snowdsp_mmx.c 39.4 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
#include "libavutil/cpu.h"
23 24 25
#include "libavutil/x86_cpu.h"
#include "libavcodec/avcodec.h"
#include "libavcodec/snow.h"
26
#include "libavcodec/dwt.h"
27
#include "dsputil_mmx.h"
28

29
static void ff_snow_horizontal_compose97i_sse2(IDWTELEM *b, int width){
30
    const int w2= (width+1)>>1;
31
    DECLARE_ALIGNED(16, IDWTELEM, temp)[width>>1];
32 33 34 35 36
    const int w_l= (width>>1);
    const int w_r= w2 - 1;
    int i;

    { // Lift 0
37 38
        IDWTELEM * const ref = b + w2 - 1;
        IDWTELEM b_0 = b[0]; //By allowing the first entry in b[0] to be calculated twice
39 40 41 42 43
        // (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;
44
        __asm__ volatile(
45
            "pcmpeqd   %%xmm7, %%xmm7         \n\t"
46 47 48
            "pcmpeqd   %%xmm3, %%xmm3         \n\t"
            "psllw         $1, %%xmm3         \n\t"
            "paddw     %%xmm7, %%xmm3         \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
49
            "psllw        $13, %%xmm3         \n\t"
50
        ::);
51
        for(; i<w_l-15; i+=16){
52
            __asm__ volatile(
53 54
                "movdqu   (%1), %%xmm1        \n\t"
                "movdqu 16(%1), %%xmm5        \n\t"
55 56 57 58 59 60
                "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"
61 62 63 64 65 66
                "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"
67 68 69 70 71 72 73 74 75
                :: "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
76
        IDWTELEM * const dst = b+w2;
77 78

        i = 0;
79
        for(; (((x86_reg)&dst[i]) & 0x1F) && i<w_r; i++){
80 81
            dst[i] = dst[i] - (b[i] + b[i + 1]);
        }
82
        for(; i<w_r-15; i+=16){
83
            __asm__ volatile(
84 85
                "movdqu   (%1), %%xmm1        \n\t"
                "movdqu 16(%1), %%xmm5        \n\t"
86 87 88 89
                "movdqu  2(%1), %%xmm2        \n\t"
                "movdqu 18(%1), %%xmm6        \n\t"
                "paddw  %%xmm1, %%xmm2        \n\t"
                "paddw  %%xmm5, %%xmm6        \n\t"
90 91
                "movdqa   (%0), %%xmm0        \n\t"
                "movdqa 16(%0), %%xmm4        \n\t"
92 93
                "psubw  %%xmm2, %%xmm0        \n\t"
                "psubw  %%xmm6, %%xmm4        \n\t"
94 95 96 97 98 99 100 101 102 103
                "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
104 105
        IDWTELEM * const ref = b+w2 - 1;
        IDWTELEM b_0 = b[0];
106 107

        i = 0;
108
        __asm__ volatile(
109
            "psllw         $15, %%xmm7        \n\t"
110 111 112
            "pcmpeqw    %%xmm6, %%xmm6        \n\t"
            "psrlw         $13, %%xmm6        \n\t"
            "paddw      %%xmm7, %%xmm6        \n\t"
113
        ::);
114
        for(; i<w_l-15; i+=16){
115
            __asm__ volatile(
116 117 118 119 120 121 122 123 124 125 126 127 128 129
                "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"
130 131
                "movdqa   (%0), %%xmm1        \n\t"
                "movdqa 16(%0), %%xmm5        \n\t"
132 133 134 135 136 137
                "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"
138 139 140 141 142 143 144
                "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);
145
        b[0] = b_0 + ((2 * ref[1] + W_BO-1 + 4 * b_0) >> W_BS);
146 147 148
    }

    { // Lift 3
149
        IDWTELEM * const src = b+w2;
150 151

        i = 0;
152
        for(; (((x86_reg)&temp[i]) & 0x1F) && i<w_r; i++){
153 154 155
            temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS);
        }
        for(; i<w_r-7; i+=8){
156
            __asm__ volatile(
157 158 159 160
                "movdqu  2(%1), %%xmm2        \n\t"
                "movdqu 18(%1), %%xmm6        \n\t"
                "paddw    (%1), %%xmm2        \n\t"
                "paddw  16(%1), %%xmm6        \n\t"
161 162
                "movdqu   (%0), %%xmm0        \n\t"
                "movdqu 16(%0), %%xmm4        \n\t"
163 164 165 166 167 168
                "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"
169 170 171 172 173 174
                "movdqa %%xmm2, (%2)          \n\t"
                "movdqa %%xmm6, 16(%2)        \n\t"
                :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
                 : "memory"
               );
        }
175
        snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
176 177 178 179 180
    }

