1 |
|
;/**************************************************************************** |
2 |
|
; * |
3 |
|
; * XVID MPEG-4 VIDEO CODEC |
4 |
|
; * - MMX and XMM forward discrete cosine transform - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
|
; * |
8 |
|
; * This program is free software; you can redistribute it and/or modify it |
9 |
|
; * under the terms of the GNU General Public License as published by |
10 |
|
; * the Free Software Foundation; either version 2 of the License, or |
11 |
|
; * (at your option) any later version. |
12 |
|
; * |
13 |
|
; * This program is distributed in the hope that it will be useful, |
14 |
|
; * but WITHOUT ANY WARRANTY; without even the implied warranty of |
15 |
|
; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
16 |
|
; * GNU General Public License for more details. |
17 |
|
; * |
18 |
|
; * You should have received a copy of the GNU General Public License |
19 |
|
; * along with this program; if not, write to the Free Software |
20 |
|
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
21 |
|
; * |
22 |
|
; * $Id$ |
23 |
|
; * |
24 |
|
; ***************************************************************************/ |
25 |
|
|
26 |
|
; **************************************************************************** |
27 |
|
; |
28 |
; Originally provided by Intel at AP-922 |
; Originally provided by Intel at AP-922 |
29 |
; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm |
; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm |
30 |
; (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) |
; (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) |
32 |
; New macro implements a column part for precise iDCT |
; New macro implements a column part for precise iDCT |
33 |
; The routine precision now satisfies IEEE standard 1180-1990. |
; The routine precision now satisfies IEEE standard 1180-1990. |
34 |
; |
; |
35 |
; Copyright (c) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
; Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
36 |
; Rounding trick Copyright (c) 2000 Michel Lespinasse <walken@zoy.org> |
; Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org> |
37 |
; |
; |
38 |
; http://www.elecard.com/peter/idct.html |
; http://www.elecard.com/peter/idct.html |
39 |
; http://www.linuxvideo.org/mpeg2dec/ |
; http://www.linuxvideo.org/mpeg2dec/ |
40 |
; |
; |
41 |
;============================================================================= |
; ***************************************************************************/ |
42 |
; |
; |
43 |
; These examples contain code fragments for first stage iDCT 8x8 |
; These examples contain code fragments for first stage iDCT 8x8 |
44 |
; (for rows) and first stage DCT 8x8 (for columns) |
; (for rows) and first stage DCT 8x8 (for columns) |
45 |
; |
; |
|
;============================================================================= |
|
|
; |
|
|
; 04.11.2001 nasm conversion; peter ross <pross@cs.rmit.edu.au> |
|
|
; |
|
46 |
|
|
47 |
bits 32 |
BITS 32 |
48 |
|
|
49 |
|
;============================================================================= |
50 |
|
; Macros and other preprocessor constants |
51 |
|
;============================================================================= |
52 |
|
|
53 |
%macro cglobal 1 |
%macro cglobal 1 |
54 |
%ifdef PREFIX |
%ifdef PREFIX |
55 |
|
%ifdef MARK_FUNCS |
56 |
|
global _%1:function |
57 |
|
%define %1 _%1:function |
58 |
|
%else |
59 |
global _%1 |
global _%1 |
60 |
%define %1 _%1 |
%define %1 _%1 |
61 |
|
%endif |
62 |
|
%else |
63 |
|
%ifdef MARK_FUNCS |
64 |
|
global %1:function |
65 |
%else |
%else |
66 |
global %1 |
global %1 |
67 |
%endif |
%endif |
68 |
|
%endif |
69 |
%endmacro |
%endmacro |
70 |
|
|
71 |
%define BITS_INV_ACC 5 ; 4 or 5 for IEEE |
%define BITS_INV_ACC 5 ; 4 or 5 for IEEE |
80 |
%define SHIFT_FRW_ROW BITS_FRW_ACC + 17 |
%define SHIFT_FRW_ROW BITS_FRW_ACC + 17 |
81 |
%define RND_FRW_ROW 262144 * (BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) |
%define RND_FRW_ROW 262144 * (BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) |
82 |
|
|
83 |
|
;============================================================================= |
84 |
|
; Local Data (Read Only) |
85 |
|
;============================================================================= |
86 |
|
|
87 |
section .data data align=16 |
%ifdef FORMAT_COFF |
88 |
|
SECTION .rodata |
89 |
align 16 |
%else |
90 |
|
SECTION .rodata align=16 |
91 |
|
%endif |
92 |
|
|
93 |
one_corr dw 1, 1, 1, 1 |
;----------------------------------------------------------------------------- |
94 |
round_inv_row dd RND_INV_ROW, RND_INV_ROW |
; Various memory constants (trigonometric values or rounding values) |
95 |
round_inv_col dw RND_INV_COL, RND_INV_COL, RND_INV_COL, RND_INV_COL |
;----------------------------------------------------------------------------- |
|
round_inv_corr dw RND_INV_CORR, RND_INV_CORR, RND_INV_CORR, RND_INV_CORR |
|
|
round_frw_row dd RND_FRW_ROW, RND_FRW_ROW |
|
|
tg_1_16 dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 |
|
|
tg_2_16 dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 |
|
|
tg_3_16 dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 |
|
|
cos_4_16 dw -19195, -19195, -19195, -19195 ; cos * (2<<16) + 0.5 |
|
|
ocos_4_16 dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 |
|
96 |
|
|
97 |
otg_3_16 dw 21895, 21895, 21895, 21895 ; tg * (2<<16) + 0.5 |
ALIGN 16 |
98 |
|
one_corr: |
99 |
|
dw 1, 1, 1, 1 |
100 |
|
round_inv_row: |
101 |
|
dd RND_INV_ROW, RND_INV_ROW |
102 |
|
round_inv_col: |
103 |
|
dw RND_INV_COL, RND_INV_COL, RND_INV_COL, RND_INV_COL |
104 |
|
round_inv_corr: |
105 |
|
dw RND_INV_CORR, RND_INV_CORR, RND_INV_CORR, RND_INV_CORR |
106 |
|
round_frw_row: |
107 |
|
dd RND_FRW_ROW, RND_FRW_ROW |
108 |
|
tg_1_16: |
109 |
|
dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 |
110 |
|
tg_2_16: |
111 |
|
dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 |
112 |
|
tg_3_16: |
113 |
|
dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 |
114 |
|
cos_4_16: |
115 |
|
dw -19195, -19195, -19195, -19195 ; cos * (2<<16) + 0.5 |
