h264dsp_mmx.c 93.6 KB
Newer Older
1 2 3
/*
 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
 *
4 5 6
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
7 8
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
9
 * version 2.1 of the License, or (at your option) any later version.
10
 *
11
 * FFmpeg is distributed in the hope that it will be useful,
12 13 14 15 16
 * 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
17
 * License along with FFmpeg; if not, write to the Free Software
18
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 20
 */

21
#include "dsputil_mmx.h"
22

23 24
DECLARE_ALIGNED(8, static const uint64_t, ff_pb_3_1  ) = 0x0103010301030103ULL;
DECLARE_ALIGNED(8, static const uint64_t, ff_pb_7_3  ) = 0x0307030703070307ULL;
25

26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
/***********************************/
/* IDCT */

#define SUMSUB_BADC( a, b, c, d ) \
    "paddw "#b", "#a" \n\t"\
    "paddw "#d", "#c" \n\t"\
    "paddw "#b", "#b" \n\t"\
    "paddw "#d", "#d" \n\t"\
    "psubw "#a", "#b" \n\t"\
    "psubw "#c", "#d" \n\t"

#define SUMSUBD2_AB( a, b, t ) \
    "movq  "#b", "#t" \n\t"\
    "psraw  $1 , "#b" \n\t"\
    "paddw "#a", "#b" \n\t"\
    "psraw  $1 , "#a" \n\t"\
    "psubw "#t", "#a" \n\t"

#define IDCT4_1D( s02, s13, d02, d13, t ) \
    SUMSUB_BA  ( s02, d02 )\
    SUMSUBD2_AB( s13, d13, t )\
    SUMSUB_BADC( d13, s02, s13, d02 )

#define STORE_DIFF_4P( p, t, z ) \
    "psraw      $6,     "#p" \n\t"\
    "movd       (%0),   "#t" \n\t"\
    "punpcklbw "#z",    "#t" \n\t"\
    "paddsw    "#t",    "#p" \n\t"\
    "packuswb  "#z",    "#p" \n\t"\
    "movd      "#p",    (%0) \n\t"

57
static void ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride)
58 59
{
    /* Load dct coeffs */
60
    __asm__ volatile(
61 62 63 64 65 66
        "movq   (%0), %%mm0 \n\t"
        "movq  8(%0), %%mm1 \n\t"
        "movq 16(%0), %%mm2 \n\t"
        "movq 24(%0), %%mm3 \n\t"
    :: "r"(block) );

67
    __asm__ volatile(
68 69 70 71 72 73 74 75 76 77 78 79 80
        /* mm1=s02+s13  mm2=s02-s13  mm4=d02+d13  mm0=d02-d13 */
        IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4 )

        "movq      %0,    %%mm6 \n\t"
        /* in: 1,4,0,2  out: 1,2,3,0 */
        TRANSPOSE4( %%mm3, %%mm1, %%mm0, %%mm2, %%mm4 )

        "paddw     %%mm6, %%mm3 \n\t"

        /* mm2=s02+s13  mm3=s02-s13  mm4=d02+d13  mm1=d02-d13 */
        IDCT4_1D( %%mm4, %%mm2, %%mm3, %%mm0, %%mm1 )

        "pxor %%mm7, %%mm7    \n\t"
81
    :: "m"(ff_pw_32));
82

83
    __asm__ volatile(
84 85 86 87 88 89 90 91
    STORE_DIFF_4P( %%mm0, %%mm1, %%mm7)
        "add %1, %0             \n\t"
    STORE_DIFF_4P( %%mm2, %%mm1, %%mm7)
        "add %1, %0             \n\t"
    STORE_DIFF_4P( %%mm3, %%mm1, %%mm7)
        "add %1, %0             \n\t"
    STORE_DIFF_4P( %%mm4, %%mm1, %%mm7)
        : "+r"(dst)
92
        : "r" ((x86_reg)stride)
93 94 95
    );
}

Loren Merritt's avatar
Loren Merritt committed
96 97
static inline void h264_idct8_1d(int16_t *block)
{
98
    __asm__ volatile(
Loren Merritt's avatar
Loren Merritt committed
99
        "movq 112(%0), %%mm7  \n\t"
100
        "movq  80(%0), %%mm0  \n\t"
Loren Merritt's avatar
Loren Merritt committed
101
        "movq  48(%0), %%mm3  \n\t"
102
        "movq  16(%0), %%mm5  \n\t"
Loren Merritt's avatar
Loren Merritt committed
103

104 105
        "movq   %%mm0, %%mm4  \n\t"
        "movq   %%mm5, %%mm1  \n\t"
Loren Merritt's avatar
Loren Merritt committed
106
        "psraw  $1,    %%mm4  \n\t"
107 108 109
        "psraw  $1,    %%mm1  \n\t"
        "paddw  %%mm0, %%mm4  \n\t"
        "paddw  %%mm5, %%mm1  \n\t"
Loren Merritt's avatar
Loren Merritt committed
110
        "paddw  %%mm7, %%mm4  \n\t"
111 112 113 114 115 116 117 118 119 120 121 122
        "paddw  %%mm0, %%mm1  \n\t"
        "psubw  %%mm5, %%mm4  \n\t"
        "paddw  %%mm3, %%mm1  \n\t"

        "psubw  %%mm3, %%mm5  \n\t"
        "psubw  %%mm3, %%mm0  \n\t"
        "paddw  %%mm7, %%mm5  \n\t"
        "psubw  %%mm7, %%mm0  \n\t"
        "psraw  $1,    %%mm3  \n\t"
        "psraw  $1,    %%mm7  \n\t"
        "psubw  %%mm3, %%mm5  \n\t"
        "psubw  %%mm7, %%mm0  \n\t"
Loren Merritt's avatar
Loren Merritt committed
123 124

        "movq   %%mm4, %%mm3  \n\t"
125 126
        "movq   %%mm1, %%mm7  \n\t"
        "psraw  $2,    %%mm1  \n\t"
Loren Merritt's avatar
Loren Merritt committed
127
        "psraw  $2,    %%mm3  \n\t"
128
        "paddw  %%mm5, %%mm3  \n\t"
Loren Merritt's avatar
Loren Merritt committed
129
        "psraw  $2,    %%mm5  \n\t"
130
        "paddw  %%mm0, %%mm1  \n\t"
Loren Merritt's avatar
Loren Merritt committed
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
        "psraw  $2,    %%mm0  \n\t"
        "psubw  %%mm4, %%mm5  \n\t"
        "psubw  %%mm0, %%mm7  \n\t"

        "movq  32(%0), %%mm2  \n\t"
        "movq  96(%0), %%mm6  \n\t"
        "movq   %%mm2, %%mm4  \n\t"
        "movq   %%mm6, %%mm0  \n\t"
        "psraw  $1,    %%mm4  \n\t"
        "psraw  $1,    %%mm6  \n\t"
        "psubw  %%mm0, %%mm4  \n\t"
        "paddw  %%mm2, %%mm6  \n\t"

        "movq    (%0), %%mm2  \n\t"
        "movq  64(%0), %%mm0  \n\t"
        SUMSUB_BA( %%mm0, %%mm2 )
        SUMSUB_BA( %%mm6, %%mm0 )
        SUMSUB_BA( %%mm4, %%mm2 )
        SUMSUB_BA( %%mm7, %%mm6 )
        SUMSUB_BA( %%mm5, %%mm4 )
        SUMSUB_BA( %%mm3, %%mm2 )
        SUMSUB_BA( %%mm1, %%mm0 )
        :: "r"(block)
    );
}

static void ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride)
{
    int i;
160
    DECLARE_ALIGNED(8, int16_t, b2)[64];
Loren Merritt's avatar
Loren Merritt committed
161 162 163 164

    block[0] += 32;

    for(i=0; i<2; i++){
165
        DECLARE_ALIGNED(8, uint64_t, tmp);
Loren Merritt's avatar
Loren Merritt committed
166 167 168

        h264_idct8_1d(block+4*i);

169
        __asm__ volatile(
Loren Merritt's avatar
Loren Merritt committed
170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
            "movq   %%mm7,    %0   \n\t"
            TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 )
            "movq   %%mm0,  8(%1)  \n\t"
            "movq   %%mm6, 24(%1)  \n\t"
            "movq   %%mm7, 40(%1)  \n\t"
            "movq   %%mm4, 56(%1)  \n\t"
            "movq    %0,    %%mm7  \n\t"
            TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 )
            "movq   %%mm7,   (%1)  \n\t"
            "movq   %%mm1, 16(%1)  \n\t"
            "movq   %%mm0, 32(%1)  \n\t"
            "movq   %%mm3, 48(%1)  \n\t"
            : "=m"(tmp)
            : "r"(b2+32*i)
            : "memory"
        );
    }

    for(i=0; i<2; i++){
        h264_idct8_1d(b2+4*i);

191
        __asm__ volatile(
Loren Merritt's avatar
Loren Merritt committed
192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216
            "psraw     $6, %%mm7  \n\t"
            "psraw     $6, %%mm6  \n\t"
            "psraw     $6, %%mm5  \n\t"
            "psraw     $6, %%mm4  \n\t"
            "psraw     $6, %%mm3  \n\t"
            "psraw     $6, %%mm2  \n\t"
            "psraw     $6, %%mm1  \n\t"
            "psraw     $6, %%mm0  \n\t"

            "movq   %%mm7,    (%0)  \n\t"
            "movq   %%mm5,  16(%0)  \n\t"
            "movq   %%mm3,  32(%0)  \n\t"
            "movq   %%mm1,  48(%0)  \n\t"
            "movq   %%mm0,  64(%0)  \n\t"
            "movq   %%mm2,  80(%0)  \n\t"
            "movq   %%mm4,  96(%0)  \n\t"
            "movq   %%mm6, 112(%0)  \n\t"
            :: "r"(b2+4*i)
            : "memory"
        );
    }

    add_pixels_clamped_mmx(b2, dst, stride);
}

Loren Merritt's avatar
Loren Merritt committed
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271
#define STORE_DIFF_8P( p, d, t, z )\
        "movq       "#d", "#t" \n"\
        "psraw       $6,  "#p" \n"\
        "punpcklbw  "#z", "#t" \n"\
        "paddsw     "#t", "#p" \n"\
        "packuswb   "#p", "#p" \n"\
        "movq       "#p", "#d" \n"

#define H264_IDCT8_1D_SSE2(a,b,c,d,e,f,g,h)\
        "movdqa     "#c", "#a" \n"\
        "movdqa     "#g", "#e" \n"\
        "psraw       $1,  "#c" \n"\
        "psraw       $1,  "#g" \n"\
        "psubw      "#e", "#c" \n"\
        "paddw      "#a", "#g" \n"\
        "movdqa     "#b", "#e" \n"\
        "psraw       $1,  "#e" \n"\
        "paddw      "#b", "#e" \n"\
        "paddw      "#d", "#e" \n"\
        "paddw      "#f", "#e" \n"\
        "movdqa     "#f", "#a" \n"\
        "psraw       $1,  "#a" \n"\
        "paddw      "#f", "#a" \n"\
        "paddw      "#h", "#a" \n"\
        "psubw      "#b", "#a" \n"\
        "psubw      "#d", "#b" \n"\
        "psubw      "#d", "#f" \n"\
        "paddw      "#h", "#b" \n"\
        "psubw      "#h", "#f" \n"\
        "psraw       $1,  "#d" \n"\
        "psraw       $1,  "#h" \n"\
        "psubw      "#d", "#b" \n"\
        "psubw      "#h", "#f" \n"\
        "movdqa     "#e", "#d" \n"\
        "movdqa     "#a", "#h" \n"\
        "psraw       $2,  "#d" \n"\
        "psraw       $2,  "#h" \n"\
        "paddw      "#f", "#d" \n"\
        "paddw      "#b", "#h" \n"\
        "psraw       $2,  "#f" \n"\
        "psraw       $2,  "#b" \n"\
        "psubw      "#f", "#e" \n"\
        "psubw      "#a", "#b" \n"\
        "movdqa 0x00(%1), "#a" \n"\
        "movdqa 0x40(%1), "#f" \n"\
        SUMSUB_BA(f, a)\
        SUMSUB_BA(g, f)\
        SUMSUB_BA(c, a)\
        SUMSUB_BA(e, g)\
        SUMSUB_BA(b, c)\
        SUMSUB_BA(h, a)\
        SUMSUB_BA(d, f)

static void ff_h264_idct8_add_sse2(uint8_t *dst, int16_t *block, int stride)
{
272
    __asm__ volatile(
Loren Merritt's avatar
Loren Merritt committed
273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299
        "movdqa   0x10(%1), %%xmm1 \n"
        "movdqa   0x20(%1), %%xmm2 \n"
        "movdqa   0x30(%1), %%xmm3 \n"
        "movdqa   0x50(%1), %%xmm5 \n"
        "movdqa   0x60(%1), %%xmm6 \n"
        "movdqa   0x70(%1), %%xmm7 \n"
        H264_IDCT8_1D_SSE2(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7)
        TRANSPOSE8(%%xmm4, %%xmm1, %%xmm7, %%xmm3, %%xmm5, %%xmm0, %%xmm2, %%xmm6, (%1))
        "paddw          %4, %%xmm4 \n"
        "movdqa     %%xmm4, 0x00(%1) \n"
        "movdqa     %%xmm2, 0x40(%1) \n"
        H264_IDCT8_1D_SSE2(%%xmm4, %%xmm0, %%xmm6, %%xmm3, %%xmm2, %%xmm5, %%xmm7, %%xmm1)
        "movdqa     %%xmm6, 0x60(%1) \n"
        "movdqa     %%xmm7, 0x70(%1) \n"
        "pxor       %%xmm7, %%xmm7 \n"
        STORE_DIFF_8P(%%xmm2, (%0),      %%xmm6, %%xmm7)
        STORE_DIFF_8P(%%xmm0, (%0,%2),   %%xmm6, %%xmm7)
        STORE_DIFF_8P(%%xmm1, (%0,%2,2), %%xmm6, %%xmm7)
        STORE_DIFF_8P(%%xmm3, (%0,%3),   %%xmm6, %%xmm7)
        "lea     (%0,%2,4), %0 \n"
        STORE_DIFF_8P(%%xmm5, (%0),      %%xmm6, %%xmm7)
        STORE_DIFF_8P(%%xmm4, (%0,%2),   %%xmm6, %%xmm7)
        "movdqa   0x60(%1), %%xmm0 \n"
        "movdqa   0x70(%1), %%xmm1 \n"
        STORE_DIFF_8P(%%xmm0, (%0,%2,2), %%xmm6, %%xmm7)
        STORE_DIFF_8P(%%xmm1, (%0,%3),   %%xmm6, %%xmm7)
        :"+r"(dst)
300
        :"r"(block), "r"((x86_reg)stride), "r"((x86_reg)3L*stride), "m"(ff_pw_32)
Loren Merritt's avatar
Loren Merritt committed
301 302 303
    );
}

