Parent Directory | Revision Log
Revision 1.2.2.1 - (view) (download)
1 : | edgomez | 1.2.2.1 | ;/**************************************************************************** |
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 : | ; * 2002 Jaan Kalda | ||
8 : | ; * | ||
9 : | ; * This program is free software; you can redistribute it and/or modify it | ||
10 : | ; * under the terms of the GNU General Public License as published by | ||
11 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
12 : | ; * (at your option) any later version. | ||
13 : | ; * | ||
14 : | ; * This program 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 | ||
17 : | ; * GNU General Public License for more details. | ||
18 : | ; * | ||
19 : | ; * You should have received a copy of the GNU General Public License | ||
20 : | ; * along with this program; if not, write to the Free Software | ||
21 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
22 : | ; * | ||
23 : | ; * $Id$ | ||
24 : | ; * | ||
25 : | ; ***************************************************************************/ | ||
26 : | |||
27 : | ; **************************************************************************** | ||
28 : | ; | ||
29 : | edgomez | 1.2 | ; Originally provided by Intel at AP-922 |
30 : | ; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm | ||
31 : | ; (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) | ||
32 : | ; but in a limited edition. | ||
33 : | ; New macro implements a column part for precise iDCT | ||
34 : | ; The routine precision now satisfies IEEE standard 1180-1990. | ||
35 : | ; | ||
36 : | edgomez | 1.2.2.1 | ; Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
37 : | ; Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org> | ||
38 : | edgomez | 1.2 | ; |
39 : | ; http://www.elecard.com/peter/idct.html | ||
40 : | ; http://www.linuxvideo.org/mpeg2dec/ | ||
41 : | ; | ||
42 : | edgomez | 1.2.2.1 | ; ***************************************************************************/ |
43 : | edgomez | 1.2 | ; |
44 : | ; These examples contain code fragments for first stage iDCT 8x8 | ||
45 : | ; (for rows) and first stage DCT 8x8 (for columns) | ||
46 : | ; | ||
47 : | |||
48 : | edgomez | 1.2.2.1 | ; ***************************************************************************/ |
49 : | edgomez | 1.2 | ; this 3dne function is compatible with iSSE, but is optimized specifically for |
50 : | edgomez | 1.2.2.1 | ; K7 pipelines (ca 5% gain), for implementation details see the idct_mmx.asm |
51 : | ; file | ||
52 : | edgomez | 1.2 | ; |
53 : | edgomez | 1.2.2.1 | ; ---------------------------------------------------------------------------- |
54 : | ; Athlon optimizations contributed by Jaan Kalda | ||
55 : | ;----------------------------------------------------------------------------- | ||
56 : | edgomez | 1.2 | |
57 : | edgomez | 1.2.2.1 | BITS 32 |
58 : | |||
59 : | ;============================================================================= | ||
60 : | ; Macros and other preprocessor constants | ||
61 : | ;============================================================================= | ||
62 : | edgomez | 1.2 | |
63 : | %macro cglobal 1 | ||
64 : | %ifdef PREFIX | ||
65 : | global _%1 | ||
66 : | %define %1 _%1 | ||
67 : | %else | ||
68 : | global %1 | ||
69 : | %endif | ||
70 : | %endmacro | ||
71 : | |||
72 : | edgomez | 1.2.2.1 | %define BITS_INV_ACC 5 ; 4 or 5 for IEEE |
73 : | edgomez | 1.2 | %define SHIFT_INV_ROW 16 - BITS_INV_ACC |
74 : | %define SHIFT_INV_COL 1 + BITS_INV_ACC | ||
75 : | %define RND_INV_ROW 1024 * (6 - BITS_INV_ACC) ; 1 << (SHIFT_INV_ROW-1) | ||
76 : | %define RND_INV_COL 16 * (BITS_INV_ACC - 3) ; 1 << (SHIFT_INV_COL-1) | ||
77 : | edgomez | 1.2.2.1 | %define RND_INV_CORR RND_INV_COL - 1 ; correction -1.0 and round |
78 : | edgomez | 1.2 | |
79 : | edgomez | 1.2.2.1 | %define BITS_FRW_ACC 3 ; 2 or 3 for accuracy |
80 : | edgomez | 1.2 | %define SHIFT_FRW_COL BITS_FRW_ACC |
81 : | %define SHIFT_FRW_ROW BITS_FRW_ACC + 17 | ||
82 : | edgomez | 1.2.2.1 | %define RND_FRW_ROW 262144*(BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) |
83 : | edgomez | 1.2 | |
84 : | edgomez | 1.2.2.1 | ;============================================================================= |
85 : | ; Local Data (Read Only) | ||
86 : | ;============================================================================= | ||
87 : | edgomez | 1.2 | |
88 : | edgomez | 1.2.2.1 | SECTION .rodata |
89 : | edgomez | 1.2 | |
90 : | edgomez | 1.2.2.1 | ;----------------------------------------------------------------------------- |
91 : | ; Various memory constants (trigonometric values or rounding values) | ||
92 : | ;----------------------------------------------------------------------------- | ||
