_bh_even_sse2(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
                      int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708
snow_inner_add_yblock_sse2_header
snow_inner_add_yblock_sse2_start_8("xmm1", "xmm5", "3", "0")
snow_inner_add_yblock_sse2_accum_8("2", "8")
snow_inner_add_yblock_sse2_accum_8("1", "128")
snow_inner_add_yblock_sse2_accum_8("0", "136")

             "mov %0, %%"REG_d"              \n\t"
             "movdqa (%%"REG_D"), %%xmm0     \n\t"
             "movdqa %%xmm1, %%xmm2          \n\t"

             "punpckhwd %%xmm7, %%xmm1       \n\t"
             "punpcklwd %%xmm7, %%xmm2       \n\t"
             "paddd %%xmm2, %%xmm0           \n\t"
             "movdqa 16(%%"REG_D"), %%xmm2   \n\t"
             "paddd %%xmm1, %%xmm2           \n\t"
             "paddd %%xmm3, %%xmm0           \n\t"
             "paddd %%xmm3, %%xmm2           \n\t"

             "mov %1, %%"REG_D"              \n\t"
             "mov "PTR_SIZE"(%%"REG_D"), %%"REG_D";\n\t"
709
             "add %3, %%"REG_D"              \n\t"
710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734

             "movdqa (%%"REG_D"), %%xmm4     \n\t"
             "movdqa %%xmm5, %%xmm6          \n\t"
             "punpckhwd %%xmm7, %%xmm5       \n\t"
             "punpcklwd %%xmm7, %%xmm6       \n\t"
             "paddd %%xmm6, %%xmm4           \n\t"
             "movdqa 16(%%"REG_D"), %%xmm6   \n\t"
             "paddd %%xmm5, %%xmm6           \n\t"
             "paddd %%xmm3, %%xmm4           \n\t"
             "paddd %%xmm3, %%xmm6           \n\t"

             "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
             "psrad $8, %%xmm2               \n\t" /* FRAC_BITS. */
             "packssdw %%xmm2, %%xmm0        \n\t"
             "packuswb %%xmm7, %%xmm0        \n\t"
             "movq %%xmm0, (%%"REG_d")       \n\t"

             "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
             "psrad $8, %%xmm6               \n\t" /* FRAC_BITS. */
             "packssdw %%xmm6, %%xmm4        \n\t"
             "packuswb %%xmm7, %%xmm4        \n\t"
             "movq %%xmm4, (%%"REG_d",%%"REG_c");\n\t"
snow_inner_add_yblock_sse2_end_8
}

735 736
static void inner_add_yblock_bw_16_obmc_32_sse2(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
                      int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
737 738 739 740 741 742 743
snow_inner_add_yblock_sse2_header
snow_inner_add_yblock_sse2_start_16("xmm1", "xmm5", "3", "0")
snow_inner_add_yblock_sse2_accum_16("2", "16")
snow_inner_add_yblock_sse2_accum_16("1", "512")
snow_inner_add_yblock_sse2_accum_16("0", "528")

             "mov %0, %%"REG_d"              \n\t"
744 745 746 747 748 749 750 751 752 753 754
             "psrlw $4, %%xmm1               \n\t"
             "psrlw $4, %%xmm5               \n\t"
             "paddw   (%%"REG_D"), %%xmm1    \n\t"
             "paddw 16(%%"REG_D"), %%xmm5    \n\t"
             "paddw %%xmm3, %%xmm1           \n\t"
             "paddw %%xmm3, %%xmm5           \n\t"
             "psraw $4, %%xmm1               \n\t" /* FRAC_BITS. */
             "psraw $4, %%xmm5               \n\t" /* FRAC_BITS. */
             "packuswb %%xmm5, %%xmm1        \n\t"

             "movdqu %%xmm1, (%%"REG_d")       \n\t"
755 756 757 758 759

snow_inner_add_yblock_sse2_end_16
}

#define snow_inner_add_yblock_mmx_header \
760
    IDWTELEM * * dst_array = sb->line + src_y;\
761
    x86_reg tmp;\
762
    asm volatile(\
763 764 765
             "mov  %7, %%"REG_c"             \n\t"\
             "mov  %6, %2                    \n\t"\
             "mov  %4, %%"REG_S"             \n\t"\
766 767
             "pxor %%mm7, %%mm7              \n\t" /* 0 */\
             "pcmpeqd %%mm3, %%mm3           \n\t"\
768 769
             "psllw $15, %%mm3               \n\t"\
             "psrlw $12, %%mm3               \n\t" /* FRAC_BITS >> 1 */\
770 771 772
             "1:                             \n\t"\
             "mov %1, %%"REG_D"              \n\t"\
             "mov (%%"REG_D"), %%"REG_D"     \n\t"\
773
             "add %3, %%"REG_D"              \n\t"
774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794

