scale_gcc.cc 55.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
/*
 *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS. All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include "libyuv/row.h"
12
#include "libyuv/scale_row.h"
13 14 15 16 17 18 19

#ifdef __cplusplus
namespace libyuv {
extern "C" {
#endif

// This module is for GCC x86 and x64.
Frank Barchard's avatar
Frank Barchard committed
20 21
#if !defined(LIBYUV_DISABLE_X86) && \
    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
22 23

// Offsets for source bytes 0 to 9
24 25
static const uvec8 kShuf0 = {0,   1,   3,   4,   5,   7,   8,   9,
                             128, 128, 128, 128, 128, 128, 128, 128};
26 27

// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
28 29
static const uvec8 kShuf1 = {3,   4,   5,   7,   8,   9,   11,  12,
                             128, 128, 128, 128, 128, 128, 128, 128};
30 31

// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
32 33
static const uvec8 kShuf2 = {5,   7,   8,   9,   11,  12,  13,  15,
                             128, 128, 128, 128, 128, 128, 128, 128};
34 35

// Offsets for source bytes 0 to 10
36
static const uvec8 kShuf01 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
37 38

// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
39 40
static const uvec8 kShuf11 = {2, 3, 4, 5,  5,  6,  6,  7,
                              8, 9, 9, 10, 10, 11, 12, 13};
41 42

// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
43 44
static const uvec8 kShuf21 = {5,  6,  6,  7,  8,  9,  9,  10,
                              10, 11, 12, 13, 13, 14, 14, 15};
45 46

// Coefficients for source bytes 0 to 10
47
static const uvec8 kMadd01 = {3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2};
48 49

// Coefficients for source bytes 10 to 21
50
static const uvec8 kMadd11 = {1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1};
51 52

// Coefficients for source bytes 21 to 31
53
static const uvec8 kMadd21 = {2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3};
54 55

// Coefficients for source bytes 21 to 31
56
static const vec16 kRound34 = {2, 2, 2, 2, 2, 2, 2, 2};
57

58 59
static const uvec8 kShuf38a = {0,   3,   6,   8,   11,  14,  128, 128,
                               128, 128, 128, 128, 128, 128, 128, 128};
60

61 62
static const uvec8 kShuf38b = {128, 128, 128, 128, 128, 128, 0,   3,
                               6,   8,   11,  14,  128, 128, 128, 128};
63 64

// Arrange words 0,3,6 into 0,1,2
65 66
static const uvec8 kShufAc = {0,   1,   6,   7,   12,  13,  128, 128,
                              128, 128, 128, 128, 128, 128, 128, 128};
67 68

// Arrange words 0,3,6 into 3,4,5
69 70
static const uvec8 kShufAc3 = {128, 128, 128, 128, 128, 128, 0,   1,
                               6,   7,   12,  13,  128, 128, 128, 128};
71 72

// Scaling values for boxes of 3x3 and 2x3
73 74
static const uvec16 kScaleAc33 = {65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9,
                                  65536 / 9, 65536 / 6, 0,         0};
75 76

// Arrange first value for pixels 0,1,2,3,4,5
77 78
static const uvec8 kShufAb0 = {0,  128, 3,  128, 6,   128, 8,   128,
                               11, 128, 14, 128, 128, 128, 128, 128};
79 80

// Arrange second value for pixels 0,1,2,3,4,5
81 82
static const uvec8 kShufAb1 = {1,  128, 4,  128, 7,   128, 9,   128,
                               12, 128, 15, 128, 128, 128, 128, 128};
83 84

// Arrange third value for pixels 0,1,2,3,4,5
85 86
static const uvec8 kShufAb2 = {2,  128, 5,   128, 128, 128, 10,  128,
                               13, 128, 128, 128, 128, 128, 128, 128};
87 88

// Scaling values for boxes of 3x2 and 2x2
89 90
static const uvec16 kScaleAb2 = {65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3,
                                 65536 / 3, 65536 / 2, 0,         0};
91 92 93 94 95

// GCC versions of row functions are verbatim conversions from Visual C.
// Generated using gcc disassembly on Visual C object file:
// objdump -D yuvscaler.obj >yuvscaler.txt

Frank Barchard's avatar
Frank Barchard committed
96 97 98 99
void ScaleRowDown2_SSSE3(const uint8* src_ptr,
                         ptrdiff_t src_stride,
                         uint8* dst_ptr,
                         int dst_width) {
100
  (void)src_stride;
101
  asm volatile (
102
    LABELALIGN
103 104 105 106 107 108 109 110 111 112 113 114 115 116
  "1:                                          \n"
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "psrlw     $0x8,%%xmm0                     \n"
    "psrlw     $0x8,%%xmm1                     \n"
    "packuswb  %%xmm1,%%xmm0                   \n"
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
    "lea       " MEMLEA(0x10,1) ",%1           \n"
    "sub       $0x10,%2                        \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
117
  :: "memory", "cc", "xmm0", "xmm1"
118 119 120
  );
}

Frank Barchard's avatar
Frank Barchard committed
121 122 123 124
void ScaleRowDown2Linear_SSSE3(const uint8* src_ptr,
                               ptrdiff_t src_stride,
                               uint8* dst_ptr,
                               int dst_width) {
125
  (void)src_stride;
126
  asm volatile (
127 128 129 130
    "pcmpeqb    %%xmm4,%%xmm4                  \n"
    "psrlw      $0xf,%%xmm4                    \n"
    "packuswb   %%xmm4,%%xmm4                  \n"
    "pxor       %%xmm5,%%xmm5                  \n"
131 132

    LABELALIGN
133 134
  "1:                                          \n"
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
135
    "movdqu    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
136
    "lea       " MEMLEA(0x20,0) ",%0           \n"
137 138 139 140 141
    "pmaddubsw  %%xmm4,%%xmm0                  \n"
    "pmaddubsw  %%xmm4,%%xmm1                  \n"
    "pavgw      %%xmm5,%%xmm0                  \n"
    "pavgw      %%xmm5,%%xmm1                  \n"
    "packuswb   %%xmm1,%%xmm0                  \n"
142 143 144 145 146 147 148
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
    "lea       " MEMLEA(0x10,1) ",%1           \n"
    "sub       $0x10,%2                        \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
149
  :: "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
150 151 152
  );
}

Frank Barchard's avatar
Frank Barchard committed
153 154 155 156
void ScaleRowDown2Box_SSSE3(const uint8* src_ptr,
                            ptrdiff_t src_stride,
                            uint8* dst_ptr,
                            int dst_width) {
157
  asm volatile (
158 159 160 161
    "pcmpeqb    %%xmm4,%%xmm4                  \n"
    "psrlw      $0xf,%%xmm4                    \n"
    "packuswb   %%xmm4,%%xmm4                  \n"
    "pxor       %%xmm5,%%xmm5                  \n"
162 163

