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>
7 ;* This file is part of Libav.
9 ;* Libav 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.
14 ;* Libav 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.
19 ;* You should have received a copy of the GNU Lesser General Public
20 ;* License along with Libav; if not, write to the Free Software
21 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 ;******************************************************************************
25 %include "x86util.asm"
39 rgb_Yrnd: times 4 dd 0x84000 ; 16.5 << 15
40 rgb_UVrnd: times 4 dd 0x404000 ; 128.5 << 15
41 bgr_Ycoeff_12x4: times 2 dw BY, GY, 0, BY
42 bgr_Ycoeff_3x56: times 2 dw RY, 0, GY, RY
43 rgb_Ycoeff_12x4: times 2 dw RY, GY, 0, RY
44 rgb_Ycoeff_3x56: times 2 dw BY, 0, GY, BY
45 bgr_Ucoeff_12x4: times 2 dw BU, GU, 0, BU
46 bgr_Ucoeff_3x56: times 2 dw RU, 0, GU, RU
47 rgb_Ucoeff_12x4: times 2 dw RU, GU, 0, RU
48 rgb_Ucoeff_3x56: times 2 dw BU, 0, GU, BU
49 bgr_Vcoeff_12x4: times 2 dw BV, GV, 0, BV
50 bgr_Vcoeff_3x56: times 2 dw RV, 0, GV, RV
51 rgb_Vcoeff_12x4: times 2 dw RV, GV, 0, RV
52 rgb_Vcoeff_3x56: times 2 dw BV, 0, GV, BV
54 shuf_rgb_12x4: db 0, 0x80, 1, 0x80, 2, 0x80, 3, 0x80, \
55 6, 0x80, 7, 0x80, 8, 0x80, 9, 0x80
56 shuf_rgb_3x56: db 2, 0x80, 3, 0x80, 4, 0x80, 5, 0x80, \
57 8, 0x80, 9, 0x80, 10, 0x80, 11, 0x80
61 ;-----------------------------------------------------------------------------
64 ; void <fmt>ToY_<opt>(uint8_t *dst, const uint8_t *src, int w);
66 ; void <fmt>toUV_<opt>(uint8_t *dstU, uint8_t *dstV, const uint8_t *src,
67 ; const uint8_t *unused, int w);
68 ;-----------------------------------------------------------------------------
70 ; %1 = nr. of XMM registers
72 %macro RGB24_TO_Y_FN 2-3
73 cglobal %2 %+ 24ToY, 3, 3, %1, dst, src, w
75 mova m5, [%2_Ycoeff_12x4]
76 mova m6, [%2_Ycoeff_3x56]
80 mova m8, [%2_Ycoeff_12x4]
81 mova m9, [%2_Ycoeff_3x56]
84 %else ; x86-32 && mmsize == 16
85 %define coeff1 [%2_Ycoeff_12x4]
86 %define coeff2 [%2_Ycoeff_3x56]
87 %endif ; x86-32/64 && mmsize == 8/16
88 %if (ARCH_X86_64 || mmsize == 8) && %0 == 3
89 jmp mangle(program_name %+ _ %+ %3 %+ 24ToY %+ SUFFIX).body
90 %else ; (ARCH_X86_64 && %0 == 3) || mmsize == 8
93 mova m7, [shuf_rgb_12x4]
96 mova m10, [shuf_rgb_3x56]
99 %define shuf_rgb2 [shuf_rgb_3x56]
101 %endif ; cpuflag(ssse3)
107 %if notcpuflag(ssse3)
109 %endif ; !cpuflag(ssse3)
113 movu m0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3]
114 movu m2, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7]
115 pshufb m1, m0, shuf_rgb2 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