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

181
        for (; (i & 0x3E) != 0x3E; i-=2){
182 183 184
            b[i+1] = temp[i>>1];
            b[i] = b[i>>1];
        }
185
        for (i-=62; i>=0; i-=64){
186
            __asm__ volatile(
187 188 189 190 191 192 193 194
                "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"
195 196 197 198
                "punpcklwd   (%2), %%xmm0       \n\t"
                "punpcklwd 16(%2), %%xmm2       \n\t"
                "punpcklwd 32(%2), %%xmm4       \n\t"
                "punpcklwd 48(%2), %%xmm6       \n\t"
199 200 201 202
                "movdqa    %%xmm0, (%0)         \n\t"
                "movdqa    %%xmm2, 32(%0)       \n\t"
                "movdqa    %%xmm4, 64(%0)       \n\t"
                "movdqa    %%xmm6, 96(%0)       \n\t"
203 204 205 206
                "punpckhwd   (%2), %%xmm1       \n\t"
                "punpckhwd 16(%2), %%xmm3       \n\t"
                "punpckhwd 32(%2), %%xmm5       \n\t"
                "punpckhwd 48(%2), %%xmm7       \n\t"
207 208 209 210 211 212 213 214 215 216 217
                "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"
               );
        }
    }
}

218
static void ff_snow_horizontal_compose97i_mmx(IDWTELEM *b, int width){
219
    const int w2= (width+1)>>1;
220
    IDWTELEM temp[width >> 1];
221 222 223 224 225
    const int w_l= (width>>1);
    const int w_r= w2 - 1;
    int i;

    { // Lift 0
226
        IDWTELEM * const ref = b + w2 - 1;
227 228 229

        i = 1;
        b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
230
        __asm__ volatile(
231
            "pcmpeqw    %%mm7, %%mm7         \n\t"
232 233 234 235
            "pcmpeqw    %%mm3, %%mm3         \n\t"
            "psllw         $1, %%mm3         \n\t"
            "paddw      %%mm7, %%mm3         \n\t"
            "psllw        $13, %%mm3         \n\t"
236
           ::);
237
        for(; i<w_l-7; i+=8){
238
            __asm__ volatile(
239 240
                "movq     (%1), %%mm2        \n\t"
                "movq    8(%1), %%mm6        \n\t"
241 242 243 244
                "paddw   2(%1), %%mm2        \n\t"
                "paddw  10(%1), %%mm6        \n\t"
                "paddw   %%mm7, %%mm2        \n\t"
                "paddw   %%mm7, %%mm6        \n\t"
245 246 247 248 249 250
                "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"
251 252 253 254 255 256 257 258
                :: "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
259
        IDWTELEM * const dst = b+w2;
260 261

        i = 0;
262
        for(; i<w_r-7; i+=8){
263
            __asm__ volatile(
264 265
                "movq     (%1), %%mm2        \n\t"
                "movq    8(%1), %%mm6        \n\t"
266 267
                "paddw   2(%1), %%mm2        \n\t"
                "paddw  10(%1), %%mm6        \n\t"
268 269
                "movq     (%0), %%mm0        \n\t"
                "movq    8(%0), %%mm4        \n\t"
270 271
                "psubw   %%mm2, %%mm0        \n\t"
                "psubw   %%mm6, %%mm4        \n\t"
272 273 274 275 276 277 278 279 280 281
                "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
282
        IDWTELEM * const ref = b+w2 - 1;
283 284

        i = 1;
285
        b[0] = b[0] + (((2 * ref[1] + W_BO) + 4 * b[0]) >> W_BS);
286
        __asm__ volatile(
287
            "psllw         $15, %%mm7        \n\t"
288 289 290
            "pcmpeqw     %%mm6, %%mm6        \n\t"
            "psrlw         $13, %%mm6        \n\t"
            "paddw       %%mm7, %%mm6        \n\t"
291
           ::);
292
        for(; i<w_l-7; i+=8){
293
            __asm__ volatile(
294 295
                "movq     (%1), %%mm0        \n\t"
                "movq    8(%1), %%mm4        \n\t"
296 297 298 299 300 301 302 303 304 305 306 307
                "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"
308 309
                "movq     (%0), %%mm1        \n\t"
                "movq    8(%0), %%mm5        \n\t"
310 311 312 313 314 315
                "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"
316 317 318 319 320 321 322 323 324 325
                "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
326
        IDWTELEM * const src = b+w2;
327 328
        i = 0;