    LABELALIGN
164 165 166
  "1:                                          \n"
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
167 168
    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
169
    "lea       " MEMLEA(0x20,0) ",%0           \n"
170 171 172 173 174 175 176 177 178 179 180
    "pmaddubsw  %%xmm4,%%xmm0                  \n"
    "pmaddubsw  %%xmm4,%%xmm1                  \n"
    "pmaddubsw  %%xmm4,%%xmm2                  \n"
    "pmaddubsw  %%xmm4,%%xmm3                  \n"
    "paddw      %%xmm2,%%xmm0                  \n"
    "paddw      %%xmm3,%%xmm1                  \n"
    "psrlw      $0x1,%%xmm0                    \n"
    "psrlw      $0x1,%%xmm1                    \n"
    "pavgw      %%xmm5,%%xmm0                  \n"
    "pavgw      %%xmm5,%%xmm1                  \n"
    "packuswb   %%xmm1,%%xmm0                  \n"
181 182 183 184 185 186 187
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
    "lea       " MEMLEA(0x10,1) ",%1           \n"
    "sub       $0x10,%2                        \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
188
  : "r"((intptr_t)(src_stride))   // %3
189 190
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
191 192 193
  );
}

194
#ifdef HAS_SCALEROWDOWN2_AVX2
Frank Barchard's avatar
Frank Barchard committed
195 196 197 198
void ScaleRowDown2_AVX2(const uint8* src_ptr,
                        ptrdiff_t src_stride,
                        uint8* dst_ptr,
                        int dst_width) {
199
  (void)src_stride;
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
  asm volatile (
    LABELALIGN
  "1:                                          \n"
    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
    "lea        " MEMLEA(0x40,0) ",%0          \n"
    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
    "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
    "lea        " MEMLEA(0x20,1) ",%1          \n"
    "sub        $0x20,%2                       \n"
    "jg         1b                             \n"
    "vzeroupper                                \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
  :: "memory", "cc", "xmm0", "xmm1"
  );
}

Frank Barchard's avatar
Frank Barchard committed
222 223 224 225
void ScaleRowDown2Linear_AVX2(const uint8* src_ptr,
                              ptrdiff_t src_stride,
                              uint8* dst_ptr,
                              int dst_width) {
226
  (void)src_stride;
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
  asm volatile (
    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
    "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
    "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"

    LABELALIGN
  "1:                                          \n"
    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
    "vmovdqu    " MEMACCESS2(0x20, 0) ",%%ymm1 \n"
    "lea        " MEMLEA(0x40,0) ",%0          \n"
    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
    "vpavgw     %%ymm5,%%ymm0,%%ymm0           \n"
    "vpavgw     %%ymm5,%%ymm1,%%ymm1           \n"
    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
    "lea        " MEMLEA(0x20,1) ",%1          \n"
    "sub        $0x20,%2                       \n"
    "jg         1b                             \n"
    "vzeroupper                                \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
  :: "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
  );
}

Frank Barchard's avatar
Frank Barchard committed
256 257 258 259
void ScaleRowDown2Box_AVX2(const uint8* src_ptr,
                           ptrdiff_t src_stride,
                           uint8* dst_ptr,
                           int dst_width) {
260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299
  asm volatile (
    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
    "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
    "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"

    LABELALIGN
  "1:                                          \n"
    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
    MEMOPREG(vmovdqu,0x00,0,3,1,ymm2)          //  vmovdqu  (%0,%3,1),%%ymm2
    MEMOPREG(vmovdqu,0x20,0,3,1,ymm3)          //  vmovdqu  0x20(%0,%3,1),%%ymm3
    "lea        " MEMLEA(0x40,0) ",%0          \n"
    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
    "vpsrlw     $0x1,%%ymm0,%%ymm0             \n"
    "vpsrlw     $0x1,%%ymm1,%%ymm1             \n"
    "vpavgw     %%ymm5,%%ymm0,%%ymm0           \n"
    "vpavgw     %%ymm5,%%ymm1,%%ymm1           \n"
    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
    "lea        " MEMLEA(0x20,1) ",%1          \n"
    "sub        $0x20,%2                       \n"
    "jg         1b                             \n"
    "vzeroupper                                \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
  : "r"((intptr_t)(src_stride))   // %3
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
  );
}
#endif  // HAS_SCALEROWDOWN2_AVX2

Frank Barchard's avatar
Frank Barchard committed
300 301 302 303
void ScaleRowDown4_SSSE3(const uint8* src_ptr,
                         ptrdiff_t src_stride,
                         uint8* dst_ptr,
                         int dst_width) {
304
  (void)src_stride;
305 306 307 308
  asm volatile (
    "pcmpeqb   %%xmm5,%%xmm5                   \n"
    "psrld     $0x18,%%xmm5                    \n"
    "pslld     $0x10,%%xmm5                    \n"
309 310

    LABELALIGN
311
  "1:                                          \n"
312 313
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
314 315 316 317 318 319 320 321 322 323 324 325 326
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "pand      %%xmm5,%%xmm0                   \n"
    "pand      %%xmm5,%%xmm1                   \n"
    "packuswb  %%xmm1,%%xmm0                   \n"
    "psrlw     $0x8,%%xmm0                     \n"
    "packuswb  %%xmm0,%%xmm0                   \n"
    "movq      %%xmm0," MEMACCESS(1) "         \n"
    "lea       " MEMLEA(0x8,1) ",%1            \n"
    "sub       $0x8,%2                         \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
327
  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
328 329 330
  );
}

Frank Barchard's avatar
Frank Barchard committed
331 332 333 334
void ScaleRowDown4Box_SSSE3(const uint8* src_ptr,
                            ptrdiff_t src_stride,
                            uint8* dst_ptr,
                            int dst_width) {
335
  intptr_t stridex3;
336
  asm volatile (
337 338 339 340 341
    "pcmpeqb    %%xmm4,%%xmm4                  \n"
    "psrlw      $0xf,%%xmm4                    \n"
    "movdqa     %%xmm4,%%xmm5                  \n"
    "packuswb   %%xmm4,%%xmm4                  \n"
    "psllw      $0x3,%%xmm5                    \n"
342
    "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
343 344

    LABELALIGN
345
  "1:                                          \n"
346 347
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
348 349
    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
350 351 352 353 354 355
    "pmaddubsw  %%xmm4,%%xmm0                  \n"
    "pmaddubsw  %%xmm4,%%xmm1                  \n"
    "pmaddubsw  %%xmm4,%%xmm2                  \n"
    "pmaddubsw  %%xmm4,%%xmm3                  \n"
    "paddw      %%xmm2,%%xmm0                  \n"
    "paddw      %%xmm3,%%xmm1                  \n"
356 357
    MEMOPREG(movdqu,0x00,0,4,2,xmm2)           //  movdqu  (%0,%4,2),%%xmm2
    MEMOPREG(movdqu,0x10,0,4,2,xmm3)           //  movdqu  0x10(%0,%4,2),%%xmm3
358 359 360 361 362 363
    "pmaddubsw  %%xmm4,%%xmm2                  \n"
    "pmaddubsw  %%xmm4,%%xmm3                  \n"
    "paddw      %%xmm2,%%xmm0                  \n"
    "paddw      %%xmm3,%%xmm1                  \n"
    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
364
    "lea       " MEMLEA(0x20,0) ",%0           \n"
365 366 367 368 369 370 371 372
    "pmaddubsw  %%xmm4,%%xmm2                  \n"
    "pmaddubsw  %%xmm4,%%xmm3                  \n"
    "paddw      %%xmm2,%%xmm0                  \n"
    "paddw      %%xmm3,%%xmm1                  \n"
    "phaddw     %%xmm1,%%xmm0                  \n"
    "paddw      %%xmm5,%%xmm0                  \n"
    "psrlw      $0x4,%%xmm0                    \n"
    "packuswb   %%xmm0,%%xmm0                  \n"
373 374 375 376 377 378 379
    "movq      %%xmm0," MEMACCESS(1) "         \n"
    "lea       " MEMLEA(0x8,1) ",%1            \n"
    "sub       $0x8,%2                         \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),     // %0
    "+r"(dst_ptr),     // %1
    "+r"(dst_width),   // %2
380
    "=&r"(stridex3)    // %3
381
  : "r"((intptr_t)(src_stride))    // %4
382
  : "memory", "cc", NACL_R14
383
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
384 385 386
  );
}