116 pshufb m0, shuf_rgb1 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
117 pshufb m3, m2, shuf_rgb2 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
118 pshufb m2, shuf_rgb1 ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
119 %else ; !cpuflag(ssse3)
120 movd m0, [srcq+0] ; (byte) { B0, G0, R0, B1 }
121 movd m1, [srcq+2] ; (byte) { R0, B1, G1, R1 }
122 movd m2, [srcq+6] ; (byte) { B2, G2, R2, B3 }
123 movd m3, [srcq+8] ; (byte) { R2, B3, G3, R3 }
124 %if mmsize == 16 ; i.e. sse2
125 punpckldq m0, m2 ; (byte) { B0, G0, R0, B1, B2, G2, R2, B3 }
126 punpckldq m1, m3 ; (byte) { R0, B1, G1, R1, R2, B3, G3, R3 }
127 movd m2, [srcq+12] ; (byte) { B4, G4, R4, B5 }
128 movd m3, [srcq+14] ; (byte) { R4, B5, G5, R5 }
129 movd m5, [srcq+18] ; (byte) { B6, G6, R6, B7 }
130 movd m6, [srcq+20] ; (byte) { R6, B7, G7, R7 }
131 punpckldq m2, m5 ; (byte) { B4, G4, R4, B5, B6, G6, R6, B7 }
132 punpckldq m3, m6 ; (byte) { R4, B5, G5, R5, R6, B7, G7, R7 }
133 %endif ; mmsize == 16
134 punpcklbw m0, m7 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
135 punpcklbw m1, m7 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
136 punpcklbw m2, m7 ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
137 punpcklbw m3, m7 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
138 %endif ; cpuflag(ssse3)
139 add srcq, 3 * mmsize / 2
140 pmaddwd m0, coeff1 ; (dword) { B0*BY + G0*GY, B1*BY, B2*BY + G2*GY, B3*BY }
141 pmaddwd m1, coeff2 ; (dword) { R0*RY, G1+GY + R1*RY, R2*RY, G3+GY + R3*RY }
142 pmaddwd m2, coeff1 ; (dword) { B4*BY + G4*GY, B5*BY, B6*BY + G6*GY, B7*BY }
143 pmaddwd m3, coeff2 ; (dword) { R4*RY, G5+GY + R5*RY, R6*RY, G7+GY + R7*RY }
144 paddd m0, m1 ; (dword) { Bx*BY + Gx*GY + Rx*RY }[0-3]
145 paddd m2, m3 ; (dword) { Bx*BY + Gx*GY + Rx*RY }[4-7]
146 paddd m0, m4 ; += rgb_Yrnd, i.e. (dword) { Y[0-3] }
147 paddd m2, m4 ; += rgb_Yrnd, i.e. (dword) { Y[4-7] }
150 packssdw m0, m2 ; (word) { Y[0-7] }
151 packuswb m0, m0 ; (byte) { Y[0-7] }
156 %endif ; (ARCH_X86_64 && %0 == 3) || mmsize == 8
159 ; %1 = nr. of XMM registers
161 %macro RGB24_TO_UV_FN 2-3
162 cglobal %2 %+ 24ToUV, 3, 4, %1, dstU, dstV, src, w
164 mova m8, [%2_Ucoeff_12x4]
165 mova m9, [%2_Ucoeff_3x56]
166 mova m10, [%2_Vcoeff_12x4]
167 mova m11, [%2_Vcoeff_3x56]
173 %define coeffU1 [%2_Ucoeff_12x4]
174 %define coeffU2 [%2_Ucoeff_3x56]
175 %define coeffV1 [%2_Vcoeff_12x4]
176 %define coeffV2 [%2_Vcoeff_3x56]
178 %if ARCH_X86_64 && %0 == 3
179 jmp mangle(program_name %+ _ %+ %3 %+ 24ToUV %+ SUFFIX).body
180 %else ; ARCH_X86_64 && %0 == 3