93 : | edgomez | 1.2 | |
94 : | edgomez | 1.2.2.1 | ALIGN 16 |
95 : | one_corr: | ||
96 : | dw 1, 1, 1, 1 | ||
97 : | round_inv_row: | ||
98 : | dd RND_INV_ROW, RND_INV_ROW | ||
99 : | round_inv_col: | ||
100 : | dw RND_INV_COL, RND_INV_COL, RND_INV_COL, RND_INV_COL | ||
101 : | round_inv_corr: | ||
102 : | dw RND_INV_CORR, RND_INV_CORR, RND_INV_CORR, RND_INV_CORR | ||
103 : | round_frw_row: | ||
104 : | dd RND_FRW_ROW, RND_FRW_ROW | ||
105 : | tg_1_16: | ||
106 : | dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 | ||
107 : | tg_2_16: | ||
108 : | dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 | ||
109 : | tg_3_16: | ||
110 : | dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 | ||
111 : | cos_4_16: | ||
112 : | dw -19195, -19195, -19195, -19195 ; cos * (2<<16) + 0.5 | ||
113 : | ocos_4_16: | ||
114 : | dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 | ||
115 : | otg_3_16: | ||
116 : | dw 21895, 21895, 21895, 21895 ; tg * (2<<16) + 0.5 | ||
117 : | edgomez | 1.2 | |
118 : | %if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 | ||
119 : | edgomez | 1.2.2.1 | rounder_0: |
120 : | dd 65536, 65536 | ||
121 : | rounder_4: | ||
122 : | dd 0, 0 | ||
123 : | rounder_1: | ||
124 : | dd 7195, 7195 | ||
125 : | rounder_7 | ||
126 : | dd 1024, 1024 | ||
127 : | rounder_2: | ||
128 : | dd 4520, 4520 | ||
129 : | rounder_6: | ||
130 : | dd 1024, 1024 | ||
131 : | rounder_3: | ||
132 : | dd 2407, 2407 | ||
133 : | rounder_5: | ||
134 : | dd 240, 240 | ||
135 : | edgomez | 1.2 | |
136 : | %elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 | ||
137 : | edgomez | 1.2.2.1 | rounder_0: |
138 : | dd 65536, 65536 | ||
139 : | rounder_4: | ||
140 : | dd 0, 0 | ||
141 : | rounder_1: | ||
142 : | dd 3597, 3597 | ||
143 : | rounder_7: | ||
144 : | dd 512, 512 | ||
145 : | rounder_2: | ||
146 : | dd 2260, 2260 | ||
147 : | rounder_6: | ||
148 : | dd 512, 512 | ||
149 : | rounder_3: | ||
150 : | dd 1203, 1203 | ||
151 : | rounder_5: | ||
152 : | dd 120, 120 | ||
153 : | edgomez | 1.2 | %else |
154 : | |||
155 : | edgomez | 1.2.2.1 | %error invalid SHIFT_INV_ROW |
156 : | edgomez | 1.2 | |
157 : | %endif | ||
158 : | |||
159 : | ;----------------------------------------------------------------------------- | ||
160 : | edgomez | 1.2.2.1 | ; Tables for xmm processors |
161 : | edgomez | 1.2 | ;----------------------------------------------------------------------------- |
162 : | |||
163 : | ; %3 for rows 0,4 - constants are multiplied by cos_4_16 | ||
164 : | edgomez | 1.2.2.1 | tab_i_04_xmm: |
165 : | dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 | ||
166 : | dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 | ||
167 : | dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 | ||
168 : | dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 | ||
169 : | dw 22725, 19266, 19266, -4520 ; w21 w20 w17 w16 | ||
170 : | dw 12873, 4520, -22725, -12873 ; w23 w22 w19 w18 | ||
171 : | dw 12873, -22725, 4520, -12873 ; w29 w28 w25 w24 | ||
172 : | dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 | ||
173 : | edgomez | 1.2 | |
174 : | ; %3 for rows 1,7 - constants are multiplied by cos_1_16 | ||
175 : | edgomez | 1.2.2.1 | tab_i_17_xmm: |
176 : | dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 | ||
177 : | dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 | ||
178 : | dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 | ||
179 : | dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 | ||
180 : | dw 31521, 26722, 26722, -6270 ; w21 w20 w17 w16 | ||
181 : | dw 17855, 6270, -31521, -17855 ; w23 w22 w19 w18 | ||
182 : | dw 17855, -31521, 6270, -17855 ; w29 w28 w25 w24 | ||
183 : | dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 | ||
184 : | edgomez | 1.2 | |
185 : | ; %3 for rows 2,6 - constants are multiplied by cos_2_16 | ||
186 : | edgomez | 1.2.2.1 | tab_i_26_xmm: |
187 : | dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 | ||
188 : | dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 | ||
189 : | dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 | ||
190 : | dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 | ||
191 : | dw 29692, 25172, 25172, -5906 ; w21 w20 w17 w16 | ||
192 : | dw 16819, 5906, -29692, -16819 ; w23 w22 w19 w18 | ||
193 : | dw 16819, -29692, 5906, -16819 ; w29 w28 w25 w24 | ||
194 : | dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 | ||
195 : | edgomez | 1.2 | |
196 : | ; %3 for rows 3,5 - constants are multiplied by cos_3_16 | ||
197 : | edgomez | 1.2.2.1 | tab_i_35_xmm: |
198 : | dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 | ||
199 : | dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 | ||
200 : | dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 | ||