304
static void ff_h264_idct_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride)
305 306
{
    int dc = (block[0] + 32) >> 6;
307
    __asm__ volatile(
308 309 310 311 312 313 314 315
        "movd          %0, %%mm0 \n\t"
        "pshufw $0, %%mm0, %%mm0 \n\t"
        "pxor       %%mm1, %%mm1 \n\t"
        "psubw      %%mm0, %%mm1 \n\t"
        "packuswb   %%mm0, %%mm0 \n\t"
        "packuswb   %%mm1, %%mm1 \n\t"
        ::"r"(dc)
    );
316
    __asm__ volatile(
317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339
        "movd          %0, %%mm2 \n\t"
        "movd          %1, %%mm3 \n\t"
        "movd          %2, %%mm4 \n\t"
        "movd          %3, %%mm5 \n\t"
        "paddusb    %%mm0, %%mm2 \n\t"
        "paddusb    %%mm0, %%mm3 \n\t"
        "paddusb    %%mm0, %%mm4 \n\t"
        "paddusb    %%mm0, %%mm5 \n\t"
        "psubusb    %%mm1, %%mm2 \n\t"
        "psubusb    %%mm1, %%mm3 \n\t"
        "psubusb    %%mm1, %%mm4 \n\t"
        "psubusb    %%mm1, %%mm5 \n\t"
        "movd       %%mm2, %0    \n\t"
        "movd       %%mm3, %1    \n\t"
        "movd       %%mm4, %2    \n\t"
        "movd       %%mm5, %3    \n\t"
        :"+m"(*(uint32_t*)(dst+0*stride)),
         "+m"(*(uint32_t*)(dst+1*stride)),
         "+m"(*(uint32_t*)(dst+2*stride)),
         "+m"(*(uint32_t*)(dst+3*stride))
    );
}

340
static void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride)
341 342 343
{
    int dc = (block[0] + 32) >> 6;
    int y;
344
    __asm__ volatile(
345 346 347 348 349 350 351 352 353
        "movd          %0, %%mm0 \n\t"
        "pshufw $0, %%mm0, %%mm0 \n\t"
        "pxor       %%mm1, %%mm1 \n\t"
        "psubw      %%mm0, %%mm1 \n\t"
        "packuswb   %%mm0, %%mm0 \n\t"
        "packuswb   %%mm1, %%mm1 \n\t"
        ::"r"(dc)
    );
    for(y=2; y--; dst += 4*stride){
354
    __asm__ volatile(
355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378
        "movq          %0, %%mm2 \n\t"
        "movq          %1, %%mm3 \n\t"
        "movq          %2, %%mm4 \n\t"
        "movq          %3, %%mm5 \n\t"
        "paddusb    %%mm0, %%mm2 \n\t"
        "paddusb    %%mm0, %%mm3 \n\t"
        "paddusb    %%mm0, %%mm4 \n\t"
        "paddusb    %%mm0, %%mm5 \n\t"
        "psubusb    %%mm1, %%mm2 \n\t"
        "psubusb    %%mm1, %%mm3 \n\t"
        "psubusb    %%mm1, %%mm4 \n\t"
        "psubusb    %%mm1, %%mm5 \n\t"
        "movq       %%mm2, %0    \n\t"
        "movq       %%mm3, %1    \n\t"
        "movq       %%mm4, %2    \n\t"
        "movq       %%mm5, %3    \n\t"
        :"+m"(*(uint64_t*)(dst+0*stride)),
         "+m"(*(uint64_t*)(dst+1*stride)),
         "+m"(*(uint64_t*)(dst+2*stride)),
         "+m"(*(uint64_t*)(dst+3*stride))
    );
    }
}

379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473
//FIXME this table is a duplicate from h264data.h, and will be removed once the tables from, h264 have been split
static const uint8_t scan8[16 + 2*4]={
 4+1*8, 5+1*8, 4+2*8, 5+2*8,
 6+1*8, 7+1*8, 6+2*8, 7+2*8,
 4+3*8, 5+3*8, 4+4*8, 5+4*8,
 6+3*8, 7+3*8, 6+4*8, 7+4*8,
 1+1*8, 2+1*8,
 1+2*8, 2+2*8,
 1+4*8, 2+4*8,
 1+5*8, 2+5*8,
};

static void ff_h264_idct_add16_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i++){
        if(nnzc[ scan8[i] ])
            ff_h264_idct_add_mmx(dst + block_offset[i], block + i*16, stride);
    }
}

static void ff_h264_idct8_add4_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i+=4){
        if(nnzc[ scan8[i] ])
            ff_h264_idct8_add_mmx(dst + block_offset[i], block + i*16, stride);
    }
}


static void ff_h264_idct_add16_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i++){
        int nnz = nnzc[ scan8[i] ];
        if(nnz){
            if(nnz==1 && block[i*16]) ff_h264_idct_dc_add_mmx2(dst + block_offset[i], block + i*16, stride);
            else                      ff_h264_idct_add_mmx    (dst + block_offset[i], block + i*16, stride);
        }
    }
}

static void ff_h264_idct_add16intra_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i++){
        if(nnzc[ scan8[i] ] || block[i*16])
            ff_h264_idct_add_mmx(dst + block_offset[i], block + i*16, stride);
    }
}

static void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i++){
        if(nnzc[ scan8[i] ]) ff_h264_idct_add_mmx    (dst + block_offset[i], block + i*16, stride);
        else if(block[i*16]) ff_h264_idct_dc_add_mmx2(dst + block_offset[i], block + i*16, stride);
    }
}

static void ff_h264_idct8_add4_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i+=4){
        int nnz = nnzc[ scan8[i] ];
        if(nnz){
            if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_mmx2(dst + block_offset[i], block + i*16, stride);
            else                      ff_h264_idct8_add_mmx    (dst + block_offset[i], block + i*16, stride);
        }
    }
}

static void ff_h264_idct8_add4_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i+=4){
        int nnz = nnzc[ scan8[i] ];
        if(nnz){
            if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_mmx2(dst + block_offset[i], block + i*16, stride);
            else                      ff_h264_idct8_add_sse2   (dst + block_offset[i], block + i*16, stride);
        }
    }
}

static void ff_h264_idct_add8_mmx(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=16; i<16+8; i++){
        if(nnzc[ scan8[i] ] || block[i*16])
            ff_h264_idct_add_mmx    (dest[(i&4)>>2] + block_offset[i], block + i*16, stride);
    }
}

static void ff_h264_idct_add8_mmx2(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=16; i<16+8; i++){
        if(nnzc[ scan8[i] ])
            ff_h264_idct_add_mmx    (dest[(i&4)>>2] + block_offset[i], block + i*16, stride);
        else if(block[i*16])
            ff_h264_idct_dc_add_mmx2(dest[(i&4)>>2] + block_offset[i], block + i*16, stride);
    }
}
474

475
#if CONFIG_GPL && HAVE_YASM
476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537
static void ff_h264_idct_dc_add8_mmx2(uint8_t *dst, int16_t *block, int stride)
{
    __asm__ volatile(
        "movd             %0, %%mm0 \n\t"   //  0 0 X D
        "punpcklwd        %1, %%mm0 \n\t"   //  x X d D
        "paddsw           %2, %%mm0 \n\t"
        "psraw            $6, %%mm0 \n\t"
        "punpcklwd     %%mm0, %%mm0 \n\t"   //  d d D D
        "pxor          %%mm1, %%mm1 \n\t"   //  0 0 0 0
        "psubw         %%mm0, %%mm1 \n\t"   // -d-d-D-D
        "packuswb      %%mm1, %%mm0 \n\t"   // -d-d-D-D d d D D
        "pshufw $0xFA, %%mm0, %%mm1 \n\t"   // -d-d-d-d-D-D-D-D
        "punpcklwd     %%mm0, %%mm0 \n\t"   //  d d d d D D D D
        ::"m"(block[ 0]),
          "m"(block[16]),
          "m"(ff_pw_32)
    );
    __asm__ volatile(
        "movq          %0, %%mm2 \n\t"
        "movq          %1, %%mm3 \n\t"
        "movq          %2, %%mm4 \n\t"
        "movq          %3, %%mm5 \n\t"
        "paddusb    %%mm0, %%mm2 \n\t"
        "paddusb    %%mm0, %%mm3 \n\t"
        "paddusb    %%mm0, %%mm4 \n\t"
        "paddusb    %%mm0, %%mm5 \n\t"
        "psubusb    %%mm1, %%mm2 \n\t"
        "psubusb    %%mm1, %%mm3 \n\t"
        "psubusb    %%mm1, %%mm4 \n\t"
        "psubusb    %%mm1, %%mm5 \n\t"
        "movq       %%mm2, %0    \n\t"
        "movq       %%mm3, %1    \n\t"
        "movq       %%mm4, %2    \n\t"
        "movq       %%mm5, %3    \n\t"
        :"+m"(*(uint64_t*)(dst+0*stride)),
         "+m"(*(uint64_t*)(dst+1*stride)),
         "+m"(*(uint64_t*)(dst+2*stride)),
         "+m"(*(uint64_t*)(dst+3*stride))
    );
}

extern void ff_x264_add8x4_idct_sse2(uint8_t *dst, int16_t *block, int stride);

static void ff_h264_idct_add16_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i+=2)
        if(nnzc[ scan8[i+0] ]|nnzc[ scan8[i+1] ])
            ff_x264_add8x4_idct_sse2 (dst + block_offset[i], block + i*16, stride);
}

static void ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
    for(i=0; i<16; i+=2){
        if(nnzc[ scan8[i+0] ]|nnzc[ scan8[i+1] ])
            ff_x264_add8x4_idct_sse2 (dst + block_offset[i], block + i*16, stride);
        else if(block[i*16]|block[i*16+16])
            ff_h264_idct_dc_add8_mmx2(dst + block_offset[i], block + i*16, stride);
    }
}

static void ff_h264_idct_add8_sse2(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){
    int i;
538
    for(i=16; i<16+8; i+=2){
539 540 541 542 543 544 545 546
        if(nnzc[ scan8[i+0] ]|nnzc[ scan8[i+1] ])
            ff_x264_add8x4_idct_sse2 (dest[(i&4)>>2] + block_offset[i], block + i*16, stride);
        else if(block[i*16]|block[i*16+16])
            ff_h264_idct_dc_add8_mmx2(dest[(i&4)>>2] + block_offset[i], block + i*16, stride);
    }
}
#endif

547 548 549 550 551 552 553 554 555 556 557 558 559
/***********************************/
/* deblocking */

// out: o = |x-y|>a
// clobbers: t
#define DIFF_GT_MMX(x,y,a,o,t)\
    "movq     "#y", "#t"  \n\t"\
    "movq     "#x", "#o"  \n\t"\
    "psubusb  "#x", "#t"  \n\t"\
    "psubusb  "#y", "#o"  \n\t"\
    "por      "#t", "#o"  \n\t"\
    "psubusb  "#a", "#o"  \n\t"

560 561 562 563 564 565 566 567 568 569 570
// out: o = |x-y|>a
// clobbers: t
#define DIFF_GT2_MMX(x,y,a,o,t)\
    "movq     "#y", "#t"  \n\t"\
    "movq     "#x", "#o"  \n\t"\
    "psubusb  "#x", "#t"  \n\t"\
    "psubusb  "#y", "#o"  \n\t"\
    "psubusb  "#a", "#t"  \n\t"\
    "psubusb  "#a", "#o"  \n\t"\
    "pcmpeqb  "#t", "#o"  \n\t"\

571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590
// in: mm0=p1 mm1=p0 mm2=q0 mm3=q1
// out: mm5=beta-1, mm7=mask
// clobbers: mm4,mm6
#define H264_DEBLOCK_MASK(alpha1, beta1) \
    "pshufw $0, "#alpha1", %%mm4 \n\t"\
    "pshufw $0, "#beta1 ", %%mm5 \n\t"\
    "packuswb  %%mm4, %%mm4      \n\t"\
    "packuswb  %%mm5, %%mm5      \n\t"\
    DIFF_GT_MMX(%%mm1, %%mm2, %%mm4, %%mm7, %%mm6) /* |p0-q0| > alpha-1 */\
    DIFF_GT_MMX(%%mm0, %%mm1, %%mm5, %%mm4, %%mm6) /* |p1-p0| > beta-1 */\
    "por       %%mm4, %%mm7      \n\t"\
    DIFF_GT_MMX(%%mm3, %%mm2, %%mm5, %%mm4, %%mm6) /* |q1-q0| > beta-1 */\
    "por       %%mm4, %%mm7      \n\t"\
    "pxor      %%mm6, %%mm6      \n\t"\
    "pcmpeqb   %%mm6, %%mm7      \n\t"

