git.sesse.net Git - vlc/blob - modules/video_chroma/i420_rgb_mmx.h

   1 /*****************************************************************************
   2  * transforms_yuvmmx.h: MMX YUV transformation assembly
   3  *****************************************************************************
   4  * Copyright (C) 1999-2004 the VideoLAN team
   5  * $Id$
   6  *
   7  * Authors: Olie Lho <ollie@sis.com.tw>
   8  *          Gaël Hendryckx <jimmy@via.ecp.fr>
   9  *          Samuel Hocevar <sam@zoy.org>
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2 of the License, or
  14  * (at your option) any later version.
  15  *
  16  * This program is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  * GNU General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU General Public License
  22  * along with this program; if not, write to the Free Software
  23  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  24  *****************************************************************************/
  25
  26 /* hope these constant values are cache line aligned */
  27 #if __GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 3)
  28 #define USED_U64(foo) \
  29     static const uint64_t foo __asm__ (#foo) __attribute__((used))
  30 #else
  31 #define USED_U64(foo) \
  32     static const uint64_t foo __asm__ (#foo) __attribute__((unused))
  33 #endif
  34 USED_U64(mmx_80w)     = 0x0080008000800080ULL;
  35 USED_U64(mmx_10w)     = 0x1010101010101010ULL;
  36 USED_U64(mmx_00ffw)   = 0x00ff00ff00ff00ffULL;
  37 USED_U64(mmx_Y_coeff) = 0x253f253f253f253fULL;
  38
  39 USED_U64(mmx_U_green) = 0xf37df37df37df37dULL;
  40 USED_U64(mmx_U_blue)  = 0x4093409340934093ULL;
  41 USED_U64(mmx_V_red)   = 0x3312331233123312ULL;
  42 USED_U64(mmx_V_green) = 0xe5fce5fce5fce5fcULL;
  43
  44 USED_U64(mmx_mask_f8) = 0xf8f8f8f8f8f8f8f8ULL;
  45 USED_U64(mmx_mask_fc) = 0xfcfcfcfcfcfcfcfcULL;
  46 #undef USED_U64
  47
  48 #define MMX_INIT_16 "                                                       \n\
  49 movd      (%1), %%mm0       # Load 4 Cb       00 00 00 00 u3 u2 u1 u0       \n\
  50 movd      (%2), %%mm1       # Load 4 Cr       00 00 00 00 v3 v2 v1 v0       \n\
  51 pxor      %%mm4, %%mm4      # zero mm4                                      \n\
  52 movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
  53 #movl      $0, (%3)         # cache preload for image                       \n\
  54 "
  55
  56 #define INTRINSICS_INIT_16 \
  57     tmp64 = *(uint32_t *)p_u; \
  58     mm0 = (__m64)tmp64; \
  59     tmp64 = *(uint32_t *)p_v; \
  60     mm1 = (__m64)tmp64; \
  61     mm4 = (__m64)(uint64_t)0; \
  62     mm6 = (__m64)*(uint64_t *)p_y; \
  63     /* *(uint16_t *)p_buffer = 0; */
  64
  65 #define MMX_INIT_16_GRAY "                                                  \n\
  66 movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
  67 #movl      $0, (%3)         # cache preload for image                       \n\
  68 "
  69
  70 #define MMX_INIT_32 "                                                       \n\
  71 movd      (%1), %%mm0       # Load 4 Cb       00 00 00 00 u3 u2 u1 u0       \n\
  72 movl      $0, (%3)          # cache preload for image                       \n\
  73 movd      (%2), %%mm1       # Load 4 Cr       00 00 00 00 v3 v2 v1 v0       \n\
  74 pxor      %%mm4, %%mm4      # zero mm4                                      \n\
