ffmpeg/libswscale/x86/input.asm

   1 ;******************************************************************************
   2 ;* x86-optimized input routines; does shuffling of packed
   3 ;* YUV formats into individual planes, and converts RGB
   4 ;* into YUV planes also.
   5 ;* Copyright (c) 2012 Ronald S. Bultje <rsbultje@gmail.com>
   6 ;*
   7 ;* This file is part of FFmpeg.
   8 ;*
   9 ;* FFmpeg is free software; you can redistribute it and/or
  10 ;* modify it under the terms of the GNU Lesser General Public
  11 ;* License as published by the Free Software Foundation; either
  12 ;* version 2.1 of the License, or (at your option) any later version.
  13 ;*
  14 ;* FFmpeg is distributed in the hope that it will be useful,
  15 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
  16 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17 ;* Lesser General Public License for more details.
  18 ;*
  19 ;* You should have received a copy of the GNU Lesser General Public
  20 ;* License along with FFmpeg; if not, write to the Free Software
  21 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22 ;******************************************************************************
  23
  24 %include "libavutil/x86/x86util.asm"
  25
  26 SECTION_RODATA
  27
  28 %define RY 0x20DE
  29 %define GY 0x4087
  30 %define BY 0x0C88
  31 %define RU 0xECFF
  32 %define GU 0xDAC8
  33 %define BU 0x3838
  34 %define RV 0x3838
  35 %define GV 0xD0E3
  36 %define BV 0xF6E4
  37
  38 rgb_Yrnd:        times 4 dd 0x80100        ;  16.5 << 15
  39 rgb_UVrnd:       times 4 dd 0x400100       ; 128.5 << 15
  40 %define bgr_Ycoeff_12x4 16*4 + 16* 0 + tableq
  41 %define bgr_Ycoeff_3x56 16*4 + 16* 1 + tableq
  42 %define rgb_Ycoeff_12x4 16*4 + 16* 2 + tableq
  43 %define rgb_Ycoeff_3x56 16*4 + 16* 3 + tableq
  44 %define bgr_Ucoeff_12x4 16*4 + 16* 4 + tableq
  45 %define bgr_Ucoeff_3x56 16*4 + 16* 5 + tableq
  46 %define rgb_Ucoeff_12x4 16*4 + 16* 6 + tableq
  47 %define rgb_Ucoeff_3x56 16*4 + 16* 7 + tableq
  48 %define bgr_Vcoeff_12x4 16*4 + 16* 8 + tableq
  49 %define bgr_Vcoeff_3x56 16*4 + 16* 9 + tableq
  50 %define rgb_Vcoeff_12x4 16*4 + 16*10 + tableq
  51 %define rgb_Vcoeff_3x56 16*4 + 16*11 + tableq
  52
  53 %define rgba_Ycoeff_rb 16*4 + 16*12 + tableq
  54 %define rgba_Ycoeff_br 16*4 + 16*13 + tableq
  55 %define rgba_Ycoeff_ga 16*4 + 16*14 + tableq
  56 %define rgba_Ycoeff_ag 16*4 + 16*15 + tableq
  57 %define rgba_Ucoeff_rb 16*4 + 16*16 + tableq
  58 %define rgba_Ucoeff_br 16*4 + 16*17 + tableq
  59 %define rgba_Ucoeff_ga 16*4 + 16*18 + tableq
  60 %define rgba_Ucoeff_ag 16*4 + 16*19 + tableq
  61 %define rgba_Vcoeff_rb 16*4 + 16*20 + tableq
  62 %define rgba_Vcoeff_br 16*4 + 16*21 + tableq
  63 %define rgba_Vcoeff_ga 16*4 + 16*22 + tableq
  64 %define rgba_Vcoeff_ag 16*4 + 16*23 + tableq
  65
  66 ; bgr_Ycoeff_12x4: times 2 dw BY, GY, 0, BY
  67 ; bgr_Ycoeff_3x56: times 2 dw RY, 0, GY, RY
  68 ; rgb_Ycoeff_12x4: times 2 dw RY, GY, 0, RY
  69 ; rgb_Ycoeff_3x56: times 2 dw BY, 0, GY, BY
  70 ; bgr_Ucoeff_12x4: times 2 dw BU, GU, 0, BU
  71 ; bgr_Ucoeff_3x56: times 2 dw RU, 0, GU, RU
  72 ; rgb_Ucoeff_12x4: times 2 dw RU, GU, 0, RU
  73 ; rgb_Ucoeff_3x56: times 2 dw BU, 0, GU, BU
  74 ; bgr_Vcoeff_12x4: times 2 dw BV, GV, 0, BV
  75 ; bgr_Vcoeff_3x56: times 2 dw RV, 0, GV, RV
  76 ; rgb_Vcoeff_12x4: times 2 dw RV, GV, 0, RV
  77 ; rgb_Vcoeff_3x56: times 2 dw BV, 0, GV, BV
  78
  79 ; rgba_Ycoeff_rb:  times 4 dw RY, BY
  80 ; rgba_Ycoeff_br:  times 4 dw BY, RY
  81 ; rgba_Ycoeff_ga:  times 4 dw GY, 0
  82 ; rgba_Ycoeff_ag:  times 4 dw 0,  GY
  83 ; rgba_Ucoeff_rb:  times 4 dw RU, BU
  84 ; rgba_Ucoeff_br:  times 4 dw BU, RU
  85 ; rgba_Ucoeff_ga:  times 4 dw GU, 0
  86 ; rgba_Ucoeff_ag:  times 4 dw 0,  GU
  87 ; rgba_Vcoeff_rb:  times 4 dw RV, BV
  88 ; rgba_Vcoeff_br:  times 4 dw BV, RV
  89 ; rgba_Vcoeff_ga:  times 4 dw GV, 0
  90 ; rgba_Vcoeff_ag:  times 4 dw 0,  GV
  91
  92 shuf_rgb_12x4:   db 0, 0x80, 1, 0x80,  2, 0x80,  3, 0x80, \
  93                     6, 0x80, 7, 0x80,  8, 0x80,  9, 0x80
  94 shuf_rgb_3x56:   db 2, 0x80, 3, 0x80,  4, 0x80,  5, 0x80, \
  95                     8, 0x80, 9, 0x80, 10, 0x80, 11, 0x80
  96
  97 SECTION .text
  98
  99 ;-----------------------------------------------------------------------------
 100 ; RGB to Y/UV.