// in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask)
// out: mm1=p0' mm2=q0'
// clobbers: mm0,3-6
#define H264_DEBLOCK_P0_Q0(pb_01, pb_3f)\
591 592 593 594 595 596 597 598 599 600
        "movq    %%mm1              , %%mm5 \n\t"\
        "pxor    %%mm2              , %%mm5 \n\t" /* p0^q0*/\
        "pand    "#pb_01"           , %%mm5 \n\t" /* (p0^q0)&1*/\
        "pcmpeqb %%mm4              , %%mm4 \n\t"\
        "pxor    %%mm4              , %%mm3 \n\t"\
        "pavgb   %%mm0              , %%mm3 \n\t" /* (p1 - q1 + 256)>>1*/\
        "pavgb   "MANGLE(ff_pb_3)"  , %%mm3 \n\t" /*(((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2*/\
        "pxor    %%mm1              , %%mm4 \n\t"\
        "pavgb   %%mm2              , %%mm4 \n\t" /* (q0 - p0 + 256)>>1*/\
        "pavgb   %%mm5              , %%mm3 \n\t"\
601
        "paddusb %%mm4              , %%mm3 \n\t" /* d+128+33*/\
Michael Niedermayer's avatar
Michael Niedermayer committed
602 603
        "movq    "MANGLE(ff_pb_A1)" , %%mm6 \n\t"\
        "psubusb %%mm3              , %%mm6 \n\t"\
604 605
        "psubusb "MANGLE(ff_pb_A1)" , %%mm3 \n\t"\
        "pminub  %%mm7              , %%mm6 \n\t"\
Michael Niedermayer's avatar
Michael Niedermayer committed
606
        "pminub  %%mm7              , %%mm3 \n\t"\
607
        "psubusb %%mm6              , %%mm1 \n\t"\
Michael Niedermayer's avatar
Michael Niedermayer committed
608 609 610
        "psubusb %%mm3              , %%mm2 \n\t"\
        "paddusb %%mm3              , %%mm1 \n\t"\
        "paddusb %%mm6              , %%mm2 \n\t"
611

612
// in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) %8=ff_bone
613
// out: (q1addr) = av_clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 )
614 615 616 617 618 619
// clobbers: q2, tmp, tc0
#define H264_DEBLOCK_Q1(p1, q2, q2addr, q1addr, tc0, tmp)\
        "movq     %%mm1,  "#tmp"   \n\t"\
        "pavgb    %%mm2,  "#tmp"   \n\t"\
        "pavgb    "#tmp", "#q2"    \n\t" /* avg(p2,avg(p0,q0)) */\
        "pxor   "q2addr", "#tmp"   \n\t"\
620
        "pand     %9,     "#tmp"   \n\t" /* (p2^avg(p0,q0))&1 */\
621 622 623 624 625 626 627 628 629 630
        "psubusb  "#tmp", "#q2"    \n\t" /* (p2+((p0+q0+1)>>1))>>1 */\
        "movq     "#p1",  "#tmp"   \n\t"\
        "psubusb  "#tc0", "#tmp"   \n\t"\
        "paddusb  "#p1",  "#tc0"   \n\t"\
        "pmaxub   "#tmp", "#q2"    \n\t"\
        "pminub   "#tc0", "#q2"    \n\t"\
        "movq     "#q2",  "q1addr" \n\t"

static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0)
{
631
    DECLARE_ALIGNED(8, uint64_t, tmp0)[2];
632

633
    __asm__ volatile(
634 635 636 637 638
        "movq    (%2,%4), %%mm0    \n\t" //p1
        "movq    (%2,%4,2), %%mm1  \n\t" //p0
        "movq    (%3),    %%mm2    \n\t" //q0
        "movq    (%3,%4), %%mm3    \n\t" //q1
        H264_DEBLOCK_MASK(%7, %8)
639

640
        "movd      %6,    %%mm4    \n\t"
641 642 643 644 645
        "punpcklbw %%mm4, %%mm4    \n\t"
        "punpcklwd %%mm4, %%mm4    \n\t"
        "pcmpeqb   %%mm3, %%mm3    \n\t"
        "movq      %%mm4, %%mm6    \n\t"
        "pcmpgtb   %%mm3, %%mm4    \n\t"
646
        "movq      %%mm6, %1       \n\t"
647 648
        "pand      %%mm4, %%mm7    \n\t"
        "movq      %%mm7, %0       \n\t"
649 650

        /* filter p1 */
651
        "movq     (%2),   %%mm3    \n\t" //p2
652
        DIFF_GT2_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4) // |p2-p0|>beta-1
653
        "pand     %%mm7,  %%mm6    \n\t" // mask & |p2-p0|<beta
654
        "pand     %1,     %%mm7    \n\t" // mask & tc0
655
        "movq     %%mm7,  %%mm4    \n\t"
656
        "psubb    %%mm6,  %%mm7    \n\t"
657
        "pand     %%mm4,  %%mm6    \n\t" // mask & |p2-p0|<beta & tc0
658
        H264_DEBLOCK_Q1(%%mm0, %%mm3, "(%2)", "(%2,%4)", %%mm6, %%mm4)
659 660

        /* filter q1 */
661
        "movq    (%3,%4,2), %%mm4  \n\t" //q2
662
        DIFF_GT2_MMX(%%mm2, %%mm4, %%mm5, %%mm6, %%mm3) // |q2-q0|>beta-1
663
        "pand     %0,     %%mm6    \n\t"
664
        "movq     %1,     %%mm5    \n\t" // can be merged with the and below but is slower then
665
        "pand     %%mm6,  %%mm5    \n\t"
666
        "psubb    %%mm6,  %%mm7    \n\t"
667 668
        "movq    (%3,%4), %%mm3    \n\t"
        H264_DEBLOCK_Q1(%%mm3, %%mm4, "(%3,%4,2)", "(%3,%4)", %%mm5, %%mm6)
669 670

        /* filter p0, q0 */
671 672 673
        H264_DEBLOCK_P0_Q0(%9, unused)
        "movq      %%mm1, (%2,%4,2) \n\t"
        "movq      %%mm2, (%3)      \n\t"
674

675
        : "=m"(tmp0[0]), "=m"(tmp0[1])
676
        : "r"(pix-3*stride), "r"(pix), "r"((x86_reg)stride),
677
          "m"(*tmp0/*unused*/), "m"(*(uint32_t*)tc0), "m"(alpha1), "m"(beta1),
678
          "m"(ff_bone)
679 680 681 682 683 684 685 686 687 688 689 690 691 692
    );
}

static void h264_v_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    if((tc0[0] & tc0[1]) >= 0)
        h264_loop_filter_luma_mmx2(pix, stride, alpha-1, beta-1, tc0);
    if((tc0[2] & tc0[3]) >= 0)
        h264_loop_filter_luma_mmx2(pix+8, stride, alpha-1, beta-1, tc0+2);
}
static void h264_h_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    //FIXME: could cut some load/stores by merging transpose with filter
    // also, it only needs to transpose 6x8
693
    DECLARE_ALIGNED(8, uint8_t, trans)[8*8];
694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
    int i;
    for(i=0; i<2; i++, pix+=8*stride, tc0+=2) {
        if((tc0[0] & tc0[1]) < 0)
            continue;
        transpose4x4(trans,       pix-4,          8, stride);
        transpose4x4(trans  +4*8, pix,            8, stride);
        transpose4x4(trans+4,     pix-4+4*stride, 8, stride);
        transpose4x4(trans+4+4*8, pix  +4*stride, 8, stride);
        h264_loop_filter_luma_mmx2(trans+4*8, 8, alpha-1, beta-1, tc0);
        transpose4x4(pix-2,          trans  +2*8, stride, 8);
        transpose4x4(pix-2+4*stride, trans+4+2*8, stride, 8);
    }
}

static inline void h264_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0)
{
710
    __asm__ volatile(
711 712 713 714 715 716 717 718 719 720 721 722
        "movq    (%0),    %%mm0     \n\t" //p1
        "movq    (%0,%2), %%mm1     \n\t" //p0
        "movq    (%1),    %%mm2     \n\t" //q0
        "movq    (%1,%2), %%mm3     \n\t" //q1
        H264_DEBLOCK_MASK(%4, %5)
        "movd      %3,    %%mm6     \n\t"
        "punpcklbw %%mm6, %%mm6     \n\t"
        "pand      %%mm6, %%mm7     \n\t" // mm7 = tc&mask
        H264_DEBLOCK_P0_Q0(%6, %7)
        "movq      %%mm1, (%0,%2)   \n\t"
        "movq      %%mm2, (%1)      \n\t"

723
        :: "r"(pix-2*stride), "r"(pix), "r"((x86_reg)stride),
724
           "r"(*(uint32_t*)tc0),
725
           "m"(alpha1), "m"(beta1), "m"(ff_bone), "m"(ff_pb_3F)
726 727 728 729 730 731 732 733 734 735 736
    );
}

static void h264_v_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    h264_loop_filter_chroma_mmx2(pix, stride, alpha-1, beta-1, tc0);
}

static void h264_h_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    //FIXME: could cut some load/stores by merging transpose with filter
737
    DECLARE_ALIGNED(8, uint8_t, trans)[8*4];
738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755
    transpose4x4(trans, pix-2, 8, stride);
    transpose4x4(trans+4, pix-2+4*stride, 8, stride);
    h264_loop_filter_chroma_mmx2(trans+2*8, 8, alpha-1, beta-1, tc0);
    transpose4x4(pix-2, trans, stride, 8);
    transpose4x4(pix-2+4*stride, trans+4, stride, 8);
}

// p0 = (p0 + q1 + 2*p1 + 2) >> 2
#define H264_FILTER_CHROMA4(p0, p1, q1, one) \
    "movq    "#p0", %%mm4  \n\t"\
    "pxor    "#q1", %%mm4  \n\t"\
    "pand   "#one", %%mm4  \n\t" /* mm4 = (p0^q1)&1 */\
    "pavgb   "#q1", "#p0"  \n\t"\
    "psubusb %%mm4, "#p0"  \n\t"\
    "pavgb   "#p1", "#p0"  \n\t" /* dst = avg(p1, avg(p0,q1) - ((p0^q1)&1)) */\

static inline void h264_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha1, int beta1)
{
756
    __asm__ volatile(
757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
        "movq    (%0),    %%mm0     \n\t"
        "movq    (%0,%2), %%mm1     \n\t"
        "movq    (%1),    %%mm2     \n\t"
        "movq    (%1,%2), %%mm3     \n\t"
        H264_DEBLOCK_MASK(%3, %4)
        "movq    %%mm1,   %%mm5     \n\t"
        "movq    %%mm2,   %%mm6     \n\t"
        H264_FILTER_CHROMA4(%%mm1, %%mm0, %%mm3, %5) //p0'
        H264_FILTER_CHROMA4(%%mm2, %%mm3, %%mm0, %5) //q0'
        "psubb   %%mm5,   %%mm1     \n\t"
        "psubb   %%mm6,   %%mm2     \n\t"
        "pand    %%mm7,   %%mm1     \n\t"
        "pand    %%mm7,   %%mm2     \n\t"
        "paddb   %%mm5,   %%mm1     \n\t"
        "paddb   %%mm6,   %%mm2     \n\t"
        "movq    %%mm1,   (%0,%2)   \n\t"
        "movq    %%mm2,   (%1)      \n\t"
774
        :: "r"(pix-2*stride), "r"(pix), "r"((x86_reg)stride),
775
           "m"(alpha1), "m"(beta1), "m"(ff_bone)
776 777 778 779 780 781 782 783 784 785 786
    );
}

static void h264_v_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta)
{
    h264_loop_filter_chroma_intra_mmx2(pix, stride, alpha-1, beta-1);
}

static void h264_h_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta)
{
    //FIXME: could cut some load/stores by merging transpose with filter
787
    DECLARE_ALIGNED(8, uint8_t, trans)[8*4];
788 789 790 791 792 793 794
    transpose4x4(trans, pix-2, 8, stride);
    transpose4x4(trans+4, pix-2+4*stride, 8, stride);
    h264_loop_filter_chroma_intra_mmx2(trans+2*8, 8, alpha-1, beta-1);
    transpose4x4(pix-2, trans, stride, 8);
    transpose4x4(pix-2+4*stride, trans+4, stride, 8);
}

795
static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
796
                                            int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
797
    int dir;
798
    __asm__ volatile(
799 800 801
        "movq %0, %%mm7 \n"
        "movq %1, %%mm6 \n"
        ::"m"(ff_pb_1), "m"(ff_pb_3)
802
    );
803
    if(field)
804
        __asm__ volatile(
805 806
            "movq %0, %%mm6 \n"
            ::"m"(ff_pb_3_1)
807
        );
808 809 810 811
    __asm__ volatile(
        "movq  %%mm6, %%mm5 \n"
        "paddb %%mm5, %%mm5 \n"
    :);
812