  75 movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
  76 "
  77
  78 #define INTRINSICS_INIT_32 \
  79     tmp64 = *(uint32_t *)p_u; \
  80     mm0 = (__m64)tmp64; \
  81     *(uint16_t *)p_buffer = 0; \
  82     tmp64 = *(uint32_t *)p_v; \
  83     mm1 = (__m64)tmp64; \
  84     mm4 = (__m64)(uint64_t)0; \
  85     mm6 = (__m64)*(uint64_t *)p_y;
  86
  87 /*
  88  * Do the multiply part of the conversion for even and odd pixels,
  89  * register usage:
  90  * mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels,
  91  * mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd  pixels,
  92  * mm6 -> Y even, mm7 -> Y odd
  93  */
  94
  95 #define MMX_YUV_MUL "                                                       \n\
  96 # convert the chroma part                                                   \n\
  97 punpcklbw %%mm4, %%mm0          # scatter 4 Cb    00 u3 00 u2 00 u1 00 u0   \n\
  98 punpcklbw %%mm4, %%mm1          # scatter 4 Cr    00 v3 00 v2 00 v1 00 v0   \n\
  99 psubsw    mmx_80w, %%mm0        # Cb -= 128                                 \n\
 100 psubsw    mmx_80w, %%mm1        # Cr -= 128                                 \n\
 101 psllw     $3, %%mm0             # Promote precision                         \n\
 102 psllw     $3, %%mm1             # Promote precision                         \n\
 103 movq      %%mm0, %%mm2          # Copy 4 Cb       00 u3 00 u2 00 u1 00 u0   \n\
 104 movq      %%mm1, %%mm3          # Copy 4 Cr       00 v3 00 v2 00 v1 00 v0   \n\
 105 pmulhw    mmx_U_green, %%mm2    # Mul Cb with green coeff -> Cb green       \n\
 106 pmulhw    mmx_V_green, %%mm3    # Mul Cr with green coeff -> Cr green       \n\
 107 pmulhw    mmx_U_blue, %%mm0     # Mul Cb -> Cblue 00 b3 00 b2 00 b1 00 b0   \n\
 108 pmulhw    mmx_V_red, %%mm1      # Mul Cr -> Cred  00 r3 00 r2 00 r1 00 r0   \n\
 109 paddsw    %%mm3, %%mm2          # Cb green + Cr green -> Cgreen             \n\
 110                                                                             \n\
 111 # convert the luma part                                                     \n\
 112 psubusb   mmx_10w, %%mm6        # Y -= 16                                   \n\
 113 movq      %%mm6, %%mm7          # Copy 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0   \n\
 114 pand      mmx_00ffw, %%mm6      # get Y even      00 Y6 00 Y4 00 Y2 00 Y0   \n\
 115 psrlw     $8, %%mm7             # get Y odd       00 Y7 00 Y5 00 Y3 00 Y1   \n\
 116 psllw     $3, %%mm6             # Promote precision                         \n\
 117 psllw     $3, %%mm7             # Promote precision                         \n\
 118 pmulhw    mmx_Y_coeff, %%mm6    # Mul 4 Y even    00 y6 00 y4 00 y2 00 y0   \n\
 119 pmulhw    mmx_Y_coeff, %%mm7    # Mul 4 Y odd     00 y7 00 y5 00 y3 00 y1   \n\
 120 "
 121
 122 #define INTRINSICS_YUV_MUL \
 123     mm0 = _mm_unpacklo_pi8(mm0, mm4); \
 124     mm1 = _mm_unpacklo_pi8(mm1, mm4); \
 125     mm0 = _mm_subs_pi16(mm0, (__m64)mmx_80w); \
 126     mm1 = _mm_subs_pi16(mm1, (__m64)mmx_80w); \
 127     mm0 = _mm_slli_pi16(mm0, 3); \
 128     mm1 = _mm_slli_pi16(mm1, 3); \
 129     mm2 = mm0; \
 130     mm3 = mm1; \
 131     mm2 = _mm_mulhi_pi16(mm2, (__m64)mmx_U_green); \
 132     mm3 = _mm_mulhi_pi16(mm3, (__m64)mmx_V_green); \
 133     mm0 = _mm_mulhi_pi16(mm0, (__m64)mmx_U_blue); \
 134     mm1 = _mm_mulhi_pi16(mm1, (__m64)mmx_V_red); \
 135     mm2 = _mm_adds_pi16(mm2, mm3); \