116 |
|
ocos_4_16: |
117 |
|
dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 |
118 |
|
otg_3_16: |
119 |
|
dw 21895, 21895, 21895, 21895 ; tg * (2<<16) + 0.5 |
120 |
|
|
121 |
%if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 |
%if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 |
122 |
rounder_0 dd 65536, 65536 |
rounder_0: |
123 |
rounder_4 dd 0, 0 |
dd 65536, 65536 |
124 |
rounder_1 dd 7195, 7195 |
rounder_4: |
125 |
rounder_7 dd 1024, 1024 |
dd 0, 0 |
126 |
rounder_2 dd 4520, 4520 |
rounder_1: |
127 |
rounder_6 dd 1024, 1024 |
dd 7195, 7195 |
128 |
rounder_3 dd 2407, 2407 |
rounder_7 |
129 |
rounder_5 dd 240, 240 |
dd 1024, 1024 |
130 |
|
rounder_2: |
131 |
|
dd 4520, 4520 |
132 |
|
rounder_6: |
133 |
|
dd 1024, 1024 |
134 |
|
rounder_3: |
135 |
|
dd 2407, 2407 |
136 |
|
rounder_5: |
137 |
|
dd 240, 240 |
138 |
|
|
139 |
%elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 |
%elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 |
140 |
rounder_0 dd 65536, 65536 |
rounder_0: |
141 |
rounder_4 dd 0, 0 |
dd 65536, 65536 |
142 |
rounder_1 dd 3597, 3597 |
rounder_4: |
143 |
rounder_7 dd 512, 512 |
dd 0, 0 |
144 |
rounder_2 dd 2260, 2260 |
rounder_1: |
145 |
rounder_6 dd 512, 512 |
dd 3597, 3597 |
146 |
rounder_3 dd 1203, 1203 |
rounder_7: |
147 |
rounder_5 dd 120, 120 |
dd 512, 512 |
148 |
|
rounder_2: |
149 |
|
dd 2260, 2260 |
150 |
|
rounder_6: |
151 |
|
dd 512, 512 |
152 |
|
rounder_3: |
153 |
|
dd 1203, 1203 |
154 |
|
rounder_5: |
155 |
|
dd 120, 120 |
156 |
%else |
%else |
157 |
|
|
158 |
%error invalid _SHIFT_INV_ROW_ |
%error invalid SHIFT_INV_ROW |
159 |
|
|
160 |
%endif |
%endif |
161 |
|
|
162 |
;============================================================================= |
;----------------------------------------------------------------------------- |
163 |
; |
; |
164 |
; The first stage iDCT 8x8 - inverse DCTs of rows |
; The first stage iDCT 8x8 - inverse DCTs of rows |
165 |
; |
; |
212 |
; For the constants used, |
; For the constants used, |
213 |
; FIX(float_const) = (short) (float_const * (1<<15) + 0.5) |
; FIX(float_const) = (short) (float_const * (1<<15) + 0.5) |
214 |
; |
; |
215 |
;============================================================================= |
;----------------------------------------------------------------------------- |
216 |
|
|
217 |
;============================================================================= |
;----------------------------------------------------------------------------- |
218 |
; MMX code |
; Tables for mmx processors |
219 |
;============================================================================= |
;----------------------------------------------------------------------------- |
220 |
|
|
221 |
; Table for rows 0,4 - constants are multiplied by cos_4_16 |
; Table for rows 0,4 - constants are multiplied by cos_4_16 |
222 |
|
tab_i_04_mmx: |
223 |
tab_i_04 dw 16384, 16384, 16384, -16384 ; movq-> w06 w04 w02 w00 |
dw 16384, 16384, 16384, -16384 ; movq-> w06 w04 w02 w00 |
224 |
dw 21407, 8867, 8867, -21407 ; w07 w05 w03 w01 |
dw 21407, 8867, 8867, -21407 ; w07 w05 w03 w01 |
225 |
dw 16384, -16384, 16384, 16384 ; w14 w12 w10 w08 |
dw 16384, -16384, 16384, 16384 ; w14 w12 w10 w08 |
226 |
dw -8867, 21407, -21407, -8867 ; w15 w13 w11 w09 |
dw -8867, 21407, -21407, -8867 ; w15 w13 w11 w09 |
230 |
dw -22725, 19266, -12873, -22725 ; w31 w29 w27 w25 |
dw -22725, 19266, -12873, -22725 ; w31 w29 w27 w25 |
231 |
|
|
232 |
; Table for rows 1,7 - constants are multiplied by cos_1_16 |
; Table for rows 1,7 - constants are multiplied by cos_1_16 |
233 |
|
tab_i_17_mmx: |
234 |
tab_i_17 dw 22725, 22725, 22725, -22725 ; movq-> w06 w04 w02 w00 |
dw 22725, 22725, 22725, -22725 ; movq-> w06 w04 w02 w00 |
235 |
dw 29692, 12299, 12299, -29692 ; w07 w05 w03 w01 |
dw 29692, 12299, 12299, -29692 ; w07 w05 w03 w01 |
236 |
dw 22725, -22725, 22725, 22725 ; w14 w12 w10 w08 |
dw 22725, -22725, 22725, 22725 ; w14 w12 w10 w08 |
237 |
dw -12299, 29692, -29692, -12299 ; w15 w13 w11 w09 |
dw -12299, 29692, -29692, -12299 ; w15 w13 w11 w09 |
241 |
dw -31521, 26722, -17855, -31521 ; w31 w29 w27 w25 |
dw -31521, 26722, -17855, -31521 ; w31 w29 w27 w25 |
242 |
|
|
243 |
; Table for rows 2,6 - constants are multiplied by cos_2_16 |
; Table for rows 2,6 - constants are multiplied by cos_2_16 |
244 |
|
tab_i_26_mmx: |
245 |
tab_i_26 dw 21407, 21407, 21407, -21407 ; movq-> w06 w04 w02 w00 |
dw 21407, 21407, 21407, -21407 ; movq-> w06 w04 w02 w00 |
246 |
dw 27969, 11585, 11585, -27969 ; w07 w05 w03 w01 |
dw 27969, 11585, 11585, -27969 ; w07 w05 w03 w01 |
247 |
dw 21407, -21407, 21407, 21407 ; w14 w12 w10 w08 |
dw 21407, -21407, 21407, 21407 ; w14 w12 w10 w08 |
248 |
dw -11585, 27969, -27969, -11585 ; w15 w13 w11 w09 |
dw -11585, 27969, -27969, -11585 ; w15 w13 w11 w09 |
252 |
dw -29692, 25172, -16819, -29692 ; w31 w29 w27 w25 |
dw -29692, 25172, -16819, -29692 ; w31 w29 w27 w25 |
253 |
|
|
254 |
; Table for rows 3,5 - constants are multiplied by cos_3_16 |
; Table for rows 3,5 - constants are multiplied by cos_3_16 |
255 |
|
tab_i_35_mmx: |
256 |
tab_i_35 dw 19266, 19266, 19266, -19266 ; movq-> w06 w04 w02 w00 |
dw 19266, 19266, 19266, -19266 ; movq-> w06 w04 w02 w00 |
257 |
dw 25172, 10426, 10426, -25172 ; w07 w05 w03 w01 |
dw 25172, 10426, 10426, -25172 ; w07 w05 w03 w01 |
258 |
dw 19266, -19266, 19266, 19266 ; w14 w12 w10 w08 |
dw 19266, -19266, 19266, 19266 ; w14 w12 w10 w08 |
259 |
dw -10426, 25172, -25172, -10426 ; w15 w13 w11 w09 |
dw -10426, 25172, -25172, -10426 ; w15 w13 w11 w09 |
263 |
dw -26722, 22654, -15137, -26722 ; w31 w29 w27 w25 |
dw -26722, 22654, -15137, -26722 ; w31 w29 w27 w25 |
264 |
|
|
265 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
266 |
|
; Tables for xmm processors |
267 |
|
;----------------------------------------------------------------------------- |
268 |
|
|
269 |
; |
; %3 for rows 0,4 - constants are multiplied by cos_4_16 |
270 |
; DCT_8_INV_ROW_1 INP, OUT, TABLE, ROUNDER |
tab_i_04_xmm: |
271 |
; |
dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 |
272 |
|
dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 |
273 |
|
dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 |
274 |
|
dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 |
275 |
|
dw 22725, 19266, 19266, -4520 ; w21 w20 w17 w16 |
276 |
|
dw 12873, 4520, -22725, -12873 ; w23 w22 w19 w18 |
277 |
|
dw 12873, -22725, 4520, -12873 ; w29 w28 w25 w24 |
278 |
|
dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 |
279 |
|
|
280 |
|
; %3 for rows 1,7 - constants are multiplied by cos_1_16 |
281 |
|
tab_i_17_xmm: |
282 |
|
dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 |
283 |
|
dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 |
284 |
|
dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 |
285 |
|
dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 |
286 |
|
dw 31521, 26722, 26722, -6270 ; w21 w20 w17 w16 |
287 |
|
dw 17855, 6270, -31521, -17855 ; w23 w22 w19 w18 |
288 |
|
dw 17855, -31521, 6270, -17855 ; w29 w28 w25 w24 |
289 |
|
dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 |
290 |
|
|
291 |
%macro DCT_8_INV_ROW_1 4 |
; %3 for rows 2,6 - constants are multiplied by cos_2_16 |
292 |
|
tab_i_26_xmm: |
293 |
|
dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 |
294 |
|
dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 |
295 |
|
dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 |
296 |
|
dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 |
297 |
|
dw 29692, 25172, 25172, -5906 ; w21 w20 w17 w16 |
298 |
|
dw 16819, 5906, -29692, -16819 ; w23 w22 w19 w18 |
299 |
|
dw 16819, -29692, 5906, -16819 ; w29 w28 w25 w24 |
300 |
|
dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 |
301 |
|
|
302 |
movq mm0, [%1] ; 0 ; x3 x2 x1 x0 |
; %3 for rows 3,5 - constants are multiplied by cos_3_16 |
303 |
|
tab_i_35_xmm: |
304 |
|
dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 |
305 |
|
dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 |
306 |
|
dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 |
307 |
|
dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 |
308 |
|
dw 26722, 22654, 22654, -5315 ; w21 w20 w17 w16 |
309 |
|
dw 15137, 5315, -26722, -15137 ; w23 w22 w19 w18 |
310 |
|
dw 15137, -26722, 5315, -15137 ; w29 w28 w25 w24 |
311 |
|
dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 |
312 |
|
|
313 |
|
;============================================================================= |
314 |
|
; Helper macros for the code |
315 |
|
;============================================================================= |
316 |
|
|
317 |
|
;----------------------------------------------------------------------------- |
318 |
|
; DCT_8_INV_ROW_MMX INP, OUT, TABLE, ROUNDER |
319 |
|
;----------------------------------------------------------------------------- |
320 |
|
|
321 |
|
%macro DCT_8_INV_ROW_MMX 4 |
322 |
|
movq mm0, [%1] ; 0 ; x3 x2 x1 x0 |
323 |
movq mm1, [%1+8] ; 1 ; x7 x6 x5 x4 |
movq mm1, [%1+8] ; 1 ; x7 x6 x5 x4 |
324 |
movq mm2, mm0 ; 2 ; x3 x2 x1 x0 |
movq mm2, mm0 ; 2 ; x3 x2 x1 x0 |
|
|
|
325 |
movq mm3, [%3] ; 3 ; w06 w04 w02 w00 |
movq mm3, [%3] ; 3 ; w06 w04 w02 w00 |
326 |
punpcklwd mm0, mm1 ; x5 x1 x4 x0 |
punpcklwd mm0, mm1 ; x5 x1 x4 x0 |
|
|
|
327 |
movq mm5, mm0 ; 5 ; x5 x1 x4 x0 |
movq mm5, mm0 ; 5 ; x5 x1 x4 x0 |
328 |
punpckldq mm0, mm0 ; x4 x0 x4 x0 |
punpckldq mm0, mm0 ; x4 x0 x4 x0 |
|
|
|
329 |
movq mm4, [%3+8] ; 4 ; w07 w05 w03 w01 |
movq mm4, [%3+8] ; 4 ; w07 w05 w03 w01 |
330 |
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 |
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 |
|
|
|
331 |
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 |
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 |
332 |
movq mm6, mm2 ; 6 ; x7 x3 x6 x2 |
movq mm6, mm2 ; 6 ; x7 x3 x6 x2 |
|
|
|
333 |
movq mm1, [%3+32] ; 1 ; w22 w20 w18 w16 |
movq mm1, [%3+32] ; 1 ; w22 w20 w18 w16 |
334 |
punpckldq mm2, mm2 ; x6 x2 x6 x2 |
punpckldq mm2, mm2 ; x6 x2 x6 x2 |
|
|
|
335 |
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 |
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 |
336 |
punpckhdq mm5, mm5 ; x5 x1 x5 x1 |
punpckhdq mm5, mm5 ; x5 x1 x5 x1 |
|
|
|
337 |
pmaddwd mm0, [%3+16] ; x4*w14+x0*w12 x4*w10+x0*w08 |
pmaddwd mm0, [%3+16] ; x4*w14+x0*w12 x4*w10+x0*w08 |
338 |
punpckhdq mm6, mm6 ; x7 x3 x7 x3 |
punpckhdq mm6, mm6 ; x7 x3 x7 x3 |
|
|
|
339 |
movq mm7, [%3+40] ; 7 ; w23 w21 w19 w17 |
movq mm7, [%3+40] ; 7 ; w23 w21 w19 w17 |
340 |
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 |
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 |
|
|
|
341 |
paddd mm3, [%4] ; +%4 |
paddd mm3, [%4] ; +%4 |
342 |
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 |
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 |
|
|
|
343 |
pmaddwd mm2, [%3+24] ; x6*w15+x2*w13 x6*w11+x2*w09 |
pmaddwd mm2, [%3+24] ; x6*w15+x2*w13 x6*w11+x2*w09 |
344 |
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) |
|
|
|
345 |
pmaddwd mm5, [%3+48] ; x5*w30+x1*w28 x5*w26+x1*w24 |
pmaddwd mm5, [%3+48] ; x5*w30+x1*w28 x5*w26+x1*w24 |
346 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
|
|
|
347 |
pmaddwd mm6, [%3+56] ; x7*w31+x3*w29 x7*w27+x3*w25 |
pmaddwd mm6, [%3+56] ; x7*w31+x3*w29 x7*w27+x3*w25 |
348 |
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) |
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) |
|
|
|
349 |
paddd mm0, [%4] ; +%4 |
paddd mm0, [%4] ; +%4 |
350 |
psubd mm3, mm1 ; a1-b1 a0-b0 |
psubd mm3, mm1 ; a1-b1 a0-b0 |
|
|
|
351 |
psrad mm3, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm3, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
352 |
paddd mm1, mm4 ; 4 ; a1+b1 a0+b0 |
paddd mm1, mm4 ; 4 ; a1+b1 a0+b0 |
|
|
|
353 |
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) |
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) |
354 |
psrad mm1, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm1, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
|
|
|
355 |
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) |
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) |
356 |
movq mm4, mm0 ; 4 ; a3 a2 |
movq mm4, mm0 ; 4 ; a3 a2 |
|
|
|
357 |
paddd mm0, mm5 ; a3+b3 a2+b2 |