 101 ;
 102 ; void <fmt>ToY_<opt>(uint8_t *dst, const uint8_t *src, int w);
 103 ; and
 104 ; void <fmt>toUV_<opt>(uint8_t *dstU, uint8_t *dstV, const uint8_t *src,
 105 ;                      const uint8_t *unused, int w);
 106 ;-----------------------------------------------------------------------------
 107
 108 ; %1 = nr. of XMM registers
 109 ; %2 = rgb or bgr
 110 %macro RGB24_TO_Y_FN 2-3
 111 cglobal %2 %+ 24ToY, 6, 6, %1, dst, src, u1, u2, w, table
 112 %if mmsize == 8
 113     mova           m5, [%2_Ycoeff_12x4]
 114     mova           m6, [%2_Ycoeff_3x56]
 115 %define coeff1 m5
 116 %define coeff2 m6
 117 %elif ARCH_X86_64
 118     mova           m8, [%2_Ycoeff_12x4]
 119     mova           m9, [%2_Ycoeff_3x56]
 120 %define coeff1 m8
 121 %define coeff2 m9
 122 %else ; x86-32 && mmsize == 16
 123 %define coeff1 [%2_Ycoeff_12x4]
 124 %define coeff2 [%2_Ycoeff_3x56]
 125 %endif ; x86-32/64 && mmsize == 8/16
 126 %if (ARCH_X86_64 || mmsize == 8) && %0 == 3
 127     jmp mangle(private_prefix %+ _ %+ %3 %+ 24ToY %+ SUFFIX).body
 128 %else ; (ARCH_X86_64 && %0 == 3) || mmsize == 8
 129 .body:
 130 %if cpuflag(ssse3)
 131     mova           m7, [shuf_rgb_12x4]
 132 %define shuf_rgb1 m7
 133 %if ARCH_X86_64
 134     mova          m10, [shuf_rgb_3x56]
 135 %define shuf_rgb2 m10
 136 %else ; x86-32
 137 %define shuf_rgb2 [shuf_rgb_3x56]
 138 %endif ; x86-32/64
 139 %endif ; cpuflag(ssse3)
 140 %if ARCH_X86_64
 141     movsxd         wq, wd
 142 %endif
 143     add            wq, wq
 144     add          dstq, wq
 145     neg            wq
 146 %if notcpuflag(ssse3)
 147     pxor           m7, m7
 148 %endif ; !cpuflag(ssse3)
 149     mova           m4, [rgb_Yrnd]
 150 .loop:
 151 %if cpuflag(ssse3)
 152     movu           m0, [srcq+0]           ; (byte) { Bx, Gx, Rx }[0-3]
 153     movu           m2, [srcq+12]          ; (byte) { Bx, Gx, Rx }[4-7]
 154     pshufb         m1, m0, shuf_rgb2      ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
 155     pshufb         m0, shuf_rgb1          ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
 156     pshufb         m3, m2, shuf_rgb2      ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
 157     pshufb         m2, shuf_rgb1          ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
 158 %else ; !cpuflag(ssse3)
 159     movd           m0, [srcq+0]           ; (byte) { B0, G0, R0, B1 }
 160     movd           m1, [srcq+2]           ; (byte) { R0, B1, G1, R1 }
 161     movd           m2, [srcq+6]           ; (byte) { B2, G2, R2, B3 }
 162     movd           m3, [srcq+8]           ; (byte) { R2, B3, G3, R3 }
 163 %if mmsize == 16 ; i.e. sse2
 164     punpckldq      m0, m2                 ; (byte) { B0, G0, R0, B1, B2, G2, R2, B3 }
 165     punpckldq      m1, m3                 ; (byte) { R0, B1, G1, R1, R2, B3, G3, R3 }
 166     movd           m2, [srcq+12]          ; (byte) { B4, G4, R4, B5 }
 167     movd           m3, [srcq+14]          ; (byte) { R4, B5, G5, R5 }
 168     movd           m5, [srcq+18]          ; (byte) { B6, G6, R6, B7 }
 169     movd           m6, [srcq+20]          ; (byte) { R6, B7, G7, R7 }
 170     punpckldq      m2, m5                 ; (byte) { B4, G4, R4, B5, B6, G6, R6, B7 }
 171     punpckldq      m3, m6                 ; (byte) { R4, B5, G5, R5, R6, B7, G7, R7 }
 172 %endif ; mmsize == 16
 173     punpcklbw      m0, m7                 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
 174     punpcklbw      m1, m7                 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
 175     punpcklbw      m2, m7                 ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
 176     punpcklbw      m3, m7                 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
 177 %endif ; cpuflag(ssse3)
 178     add          srcq, 3 * mmsize / 2
 179     pmaddwd        m0, coeff1             ; (dword) { B0*BY + G0*GY, B1*BY, B2*BY + G2*GY, B3*BY }