 136     \
 137     mm6 = _mm_subs_pu8(mm6, (__m64)mmx_10w); \
 138     mm7 = mm6; \
 139     mm6 = _mm_and_si64(mm6, (__m64)mmx_00ffw); \
 140     mm7 = _mm_srli_pi16(mm7, 8); \
 141     mm6 = _mm_slli_pi16(mm6, 3); \
 142     mm7 = _mm_slli_pi16(mm7, 3); \
 143     mm6 = _mm_mulhi_pi16(mm6, (__m64)mmx_Y_coeff); \
 144     mm7 = _mm_mulhi_pi16(mm7, (__m64)mmx_Y_coeff);
 145
 146 /*
 147  * Do the addition part of the conversion for even and odd pixels,
 148  * register usage:
 149  * mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels,
 150  * mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd  pixels,
 151  * mm6 -> Y even, mm7 -> Y odd
 152  */
 153
 154 #define MMX_YUV_ADD "                                                       \n\
 155 # Do horizontal and vertical scaling                                        \n\
 156 movq      %%mm0, %%mm3          # Copy Cblue                                \n\
 157 movq      %%mm1, %%mm4          # Copy Cred                                 \n\
 158 movq      %%mm2, %%mm5          # Copy Cgreen                               \n\
 159 paddsw    %%mm6, %%mm0          # Y even + Cblue  00 B6 00 B4 00 B2 00 B0   \n\
 160 paddsw    %%mm7, %%mm3          # Y odd  + Cblue  00 B7 00 B5 00 B3 00 B1   \n\
 161 paddsw    %%mm6, %%mm1          # Y even + Cred   00 R6 00 R4 00 R2 00 R0   \n\
 162 paddsw    %%mm7, %%mm4          # Y odd  + Cred   00 R7 00 R5 00 R3 00 R1   \n\
 163 paddsw    %%mm6, %%mm2          # Y even + Cgreen 00 G6 00 G4 00 G2 00 G0   \n\
 164 paddsw    %%mm7, %%mm5          # Y odd  + Cgreen 00 G7 00 G5 00 G3 00 G1   \n\
 165                                                                             \n\
 166 # Limit RGB even to 0..255                                                  \n\
 167 packuswb  %%mm0, %%mm0          # B6 B4 B2 B0 / B6 B4 B2 B0                 \n\
 168 packuswb  %%mm1, %%mm1          # R6 R4 R2 R0 / R6 R4 R2 R0                 \n\
 169 packuswb  %%mm2, %%mm2          # G6 G4 G2 G0 / G6 G4 G2 G0                 \n\
 170                                                                             \n\
 171 # Limit RGB odd to 0..255                                                   \n\
 172 packuswb  %%mm3, %%mm3          # B7 B5 B3 B1 / B7 B5 B3 B1                 \n\
 173 packuswb  %%mm4, %%mm4          # R7 R5 R3 R1 / R7 R5 R3 R1                 \n\
 174 packuswb  %%mm5, %%mm5          # G7 G5 G3 G1 / G7 G5 G3 G1                 \n\
 175                                                                             \n\
 176 # Interleave RGB even and odd                                               \n\
 177 punpcklbw %%mm3, %%mm0          #                 B7 B6 B5 B4 B3 B2 B1 B0   \n\
 178 punpcklbw %%mm4, %%mm1          #                 R7 R6 R5 R4 R3 R2 R1 R0   \n\
 179 punpcklbw %%mm5, %%mm2          #                 G7 G6 G5 G4 G3 G2 G1 G0   \n\
 180 "
 181
 182 #define INTRINSICS_YUV_ADD \
 183     mm3 = mm0; \
 184     mm4 = mm1; \
 185     mm5 = mm2; \
 186     mm0 = _mm_adds_pi16(mm0, mm6); \
 187     mm3 = _mm_adds_pi16(mm3, mm7); \
 188     mm1 = _mm_adds_pi16(mm1, mm6); \
 189     mm4 = _mm_adds_pi16(mm4, mm7); \
 190     mm2 = _mm_adds_pi16(mm2, mm6); \
 191     mm5 = _mm_adds_pi16(mm5, mm7); \
 192     \
 193     mm0 = _mm_packs_pu16(mm0, mm0); \
 194     mm1 = _mm_packs_pu16(mm1, mm1); \
 195     mm2 = _mm_packs_pu16(mm2, mm2); \
 196     \
