dsputil_mmx.c 19.1 KB
Newer Older
Fabrice Bellard's avatar
Fabrice Bellard committed
1 2
/*
 * MMX optimized DSP utils
3
 * Copyright (c) 2000, 2001 Fabrice Bellard
4
 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
Fabrice Bellard's avatar
Fabrice Bellard committed
5
 *
6 7
 * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
 *
8
 * This file is part of Libav.
9
 *
10
 * Libav is free software; you can redistribute it and/or
11 12
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
13
 * version 2.1 of the License, or (at your option) any later version.
Fabrice Bellard's avatar
Fabrice Bellard committed
14
 *
15
 * Libav is distributed in the hope that it will be useful,
Fabrice Bellard's avatar
Fabrice Bellard committed
16
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 18
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
Fabrice Bellard's avatar
Fabrice Bellard committed
19
 *
20
 * You should have received a copy of the GNU Lesser General Public
21
 * License along with Libav; if not, write to the Free Software
22
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Fabrice Bellard's avatar
Fabrice Bellard committed
23 24
 */

25
#include "config.h"
26
#include "libavutil/cpu.h"
27
#include "libavutil/x86/asm.h"
28
#include "constants.h"
29
#include "dsputil_x86.h"
30

31 32
#if HAVE_INLINE_ASM

Diego Biurrun's avatar
Diego Biurrun committed
33
void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
34
                               int line_size)
Fabrice Bellard's avatar
Fabrice Bellard committed
35
{
Diego Biurrun's avatar
Diego Biurrun committed
36
    const int16_t *p;
37
    uint8_t *pix;
Fabrice Bellard's avatar
Fabrice Bellard committed
38 39

    /* read the pixels */
40
    p   = block;
Fabrice Bellard's avatar
Fabrice Bellard committed
41
    pix = pixels;
42
    /* unrolled loop */
43
    __asm__ volatile (
44 45 46 47 48 49 50 51
        "movq      (%3), %%mm0          \n\t"
        "movq     8(%3), %%mm1          \n\t"
        "movq    16(%3), %%mm2          \n\t"
        "movq    24(%3), %%mm3          \n\t"
        "movq    32(%3), %%mm4          \n\t"
        "movq    40(%3), %%mm5          \n\t"
        "movq    48(%3), %%mm6          \n\t"
        "movq    56(%3), %%mm7          \n\t"
52 53 54 55 56 57 58 59 60
        "packuswb %%mm1, %%mm0          \n\t"
        "packuswb %%mm3, %%mm2          \n\t"
        "packuswb %%mm5, %%mm4          \n\t"
        "packuswb %%mm7, %%mm6          \n\t"
        "movq     %%mm0, (%0)           \n\t"
        "movq     %%mm2, (%0, %1)       \n\t"
        "movq     %%mm4, (%0, %1, 2)    \n\t"
        "movq     %%mm6, (%0, %2)       \n\t"
        :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
61
           "r"(p)
62
        : "memory");
63 64
    pix += line_size * 4;
    p   += 32;
65 66 67 68

    // if here would be an exact copy of the code above
    // compiler would generate some very strange code
    // thus using "r"
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
    __asm__ volatile (
        "movq       (%3), %%mm0         \n\t"
        "movq      8(%3), %%mm1         \n\t"
        "movq     16(%3), %%mm2         \n\t"
        "movq     24(%3), %%mm3         \n\t"
        "movq     32(%3), %%mm4         \n\t"
        "movq     40(%3), %%mm5         \n\t"
        "movq     48(%3), %%mm6         \n\t"
        "movq     56(%3), %%mm7         \n\t"
        "packuswb  %%mm1, %%mm0         \n\t"
        "packuswb  %%mm3, %%mm2         \n\t"
        "packuswb  %%mm5, %%mm4         \n\t"
        "packuswb  %%mm7, %%mm6         \n\t"
        "movq      %%mm0, (%0)          \n\t"
        "movq      %%mm2, (%0, %1)      \n\t"
        "movq      %%mm4, (%0, %1, 2)   \n\t"
        "movq      %%mm6, (%0, %2)      \n\t"
        :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
        : "memory");
Fabrice Bellard's avatar
Fabrice Bellard committed
88 89
}