387
#ifdef HAS_SCALEROWDOWN4_AVX2
Frank Barchard's avatar
Frank Barchard committed
388 389 390 391
void ScaleRowDown4_AVX2(const uint8* src_ptr,
                        ptrdiff_t src_stride,
                        uint8* dst_ptr,
                        int dst_width) {
392
  (void)src_stride;
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
  asm volatile (
    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
    "vpsrld     $0x18,%%ymm5,%%ymm5            \n"
    "vpslld     $0x10,%%ymm5,%%ymm5            \n"
    LABELALIGN
  "1:                                          \n"
    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
    "lea        " MEMLEA(0x40,0) ",%0          \n"
    "vpand      %%ymm5,%%ymm0,%%ymm0           \n"
    "vpand      %%ymm5,%%ymm1,%%ymm1           \n"
    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vmovdqu    %%xmm0," MEMACCESS(1) "        \n"
    "lea        " MEMLEA(0x10,1) ",%1          \n"
    "sub        $0x10,%2                       \n"
    "jg         1b                             \n"
    "vzeroupper                                \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
  );
}

Frank Barchard's avatar
Frank Barchard committed
421 422 423 424
void ScaleRowDown4Box_AVX2(const uint8* src_ptr,
                           ptrdiff_t src_stride,
                           uint8* dst_ptr,
                           int dst_width) {
425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477
  asm volatile (
    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
    "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
    "vpsllw     $0x3,%%ymm4,%%ymm5             \n"
    "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"

    LABELALIGN
  "1:                                          \n"
    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
    MEMOPREG(vmovdqu,0x00,0,3,1,ymm2)          //  vmovdqu  (%0,%3,1),%%ymm2
    MEMOPREG(vmovdqu,0x20,0,3,1,ymm3)          //  vmovdqu  0x20(%0,%3,1),%%ymm3
    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
    MEMOPREG(vmovdqu,0x00,0,3,2,ymm2)          //  vmovdqu  (%0,%3,2),%%ymm2
    MEMOPREG(vmovdqu,0x20,0,3,2,ymm3)          //  vmovdqu  0x20(%0,%3,2),%%ymm3
    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
    MEMOPREG(vmovdqu,0x00,0,4,1,ymm2)          //  vmovdqu  (%0,%4,1),%%ymm2
    MEMOPREG(vmovdqu,0x20,0,4,1,ymm3)          //  vmovdqu  0x20(%0,%4,1),%%ymm3
    "lea        " MEMLEA(0x40,0) ",%0          \n"
    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
    "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
    "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"
    "vpsrlw     $0x4,%%ymm0,%%ymm0             \n"
    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vmovdqu    %%xmm0," MEMACCESS(1) "        \n"
    "lea        " MEMLEA(0x10,1) ",%1          \n"
    "sub        $0x10,%2                       \n"
    "jg         1b                             \n"
    "vzeroupper                                \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
  : "r"((intptr_t)(src_stride)),  // %3
    "r"((intptr_t)(src_stride * 3))   // %4
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
  );
}
#endif  // HAS_SCALEROWDOWN4_AVX2

Frank Barchard's avatar
Frank Barchard committed
478 479 480 481
void ScaleRowDown34_SSSE3(const uint8* src_ptr,
                          ptrdiff_t src_stride,
                          uint8* dst_ptr,
                          int dst_width) {
482
  (void)src_stride;
Frank Barchard's avatar
Frank Barchard committed
483 484 485 486 487 488 489 490 491
  asm volatile(
      "movdqa    %0,%%xmm3                       \n"
      "movdqa    %1,%%xmm4                       \n"
      "movdqa    %2,%%xmm5                       \n"
      :
      : "m"(kShuf0),  // %0
        "m"(kShuf1),  // %1
        "m"(kShuf2)   // %2
      );
492
  asm volatile (
493
    LABELALIGN
494
  "1:                                          \n"
495 496
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
497 498 499 500 501 502 503 504 505 506 507 508 509 510 511
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "movdqa    %%xmm2,%%xmm1                   \n"
    "palignr   $0x8,%%xmm0,%%xmm1              \n"
    "pshufb    %%xmm3,%%xmm0                   \n"
    "pshufb    %%xmm4,%%xmm1                   \n"
    "pshufb    %%xmm5,%%xmm2                   \n"
    "movq      %%xmm0," MEMACCESS(1) "         \n"
    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
    "lea       " MEMLEA(0x18,1) ",%1           \n"
    "sub       $0x18,%2                        \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),   // %0
    "+r"(dst_ptr),   // %1
    "+r"(dst_width)  // %2
512
  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
513 514 515 516 517
  );
}

void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
                                ptrdiff_t src_stride,
Frank Barchard's avatar
Frank Barchard committed
518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537
                                uint8* dst_ptr,
                                int dst_width) {
  asm volatile(
      "movdqa    %0,%%xmm2                       \n"  // kShuf01
      "movdqa    %1,%%xmm3                       \n"  // kShuf11
      "movdqa    %2,%%xmm4                       \n"  // kShuf21
      :
      : "m"(kShuf01),  // %0
        "m"(kShuf11),  // %1
        "m"(kShuf21)   // %2
      );
  asm volatile(
      "movdqa    %0,%%xmm5                       \n"  // kMadd01
      "movdqa    %1,%%xmm0                       \n"  // kMadd11
      "movdqa    %2,%%xmm1                       \n"  // kRound34
      :
      : "m"(kMadd01),  // %0
        "m"(kMadd11),  // %1
        "m"(kRound34)  // %2
      );
538
  asm volatile (
539
    LABELALIGN
540
  "1:                                          \n"
541
    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
542
    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3),%%xmm7
543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558
    "pavgb     %%xmm7,%%xmm6                   \n"
    "pshufb    %%xmm2,%%xmm6                   \n"
    "pmaddubsw %%xmm5,%%xmm6                   \n"
    "paddsw    %%xmm1,%%xmm6                   \n"
    "psrlw     $0x2,%%xmm6                     \n"
    "packuswb  %%xmm6,%%xmm6                   \n"
    "movq      %%xmm6," MEMACCESS(1) "         \n"
    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3),%%xmm7
    "pavgb     %%xmm7,%%xmm6                   \n"
    "pshufb    %%xmm3,%%xmm6                   \n"
    "pmaddubsw %%xmm0,%%xmm6                   \n"
    "paddsw    %%xmm1,%%xmm6                   \n"
    "psrlw     $0x2,%%xmm6                     \n"
    "packuswb  %%xmm6,%%xmm6                   \n"
    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
559
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
560
    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3),%%xmm7
561 562 563 564 565 566 567 568 569 570 571 572 573 574
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "pavgb     %%xmm7,%%xmm6                   \n"
    "pshufb    %%xmm4,%%xmm6                   \n"
    "pmaddubsw %4,%%xmm6                       \n"
    "paddsw    %%xmm1,%%xmm6                   \n"
    "psrlw     $0x2,%%xmm6                     \n"
    "packuswb  %%xmm6,%%xmm6                   \n"
    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
    "lea       " MEMLEA(0x18,1) ",%1           \n"
    "sub       $0x18,%2                        \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),   // %0
    "+r"(dst_ptr),   // %1
    "+r"(dst_width)  // %2
575
  : "r"((intptr_t)(src_stride)),  // %3
576
    "m"(kMadd21)     // %4
577 578
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
579 580 581 582 583
  );
}