813 814 815
    // could do a special case for dir==0 && edges==1, but it only reduces the
    // average filter time by 1.2%
    for( dir=1; dir>=0; dir-- ) {
816
        const x86_reg d_idx = dir ? -8 : -1;
817
        const int mask_mv = dir ? mask_mv1 : mask_mv0;
818
        DECLARE_ALIGNED(8, const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL;
819
        int b_idx, edge;
820
        for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) {
821
            __asm__ volatile(
822 823 824 825
                "pand %0, %%mm0 \n\t"
                ::"m"(mask_dir)
            );
            if(!(mask_mv & edge)) {
826
                if(bidir) {
827
                    __asm__ volatile(
828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862
                        "movd         (%1,%0), %%mm2 \n"
                        "punpckldq  40(%1,%0), %%mm2 \n" // { ref0[bn], ref1[bn] }
                        "pshufw $0x44,   (%1), %%mm0 \n" // { ref0[b], ref0[b] }
                        "pshufw $0x44, 40(%1), %%mm1 \n" // { ref1[b], ref1[b] }
                        "pshufw $0x4E, %%mm2, %%mm3 \n"
                        "psubb         %%mm2, %%mm0 \n" // { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] }
                        "psubb         %%mm3, %%mm1 \n" // { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] }
                        "1: \n"
                        "por           %%mm1, %%mm0 \n"
                        "movq      (%2,%0,4), %%mm1 \n"
                        "movq     8(%2,%0,4), %%mm2 \n"
                        "movq          %%mm1, %%mm3 \n"
                        "movq          %%mm2, %%mm4 \n"
                        "psubw          (%2), %%mm1 \n"
                        "psubw         8(%2), %%mm2 \n"
                        "psubw       160(%2), %%mm3 \n"
                        "psubw       168(%2), %%mm4 \n"
                        "packsswb      %%mm2, %%mm1 \n"
                        "packsswb      %%mm4, %%mm3 \n"
                        "paddb         %%mm6, %%mm1 \n"
                        "paddb         %%mm6, %%mm3 \n"
                        "psubusb       %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
                        "psubusb       %%mm5, %%mm3 \n"
                        "packsswb      %%mm3, %%mm1 \n"
                        "add $40, %0 \n"
                        "cmp $40, %0 \n"
                        "jl 1b \n"
                        "sub $80, %0 \n"
                        "pshufw $0x4E, %%mm1, %%mm1 \n"
                        "por           %%mm1, %%mm0 \n"
                        "pshufw $0x4E, %%mm0, %%mm1 \n"
                        "pminub        %%mm1, %%mm0 \n"
                        ::"r"(d_idx),
                          "r"(ref[0]+b_idx),
                          "r"(mv[0]+b_idx)
863
                    );
864
                } else {
865
                    __asm__ volatile(
866 867 868 869 870 871 872 873 874 875 876 877 878 879
                        "movd        (%1), %%mm0 \n"
                        "psubb    (%1,%0), %%mm0 \n" // ref[b] != ref[bn]
                        "movq        (%2), %%mm1 \n"
                        "movq       8(%2), %%mm2 \n"
                        "psubw  (%2,%0,4), %%mm1 \n"
                        "psubw 8(%2,%0,4), %%mm2 \n"
                        "packsswb   %%mm2, %%mm1 \n"
                        "paddb      %%mm6, %%mm1 \n"
                        "psubusb    %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
                        "packsswb   %%mm1, %%mm1 \n"
                        "por        %%mm1, %%mm0 \n"
                        ::"r"(d_idx),
                          "r"(ref[0]+b_idx),
                          "r"(mv[0]+b_idx)
880 881
                    );
                }
882
            }
883
            __asm__ volatile(
884 885
                "movd %0, %%mm1 \n"
                "por  %1, %%mm1 \n" // nnz[b] || nnz[bn]
886 887 888
                ::"m"(nnz[b_idx]),
                  "m"(nnz[b_idx+d_idx])
            );
889
            __asm__ volatile(
890 891 892 893 894 895 896
                "pminub    %%mm7, %%mm1 \n"
                "pminub    %%mm7, %%mm0 \n"
                "psllw        $1, %%mm1 \n"
                "pxor      %%mm2, %%mm2 \n"
                "pmaxub    %%mm0, %%mm1 \n"
                "punpcklbw %%mm2, %%mm1 \n"
                "movq      %%mm1, %0    \n"
897 898 899 900 901 902 903
                :"=m"(*bS[dir][edge])
                ::"memory"
            );
        }
        edges = 4;
        step = 1;
    }
904
    __asm__ volatile(
905 906 907 908 909 910 911 912 913 914 915 916 917
        "movq   (%0), %%mm0 \n\t"
        "movq  8(%0), %%mm1 \n\t"
        "movq 16(%0), %%mm2 \n\t"
        "movq 24(%0), %%mm3 \n\t"
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
        "movq %%mm0,   (%0) \n\t"
        "movq %%mm3,  8(%0) \n\t"
        "movq %%mm4, 16(%0) \n\t"
        "movq %%mm2, 24(%0) \n\t"
        ::"r"(bS[0])
        :"memory"
    );
}
918 919 920 921

/***********************************/
/* motion compensation */

922 923 924 925 926 927 928 929 930
#define QPEL_H264V_MM(A,B,C,D,E,F,OP,T,Z,d,q)\
        "mov"#q" "#C", "#T"         \n\t"\
        "mov"#d" (%0), "#F"         \n\t"\
        "paddw "#D", "#T"           \n\t"\
        "psllw $2, "#T"             \n\t"\
        "psubw "#B", "#T"           \n\t"\
        "psubw "#E", "#T"           \n\t"\
        "punpcklbw "#Z", "#F"       \n\t"\
        "pmullw %4, "#T"            \n\t"\
931
        "paddw %5, "#A"             \n\t"\
932
        "add %2, %0                 \n\t"\
933
        "paddw "#F", "#A"           \n\t"\
934 935 936 937
        "paddw "#A", "#T"           \n\t"\
        "psraw $5, "#T"             \n\t"\
        "packuswb "#T", "#T"        \n\t"\
        OP(T, (%1), A, d)\
938
        "add %3, %1                 \n\t"
939

940 941 942 943 944
#define QPEL_H264HV_MM(A,B,C,D,E,F,OF,T,Z,d,q)\
        "mov"#q" "#C", "#T"         \n\t"\
        "mov"#d" (%0), "#F"         \n\t"\
        "paddw "#D", "#T"           \n\t"\
        "psllw $2, "#T"             \n\t"\
945
        "paddw %4, "#A"             \n\t"\
946 947 948 949
        "psubw "#B", "#T"           \n\t"\
        "psubw "#E", "#T"           \n\t"\
        "punpcklbw "#Z", "#F"       \n\t"\
        "pmullw %3, "#T"            \n\t"\
950
        "paddw "#F", "#A"           \n\t"\
951 952 953 954 955 956 957 958 959
        "add %2, %0                 \n\t"\
        "paddw "#A", "#T"           \n\t"\
        "mov"#q" "#T", "#OF"(%1)    \n\t"

#define QPEL_H264V(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%mm6,%%mm7,d,q)
#define QPEL_H264HV(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%mm6,%%mm7,d,q)
#define QPEL_H264V_XMM(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%xmm6,%%xmm7,q,dqa)
#define QPEL_H264HV_XMM(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%xmm6,%%xmm7,q,dqa)