90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
#define put_signed_pixels_clamped_mmx_half(off)             \
    "movq          "#off"(%2), %%mm1        \n\t"           \
    "movq     16 + "#off"(%2), %%mm2        \n\t"           \
    "movq     32 + "#off"(%2), %%mm3        \n\t"           \
    "movq     48 + "#off"(%2), %%mm4        \n\t"           \
    "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
    "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
    "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
    "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
    "paddb              %%mm0, %%mm1        \n\t"           \
    "paddb              %%mm0, %%mm2        \n\t"           \
    "paddb              %%mm0, %%mm3        \n\t"           \
    "paddb              %%mm0, %%mm4        \n\t"           \
    "movq               %%mm1, (%0)         \n\t"           \
    "movq               %%mm2, (%0, %3)     \n\t"           \
    "movq               %%mm3, (%0, %3, 2)  \n\t"           \
    "movq               %%mm4, (%0, %1)     \n\t"
107

Diego Biurrun's avatar
Diego Biurrun committed
108
void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
109
                                      int line_size)
110
{
111
    x86_reg line_skip = line_size;
112
    x86_reg line_skip3;
113 114

    __asm__ volatile (
115 116 117 118 119 120 121 122
        "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
        "lea         (%3, %3, 2), %1        \n\t"
        put_signed_pixels_clamped_mmx_half(0)
        "lea         (%0, %3, 4), %0        \n\t"
        put_signed_pixels_clamped_mmx_half(64)
        : "+&r"(pixels), "=&r"(line_skip3)
        : "r"(block), "r"(line_skip)
        : "memory");
123 124
}

Diego Biurrun's avatar
Diego Biurrun committed
125
void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
126
                               int line_size)
Fabrice Bellard's avatar
Fabrice Bellard committed
127
{
Diego Biurrun's avatar
Diego Biurrun committed
128
    const int16_t *p;
129
    uint8_t *pix;
Fabrice Bellard's avatar
Fabrice Bellard committed
130 131 132
    int i;

    /* read the pixels */
133
    p   = block;
Fabrice Bellard's avatar
Fabrice Bellard committed
134
    pix = pixels;
135 136
    MOVQ_ZERO(mm7);
    i = 4;
137
    do {
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
        __asm__ volatile (
            "movq        (%2), %%mm0    \n\t"
            "movq       8(%2), %%mm1    \n\t"
            "movq      16(%2), %%mm2    \n\t"
            "movq      24(%2), %%mm3    \n\t"
            "movq          %0, %%mm4    \n\t"
            "movq          %1, %%mm6    \n\t"
            "movq       %%mm4, %%mm5    \n\t"
            "punpcklbw  %%mm7, %%mm4    \n\t"
            "punpckhbw  %%mm7, %%mm5    \n\t"
            "paddsw     %%mm4, %%mm0    \n\t"
            "paddsw     %%mm5, %%mm1    \n\t"
            "movq       %%mm6, %%mm5    \n\t"
            "punpcklbw  %%mm7, %%mm6    \n\t"
            "punpckhbw  %%mm7, %%mm5    \n\t"
            "paddsw     %%mm6, %%mm2    \n\t"
            "paddsw     %%mm5, %%mm3    \n\t"
            "packuswb   %%mm1, %%mm0    \n\t"
            "packuswb   %%mm3, %%mm2    \n\t"
            "movq       %%mm0, %0       \n\t"
            "movq       %%mm2, %1       \n\t"
            : "+m"(*pix), "+m"(*(pix + line_size))
            : "r"(p)
            : "memory");
162 163
        pix += line_size * 2;
        p   += 16;
164
    } while (--i);
Fabrice Bellard's avatar
Fabrice Bellard committed
165 166
}

167
#define CLEAR_BLOCKS(name, n)                           \
168
void name(int16_t *blocks)                              \
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183
{                                                       \
    __asm__ volatile (                                  \
        "pxor %%mm7, %%mm7              \n\t"           \
        "mov     %1,        %%"REG_a"   \n\t"           \
        "1:                             \n\t"           \
        "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
        "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
        "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
        "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
        "add    $32, %%"REG_a"          \n\t"           \
        "js      1b                     \n\t"           \
        :: "r"(((uint8_t *)blocks) + 128 * n),          \
           "i"(-128 * n)                                \
        : "%"REG_a                                      \
        );                                              \
Loren Merritt's avatar
Loren Merritt committed
184
}
185 186
CLEAR_BLOCKS(ff_clear_blocks_mmx, 6)
CLEAR_BLOCKS(ff_clear_block_mmx, 1)
Loren Merritt's avatar
Loren Merritt committed
187