#define snow_inner_add_yblock_mmx_start(out_reg1, out_reg2, ptr_offset, s_offset, d_offset)\
             "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
             "movd "d_offset"(%%"REG_d"), %%"out_reg1" \n\t"\
             "movd "d_offset"+4(%%"REG_d"), %%"out_reg2" \n\t"\
             "punpcklbw %%mm7, %%"out_reg1" \n\t"\
             "punpcklbw %%mm7, %%"out_reg2" \n\t"\
             "movd "s_offset"(%%"REG_S"), %%mm0 \n\t"\
             "movd "s_offset"+4(%%"REG_S"), %%mm4 \n\t"\
             "punpcklbw %%mm7, %%mm0       \n\t"\
             "punpcklbw %%mm7, %%mm4       \n\t"\
             "pmullw %%mm0, %%"out_reg1"    \n\t"\
             "pmullw %%mm4, %%"out_reg2"    \n\t"

#define snow_inner_add_yblock_mmx_accum(ptr_offset, s_offset, d_offset) \
             snow_inner_add_yblock_mmx_start("mm2", "mm6", ptr_offset, s_offset, d_offset)\
             "paddusw %%mm2, %%mm1         \n\t"\
             "paddusw %%mm6, %%mm5         \n\t"

#define snow_inner_add_yblock_mmx_mix(read_offset, write_offset)\
             "mov %0, %%"REG_d"              \n\t"\
795 796 797 798 799 800 801 802 803 804
             "psrlw $4, %%mm1                \n\t"\
             "psrlw $4, %%mm5                \n\t"\
             "paddw "read_offset"(%%"REG_D"), %%mm1 \n\t"\
             "paddw "read_offset"+8(%%"REG_D"), %%mm5 \n\t"\
             "paddw %%mm3, %%mm1             \n\t"\
             "paddw %%mm3, %%mm5             \n\t"\
             "psraw $4, %%mm1                \n\t"\
             "psraw $4, %%mm5                \n\t"\
             "packuswb %%mm5, %%mm1          \n\t"\
             "movq %%mm1, "write_offset"(%%"REG_d") \n\t"
805 806 807 808 809 810 811 812 813

#define snow_inner_add_yblock_mmx_end(s_step)\
             "add $"s_step", %%"REG_S"             \n\t"\
             "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
             "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
             "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
             "add %%"REG_c", (%%"REG_a")     \n\t"\
             "add $"PTR_SIZE"*1, %1          \n\t"\
             "add %%"REG_c", %0              \n\t"\
814
             "dec %2                         \n\t"\
815
             "jnz 1b                         \n\t"\
816
             :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
817
             :\
818
             "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"((x86_reg)b_h),"m"((x86_reg)src_stride):\
819
             "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
820

821 822
static void inner_add_yblock_bw_8_obmc_16_mmx(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
                      int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
823 824 825 826 827 828 829 830 831
snow_inner_add_yblock_mmx_header
snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
snow_inner_add_yblock_mmx_accum("2", "8", "0")
snow_inner_add_yblock_mmx_accum("1", "128", "0")
snow_inner_add_yblock_mmx_accum("0", "136", "0")
snow_inner_add_yblock_mmx_mix("0", "0")
snow_inner_add_yblock_mmx_end("16")
}

832 833
static void inner_add_yblock_bw_16_obmc_32_mmx(const uint8_t *obmc, const x86_reg obmc_stride, uint8_t * * block, int b_w, x86_reg b_h,
                      int src_x, int src_y, x86_reg src_stride, slice_buffer * sb, int add, uint8_t * dst8){
834 835 836 837 838 839 840 841 842 843 844
snow_inner_add_yblock_mmx_header
snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
snow_inner_add_yblock_mmx_accum("2", "16", "0")
snow_inner_add_yblock_mmx_accum("1", "512", "0")
snow_inner_add_yblock_mmx_accum("0", "528", "0")
snow_inner_add_yblock_mmx_mix("0", "0")

snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "8", "8")
snow_inner_add_yblock_mmx_accum("2", "24", "8")
snow_inner_add_yblock_mmx_accum("1", "520", "8")
snow_inner_add_yblock_mmx_accum("0", "536", "8")
845
snow_inner_add_yblock_mmx_mix("16", "8")
846 847 848
snow_inner_add_yblock_mmx_end("32")
}

849
void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
850 851 852 853 854 855 856 857 858 859 860 861 862
                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){

    if (b_w == 16)
        inner_add_yblock_bw_16_obmc_32_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
    else if (b_w == 8 && obmc_stride == 16) {
        if (!(b_h & 1))
            inner_add_yblock_bw_8_obmc_16_bh_even_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
        else
            inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
    } else
         ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
}

863
void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
864 865 866 867 868 869 870 871
                          int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
    if (b_w == 16)
        inner_add_yblock_bw_16_obmc_32_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
    else if (b_w == 8 && obmc_stride == 16)
        inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
    else
        ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
}