paddd mm0, mm5 ; a3+b3 a2+b2 |
358 |
psubd mm4, mm5 ; 5 ; a3-b3 a2-b2 |
psubd mm4, mm5 ; 5 ; a3-b3 a2-b2 |
|
|
|
359 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
360 |
psrad mm4, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm4, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
|
|
|
361 |
packssdw mm1, mm0 ; 0 ; y3 y2 y1 y0 |
packssdw mm1, mm0 ; 0 ; y3 y2 y1 y0 |
362 |
packssdw mm4, mm3 ; 3 ; y6 y7 y4 y5 |
packssdw mm4, mm3 ; 3 ; y6 y7 y4 y5 |
|
|
|
363 |
movq mm7, mm4 ; 7 ; y6 y7 y4 y5 |
movq mm7, mm4 ; 7 ; y6 y7 y4 y5 |
364 |
psrld mm4, 16 ; 0 y6 0 y4 |
psrld mm4, 16 ; 0 y6 0 y4 |
|
|
|
365 |
pslld mm7, 16 ; y7 0 y5 0 |
pslld mm7, 16 ; y7 0 y5 0 |
366 |
movq [%2], mm1 ; 1 ; save y3 y2 y1 y0 |
movq [%2], mm1 ; 1 ; save y3 y2 y1 y0 |
|
|
|
367 |
por mm7, mm4 ; 4 ; y7 y6 y5 y4 |
por mm7, mm4 ; 4 ; y7 y6 y5 y4 |
368 |
movq [%2+8], mm7 ; 7 ; save y7 y6 y5 y4 |
movq [%2+8], mm7 ; 7 ; save y7 y6 y5 y4 |
369 |
%endmacro |
%endmacro |
370 |
|
|
371 |
|
;----------------------------------------------------------------------------- |
372 |
|
; DCT_8_INV_ROW_XMM INP, OUT, TABLE, ROUNDER |
|
|
|
|
;============================================================================= |
|
|
; code for Pentium III |
|
|
;============================================================================= |
|
|
|
|
|
; %3 for rows 0,4 - constants are multiplied by cos_4_16 |
|
|
|
|
|
tab_i_04_sse dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 |
|
|
dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 |
|
|
dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 |
|
|
dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 |
|
|
dw 22725, 19266, 19266, -4520 ; w21 w20 w17 w16 |
|
|
dw 12873, 4520, -22725, -12873 ; w23 w22 w19 w18 |
|
|
dw 12873, -22725, 4520, -12873 ; w29 w28 w25 w24 |
|
|
dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 |
|
|
|
|
|
; %3 for rows 1,7 - constants are multiplied by cos_1_16 |
|
|
|
|
|
tab_i_17_sse dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 |
|
|
dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 |
|
|
dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 |
|
|
dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 |
|
|
dw 31521, 26722, 26722, -6270 ; w21 w20 w17 w16 |
|
|
dw 17855, 6270, -31521, -17855 ; w23 w22 w19 w18 |
|
|
dw 17855, -31521, 6270, -17855 ; w29 w28 w25 w24 |
|
|
dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 |
|
|
|
|
|
; %3 for rows 2,6 - constants are multiplied by cos_2_16 |
|
|
|
|
|
tab_i_26_sse dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 |
|
|
dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 |
|
|
dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 |
|
|
dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 |
|
|
dw 29692, 25172, 25172, -5906 ; w21 w20 w17 w16 |
|
|
dw 16819, 5906, -29692, -16819 ; w23 w22 w19 w18 |
|
|
dw 16819, -29692, 5906, -16819 ; w29 w28 w25 w24 |
|
|
dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 |
|
|
|
|
|
; %3 for rows 3,5 - constants are multiplied by cos_3_16 |
|
|
|
|
|
tab_i_35_sse dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 |
|
|
dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 |
|
|
dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 |
|
|
dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 |
|
|
dw 26722, 22654, 22654, -5315 ; w21 w20 w17 w16 |
|
|
dw 15137, 5315, -26722, -15137 ; w23 w22 w19 w18 |
|
|
dw 15137, -26722, 5315, -15137 ; w29 w28 w25 w24 |
|
|
dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 |
|
|
|
|
373 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
374 |
|
|
375 |
; |
%macro DCT_8_INV_ROW_XMM 4 |
|
; DCT_8_INV_ROW_1_sse INP, OUT, TABLE, ROUNDER |
|
|
; |
|
|
|
|
|
%macro DCT_8_INV_ROW_1_sse 4 |
|
|
|
|
376 |
movq mm0, [%1] ; 0 ; x3 x2 x1 x0 |
movq mm0, [%1] ; 0 ; x3 x2 x1 x0 |
|
|
|
377 |
movq mm1, [%1+8] ; 1 ; x7 x6 x5 x4 |
movq mm1, [%1+8] ; 1 ; x7 x6 x5 x4 |
378 |
movq mm2, mm0 ; 2 ; x3 x2 x1 x0 |
movq mm2, mm0 ; 2 ; x3 x2 x1 x0 |
|
|
|
379 |
movq mm3, [%3] ; 3 ; w05 w04 w01 w00 |
movq mm3, [%3] ; 3 ; w05 w04 w01 w00 |
380 |
pshufw mm0, mm0, 10001000b ; x2 x0 x2 x0 |
pshufw mm0, mm0, 10001000b ; x2 x0 x2 x0 |
|
|
|
381 |
movq mm4, [%3+8] ; 4 ; w07 w06 w03 w02 |
movq mm4, [%3+8] ; 4 ; w07 w06 w03 w02 |
382 |
movq mm5, mm1 ; 5 ; x7 x6 x5 x4 |
movq mm5, mm1 ; 5 ; x7 x6 x5 x4 |
383 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
|
|
|
384 |
movq mm6, [%3+32] ; 6 ; w21 w20 w17 w16 |
movq mm6, [%3+32] ; 6 ; w21 w20 w17 w16 |
385 |
pshufw mm1, mm1, 10001000b ; x6 x4 x6 x4 |
pshufw mm1, mm1, 10001000b ; x6 x4 x6 x4 |
386 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
|
|
|
387 |
movq mm7, [%3+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [%3+40] ; 7 ; w23 w22 w19 w18 |
388 |
pshufw mm2, mm2, 11011101b ; x3 x1 x3 x1 |
pshufw mm2, mm2, 11011101b ; x3 x1 x3 x1 |
389 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
|
|
|
390 |
pshufw mm5, mm5, 11011101b ; x7 x5 x7 x5 |
pshufw mm5, mm5, 11011101b ; x7 x5 x7 x5 |
391 |
pmaddwd mm7, mm5 ; x7*w23+x5*w22 x7*w19+x5*w18 |
pmaddwd mm7, mm5 ; x7*w23+x5*w22 x7*w19+x5*w18 |
|
|
|
392 |
paddd mm3, [%4] ; +%4 |
paddd mm3, [%4] ; +%4 |
|
|
|
393 |
pmaddwd mm0, [%3+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm0, [%3+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
394 |
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) |
|
|
|
395 |
pmaddwd mm1, [%3+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [%3+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
396 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
|
|
|
397 |
pmaddwd mm2, [%3+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [%3+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
398 |
paddd mm6, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) |
|
|
|
399 |
pmaddwd mm5, [%3+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm5, [%3+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
400 |
paddd mm3, mm6 ; a1+b1 a0+b0 |
paddd mm3, mm6 ; a1+b1 a0+b0 |
|
|
|
401 |
paddd mm0, [%4] ; +%4 |
paddd mm0, [%4] ; +%4 |
402 |