188
void ff_clear_block_sse(int16_t *block)
189
{
190 191 192 193 194 195 196 197 198 199
    __asm__ volatile (
        "xorps  %%xmm0, %%xmm0          \n"
        "movaps %%xmm0,    (%0)         \n"
        "movaps %%xmm0,  16(%0)         \n"
        "movaps %%xmm0,  32(%0)         \n"
        "movaps %%xmm0,  48(%0)         \n"
        "movaps %%xmm0,  64(%0)         \n"
        "movaps %%xmm0,  80(%0)         \n"
        "movaps %%xmm0,  96(%0)         \n"
        "movaps %%xmm0, 112(%0)         \n"
Loren Merritt's avatar
Loren Merritt committed
200 201 202
        :: "r"(block)
        : "memory"
    );
203 204
}

205
void ff_clear_blocks_sse(int16_t *blocks)
206
{
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222
    __asm__ volatile (
        "xorps  %%xmm0, %%xmm0              \n"
        "mov        %1,         %%"REG_a"   \n"
        "1:                                 \n"
        "movaps %%xmm0,    (%0, %%"REG_a")  \n"
        "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
        "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
        "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
        "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
        "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
        "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
        "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
        "add      $128,         %%"REG_a"   \n"
        "js         1b                      \n"
        :: "r"(((uint8_t *)blocks) + 128 * 6),
           "i"(-128 * 6)
223 224 225 226
        : "%"REG_a
    );
}

227
void ff_add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
228 229
{
    x86_reg i = 0;
230 231
    __asm__ volatile (
        "jmp          2f                \n\t"
232
        "1:                             \n\t"
233 234 235 236 237 238 239 240 241
        "movq   (%1, %0), %%mm0         \n\t"
        "movq   (%2, %0), %%mm1         \n\t"
        "paddb     %%mm0, %%mm1         \n\t"
        "movq      %%mm1, (%2, %0)      \n\t"
        "movq  8(%1, %0), %%mm0         \n\t"
        "movq  8(%2, %0), %%mm1         \n\t"
        "paddb     %%mm0, %%mm1         \n\t"
        "movq      %%mm1, 8(%2, %0)     \n\t"
        "add         $16, %0            \n\t"
242
        "2:                             \n\t"
243 244 245 246
        "cmp          %3, %0            \n\t"
        "js           1b                \n\t"
        : "+r"(i)
        : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
Michael Niedermayer's avatar
Michael Niedermayer committed
247
    );
248 249
    for ( ; i < w; i++)
        dst[i + 0] += src[i + 0];
Michael Niedermayer's avatar
Michael Niedermayer committed
250 251
}

252 253
/* Draw the edges of width 'w' of an image of size width, height
 * this MMX version can only handle w == 8 || w == 16. */
254 255
void ff_draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
                       int w, int h, int sides)
256 257 258 259 260 261 262
{
    uint8_t *ptr, *last_line;
    int i;

    last_line = buf + (height - 1) * wrap;
    /* left and right */
    ptr = buf;
263
    if (w == 8) {
264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281
        __asm__ volatile (
            "1:                             \n\t"
            "movd            (%0), %%mm0    \n\t"
            "punpcklbw      %%mm0, %%mm0    \n\t"
            "punpcklwd      %%mm0, %%mm0    \n\t"
            "punpckldq      %%mm0, %%mm0    \n\t"
            "movq           %%mm0, -8(%0)   \n\t"
            "movq      -8(%0, %2), %%mm1    \n\t"
            "punpckhbw      %%mm1, %%mm1    \n\t"
            "punpckhwd      %%mm1, %%mm1    \n\t"
            "punpckhdq      %%mm1, %%mm1    \n\t"
            "movq           %%mm1, (%0, %2) \n\t"
            "add               %1, %0       \n\t"
            "cmp               %3, %0       \n\t"
            "jb                1b           \n\t"
            : "+r"(ptr)
            : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
            );
282
    } else {
283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
        __asm__ volatile (
            "1:                                 \n\t"
            "movd            (%0), %%mm0        \n\t"
            "punpcklbw      %%mm0, %%mm0        \n\t"
            "punpcklwd      %%mm0, %%mm0        \n\t"
            "punpckldq      %%mm0, %%mm0        \n\t"
            "movq           %%mm0, -8(%0)       \n\t"
            "movq           %%mm0, -16(%0)      \n\t"
            "movq      -8(%0, %2), %%mm1        \n\t"
            "punpckhbw      %%mm1, %%mm1        \n\t"
            "punpckhwd      %%mm1, %%mm1        \n\t"
            "punpckhdq      %%mm1, %%mm1        \n\t"
            "movq           %%mm1,  (%0, %2)    \n\t"
            "movq           %%mm1, 8(%0, %2)    \n\t"
            "add               %1, %0           \n\t"
            "cmp               %3, %0           \n\t"
            "jb                1b               \n\t"
            : "+r"(ptr)
            : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
            );
303 304
    }