void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
                                ptrdiff_t src_stride,
Frank Barchard's avatar
Frank Barchard committed
584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603
                                uint8* dst_ptr,
                                int dst_width) {
  asm volatile(
      "movdqa    %0,%%xmm2                       \n"  // kShuf01
      "movdqa    %1,%%xmm3                       \n"  // kShuf11
      "movdqa    %2,%%xmm4                       \n"  // kShuf21
      :
      : "m"(kShuf01),  // %0
        "m"(kShuf11),  // %1
        "m"(kShuf21)   // %2
      );
  asm volatile(
      "movdqa    %0,%%xmm5                       \n"  // kMadd01
      "movdqa    %1,%%xmm0                       \n"  // kMadd11
      "movdqa    %2,%%xmm1                       \n"  // kRound34
      :
      : "m"(kMadd01),  // %0
        "m"(kMadd11),  // %1
        "m"(kRound34)  // %2
      );
604 605

  asm volatile (
606
    LABELALIGN
607
  "1:                                          \n"
608
    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
609
    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3,1),%%xmm7
610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627
    "pavgb     %%xmm6,%%xmm7                   \n"
    "pavgb     %%xmm7,%%xmm6                   \n"
    "pshufb    %%xmm2,%%xmm6                   \n"
    "pmaddubsw %%xmm5,%%xmm6                   \n"
    "paddsw    %%xmm1,%%xmm6                   \n"
    "psrlw     $0x2,%%xmm6                     \n"
    "packuswb  %%xmm6,%%xmm6                   \n"
    "movq      %%xmm6," MEMACCESS(1) "         \n"
    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3,1),%%xmm7
    "pavgb     %%xmm6,%%xmm7                   \n"
    "pavgb     %%xmm7,%%xmm6                   \n"
    "pshufb    %%xmm3,%%xmm6                   \n"
    "pmaddubsw %%xmm0,%%xmm6                   \n"
    "paddsw    %%xmm1,%%xmm6                   \n"
    "psrlw     $0x2,%%xmm6                     \n"
    "packuswb  %%xmm6,%%xmm6                   \n"
    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
628
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
629
    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3,1),%%xmm7
630 631 632 633 634 635 636 637 638 639 640 641 642 643 644
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "pavgb     %%xmm6,%%xmm7                   \n"
    "pavgb     %%xmm7,%%xmm6                   \n"
    "pshufb    %%xmm4,%%xmm6                   \n"
    "pmaddubsw %4,%%xmm6                       \n"
    "paddsw    %%xmm1,%%xmm6                   \n"
    "psrlw     $0x2,%%xmm6                     \n"
    "packuswb  %%xmm6,%%xmm6                   \n"
    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
    "lea       " MEMLEA(0x18,1) ",%1           \n"
    "sub       $0x18,%2                        \n"
    "jg        1b                              \n"
    : "+r"(src_ptr),   // %0
      "+r"(dst_ptr),   // %1
      "+r"(dst_width)  // %2
645
    : "r"((intptr_t)(src_stride)),  // %3
646
      "m"(kMadd21)     // %4
647 648
    : "memory", "cc", NACL_R14
      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
649 650 651
  );
}

Frank Barchard's avatar
Frank Barchard committed
652 653 654 655
void ScaleRowDown38_SSSE3(const uint8* src_ptr,
                          ptrdiff_t src_stride,
                          uint8* dst_ptr,
                          int dst_width) {
656
  (void)src_stride;
657 658 659
  asm volatile (
    "movdqa    %3,%%xmm4                       \n"
    "movdqa    %4,%%xmm5                       \n"
660 661

    LABELALIGN
662
  "1:                                          \n"
663 664
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
665 666 667 668 669 670 671 672 673 674 675 676 677 678 679
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "pshufb    %%xmm4,%%xmm0                   \n"
    "pshufb    %%xmm5,%%xmm1                   \n"
    "paddusb   %%xmm1,%%xmm0                   \n"
    "movq      %%xmm0," MEMACCESS(1) "         \n"
    "movhlps   %%xmm0,%%xmm1                   \n"
    "movd      %%xmm1," MEMACCESS2(0x8,1) "    \n"
    "lea       " MEMLEA(0xc,1) ",%1            \n"
    "sub       $0xc,%2                         \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),   // %0
    "+r"(dst_ptr),   // %1
    "+r"(dst_width)  // %2
  : "m"(kShuf38a),   // %3
    "m"(kShuf38b)    // %4
680
  : "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
681 682 683 684 685
  );
}

void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
                                ptrdiff_t src_stride,
Frank Barchard's avatar
Frank Barchard committed
686 687 688 689 690 691 692 693 694 695 696 697 698
                                uint8* dst_ptr,
                                int dst_width) {
  asm volatile(
      "movdqa    %0,%%xmm2                       \n"
      "movdqa    %1,%%xmm3                       \n"
      "movdqa    %2,%%xmm4                       \n"
      "movdqa    %3,%%xmm5                       \n"
      :
      : "m"(kShufAb0),  // %0
        "m"(kShufAb1),  // %1
        "m"(kShufAb2),  // %2
        "m"(kScaleAb2)  // %3
      );
699
  asm volatile (
700
    LABELALIGN
701
  "1:                                          \n"
702
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
703
    MEMOPREG(movdqu,0x00,0,3,1,xmm1)           //  movdqu  (%0,%3,1),%%xmm1
704
    "lea       " MEMLEA(0x10,0) ",%0           \n"
705
    "pavgb     %%xmm1,%%xmm0                   \n"
706 707 708 709 710 711 712 713 714 715 716 717 718
    "movdqa    %%xmm0,%%xmm1                   \n"
    "pshufb    %%xmm2,%%xmm1                   \n"
    "movdqa    %%xmm0,%%xmm6                   \n"
    "pshufb    %%xmm3,%%xmm6                   \n"
    "paddusw   %%xmm6,%%xmm1                   \n"
    "pshufb    %%xmm4,%%xmm0                   \n"
    "paddusw   %%xmm0,%%xmm1                   \n"
    "pmulhuw   %%xmm5,%%xmm1                   \n"
    "packuswb  %%xmm1,%%xmm1                   \n"
    "movd      %%xmm1," MEMACCESS(1) "         \n"
    "psrlq     $0x10,%%xmm1                    \n"
    "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
    "lea       " MEMLEA(0x6,1) ",%1            \n"
719
    "sub       $0x6,%2                         \n"
720 721 722 723
    "jg        1b                              \n"
  : "+r"(src_ptr),     // %0
    "+r"(dst_ptr),     // %1
    "+r"(dst_width)    // %2
724
  : "r"((intptr_t)(src_stride))  // %3
725 726
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
727 728 729 730 731
  );
}