 180     pmaddwd        m1, coeff2             ; (dword) { R0*RY, G1+GY + R1*RY, R2*RY, G3+GY + R3*RY }
 181     pmaddwd        m2, coeff1             ; (dword) { B4*BY + G4*GY, B5*BY, B6*BY + G6*GY, B7*BY }
 182     pmaddwd        m3, coeff2             ; (dword) { R4*RY, G5+GY + R5*RY, R6*RY, G7+GY + R7*RY }
 183     paddd          m0, m1                 ; (dword) { Bx*BY + Gx*GY + Rx*RY }[0-3]
 184     paddd          m2, m3                 ; (dword) { Bx*BY + Gx*GY + Rx*RY }[4-7]
 185     paddd          m0, m4                 ; += rgb_Yrnd, i.e. (dword) { Y[0-3] }
 186     paddd          m2, m4                 ; += rgb_Yrnd, i.e. (dword) { Y[4-7] }
 187     psrad          m0, 9
 188     psrad          m2, 9
 189     packssdw       m0, m2                 ; (word) { Y[0-7] }
 190     mova    [dstq+wq], m0
 191     add            wq, mmsize
 192     jl .loop
 193     REP_RET
 194 %endif ; (ARCH_X86_64 && %0 == 3) || mmsize == 8
 195 %endmacro
 196
 197 ; %1 = nr. of XMM registers
 198 ; %2 = rgb or bgr
 199 %macro RGB24_TO_UV_FN 2-3
 200 cglobal %2 %+ 24ToUV, 7, 7, %1, dstU, dstV, u1, src, u2, w, table
 201 %if ARCH_X86_64
 202     mova           m8, [%2_Ucoeff_12x4]
 203     mova           m9, [%2_Ucoeff_3x56]
 204     mova          m10, [%2_Vcoeff_12x4]
 205     mova          m11, [%2_Vcoeff_3x56]
 206 %define coeffU1 m8
 207 %define coeffU2 m9
 208 %define coeffV1 m10
 209 %define coeffV2 m11
 210 %else ; x86-32
 211 %define coeffU1 [%2_Ucoeff_12x4]
 212 %define coeffU2 [%2_Ucoeff_3x56]
 213 %define coeffV1 [%2_Vcoeff_12x4]
 214 %define coeffV2 [%2_Vcoeff_3x56]
 215 %endif ; x86-32/64
 216 %if ARCH_X86_64 && %0 == 3
 217     jmp mangle(private_prefix %+ _ %+ %3 %+ 24ToUV %+ SUFFIX).body
 218 %else ; ARCH_X86_64 && %0 == 3
 219 .body:
 220 %if cpuflag(ssse3)
 221     mova           m7, [shuf_rgb_12x4]
 222 %define shuf_rgb1 m7
 223 %if ARCH_X86_64
 224     mova          m12, [shuf_rgb_3x56]
 225 %define shuf_rgb2 m12
 226 %else ; x86-32
 227 %define shuf_rgb2 [shuf_rgb_3x56]
 228 %endif ; x86-32/64
 229 %endif ; cpuflag(ssse3)
 230 %if ARCH_X86_64
 231     movsxd         wq, dword r5m
 232 %else ; x86-32
 233     mov            wq, r5m
 234 %endif
 235     add            wq, wq
 236     add         dstUq, wq
 237     add         dstVq, wq
 238     neg            wq
 239     mova           m6, [rgb_UVrnd]
 240 %if notcpuflag(ssse3)
 241     pxor           m7, m7
 242 %endif
 243 .loop:
 244 %if cpuflag(ssse3)
 245     movu           m0, [srcq+0]           ; (byte) { Bx, Gx, Rx }[0-3]
 246     movu           m4, [srcq+12]          ; (byte) { Bx, Gx, Rx }[4-7]
 247     pshufb         m1, m0, shuf_rgb2      ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
 248     pshufb         m0, shuf_rgb1          ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
 249 %else ; !cpuflag(ssse3)
 250     movd           m0, [srcq+0]           ; (byte) { B0, G0, R0, B1 }
 251     movd           m1, [srcq+2]           ; (byte) { R0, B1, G1, R1 }
 252     movd           m4, [srcq+6]           ; (byte) { B2, G2, R2, B3 }
 253     movd           m5, [srcq+8]           ; (byte) { R2, B3, G3, R3 }
 254 %if mmsize == 16
 255     punpckldq      m0, m4                 ; (byte) { B0, G0, R0, B1, B2, G2, R2, B3 }
 256     punpckldq      m1, m5                 ; (byte) { R0, B1, G1, R1, R2, B3, G3, R3 }
 257     movd           m4, [srcq+12]          ; (byte) { B4, G4, R4, B5 }
 258     movd           m5, [srcq+14]          ; (byte) { R4, B5, G5, R5 }
 259 %endif ; mmsize == 16
 260     punpcklbw      m0, m7                 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
 261     punpcklbw      m1, m7                 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
 262 %endif ; cpuflag(ssse3)
 263     pmaddwd        m2, m0, coeffV1        ; (dword) { B0*BV + G0*GV, B1*BV, B2*BV + G2*GV, B3*BV }