201 : | dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 | ||
202 : | dw 26722, 22654, 22654, -5315 ; w21 w20 w17 w16 | ||
203 : | dw 15137, 5315, -26722, -15137 ; w23 w22 w19 w18 | ||
204 : | dw 15137, -26722, 5315, -15137 ; w29 w28 w25 w24 | ||
205 : | dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 | ||
206 : | edgomez | 1.2 | |
207 : | ;============================================================================= | ||
208 : | edgomez | 1.2.2.1 | ; Code |
209 : | edgomez | 1.2 | ;============================================================================= |
210 : | |||
211 : | edgomez | 1.2.2.1 | SECTION .text |
212 : | |||
213 : | edgomez | 1.2 | ;----------------------------------------------------------------------------- |
214 : | edgomez | 1.2.2.1 | ; void idct_3dne(uint16_t block[64]); |
215 : | edgomez | 1.2 | ;----------------------------------------------------------------------------- |
216 : | |||
217 : | edgomez | 1.2.2.1 | ALIGN 16 |
218 : | cglobal idct_3dne | ||
219 : | idct_3dne: | ||
220 : | mov eax, [esp+4] | ||
221 : | edgomez | 1.2 | |
222 : | edgomez | 1.2.2.1 | ; DCT_8_INV_ROW_1_s [eax+64], [eax+64], tab_i_04_sse, rounder_4 ;rounder_4=0 |
223 : | pshufw mm0, [eax+64],10001000b ; x2 x0 x2 x0 | ||
224 : | movq mm3, [tab_i_04_xmm] ; 3 ; w05 w04 w01 w00 | ||
225 : | pshufw mm1, [eax+64+8],10001000b ; x6 x4 x6 x4 | ||
226 : | movq mm4, [tab_i_04_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
227 : | pshufw mm2, [eax+64],11011101b ; x3 x1 x3 x1 | ||
228 : | pshufw mm5, [eax+64+8],11011101b ; x7 x5 x7 x5 | ||
229 : | movq mm6, [tab_i_04_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
230 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
231 : | movq mm7, [tab_i_04_xmm+40] ; 7 ; w23 w22 w19 w18 ; | ||
232 : | pmaddwd mm0, [tab_i_04_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
233 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
234 : | pmaddwd mm1, [tab_i_04_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
235 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
236 : | pmaddwd mm2, [tab_i_04_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
237 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
238 : | pmaddwd mm5, [tab_i_04_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
239 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
240 : | paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
241 : | pshufw mm1, [eax+80+8],10001000b ; x6 x4 x6 x4 | ||
242 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
243 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
244 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
245 : | pshufw mm5, [eax+80],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
246 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
247 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
248 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
249 : | movq mm3, [tab_i_35_xmm] ; 3 ; w05 w04 w01 w00 | ||
250 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
251 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
252 : | pshufw mm2, [eax+80],11011101b ; x3 x1 x3 x1 | ||
253 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
254 : | pmaddwd mm5, [tab_i_35_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
255 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
256 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
257 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
258 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
259 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
260 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
261 : | pshufw mm0, [eax+80+8],11011101b ; x7 x5 x7 x5 | ||
262 : | movq [eax+64], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
263 : | |||
264 : | ; DCT_8_INV_ROW_1_s [eax+80], [eax+80], tab_i_35_xmm, rounder_5 | ||
265 : | movq mm4, [tab_i_35_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
266 : | movq mm6, [tab_i_35_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
267 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
268 : | paddd mm3, [rounder_5] ; +rounder stall 6 | ||
269 : | paddd mm5, [rounder_5] ; +rounder | ||
270 : | movq [eax+64+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
271 : | movq mm7, [tab_i_35_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
272 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
273 : | pmaddwd mm1, [tab_i_35_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
274 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
275 : | pmaddwd mm2, [tab_i_35_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
276 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
277 : | pmaddwd mm0, [tab_i_35_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