void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
                                ptrdiff_t src_stride,
Frank Barchard's avatar
Frank Barchard committed
732 733 734 735 736 737 738 739 740 741 742 743
                                uint8* dst_ptr,
                                int dst_width) {
  asm volatile(
      "movdqa    %0,%%xmm2                       \n"
      "movdqa    %1,%%xmm3                       \n"
      "movdqa    %2,%%xmm4                       \n"
      "pxor      %%xmm5,%%xmm5                   \n"
      :
      : "m"(kShufAc),    // %0
        "m"(kShufAc3),   // %1
        "m"(kScaleAc33)  // %2
      );
744
  asm volatile (
745
    LABELALIGN
746
  "1:                                          \n"
747
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
748
    MEMOPREG(movdqu,0x00,0,3,1,xmm6)           //  movdqu  (%0,%3,1),%%xmm6
749 750 751 752 753 754 755 756
    "movhlps   %%xmm0,%%xmm1                   \n"
    "movhlps   %%xmm6,%%xmm7                   \n"
    "punpcklbw %%xmm5,%%xmm0                   \n"
    "punpcklbw %%xmm5,%%xmm1                   \n"
    "punpcklbw %%xmm5,%%xmm6                   \n"
    "punpcklbw %%xmm5,%%xmm7                   \n"
    "paddusw   %%xmm6,%%xmm0                   \n"
    "paddusw   %%xmm7,%%xmm1                   \n"
757
    MEMOPREG(movdqu,0x00,0,3,2,xmm6)           //  movdqu  (%0,%3,2),%%xmm6
758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782
    "lea       " MEMLEA(0x10,0) ",%0           \n"
    "movhlps   %%xmm6,%%xmm7                   \n"
    "punpcklbw %%xmm5,%%xmm6                   \n"
    "punpcklbw %%xmm5,%%xmm7                   \n"
    "paddusw   %%xmm6,%%xmm0                   \n"
    "paddusw   %%xmm7,%%xmm1                   \n"
    "movdqa    %%xmm0,%%xmm6                   \n"
    "psrldq    $0x2,%%xmm0                     \n"
    "paddusw   %%xmm0,%%xmm6                   \n"
    "psrldq    $0x2,%%xmm0                     \n"
    "paddusw   %%xmm0,%%xmm6                   \n"
    "pshufb    %%xmm2,%%xmm6                   \n"
    "movdqa    %%xmm1,%%xmm7                   \n"
    "psrldq    $0x2,%%xmm1                     \n"
    "paddusw   %%xmm1,%%xmm7                   \n"
    "psrldq    $0x2,%%xmm1                     \n"
    "paddusw   %%xmm1,%%xmm7                   \n"
    "pshufb    %%xmm3,%%xmm7                   \n"
    "paddusw   %%xmm7,%%xmm6                   \n"
    "pmulhuw   %%xmm4,%%xmm6                   \n"
    "packuswb  %%xmm6,%%xmm6                   \n"
    "movd      %%xmm6," MEMACCESS(1) "         \n"
    "psrlq     $0x10,%%xmm6                    \n"
    "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
    "lea       " MEMLEA(0x6,1) ",%1            \n"
783
    "sub       $0x6,%2                         \n"
784 785 786 787
    "jg        1b                              \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
788
  : "r"((intptr_t)(src_stride))   // %3
789 790
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
791 792 793
  );
}

794
// Reads 16xN bytes and produces 16 shorts at a time.
795
void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
796
  asm volatile (
797
    "pxor      %%xmm5,%%xmm5                   \n"
798 799

    LABELALIGN
800
  "1:                                          \n"
801 802 803 804 805 806 807
    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
    "lea       " MEMLEA(0x10,0) ",%0           \n"  // src_ptr += 16
    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm1   \n"
    "movdqa    %%xmm3,%%xmm2                   \n"
    "punpcklbw %%xmm5,%%xmm2                   \n"
    "punpckhbw %%xmm5,%%xmm3                   \n"
808 809
    "paddusw   %%xmm2,%%xmm0                   \n"
    "paddusw   %%xmm3,%%xmm1                   \n"
810 811
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
812
    "lea       " MEMLEA(0x20,1) ",%1           \n"
813
    "sub       $0x10,%2                        \n"
814 815 816
    "jg        1b                              \n"
  : "+r"(src_ptr),     // %0
    "+r"(dst_ptr),     // %1
817 818 819
    "+r"(src_width)    // %2
  :
  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
820 821 822
  );
}

823 824 825 826
#ifdef HAS_SCALEADDROW_AVX2
// Reads 32 bytes and accumulates to 32 shorts at a time.
void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
  asm volatile (
827
    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
828 829 830 831

    LABELALIGN
  "1:                                          \n"
    "vmovdqu    " MEMACCESS(0) ",%%ymm3        \n"
832
    "lea        " MEMLEA(0x20,0) ",%0          \n"  // src_ptr += 32
833 834 835 836
    "vpermq     $0xd8,%%ymm3,%%ymm3            \n"
    "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
    "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
    "vpaddusw   " MEMACCESS(1) ",%%ymm2,%%ymm0 \n"
837
    "vpaddusw   " MEMACCESS2(0x20,1) ",%%ymm3,%%ymm1 \n"
838
    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
839
    "vmovdqu    %%ymm1," MEMACCESS2(0x20,1) "  \n"
840 841 842 843 844 845 846 847 848 849 850 851 852
    "lea       " MEMLEA(0x40,1) ",%1           \n"
    "sub       $0x20,%2                        \n"
    "jg        1b                              \n"
    "vzeroupper                                \n"
  : "+r"(src_ptr),     // %0
    "+r"(dst_ptr),     // %1
    "+r"(src_width)    // %2
  :
  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
  );
}
#endif  // HAS_SCALEADDROW_AVX2

853 854
// Constant for making pixels signed to avoid pmaddubsw
// saturation.
855 856
static const uvec8 kFsub80 = {0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
                              0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
857 858

// Constant for making pixels unsigned and adding .5 for rounding.
859 860
static const uvec16 kFadd40 = {0x4040, 0x4040, 0x4040, 0x4040,
                               0x4040, 0x4040, 0x4040, 0x4040};
861

862
// Bilinear column filtering. SSSE3 version.
Frank Barchard's avatar
Frank Barchard committed
863 864 865 866 867
void ScaleFilterCols_SSSE3(uint8* dst_ptr,
                           const uint8* src_ptr,
                           int dst_width,
                           int x,
                           int dx) {
868
  intptr_t x0, x1, temp_pixel;
869 870 871 872 873 874
  asm volatile (
    "movd      %6,%%xmm2                       \n"
    "movd      %7,%%xmm3                       \n"
    "movl      $0x04040000,%k2                 \n"
    "movd      %k2,%%xmm5                      \n"
    "pcmpeqb   %%xmm6,%%xmm6                   \n"
875 876 877 878
    "psrlw     $0x9,%%xmm6                     \n"  // 0x007f007f
    "pcmpeqb   %%xmm7,%%xmm7                   \n"
    "psrlw     $15,%%xmm7                      \n"  // 0x00010001