960

961
#define QPEL_H264(OPNAME, OP, MMX)\
962
static av_noinline void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
963 964
    int h=4;\
\
965
    __asm__ volatile(\
966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991
        "pxor %%mm7, %%mm7          \n\t"\
        "movq %5, %%mm4             \n\t"\
        "movq %6, %%mm5             \n\t"\
        "1:                         \n\t"\
        "movd  -1(%0), %%mm1        \n\t"\
        "movd    (%0), %%mm2        \n\t"\
        "movd   1(%0), %%mm3        \n\t"\
        "movd   2(%0), %%mm0        \n\t"\
        "punpcklbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "paddw %%mm0, %%mm1         \n\t"\
        "paddw %%mm3, %%mm2         \n\t"\
        "movd  -2(%0), %%mm0        \n\t"\
        "movd   3(%0), %%mm3        \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "paddw %%mm3, %%mm0         \n\t"\
        "psllw $2, %%mm2            \n\t"\
        "psubw %%mm1, %%mm2         \n\t"\
        "pmullw %%mm4, %%mm2        \n\t"\
        "paddw %%mm5, %%mm0         \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "psraw $5, %%mm0            \n\t"\
        "packuswb %%mm0, %%mm0      \n\t"\
992
        OP(%%mm0, (%1),%%mm6, d)\
993 994 995 996
        "add %3, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "decl %2                    \n\t"\
        " jnz 1b                    \n\t"\
997
        : "+a"(src), "+c"(dst), "+g"(h)\
998
        : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
999 1000 1001
        : "memory"\
    );\
}\
1002
static av_noinline void OPNAME ## h264_qpel4_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
Loren Merritt's avatar
Loren Merritt committed
1003
    int h=4;\
1004
    __asm__ volatile(\
Loren Merritt's avatar
Loren Merritt committed
1005
        "pxor %%mm7, %%mm7          \n\t"\
Loren Merritt's avatar
Loren Merritt committed
1006 1007 1008 1009 1010
        "movq %0, %%mm4             \n\t"\
        "movq %1, %%mm5             \n\t"\
        :: "m"(ff_pw_5), "m"(ff_pw_16)\
    );\
    do{\
1011
    __asm__ volatile(\
Loren Merritt's avatar
Loren Merritt committed
1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036
        "movd  -1(%0), %%mm1        \n\t"\
        "movd    (%0), %%mm2        \n\t"\
        "movd   1(%0), %%mm3        \n\t"\
        "movd   2(%0), %%mm0        \n\t"\
        "punpcklbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "paddw %%mm0, %%mm1         \n\t"\
        "paddw %%mm3, %%mm2         \n\t"\
        "movd  -2(%0), %%mm0        \n\t"\
        "movd   3(%0), %%mm3        \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "paddw %%mm3, %%mm0         \n\t"\
        "psllw $2, %%mm2            \n\t"\
        "psubw %%mm1, %%mm2         \n\t"\
        "pmullw %%mm4, %%mm2        \n\t"\
        "paddw %%mm5, %%mm0         \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "movd   (%2), %%mm3         \n\t"\
        "psraw $5, %%mm0            \n\t"\
        "packuswb %%mm0, %%mm0      \n\t"\
        PAVGB" %%mm3, %%mm0         \n\t"\
        OP(%%mm0, (%1),%%mm6, d)\
Loren Merritt's avatar
Loren Merritt committed
1037 1038 1039 1040
        "add %4, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "add %3, %2                 \n\t"\
        : "+a"(src), "+c"(dst), "+d"(src2)\
1041
        : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride)\
Loren Merritt's avatar
Loren Merritt committed
1042 1043
        : "memory"\
    );\
Loren Merritt's avatar
Loren Merritt committed
1044
    }while(--h);\
Loren Merritt's avatar
Loren Merritt committed
1045
}\
1046
static av_noinline void OPNAME ## h264_qpel4_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
1047
    src -= 2*srcStride;\
1048
    __asm__ volatile(\
1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064
        "pxor %%mm7, %%mm7          \n\t"\
        "movd (%0), %%mm0           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm1           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm2           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm3           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm4           \n\t"\
        "add %2, %0                 \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpcklbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm4     \n\t"\
1065 1066 1067 1068 1069 1070
        QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
        QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
        QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
        QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
         \
        : "+a"(src), "+c"(dst)\
1071
        : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1072 1073 1074
        : "memory"\
    );\
}\
1075
static av_noinline void OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
1076 1077 1078 1079
    int h=4;\
    int w=3;\
    src -= 2*srcStride+2;\
    while(w--){\
1080
        __asm__ volatile(\
1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
            "pxor %%mm7, %%mm7      \n\t"\
            "movd (%0), %%mm0       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm1       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm2       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm3       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm4       \n\t"\
            "add %2, %0             \n\t"\
            "punpcklbw %%mm7, %%mm0 \n\t"\
            "punpcklbw %%mm7, %%mm1 \n\t"\
            "punpcklbw %%mm7, %%mm2 \n\t"\
            "punpcklbw %%mm7, %%mm3 \n\t"\
            "punpcklbw %%mm7, %%mm4 \n\t"\
1097 1098 1099 1100 1101 1102
            QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*3)\
            QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*3)\
            QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*3)\
            QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*3)\
             \
            : "+a"(src)\
1103
            : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1104 1105 1106 1107 1108 1109
            : "memory"\
        );\
        tmp += 4;\
        src += 4 - 9*srcStride;\
    }\
    tmp -= 3*4;\
1110
    __asm__ volatile(\
1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121
        "1:                         \n\t"\
        "movq     (%0), %%mm0       \n\t"\
        "paddw  10(%0), %%mm0       \n\t"\
        "movq    2(%0), %%mm1       \n\t"\
        "paddw   8(%0), %%mm1       \n\t"\
        "movq    4(%0), %%mm2       \n\t"\
        "paddw   6(%0), %%mm2       \n\t"\
        "psubw %%mm1, %%mm0         \n\t"/*a-b   (abccba)*/\
        "psraw $2, %%mm0            \n\t"/*(a-b)/4 */\
        "psubw %%mm1, %%mm0         \n\t"/*(a-b)/4-b */\
        "paddsw %%mm2, %%mm0        \n\t"\
1122
        "psraw $2, %%mm0            \n\t"/*((a-b)/4-b+c)/4 */\
1123
        "paddw %%mm2, %%mm0         \n\t"/*(a-5*b+20*c)/16 */\
1124 1125
        "psraw $6, %%mm0            \n\t"\
        "packuswb %%mm0, %%mm0      \n\t"\
1126
        OP(%%mm0, (%1),%%mm7, d)\
1127 1128 1129 1130
        "add $24, %0                \n\t"\
        "add %3, %1                 \n\t"\
        "decl %2                    \n\t"\
        " jnz 1b                    \n\t"\
1131
        : "+a"(tmp), "+c"(dst), "+g"(h)\
1132
        : "S"((x86_reg)dstStride)\
1133 1134 1135 1136
        : "memory"\
    );\
}\
\
1137
static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
1138
    int h=8;\
1139
    __asm__ volatile(\
1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182
        "pxor %%mm7, %%mm7          \n\t"\
        "movq %5, %%mm6             \n\t"\
        "1:                         \n\t"\
        "movq    (%0), %%mm0        \n\t"\
        "movq   1(%0), %%mm2        \n\t"\
        "movq %%mm0, %%mm1          \n\t"\
        "movq %%mm2, %%mm3          \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpckhbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpckhbw %%mm7, %%mm3     \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "paddw %%mm3, %%mm1         \n\t"\
        "psllw $2, %%mm0            \n\t"\
        "psllw $2, %%mm1            \n\t"\
        "movq   -1(%0), %%mm2       \n\t"\
        "movq    2(%0), %%mm4       \n\t"\
        "movq %%mm2, %%mm3          \n\t"\
        "movq %%mm4, %%mm5          \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpckhbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm4     \n\t"\
        "punpckhbw %%mm7, %%mm5     \n\t"\
        "paddw %%mm4, %%mm2         \n\t"\
        "paddw %%mm3, %%mm5         \n\t"\
        "psubw %%mm2, %%mm0         \n\t"\
        "psubw %%mm5, %%mm1         \n\t"\
        "pmullw %%mm6, %%mm0        \n\t"\
        "pmullw %%mm6, %%mm1        \n\t"\
        "movd   -2(%0), %%mm2       \n\t"\
        "movd    7(%0), %%mm5       \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm5     \n\t"\
        "paddw %%mm3, %%mm2         \n\t"\
        "paddw %%mm5, %%mm4         \n\t"\
        "movq %6, %%mm5             \n\t"\
        "paddw %%mm5, %%mm2         \n\t"\
        "paddw %%mm5, %%mm4         \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "paddw %%mm4, %%mm1         \n\t"\
        "psraw $5, %%mm0            \n\t"\
        "psraw $5, %%mm1            \n\t"\
        "packuswb %%mm1, %%mm0      \n\t"\
1183
        OP(%%mm0, (%1),%%mm5, q)\
1184 1185 1186 1187
        "add %3, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "decl %2                    \n\t"\
        " jnz 1b                    \n\t"\
1188
        : "+a"(src), "+c"(dst), "+g"(h)\
1189
        : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1190 1191 1192 1193
        : "memory"\
    );\
}\
\
1194
static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
Loren Merritt's avatar
Loren Merritt committed
1195
    int h=8;\
1196
    __asm__ volatile(\
Loren Merritt's avatar
Loren Merritt committed
1197
        "pxor %%mm7, %%mm7          \n\t"\
Loren Merritt's avatar
Loren Merritt committed
1198 1199 1200 1201
        "movq %0, %%mm6             \n\t"\
        :: "m"(ff_pw_5)\
    );\
    do{\
1202
    __asm__ volatile(\
Loren Merritt's avatar
Loren Merritt committed
1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234
        "movq    (%0), %%mm0        \n\t"\
        "movq   1(%0), %%mm2        \n\t"\
        "movq %%mm0, %%mm1          \n\t"\
        "movq %%mm2, %%mm3          \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpckhbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpckhbw %%mm7, %%mm3     \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "paddw %%mm3, %%mm1         \n\t"\
        "psllw $2, %%mm0            \n\t"\
        "psllw $2, %%mm1            \n\t"\
        "movq   -1(%0), %%mm2       \n\t"\
        "movq    2(%0), %%mm4       \n\t"\
        "movq %%mm2, %%mm3          \n\t"\
        "movq %%mm4, %%mm5          \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpckhbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm4     \n\t"\
        "punpckhbw %%mm7, %%mm5     \n\t"\
        "paddw %%mm4, %%mm2         \n\t"\
        "paddw %%mm3, %%mm5         \n\t"\
        "psubw %%mm2, %%mm0         \n\t"\
        "psubw %%mm5, %%mm1         \n\t"\
        "pmullw %%mm6, %%mm0        \n\t"\
        "pmullw %%mm6, %%mm1        \n\t"\
        "movd   -2(%0), %%mm2       \n\t"\
        "movd    7(%0), %%mm5       \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm5     \n\t"\
        "paddw %%mm3, %%mm2         \n\t"\
        "paddw %%mm5, %%mm4         \n\t"\
Loren Merritt's avatar
Loren Merritt committed
1235
        "movq %5, %%mm5             \n\t"\
Loren Merritt's avatar
Loren Merritt committed
1236 1237 1238 1239 1240 1241 1242 1243 1244 1245
        "paddw %%mm5, %%mm2         \n\t"\
        "paddw %%mm5, %%mm4         \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "paddw %%mm4, %%mm1         \n\t"\
        "psraw $5, %%mm0            \n\t"\
        "psraw $5, %%mm1            \n\t"\
        "movq (%2), %%mm4           \n\t"\
        "packuswb %%mm1, %%mm0      \n\t"\
        PAVGB" %%mm4, %%mm0         \n\t"\
        OP(%%mm0, (%1),%%mm5, q)\
Loren Merritt's avatar
Loren Merritt committed
1246 1247 1248 1249
        "add %4, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "add %3, %2                 \n\t"\
        : "+a"(src), "+c"(dst), "+d"(src2)\
1250
        : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\
Loren Merritt's avatar
Loren Merritt committed
1251
          "m"(ff_pw_16)\
Loren Merritt's avatar
Loren Merritt committed
1252 1253
        : "memory"\
    );\
Loren Merritt's avatar
Loren Merritt committed
1254
    }while(--h);\
Loren Merritt's avatar
Loren Merritt committed
1255 1256
}\
\
1257
static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
1258
    int w= 2;\
1259 1260
    src -= 2*srcStride;\
    \
1261
    while(w--){\
1262
      __asm__ volatile(\
1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278
        "pxor %%mm7, %%mm7          \n\t"\
        "movd (%0), %%mm0           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm1           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm2           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm3           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm4           \n\t"\
        "add %2, %0                 \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpcklbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm4     \n\t"\
1279 1280 1281 1282 1283 1284 1285 1286 1287 1288
        QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
        QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
        QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
        QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
        QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
        QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
        QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
        QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
         \
        : "+a"(src), "+c"(dst)\
1289
        : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1290 1291
        : "memory"\
     );\
1292
     if(h==16){\
1293
        __asm__ volatile(\
1294 1295 1296 1297 1298 1299 1300 1301 1302 1303
            QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
            QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
            QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
            QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
            QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
            QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
            QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
            QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
            \
           : "+a"(src), "+c"(dst)\
1304
           : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1305 1306 1307 1308 1309
           : "memory"\
        );\
     }\
     src += 4-(h+5)*srcStride;\
     dst += 4-h*dstStride;\
1310 1311
   }\
}\
1312
static av_always_inline void OPNAME ## h264_qpel8or16_hv1_lowpass_ ## MMX(int16_t *tmp, uint8_t *src, int tmpStride, int srcStride, int size){\
1313
    int w = (size+8)>>2;\
1314 1315
    src -= 2*srcStride+2;\
    while(w--){\
1316
        __asm__ volatile(\
1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332
            "pxor %%mm7, %%mm7      \n\t"\
            "movd (%0), %%mm0       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm1       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm2       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm3       \n\t"\
            "add %2, %0             \n\t"\
            "movd (%0), %%mm4       \n\t"\
            "add %2, %0             \n\t"\
            "punpcklbw %%mm7, %%mm0 \n\t"\
            "punpcklbw %%mm7, %%mm1 \n\t"\
            "punpcklbw %%mm7, %%mm2 \n\t"\
            "punpcklbw %%mm7, %%mm3 \n\t"\
            "punpcklbw %%mm7, %%mm4 \n\t"\
1333 1334 1335 1336 1337 1338 1339 1340
            QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*48)\
            QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*48)\
            QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*48)\
            QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*48)\
            QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 4*48)\
            QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 5*48)\
            QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 6*48)\
            QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 7*48)\
1341
            : "+a"(src)\
1342
            : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1343 1344
            : "memory"\
        );\
1345
        if(size==16){\
1346
            __asm__ volatile(\
1347 1348 1349 1350 1351 1352 1353 1354 1355
                QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1,  8*48)\
                QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2,  9*48)\
                QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 10*48)\
                QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 11*48)\
                QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 12*48)\
                QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 13*48)\
                QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 14*48)\
                QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 15*48)\
                : "+a"(src)\
1356
                : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1357 1358 1359
                : "memory"\
            );\
        }\
1360
        tmp += 4;\
1361
        src += 4 - (size+5)*srcStride;\
1362
    }\
1363 1364 1365
}\
static av_always_inline void OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, int dstStride, int tmpStride, int size){\
    int w = size>>4;\
1366
    do{\
1367
    int h = size;\
1368
    __asm__ volatile(\
1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396
        "1:                         \n\t"\
        "movq     (%0), %%mm0       \n\t"\
        "movq    8(%0), %%mm3       \n\t"\
        "movq    2(%0), %%mm1       \n\t"\
        "movq   10(%0), %%mm4       \n\t"\
        "paddw   %%mm4, %%mm0       \n\t"\
        "paddw   %%mm3, %%mm1       \n\t"\
        "paddw  18(%0), %%mm3       \n\t"\
        "paddw  16(%0), %%mm4       \n\t"\
        "movq    4(%0), %%mm2       \n\t"\
        "movq   12(%0), %%mm5       \n\t"\
        "paddw   6(%0), %%mm2       \n\t"\
        "paddw  14(%0), %%mm5       \n\t"\
        "psubw %%mm1, %%mm0         \n\t"\
        "psubw %%mm4, %%mm3         \n\t"\
        "psraw $2, %%mm0            \n\t"\
        "psraw $2, %%mm3            \n\t"\
        "psubw %%mm1, %%mm0         \n\t"\
        "psubw %%mm4, %%mm3         \n\t"\
        "paddsw %%mm2, %%mm0        \n\t"\
        "paddsw %%mm5, %%mm3        \n\t"\
        "psraw $2, %%mm0            \n\t"\
        "psraw $2, %%mm3            \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "paddw %%mm5, %%mm3         \n\t"\
        "psraw $6, %%mm0            \n\t"\
        "psraw $6, %%mm3            \n\t"\
        "packuswb %%mm3, %%mm0      \n\t"\
1397
        OP(%%mm0, (%1),%%mm7, q)\
1398
        "add $48, %0                \n\t"\
1399 1400 1401
        "add %3, %1                 \n\t"\
        "decl %2                    \n\t"\
        " jnz 1b                    \n\t"\
1402
        : "+a"(tmp), "+c"(dst), "+g"(h)\
1403
        : "S"((x86_reg)dstStride)\
1404 1405
        : "memory"\
    );\
1406 1407 1408
    tmp += 8 - size*24;\
    dst += 8 - size*dstStride;\
    }while(w--);\
1409
}\
1410 1411 1412 1413
\
static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst  , src  , dstStride, srcStride, 8);\
}\
1414
static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
1415 1416
    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst  , src  , dstStride, srcStride, 16);\
    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
1417 1418
}\
\
1419
static void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
1420 1421 1422 1423 1424 1425 1426 1427
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst  , src  , dstStride, srcStride);\
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
    src += 8*srcStride;\
    dst += 8*dstStride;\
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst  , src  , dstStride, srcStride);\
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
}\
\
1428
static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
Loren Merritt's avatar
Loren Merritt committed
1429 1430 1431 1432 1433 1434 1435 1436 1437
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst  , src  , src2  , dstStride, src2Stride);\
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
    src += 8*dstStride;\
    dst += 8*dstStride;\
    src2 += 8*src2Stride;\
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst  , src  , src2  , dstStride, src2Stride);\
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
}\
\
1438 1439 1440 1441
static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\
          put_h264_qpel8or16_hv1_lowpass_ ## MMX(tmp, src, tmpStride, srcStride, size);\
    OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(dst, tmp, dstStride, tmpStride, size);\
}\
1442 1443 1444 1445
static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
    OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst  , tmp  , src  , dstStride, tmpStride, srcStride, 8);\
}\
\
1446
static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
1447
    OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst  , tmp  , src  , dstStride, tmpStride, srcStride, 16);\
1448
}\
1449
\
1450
static av_noinline void OPNAME ## pixels4_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
1451
{\
1452
    __asm__ volatile(\
1453 1454
        "movq      (%1), %%mm0          \n\t"\
        "movq    24(%1), %%mm1          \n\t"\
Loren Merritt's avatar
Loren Merritt committed
1455 1456
        "psraw      $5,  %%mm0          \n\t"\
        "psraw      $5,  %%mm1          \n\t"\
1457 1458
        "packuswb %%mm0, %%mm0          \n\t"\
        "packuswb %%mm1, %%mm1          \n\t"\
1459 1460 1461 1462 1463 1464 1465 1466
        PAVGB"     (%0), %%mm0          \n\t"\
        PAVGB"  (%0,%3), %%mm1          \n\t"\
        OP(%%mm0, (%2),    %%mm4, d)\
        OP(%%mm1, (%2,%4), %%mm5, d)\
        "lea  (%0,%3,2), %0             \n\t"\
        "lea  (%2,%4,2), %2             \n\t"\
        "movq    48(%1), %%mm0          \n\t"\
        "movq    72(%1), %%mm1          \n\t"\
Loren Merritt's avatar
Loren Merritt committed
1467 1468
        "psraw      $5,  %%mm0          \n\t"\
        "psraw      $5,  %%mm1          \n\t"\
1469 1470
        "packuswb %%mm0, %%mm0          \n\t"\
        "packuswb %%mm1, %%mm1          \n\t"\
1471 1472 1473 1474 1475
        PAVGB"     (%0), %%mm0          \n\t"\
        PAVGB"  (%0,%3), %%mm1          \n\t"\
        OP(%%mm0, (%2),    %%mm4, d)\
        OP(%%mm1, (%2,%4), %%mm5, d)\
        :"+a"(src8), "+c"(src16), "+d"(dst)\
1476
        :"S"((x86_reg)src8Stride), "D"((x86_reg)dstStride)\
1477 1478
        :"memory");\
}\
1479
static av_noinline void OPNAME ## pixels8_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
1480
{\
1481
    do{\
1482
    __asm__ volatile(\
1483 1484
        "movq      (%1), %%mm0          \n\t"\
        "movq     8(%1), %%mm1          \n\t"\
1485 1486
        "movq    48(%1), %%mm2          \n\t"\
        "movq  8+48(%1), %%mm3          \n\t"\
Loren Merritt's avatar
Loren Merritt committed
1487 1488
        "psraw      $5,  %%mm0          \n\t"\
        "psraw      $5,  %%mm1          \n\t"\
1489 1490
        "psraw      $5,  %%mm2          \n\t"\
        "psraw      $5,  %%mm3          \n\t"\
1491
        "packuswb %%mm1, %%mm0          \n\t"\
1492
        "packuswb %%mm3, %%mm2          \n\t"\
1493
        PAVGB"     (%0), %%mm0          \n\t"\
1494
        PAVGB"  (%0,%3), %%mm2          \n\t"\
1495
        OP(%%mm0, (%2), %%mm5, q)\
1496 1497
        OP(%%mm2, (%2,%4), %%mm5, q)\
        ::"a"(src8), "c"(src16), "d"(dst),\
1498
          "r"((x86_reg)src8Stride), "r"((x86_reg)dstStride)\
1499
        :"memory");\
1500 1501 1502 1503
        src8 += 2L*src8Stride;\
        src16 += 48;\
        dst += 2L*dstStride;\
    }while(h-=2);\
1504 1505 1506 1507 1508 1509 1510
}\
static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
{\
    OPNAME ## pixels8_l2_shift5_ ## MMX(dst  , src16  , src8  , dstStride, src8Stride, h);\
    OPNAME ## pixels8_l2_shift5_ ## MMX(dst+8, src16+8, src8+8, dstStride, src8Stride, h);\
}\