 197     mm3 = _mm_packs_pu16(mm3, mm3); \
 198     mm4 = _mm_packs_pu16(mm4, mm4); \
 199     mm5 = _mm_packs_pu16(mm5, mm5); \
 200     \
 201     mm0 = _mm_unpacklo_pi8(mm0, mm3); \
 202     mm1 = _mm_unpacklo_pi8(mm1, mm4); \
 203     mm2 = _mm_unpacklo_pi8(mm2, mm5);
 204
 205 /*
 206  * Grayscale case, only use Y
 207  */
 208
 209 #define MMX_YUV_GRAY "                                                      \n\
 210 # convert the luma part                                                     \n\
 211 psubusb   mmx_10w, %%mm6                                                    \n\
 212 movq      %%mm6, %%mm7                                                      \n\
 213 pand      mmx_00ffw, %%mm6                                                  \n\
 214 psrlw     $8, %%mm7                                                         \n\
 215 psllw     $3, %%mm6                                                         \n\
 216 psllw     $3, %%mm7                                                         \n\
 217 pmulhw    mmx_Y_coeff, %%mm6                                                \n\
 218 pmulhw    mmx_Y_coeff, %%mm7                                                \n\
 219 packuswb  %%mm6, %%mm6                                                      \n\
 220 packuswb  %%mm7, %%mm7                                                      \n\
 221 punpcklbw %%mm7, %%mm6                                                      \n\
 222 "
 223
 224 #define MMX_UNPACK_16_GRAY "                                                \n\
 225 movq      %%mm6, %%mm5                                                      \n\
 226 pand      mmx_mask_f8, %%mm6                                                \n\
 227 pand      mmx_mask_fc, %%mm5                                                \n\
 228 movq      %%mm6, %%mm7                                                      \n\
 229 psrlw     $3, %%mm7                                                         \n\
 230 pxor      %%mm3, %%mm3                                                      \n\
 231 movq      %%mm7, %%mm2                                                      \n\
 232 movq      %%mm5, %%mm0                                                      \n\
 233 punpcklbw %%mm3, %%mm5                                                      \n\
 234 punpcklbw %%mm6, %%mm7                                                      \n\
 235 psllw     $3, %%mm5                                                         \n\
 236 por       %%mm5, %%mm7                                                      \n\
 237 movq      %%mm7, (%3)                                                       \n\
 238 punpckhbw %%mm3, %%mm0                                                      \n\
 239 punpckhbw %%mm6, %%mm2                                                      \n\
 240 psllw     $3, %%mm0                                                         \n\
 241 movq      8(%0), %%mm6                                                      \n\
 242 por       %%mm0, %%mm2                                                      \n\
 243 movq      %%mm2, 8(%3)                                                      \n\
 244 "
 245
 246
 247 /*
 248  * convert RGB plane to RGB 15 bits,
 249  * mm0 -> B, mm1 -> R, mm2 -> G,
 250  * mm4 -> GB, mm5 -> AR pixel 4-7,
 251  * mm6 -> GB, mm7 -> AR pixel 0-3
 252  */
 253
 254 #define MMX_UNPACK_15 "                                                     \n\