 264     pmaddwd        m3, m1, coeffV2        ; (dword) { R0*BV, G1*GV + R1*BV, R2*BV, G3*GV + R3*BV }
 265     pmaddwd        m0, coeffU1            ; (dword) { B0*BU + G0*GU, B1*BU, B2*BU + G2*GU, B3*BU }
 266     pmaddwd        m1, coeffU2            ; (dword) { R0*BU, G1*GU + R1*BU, R2*BU, G3*GU + R3*BU }
 267     paddd          m0, m1                 ; (dword) { Bx*BU + Gx*GU + Rx*RU }[0-3]
 268     paddd          m2, m3                 ; (dword) { Bx*BV + Gx*GV + Rx*RV }[0-3]
 269 %if cpuflag(ssse3)
 270     pshufb         m5, m4, shuf_rgb2      ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
 271     pshufb         m4, shuf_rgb1          ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
 272 %else ; !cpuflag(ssse3)
 273 %if mmsize == 16
 274     movd           m1, [srcq+18]          ; (byte) { B6, G6, R6, B7 }
 275     movd           m3, [srcq+20]          ; (byte) { R6, B7, G7, R7 }
 276     punpckldq      m4, m1                 ; (byte) { B4, G4, R4, B5, B6, G6, R6, B7 }
 277     punpckldq      m5, m3                 ; (byte) { R4, B5, G5, R5, R6, B7, G7, R7 }
 278 %endif ; mmsize == 16 && !cpuflag(ssse3)
 279     punpcklbw      m4, m7                 ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
 280     punpcklbw      m5, m7                 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
 281 %endif ; cpuflag(ssse3)
 282     add          srcq, 3 * mmsize / 2
 283     pmaddwd        m1, m4, coeffU1        ; (dword) { B4*BU + G4*GU, B5*BU, B6*BU + G6*GU, B7*BU }
 284     pmaddwd        m3, m5, coeffU2        ; (dword) { R4*BU, G5*GU + R5*BU, R6*BU, G7*GU + R7*BU }
 285     pmaddwd        m4, coeffV1            ; (dword) { B4*BV + G4*GV, B5*BV, B6*BV + G6*GV, B7*BV }
 286     pmaddwd        m5, coeffV2            ; (dword) { R4*BV, G5*GV + R5*BV, R6*BV, G7*GV + R7*BV }
 287     paddd          m1, m3                 ; (dword) { Bx*BU + Gx*GU + Rx*RU }[4-7]
 288     paddd          m4, m5                 ; (dword) { Bx*BV + Gx*GV + Rx*RV }[4-7]
 289     paddd          m0, m6                 ; += rgb_UVrnd, i.e. (dword) { U[0-3] }
 290     paddd          m2, m6                 ; += rgb_UVrnd, i.e. (dword) { V[0-3] }
 291     paddd          m1, m6                 ; += rgb_UVrnd, i.e. (dword) { U[4-7] }
 292     paddd          m4, m6                 ; += rgb_UVrnd, i.e. (dword) { V[4-7] }
 293     psrad          m0, 9
 294     psrad          m2, 9
 295     psrad          m1, 9
 296     psrad          m4, 9
 297     packssdw       m0, m1                 ; (word) { U[0-7] }
 298     packssdw       m2, m4                 ; (word) { V[0-7] }
 299 %if mmsize == 8
 300     mova   [dstUq+wq], m0
 301     mova   [dstVq+wq], m2
 302 %else ; mmsize == 16
 303     mova   [dstUq+wq], m0
 304     mova   [dstVq+wq], m2
 305 %endif ; mmsize == 8/16
 306     add            wq, mmsize
 307     jl .loop
 308     REP_RET
 309 %endif ; ARCH_X86_64 && %0 == 3
 310 %endmacro
 311
 312 ; %1 = nr. of XMM registers for rgb-to-Y func
 313 ; %2 = nr. of XMM registers for rgb-to-UV func
 314 %macro RGB24_FUNCS 2
 315 RGB24_TO_Y_FN %1, rgb
 316 RGB24_TO_Y_FN %1, bgr, rgb
 317 RGB24_TO_UV_FN %2, rgb
 318 RGB24_TO_UV_FN %2, bgr, rgb
 319 %endmacro
 320
 321 %if ARCH_X86_32
 322 INIT_MMX mmx
 323 RGB24_FUNCS 0, 0
 324 %endif
 325
 326 INIT_XMM sse2
 327 RGB24_FUNCS 10, 12
 328
 329 INIT_XMM ssse3
 330 RGB24_FUNCS 11, 13
 331
 332 %if HAVE_AVX_EXTERNAL
 333 INIT_XMM avx
 334 RGB24_FUNCS 11, 13
 335 %endif
 336
 337 ; %1 = nr. of XMM registers
 338 ; %2-5 = rgba, bgra, argb or abgr (in individual characters)
 339 %macro RGB32_TO_Y_FN 5-6
 340 cglobal %2%3%4%5 %+ ToY, 6, 6, %1, dst, src, u1, u2, w, table
 341     mova           m5, [rgba_Ycoeff_%2%4]
 342     mova           m6, [rgba_Ycoeff_%3%5]
 343 %if %0 == 6
 344     jmp mangle(private_prefix %+ _ %+ %6 %+ ToY %+ SUFFIX).body
 345 %else ; %0 == 6
 346 .body:
 347 %if ARCH_X86_64
 348     movsxd         wq, wd
 349 %endif
 350     add            wq, wq
 351     sub            wq, mmsize - 1
 352     lea          srcq, [srcq+wq*2]
 353     add          dstq, wq
 354     neg            wq
 355     mova           m4, [rgb_Yrnd]
 356     pcmpeqb        m7, m7
 357     psrlw          m7, 8                  ; (word) { 0x00ff } x4
 358 .loop:
 359     ; FIXME check alignment and use mova
 360     movu           m0, [srcq+wq*2+0]      ; (byte) { Bx, Gx, Rx, xx }[0-3]
 361     movu           m2, [srcq+wq*2+mmsize] ; (byte) { Bx, Gx, Rx, xx }[4-7]
 362     DEINTB          1,  0,  3,  2,  7     ; (word) { Gx, xx (m0/m2) or Bx, Rx (m1/m3) }[0-3]/[4-7]
 363     pmaddwd        m1, m5                 ; (dword) { Bx*BY + Rx*RY }[0-3]
 364     pmaddwd        m0, m6                 ; (dword) { Gx*GY }[0-3]
 365     pmaddwd        m3, m5                 ; (dword) { Bx*BY + Rx*RY }[4-7]
 366     pmaddwd        m2, m6                 ; (dword) { Gx*GY }[4-7]
 367     paddd          m0, m4                 ; += rgb_Yrnd
 368     paddd          m2, m4                 ; += rgb_Yrnd
 369     paddd          m0, m1                 ; (dword) { Y[0-3] }
 370     paddd          m2, m3                 ; (dword) { Y[4-7] }
 371     psrad          m0, 9
 372     psrad          m2, 9
 373     packssdw       m0, m2                 ; (word) { Y[0-7] }
 374     mova    [dstq+wq], m0
 375     add            wq, mmsize
 376     jl .loop
 377     sub            wq, mmsize - 1
 378     jz .end
 379     add            srcq, 2*mmsize - 2
 380     add            dstq, mmsize - 1
 381 .loop2:
 382     movd           m0, [srcq+wq*2+0]      ; (byte) { Bx, Gx, Rx, xx }[0-3]
 383     DEINTB          1,  0,  3,  2,  7     ; (word) { Gx, xx (m0/m2) or Bx, Rx (m1/m3) }[0-3]/[4-7]
 384     pmaddwd        m1, m5                 ; (dword) { Bx*BY + Rx*RY }[0-3]
 385     pmaddwd        m0, m6                 ; (dword) { Gx*GY }[0-3]
 386     paddd          m0, m4                 ; += rgb_Yrnd
 387     paddd          m0, m1                 ; (dword) { Y[0-3] }
 388     psrad          m0, 9
 389     packssdw       m0, m0                 ; (word) { Y[0-7] }
 390     movd    [dstq+wq], m0
 391     add            wq, 2
 392     jl .loop2
 393 .end:
 394     REP_RET
 395 %endif ; %0 == 3
 396 %endmacro
 397
 398 ; %1 = nr. of XMM registers
 399 ; %2-5 = rgba, bgra, argb or abgr (in individual characters)
 400 %macro RGB32_TO_UV_FN 5-6
 401 cglobal %2%3%4%5 %+ ToUV, 7, 7, %1, dstU, dstV, u1, src, u2, w, table
 402 %if ARCH_X86_64
 403     mova           m8, [rgba_Ucoeff_%2%4]
 404     mova           m9, [rgba_Ucoeff_%3%5]
 405     mova          m10, [rgba_Vcoeff_%2%4]
 406     mova          m11, [rgba_Vcoeff_%3%5]
 407 %define coeffU1 m8
 408 %define coeffU2 m9
 409 %define coeffV1 m10
 410 %define coeffV2 m11
 411 %else ; x86-32
 412 %define coeffU1 [rgba_Ucoeff_%2%4]
 413 %define coeffU2 [rgba_Ucoeff_%3%5]
 414 %define coeffV1 [rgba_Vcoeff_%2%4]
 415 %define coeffV2 [rgba_Vcoeff_%3%5]
 416 %endif ; x86-64/32
 417 %if ARCH_X86_64 && %0 == 6
 418     jmp mangle(private_prefix %+ _ %+ %6 %+ ToUV %+ SUFFIX).body
 419 %else ; ARCH_X86_64 && %0 == 6
 420 .body:
 421 %if ARCH_X86_64
 422     movsxd         wq, dword r5m
 423 %else ; x86-32
 424     mov            wq, r5m
 425 %endif
 426     add            wq, wq
 427     sub            wq, mmsize - 1
 428     add         dstUq, wq
 429     add         dstVq, wq
 430     lea          srcq, [srcq+wq*2]
 431     neg            wq
 432     pcmpeqb        m7, m7
 433     psrlw          m7, 8                  ; (word) { 0x00ff } x4
 434     mova           m6, [rgb_UVrnd]
 435 .loop:
 436     ; FIXME check alignment and use mova
 437     movu           m0, [srcq+wq*2+0]      ; (byte) { Bx, Gx, Rx, xx }[0-3]
 438     movu           m4, [srcq+wq*2+mmsize] ; (byte) { Bx, Gx, Rx, xx }[4-7]