879 880 881 882 883 884 885 886 887
    "pextrw    $0x1,%%xmm2,%k3                 \n"
    "subl      $0x2,%5                         \n"
    "jl        29f                             \n"
    "movdqa    %%xmm2,%%xmm0                   \n"
    "paddd     %%xmm3,%%xmm0                   \n"
    "punpckldq %%xmm0,%%xmm2                   \n"
    "punpckldq %%xmm3,%%xmm3                   \n"
    "paddd     %%xmm3,%%xmm3                   \n"
    "pextrw    $0x3,%%xmm2,%k4                 \n"
888 889

    LABELALIGN
890 891 892
  "2:                                          \n"
    "movdqa    %%xmm2,%%xmm1                   \n"
    "paddd     %%xmm3,%%xmm2                   \n"
893
    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
894 895
    "movd      %k2,%%xmm0                      \n"
    "psrlw     $0x9,%%xmm1                     \n"
896
    MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
897 898 899
    "movd      %k2,%%xmm4                      \n"
    "pshufb    %%xmm5,%%xmm1                   \n"
    "punpcklwd %%xmm4,%%xmm0                   \n"
900
    "psubb     %8,%%xmm0                       \n"  // make pixels signed.
901
    "pxor      %%xmm6,%%xmm1                   \n"  // 128 - f = (f ^ 127 ) + 1
902 903
    "paddusb   %%xmm7,%%xmm1                   \n"
    "pmaddubsw %%xmm0,%%xmm1                   \n"
904 905
    "pextrw    $0x1,%%xmm2,%k3                 \n"
    "pextrw    $0x3,%%xmm2,%k4                 \n"
906 907 908 909
    "paddw     %9,%%xmm1                       \n"  // make pixels unsigned.
    "psrlw     $0x7,%%xmm1                     \n"
    "packuswb  %%xmm1,%%xmm1                   \n"
    "movd      %%xmm1,%k2                      \n"
910 911
    "mov       %w2," MEMACCESS(0) "            \n"
    "lea       " MEMLEA(0x2,0) ",%0            \n"
912
    "subl      $0x2,%5                         \n"
913
    "jge       2b                              \n"
914 915

    LABELALIGN
916 917 918
  "29:                                         \n"
    "addl      $0x1,%5                         \n"
    "jl        99f                             \n"
919
    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
920 921 922
    "movd      %k2,%%xmm0                      \n"
    "psrlw     $0x9,%%xmm2                     \n"
    "pshufb    %%xmm5,%%xmm2                   \n"
923
    "psubb     %8,%%xmm0                       \n"  // make pixels signed.
924
    "pxor      %%xmm6,%%xmm2                   \n"
925 926 927 928 929 930
    "paddusb   %%xmm7,%%xmm2                   \n"
    "pmaddubsw %%xmm0,%%xmm2                   \n"
    "paddw     %9,%%xmm2                       \n"  // make pixels unsigned.
    "psrlw     $0x7,%%xmm2                     \n"
    "packuswb  %%xmm2,%%xmm2                   \n"
    "movd      %%xmm2,%k2                      \n"
931 932
    "mov       %b2," MEMACCESS(0) "            \n"
  "99:                                         \n"
933 934 935 936 937
  : "+r"(dst_ptr),      // %0
    "+r"(src_ptr),      // %1
    "=&a"(temp_pixel),  // %2
    "=&r"(x0),          // %3
    "=&r"(x1),          // %4
938
#if defined(__x86_64__)
939
    "+rm"(dst_width)    // %5
940 941 942
#else
    "+m"(dst_width)    // %5
#endif
943
  : "rm"(x),            // %6
944 945 946 947 948 949 950 951
    "rm"(dx),           // %7
#if defined(__x86_64__)
    "x"(kFsub80),       // %8
    "x"(kFadd40)        // %9
#else
    "m"(kFsub80),       // %8
    "m"(kFadd40)        // %9
#endif
952
  : "memory", "cc", NACL_R14
953
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
954 955 956 957 958
  );
}

// Reads 4 pixels, duplicates them and writes 8 pixels.
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
Frank Barchard's avatar
Frank Barchard committed
959 960 961 962 963
void ScaleColsUp2_SSE2(uint8* dst_ptr,
                       const uint8* src_ptr,
                       int dst_width,
                       int x,
                       int dx) {
964 965
  (void)x;
  (void)dx;
966
  asm volatile (
967
    LABELALIGN
968
  "1:                                          \n"
969
    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
970 971 972 973
    "lea       " MEMLEA(0x10,1) ",%1           \n"
    "movdqa    %%xmm0,%%xmm1                   \n"
    "punpcklbw %%xmm0,%%xmm0                   \n"
    "punpckhbw %%xmm1,%%xmm1                   \n"
974 975
    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
976
    "lea       " MEMLEA(0x20,0) ",%0           \n"
977
    "sub       $0x20,%2                         \n"
978 979 980 981 982
    "jg        1b                              \n"

  : "+r"(dst_ptr),     // %0
    "+r"(src_ptr),     // %1
    "+r"(dst_width)    // %2
983
  :: "memory", "cc", "xmm0", "xmm1"
984 985 986 987
  );
}

void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
988
                            ptrdiff_t src_stride,
Frank Barchard's avatar
Frank Barchard committed
989 990
                            uint8* dst_argb,
                            int dst_width) {
991
  (void)src_stride;
992
  asm volatile (
993
    LABELALIGN
994
  "1:                                          \n"
995 996
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
997 998
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "shufps    $0xdd,%%xmm1,%%xmm0             \n"
999
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
1000
    "lea       " MEMLEA(0x10,1) ",%1           \n"
1001
    "sub       $0x4,%2                         \n"
1002 1003 1004 1005
    "jg        1b                              \n"
  : "+r"(src_argb),  // %0
    "+r"(dst_argb),  // %1
    "+r"(dst_width)  // %2
1006
  :: "memory", "cc", "xmm0", "xmm1"
1007 1008 1009 1010
  );
}

void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
1011
                                  ptrdiff_t src_stride,
Frank Barchard's avatar
Frank Barchard committed
1012 1013
                                  uint8* dst_argb,
                                  int dst_width) {
1014
  (void)src_stride;
1015
  asm volatile (
1016
    LABELALIGN
1017
  "1:                                          \n"
1018 1019
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
1020 1021 1022 1023 1024
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "movdqa    %%xmm0,%%xmm2                   \n"
    "shufps    $0x88,%%xmm1,%%xmm0             \n"
    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
    "pavgb     %%xmm2,%%xmm0                   \n"
1025
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
1026
    "lea       " MEMLEA(0x10,1) ",%1           \n"
1027
    "sub       $0x4,%2                         \n"
1028 1029 1030 1031
    "jg        1b                              \n"
  : "+r"(src_argb),  // %0
    "+r"(dst_argb),  // %1
    "+r"(dst_width)  // %2
1032
  :: "memory", "cc", "xmm0", "xmm1"
1033 1034 1035 1036 1037
  );
}