305
    /* top and bottom (and hopefully also the corners) */
306 307 308
    if (sides & EDGE_TOP) {
        for (i = 0; i < h; i += 4) {
            ptr = buf - (i + 1) * wrap - w;
309 310 311 312 313 314 315 316 317 318 319 320 321 322
            __asm__ volatile (
                "1:                             \n\t"
                "movq (%1, %0), %%mm0           \n\t"
                "movq    %%mm0, (%0)            \n\t"
                "movq    %%mm0, (%0, %2)        \n\t"
                "movq    %%mm0, (%0, %2, 2)     \n\t"
                "movq    %%mm0, (%0, %3)        \n\t"
                "add        $8, %0              \n\t"
                "cmp        %4, %0              \n\t"
                "jb         1b                  \n\t"
                : "+r"(ptr)
                : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
                  "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
                );
323 324 325
        }
    }

326 327 328
    if (sides & EDGE_BOTTOM) {
        for (i = 0; i < h; i += 4) {
            ptr = last_line + (i + 1) * wrap - w;
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
            __asm__ volatile (
                "1:                             \n\t"
                "movq (%1, %0), %%mm0           \n\t"
                "movq    %%mm0, (%0)            \n\t"
                "movq    %%mm0, (%0, %2)        \n\t"
                "movq    %%mm0, (%0, %2, 2)     \n\t"
                "movq    %%mm0, (%0, %3)        \n\t"
                "add        $8, %0              \n\t"
                "cmp        %4, %0              \n\t"
                "jb         1b                  \n\t"
                : "+r"(ptr)
                : "r"((x86_reg)last_line - (x86_reg)ptr - w),
                  "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
                  "r"(ptr + width + 2 * w)
                );
344
        }
345 346
    }
}
Michael Niedermayer's avatar
Michael Niedermayer committed
347

348 349 350 351
void ff_gmc_mmx(uint8_t *dst, uint8_t *src,
                int stride, int h, int ox, int oy,
                int dxx, int dxy, int dyx, int dyy,
                int shift, int r, int width, int height)
352
{
353 354 355 356 357 358 359 360 361 362 363 364 365
    const int w    = 8;
    const int ix   = ox  >> (16 + shift);
    const int iy   = oy  >> (16 + shift);
    const int oxs  = ox  >> 4;
    const int oys  = oy  >> 4;
    const int dxxs = dxx >> 4;
    const int dxys = dxy >> 4;
    const int dyxs = dyx >> 4;
    const int dyys = dyy >> 4;
    const uint16_t r4[4]   = { r, r, r, r };
    const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
    const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
    const uint64_t shift2 = 2 * shift;
366 367
    int x, y;

368 369 370 371 372 373 374
    const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
    const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
    const int dxh = dxy * (h - 1);
    const int dyw = dyx * (w - 1);
    if ( // non-constant fullpel offset (3% of blocks)
        ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
         (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
375
        // uses more than 16 bits of subpel mv (only at huge resolution)
376 377 378
        || (dxx | dxy | dyx | dyy) & 15 ||
        (unsigned)ix >= width  - w ||
        (unsigned)iy >= height - h) {
379
        // FIXME could still use mmx for some of the rows
380 381
        ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
                 shift, r, width, height);
382 383 384
        return;
    }

385
    src += ix + iy * stride;
Loren Merritt's avatar
Loren Merritt committed
386

387 388 389 390 391
    __asm__ volatile (
        "movd         %0, %%mm6         \n\t"
        "pxor      %%mm7, %%mm7         \n\t"
        "punpcklwd %%mm6, %%mm6         \n\t"
        "punpcklwd %%mm6, %%mm6         \n\t"
Loren Merritt's avatar
Loren Merritt committed
392 393
        :: "r"(1<<shift)
    );
394