 439     DEINTB          1,  0,  5,  4,  7     ; (word) { Gx, xx (m0/m4) or Bx, Rx (m1/m5) }[0-3]/[4-7]
 440     pmaddwd        m3, m1, coeffV1        ; (dword) { Bx*BV + Rx*RV }[0-3]
 441     pmaddwd        m2, m0, coeffV2        ; (dword) { Gx*GV }[0-3]
 442     pmaddwd        m1, coeffU1            ; (dword) { Bx*BU + Rx*RU }[0-3]
 443     pmaddwd        m0, coeffU2            ; (dword) { Gx*GU }[0-3]
 444     paddd          m3, m6                 ; += rgb_UVrnd
 445     paddd          m1, m6                 ; += rgb_UVrnd
 446     paddd          m2, m3                 ; (dword) { V[0-3] }
 447     paddd          m0, m1                 ; (dword) { U[0-3] }
 448     pmaddwd        m3, m5, coeffV1        ; (dword) { Bx*BV + Rx*RV }[4-7]
 449     pmaddwd        m1, m4, coeffV2        ; (dword) { Gx*GV }[4-7]
 450     pmaddwd        m5, coeffU1            ; (dword) { Bx*BU + Rx*RU }[4-7]
 451     pmaddwd        m4, coeffU2            ; (dword) { Gx*GU }[4-7]
 452     paddd          m3, m6                 ; += rgb_UVrnd
 453     paddd          m5, m6                 ; += rgb_UVrnd
 454     psrad          m0, 9
 455     paddd          m1, m3                 ; (dword) { V[4-7] }
 456     paddd          m4, m5                 ; (dword) { U[4-7] }
 457     psrad          m2, 9
 458     psrad          m4, 9
 459     psrad          m1, 9
 460     packssdw       m0, m4                 ; (word) { U[0-7] }
 461     packssdw       m2, m1                 ; (word) { V[0-7] }
 462 %if mmsize == 8
 463     mova   [dstUq+wq], m0
 464     mova   [dstVq+wq], m2
 465 %else ; mmsize == 16
 466     mova   [dstUq+wq], m0
 467     mova   [dstVq+wq], m2
 468 %endif ; mmsize == 8/16
 469     add            wq, mmsize
 470     jl .loop
 471     sub            wq, mmsize - 1
 472     jz .end
 473     add            srcq , 2*mmsize - 2
 474     add            dstUq, mmsize - 1
 475     add            dstVq, mmsize - 1
 476 .loop2:
 477     movd           m0, [srcq+wq*2]        ; (byte) { Bx, Gx, Rx, xx }[0-3]
 478     DEINTB          1,  0,  5,  4,  7     ; (word) { Gx, xx (m0/m4) or Bx, Rx (m1/m5) }[0-3]/[4-7]
 479     pmaddwd        m3, m1, coeffV1        ; (dword) { Bx*BV + Rx*RV }[0-3]
 480     pmaddwd        m2, m0, coeffV2        ; (dword) { Gx*GV }[0-3]
 481     pmaddwd        m1, coeffU1            ; (dword) { Bx*BU + Rx*RU }[0-3]
 482     pmaddwd        m0, coeffU2            ; (dword) { Gx*GU }[0-3]
 483     paddd          m3, m6                 ; += rgb_UVrnd
 484     paddd          m1, m6                 ; += rgb_UVrnd
 485     paddd          m2, m3                 ; (dword) { V[0-3] }
 486     paddd          m0, m1                 ; (dword) { U[0-3] }
 487     psrad          m0, 9
 488     psrad          m2, 9
 489     packssdw       m0, m0                 ; (word) { U[0-7] }
 490     packssdw       m2, m2                 ; (word) { V[0-7] }
 491     movd   [dstUq+wq], m0
 492     movd   [dstVq+wq], m2
 493     add            wq, 2
 494     jl .loop2
 495 .end:
 496     REP_RET
 497 %endif ; ARCH_X86_64 && %0 == 3
 498 %endmacro
 499
 500 ; %1 = nr. of XMM registers for rgb-to-Y func
 501 ; %2 = nr. of XMM registers for rgb-to-UV func
 502 %macro RGB32_FUNCS 2
 503 RGB32_TO_Y_FN %1, r, g, b, a
 504 RGB32_TO_Y_FN %1, b, g, r, a, rgba
 505 RGB32_TO_Y_FN %1, a, r, g, b, rgba
 506 RGB32_TO_Y_FN %1, a, b, g, r, rgba
 507
 508 RGB32_TO_UV_FN %2, r, g, b, a
 509 RGB32_TO_UV_FN %2, b, g, r, a, rgba
 510 RGB32_TO_UV_FN %2, a, r, g, b, rgba
 511 RGB32_TO_UV_FN %2, a, b, g, r, rgba
 512 %endmacro
 513
 514 %if ARCH_X86_32
 515 INIT_MMX mmx
 516 RGB32_FUNCS 0, 0
 517 %endif
 518
 519 INIT_XMM sse2
 520 RGB32_FUNCS 8, 12
 521
 522 %if HAVE_AVX_EXTERNAL
 523 INIT_XMM avx
 524 RGB32_FUNCS 8, 12
 525 %endif
 526
 527 ;-----------------------------------------------------------------------------
 528 ; YUYV/UYVY/NV12/NV21 packed pixel shuffling.