void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
                               ptrdiff_t src_stride,
Frank Barchard's avatar
Frank Barchard committed
1038 1039
                               uint8* dst_argb,
                               int dst_width) {
1040
  asm volatile (
1041
    LABELALIGN
1042
  "1:                                          \n"
1043 1044
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
1045 1046
    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu   (%0,%3,1),%%xmm2
    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu   0x10(%0,%3,1),%%xmm3
1047 1048 1049 1050 1051 1052 1053
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "pavgb     %%xmm2,%%xmm0                   \n"
    "pavgb     %%xmm3,%%xmm1                   \n"
    "movdqa    %%xmm0,%%xmm2                   \n"
    "shufps    $0x88,%%xmm1,%%xmm0             \n"
    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
    "pavgb     %%xmm2,%%xmm0                   \n"
1054
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
1055
    "lea       " MEMLEA(0x10,1) ",%1           \n"
1056
    "sub       $0x4,%2                         \n"
1057 1058 1059 1060
    "jg        1b                              \n"
  : "+r"(src_argb),   // %0
    "+r"(dst_argb),   // %1
    "+r"(dst_width)   // %2
1061
  : "r"((intptr_t)(src_stride))   // %3
1062 1063
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3"
1064 1065 1066 1067 1068
  );
}

// Reads 4 pixels at a time.
// Alignment requirement: dst_argb 16 byte aligned.
Frank Barchard's avatar
Frank Barchard committed
1069 1070 1071 1072 1073
void ScaleARGBRowDownEven_SSE2(const uint8* src_argb,
                               ptrdiff_t src_stride,
                               int src_stepx,
                               uint8* dst_argb,
                               int dst_width) {
1074
  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
1075
  intptr_t src_stepx_x12;
1076
  (void)src_stride;
1077 1078 1079
  asm volatile (
    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
1080
    LABELALIGN
1081 1082 1083 1084 1085 1086 1087 1088 1089
  "1:                                          \n"
    "movd      " MEMACCESS(0) ",%%xmm0         \n"
    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
    "punpckldq %%xmm1,%%xmm0                   \n"
    MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
    MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
    "punpckldq %%xmm3,%%xmm2                   \n"
    "punpcklqdq %%xmm2,%%xmm0                  \n"
1090
    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
1091
    "lea       " MEMLEA(0x10,2) ",%2           \n"
1092
    "sub       $0x4,%3                         \n"
1093
    "jg        1b                              \n"
1094 1095 1096 1097 1098
  : "+r"(src_argb),       // %0
    "+r"(src_stepx_x4),   // %1
    "+r"(dst_argb),       // %2
    "+r"(dst_width),      // %3
    "=&r"(src_stepx_x12)  // %4
1099 1100
  :: "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3"
1101 1102 1103 1104 1105 1106
  );
}

// Blends four 2x2 to 4x1.
// Alignment requirement: dst_argb 16 byte aligned.
void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
Frank Barchard's avatar
Frank Barchard committed
1107 1108 1109 1110
                                  ptrdiff_t src_stride,
                                  int src_stepx,
                                  uint8* dst_argb,
                                  int dst_width) {
1111
  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
1112
  intptr_t src_stepx_x12;
1113
  intptr_t row1 = (intptr_t)(src_stride);
1114 1115 1116 1117
  asm volatile (
    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
    "lea       " MEMLEA4(0x00,0,5,1) ",%5      \n"
1118 1119

    LABELALIGN
1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136
  "1:                                          \n"
    "movq      " MEMACCESS(0) ",%%xmm0         \n"
    MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
    MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
    MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
    "movq      " MEMACCESS(5) ",%%xmm2         \n"
    MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
    MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
    MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
    "lea       " MEMLEA4(0x00,5,1,4) ",%5      \n"
    "pavgb     %%xmm2,%%xmm0                   \n"
    "pavgb     %%xmm3,%%xmm1                   \n"
    "movdqa    %%xmm0,%%xmm2                   \n"
    "shufps    $0x88,%%xmm1,%%xmm0             \n"
    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
    "pavgb     %%xmm2,%%xmm0                   \n"
1137
    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
1138
    "lea       " MEMLEA(0x10,2) ",%2           \n"
1139
    "sub       $0x4,%3                         \n"
1140
    "jg        1b                              \n"
1141 1142 1143 1144 1145 1146
  : "+r"(src_argb),        // %0
    "+r"(src_stepx_x4),    // %1
    "+r"(dst_argb),        // %2
    "+rm"(dst_width),      // %3
    "=&r"(src_stepx_x12),  // %4
    "+r"(row1)             // %5
1147 1148
  :: "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3"
1149 1150 1151
  );
}

Frank Barchard's avatar
Frank Barchard committed
1152 1153 1154 1155 1156
void ScaleARGBCols_SSE2(uint8* dst_argb,
                        const uint8* src_argb,
                        int dst_width,
                        int x,
                        int dx) {
1157
  intptr_t x0, x1;
1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174
  asm volatile (
    "movd      %5,%%xmm2                       \n"
    "movd      %6,%%xmm3                       \n"
    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
    "pshufd    $0x11,%%xmm3,%%xmm0             \n"
    "paddd     %%xmm0,%%xmm2                   \n"
    "paddd     %%xmm3,%%xmm3                   \n"
    "pshufd    $0x5,%%xmm3,%%xmm0              \n"
    "paddd     %%xmm0,%%xmm2                   \n"
    "paddd     %%xmm3,%%xmm3                   \n"
    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
    "pextrw    $0x1,%%xmm2,%k0                 \n"
    "pextrw    $0x3,%%xmm2,%k1                 \n"
    "cmp       $0x0,%4                         \n"
    "jl        99f                             \n"
    "sub       $0x4,%4                         \n"
    "jl        49f                             \n"
1175 1176

    LABELALIGN
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191
  "40:                                         \n"
    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
    "pextrw    $0x5,%%xmm2,%k0                 \n"
    "pextrw    $0x7,%%xmm2,%k1                 \n"
    "paddd     %%xmm3,%%xmm2                   \n"
    "punpckldq %%xmm1,%%xmm0                   \n"
    MEMOPREG(movd,0x00,3,0,4,xmm1)             //  movd      (%3,%0,4),%%xmm1
    MEMOPREG(movd,0x00,3,1,4,xmm4)             //  movd      (%3,%1,4),%%xmm4
    "pextrw    $0x1,%%xmm2,%k0                 \n"
    "pextrw    $0x3,%%xmm2,%k1                 \n"
    "punpckldq %%xmm4,%%xmm1                   \n"
    "punpcklqdq %%xmm1,%%xmm0                  \n"
    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
    "lea       " MEMLEA(0x10,2) ",%2           \n"
1192
    "sub       $0x4,%4                         \n"
1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209
    "jge       40b                             \n"

  "49:                                         \n"
    "test      $0x2,%4                         \n"
    "je        29f                             \n"
    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
    "pextrw    $0x5,%%xmm2,%k0                 \n"
    "punpckldq %%xmm1,%%xmm0                   \n"
    "movq      %%xmm0," MEMACCESS(2) "         \n"
    "lea       " MEMLEA(0x8,2) ",%2            \n"
  "29:                                         \n"
    "test      $0x1,%4                         \n"
    "je        99f                             \n"
    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
    "movd      %%xmm0," MEMACCESS(2) "         \n"
  "99:                                         \n"
1210 1211
  : "=&a"(x0),         // %0
    "=&d"(x1),         // %1
1212 1213 1214 1215 1216
    "+r"(dst_argb),    // %2
    "+r"(src_argb),    // %3
    "+r"(dst_width)    // %4
  : "rm"(x),           // %5
    "rm"(dx)           // %6
1217 1218
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
1219 1220 1221 1222 1223
  );
}