183 mova m7, [shuf_rgb_12x4]
186 mova m12, [shuf_rgb_3x56]
187 %define shuf_rgb2 m12
189 %define shuf_rgb2 [shuf_rgb_3x56]
191 %endif ; cpuflag(ssse3)
201 %if notcpuflag(ssse3)
206 movu m0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3]
207 movu m4, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7]
208 pshufb m1, m0, shuf_rgb2 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
209 pshufb m0, shuf_rgb1 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
210 %else ; !cpuflag(ssse3)
211 movd m0, [srcq+0] ; (byte) { B0, G0, R0, B1 }
212 movd m1, [srcq+2] ; (byte) { R0, B1, G1, R1 }
213 movd m4, [srcq+6] ; (byte) { B2, G2, R2, B3 }
214 movd m5, [srcq+8] ; (byte) { R2, B3, G3, R3 }
216 punpckldq m0, m4 ; (byte) { B0, G0, R0, B1, B2, G2, R2, B3 }
217 punpckldq m1, m5 ; (byte) { R0, B1, G1, R1, R2, B3, G3, R3 }
218 movd m4, [srcq+12] ; (byte) { B4, G4, R4, B5 }
219 movd m5, [srcq+14] ; (byte) { R4, B5, G5, R5 }
220 %endif ; mmsize == 16
221 punpcklbw m0, m7 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 }
222 punpcklbw m1, m7 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 }
223 %endif ; cpuflag(ssse3)
224 pmaddwd m2, m0, coeffV1 ; (dword) { B0*BV + G0*GV, B1*BV, B2*BV + G2*GV, B3*BV }
225 pmaddwd m3, m1, coeffV2 ; (dword) { R0*BV, G1*GV + R1*BV, R2*BV, G3*GV + R3*BV }
226 pmaddwd m0, coeffU1 ; (dword) { B0*BU + G0*GU, B1*BU, B2*BU + G2*GU, B3*BU }
227 pmaddwd m1, coeffU2 ; (dword) { R0*BU, G1*GU + R1*BU, R2*BU, G3*GU + R3*BU }
228 paddd m0, m1 ; (dword) { Bx*BU + Gx*GU + Rx*RU }[0-3]
229 paddd m2, m3 ; (dword) { Bx*BV + Gx*GV + Rx*RV }[0-3]
231 pshufb m5, m4, shuf_rgb2 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
232 pshufb m4, shuf_rgb1 ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
233 %else ; !cpuflag(ssse3)
235 movd m1, [srcq+18] ; (byte) { B6, G6, R6, B7 }
236 movd m3, [srcq+20] ; (byte) { R6, B7, G7, R7 }
237 punpckldq m4, m1 ; (byte) { B4, G4, R4, B5, B6, G6, R6, B7 }
238 punpckldq m5, m3 ; (byte) { R4, B5, G5, R5, R6, B7, G7, R7 }
239 %endif ; mmsize == 16 && !cpuflag(ssse3)
240 punpcklbw m4, m7 ; (word) { B4, G4, R4, B5, B6, G6, R6, B7 }
241 punpcklbw m5, m7 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 }
242 %endif ; cpuflag(ssse3)
243 add srcq, 3 * mmsize / 2
244 pmaddwd m1, m4, coeffU1 ; (dword) { B4*BU + G4*GU, B5*BU, B6*BU + G6*GU, B7*BU }
245 pmaddwd m3, m5, coeffU2 ; (dword) { R4*BU, G5*GU + R5*BU, R6*BU, G7*GU + R7*BU }
246 pmaddwd m4, coeffV1 ; (dword) { B4*BV + G4*GV, B5*BV, B6*BV + G6*GV, B7*BV }
247 pmaddwd m5, coeffV2 ; (dword) { R4*BV, G5*GV + R5*BV, R6*BV, G7*GV + R7*BV }