278 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
279 : | paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
280 : | pshufw mm1, [eax+96+8],10001000b ; x6 x4 x6 x4 | ||
281 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
282 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
283 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
284 : | pshufw mm0, [eax+96],10001000b ; x2 x0 x2 x0 | ||
285 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
286 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 | ||
287 : | paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
288 : | movq mm3, [tab_i_26_xmm] ; 3 ; w05 w04 w01 w00 | ||
289 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
290 : | paddd mm5, mm2 ; 0 free a3+b3 a2+b2 | ||
291 : | pshufw mm2, [eax+96],11011101b ; x3 x1 x3 x1 | ||
292 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
293 : | pmaddwd mm0, [tab_i_26_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
294 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
295 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
296 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
297 : | psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
298 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
299 : | packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 | ||
300 : | pshufw mm5, [eax+96+8],11011101b ; x7 x5 x7 x5 | ||
301 : | movq [eax+80], mm6 ; 3 ; save y3 y2 y1 y0 | ||
302 : | |||
303 : | ; DCT_8_INV_ROW_1_s [eax+96], [eax+96], tab_i_26_xmm, rounder_6 | ||
304 : | movq mm4, [tab_i_26_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
305 : | movq mm6, [tab_i_26_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
306 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 | ||
307 : | paddd mm3, [rounder_6] ; +rounder | ||
308 : | paddd mm0, [rounder_6] ; +rounder | ||
309 : | movq [eax+80+8], mm7 ; 7 ; save y7 y6 | ||
310 : | movq mm7, [tab_i_26_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
311 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
312 : | pmaddwd mm1, [tab_i_26_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
313 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
314 : | pmaddwd mm2, [tab_i_26_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
315 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
316 : | pmaddwd mm5, [tab_i_26_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
317 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
318 : | paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
319 : | pshufw mm1, [eax+112+8],10001000b ; x6 x4 x6 x4 | ||
320 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
321 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
322 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
323 : | pshufw mm5, [eax+112],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
324 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
325 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
326 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
327 : | movq mm3, [tab_i_17_xmm] ; 3 ; w05 w04 w01 w00 | ||
328 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
329 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
330 : | pshufw mm2, [eax+112],11011101b; x3 x1 x3 x1 | ||
331 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
332 : | pmaddwd mm5, [tab_i_17_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
333 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
334 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
335 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
336 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
337 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
338 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
339 : | pshufw mm0, [eax+112+8],11011101b ; x7 x5 x7 x5 | ||
340 : | movq [eax+96], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
341 : | |||
342 : | ; DCT_8_INV_ROW_1_s [eax+112], [eax+112], tab_i_17_xmm, rounder_7 | ||
343 : | movq mm4, [tab_i_17_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
344 : | movq mm6, [tab_i_17_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
345 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
346 : | paddd mm3, [rounder_7] ; +rounder stall 6 | ||
347 : | paddd mm5, [rounder_7] ; +rounder | ||
348 : | movq [eax+96+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
349 : | movq mm7, [tab_i_17_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
350 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
351 : | pmaddwd mm1, [tab_i_17_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
352 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
353 : | pmaddwd mm2, [tab_i_17_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