1511

1512
#if ARCH_X86_64
1513
#define QPEL_H264_H16_XMM(OPNAME, OP, MMX)\
1514 1515
static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
    int h=16;\
1516
    __asm__ volatile(\
1517 1518 1519 1520
        "pxor %%xmm15, %%xmm15      \n\t"\
        "movdqa %6, %%xmm14         \n\t"\
        "movdqa %7, %%xmm13         \n\t"\
        "1:                         \n\t"\
1521 1522
        "lddqu    6(%0), %%xmm1     \n\t"\
        "lddqu   -2(%0), %%xmm7     \n\t"\
1523 1524 1525 1526 1527 1528 1529 1530 1531 1532
        "movdqa  %%xmm1, %%xmm0     \n\t"\
        "punpckhbw %%xmm15, %%xmm1  \n\t"\
        "punpcklbw %%xmm15, %%xmm0  \n\t"\
        "punpcklbw %%xmm15, %%xmm7  \n\t"\
        "movdqa  %%xmm1, %%xmm2     \n\t"\
        "movdqa  %%xmm0, %%xmm6     \n\t"\
        "movdqa  %%xmm1, %%xmm3     \n\t"\
        "movdqa  %%xmm0, %%xmm8     \n\t"\
        "movdqa  %%xmm1, %%xmm4     \n\t"\
        "movdqa  %%xmm0, %%xmm9     \n\t"\
1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546
        "movdqa  %%xmm0, %%xmm12    \n\t"\
        "movdqa  %%xmm1, %%xmm11    \n\t"\
        "palignr $10,%%xmm0, %%xmm11\n\t"\
        "palignr $10,%%xmm7, %%xmm12\n\t"\
        "palignr $2, %%xmm0, %%xmm4 \n\t"\
        "palignr $2, %%xmm7, %%xmm9 \n\t"\
        "palignr $4, %%xmm0, %%xmm3 \n\t"\
        "palignr $4, %%xmm7, %%xmm8 \n\t"\
        "palignr $6, %%xmm0, %%xmm2 \n\t"\
        "palignr $6, %%xmm7, %%xmm6 \n\t"\
        "paddw   %%xmm0 ,%%xmm11    \n\t"\
        "palignr $8, %%xmm0, %%xmm1 \n\t"\
        "palignr $8, %%xmm7, %%xmm0 \n\t"\
        "paddw   %%xmm12,%%xmm7     \n\t"\
1547 1548 1549 1550 1551 1552 1553 1554
        "paddw   %%xmm3, %%xmm2     \n\t"\
        "paddw   %%xmm8, %%xmm6     \n\t"\
        "paddw   %%xmm4, %%xmm1     \n\t"\
        "paddw   %%xmm9, %%xmm0     \n\t"\
        "psllw   $2,     %%xmm2     \n\t"\
        "psllw   $2,     %%xmm6     \n\t"\
        "psubw   %%xmm1, %%xmm2     \n\t"\
        "psubw   %%xmm0, %%xmm6     \n\t"\
1555 1556
        "paddw   %%xmm13,%%xmm11    \n\t"\
        "paddw   %%xmm13,%%xmm7     \n\t"\
1557 1558 1559
        "pmullw  %%xmm14,%%xmm2     \n\t"\
        "pmullw  %%xmm14,%%xmm6     \n\t"\
        "lddqu   (%2),   %%xmm3     \n\t"\
1560 1561
        "paddw   %%xmm11,%%xmm2     \n\t"\
        "paddw   %%xmm7, %%xmm6     \n\t"\
1562 1563 1564 1565 1566 1567 1568 1569
        "psraw   $5,     %%xmm2     \n\t"\
        "psraw   $5,     %%xmm6     \n\t"\
        "packuswb %%xmm2,%%xmm6     \n\t"\
        "pavgb   %%xmm3, %%xmm6     \n\t"\
        OP(%%xmm6, (%1), %%xmm4, dqa)\
        "add %5, %0                 \n\t"\
        "add %5, %1                 \n\t"\
        "add %4, %2                 \n\t"\
1570
        "decl %3                    \n\t"\
1571 1572
        "jg 1b                      \n\t"\
        : "+a"(src), "+c"(dst), "+d"(src2), "+g"(h)\
1573
        : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\
1574 1575 1576 1577 1578
          "m"(ff_pw_5), "m"(ff_pw_16)\
        : "memory"\
    );\
}
#else // ARCH_X86_64
1579
#define QPEL_H264_H16_XMM(OPNAME, OP, MMX)\
1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590
static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst  , src  , src2  , dstStride, src2Stride);\
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
    src += 8*dstStride;\
    dst += 8*dstStride;\
    src2 += 8*src2Stride;\
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst  , src  , src2  , dstStride, src2Stride);\
    OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
}
#endif // ARCH_X86_64

1591
#define QPEL_H264_H_XMM(OPNAME, OP, MMX)\
1592 1593
static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
    int h=8;\
1594
    __asm__ volatile(\
1595 1596 1597 1598 1599
        "pxor %%xmm7, %%xmm7        \n\t"\
        "movdqa %0, %%xmm6          \n\t"\
        :: "m"(ff_pw_5)\
    );\
    do{\
1600
    __asm__ volatile(\
1601
        "lddqu   -2(%0), %%xmm1     \n\t"\
1602 1603 1604 1605 1606 1607 1608
        "movdqa  %%xmm1, %%xmm0     \n\t"\
        "punpckhbw %%xmm7, %%xmm1   \n\t"\
        "punpcklbw %%xmm7, %%xmm0   \n\t"\
        "movdqa  %%xmm1, %%xmm2     \n\t"\
        "movdqa  %%xmm1, %%xmm3     \n\t"\
        "movdqa  %%xmm1, %%xmm4     \n\t"\
        "movdqa  %%xmm1, %%xmm5     \n\t"\
1609 1610 1611 1612 1613 1614
        "palignr $2, %%xmm0, %%xmm4 \n\t"\
        "palignr $4, %%xmm0, %%xmm3 \n\t"\
        "palignr $6, %%xmm0, %%xmm2 \n\t"\
        "palignr $8, %%xmm0, %%xmm1 \n\t"\
        "palignr $10,%%xmm0, %%xmm5 \n\t"\
        "paddw   %%xmm5, %%xmm0     \n\t"\
1615 1616 1617 1618 1619
        "paddw   %%xmm3, %%xmm2     \n\t"\
        "paddw   %%xmm4, %%xmm1     \n\t"\
        "psllw   $2,     %%xmm2     \n\t"\
        "movq    (%2),   %%xmm3     \n\t"\
        "psubw   %%xmm1, %%xmm2     \n\t"\
1620
        "paddw   %5,     %%xmm0     \n\t"\
1621
        "pmullw  %%xmm6, %%xmm2     \n\t"\
1622
        "paddw   %%xmm0, %%xmm2     \n\t"\
1623 1624 1625 1626 1627 1628 1629 1630
        "psraw   $5,     %%xmm2     \n\t"\
        "packuswb %%xmm2, %%xmm2    \n\t"\
        "pavgb   %%xmm3, %%xmm2     \n\t"\
        OP(%%xmm2, (%1), %%xmm4, q)\
        "add %4, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "add %3, %2                 \n\t"\
        : "+a"(src), "+c"(dst), "+d"(src2)\
1631
        : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\
1632 1633 1634 1635 1636
          "m"(ff_pw_16)\
        : "memory"\
    );\
    }while(--h);\
}\
1637
QPEL_H264_H16_XMM(OPNAME, OP, MMX)\
1638
\
1639 1640
static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    int h=8;\
1641
    __asm__ volatile(\
1642 1643 1644
        "pxor %%xmm7, %%xmm7        \n\t"\
        "movdqa %5, %%xmm6          \n\t"\
        "1:                         \n\t"\
1645
        "lddqu   -2(%0), %%xmm1     \n\t"\
1646 1647 1648 1649 1650 1651 1652
        "movdqa  %%xmm1, %%xmm0     \n\t"\
        "punpckhbw %%xmm7, %%xmm1   \n\t"\
        "punpcklbw %%xmm7, %%xmm0   \n\t"\
        "movdqa  %%xmm1, %%xmm2     \n\t"\
        "movdqa  %%xmm1, %%xmm3     \n\t"\
        "movdqa  %%xmm1, %%xmm4     \n\t"\
        "movdqa  %%xmm1, %%xmm5     \n\t"\
1653 1654 1655 1656 1657 1658
        "palignr $2, %%xmm0, %%xmm4 \n\t"\
        "palignr $4, %%xmm0, %%xmm3 \n\t"\
        "palignr $6, %%xmm0, %%xmm2 \n\t"\
        "palignr $8, %%xmm0, %%xmm1 \n\t"\
        "palignr $10,%%xmm0, %%xmm5 \n\t"\
        "paddw   %%xmm5, %%xmm0     \n\t"\
1659 1660 1661 1662
        "paddw   %%xmm3, %%xmm2     \n\t"\
        "paddw   %%xmm4, %%xmm1     \n\t"\
        "psllw   $2,     %%xmm2     \n\t"\
        "psubw   %%xmm1, %%xmm2     \n\t"\
1663
        "paddw   %6,     %%xmm0     \n\t"\
1664
        "pmullw  %%xmm6, %%xmm2     \n\t"\
1665
        "paddw   %%xmm0, %%xmm2     \n\t"\
1666 1667 1668 1669 1670 1671 1672 1673
        "psraw   $5,     %%xmm2     \n\t"\
        "packuswb %%xmm2, %%xmm2    \n\t"\
        OP(%%xmm2, (%1), %%xmm4, q)\
        "add %3, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "decl %2                    \n\t"\
        " jnz 1b                    \n\t"\
        : "+a"(src), "+c"(dst), "+g"(h)\
1674
        : "D"((x86_reg)srcStride), "S"((x86_reg)dstStride),\
1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688
          "m"(ff_pw_5), "m"(ff_pw_16)\
        : "memory"\
    );\
}\
static void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst  , src  , dstStride, srcStride);\
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
    src += 8*srcStride;\
    dst += 8*dstStride;\
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst  , src  , dstStride, srcStride);\
    OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
}\

#define QPEL_H264_V_XMM(OPNAME, OP, MMX)\
1689 1690 1691
static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
    src -= 2*srcStride;\
    \
1692
    __asm__ volatile(\
1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718
        "pxor %%xmm7, %%xmm7        \n\t"\
        "movq (%0), %%xmm0          \n\t"\
        "add %2, %0                 \n\t"\
        "movq (%0), %%xmm1          \n\t"\
        "add %2, %0                 \n\t"\
        "movq (%0), %%xmm2          \n\t"\
        "add %2, %0                 \n\t"\
        "movq (%0), %%xmm3          \n\t"\
        "add %2, %0                 \n\t"\
        "movq (%0), %%xmm4          \n\t"\
        "add %2, %0                 \n\t"\
        "punpcklbw %%xmm7, %%xmm0   \n\t"\
        "punpcklbw %%xmm7, %%xmm1   \n\t"\
        "punpcklbw %%xmm7, %%xmm2   \n\t"\
        "punpcklbw %%xmm7, %%xmm3   \n\t"\
        "punpcklbw %%xmm7, %%xmm4   \n\t"\
        QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
        QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
        QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
        QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
        QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
        QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
        QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
        QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
         \
        : "+a"(src), "+c"(dst)\
1719
        : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1720 1721 1722
        : "memory"\
    );\
    if(h==16){\
1723
        __asm__ volatile(\
1724 1725 1726 1727 1728 1729 1730 1731 1732 1733
            QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
            QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
            QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
            QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
            QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
            QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
            QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
            QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
            \
            : "+a"(src), "+c"(dst)\
1734
            : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
1735 1736 1737 1738
            : "memory"\
        );\
    }\
}\
1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750
static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst  , src  , dstStride, srcStride, 8);\
}\
static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst  , src  , dstStride, srcStride, 16);\
    OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
}

static av_always_inline void put_h264_qpel8or16_hv1_lowpass_sse2(int16_t *tmp, uint8_t *src, int tmpStride, int srcStride, int size){
    int w = (size+8)>>3;
    src -= 2*srcStride+2;
    while(w--){
1751
        __asm__ volatile(
1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776
            "pxor %%xmm7, %%xmm7        \n\t"
            "movq (%0), %%xmm0          \n\t"
            "add %2, %0                 \n\t"
            "movq (%0), %%xmm1          \n\t"
            "add %2, %0                 \n\t"
            "movq (%0), %%xmm2          \n\t"
            "add %2, %0                 \n\t"
            "movq (%0), %%xmm3          \n\t"
            "add %2, %0                 \n\t"
            "movq (%0), %%xmm4          \n\t"
            "add %2, %0                 \n\t"
            "punpcklbw %%xmm7, %%xmm0   \n\t"
            "punpcklbw %%xmm7, %%xmm1   \n\t"
            "punpcklbw %%xmm7, %%xmm2   \n\t"
            "punpcklbw %%xmm7, %%xmm3   \n\t"
            "punpcklbw %%xmm7, %%xmm4   \n\t"
            QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 0*48)
            QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 1*48)
            QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 2*48)
            QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 3*48)
            QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 4*48)
            QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 5*48)
            QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 6*48)
            QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 7*48)
            : "+a"(src)
1777
            : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)
1778 1779 1780
            : "memory"
        );
        if(size==16){
1781
            __asm__ volatile(
1782 1783 1784 1785 1786 1787 1788 1789 1790
                QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1,  8*48)
                QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2,  9*48)
                QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 10*48)
                QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 11*48)
                QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 12*48)
                QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 13*48)
                QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 14*48)
                QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 15*48)
                : "+a"(src)
1791
                : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)
1792 1793 1794 1795 1796 1797 1798 1799 1800 1801
                : "memory"
            );
        }
        tmp += 8;
        src += 8 - (size+5)*srcStride;
    }
}