395 396 397 398 399 400 401 402 403 404 405
    for (x = 0; x < w; x += 4) {
        uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
                            oxs - dxys + dxxs * (x + 1),
                            oxs - dxys + dxxs * (x + 2),
                            oxs - dxys + dxxs * (x + 3) };
        uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
                            oys - dyys + dyxs * (x + 1),
                            oys - dyys + dyxs * (x + 2),
                            oys - dyys + dyxs * (x + 3) };

        for (y = 0; y < h; y++) {
406 407 408 409 410 411 412 413 414
            __asm__ volatile (
                "movq      %0, %%mm4    \n\t"
                "movq      %1, %%mm5    \n\t"
                "paddw     %2, %%mm4    \n\t"
                "paddw     %3, %%mm5    \n\t"
                "movq   %%mm4, %0       \n\t"
                "movq   %%mm5, %1       \n\t"
                "psrlw    $12, %%mm4    \n\t"
                "psrlw    $12, %%mm5    \n\t"
415 416 417 418
                : "+m"(*dx4), "+m"(*dy4)
                : "m"(*dxy4), "m"(*dyy4)
            );

419 420 421 422 423 424 425 426 427 428 429 430 431 432
            __asm__ volatile (
                "movq      %%mm6, %%mm2 \n\t"
                "movq      %%mm6, %%mm1 \n\t"
                "psubw     %%mm4, %%mm2 \n\t"
                "psubw     %%mm5, %%mm1 \n\t"
                "movq      %%mm2, %%mm0 \n\t"
                "movq      %%mm4, %%mm3 \n\t"
                "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
                "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
                "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
                "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)

                "movd         %4, %%mm5 \n\t"
                "movd         %3, %%mm4 \n\t"
433 434
                "punpcklbw %%mm7, %%mm5 \n\t"
                "punpcklbw %%mm7, %%mm4 \n\t"
435 436
                "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
                "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
437

438 439
                "movd         %2, %%mm5 \n\t"
                "movd         %1, %%mm4 \n\t"
440 441
                "punpcklbw %%mm7, %%mm5 \n\t"
                "punpcklbw %%mm7, %%mm4 \n\t"
442 443 444 445 446 447 448 449 450 451 452 453
                "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
                "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
                "paddw        %5, %%mm1 \n\t"
                "paddw     %%mm3, %%mm2 \n\t"
                "paddw     %%mm1, %%mm0 \n\t"
                "paddw     %%mm2, %%mm0 \n\t"

                "psrlw        %6, %%mm0 \n\t"
                "packuswb  %%mm0, %%mm0 \n\t"
                "movd      %%mm0, %0    \n\t"

                : "=m"(dst[x + y * stride])
454
                : "m"(src[0]), "m"(src[1]),
455
                  "m"(src[stride]), "m"(src[stride + 1]),
456 457 458 459
                  "m"(*r4), "m"(shift2)
            );
            src += stride;
        }
460
        src += 4 - h * stride;
461 462
    }
}
463

464 465
void ff_vector_clipf_sse(float *dst, const float *src,
                         float min, float max, int len)
466
{
467
    x86_reg i = (len - 16) * 4;
468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
    __asm__ volatile (
        "movss          %3, %%xmm4      \n\t"
        "movss          %4, %%xmm5      \n\t"
        "shufps $0, %%xmm4, %%xmm4      \n\t"
        "shufps $0, %%xmm5, %%xmm5      \n\t"
        "1:                             \n\t"
        "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
        "movaps 16(%2, %0), %%xmm1      \n\t"
        "movaps 32(%2, %0), %%xmm2      \n\t"
        "movaps 48(%2, %0), %%xmm3      \n\t"
        "maxps      %%xmm4, %%xmm0      \n\t"
        "maxps      %%xmm4, %%xmm1      \n\t"
        "maxps      %%xmm4, %%xmm2      \n\t"
        "maxps      %%xmm4, %%xmm3      \n\t"
        "minps      %%xmm5, %%xmm0      \n\t"
        "minps      %%xmm5, %%xmm1      \n\t"
        "minps      %%xmm5, %%xmm2      \n\t"
        "minps      %%xmm5, %%xmm3      \n\t"
        "movaps     %%xmm0,   (%1, %0)  \n\t"
        "movaps     %%xmm1, 16(%1, %0)  \n\t"
        "movaps     %%xmm2, 32(%1, %0)  \n\t"
        "movaps     %%xmm3, 48(%1, %0)  \n\t"
        "sub           $64, %0          \n\t"
        "jge            1b              \n\t"
        : "+&r"(i)
        : "r"(dst), "r"(src), "m"(min), "m"(max)
        : "memory"
495 496 497
    );
}

498
#endif /* HAVE_INLINE_ASM */