 529 ;
 530 ; void <fmt>ToY_<opt>(uint8_t *dst, const uint8_t *src, int w);
 531 ; and
 532 ; void <fmt>toUV_<opt>(uint8_t *dstU, uint8_t *dstV, const uint8_t *src,
 533 ;                      const uint8_t *unused, int w);
 534 ;-----------------------------------------------------------------------------
 535
 536 ; %1 = a (aligned) or u (unaligned)
 537 ; %2 = yuyv or uyvy
 538 %macro LOOP_YUYV_TO_Y 2
 539 .loop_%1:
 540     mov%1          m0, [srcq+wq*2]        ; (byte) { Y0, U0, Y1, V0, ... }
 541     mov%1          m1, [srcq+wq*2+mmsize] ; (byte) { Y8, U4, Y9, V4, ... }
 542 %ifidn %2, yuyv
 543     pand           m0, m2                 ; (word) { Y0, Y1, ..., Y7 }
 544     pand           m1, m2                 ; (word) { Y8, Y9, ..., Y15 }
 545 %else ; uyvy
 546     psrlw          m0, 8                  ; (word) { Y0, Y1, ..., Y7 }
 547     psrlw          m1, 8                  ; (word) { Y8, Y9, ..., Y15 }
 548 %endif ; yuyv/uyvy
 549     packuswb       m0, m1                 ; (byte) { Y0, ..., Y15 }
 550     mova    [dstq+wq], m0
 551     add            wq, mmsize
 552     jl .loop_%1
 553     REP_RET
 554 %endmacro
 555
 556 ; %1 = nr. of XMM registers
 557 ; %2 = yuyv or uyvy
 558 ; %3 = if specified, it means that unaligned and aligned code in loop
 559 ;      will be the same (i.e. YUYV+AVX), and thus we don't need to
 560 ;      split the loop in an aligned and unaligned case
 561 %macro YUYV_TO_Y_FN 2-3
 562 cglobal %2ToY, 5, 5, %1, dst, unused0, unused1, src, w
 563 %if ARCH_X86_64
 564     movsxd         wq, wd
 565 %endif
 566     add          dstq, wq
 567 %if mmsize == 16
 568     test         srcq, 15
 569 %endif
 570     lea          srcq, [srcq+wq*2]
 571 %ifidn %2, yuyv
 572     pcmpeqb        m2, m2                 ; (byte) { 0xff } x 16
 573     psrlw          m2, 8                  ; (word) { 0x00ff } x 8
 574 %endif ; yuyv
 575 %if mmsize == 16
 576     jnz .loop_u_start
 577     neg            wq
 578     LOOP_YUYV_TO_Y  a, %2
 579 .loop_u_start:
 580     neg            wq
 581     LOOP_YUYV_TO_Y  u, %2
 582 %else ; mmsize == 8
 583     neg            wq
 584     LOOP_YUYV_TO_Y  a, %2
 585 %endif ; mmsize == 8/16
 586 %endmacro
 587
 588 ; %1 = a (aligned) or u (unaligned)
 589 ; %2 = yuyv or uyvy
 590 %macro LOOP_YUYV_TO_UV 2
 591 .loop_%1:
 592 %ifidn %2, yuyv
 593     mov%1          m0, [srcq+wq*4]        ; (byte) { Y0, U0, Y1, V0, ... }
 594     mov%1          m1, [srcq+wq*4+mmsize] ; (byte) { Y8, U4, Y9, V4, ... }
 595     psrlw          m0, 8                  ; (word) { U0, V0, ..., U3, V3 }
 596     psrlw          m1, 8                  ; (word) { U4, V4, ..., U7, V7 }
 597 %else ; uyvy
 598 %if cpuflag(avx)
 599     vpand          m0, m2, [srcq+wq*4]        ; (word) { U0, V0, ..., U3, V3 }
 600     vpand          m1, m2, [srcq+wq*4+mmsize] ; (word) { U4, V4, ..., U7, V7 }
 601 %else
 602     mov%1          m0, [srcq+wq*4]        ; (byte) { Y0, U0, Y1, V0, ... }
 603     mov%1          m1, [srcq+wq*4+mmsize] ; (byte) { Y8, U4, Y9, V4, ... }
 604     pand           m0, m2                 ; (word) { U0, V0, ..., U3, V3 }
 605     pand           m1, m2                 ; (word) { U4, V4, ..., U7, V7 }
 606 %endif
 607 %endif ; yuyv/uyvy
 608     packuswb       m0, m1                 ; (byte) { U0, V0, ..., U7, V7 }
 609     pand           m1, m0, m2             ; (word) { U0, U1, ..., U7 }
 610     psrlw          m0, 8                  ; (word) { V0, V1, ..., V7 }
 611 %if mmsize == 16
 612     packuswb       m1, m0                 ; (byte) { U0, ... U7, V1, ... V7 }
 613     movh   [dstUq+wq], m1
 614     movhps [dstVq+wq], m1
 615 %else ; mmsize == 8
 616     packuswb       m1, m1                 ; (byte) { U0, ... U3 }
 617     packuswb       m0, m0                 ; (byte) { V0, ... V3 }
 618     movh   [dstUq+wq], m1
 619     movh   [dstVq+wq], m0