 255 # mask unneeded bits off                                                    \n\
 256 pand      mmx_mask_f8, %%mm0    # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
 257 psrlw     $3,%%mm0              # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
 258 pand      mmx_mask_f8, %%mm2    # g7g6g5g4 g3______ g7g6g5g4 g3______       \n\
 259 pand      mmx_mask_f8, %%mm1    # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
 260 psrlw     $1,%%mm1              # __r7r6r5 r4r3____ __r7r6r5 r4r3____       \n\
 261 pxor      %%mm4, %%mm4          # zero mm4                                  \n\
 262 movq      %%mm0, %%mm5          # Copy B7-B0                                \n\
 263 movq      %%mm2, %%mm7          # Copy G7-G0                                \n\
 264                                                                             \n\
 265 # convert rgb24 plane to rgb15 pack for pixel 0-3                           \n\
 266 punpcklbw %%mm4, %%mm2          # ________ ________ g7g6g5g4 g3______       \n\
 267 punpcklbw %%mm1, %%mm0          # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
 268 psllw     $2,%%mm2              # ________ ____g7g6 g5g4g3__ ________       \n\
 269 por       %%mm2, %%mm0          # r7r6r5r4 r3__g7g6 g5g4g3b7 b6b5b4b3       \n\
 270 movq      8(%0), %%mm6          # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0   \n\
 271 movq      %%mm0, (%3)           # store pixel 0-3                           \n\
 272                                                                             \n\
 273 # convert rgb24 plane to rgb16 pack for pixel 0-3                           \n\
 274 punpckhbw %%mm4, %%mm7          # ________ ________ g7g6g5g4 g3______       \n\
 275 punpckhbw %%mm1, %%mm5          # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
 276 psllw     $2,%%mm7              # ________ ____g7g6 g5g4g3__ ________       \n\
 277 movd      4(%1), %%mm0          # Load 4 Cb       __ __ __ __ u3 u2 u1 u0   \n\
 278 por       %%mm7, %%mm5          # r7r6r5r4 r3__g7g6 g5g4g3b7 b6b5b4b3       \n\
 279 movd      4(%2), %%mm1          # Load 4 Cr       __ __ __ __ v3 v2 v1 v0   \n\
 280 movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
 281 "
 282
 283 #define INTRINSICS_UNPACK_15 \
 284     mm0 = _mm_and_si64(mm0, (__m64)mmx_mask_f8); \
 285     mm0 = _mm_srli_pi16(mm0, 3); \
 286     mm2 = _mm_and_si64(mm2, (__m64)mmx_mask_f8); \
 287     mm1 = _mm_and_si64(mm1, (__m64)mmx_mask_f8); \
 288     mm1 = _mm_srli_pi16(mm1, 1); \
 289     mm4 = (__m64)(uint64_t)0; \
 290     mm5 = mm0; \
 291     mm7 = mm2; \
 292     \
 293     mm2 = _mm_unpacklo_pi8(mm2, mm4); \
 294     mm0 = _mm_unpacklo_pi8(mm0, mm1); \
 295     mm2 = _mm_slli_pi16(mm2, 2); \
 296     mm0 = _mm_or_si64(mm0, mm2); \
 297     tmp64 = *(uint64_t *)(p_y + 8); \
 298     mm6 = (__m64)tmp64; \
 299     *(uint64_t *)p_buffer = (uint64_t)mm0; \
 300     \
 301     mm7 = _mm_unpackhi_pi8(mm7, mm4); \
 302     mm5 = _mm_unpackhi_pi8(mm5, mm1); \
 303     mm7 = _mm_slli_pi16(mm7, 2); \
 304     tmp64 = (uint64_t)*(uint32_t *)(p_u + 4); \
 305     mm0 = (__m64)tmp64; \
 306     mm5 = _mm_or_si64(mm5, mm7); \
 307     tmp64 = (uint64_t)*(uint32_t *)(p_v + 4); \
 308     mm1 = (__m64)tmp64; \
 309     *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
 310
 311 /*
 312  * convert RGB plane to RGB 16 bits,