248 paddd m1, m3 ; (dword) { Bx*BU + Gx*GU + Rx*RU }[4-7]
249 paddd m4, m5 ; (dword) { Bx*BV + Gx*GV + Rx*RV }[4-7]
250 paddd m0, m6 ; += rgb_UVrnd, i.e. (dword) { U[0-3] }
251 paddd m2, m6 ; += rgb_UVrnd, i.e. (dword) { V[0-3] }
252 paddd m1, m6 ; += rgb_UVrnd, i.e. (dword) { U[4-7] }
253 paddd m4, m6 ; += rgb_UVrnd, i.e. (dword) { V[4-7] }
258 packssdw m0, m1 ; (word) { U[0-7] }
259 packssdw m2, m4 ; (word) { V[0-7] }
261 packuswb m0, m0 ; (byte) { U[0-3] }
262 packuswb m2, m2 ; (byte) { V[0-3] }
266 packuswb m0, m2 ; (byte) { U[0-7], V[0-7] }
268 movhps [dstVq+wq], m0
269 %endif ; mmsize == 8/16
273 %endif ; ARCH_X86_64 && %0 == 3
279 RGB24_TO_Y_FN 0, bgr, rgb
280 RGB24_TO_UV_FN 0, rgb
281 RGB24_TO_UV_FN 0, bgr, rgb
285 RGB24_TO_Y_FN 10, rgb
286 RGB24_TO_Y_FN 10, bgr, rgb
287 RGB24_TO_UV_FN 12, rgb
288 RGB24_TO_UV_FN 12, bgr, rgb
291 RGB24_TO_Y_FN 11, rgb
292 RGB24_TO_Y_FN 11, bgr, rgb
293 RGB24_TO_UV_FN 13, rgb
294 RGB24_TO_UV_FN 13, bgr, rgb
297 RGB24_TO_Y_FN 11, rgb
298 RGB24_TO_Y_FN 11, bgr, rgb
299 RGB24_TO_UV_FN 13, rgb
300 RGB24_TO_UV_FN 13, bgr, rgb
302 ;-----------------------------------------------------------------------------
303 ; YUYV/UYVY/NV12/NV21 packed pixel shuffling.
305 ; void <fmt>ToY_<opt>(uint8_t *dst, const uint8_t *src, int w);
307 ; void <fmt>toUV_<opt>(uint8_t *dstU, uint8_t *dstV, const uint8_t *src,
308 ; const uint8_t *unused, int w);
309 ;-----------------------------------------------------------------------------
311 ; %1 = a (aligned) or u (unaligned)
313 %macro LOOP_YUYV_TO_Y 2
315 mov%1 m0, [srcq+wq*2] ; (byte) { Y0, U0, Y1, V0, ... }
316 mov%1 m1, [srcq+wq*2+mmsize] ; (byte) { Y8, U4, Y9, V4, ... }
318 pand m0, m2 ; (word) { Y0, Y1, ..., Y7 }
319 pand m1, m2 ; (word) { Y8, Y9, ..., Y15 }
321 psrlw m0, 8 ; (word) { Y0, Y1, ..., Y7 }
322 psrlw m1, 8 ; (word) { Y8, Y9, ..., Y15 }
324 packuswb m0, m1 ; (byte) { Y0, ..., Y15 }
331 ; %1 = nr. of XMM registers
333 ; %3 = if specified, it means that unaligned and aligned code in loop
334 ; will be the same (i.e. YUYV+AVX), and thus we don't need to
335 ; split the loop in an aligned and unaligned case
336 %macro YUYV_TO_Y_FN 2-3
337 cglobal %2ToY, 3, 3, %1, dst, src, w
345 lea srcq, [srcq+wq*2]
347 pcmpeqb m2, m2 ; (byte) { 0xff } x 16
348 psrlw m2, 8 ; (word) { 0x00ff } x 8
360 %endif ; mmsize == 8/16
363 ; %1 = a (aligned) or u (unaligned)
365 %macro LOOP_YUYV_TO_UV 2
368 mov%1 m0, [srcq+wq*4] ; (byte) { Y0, U0, Y1, V0, ... }
369 mov%1 m1, [srcq+wq*4+mmsize] ; (byte) { Y8, U4, Y9, V4, ... }
370 psrlw m0, 8 ; (word) { U0, V0, ..., U3, V3 }