 620 %endif ; mmsize == 8/16
 621     add            wq, mmsize / 2
 622     jl .loop_%1
 623     REP_RET
 624 %endmacro
 625
 626 ; %1 = nr. of XMM registers
 627 ; %2 = yuyv or uyvy
 628 ; %3 = if specified, it means that unaligned and aligned code in loop
 629 ;      will be the same (i.e. UYVY+AVX), and thus we don't need to
 630 ;      split the loop in an aligned and unaligned case
 631 %macro YUYV_TO_UV_FN 2-3
 632 cglobal %2ToUV, 4, 5, %1, dstU, dstV, unused, src, w
 633 %if ARCH_X86_64
 634     movsxd         wq, dword r5m
 635 %else ; x86-32
 636     mov            wq, r5m
 637 %endif
 638     add         dstUq, wq
 639     add         dstVq, wq
 640 %if mmsize == 16 && %0 == 2
 641     test         srcq, 15
 642 %endif
 643     lea          srcq, [srcq+wq*4]
 644     pcmpeqb        m2, m2                 ; (byte) { 0xff } x 16
 645     psrlw          m2, 8                  ; (word) { 0x00ff } x 8
 646     ; NOTE: if uyvy+avx, u/a are identical
 647 %if mmsize == 16 && %0 == 2
 648     jnz .loop_u_start
 649     neg            wq
 650     LOOP_YUYV_TO_UV a, %2
 651 .loop_u_start:
 652     neg            wq
 653     LOOP_YUYV_TO_UV u, %2
 654 %else ; mmsize == 8
 655     neg            wq
 656     LOOP_YUYV_TO_UV a, %2
 657 %endif ; mmsize == 8/16
 658 %endmacro
 659
 660 ; %1 = a (aligned) or u (unaligned)
 661 ; %2 = nv12 or nv21
 662 %macro LOOP_NVXX_TO_UV 2
 663 .loop_%1:
 664     mov%1          m0, [srcq+wq*2]        ; (byte) { U0, V0, U1, V1, ... }
 665     mov%1          m1, [srcq+wq*2+mmsize] ; (byte) { U8, V8, U9, V9, ... }
 666     pand           m2, m0, m5             ; (word) { U0, U1, ..., U7 }
 667     pand           m3, m1, m5             ; (word) { U8, U9, ..., U15 }
 668     psrlw          m0, 8                  ; (word) { V0, V1, ..., V7 }
 669     psrlw          m1, 8                  ; (word) { V8, V9, ..., V15 }
 670     packuswb       m2, m3                 ; (byte) { U0, ..., U15 }
 671     packuswb       m0, m1                 ; (byte) { V0, ..., V15 }
 672 %ifidn %2, nv12
 673     mova   [dstUq+wq], m2
 674     mova   [dstVq+wq], m0
 675 %else ; nv21
 676     mova   [dstVq+wq], m2
 677     mova   [dstUq+wq], m0
 678 %endif ; nv12/21
 679     add            wq, mmsize
 680     jl .loop_%1
 681     REP_RET
 682 %endmacro
 683
 684 ; %1 = nr. of XMM registers
 685 ; %2 = nv12 or nv21
 686 %macro NVXX_TO_UV_FN 2
 687 cglobal %2ToUV, 4, 5, %1, dstU, dstV, unused, src, w
 688 %if ARCH_X86_64
 689     movsxd         wq, dword r5m
 690 %else ; x86-32
 691     mov            wq, r5m
 692 %endif
 693     add         dstUq, wq
 694     add         dstVq, wq
 695 %if mmsize == 16
 696     test         srcq, 15
 697 %endif
 698     lea          srcq, [srcq+wq*2]
 699     pcmpeqb        m5, m5                 ; (byte) { 0xff } x 16
 700     psrlw          m5, 8                  ; (word) { 0x00ff } x 8
 701 %if mmsize == 16
 702     jnz .loop_u_start
 703     neg            wq
 704     LOOP_NVXX_TO_UV a, %2
 705 .loop_u_start:
 706     neg            wq
 707     LOOP_NVXX_TO_UV u, %2
 708 %else ; mmsize == 8
 709     neg            wq
 710     LOOP_NVXX_TO_UV a, %2
 711 %endif ; mmsize == 8/16
 712 %endmacro
 713
 714 %if ARCH_X86_32
 715 INIT_MMX mmx
 716 YUYV_TO_Y_FN  0, yuyv
 717 YUYV_TO_Y_FN  0, uyvy
 718 YUYV_TO_UV_FN 0, yuyv
 719 YUYV_TO_UV_FN 0, uyvy
 720 NVXX_TO_UV_FN 0, nv12
 721 NVXX_TO_UV_FN 0, nv21
 722 %endif
 723
 724 INIT_XMM sse2
 725 YUYV_TO_Y_FN  3, yuyv
 726 YUYV_TO_Y_FN  2, uyvy
 727 YUYV_TO_UV_FN 3, yuyv
 728 YUYV_TO_UV_FN 3, uyvy
 729 NVXX_TO_UV_FN 5, nv12
 730 NVXX_TO_UV_FN 5, nv21
 731
 732 %if HAVE_AVX_EXTERNAL
 733 INIT_XMM avx
 734 ; in theory, we could write a yuy2-to-y using vpand (i.e. AVX), but
 735 ; that's not faster in practice
 736 YUYV_TO_UV_FN 3, yuyv
 737 YUYV_TO_UV_FN 3, uyvy, 1
 738 NVXX_TO_UV_FN 5, nv12
 739 NVXX_TO_UV_FN 5, nv21
 740 %endif