329
        for(; i<w_r-7; i+=8){
330
            __asm__ volatile(
331 332 333 334
                "movq    2(%1), %%mm2        \n\t"
                "movq   10(%1), %%mm6        \n\t"
                "paddw    (%1), %%mm2        \n\t"
                "paddw   8(%1), %%mm6        \n\t"
335 336
                "movq     (%0), %%mm0        \n\t"
                "movq    8(%0), %%mm4        \n\t"
337 338 339 340 341 342
                "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"
343 344 345 346 347 348
                "movq    %%mm2, (%2)         \n\t"
                "movq    %%mm6, 8(%2)        \n\t"
                :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
                 : "memory"
               );
        }
349
        snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS);
350 351 352 353 354
    }

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

355
        for (; (i & 0x1E) != 0x1E; i-=2){
356 357 358
            b[i+1] = temp[i>>1];
            b[i] = b[i>>1];
        }
359
        for (i-=30; i>=0; i-=32){
360
            __asm__ volatile(
361 362 363 364 365 366 367 368
                "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"
369 370 371 372
                "punpcklwd   (%2), %%mm0       \n\t"
                "punpcklwd  8(%2), %%mm2       \n\t"
                "punpcklwd 16(%2), %%mm4       \n\t"
                "punpcklwd 24(%2), %%mm6       \n\t"
373 374 375 376
                "movq       %%mm0, (%0)        \n\t"
                "movq       %%mm2, 16(%0)      \n\t"
                "movq       %%mm4, 32(%0)      \n\t"
                "movq       %%mm6, 48(%0)      \n\t"
377 378 379 380
                "punpckhwd   (%2), %%mm1       \n\t"
                "punpckhwd  8(%2), %%mm3       \n\t"
                "punpckhwd 16(%2), %%mm5       \n\t"
                "punpckhwd 24(%2), %%mm7       \n\t"
381 382 383 384 385 386 387 388 389 390 391
                "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"
               );
        }
    }
}

392
#if HAVE_7REGS
393
#define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\
394 395 396 397
        ""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"
398 399 400 401 402

#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)\
403
        snow_vertical_compose_sse2_load_add("paddw",r,t0,t1,t2,t3)
404

Michael Niedermayer's avatar
Michael Niedermayer committed
405
#define snow_vertical_compose_r2r_sub(s0,s1,s2,s3,t0,t1,t2,t3)\
406 407 408 409
        "psubw %%"s0", %%"t0" \n\t"\
        "psubw %%"s1", %%"t1" \n\t"\
        "psubw %%"s2", %%"t2" \n\t"\
        "psubw %%"s3", %%"t3" \n\t"
410 411

#define snow_vertical_compose_sse2_store(w,s0,s1,s2,s3)\
412 413 414 415
        "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"
416

Michael Niedermayer's avatar
Michael Niedermayer committed
417
#define snow_vertical_compose_sra(n,t0,t1,t2,t3)\
418 419 420 421
        "psraw $"n", %%"t0" \n\t"\
        "psraw $"n", %%"t1" \n\t"\
        "psraw $"n", %%"t2" \n\t"\
        "psraw $"n", %%"t3" \n\t"
422

Michael Niedermayer's avatar
Michael Niedermayer committed
423
#define snow_vertical_compose_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\
424 425 426 427
        "paddw %%"s0", %%"t0" \n\t"\
        "paddw %%"s1", %%"t1" \n\t"\
        "paddw %%"s2", %%"t2" \n\t"\
        "paddw %%"s3", %%"t3" \n\t"
428

429 430 431 432 433 434
#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"

435 436 437 438 439 440
#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"

441
static void ff_snow_vertical_compose97i_sse2(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
442
    x86_reg i = width;
443

444
    while(i & 0x1F)
445 446 447 448 449 450 451
    {
        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;
    }
452
    i+=i;
453

454
         __asm__ volatile (
455 456
        "jmp 2f                                      \n\t"
        "1:                                          \n\t"
457 458
        snow_vertical_compose_sse2_load("%4","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sse2_add("%6","xmm0","xmm2","xmm4","xmm6")
459 460


461 462 463 464 465 466 467 468
        "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")
469 470 471
        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
472
        snow_vertical_compose_r2r_sub("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
473
        snow_vertical_compose_sse2_store("%4","xmm0","xmm2","xmm4","xmm6")
474 475 476 477 478 479