371 psrlw m1, 8 ; (word) { U4, V4, ..., U7, V7 }
374 vpand m0, m2, [srcq+wq*4] ; (word) { U0, V0, ..., U3, V3 }
375 vpand m1, m2, [srcq+wq*4+mmsize] ; (word) { U4, V4, ..., U7, V7 }
377 mov%1 m0, [srcq+wq*4] ; (byte) { Y0, U0, Y1, V0, ... }
378 mov%1 m1, [srcq+wq*4+mmsize] ; (byte) { Y8, U4, Y9, V4, ... }
379 pand m0, m2 ; (word) { U0, V0, ..., U3, V3 }
380 pand m1, m2 ; (word) { U4, V4, ..., U7, V7 }
383 packuswb m0, m1 ; (byte) { U0, V0, ..., U7, V7 }
384 pand m1, m0, m2 ; (word) { U0, U1, ..., U7 }
385 psrlw m0, 8 ; (word) { V0, V1, ..., V7 }
387 packuswb m1, m0 ; (byte) { U0, ... U7, V1, ... V7 }
389 movhps [dstVq+wq], m1
391 packuswb m1, m1 ; (byte) { U0, ... U3 }
392 packuswb m0, m0 ; (byte) { V0, ... V3 }
395 %endif ; mmsize == 8/16
401 ; %1 = nr. of XMM registers
403 ; %3 = if specified, it means that unaligned and aligned code in loop
404 ; will be the same (i.e. UYVY+AVX), and thus we don't need to
405 ; split the loop in an aligned and unaligned case
406 %macro YUYV_TO_UV_FN 2-3
407 cglobal %2ToUV, 3, 4, %1, dstU, dstV, src, w
415 %if mmsize == 16 && %0 == 2
418 lea srcq, [srcq+wq*4]
419 pcmpeqb m2, m2 ; (byte) { 0xff } x 16
420 psrlw m2, 8 ; (word) { 0x00ff } x 8
421 ; NOTE: if uyvy+avx, u/a are identical
422 %if mmsize == 16 && %0 == 2
425 LOOP_YUYV_TO_UV a, %2
428 LOOP_YUYV_TO_UV u, %2
431 LOOP_YUYV_TO_UV a, %2
432 %endif ; mmsize == 8/16
435 ; %1 = a (aligned) or u (unaligned)
437 %macro LOOP_NVXX_TO_UV 2
439 mov%1 m0, [srcq+wq*2] ; (byte) { U0, V0, U1, V1, ... }
440 mov%1 m1, [srcq+wq*2+mmsize] ; (byte) { U8, V8, U9, V9, ... }
441 pand m2, m0, m4 ; (word) { U0, U1, ..., U7 }
442 pand m3, m1, m4 ; (word) { U8, U9, ..., U15 }
443 psrlw m0, 8 ; (word) { V0, V1, ..., V7 }
444 psrlw m1, 8 ; (word) { V8, V9, ..., V15 }
445 packuswb m2, m3 ; (byte) { U0, ..., U15 }
446 packuswb m0, m1 ; (byte) { V0, ..., V15 }
459 ; %1 = nr. of XMM registers
461 %macro NVXX_TO_UV_FN 2
462 cglobal %2ToUV, 3, 4, %1, dstU, dstV, src, w
473 lea srcq, [srcq+wq*2]
474 pcmpeqb m4, m4 ; (byte) { 0xff } x 16
475 psrlw m4, 8 ; (word) { 0x00ff } x 8
479 LOOP_NVXX_TO_UV a, %2
482 LOOP_NVXX_TO_UV u, %2
485 LOOP_NVXX_TO_UV a, %2
486 %endif ; mmsize == 8/16
493 YUYV_TO_UV_FN 0, yuyv
494 YUYV_TO_UV_FN 0, uyvy
495 NVXX_TO_UV_FN 0, nv12
496 NVXX_TO_UV_FN 0, nv21
502 YUYV_TO_UV_FN 3, yuyv
503 YUYV_TO_UV_FN 3, uyvy
504 NVXX_TO_UV_FN 5, nv12
505 NVXX_TO_UV_FN 5, nv21
508 ; in theory, we could write a yuy2-to-y using vpand (i.e. AVX), but
509 ; that's not faster in practice
510 YUYV_TO_UV_FN 3, yuyv
511 YUYV_TO_UV_FN 3, uyvy, 1
512 NVXX_TO_UV_FN 5, nv12
513 NVXX_TO_UV_FN 5, nv21