#define QPEL_H264_HV2_XMM(OPNAME, OP, MMX)\
static av_always_inline void OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, int dstStride, int tmpStride, int size){\
1802
    int h = size;\
1803
    if(size == 16){\
1804
        __asm__ volatile(\
1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854
            "1:                         \n\t"\
            "movdqa 32(%0), %%xmm4      \n\t"\
            "movdqa 16(%0), %%xmm5      \n\t"\
            "movdqa   (%0), %%xmm7      \n\t"\
            "movdqa %%xmm4, %%xmm3      \n\t"\
            "movdqa %%xmm4, %%xmm2      \n\t"\
            "movdqa %%xmm4, %%xmm1      \n\t"\
            "movdqa %%xmm4, %%xmm0      \n\t"\
            "palignr $10, %%xmm5, %%xmm0 \n\t"\
            "palignr  $8, %%xmm5, %%xmm1 \n\t"\
            "palignr  $6, %%xmm5, %%xmm2 \n\t"\
            "palignr  $4, %%xmm5, %%xmm3 \n\t"\
            "palignr  $2, %%xmm5, %%xmm4 \n\t"\
            "paddw  %%xmm5, %%xmm0      \n\t"\
            "paddw  %%xmm4, %%xmm1      \n\t"\
            "paddw  %%xmm3, %%xmm2      \n\t"\
            "movdqa %%xmm5, %%xmm6      \n\t"\
            "movdqa %%xmm5, %%xmm4      \n\t"\
            "movdqa %%xmm5, %%xmm3      \n\t"\
            "palignr  $8, %%xmm7, %%xmm4 \n\t"\
            "palignr  $2, %%xmm7, %%xmm6 \n\t"\
            "palignr $10, %%xmm7, %%xmm3 \n\t"\
            "paddw  %%xmm6, %%xmm4      \n\t"\
            "movdqa %%xmm5, %%xmm6      \n\t"\
            "palignr  $6, %%xmm7, %%xmm5 \n\t"\
            "palignr  $4, %%xmm7, %%xmm6 \n\t"\
            "paddw  %%xmm7, %%xmm3      \n\t"\
            "paddw  %%xmm6, %%xmm5      \n\t"\
            \
            "psubw  %%xmm1, %%xmm0      \n\t"\
            "psubw  %%xmm4, %%xmm3      \n\t"\
            "psraw      $2, %%xmm0      \n\t"\
            "psraw      $2, %%xmm3      \n\t"\
            "psubw  %%xmm1, %%xmm0      \n\t"\
            "psubw  %%xmm4, %%xmm3      \n\t"\
            "paddw  %%xmm2, %%xmm0      \n\t"\
            "paddw  %%xmm5, %%xmm3      \n\t"\
            "psraw      $2, %%xmm0      \n\t"\
            "psraw      $2, %%xmm3      \n\t"\
            "paddw  %%xmm2, %%xmm0      \n\t"\
            "paddw  %%xmm5, %%xmm3      \n\t"\
            "psraw      $6, %%xmm0      \n\t"\
            "psraw      $6, %%xmm3      \n\t"\
            "packuswb %%xmm0, %%xmm3    \n\t"\
            OP(%%xmm3, (%1), %%xmm7, dqa)\
            "add $48, %0                \n\t"\
            "add %3, %1                 \n\t"\
            "decl %2                    \n\t"\
            " jnz 1b                    \n\t"\
            : "+a"(tmp), "+c"(dst), "+g"(h)\
1855
            : "S"((x86_reg)dstStride)\
1856 1857 1858
            : "memory"\
        );\
    }else{\
1859
        __asm__ volatile(\
1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888
            "1:                         \n\t"\
            "movdqa 16(%0), %%xmm1      \n\t"\
            "movdqa   (%0), %%xmm0      \n\t"\
            "movdqa %%xmm1, %%xmm2      \n\t"\
            "movdqa %%xmm1, %%xmm3      \n\t"\
            "movdqa %%xmm1, %%xmm4      \n\t"\
            "movdqa %%xmm1, %%xmm5      \n\t"\
            "palignr $10, %%xmm0, %%xmm5 \n\t"\
            "palignr  $8, %%xmm0, %%xmm4 \n\t"\
            "palignr  $6, %%xmm0, %%xmm3 \n\t"\
            "palignr  $4, %%xmm0, %%xmm2 \n\t"\
            "palignr  $2, %%xmm0, %%xmm1 \n\t"\
            "paddw  %%xmm5, %%xmm0      \n\t"\
            "paddw  %%xmm4, %%xmm1      \n\t"\
            "paddw  %%xmm3, %%xmm2      \n\t"\
            "psubw  %%xmm1, %%xmm0      \n\t"\
            "psraw      $2, %%xmm0      \n\t"\
            "psubw  %%xmm1, %%xmm0      \n\t"\
            "paddw  %%xmm2, %%xmm0      \n\t"\
            "psraw      $2, %%xmm0      \n\t"\
            "paddw  %%xmm2, %%xmm0      \n\t"\
            "psraw      $6, %%xmm0      \n\t"\
            "packuswb %%xmm0, %%xmm0    \n\t"\
            OP(%%xmm0, (%1), %%xmm7, q)\
            "add $48, %0                \n\t"\
            "add %3, %1                 \n\t"\
            "decl %2                    \n\t"\
            " jnz 1b                    \n\t"\
            : "+a"(tmp), "+c"(dst), "+g"(h)\
1889
            : "S"((x86_reg)dstStride)\
1890 1891 1892
            : "memory"\
        );\
    }\
1893 1894 1895 1896 1897 1898
}

#define QPEL_H264_HV_XMM(OPNAME, OP, MMX)\
static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\
          put_h264_qpel8or16_hv1_lowpass_sse2(tmp, src, tmpStride, srcStride, size);\
    OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(dst, tmp, dstStride, tmpStride, size);\
1899 1900
}\
static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
1901
    OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst, tmp, src, dstStride, tmpStride, srcStride, 8);\
1902 1903
}\
static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
1904
    OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst, tmp, src, dstStride, tmpStride, srcStride, 16);\
1905 1906
}\

1907 1908 1909 1910
#define put_pixels8_l2_sse2 put_pixels8_l2_mmx2
#define avg_pixels8_l2_sse2 avg_pixels8_l2_mmx2
#define put_pixels16_l2_sse2 put_pixels16_l2_mmx2
#define avg_pixels16_l2_sse2 avg_pixels16_l2_mmx2
1911 1912
#define put_pixels8_l2_ssse3 put_pixels8_l2_mmx2
#define avg_pixels8_l2_ssse3 avg_pixels8_l2_mmx2
1913
#define put_pixels16_l2_ssse3 put_pixels16_l2_mmx2
1914 1915
#define avg_pixels16_l2_ssse3 avg_pixels16_l2_mmx2

1916 1917 1918 1919
#define put_pixels8_l2_shift5_sse2 put_pixels8_l2_shift5_mmx2
#define avg_pixels8_l2_shift5_sse2 avg_pixels8_l2_shift5_mmx2
#define put_pixels16_l2_shift5_sse2 put_pixels16_l2_shift5_mmx2
#define avg_pixels16_l2_shift5_sse2 avg_pixels16_l2_shift5_mmx2
1920 1921
#define put_pixels8_l2_shift5_ssse3 put_pixels8_l2_shift5_mmx2
#define avg_pixels8_l2_shift5_ssse3 avg_pixels8_l2_shift5_mmx2
1922
#define put_pixels16_l2_shift5_ssse3 put_pixels16_l2_shift5_mmx2
1923 1924
#define avg_pixels16_l2_shift5_ssse3 avg_pixels16_l2_shift5_mmx2

1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936
#define put_h264_qpel8_h_lowpass_l2_sse2 put_h264_qpel8_h_lowpass_l2_mmx2
#define avg_h264_qpel8_h_lowpass_l2_sse2 avg_h264_qpel8_h_lowpass_l2_mmx2
#define put_h264_qpel16_h_lowpass_l2_sse2 put_h264_qpel16_h_lowpass_l2_mmx2
#define avg_h264_qpel16_h_lowpass_l2_sse2 avg_h264_qpel16_h_lowpass_l2_mmx2

#define put_h264_qpel8_v_lowpass_ssse3 put_h264_qpel8_v_lowpass_sse2
#define avg_h264_qpel8_v_lowpass_ssse3 avg_h264_qpel8_v_lowpass_sse2
#define put_h264_qpel16_v_lowpass_ssse3 put_h264_qpel16_v_lowpass_sse2
#define avg_h264_qpel16_v_lowpass_ssse3 avg_h264_qpel16_v_lowpass_sse2

#define put_h264_qpel8or16_hv2_lowpass_sse2 put_h264_qpel8or16_hv2_lowpass_mmx2
#define avg_h264_qpel8or16_hv2_lowpass_sse2 avg_h264_qpel8or16_hv2_lowpass_mmx2
1937

1938
#define H264_MC(OPNAME, SIZE, MMX, ALIGN) \
1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953
H264_MC_C(OPNAME, SIZE, MMX, ALIGN)\
H264_MC_V(OPNAME, SIZE, MMX, ALIGN)\
H264_MC_H(OPNAME, SIZE, MMX, ALIGN)\
H264_MC_HV(OPNAME, SIZE, MMX, ALIGN)\

static void put_h264_qpel16_mc00_sse2 (uint8_t *dst, uint8_t *src, int stride){
    put_pixels16_sse2(dst, src, stride, 16);
}
static void avg_h264_qpel16_mc00_sse2 (uint8_t *dst, uint8_t *src, int stride){
    avg_pixels16_sse2(dst, src, stride, 16);
}
#define put_h264_qpel8_mc00_sse2 put_h264_qpel8_mc00_mmx2
#define avg_h264_qpel8_mc00_sse2 avg_h264_qpel8_mc00_mmx2

#define H264_MC_C(OPNAME, SIZE, MMX, ALIGN) \
1954
static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
1955
    OPNAME ## pixels ## SIZE ## _ ## MMX(dst, src, stride, SIZE);\
1956
}\
1957 1958

#define H264_MC_H(OPNAME, SIZE, MMX, ALIGN) \
1959
static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
Loren Merritt's avatar
Loren Merritt committed
1960
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src, stride, stride);\
1961 1962 1963 1964 1965 1966 1967
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## MMX(dst, src, stride, stride);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
Loren Merritt's avatar
Loren Merritt committed
1968
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src+1, stride, stride);\
1969
}\
1970 1971

#define H264_MC_V(OPNAME, SIZE, MMX, ALIGN) \
1972
static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
1973
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
1974 1975
    put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
    OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src, temp, stride, stride, SIZE);\
1976 1977 1978 1979 1980 1981 1982
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
    OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## MMX(dst, src, stride, stride);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
1983
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
1984 1985
    put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
    OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src+stride, temp, stride, stride, SIZE);\
1986
}\
1987 1988

#define H264_MC_HV(OPNAME, SIZE, MMX, ALIGN) \
1989
static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
1990
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
1991 1992
    put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, temp, stride, SIZE);\
1993 1994 1995
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
1996
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
1997 1998
    put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src+1, SIZE, stride);\
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, temp, stride, SIZE);\
1999 2000 2001
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2002
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
2003 2004
    put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, temp, stride, SIZE);\
2005 2006 2007
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2008
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
2009 2010
    put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src+1, SIZE, stride);\
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, temp, stride, SIZE);\
2011 2012 2013
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2014
    DECLARE_ALIGNED(ALIGN, uint16_t, temp)[SIZE*(SIZE<8?12:24)];\
2015
    OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(dst, temp, src, stride, SIZE, stride);\
2016 2017 2018
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2019
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
2020 2021
    uint8_t * const halfHV= temp;\
    int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
2022
    assert(((int)temp & 7) == 0);\
2023
    put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
Loren Merritt's avatar
Loren Merritt committed
2024
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfHV, stride, SIZE);\
2025 2026 2027
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2028
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
2029 2030
    uint8_t * const halfHV= temp;\
    int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
2031
    assert(((int)temp & 7) == 0);\
2032
    put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
Loren Merritt's avatar
Loren Merritt committed
2033
    OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfHV, stride, SIZE);\
2034 2035 2036
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2037
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
2038 2039
    uint8_t * const halfHV= temp;\
    int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
2040
    assert(((int)temp & 7) == 0);\
2041 2042
    put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
    OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+2, halfHV, stride, SIZE, SIZE);\
2043 2044 2045
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
2046
    DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
2047 2048
    uint8_t * const halfHV= temp;\
    int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
2049
    assert(((int)temp & 7) == 0);\
2050 2051
    put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
    OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+3, halfHV, stride, SIZE, SIZE);\
2052 2053
}\

2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067
#define H264_MC_4816(MMX)\
H264_MC(put_, 4, MMX, 8)\
H264_MC(put_, 8, MMX, 8)\
H264_MC(put_, 16,MMX, 8)\
H264_MC(avg_, 4, MMX, 8)\
H264_MC(avg_, 8, MMX, 8)\
H264_MC(avg_, 16,MMX, 8)\

#define H264_MC_816(QPEL, XMM)\
QPEL(put_, 8, XMM, 16)\
QPEL(put_, 16,XMM, 16)\
QPEL(avg_, 8, XMM, 16)\
QPEL(avg_, 16,XMM, 16)\