354 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
355 : | pmaddwd mm0, [tab_i_17_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
356 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
357 : | paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
358 : | pshufw mm1, [eax+0+8],10001000b; x6 x4 x6 x4 | ||
359 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
360 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
361 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
362 : | pshufw mm0, [eax+0],10001000b ; x2 x0 x2 x0 | ||
363 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
364 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 | ||
365 : | paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
366 : | movq mm3, [tab_i_04_xmm] ; 3 ; w05 w04 w01 w00 | ||
367 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
368 : | paddd mm5, mm2 ; 0 free a3+b3 a2+b2 | ||
369 : | pshufw mm2, [eax+0],11011101b ; x3 x1 x3 x1 | ||
370 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
371 : | pmaddwd mm0, [tab_i_04_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
372 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
373 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
374 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
375 : | psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
376 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
377 : | packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 | ||
378 : | pshufw mm5, [eax+0+8],11011101b; x7 x5 x7 x5 | ||
379 : | movq [eax+112], mm6 ; 3 ; save y3 y2 y1 y0 | ||
380 : | |||
381 : | ; DCT_8_INV_ROW_1_s [eax+0], 0, tab_i_04_xmm, rounder_0 | ||
382 : | movq mm4, [tab_i_04_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
383 : | movq mm6, [tab_i_04_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
384 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 | ||
385 : | paddd mm3, [rounder_0] ; +rounder | ||
386 : | paddd mm0, [rounder_0] ; +rounder | ||
387 : | movq [eax+112+8], mm7 ; 7 ; save y7 y6 | ||
388 : | movq mm7, [tab_i_04_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
389 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
390 : | pmaddwd mm1, [tab_i_04_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
391 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
392 : | pmaddwd mm2, [tab_i_04_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
393 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
394 : | pmaddwd mm5, [tab_i_04_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
395 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
396 : | paddd mm0, mm1 ; 1 | ||
397 : | pshufw mm1, [eax+16+8],10001000b ; x6 x4 x6 x4 | ||
398 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
399 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
400 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
401 : | pshufw mm5, [eax+16],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
402 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
403 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
404 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
405 : | movq mm3, [tab_i_17_xmm] ; 3 ; w05 w04 w01 w00 | ||
406 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
407 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
408 : | pshufw mm2, [eax+16],11011101b ; x3 x1 x3 x1 | ||
409 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
410 : | pmaddwd mm5, [tab_i_17_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
411 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
412 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
413 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
414 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
415 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
416 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
417 : | pshufw mm0, [eax+16+8],11011101b ; x7 x5 x7 x5 | ||
418 : | movq [eax+0], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
419 : | |||
420 : | ; DCT_8_INV_ROW_1_s [eax+16], 16, tab_i_17_xmm, rounder_1 | ||
421 : | movq mm4, [tab_i_17_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
422 : | movq mm6, [tab_i_17_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
423 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
424 : | paddd mm3, [rounder_1] ; +rounder stall 6 | ||
425 : | paddd mm5, [rounder_1] ; +rounder | ||
426 : | movq [eax+0+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
427 : | movq mm7, [tab_i_17_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
428 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
429 : | pmaddwd mm1, [tab_i_17_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