 313  * mm0 -> B, mm1 -> R, mm2 -> G,
 314  * mm4 -> GB, mm5 -> AR pixel 4-7,
 315  * mm6 -> GB, mm7 -> AR pixel 0-3
 316  */
 317
 318 #define MMX_UNPACK_16 "                                                     \n\
 319 # mask unneeded bits off                                                    \n\
 320 pand      mmx_mask_f8, %%mm0    # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
 321 pand      mmx_mask_fc, %%mm2    # g7g6g5g4 g3g2____ g7g6g5g4 g3g2____       \n\
 322 pand      mmx_mask_f8, %%mm1    # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
 323 psrlw     $3,%%mm0              # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
 324 pxor      %%mm4, %%mm4          # zero mm4                                  \n\
 325 movq      %%mm0, %%mm5          # Copy B7-B0                                \n\
 326 movq      %%mm2, %%mm7          # Copy G7-G0                                \n\
 327                                                                             \n\
 328 # convert rgb24 plane to rgb16 pack for pixel 0-3                           \n\
 329 punpcklbw %%mm4, %%mm2          # ________ ________ g7g6g5g4 g3g2____       \n\
 330 punpcklbw %%mm1, %%mm0          # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
 331 psllw     $3,%%mm2              # ________ __g7g6g5 g4g3g2__ ________       \n\
 332 por       %%mm2, %%mm0          # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
 333 movq      8(%0), %%mm6          # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0   \n\
 334 movq      %%mm0, (%3)           # store pixel 0-3                           \n\
 335                                                                             \n\
 336 # convert rgb24 plane to rgb16 pack for pixel 0-3                           \n\
 337 punpckhbw %%mm4, %%mm7          # ________ ________ g7g6g5g4 g3g2____       \n\
 338 punpckhbw %%mm1, %%mm5          # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
 339 psllw     $3,%%mm7              # ________ __g7g6g5 g4g3g2__ ________       \n\
 340 movd      4(%1), %%mm0          # Load 4 Cb       __ __ __ __ u3 u2 u1 u0   \n\
 341 por       %%mm7, %%mm5          # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
 342 movd      4(%2), %%mm1          # Load 4 Cr       __ __ __ __ v3 v2 v1 v0   \n\
 343 movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
 344 "
 345
 346 #define INTRINSICS_UNPACK_16 \
 347     mm0 = _mm_and_si64(mm0, (__m64)mmx_mask_f8); \
 348     mm2 = _mm_and_si64(mm2, (__m64)mmx_mask_fc); \
 349     mm1 = _mm_and_si64(mm1, (__m64)mmx_mask_f8); \
 350     mm0 = _mm_srli_pi16(mm0, 3); \
 351     mm4 = (__m64)(uint64_t)0; \
 352     mm5 = mm0; \
 353     mm7 = mm2; \
 354     \
 355     mm2 = _mm_unpacklo_pi8(mm2, mm4); \
 356     mm0 = _mm_unpacklo_pi8(mm0, mm1); \
 357     mm2 = _mm_slli_pi16(mm2, 3); \
 358     mm0 = _mm_or_si64(mm0, mm2); \
 359     tmp64 = *(uint64_t *)(p_y + 8); \
 360     mm6 = (__m64)tmp64; \
 361     *(uint64_t *)p_buffer = (uint64_t)mm0; \
 362     \
 363     mm7 = _mm_unpackhi_pi8(mm7, mm4); \
 364     mm5 = _mm_unpackhi_pi8(mm5, mm1); \
 365     mm7 = _mm_slli_pi16(mm7, 3); \
 366     tmp64 = (uint64_t)*(uint32_t *)(p_u + 4); \
 367     mm0 = (__m64)tmp64; \
 368     mm5 = _mm_or_si64(mm5, mm7); \
 369     tmp64 = (uint64_t)*(uint32_t *)(p_v + 4); \
 370     mm1 = (__m64)tmp64; \
 371     *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
 372
 373 /*
 374  * convert RGB plane to RGB packed format,
 375  * mm0 -> B, mm1 -> R, mm2 -> G, mm3 -> 0,
 376  * mm4 -> GB, mm5 -> AR pixel 4-7,
 377  * mm6 -> GB, mm7 -> AR pixel 0-3
 378  */
 379
 380 #define MMX_UNPACK_32 "                                                     \n\
 381 pxor      %%mm3, %%mm3  # zero mm3                                          \n\
 382 movq      %%mm0, %%mm6  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
 383 movq      %%mm1, %%mm7  #                 R7 R6 R5 R4 R3 R2 R1 R0           \n\
 384 movq      %%mm0, %%mm4  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
 385 movq      %%mm1, %%mm5  #                 R7 R6 R5 R4 R3 R2 R1 R0           \n\
 386 punpcklbw %%mm2, %%mm6  #                 G3 B3 G2 B2 G1 B1 G0 B0           \n\
 387 punpcklbw %%mm3, %%mm7  #                 00 R3 00 R2 00 R1 00 R0           \n\
 388 punpcklwd %%mm7, %%mm6  #                 00 R1 B1 G1 00 R0 B0 G0           \n\
 389 movq      %%mm6, (%3)   # Store ARGB1 ARGB0                                 \n\
 390 movq      %%mm0, %%mm6  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
 391 punpcklbw %%mm2, %%mm6  #                 G3 B3 G2 B2 G1 B1 G0 B0           \n\
 392 punpckhwd %%mm7, %%mm6  #                 00 R3 G3 B3 00 R2 B3 G2           \n\
 393 movq      %%mm6, 8(%3)  # Store ARGB3 ARGB2                                 \n\
 394 punpckhbw %%mm2, %%mm4  #                 G7 B7 G6 B6 G5 B5 G4 B4           \n\
 395 punpckhbw %%mm3, %%mm5  #                 00 R7 00 R6 00 R5 00 R4           \n\
 396 punpcklwd %%mm5, %%mm4  #                 00 R5 B5 G5 00 R4 B4 G4           \n\
 397 movq      %%mm4, 16(%3) # Store ARGB5 ARGB4                                 \n\
 398 movq      %%mm0, %%mm4  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
 399 punpckhbw %%mm2, %%mm4  #                 G7 B7 G6 B6 G5 B5 G4 B4           \n\
 400 punpckhwd %%mm5, %%mm4  #                 00 R7 G7 B7 00 R6 B6 G6           \n\
 401 movq      %%mm4, 24(%3) # Store ARGB7 ARGB6                                 \n\
 402                                                                             \n\
 403 #movd      4(%1), %%mm0  # Load 4 Cb       00 00 00 00 u3 u2 u1 u0           \n\
 404 #movd      4(%2), %%mm1  # Load 4 Cr       00 00 00 00 v3 v2 v1 v0           \n\
 405 #pxor      %%mm4, %%mm4  # zero mm4                                          \n\
 406 #movq      8(%0), %%mm6  # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0           \n\
 407 "
 408
 409 #define INTRINSICS_UNPACK_32 \
 410     mm3 = (__m64)(uint64_t)0; \
 411     mm6 = mm0; \
 412     mm7 = mm1; \
 413     mm4 = mm0; \
 414     mm5 = mm1; \
 415     mm6 = _mm_unpacklo_pi8(mm6, mm2); \
 416     mm7 = _mm_unpacklo_pi8(mm7, mm3); \
 417     mm6 = _mm_unpacklo_pi16(mm6, mm7); \
 418     *(uint64_t *)p_buffer = (uint64_t)mm6; \
 419     mm6 = mm0; \
 420     mm6 = _mm_unpacklo_pi8(mm6, mm2); \
 421     mm6 = _mm_unpackhi_pi16(mm6, mm7); \
 422     *(uint64_t *)(p_buffer + 2) = (uint64_t)mm6; \
 423     mm4 = _mm_unpackhi_pi8(mm4, mm2); \
 424     mm5 = _mm_unpackhi_pi8(mm5, mm3); \
 425     mm4 = _mm_unpacklo_pi16(mm4, mm5); \
 426     *(uint64_t *)(p_buffer + 4) = (uint64_t)mm4; \
 427     mm4 = mm0; \
 428     mm4 = _mm_unpackhi_pi8(mm4, mm2); \
 429     mm4 = _mm_unpackhi_pi16(mm4, mm5); \
 430     *(uint64_t *)(p_buffer + 6) = (uint64_t)mm4; \
 431