        "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
480
        snow_vertical_compose_r2r_add("xmm5","xmm5","xmm5","xmm5","xmm0","xmm2","xmm4","xmm6")
481 482
        "movq   (%2,%%"REG_d"), %%xmm1        \n\t"
        "movq  8(%2,%%"REG_d"), %%xmm3        \n\t"
483 484 485 486
        "paddw %%xmm7, %%xmm1                        \n\t"
        "paddw %%xmm7, %%xmm3                        \n\t"
        "pavgw %%xmm1, %%xmm0                        \n\t"
        "pavgw %%xmm3, %%xmm2                        \n\t"
487 488
        "movq 16(%2,%%"REG_d"), %%xmm1        \n\t"
        "movq 24(%2,%%"REG_d"), %%xmm3        \n\t"
489 490 491 492
        "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
493 494
        snow_vertical_compose_r2r_sub("xmm7","xmm7","xmm7","xmm7","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
495
        snow_vertical_compose_sse2_add("%3","xmm0","xmm2","xmm4","xmm6")
496

Michael Niedermayer's avatar
Michael Niedermayer committed
497
        snow_vertical_compose_sra("2","xmm0","xmm2","xmm4","xmm6")
498 499 500
        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")
501
        snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7")
Michael Niedermayer's avatar
Michael Niedermayer committed
502 503
        snow_vertical_compose_sra("1","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6")
504 505
        snow_vertical_compose_sse2_add("%2","xmm0","xmm2","xmm4","xmm6")
        snow_vertical_compose_sse2_store("%2","xmm0","xmm2","xmm4","xmm6")
506 507

        "2:                                          \n\t"
508
        "sub $64, %%"REG_d"                          \n\t"
509 510
        "jge 1b                                      \n\t"
        :"+d"(i)
511
        :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
512 513 514
}

#define snow_vertical_compose_mmx_load_add(op,r,t0,t1,t2,t3)\
515 516 517 518
        ""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"
519 520 521 522 523

#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)\
524
        snow_vertical_compose_mmx_load_add("paddw",r,t0,t1,t2,t3)
525 526

#define snow_vertical_compose_mmx_store(w,s0,s1,s2,s3)\
527 528 529 530
        "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"
531 532 533 534 535 536 537

#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"

538

539
static void ff_snow_vertical_compose97i_mmx(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width){
540
    x86_reg i = width;
541
    while(i & 15)
542 543 544 545 546 547 548
    {
        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;
    }
549
    i+=i;
550
    __asm__ volatile(
551 552 553
        "jmp 2f                                      \n\t"
        "1:                                          \n\t"

554 555 556 557 558 559 560 561 562 563
        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")
564 565 566
        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
567
        snow_vertical_compose_r2r_sub("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
568
        snow_vertical_compose_mmx_store("%4","mm0","mm2","mm4","mm6")
569 570 571 572 573
        "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
574
        snow_vertical_compose_r2r_add("mm5","mm5","mm5","mm5","mm0","mm2","mm4","mm6")
575 576
        "movq   (%2,%%"REG_d"), %%mm1         \n\t"
        "movq  8(%2,%%"REG_d"), %%mm3         \n\t"
577 578 579 580
        "paddw %%mm7, %%mm1                          \n\t"
        "paddw %%mm7, %%mm3                          \n\t"
        "pavgw %%mm1, %%mm0                          \n\t"
        "pavgw %%mm3, %%mm2                          \n\t"
581 582
        "movq 16(%2,%%"REG_d"), %%mm1         \n\t"
        "movq 24(%2,%%"REG_d"), %%mm3         \n\t"
583 584 585 586
        "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
587 588
        snow_vertical_compose_r2r_sub("mm7","mm7","mm7","mm7","mm0","mm2","mm4","mm6")
        snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
589
        snow_vertical_compose_mmx_add("%3","mm0","mm2","mm4","mm6")
590

Michael Niedermayer's avatar
Michael Niedermayer committed
591
        snow_vertical_compose_sra("2","mm0","mm2","mm4","mm6")
592 593 594
        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")
595
        snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7")
Michael Niedermayer's avatar
Michael Niedermayer committed
596 597
        snow_vertical_compose_sra("1","mm0","mm2","mm4","mm6")
        snow_vertical_compose_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6")
598 599
        snow_vertical_compose_mmx_add("%2","mm0","mm2","mm4","mm6")
        snow_vertical_compose_mmx_store("%2","mm0","mm2","mm4","mm6")
600 601

        "2:                                          \n\t"
602
        "sub $32, %%"REG_d"                          \n\t"
603 604
        "jge 1b                                      \n\t"
        :"+d"(i)
605
        :"r"(b0),"r"(b1),"r"(b2),"r"(b3),"r"(b4),"r"(b5));
606
}
607
#endif //HAVE_7REGS
608 609

#define snow_inner_add_yblock_sse2_header \
610
    IDWTELEM * * dst_array = sb->line + src_y;\
611
    x86_reg tmp;\
612
    __asm__ volatile(\
613 614 615
             "mov  %7, %%"REG_c"             \n\t"\
             "mov  %6, %2                    \n\t"\
             "mov  %4, %%"REG_S"             \n\t"\
616 617
             "pxor %%xmm7, %%xmm7            \n\t" /* 0 */\
             "pcmpeqd %%xmm3, %%xmm3         \n\t"\
618 619
             "psllw $15, %%xmm3              \n\t"\
             "psrlw $12, %%xmm3              \n\t" /* FRAC_BITS >> 1 */\
620 621 622
             "1:                             \n\t"\
             "mov %1, %%"REG_D"              \n\t"\
             "mov (%%"REG_D"), %%"REG_D"     \n\t"\
623
             "add %3, %%"REG_D"              \n\t"
624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670

#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"\
671
             :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
672
             :\
673
             "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
674
             "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
675 676 677 678 679 680

#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"\
681
             "sub $2, %2                     \n\t"\
682 683 684 685 686
             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\
687
             "dec %2                         \n\t"\
688 689
             snow_inner_add_yblock_sse2_end_common2

690 691
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){
692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711
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"
712
             "add %3, %%"REG_D"              \n\t"
713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737

             "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
}

738 739
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){
740 741 742 743 744 745 746
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"
747 748 749 750 751 752 753 754 755 756 757
             "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"
758 759 760 761 762

snow_inner_add_yblock_sse2_end_16
}

#define snow_inner_add_yblock_mmx_header \
763
    IDWTELEM * * dst_array = sb->line + src_y;\
764
    x86_reg tmp;\
765
    __asm__ volatile(\
766 767 768
             "mov  %7, %%"REG_c"             \n\t"\
             "mov  %6, %2                    \n\t"\
             "mov  %4, %%"REG_S"             \n\t"\
769 770
             "pxor %%mm7, %%mm7              \n\t" /* 0 */\
             "pcmpeqd %%mm3, %%mm3           \n\t"\
771 772
             "psllw $15, %%mm3               \n\t"\
             "psrlw $12, %%mm3               \n\t" /* FRAC_BITS >> 1 */\
773 774 775
             "1:                             \n\t"\
             "mov %1, %%"REG_D"              \n\t"\
             "mov (%%"REG_D"), %%"REG_D"     \n\t"\
776
             "add %3, %%"REG_D"              \n\t"
777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797

#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"\
798 799 800 801 802 803 804 805 806 807
             "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"
808 809 810 811 812 813 814

#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"\
815
             "add"OPSIZE " $"PTR_SIZE"*1, %1 \n\t"\
816
             "add %%"REG_c", %0              \n\t"\
817
             "dec %2                         \n\t"\
818
             "jnz 1b                         \n\t"\
819
             :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\
820
             :\
821
             "rm"((x86_reg)(src_x<<1)),"m"(obmc),"a"(block),"m"(b_h),"m"(src_stride):\
822
             "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
823

824 825
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){
826 827 828 829 830 831 832 833 834
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")
}

835 836
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){
837 838 839 840 841 842 843 844 845 846 847
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")
848
snow_inner_add_yblock_mmx_mix("16", "8")
849 850 851
snow_inner_add_yblock_mmx_end("32")
}

852
static void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
853 854 855 856 857 858 859 860 861 862 863 864 865
                           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);
}

866
static void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
867 868 869 870 871 872 873 874
                          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);
}
875 876 877

void ff_dwt_init_x86(DWTContext *c)
{
878
    int mm_flags = av_get_cpu_flags();
879

880 881
    if (mm_flags & AV_CPU_FLAG_MMX) {
        if(mm_flags & AV_CPU_FLAG_SSE2 & 0){
882 883 884 885 886 887 888
            c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2;
#if HAVE_7REGS
            c->vertical_compose97i = ff_snow_vertical_compose97i_sse2;
#endif
            c->inner_add_yblock = ff_snow_inner_add_yblock_sse2;
        }
        else{
889
            if(mm_flags & AV_CPU_FLAG_MMX2){
890 891 892 893 894 895 896 897 898
            c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx;
#if HAVE_7REGS
            c->vertical_compose97i = ff_snow_vertical_compose97i_mmx;
#endif
            }
            c->inner_add_yblock = ff_snow_inner_add_yblock_mmx;
        }
    }
}