430 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
431 : | pmaddwd mm2, [tab_i_17_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
432 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
433 : | pmaddwd mm0, [tab_i_17_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
434 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
435 : | paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
436 : | pshufw mm1, [eax+32+8],10001000b ; x6 x4 x6 x4 | ||
437 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
438 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
439 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
440 : | pshufw mm0, [eax+32],10001000b ; x2 x0 x2 x0 | ||
441 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
442 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 | ||
443 : | paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
444 : | movq mm3, [tab_i_26_xmm] ; 3 ; w05 w04 w01 w00 | ||
445 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
446 : | paddd mm5, mm2 ; 0 free a3+b3 a2+b2 | ||
447 : | pshufw mm2, [eax+32],11011101b ; x3 x1 x3 x1 | ||
448 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
449 : | pmaddwd mm0, [tab_i_26_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
450 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
451 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
452 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
453 : | psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
454 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
455 : | packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 | ||
456 : | pshufw mm5, [eax+32+8],11011101b ; x7 x5 x7 x5 | ||
457 : | movq [eax+16], mm6 ; 3 ; save y3 y2 y1 y0 | ||
458 : | |||
459 : | ; DCT_8_INV_ROW_1_s [eax+32], 32, tab_i_26_xmm, rounder_2 | ||
460 : | movq mm4, [tab_i_26_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
461 : | movq mm6, [tab_i_26_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
462 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 | ||
463 : | paddd mm3, [rounder_2] ; +rounder | ||
464 : | paddd mm0, [rounder_2] ; +rounder | ||
465 : | movq [eax+16+8], mm7 ; 7 ; save y7 y6 | ||
466 : | movq mm7, [tab_i_26_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
467 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
468 : | pmaddwd mm1, [tab_i_26_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
469 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
470 : | pmaddwd mm2, [tab_i_26_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
471 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
472 : | pmaddwd mm5, [tab_i_26_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
473 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
474 : | paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
475 : | pshufw mm1, [eax+48+8],10001000b ; x6 x4 x6 x4 | ||
476 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
477 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
478 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
479 : | pshufw mm5, [eax+48],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
480 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
481 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
482 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
483 : | movq mm3, [tab_i_35_xmm] ; 3 ; w05 w04 w01 w00 | ||
484 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
485 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
486 : | pshufw mm2, [eax+48],11011101b ; x3 x1 x3 x1 | ||
487 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
488 : | pmaddwd mm5, [tab_i_35_xmm+16] ; x2*w13+x0*w12 x2*w09+x0*w08 | ||
489 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
490 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
491 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
492 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
493 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
494 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
495 : | pshufw mm0, [eax+48+8],11011101b ; x7 x5 x7 x5 | ||
496 : | movq [eax+32], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
497 : | |||
498 : | ; DCT_8_INV_ROW_1_s [eax+48], [eax+48], tab_i_35_xmm, rounder_3 | ||
499 : | movq mm4, [tab_i_35_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
500 : | movq mm6, [tab_i_35_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
501 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
502 : | paddd mm3, [rounder_3] ; +rounder stall 6 | ||
503 : | paddd mm5, [rounder_3] ; +rounder | ||
504 : | movq [eax+32+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
505 : | movq mm7, [tab_i_35_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