2068 2069

#define AVG_3DNOW_OP(a,b,temp, size) \
2070 2071 2072
"mov" #size " " #b ", " #temp "   \n\t"\
"pavgusb " #temp ", " #a "        \n\t"\
"mov" #size " " #a ", " #b "      \n\t"
2073
#define AVG_MMX2_OP(a,b,temp, size) \
2074 2075 2076
"mov" #size " " #b ", " #temp "   \n\t"\
"pavgb " #temp ", " #a "          \n\t"\
"mov" #size " " #a ", " #b "      \n\t"
2077

2078
#define PAVGB "pavgusb"
2079 2080
QPEL_H264(put_,       PUT_OP, 3dnow)
QPEL_H264(avg_, AVG_3DNOW_OP, 3dnow)
2081 2082
#undef PAVGB
#define PAVGB "pavgb"
2083 2084
QPEL_H264(put_,       PUT_OP, mmx2)
QPEL_H264(avg_,  AVG_MMX2_OP, mmx2)
2085 2086 2087 2088
QPEL_H264_V_XMM(put_,       PUT_OP, sse2)
QPEL_H264_V_XMM(avg_,  AVG_MMX2_OP, sse2)
QPEL_H264_HV_XMM(put_,       PUT_OP, sse2)
QPEL_H264_HV_XMM(avg_,  AVG_MMX2_OP, sse2)
2089
#if HAVE_SSSE3
2090 2091 2092 2093 2094 2095
QPEL_H264_H_XMM(put_,       PUT_OP, ssse3)
QPEL_H264_H_XMM(avg_,  AVG_MMX2_OP, ssse3)
QPEL_H264_HV2_XMM(put_,       PUT_OP, ssse3)
QPEL_H264_HV2_XMM(avg_,  AVG_MMX2_OP, ssse3)
QPEL_H264_HV_XMM(put_,       PUT_OP, ssse3)
QPEL_H264_HV_XMM(avg_,  AVG_MMX2_OP, ssse3)
2096
#endif
2097
#undef PAVGB
2098

2099 2100 2101 2102
H264_MC_4816(3dnow)
H264_MC_4816(mmx2)
H264_MC_816(H264_MC_V, sse2)
H264_MC_816(H264_MC_HV, sse2)
2103
#if HAVE_SSSE3
2104 2105
H264_MC_816(H264_MC_H, ssse3)
H264_MC_816(H264_MC_HV, ssse3)
2106 2107
#endif

2108
/* rnd interleaved with rnd div 8, use p+1 to access rnd div 8 */
2109
DECLARE_ALIGNED(8, static const uint64_t, h264_rnd_reg)[4] = {
2110 2111
    0x0020002000200020ULL, 0x0004000400040004ULL, 0x001C001C001C001CULL, 0x0003000300030003ULL
};
2112 2113

#define H264_CHROMA_OP(S,D)
2114
#define H264_CHROMA_OP4(S,D,T)
2115 2116
#define H264_CHROMA_MC8_TMPL put_h264_chroma_generic_mc8_mmx
#define H264_CHROMA_MC4_TMPL put_h264_chroma_generic_mc4_mmx
2117
#define H264_CHROMA_MC2_TMPL put_h264_chroma_mc2_mmx2
2118
#define H264_CHROMA_MC8_MV0 put_pixels8_mmx
2119
#include "dsputil_h264_template_mmx.c"
2120 2121 2122

static void put_h264_chroma_mc8_mmx_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
2123
    put_h264_chroma_generic_mc8_mmx(dst, src, stride, h, x, y, h264_rnd_reg);
2124
}
2125
static void put_vc1_chroma_mc8_mmx_nornd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
2126
{
2127 2128 2129 2130 2131
    put_h264_chroma_generic_mc8_mmx(dst, src, stride, h, x, y, h264_rnd_reg+2);
}
static void put_h264_chroma_mc4_mmx(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
    put_h264_chroma_generic_mc4_mmx(dst, src, stride, h, x, y, h264_rnd_reg);
2132 2133
}

2134
#undef H264_CHROMA_OP
2135
#undef H264_CHROMA_OP4
2136
#undef H264_CHROMA_MC8_TMPL
2137
#undef H264_CHROMA_MC4_TMPL
2138
#undef H264_CHROMA_MC2_TMPL
2139
#undef H264_CHROMA_MC8_MV0
2140 2141

#define H264_CHROMA_OP(S,D) "pavgb " #S ", " #D " \n\t"
2142 2143
#define H264_CHROMA_OP4(S,D,T) "movd  " #S ", " #T " \n\t"\
                               "pavgb " #T ", " #D " \n\t"
2144 2145
#define H264_CHROMA_MC8_TMPL avg_h264_chroma_generic_mc8_mmx2
#define H264_CHROMA_MC4_TMPL avg_h264_chroma_generic_mc4_mmx2
2146
#define H264_CHROMA_MC2_TMPL avg_h264_chroma_mc2_mmx2
2147
#define H264_CHROMA_MC8_MV0 avg_pixels8_mmx2
2148
#include "dsputil_h264_template_mmx.c"
2149 2150
static void avg_h264_chroma_mc8_mmx2_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
2151 2152
    avg_h264_chroma_generic_mc8_mmx2(dst, src, stride, h, x, y, h264_rnd_reg);
}
2153 2154 2155 2156
static void avg_vc1_chroma_mc8_mmx2_nornd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
    avg_h264_chroma_generic_mc8_mmx2(dst, src, stride, h, x, y, h264_rnd_reg+2);
}
2157 2158 2159
static void avg_h264_chroma_mc4_mmx2(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
    avg_h264_chroma_generic_mc4_mmx2(dst, src, stride, h, x, y, h264_rnd_reg);
2160
}
2161
#undef H264_CHROMA_OP
2162
#undef H264_CHROMA_OP4
2163
#undef H264_CHROMA_MC8_TMPL
2164
#undef H264_CHROMA_MC4_TMPL
2165
#undef H264_CHROMA_MC2_TMPL
2166
#undef H264_CHROMA_MC8_MV0
2167 2168

#define H264_CHROMA_OP(S,D) "pavgusb " #S ", " #D " \n\t"
2169 2170
#define H264_CHROMA_OP4(S,D,T) "movd " #S ", " #T " \n\t"\
                               "pavgusb " #T ", " #D " \n\t"
2171 2172
#define H264_CHROMA_MC8_TMPL avg_h264_chroma_generic_mc8_3dnow
#define H264_CHROMA_MC4_TMPL avg_h264_chroma_generic_mc4_3dnow
2173
#define H264_CHROMA_MC8_MV0 avg_pixels8_3dnow
2174
#include "dsputil_h264_template_mmx.c"
2175 2176
static void avg_h264_chroma_mc8_3dnow_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
2177 2178 2179 2180 2181
    avg_h264_chroma_generic_mc8_3dnow(dst, src, stride, h, x, y, h264_rnd_reg);
}
static void avg_h264_chroma_mc4_3dnow(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
    avg_h264_chroma_generic_mc4_3dnow(dst, src, stride, h, x, y, h264_rnd_reg);
2182
}
2183
#undef H264_CHROMA_OP
2184
#undef H264_CHROMA_OP4
2185
#undef H264_CHROMA_MC8_TMPL
2186 2187
#undef H264_CHROMA_MC4_TMPL
#undef H264_CHROMA_MC8_MV0
2188

2189
#if HAVE_SSSE3
Loren Merritt's avatar
Loren Merritt committed
2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200
#define AVG_OP(X)
#undef H264_CHROMA_MC8_TMPL
#undef H264_CHROMA_MC4_TMPL
#define H264_CHROMA_MC8_TMPL put_h264_chroma_mc8_ssse3
#define H264_CHROMA_MC4_TMPL put_h264_chroma_mc4_ssse3
#define H264_CHROMA_MC8_MV0 put_pixels8_mmx
#include "dsputil_h264_template_ssse3.c"
static void put_h264_chroma_mc8_ssse3_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
    put_h264_chroma_mc8_ssse3(dst, src, stride, h, x, y, 1);
}
2201
static void put_vc1_chroma_mc8_ssse3_nornd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
Loren Merritt's avatar
Loren Merritt committed
2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218
{
    put_h264_chroma_mc8_ssse3(dst, src, stride, h, x, y, 0);
}

#undef AVG_OP
#undef H264_CHROMA_MC8_TMPL
#undef H264_CHROMA_MC4_TMPL
#undef H264_CHROMA_MC8_MV0
#define AVG_OP(X) X
#define H264_CHROMA_MC8_TMPL avg_h264_chroma_mc8_ssse3
#define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_ssse3
#define H264_CHROMA_MC8_MV0 avg_pixels8_mmx2
#include "dsputil_h264_template_ssse3.c"
static void avg_h264_chroma_mc8_ssse3_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
    avg_h264_chroma_mc8_ssse3(dst, src, stride, h, x, y, 1);
}
2219 2220 2221 2222
static void avg_vc1_chroma_mc8_ssse3_nornd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
{
    avg_h264_chroma_mc8_ssse3(dst, src, stride, h, x, y, 0);
}
Loren Merritt's avatar
Loren Merritt committed
2223 2224 2225 2226 2227 2228
#undef AVG_OP
#undef H264_CHROMA_MC8_TMPL
#undef H264_CHROMA_MC4_TMPL
#undef H264_CHROMA_MC8_MV0
#endif

2229 2230 2231 2232 2233 2234 2235 2236
/***********************************/
/* weighted prediction */

static inline void ff_h264_weight_WxH_mmx2(uint8_t *dst, int stride, int log2_denom, int weight, int offset, int w, int h)
{
    int x, y;
    offset <<= log2_denom;
    offset += (1 << log2_denom) >> 1;
2237
    __asm__ volatile(
2238 2239 2240 2241 2242 2243 2244 2245 2246 2247
        "movd    %0, %%mm4        \n\t"
        "movd    %1, %%mm5        \n\t"
        "movd    %2, %%mm6        \n\t"
        "pshufw  $0, %%mm4, %%mm4 \n\t"
        "pshufw  $0, %%mm5, %%mm5 \n\t"
        "pxor    %%mm7, %%mm7     \n\t"
        :: "g"(weight), "g"(offset), "g"(log2_denom)
    );
    for(y=0; y<h; y+=2){
        for(x=0; x<w; x+=4){
2248
            __asm__ volatile(
2249 2250 2251 2252 2253 2254
                "movd      %0,    %%mm0 \n\t"
                "movd      %1,    %%mm1 \n\t"
                "punpcklbw %%mm7, %%mm0 \n\t"
                "punpcklbw %%mm7, %%mm1 \n\t"
                "pmullw    %%mm4, %%mm0 \n\t"
                "pmullw    %%mm4, %%mm1 \n\t"
2255 2256
                "paddsw    %%mm5, %%mm0 \n\t"
                "paddsw    %%mm5, %%mm1 \n\t"
2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270
                "psraw     %%mm6, %%mm0 \n\t"
                "psraw     %%mm6, %%mm1 \n\t"
                "packuswb  %%mm7, %%mm0 \n\t"
                "packuswb  %%mm7, %%mm1 \n\t"
                "movd      %%mm0, %0    \n\t"
                "movd      %%mm1, %1    \n\t"
                : "+m"(*(uint32_t*)(dst+x)),
                  "+m"(*(uint32_t*)(dst+x+stride))
            );
        }
        dst += 2*stride;
    }
}

Loren Merritt's avatar
Loren Merritt committed
2271
static inline void ff_h264_biweight_WxH_mmx2(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset, int w, int h)
2272 2273
{
    int x, y;
Loren Merritt's avatar
Loren Merritt committed
2274
    offset = ((offset + 1) | 1) << log2_denom;
2275
    __asm__ volatile(
2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287
        "movd    %0, %%mm3        \n\t"
        "movd    %1, %%mm4        \n\t"
        "movd    %2, %%mm5        \n\t"
        "movd    %3, %%mm6        \n\t"
        "pshufw  $0, %%mm3, %%mm3 \n\t"
        "pshufw  $0, %%mm4, %%mm4 \n\t"
        "pshufw  $0, %%mm5, %%mm5 \n\t"
        "pxor    %%mm7, %%mm7     \n\t"
        :: "g"(weightd), "g"(weights), "g"(offset), "g"(log2_denom+1)
    );
    for(y=0; y<h; y++){
        for(x=0; x<w; x+=4){
2288
            __asm__ volatile(
2289 2290 2291 2292 2293 2294
                "movd      %0,    %%mm0 \n\t"
                "movd      %1,    %%mm1 \n\t"
                "punpcklbw %%mm7, %%mm0 \n\t"
                "punpcklbw %%mm7, %%mm1 \n\t"
                "pmullw    %%mm3, %%mm0 \n\t"
                "pmullw    %%mm4, %%mm1 \n\t"
2295 2296
                "paddsw    %%mm1, %%mm0 \n\t"
                "paddsw    %%mm5, %%mm0 \n\t"
2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309
                "psraw     %%mm6, %%mm0 \n\t"
                "packuswb  %%mm0, %%mm0 \n\t"
                "movd      %%mm0, %0    \n\t"
                : "+m"(*(uint32_t*)(dst+x))
                :  "m"(*(uint32_t*)(src+x))
            );
        }
        src += stride;
        dst += stride;
    }
}

#define H264_WEIGHT(W,H) \
Loren Merritt's avatar
Loren Merritt committed
2310 2311
static void ff_h264_biweight_ ## W ## x ## H ## _mmx2(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset){ \
    ff_h264_biweight_WxH_mmx2(dst, src, stride, log2_denom, weightd, weights, offset, W, H); \
2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325
} \
static void ff_h264_weight_ ## W ## x ## H ## _mmx2(uint8_t *dst, int stride, int log2_denom, int weight, int offset){ \
    ff_h264_weight_WxH_mmx2(dst, stride, log2_denom, weight, offset, W, H); \
}

H264_WEIGHT(16,16)
H264_WEIGHT(16, 8)
H264_WEIGHT( 8,16)
H264_WEIGHT( 8, 8)
H264_WEIGHT( 8, 4)
H264_WEIGHT( 4, 8)
H264_WEIGHT( 4, 4)
H264_WEIGHT( 4, 2)