// Reads 4 pixels, duplicates them and writes 8 pixels.
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
Frank Barchard's avatar
Frank Barchard committed
1224 1225 1226 1227 1228
void ScaleARGBColsUp2_SSE2(uint8* dst_argb,
                           const uint8* src_argb,
                           int dst_width,
                           int x,
                           int dx) {
1229 1230
  (void)x;
  (void)dx;
1231
  asm volatile (
1232
    LABELALIGN
1233
  "1:                                          \n"
1234
    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
1235 1236 1237 1238
    "lea       " MEMLEA(0x10,1) ",%1           \n"
    "movdqa    %%xmm0,%%xmm1                   \n"
    "punpckldq %%xmm0,%%xmm0                   \n"
    "punpckhdq %%xmm1,%%xmm1                   \n"
1239 1240
    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
1241
    "lea       " MEMLEA(0x20,0) ",%0           \n"
1242
    "sub       $0x8,%2                         \n"
1243 1244 1245 1246 1247
    "jg        1b                              \n"

  : "+r"(dst_argb),    // %0
    "+r"(src_argb),    // %1
    "+r"(dst_width)    // %2
1248 1249
  :: "memory", "cc", NACL_R14
    "xmm0", "xmm1"
1250 1251 1252 1253
  );
}

// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
1254
static const uvec8 kShuffleColARGB = {
Frank Barchard's avatar
Frank Barchard committed
1255 1256
    0u, 4u,  1u, 5u,  2u,  6u,  3u,  7u,  // bbggrraa 1st pixel
    8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
1257 1258 1259
};

// Shuffle table for duplicating 2 fractions into 8 bytes each
1260
static const uvec8 kShuffleFractions = {
Frank Barchard's avatar
Frank Barchard committed
1261
    0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
1262 1263 1264
};

// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
Frank Barchard's avatar
Frank Barchard committed
1265 1266 1267 1268 1269
void ScaleARGBFilterCols_SSSE3(uint8* dst_argb,
                               const uint8* src_argb,
                               int dst_width,
                               int x,
                               int dx) {
1270
  intptr_t x0, x1;
Frank Barchard's avatar
Frank Barchard committed
1271 1272 1273 1274 1275 1276 1277
  asm volatile(
      "movdqa    %0,%%xmm4                       \n"
      "movdqa    %1,%%xmm5                       \n"
      :
      : "m"(kShuffleColARGB),   // %0
        "m"(kShuffleFractions)  // %1
      );
1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293

  asm volatile (
    "movd      %5,%%xmm2                       \n"
    "movd      %6,%%xmm3                       \n"
    "pcmpeqb   %%xmm6,%%xmm6                   \n"
    "psrlw     $0x9,%%xmm6                     \n"
    "pextrw    $0x1,%%xmm2,%k3                 \n"
    "sub       $0x2,%2                         \n"
    "jl        29f                             \n"
    "movdqa    %%xmm2,%%xmm0                   \n"
    "paddd     %%xmm3,%%xmm0                   \n"
    "punpckldq %%xmm0,%%xmm2                   \n"
    "punpckldq %%xmm3,%%xmm3                   \n"
    "paddd     %%xmm3,%%xmm3                   \n"
    "pextrw    $0x3,%%xmm2,%k4                 \n"

1294
    LABELALIGN
1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313
  "2:                                          \n"
    "movdqa    %%xmm2,%%xmm1                   \n"
    "paddd     %%xmm3,%%xmm2                   \n"
    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
    "psrlw     $0x9,%%xmm1                     \n"
    MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
    "pshufb    %%xmm5,%%xmm1                   \n"
    "pshufb    %%xmm4,%%xmm0                   \n"
    "pxor      %%xmm6,%%xmm1                   \n"
    "pmaddubsw %%xmm1,%%xmm0                   \n"
    "psrlw     $0x7,%%xmm0                     \n"
    "pextrw    $0x1,%%xmm2,%k3                 \n"
    "pextrw    $0x3,%%xmm2,%k4                 \n"
    "packuswb  %%xmm0,%%xmm0                   \n"
    "movq      %%xmm0," MEMACCESS(0) "         \n"
    "lea       " MEMLEA(0x8,0) ",%0            \n"
    "sub       $0x2,%2                         \n"
    "jge       2b                              \n"

1314
    LABELALIGN
1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327
  "29:                                         \n"
    "add       $0x1,%2                         \n"
    "jl        99f                             \n"
    "psrlw     $0x9,%%xmm2                     \n"
    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
    "pshufb    %%xmm5,%%xmm2                   \n"
    "pshufb    %%xmm4,%%xmm0                   \n"
    "pxor      %%xmm6,%%xmm2                   \n"
    "pmaddubsw %%xmm2,%%xmm0                   \n"
    "psrlw     $0x7,%%xmm0                     \n"
    "packuswb  %%xmm0,%%xmm0                   \n"
    "movd      %%xmm0," MEMACCESS(0) "         \n"

1328
    LABELALIGN
1329 1330 1331 1332
  "99:                                         \n"
  : "+r"(dst_argb),    // %0
    "+r"(src_argb),    // %1
    "+rm"(dst_width),  // %2
1333 1334
    "=&r"(x0),         // %3
    "=&r"(x1)          // %4
1335 1336
  : "rm"(x),           // %5
    "rm"(dx)           // %6
1337 1338
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
1339 1340 1341
  );
}

1342 1343
// Divide num by div and return as 16.16 fixed point result.
int FixedDiv_X86(int num, int div) {
Frank Barchard's avatar
Frank Barchard committed
1344 1345 1346 1347 1348 1349 1350 1351 1352
  asm volatile(
      "cdq                                       \n"
      "shld      $0x10,%%eax,%%edx               \n"
      "shl       $0x10,%%eax                     \n"
      "idiv      %1                              \n"
      "mov       %0, %%eax                       \n"
      : "+a"(num)  // %0
      : "c"(div)   // %1
      : "memory", "cc", "edx");
1353 1354 1355 1356 1357
  return num;
}

// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
int FixedDiv1_X86(int num, int div) {
Frank Barchard's avatar
Frank Barchard committed
1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369
  asm volatile(
      "cdq                                       \n"
      "shld      $0x10,%%eax,%%edx               \n"
      "shl       $0x10,%%eax                     \n"
      "sub       $0x10001,%%eax                  \n"
      "sbb       $0x0,%%edx                      \n"
      "sub       $0x1,%1                         \n"
      "idiv      %1                              \n"
      "mov       %0, %%eax                       \n"
      : "+a"(num)  // %0
      : "c"(div)   // %1
      : "memory", "cc", "edx");
1370 1371 1372
  return num;
}

1373 1374 1375 1376 1377 1378
#endif  // defined(__x86_64__) || defined(__i386__)

#ifdef __cplusplus
}  // extern "C"
}  // namespace libyuv
#endif