506 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
507 : | pmaddwd mm1, [tab_i_35_xmm+24] ; x6*w15+x4*w14 x6*w11+x4*w10 | ||
508 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
509 : | pmaddwd mm2, [tab_i_35_xmm+48] ; x3*w29+x1*w28 x3*w25+x1*w24 | ||
510 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
511 : | pmaddwd mm0, [tab_i_35_xmm+56] ; x7*w31+x5*w30 x7*w27+x5*w26 | ||
512 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
513 : | paddd mm5, mm1 ; mm1 free ; a3=sum(even3) a2=sum(even2) | ||
514 : | movq mm1, [tg_3_16] | ||
515 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
516 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
517 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
518 : | movq mm0, [tg_3_16] | ||
519 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
520 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
521 : | paddd mm3, mm6 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
522 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
523 : | paddd mm2, mm5 ; 0 free a3+b3 a2+b2 | ||
524 : | movq mm5, [eax+16*5] | ||
525 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
526 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
527 : | psrad mm3, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
528 : | psrad mm2, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
529 : | movq mm6, [eax+16*1] | ||
530 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
531 : | movq mm4, [tg_1_16] | ||
532 : | packssdw mm3, mm2 ; 0 free ; y3 y2 y1 y0 | ||
533 : | pshufw mm2, mm7, 10110001b ; y7 y6 y5 y4 | ||
534 : | |||
535 : | ; DCT_8_INV_COL_4 [eax+0],[eax+0] | ||
536 : | ; movq mm3,mmword ptr [eax+16*3] | ||
537 : | movq mm7, [eax+16*7] | ||
538 : | edgomez | 1.2 | pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
539 : | pmulhw mm1, mm5 ; x5*(tg_3_16-1) | ||
540 : | edgomez | 1.2.2.1 | movq [eax+48+8], mm2 ; 7 ; save y7 y6 y5 y4 |
541 : | edgomez | 1.2 | movq mm2, mm4 ; tg_1_16 |
542 : | pmulhw mm4, mm7 ; x7*tg_1_16 | ||
543 : | paddsw mm0, mm3 ; x3*tg_3_16 | ||
544 : | pmulhw mm2, mm6 ; x1*tg_1_16 | ||
545 : | paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) | ||
546 : | psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 | ||
547 : | edgomez | 1.2.2.1 | movq [eax+48], mm3 ; 3 ; save y3 y2 y1 y0 |
548 : | movq mm3, [ocos_4_16] | ||
549 : | edgomez | 1.2 | paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
550 : | paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 | ||
551 : | psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 | ||
552 : | movq mm5, mm4 ; tp17 | ||
553 : | movq mm6, mm2 ; tm17 | ||
554 : | paddsw mm5, mm1 ; tp17+tp35 = b0 | ||
555 : | psubsw mm6, mm0 ; tm17-tm35 = b3 | ||
556 : | psubsw mm4, mm1 ; tp17-tp35 = t1 | ||
557 : | paddsw mm2, mm0 ; tm17+tm35 = t2 | ||
558 : | edgomez | 1.2.2.1 | movq mm7, [tg_2_16] |
559 : | edgomez | 1.2 | movq mm1, mm4 ; t1 |
560 : | edgomez | 1.2.2.1 | movq [eax+3*16], mm5 ; save b0 |
561 : | edgomez | 1.2 | paddsw mm1, mm2 ; t1+t2 |
562 : | edgomez | 1.2.2.1 | movq [eax+5*16], mm6 ; save b3 |
563 : | edgomez | 1.2 | psubsw mm4, mm2 ; t1-t2 |
564 : | edgomez | 1.2.2.1 | movq mm5, [eax+2*16] |
565 : | edgomez | 1.2 | movq mm0, mm7 ; tg_2_16 |
566 : | edgomez | 1.2.2.1 | movq mm6, [eax+6*16] |
567 : | edgomez | 1.2 | pmulhw mm0, mm5 ; x2*tg_2_16 |
568 : | pmulhw mm7, mm6 ; x6*tg_2_16 | ||
569 : | ; slot | ||
570 : | pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 | ||
571 : | ; slot | ||
572 : | edgomez | 1.2.2.1 | movq mm2, [eax+0*16] |
573 : | edgomez | 1.2 | pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
574 : | psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 | ||
575 : | edgomez | 1.2.2.1 | movq mm3, [eax+0*16] ; x0 |
576 : | movq mm6, [eax+4*16] | ||
577 : | edgomez | 1.2 | paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
578 : | paddsw mm2, mm6 ; x0+x4 = tp04 | ||
579 : | psubsw mm3, mm6 ; x0-x4 = tm04 | ||
580 : | movq mm5, mm2 ; tp04 | ||
581 : | movq mm6, mm3 ; tm04 | ||
582 : | psubsw mm2, mm7 ; tp04-tp26 = a3 | ||
583 : | paddsw mm3, mm0 ; tm04+tm26 = a1 | ||
584 : | edgomez | 1.2.2.1 | paddsw mm1, mm1 ; b1 |
585 : | paddsw mm4, mm4 ; b2 | ||
586 : | edgomez | 1.2 | paddsw mm5, mm7 ; tp04+tp26 = a0 |
587 : | psubsw mm6, mm0 ; tm04-tm26 = a2 | ||
588 : | movq mm7, mm3 ; a1 | ||
589 : | movq mm0, mm6 ; a2 | ||
590 : | paddsw mm3, mm1 ; a1+b1 | ||
591 : | paddsw mm6, mm4 ; a2+b2 | ||
592 : | edgomez | 1.2.2.1 | psraw mm3, SHIFT_INV_COL ; dst1 |
593 : | edgomez | 1.2 | psubsw mm7, mm1 ; a1-b1 |
594 : | edgomez | 1.2.2.1 | psraw mm6, SHIFT_INV_COL ; dst2 |
595 : | edgomez | 1.2 | psubsw mm0, mm4 ; a2-b2 |
596 : | edgomez | 1.2.2.1 | movq mm1, [eax+3*16] ; load b0 |
597 : | psraw mm7, SHIFT_INV_COL ; dst6 | ||