psrad mm3, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm3, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
|
|
|
403 |
paddd mm0, mm1 ; 1 ; a3=sum(even3) a2=sum(even2) |
paddd mm0, mm1 ; 1 ; a3=sum(even3) a2=sum(even2) |
404 |
psubd mm4, mm6 ; 6 ; a1-b1 a0-b0 |
psubd mm4, mm6 ; 6 ; a1-b1 a0-b0 |
|
|
|
405 |
movq mm7, mm0 ; 7 ; a3 a2 |
movq mm7, mm0 ; 7 ; a3 a2 |
406 |
paddd mm2, mm5 ; 5 ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm5 ; 5 ; b3=sum(odd3) b2=sum(odd2) |
|
|
|
407 |
paddd mm0, mm2 ; a3+b3 a2+b2 |
paddd mm0, mm2 ; a3+b3 a2+b2 |
408 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
|
|
|
409 |
psubd mm7, mm2 ; 2 ; a3-b3 a2-b2 |
psubd mm7, mm2 ; 2 ; a3-b3 a2-b2 |
410 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
|
|
|
411 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
|
|
|
412 |
packssdw mm3, mm0 ; 0 ; y3 y2 y1 y0 |
packssdw mm3, mm0 ; 0 ; y3 y2 y1 y0 |
|
|
|
413 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
|
|
|
414 |
movq [%2], mm3 ; 3 ; save y3 y2 y1 y0 |
movq [%2], mm3 ; 3 ; save y3 y2 y1 y0 |
415 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
|
|
|
416 |
movq [%2+8], mm7 ; 7 ; save y7 y6 y5 y4 |
movq [%2+8], mm7 ; 7 ; save y7 y6 y5 y4 |
|
|
|
417 |
%endmacro |
%endmacro |
418 |
|
|
419 |
|
;----------------------------------------------------------------------------- |
|
;============================================================================= |
|
|
; |
|
|
;============================================================================= |
|
|
|
|
|
;============================================================================= |
|
420 |
; |
; |
421 |
; The first stage DCT 8x8 - forward DCTs of columns |
; The first stage DCT 8x8 - forward DCTs of columns |
422 |
; |
; |
473 |
; y[3] = tm765 - tm465 * tg_3_16; |
; y[3] = tm765 - tm465 * tg_3_16; |
474 |
;} |
;} |
475 |
; |
; |
476 |
;============================================================================= |
;----------------------------------------------------------------------------- |
|
|
|
|
|
|
|
; |
|
|
; DCT_8_FRW_COL_4 INP, OUT |
|
|
; |
|
|
|
|
|
%macro DCT_8_FRW_COL_4 2 |
|
|
|
|
|
LOCAL x0, x1, x2, x3, x4, x5, x6, x7 |
|
|
LOCAL y0, y1, y2, y3, y4, y5, y6, y7 |
|
|
x0 equ [%1 + 0*16] |
|
|
x1 equ [%1 + 1*16] |
|
|
x2 equ [%1 + 2*16] |
|
|
x3 equ [%1 + 3*16] |
|
|
x4 equ [%1 + 4*16] |
|
|
x5 equ [%1 + 5*16] |
|
|
x6 equ [%1 + 6*16] |
|
|
x7 equ [%1 + 7*16] |
|
|
y0 equ [%2 + 0*16] |
|
|
y1 equ [%2 + 1*16] |
|
|
y2 equ [%2 + 2*16] |
|
|
y3 equ [%2 + 3*16] |
|
|
y4 equ [%2 + 4*16] |
|
|
y5 equ [%2 + 5*16] |
|
|
y6 equ [%2 + 6*16] |
|
|
y7 equ [%2 + 7*16] |
|
|
movq mm0, x1 ; 0 ; x1 |
|
|
movq mm1, x6 ; 1 ; x6 |
|
|
movq mm2, mm0 ; 2 ; x1 |
|
|
movq mm3, x2 ; 3 ; x2 |
|
|
paddsw mm0, mm1 ; t1 = x[1] + x[6] |
|
|
movq mm4, x5 ; 4 ; x5 |
|
|
psllw mm0, SHIFT_FRW_COL ; t1 |
|
|
movq mm5, x0 ; 5 ; x0 |
|
|
paddsw mm4, mm3 ; t2 = x[2] + x[5] |
|
|
paddsw mm5, x7 ; t0 = x[0] + x[7] |
|
|
psllw mm4, SHIFT_FRW_COL ; t2 |
|
|
movq mm6, mm0 ; 6 ; t1 |
|
|
psubsw mm2, mm1 ; 1 ; t6 = x[1] - x[6] |
|
|
movq mm1, [tg_2_16] ; 1 ; tg_2_16 |
|
|
psubsw mm0, mm4 ; tm12 = t1 - t2 |
|
|
movq mm7, x3 ; 7 ; x3 |
|
|
pmulhw mm1, mm0 ; tm12*tg_2_16 |
|
|
paddsw mm7, x4 ; t3 = x[3] + x[4] |
|
|
psllw mm5, SHIFT_FRW_COL ; t0 |
|
|
paddsw mm6, mm4 ; 4 ; tp12 = t1 + t2 |
|
|
psllw mm7, SHIFT_FRW_COL ; t3 |
|
|
movq mm4, mm5 ; 4 ; t0 |
|
|
psubsw mm5, mm7 ; tm03 = t0 - t3 |
|
|
paddsw mm1, mm5 ; y2 = tm03 + tm12*tg_2_16 |
|
|
paddsw mm4, mm7 ; 7 ; tp03 = t0 + t3 |
|
|
por mm1, [one_corr] ; correction y2 +0.5 |
|
|
psllw mm2, SHIFT_FRW_COL+1 ; t6 |
|
|
pmulhw mm5, [tg_2_16] ; tm03*tg_2_16 |
|
|
movq mm7, mm4 ; 7 ; tp03 |
|
|
psubsw mm3, x5 ; t5 = x[2] - x[5] |
|
|
psubsw mm4, mm6 ; y4 = tp03 - tp12 |
|
|
movq y2, mm1 ; 1 ; save y2 |
|
|
paddsw mm7, mm6 ; 6 ; y0 = tp03 + tp12 |
|
|
movq mm1, x3 ; 1 ; x3 |
|
|
psllw mm3, SHIFT_FRW_COL+1 ; t5 |
|
|
psubsw mm1, x4 ; t4 = x[3] - x[4] |
|
|
movq mm6, mm2 ; 6 ; t6 |
|
|
movq y4, mm4 ; 4 ; save y4 |
|
|
paddsw mm2, mm3 ; t6 + t5 |
|
|
pmulhw mm2, [ocos_4_16] ; tp65 = (t6 + t5)*cos_4_16 |
|
|
psubsw mm6, mm3 ; 3 ; t6 - t5 |
|
|
pmulhw mm6, [ocos_4_16] ; tm65 = (t6 - t5)*cos_4_16 |
|
|
psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12 |
|
|
por mm5, [one_corr] ; correction y6 +0.5 |
|
|
psllw mm1, SHIFT_FRW_COL ; t4 |
|
|
por mm2, [one_corr] ; correction tp65 +0.5 |
|
|
movq mm4, mm1 ; 4 ; t4 |
|
|
movq mm3, x0 ; 3 ; x0 |
|
|
paddsw mm1, mm6 ; tp465 = t4 + tm65 |
|
|
psubsw mm3, x7 ; t7 = x[0] - x[7] |
|
|
psubsw mm4, mm6 ; 6 ; tm465 = t4 - tm65 |
|
|
movq mm0, [tg_1_16] ; 0 ; tg_1_16 |
|
|
psllw mm3, SHIFT_FRW_COL ; t7 |
|
|
movq mm6, [tg_3_16] ; 6 ; tg_3_16 |
|
|
pmulhw mm0, mm1 ; tp465*tg_1_16 |
|
|
movq y0, mm7 ; 7 ; save y0 |
|
|
pmulhw mm6, mm4 ; tm465*tg_3_16 |
|
|
movq y6, mm5 ; 5 ; save y6 |
|
|
movq mm7, mm3 ; 7 ; t7 |
|
|
movq mm5, [tg_3_16] ; 5 ; tg_3_16 |
|
|
psubsw mm7, mm2 ; tm765 = t7 - tp65 |
|
|
paddsw mm3, mm2 ; 2 ; tp765 = t7 + tp65 |
|
|
pmulhw mm5, mm7 ; tm765*tg_3_16 |
|
|
paddsw mm0, mm3 ; y1 = tp765 + tp465*tg_1_16 |
|
|
paddsw mm6, mm4 ; tm465*tg_3_16 |
|
|
pmulhw mm3, [tg_1_16] ; tp765*tg_1_16 |
|
|
por mm0, [one_corr] ; correction y1 +0.5 |
|
|
paddsw mm5, mm7 ; tm765*tg_3_16 |
|
|
psubsw mm7, mm6 ; 6 ; y3 = tm765 - tm465*tg_3_16 |
|
|
movq y1, mm0 ; 0 ; save y1 |
|
|
paddsw mm5, mm4 ; 4 ; y5 = tm765*tg_3_16 + tm465 |
|
|
movq y3, mm7 ; 7 ; save y3 |
|
|
psubsw mm3, mm1 ; 1 ; y7 = tp765*tg_1_16 - tp465 |
|
|
movq y5, mm5 ; 5 ; save y5 |
|
|
movq y7, mm3 ; 3 ; save y7 |
|
|
%endmacro |
|
|
|
|
477 |
|
|
478 |
; |
;----------------------------------------------------------------------------- |
479 |
; DCT_8_INV_COL_4 INP,OUT |
; DCT_8_INV_COL_4 INP,OUT |
480 |
; |
;----------------------------------------------------------------------------- |
481 |
|
|
482 |
%macro DCT_8_INV_COL_4 2 |
%macro DCT_8_INV_COL 2 |
483 |
movq mm0, [tg_3_16] |
movq mm0, [tg_3_16] |
|
|
|
484 |
movq mm3, [%1+16*3] |
movq mm3, [%1+16*3] |
485 |
movq mm1, mm0 ; tg_3_16 |
movq mm1, mm0 ; tg_3_16 |
|
|
|
486 |
movq mm5, [%1+16*5] |
movq mm5, [%1+16*5] |
487 |
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
|
|
|
488 |
movq mm4, [tg_1_16] |
movq mm4, [tg_1_16] |
489 |
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
|
|
|
490 |
movq mm7, [%1+16*7] |
movq mm7, [%1+16*7] |
491 |
movq mm2, mm4 ; tg_1_16 |
movq mm2, mm4 ; tg_1_16 |
|
|
|
492 |
movq mm6, [%1+16*1] |
movq mm6, [%1+16*1] |
493 |
pmulhw mm4, mm7 ; x7*tg_1_16 |
pmulhw mm4, mm7 ; x7*tg_1_16 |
|
|
|
494 |
paddsw mm0, mm3 ; x3*tg_3_16 |
paddsw mm0, mm3 ; x3*tg_3_16 |
495 |
pmulhw mm2, mm6 ; x1*tg_1_16 |
pmulhw mm2, mm6 ; x1*tg_1_16 |
|
|
|
496 |
paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) |
paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) |
497 |
psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 |
psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 |
|
|
|
498 |
movq mm3, [ocos_4_16] |
movq mm3, [ocos_4_16] |
499 |
paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
|
|
|
500 |
paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 |
paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 |
501 |
psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 |
psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 |
|
|
|
502 |
movq mm5, mm4 ; tp17 |
movq mm5, mm4 ; tp17 |
503 |
movq mm6, mm2 ; tm17 |
movq mm6, mm2 ; tm17 |
|
|
|
504 |
paddsw mm5, mm1 ; tp17+tp35 = b0 |
paddsw mm5, mm1 ; tp17+tp35 = b0 |
505 |
psubsw mm6, mm0 ; tm17-tm35 = b3 |
psubsw mm6, mm0 ; tm17-tm35 = b3 |
|
|
|
506 |
psubsw mm4, mm1 ; tp17-tp35 = t1 |
psubsw mm4, mm1 ; tp17-tp35 = t1 |
507 |
paddsw mm2, mm0 ; tm17+tm35 = t2 |
paddsw mm2, mm0 ; tm17+tm35 = t2 |
|
|
|
508 |
movq mm7, [tg_2_16] |
movq mm7, [tg_2_16] |
509 |
movq mm1, mm4 ; t1 |
movq mm1, mm4 ; t1 |
|
|
|
510 |
; movq [SCRATCH+0], mm5 ; save b0 |
; movq [SCRATCH+0], mm5 ; save b0 |
511 |
movq [%2+3*16], mm5 ; save b0 |
movq [%2+3*16], mm5 ; save b0 |
512 |
paddsw mm1, mm2 ; t1+t2 |
paddsw mm1, mm2 ; t1+t2 |
|
|
|
513 |
; movq [SCRATCH+8], mm6 ; save b3 |
; movq [SCRATCH+8], mm6 ; save b3 |
514 |
movq [%2+5*16], mm6 ; save b3 |
movq [%2+5*16], mm6 ; save b3 |
515 |
psubsw mm4, mm2 ; t1-t2 |
psubsw mm4, mm2 ; t1-t2 |
|
|
|
516 |
movq mm5, [%1+2*16] |
movq mm5, [%1+2*16] |
517 |
movq mm0, mm7 ; tg_2_16 |
movq mm0, mm7 ; tg_2_16 |
|
|
|
518 |
movq mm6, [%1+6*16] |
movq mm6, [%1+6*16] |
519 |
pmulhw mm0, mm5 ; x2*tg_2_16 |
pmulhw mm0, mm5 ; x2*tg_2_16 |
|
|
|
520 |
pmulhw mm7, mm6 ; x6*tg_2_16 |
pmulhw mm7, mm6 ; x6*tg_2_16 |
521 |
; slot |
; slot |
522 |
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
523 |
; slot |
; slot |
524 |
movq mm2, [%1+0*16] |
movq mm2, [%1+0*16] |
525 |
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
|
|
|
526 |
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
527 |
movq mm3, mm2 ; x0 |
movq mm3, mm2 ; x0 |
|
|
|
528 |
movq mm6, [%1+4*16] |
movq mm6, [%1+4*16] |
529 |
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
|
|
|
530 |
paddsw mm2, mm6 ; x0+x4 = tp04 |
paddsw mm2, mm6 ; x0+x4 = tp04 |
531 |
psubsw mm3, mm6 ; x0-x4 = tm04 |
psubsw mm3, mm6 ; x0-x4 = tm04 |
|
|
|
532 |
movq mm5, mm2 ; tp04 |
movq mm5, mm2 ; tp04 |
533 |
movq mm6, mm3 ; tm04 |
movq mm6, mm3 ; tm04 |
|
|
|
534 |
psubsw mm2, mm7 ; tp04-tp26 = a3 |
psubsw mm2, mm7 ; tp04-tp26 = a3 |
535 |
paddsw mm3, mm0 ; tm04+tm26 = a1 |
paddsw mm3, mm0 ; tm04+tm26 = a1 |
|
|
|
536 |
paddsw mm1, mm1 ; b1 |
paddsw mm1, mm1 ; b1 |
537 |
paddsw mm4, mm4 ; b2 |
paddsw mm4, mm4 ; b2 |
|
|
|
538 |
paddsw mm5, mm7 ; tp04+tp26 = a0 |
paddsw mm5, mm7 ; tp04+tp26 = a0 |
539 |
psubsw mm6, mm0 ; tm04-tm26 = a2 |
psubsw mm6, mm0 ; tm04-tm26 = a2 |
|
|
|
540 |
movq mm7, mm3 ; a1 |
movq mm7, mm3 ; a1 |
541 |
movq mm0, mm6 ; a2 |
movq mm0, mm6 ; a2 |
|
|
|
542 |
paddsw mm3, mm1 ; a1+b1 |
paddsw mm3, mm1 ; a1+b1 |
543 |
paddsw mm6, mm4 ; a2+b2 |
paddsw mm6, mm4 ; a2+b2 |
|
|
|
544 |
psraw mm3, SHIFT_INV_COL ; dst1 |
psraw mm3, SHIFT_INV_COL ; dst1 |
545 |
psubsw mm7, mm1 ; a1-b1 |
psubsw mm7, mm1 ; a1-b1 |
|
|
|
546 |
psraw mm6, SHIFT_INV_COL ; dst2 |
psraw mm6, SHIFT_INV_COL ; dst2 |
547 |
psubsw mm0, mm4 ; a2-b2 |
psubsw mm0, mm4 ; a2-b2 |
|
|
|
548 |
; movq mm1, [SCRATCH+0] ; load b0 |
; movq mm1, [SCRATCH+0] ; load b0 |
549 |
movq mm1, [%2+3*16] ; load b0 |
movq mm1, [%2+3*16] ; load b0 |
550 |
psraw mm7, SHIFT_INV_COL ; dst6 |
psraw mm7, SHIFT_INV_COL ; dst6 |
|
|
|
551 |
movq mm4, mm5 ; a0 |
movq mm4, mm5 ; a0 |
552 |
psraw mm0, SHIFT_INV_COL ; dst5 |
psraw mm0, SHIFT_INV_COL ; dst5 |
|
|
|
553 |
movq [%2+1*16], mm3 |
movq [%2+1*16], mm3 |
554 |
paddsw mm5, mm1 ; a0+b0 |
paddsw mm5, mm1 ; a0+b0 |
|
|
|
555 |
movq [%2+2*16], mm6 |
movq [%2+2*16], mm6 |
556 |
psubsw mm4, mm1 ; a0-b0 |
psubsw mm4, mm1 ; a0-b0 |
|
|
|
557 |
; movq mm3, [SCRATCH+8] ; load b3 |
; movq mm3, [SCRATCH+8] ; load b3 |
558 |
movq mm3, [%2+5*16] ; load b3 |
movq mm3, [%2+5*16] ; load b3 |
559 |
psraw mm5, SHIFT_INV_COL ; dst0 |
psraw mm5, SHIFT_INV_COL ; dst0 |
|
|
|
560 |
movq mm6, mm2 ; a3 |
movq mm6, mm2 ; a3 |
561 |
psraw mm4, SHIFT_INV_COL ; dst7 |
psraw mm4, SHIFT_INV_COL ; dst7 |
|
|
|
562 |
movq [%2+5*16], mm0 |
movq [%2+5*16], mm0 |
563 |
paddsw mm2, mm3 ; a3+b3 |
paddsw mm2, mm3 ; a3+b3 |
|
|
|
564 |
movq [%2+6*16], mm7 |
movq [%2+6*16], mm7 |
565 |
psubsw mm6, mm3 ; a3-b3 |
psubsw mm6, mm3 ; a3-b3 |
|
|
|
566 |
movq [%2+0*16], mm5 |
movq [%2+0*16], mm5 |
567 |
psraw mm2, SHIFT_INV_COL ; dst3 |
psraw mm2, SHIFT_INV_COL ; dst3 |
|
|
|
568 |
movq [%2+7*16], mm4 |
movq [%2+7*16], mm4 |
569 |
psraw mm6, SHIFT_INV_COL ; dst4 |
psraw mm6, SHIFT_INV_COL ; dst4 |
|
|
|
570 |
movq [%2+3*16], mm2 |
movq [%2+3*16], mm2 |
|
|
|
571 |
movq [%2+4*16], mm6 |
movq [%2+4*16], mm6 |
572 |
%endmacro |
%endmacro |
573 |
|
|
|
|
|
|
|
|
|
section .text |
|
|
|
|
574 |
;============================================================================= |
;============================================================================= |
575 |
; |
; Code |
|
; void idct_mmx (short * const src_result); |
|
|
; |
|
576 |
;============================================================================= |
;============================================================================= |
577 |
|
|
578 |
align 16 |
SECTION .text |
|
cglobal idct_mmx |
|
|
idct_mmx |
|
|
mov eax, dword [esp + 4] |
|
|
|
|
|
DCT_8_INV_ROW_1 eax+0, eax+0, tab_i_04, rounder_0 |
|
|
DCT_8_INV_ROW_1 eax+16, eax+16, tab_i_17, rounder_1 |
|
|
DCT_8_INV_ROW_1 eax+32, eax+32, tab_i_26, rounder_2 |
|
|
DCT_8_INV_ROW_1 eax+48, eax+48, tab_i_35, rounder_3 |
|
|
DCT_8_INV_ROW_1 eax+64, eax+64, tab_i_04, rounder_4 |
|
|
DCT_8_INV_ROW_1 eax+80, eax+80, tab_i_35, rounder_5 |
|
|
DCT_8_INV_ROW_1 eax+96, eax+96, tab_i_26, rounder_6 |
|
|
DCT_8_INV_ROW_1 eax+112, eax+112, tab_i_17, rounder_7 |
|
|
|
|
|
DCT_8_INV_COL_4 eax+0,eax+0 |
|
|
DCT_8_INV_COL_4 eax+8,eax+8 |
|
|
|
|
|
ret |
|
|
|
|
579 |
|
|
580 |
|
cglobal idct_mmx |
581 |
|
cglobal idct_xmm |
582 |
|
|
583 |
;============================================================================= |
;----------------------------------------------------------------------------- |
584 |
; |
; void idct_mmx(uint16_t block[64]); |
585 |
; void idct_sse (short * const src_result); |
;----------------------------------------------------------------------------- |
|
; |
|
|
;============================================================================= |
|
586 |
|
|
587 |
align 16 |
ALIGN 16 |
588 |
cglobal idct_xmm |
idct_mmx: |
|
idct_xmm |
|
589 |
mov eax, dword [esp + 4] |
mov eax, dword [esp + 4] |
590 |
|
|
591 |
DCT_8_INV_ROW_1_sse eax+0, eax+0, tab_i_04_sse, rounder_0 |
;; Process each row |
592 |
DCT_8_INV_ROW_1_sse eax+16, eax+16, tab_i_17_sse, rounder_1 |
DCT_8_INV_ROW_MMX eax+0*16, eax+0*16, tab_i_04_mmx, rounder_0 |
593 |
DCT_8_INV_ROW_1_sse eax+32, eax+32, tab_i_26_sse, rounder_2 |
DCT_8_INV_ROW_MMX eax+1*16, eax+1*16, tab_i_17_mmx, rounder_1 |
594 |
DCT_8_INV_ROW_1_sse eax+48, eax+48, tab_i_35_sse, rounder_3 |
DCT_8_INV_ROW_MMX eax+2*16, eax+2*16, tab_i_26_mmx, rounder_2 |
595 |
DCT_8_INV_ROW_1_sse eax+64, eax+64, tab_i_04_sse, rounder_4 |
DCT_8_INV_ROW_MMX eax+3*16, eax+3*16, tab_i_35_mmx, rounder_3 |
596 |
DCT_8_INV_ROW_1_sse eax+80, eax+80, tab_i_35_sse, rounder_5 |
DCT_8_INV_ROW_MMX eax+4*16, eax+4*16, tab_i_04_mmx, rounder_4 |
597 |
DCT_8_INV_ROW_1_sse eax+96, eax+96, tab_i_26_sse, rounder_6 |
DCT_8_INV_ROW_MMX eax+5*16, eax+5*16, tab_i_35_mmx, rounder_5 |
598 |
DCT_8_INV_ROW_1_sse eax+112, eax+112, tab_i_17_sse, rounder_7 |
DCT_8_INV_ROW_MMX eax+6*16, eax+6*16, tab_i_26_mmx, rounder_6 |
599 |
|
DCT_8_INV_ROW_MMX eax+7*16, eax+7*16, tab_i_17_mmx, rounder_7 |
600 |
DCT_8_INV_COL_4 eax+0, eax+0 |
|
601 |
DCT_8_INV_COL_4 eax+8, eax+8 |
;; Process the columns (4 at a time) |
602 |
|
DCT_8_INV_COL eax+0, eax+0 |
603 |
|
DCT_8_INV_COL eax+8, eax+8 |
604 |
|
|
605 |
ret |
ret |
606 |
|
|
607 |
;============================================================================= |
;----------------------------------------------------------------------------- |
608 |
; The code below this line is for SSE2-equipped processors |
; void idct_xmm(uint16_t block[64]); |
609 |
; By Dmitry Rozhdestvensky |
;----------------------------------------------------------------------------- |
|
;============================================================================= |
|
|
|
|
|
section .data |
|
|
|
|
|
align 16 |
|
|
|
|
|
tab_i_04_s2 dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 |
|
|
dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 |
|
|
dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 |
|
|
dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 |
|
|
dw 22725, 19266, 19266, -4520 ; w21 w20 w17 w16 |
|
|
dw 12873, -22725, 4520, -12873 ; w29 w28 w25 w24 |
|
|
dw 12873, 4520, -22725, -12873 ; w23 w22 w19 w18 |
|
|
dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 |
|
|
|
|
|
; Table for rows 1,7 - constants are multiplied by cos_1_16 |
|
|
|
|
|
tab_i_17_s2 dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 |
|
|
dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 |
|
|
dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 |
|
|
dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 |
|
|
dw 31521, 26722, 26722, -6270 ; w21 w20 w17 w16 |
|
|
dw 17855, -31521, 6270, -17855 ; w29 w28 w25 w24 |
|
|
dw 17855, 6270, -31521, -17855 ; w23 w22 w19 w18 |
|
|
dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 |
|
|
|
|
|
; Table for rows 2,6 - constants are multiplied by cos_2_16 |
|
|
|
|
|
tab_i_26_s2 dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 |
|
|
dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 |
|
|
dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 |
|
|
dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 |
|
|
dw 29692, 25172, 25172, -5906 ; w21 w20 w17 w16 |
|
|
dw 16819, -29692, 5906, -16819 ; w29 w28 w25 w24 |
|
|
dw 16819, 5906, -29692, -16819 ; w23 w22 w19 w18 |
|
|
dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 |
|
|
|
|
|
; Table for rows 3,5 - constants are multiplied by cos_3_16 |
|
|
|
|
|
tab_i_35_s2 dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 |
|
|
dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 |
|
|
dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 |
|
|
dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 |
|
|
dw 26722, 22654, 22654, -5315 ; w21 w20 w17 w16 |
|
|
dw 15137, -26722, 5315, -15137 ; w29 w28 w25 w24 |
|
|
dw 15137, 5315, -26722, -15137 ; w23 w22 w19 w18 |
|
|
dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 |
|
|
|
|
|
%if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 |
|
|
rounder_2_0 dd 65536, 65536 |
|
|
dd 65536, 65536 |
|
|
rounder_2_4 dd 0, 0 |
|
|
dd 0, 0 |
|
|
rounder_2_1 dd 7195, 7195 |
|
|
dd 7195, 7195 |
|
|
rounder_2_7 dd 1024, 1024 |
|
|
dd 1024, 1024 |
|
|
rounder_2_2 dd 4520, 4520 |
|
|
dd 4520, 4520 |
|
|
rounder_2_6 dd 1024, 1024 |
|
|
dd 1024, 1024 |
|
|
rounder_2_3 dd 2407, 2407 |
|
|
dd 2407, 2407 |
|
|
rounder_2_5 dd 240, 240 |
|
|
dd 240, 240 |
|
|
|
|
|
%elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 |
|
|
rounder_2_0 dd 65536, 65536 |
|
|
dd 65536, 65536 |
|
|
rounder_2_4 dd 0, 0 |
|
|
dd 0, 0 |
|
|
rounder_2_1 dd 3597, 3597 |
|
|
dd 3597, 3597 |
|
|
rounder_2_7 dd 512, 512 |
|
|
dd 512, 512 |
|
|
rounder_2_2 dd 2260, 2260 |
|
|
dd 2260, 2260 |
|
|
rounder_2_6 dd 512, 512 |
|
|
dd 512, 512 |
|
|
rounder_2_3 dd 1203, 1203 |
|
|
dd 1203, 1203 |
|
|
rounder_2_5 dd 120, 120 |
|
|
dd 120, 120 |
|
|
%else |
|
|
|
|
|
%error invalid _SHIFT_INV_ROW_ |
|
|
|
|
|
%endif |
|
|
|
|
|
tg_1_16_2 dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 |
|
|
dw 13036, 13036, 13036, 13036 |
|
|
tg_2_16_2 dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 |
|
|
dw 27146, 27146, 27146, 27146 |
|
|
tg_3_16_2 dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 |
|
|
dw -21746, -21746, -21746, -21746 |
|
|
ocos_4_16_2 dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 |
|
|
dw 23170, 23170, 23170, 23170 |
|
|
|
|
|
%macro DCT_8_INV_ROW_1_sse2 4 |
|
|
|
|
|
pshufhw xmm1,[%1],11011000b ;x 75643210 |
|
|
pshuflw xmm1,xmm1,11011000b ;x 75643120 |
|
|
pshufd xmm0,xmm1,00000000b ;x 20202020 |
|
|
pmaddwd xmm0,[%3] ;w 13 12 9 8 5410 |
|
|
;a 3210 first part |
|
|
|
|
|
pshufd xmm2,xmm1,10101010b ;x 64646464 |
|
|
pmaddwd xmm2,[%3+16] ;w 15 14 11 10 7632 |
|
|
;a 3210 second part |
|
|
|
|
|
paddd xmm2,xmm0 ;a 3210 ready |
|
|
paddd xmm2,[%4] ;must be 4 dwords long, not 2 as for sse1 |
|
|
movdqa xmm5,xmm2 |
|
|
|
|
|
pshufd xmm3,xmm1,01010101b ;x 31313131 |
|
|
pmaddwd xmm3,[%3+32] ;w 29 28 25 24 21 20 17 16 |
|
|
;b 3210 first part |
|
|
|
|
|
pshufd xmm4,xmm1,11111111b ;x 75757575 |
|
|
pmaddwd xmm4,[%3+48] ;w 31 30 27 26 23 22 19 18 |
|
|
;b 3210 second part |
|
|
paddd xmm3,xmm4 ;b 3210 ready |
|
|
|
|
|
paddd xmm2,xmm3 ;will be y 3210 |
|
|
psubd xmm5,xmm3 ;will be y 4567 |
|
|
psrad xmm2,SHIFT_INV_ROW |
|
|
psrad xmm5,SHIFT_INV_ROW |
|
|
packssdw xmm2,xmm5 ;y 45673210 |
|
|
pshufhw xmm6,xmm2,00011011b ;y 76543210 |
|
|
movdqa [%2],xmm6 |
|
|
|
|
|
%endmacro |
|
|
|
|
|
%macro DCT_8_INV_COL_4_sse2 2 |
|
|
|
|
|
movdqa xmm0,[%1+16*0] ;x0 (all columns) |
|
|
movdqa xmm2,[%1+16*4] ;x4 |
|
|
movdqa xmm1,xmm0 |
|
|
|
|
|
movdqa xmm4,[%1+16*2] ;x2 |
|
|
movdqa xmm5,[%1+16*6] ;x6 |
|
|
movdqa xmm6,[tg_2_16_2] |
|
|
movdqa xmm7,xmm6 |
|
|
|
|
|
paddsw xmm0,xmm2 ;u04=x0+x4 |
|
|
psubsw xmm1,xmm2 ;v04=x0-x4 |
|
|
movdqa xmm3,xmm0 |
|
|
movdqa xmm2,xmm1 |
|
|
|
|
|
pmulhw xmm6,xmm4 |
|
|
pmulhw xmm7,xmm5 |
|
|
psubsw xmm6,xmm5 ;v26=x2*T2-x6 |
|
|
paddsw xmm7,xmm4 ;u26=x6*T2+x2 |
|
|
|
|
|
paddsw xmm1,xmm6 ;a1=v04+v26 |
|
|
paddsw xmm0,xmm7 ;a0=u04+u26 |
|
|
psubsw xmm2,xmm6 ;a2=v04-v26 |
|
|
psubsw xmm3,xmm7 ;a3=u04-u26 |
|
|
|
|
|
movdqa [%2+16*0],xmm0 ;store a3-a0 to |
|
|
movdqa [%2+16*6],xmm1 ;free registers |
|
|
movdqa [%2+16*2],xmm2 |
|
|
movdqa [%2+16*4],xmm3 |
|
|
|
|
|
movdqa xmm0,[%1+16*1] ;x1 |
|
|
movdqa xmm1,[%1+16*7] ;x7 |
|
|
movdqa xmm2,[tg_1_16_2] |
|
|
movdqa xmm3,xmm2 |
|
|
|
|
|
movdqa xmm4,[%1+16*3] ;x3 |
|
|
movdqa xmm5,[%1+16*5] ;x5 |
|
|
movdqa xmm6,[tg_3_16_2] |
|
|
movdqa xmm7,xmm6 |
|
|
|
|
|
pmulhw xmm2,xmm0 |
|
|
pmulhw xmm3,xmm1 |
|
|
psubsw xmm2,xmm1 ;v17=x1*T1-x7 |
|
|
paddsw xmm3,xmm0 ;u17=x7*T1+x1 |
|
|
movdqa xmm0,xmm3 ;u17 |
|
|
movdqa xmm1,xmm2 ;v17 |
|
|
|
|
|
pmulhw xmm6,xmm4 ;x3*(t3-1) |
|
|
pmulhw xmm7,xmm5 ;x5*(t3-1) |
|
|
paddsw xmm6,xmm4 |
|
|
paddsw xmm7,xmm5 |
|
|
psubsw xmm6,xmm5 ;v35=x3*T3-x5 |
|
|
paddsw xmm7,xmm4 ;u35=x5*T3+x3 |
|
|
|
|
|
movdqa xmm4,[ocos_4_16_2] |
|
|
|
|
|
paddsw xmm0,xmm7 ;b0=u17+u35 |
|
|
psubsw xmm1,xmm6 ;b3=v17-v35 |
|
|
psubsw xmm3,xmm7 ;u12=u17-v35 |
|
|
paddsw xmm2,xmm6 ;v12=v17+v35 |
|
|
|
|
|
movdqa xmm5,xmm3 |
|
|
paddsw xmm3,xmm2 ;tb1 |
|
|
psubsw xmm5,xmm2 ;tb2 |
|
|
pmulhw xmm5,xmm4 |
|
|
pmulhw xmm4,xmm3 |
|
|
paddsw xmm5,xmm5 |
|
|
paddsw xmm4,xmm4 |
|
|
|
|
|
movdqa xmm6,[%2+16*0] ;a0 |
|
|
movdqa xmm7,xmm6 |
|
|
movdqa xmm2,[%2+16*4] ;a3 |
|
|
movdqa xmm3,xmm2 |
|
|
|
|
|
paddsw xmm6,xmm0 |
|
|
psubsw xmm7,xmm0 |
|
|
psraw xmm6,SHIFT_INV_COL ;y0=a0+b0 |
|
|
psraw xmm7,SHIFT_INV_COL ;y7=a0-b0 |
|
|
movdqa [%2+16*0],xmm6 |
|
|
movdqa [%2+16*7],xmm7 |
|
|
|
|
|
paddsw xmm2,xmm1 |
|
|
psubsw xmm3,xmm1 |
|
|
psraw xmm2,SHIFT_INV_COL ;y3=a3+b3 |
|
|
psraw xmm3,SHIFT_INV_COL ;y4=a3-b3 |
|
|
movdqa [%2+16*3],xmm2 |
|
|
movdqa [%2+16*4],xmm3 |
|
|
|
|
|
movdqa xmm0,[%2+16*6] ;a1 |
|
|
movdqa xmm1,xmm0 |
|
|
movdqa xmm6,[%2+16*2] ;a2 |
|
|
movdqa xmm7,xmm6 |
|
|
|
|
|
|
|
|
paddsw xmm0,xmm4 |
|
|
psubsw xmm1,xmm4 |
|
|
psraw xmm0,SHIFT_INV_COL ;y1=a1+b1 |
|
|
psraw xmm1,SHIFT_INV_COL ;y6=a1-b1 |
|
|
movdqa [%2+16*1],xmm0 |
|
|
movdqa [%2+16*6],xmm1 |
|
|
|
|
|
paddsw xmm6,xmm5 |
|
|
psubsw xmm7,xmm5 |
|
|
psraw xmm6,SHIFT_INV_COL ;y2=a2+b2 |
|
|
psraw xmm7,SHIFT_INV_COL ;y5=a2-b2 |
|
|
movdqa [%2+16*2],xmm6 |
|
|
movdqa [%2+16*5],xmm7 |
|
|
|
|
|
%endmacro |
|
|
|
|
|
section .text |
|
|
|
|
|
align 16 |
|
|
cglobal idct_sse2 |
|
|
idct_sse2 |
|
610 |
|
|
611 |
|
ALIGN 16 |
612 |
|
idct_xmm: |
613 |
mov eax, dword [esp + 4] |
mov eax, dword [esp + 4] |
614 |
|
|
615 |
DCT_8_INV_ROW_1_sse2 eax+0, eax+0, tab_i_04_s2, rounder_2_0 |
;; Process each row |
616 |
DCT_8_INV_ROW_1_sse2 eax+16, eax+16, tab_i_17_s2, rounder_2_1 |
DCT_8_INV_ROW_XMM eax+0*16, eax+0*16, tab_i_04_xmm, rounder_0 |
617 |
DCT_8_INV_ROW_1_sse2 eax+32, eax+32, tab_i_26_s2, rounder_2_2 |
DCT_8_INV_ROW_XMM eax+1*16, eax+1*16, tab_i_17_xmm, rounder_1 |
618 |
DCT_8_INV_ROW_1_sse2 eax+48, eax+48, tab_i_35_s2, rounder_2_3 |
DCT_8_INV_ROW_XMM eax+2*16, eax+2*16, tab_i_26_xmm, rounder_2 |
619 |
DCT_8_INV_ROW_1_sse2 eax+64, eax+64, tab_i_04_s2, rounder_2_4 |
DCT_8_INV_ROW_XMM eax+3*16, eax+3*16, tab_i_35_xmm, rounder_3 |
620 |
DCT_8_INV_ROW_1_sse2 eax+80, eax+80, tab_i_35_s2, rounder_2_5 |
DCT_8_INV_ROW_XMM eax+4*16, eax+4*16, tab_i_04_xmm, rounder_4 |
621 |
DCT_8_INV_ROW_1_sse2 eax+96, eax+96, tab_i_26_s2, rounder_2_6 |
DCT_8_INV_ROW_XMM eax+5*16, eax+5*16, tab_i_35_xmm, rounder_5 |
622 |
DCT_8_INV_ROW_1_sse2 eax+112, eax+112, tab_i_17_s2, rounder_2_7 |
DCT_8_INV_ROW_XMM eax+6*16, eax+6*16, tab_i_26_xmm, rounder_6 |
623 |
|
DCT_8_INV_ROW_XMM eax+7*16, eax+7*16, tab_i_17_xmm, rounder_7 |
624 |
DCT_8_INV_COL_4_sse2 eax, eax |
|
625 |
;DCT_8_INV_COL_4 eax+8, eax+8 |
;; Process the columns (4 at a time) |
626 |
|
DCT_8_INV_COL eax+0, eax+0 |
627 |
|
DCT_8_INV_COL eax+8, eax+8 |
628 |
|
|
629 |
ret |
ret |