598 : | edgomez | 1.2 | movq mm4, mm5 ; a0 |
599 : | edgomez | 1.2.2.1 | psraw mm0, SHIFT_INV_COL ; dst5 |
600 : | movq [eax+1*16], mm3 | ||
601 : | edgomez | 1.2 | paddsw mm5, mm1 ; a0+b0 |
602 : | edgomez | 1.2.2.1 | movq [eax+2*16], mm6 |
603 : | edgomez | 1.2 | psubsw mm4, mm1 ; a0-b0 |
604 : | edgomez | 1.2.2.1 | movq mm3, [eax+5*16] ; load b3 |
605 : | psraw mm5, SHIFT_INV_COL ; dst0 | ||
606 : | edgomez | 1.2 | movq mm6, mm2 ; a3 |
607 : | edgomez | 1.2.2.1 | psraw mm4, SHIFT_INV_COL ; dst7 |
608 : | movq [eax+5*16], mm0 | ||
609 : | movq mm0, [tg_3_16] | ||
610 : | edgomez | 1.2 | paddsw mm2, mm3 ; a3+b3 |
611 : | edgomez | 1.2.2.1 | movq [eax+6*16], mm7 |
612 : | edgomez | 1.2 | psubsw mm6, mm3 ; a3-b3 |
613 : | edgomez | 1.2.2.1 | movq mm3, [eax+8+16*3] |
614 : | movq [eax+0*16], mm5 | ||
615 : | psraw mm2, SHIFT_INV_COL ; dst3 | ||
616 : | movq [eax+7*16], mm4 | ||
617 : | edgomez | 1.2 | |
618 : | ; DCT_8_INV_COL_4 [eax+8],[eax+8] | ||
619 : | edgomez | 1.2.2.1 | movq mm1, mm0 ; tg_3_16 |
620 : | movq mm5, [eax+8+16*5] | ||
621 : | psraw mm6, SHIFT_INV_COL ; dst4 | ||
622 : | pmulhw mm0, mm3 ; x3*(tg_3_16-1) | ||
623 : | movq mm4, [tg_1_16] | ||
624 : | pmulhw mm1, mm5 ; x5*(tg_3_16-1) | ||
625 : | movq mm7, [eax+8+16*7] | ||
626 : | movq [eax+3*16], mm2 | ||
627 : | movq mm2, mm4 ; tg_1_16 | ||
628 : | movq [eax+4*16], mm6 | ||
629 : | movq mm6, [eax+8+16*1] | ||
630 : | pmulhw mm4, mm7 ; x7*tg_1_16 | ||
631 : | paddsw mm0, mm3 ; x3*tg_3_16 | ||
632 : | pmulhw mm2, mm6 ; x1*tg_1_16 | ||
633 : | paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) | ||
634 : | psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 | ||
635 : | movq mm3, [ocos_4_16] | ||
636 : | paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 | ||
637 : | paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 | ||
638 : | psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 | ||
639 : | movq mm5, mm4 ; tp17 | ||
640 : | movq mm6, mm2 ; tm17 | ||
641 : | paddsw mm5, mm1 ; tp17+tp35 = b0 | ||
642 : | psubsw mm4, mm1 ; tp17-tp35 = t1 | ||
643 : | paddsw mm2, mm0 ; tm17+tm35 = t2 | ||
644 : | movq mm7, [tg_2_16] | ||
645 : | movq mm1, mm4 ; t1 | ||
646 : | psubsw mm6, mm0 ; tm17-tm35 = b3 | ||
647 : | movq [eax+8+3*16], mm5 ; save b0 | ||
648 : | movq [eax+8+5*16], mm6 ; save b3 | ||
649 : | psubsw mm4, mm2 ; t1-t2 | ||
650 : | movq mm5, [eax+8+2*16] | ||
651 : | movq mm0, mm7 ; tg_2_16 | ||
652 : | movq mm6, [eax+8+6*16] | ||
653 : | paddsw mm1, mm2 ; t1+t2 | ||
654 : | pmulhw mm0, mm5 ; x2*tg_2_16 | ||
655 : | pmulhw mm7, mm6 ; x6*tg_2_16 | ||
656 : | movq mm2, [eax+8+0*16] | ||
657 : | pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 | ||
658 : | psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 | ||
659 : | edgomez | 1.2 | ; slot |
660 : | edgomez | 1.2.2.1 | pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
661 : | edgomez | 1.2 | ; slot |
662 : | edgomez | 1.2.2.1 | movq mm3, [eax+8+0*16] ; x0 |
663 : | movq mm6, [eax+8+4*16] | ||
664 : | paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 | ||
665 : | paddsw mm2, mm6 ; x0+x4 = tp04 | ||
666 : | psubsw mm3, mm6 ; x0-x4 = tm04 | ||
667 : | movq mm5, mm2 ; tp04 | ||
668 : | movq mm6, mm3 ; tm04 | ||
669 : | psubsw mm2, mm7 ; tp04-tp26 = a3 | ||
670 : | paddsw mm3, mm0 ; tm04+tm26 = a1 | ||
671 : | paddsw mm1, mm1 ; b1 | ||
672 : | paddsw mm4, mm4 ; b2 | ||
673 : | paddsw mm5, mm7 ; tp04+tp26 = a0 | ||
674 : | psubsw mm6, mm0 ; tm04-tm26 = a2 | ||
675 : | movq mm7, mm3 ; a1 | ||
676 : | movq mm0, mm6 ; a2 | ||
677 : | paddsw mm3, mm1 ; a1+b1 | ||
678 : | paddsw mm6, mm4 ; a2+b2 | ||
679 : | psraw mm3, SHIFT_INV_COL ; dst1 | ||
680 : | psubsw mm7, mm1 ; a1-b1 | ||
681 : | psraw mm6, SHIFT_INV_COL ; dst2 | ||
682 : | psubsw mm0, mm4 ; a2-b2 | ||
683 : | movq mm1, [eax+8+3*16] ; load b0 | ||
684 : | psraw mm7, SHIFT_INV_COL ; dst6 | ||
685 : | movq mm4, mm5 ; a0 | ||
686 : | psraw mm0, SHIFT_INV_COL ; dst5 | ||
687 : | movq [eax+8+1*16], mm3 | ||
688 : | paddsw mm5, mm1 ; a0+b0 | ||
689 : | movq [eax+8+2*16], mm6 | ||
690 : | psubsw mm4, mm1 ; a0-b0 | ||
691 : | movq mm3, [eax+8+5*16] ; load b3 | ||
692 : | psraw mm5, SHIFT_INV_COL ; dst0 | ||
693 : | movq mm6, mm2 ; a3 | ||
694 : | psraw mm4, SHIFT_INV_COL ; dst7 | ||
695 : | movq [eax+8+5*16], mm0 | ||
696 : | paddsw mm2, mm3 ; a3+b3 | ||
697 : | movq [eax+8+6*16], mm7 | ||
698 : | psubsw mm6, mm3 ; a3-b3 | ||
699 : | movq [eax+8+0*16], mm5 | ||
700 : | psraw mm2, SHIFT_INV_COL ; dst3 | ||
701 : | movq [eax+8+7*16], mm4 | ||
702 : | psraw mm6, SHIFT_INV_COL ; dst4 | ||
703 : | movq [eax+8+3*16], mm2 | ||
704 : | movq [eax+8+4*16], mm6 | ||
705 : | edgomez | 1.2 | |
706 : | edgomez | 1.2.2.1 | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |