Mercurial > libavcodec.hg
annotate x86/idct_mmx_xvid.c @ 12492:58a960d6e34c libavcodec
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now
coded in asm instead of C, this is (depending on the function) up to 50%
faster for cases where gcc didn't do a great job at looping.
Since h264_idct_add8() is now faster than the manual loop setup in h264.c,
in-asm idct calling can now be enabled for chroma as well (see r16207). For
MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does
the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%.
author | rbultje |
---|---|
date | Tue, 14 Sep 2010 13:36:26 +0000 |
parents | b57409c0c286 |
children |
rev | line source |
---|---|
8430 | 1 /* |
2 * XVID MPEG-4 VIDEO CODEC | |
3 * - MMX and XMM forward discrete cosine transform - | |
4 * | |
5 * Copyright(C) 2001 Peter Ross <pross@xvid.org> | |
6 * | |
7 * Originally provided by Intel at AP-922 | |
8 * http://developer.intel.com/vtune/cbts/strmsimd/922down.htm | |
9 * (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) | |
10 * but in a limited edition. | |
11 * New macro implements a column part for precise iDCT | |
12 * The routine precision now satisfies IEEE standard 1180-1990. | |
13 * | |
14 * Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru> | |
15 * Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org> | |
16 * | |
17 * http://www.elecard.com/peter/idct.html | |
18 * http://www.linuxvideo.org/mpeg2dec/ | |
19 * | |
20 * These examples contain code fragments for first stage iDCT 8x8 | |
21 * (for rows) and first stage DCT 8x8 (for columns) | |
22 * | |
23 * conversion to gcc syntax by Michael Niedermayer | |
24 * | |
25 * This file is part of FFmpeg. | |
26 * | |
27 * FFmpeg is free software; you can redistribute it and/or | |
28 * modify it under the terms of the GNU Lesser General Public | |
29 * License as published by the Free Software Foundation; either | |
30 * version 2.1 of the License, or (at your option) any later version. | |
31 * | |
32 * FFmpeg is distributed in the hope that it will be useful, | |
33 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
34 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
35 * Lesser General Public License for more details. | |
36 * | |
37 * You should have received a copy of the GNU Lesser General Public License | |
38 * along with FFmpeg; if not, write to the Free Software Foundation, | |
39 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
40 */ | |
41 | |
42 #include <inttypes.h> | |
43 #include "libavcodec/avcodec.h" | |
11373 | 44 #include "idct_xvid.h" |
8430 | 45 |
46 //============================================================================= | |
47 // Macros and other preprocessor constants | |
48 //============================================================================= | |
49 | |
50 #define BITS_INV_ACC 5 // 4 or 5 for IEEE | |
51 #define SHIFT_INV_ROW (16 - BITS_INV_ACC) //11 | |
52 #define SHIFT_INV_COL (1 + BITS_INV_ACC) //6 | |
53 #define RND_INV_ROW (1024 * (6 - BITS_INV_ACC)) | |
54 #define RND_INV_COL (16 * (BITS_INV_ACC - 3)) | |
55 #define RND_INV_CORR (RND_INV_COL - 1) | |
56 | |
57 #define BITS_FRW_ACC 3 // 2 or 3 for accuracy | |
58 #define SHIFT_FRW_COL BITS_FRW_ACC | |
59 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17) | |
60 #define RND_FRW_ROW (262144*(BITS_FRW_ACC - 1)) | |
61 | |
62 | |
63 //----------------------------------------------------------------------------- | |
64 // Various memory constants (trigonometric values or rounding values) | |
65 //----------------------------------------------------------------------------- | |
66 | |
67 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
8430
diff
changeset
|
68 DECLARE_ALIGNED(8, static const int16_t, tg_1_16)[4*4] = { |
8430 | 69 13036,13036,13036,13036, // tg * (2<<16) + 0.5 |
70 27146,27146,27146,27146, // tg * (2<<16) + 0.5 | |
71 -21746,-21746,-21746,-21746, // tg * (2<<16) + 0.5 | |
72 23170,23170,23170,23170}; // cos * (2<<15) + 0.5 | |
73 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
8430
diff
changeset
|
74 DECLARE_ALIGNED(8, static const int32_t, rounder_0)[2*8] = { |
8430 | 75 65536,65536, |
76 3597,3597, | |
77 2260,2260, | |
78 1203,1203, | |
79 0,0, | |
80 120,120, | |
81 512,512, | |
82 512,512}; | |
83 | |
84 //----------------------------------------------------------------------------- | |
85 // | |
86 // The first stage iDCT 8x8 - inverse DCTs of rows | |
87 // | |
88 //----------------------------------------------------------------------------- | |
89 // The 8-point inverse DCT direct algorithm | |
90 //----------------------------------------------------------------------------- | |
91 // | |
92 // static const short w[32] = { | |
93 // FIX(cos_4_16), FIX(cos_2_16), FIX(cos_4_16), FIX(cos_6_16), | |
94 // FIX(cos_4_16), FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16), | |
95 // FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16), FIX(cos_2_16), | |
96 // FIX(cos_4_16), -FIX(cos_2_16), FIX(cos_4_16), -FIX(cos_6_16), | |
97 // FIX(cos_1_16), FIX(cos_3_16), FIX(cos_5_16), FIX(cos_7_16), | |
98 // FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16), | |
99 // FIX(cos_5_16), -FIX(cos_1_16), FIX(cos_7_16), FIX(cos_3_16), | |
100 // FIX(cos_7_16), -FIX(cos_5_16), FIX(cos_3_16), -FIX(cos_1_16) }; | |
101 // | |
102 // #define DCT_8_INV_ROW(x, y) | |
103 // { | |
104 // int a0, a1, a2, a3, b0, b1, b2, b3; | |
105 // | |
106 // a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3]; | |
107 // a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7]; | |
108 // a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11]; | |
109 // a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15]; | |
110 // b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19]; | |
111 // b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23]; | |
112 // b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27]; | |
113 // b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31]; | |
114 // | |
115 // y[0] = SHIFT_ROUND ( a0 + b0 ); | |
116 // y[1] = SHIFT_ROUND ( a1 + b1 ); | |
117 // y[2] = SHIFT_ROUND ( a2 + b2 ); | |
118 // y[3] = SHIFT_ROUND ( a3 + b3 ); | |
119 // y[4] = SHIFT_ROUND ( a3 - b3 ); | |
120 // y[5] = SHIFT_ROUND ( a2 - b2 ); | |
121 // y[6] = SHIFT_ROUND ( a1 - b1 ); | |
122 // y[7] = SHIFT_ROUND ( a0 - b0 ); | |
123 // } | |
124 // | |
125 //----------------------------------------------------------------------------- | |
126 // | |
127 // In this implementation the outputs of the iDCT-1D are multiplied | |
128 // for rows 0,4 - by cos_4_16, | |
129 // for rows 1,7 - by cos_1_16, | |
130 // for rows 2,6 - by cos_2_16, | |
131 // for rows 3,5 - by cos_3_16 | |
132 // and are shifted to the left for better accuracy | |
133 // | |
134 // For the constants used, | |
135 // FIX(float_const) = (short) (float_const * (1<<15) + 0.5) | |
136 // | |
137 //----------------------------------------------------------------------------- | |
138 | |
139 //----------------------------------------------------------------------------- | |
140 // Tables for mmx processors | |
141 //----------------------------------------------------------------------------- | |
142 | |
143 // Table for rows 0,4 - constants are multiplied by cos_4_16 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
8430
diff
changeset
|
144 DECLARE_ALIGNED(8, static const int16_t, tab_i_04_mmx)[32*4] = { |
8430 | 145 16384,16384,16384,-16384, // movq-> w06 w04 w02 w00 |
146 21407,8867,8867,-21407, // w07 w05 w03 w01 | |
147 16384,-16384,16384,16384, // w14 w12 w10 w08 | |
148 -8867,21407,-21407,-8867, // w15 w13 w11 w09 | |
149 22725,12873,19266,-22725, // w22 w20 w18 w16 | |
150 19266,4520,-4520,-12873, // w23 w21 w19 w17 | |
151 12873,4520,4520,19266, // w30 w28 w26 w24 | |
152 -22725,19266,-12873,-22725, // w31 w29 w27 w25 | |
153 // Table for rows 1,7 - constants are multiplied by cos_1_16 | |
154 22725,22725,22725,-22725, // movq-> w06 w04 w02 w00 | |
155 29692,12299,12299,-29692, // w07 w05 w03 w01 | |
156 22725,-22725,22725,22725, // w14 w12 w10 w08 | |
157 -12299,29692,-29692,-12299, // w15 w13 w11 w09 | |
158 31521,17855,26722,-31521, // w22 w20 w18 w16 | |
159 26722,6270,-6270,-17855, // w23 w21 w19 w17 | |
160 17855,6270,6270,26722, // w30 w28 w26 w24 | |
161 -31521,26722,-17855,-31521, // w31 w29 w27 w25 | |
162 // Table for rows 2,6 - constants are multiplied by cos_2_16 | |
163 21407,21407,21407,-21407, // movq-> w06 w04 w02 w00 | |
164 27969,11585,11585,-27969, // w07 w05 w03 w01 | |
165 21407,-21407,21407,21407, // w14 w12 w10 w08 | |
166 -11585,27969,-27969,-11585, // w15 w13 w11 w09 | |
167 29692,16819,25172,-29692, // w22 w20 w18 w16 | |
168 25172,5906,-5906,-16819, // w23 w21 w19 w17 | |
169 16819,5906,5906,25172, // w30 w28 w26 w24 | |
170 -29692,25172,-16819,-29692, // w31 w29 w27 w25 | |
171 // Table for rows 3,5 - constants are multiplied by cos_3_16 | |
172 19266,19266,19266,-19266, // movq-> w06 w04 w02 w00 | |
173 25172,10426,10426,-25172, // w07 w05 w03 w01 | |
174 19266,-19266,19266,19266, // w14 w12 w10 w08 | |
175 -10426,25172,-25172,-10426, // w15 w13 w11 w09 | |
176 26722,15137,22654,-26722, // w22 w20 w18 w16 | |
177 22654,5315,-5315,-15137, // w23 w21 w19 w17 | |
178 15137,5315,5315,22654, // w30 w28 w26 w24 | |
179 -26722,22654,-15137,-26722, // w31 w29 w27 w25 | |
180 }; | |
181 //----------------------------------------------------------------------------- | |
182 // Tables for xmm processors | |
183 //----------------------------------------------------------------------------- | |
184 | |
185 // %3 for rows 0,4 - constants are multiplied by cos_4_16 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
8430
diff
changeset
|
186 DECLARE_ALIGNED(8, static const int16_t, tab_i_04_xmm)[32*4] = { |
8430 | 187 16384,21407,16384,8867, // movq-> w05 w04 w01 w00 |
188 16384,8867,-16384,-21407, // w07 w06 w03 w02 | |
189 16384,-8867,16384,-21407, // w13 w12 w09 w08 | |
190 -16384,21407,16384,-8867, // w15 w14 w11 w10 | |
191 22725,19266,19266,-4520, // w21 w20 w17 w16 | |
192 12873,4520,-22725,-12873, // w23 w22 w19 w18 | |
193 12873,-22725,4520,-12873, // w29 w28 w25 w24 | |
194 4520,19266,19266,-22725, // w31 w30 w27 w26 | |
195 // %3 for rows 1,7 - constants are multiplied by cos_1_16 | |
196 22725,29692,22725,12299, // movq-> w05 w04 w01 w00 | |
197 22725,12299,-22725,-29692, // w07 w06 w03 w02 | |
198 22725,-12299,22725,-29692, // w13 w12 w09 w08 | |
199 -22725,29692,22725,-12299, // w15 w14 w11 w10 | |
200 31521,26722,26722,-6270, // w21 w20 w17 w16 | |
201 17855,6270,-31521,-17855, // w23 w22 w19 w18 | |
202 17855,-31521,6270,-17855, // w29 w28 w25 w24 | |
203 6270,26722,26722,-31521, // w31 w30 w27 w26 | |
204 // %3 for rows 2,6 - constants are multiplied by cos_2_16 | |
205 21407,27969,21407,11585, // movq-> w05 w04 w01 w00 | |
206 21407,11585,-21407,-27969, // w07 w06 w03 w02 | |
207 21407,-11585,21407,-27969, // w13 w12 w09 w08 | |
208 -21407,27969,21407,-11585, // w15 w14 w11 w10 | |
209 29692,25172,25172,-5906, // w21 w20 w17 w16 | |
210 16819,5906,-29692,-16819, // w23 w22 w19 w18 | |
211 16819,-29692,5906,-16819, // w29 w28 w25 w24 | |
212 5906,25172,25172,-29692, // w31 w30 w27 w26 | |
213 // %3 for rows 3,5 - constants are multiplied by cos_3_16 | |
214 19266,25172,19266,10426, // movq-> w05 w04 w01 w00 | |
215 19266,10426,-19266,-25172, // w07 w06 w03 w02 | |
216 19266,-10426,19266,-25172, // w13 w12 w09 w08 | |
217 -19266,25172,19266,-10426, // w15 w14 w11 w10 | |
218 26722,22654,22654,-5315, // w21 w20 w17 w16 | |
219 15137,5315,-26722,-15137, // w23 w22 w19 w18 | |
220 15137,-26722,5315,-15137, // w29 w28 w25 w24 | |
221 5315,22654,22654,-26722, // w31 w30 w27 w26 | |
222 }; | |
223 //============================================================================= | |
224 // Helper macros for the code | |
225 //============================================================================= | |
226 | |
227 //----------------------------------------------------------------------------- | |
228 // DCT_8_INV_ROW_MMX( INP, OUT, TABLE, ROUNDER | |
229 //----------------------------------------------------------------------------- | |
230 | |
231 #define DCT_8_INV_ROW_MMX(A1,A2,A3,A4)\ | |
232 "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\ | |
233 "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\ | |
234 "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\ | |
235 "movq " #A3 ",%%mm3 \n\t"/* 3 ; w06 w04 w02 w00*/\ | |
236 "punpcklwd %%mm1,%%mm0 \n\t"/* x5 x1 x4 x0*/\ | |
237 "movq %%mm0,%%mm5 \n\t"/* 5 ; x5 x1 x4 x0*/\ | |
238 "punpckldq %%mm0,%%mm0 \n\t"/* x4 x0 x4 x0*/\ | |
239 "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w05 w03 w01*/\ | |
240 "punpckhwd %%mm1,%%mm2 \n\t"/* 1 ; x7 x3 x6 x2*/\ | |
241 "pmaddwd %%mm0,%%mm3 \n\t"/* x4*w06+x0*w04 x4*w02+x0*w00*/\ | |
242 "movq %%mm2,%%mm6 \n\t"/* 6 ; x7 x3 x6 x2*/\ | |
243 "movq 32+" #A3 ",%%mm1 \n\t"/* 1 ; w22 w20 w18 w16*/\ | |
244 "punpckldq %%mm2,%%mm2 \n\t"/* x6 x2 x6 x2*/\ | |
245 "pmaddwd %%mm2,%%mm4 \n\t"/* x6*w07+x2*w05 x6*w03+x2*w01*/\ | |
246 "punpckhdq %%mm5,%%mm5 \n\t"/* x5 x1 x5 x1*/\ | |
247 "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x4*w14+x0*w12 x4*w10+x0*w08*/\ | |
248 "punpckhdq %%mm6,%%mm6 \n\t"/* x7 x3 x7 x3*/\ | |
249 "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w21 w19 w17*/\ | |
250 "pmaddwd %%mm5,%%mm1 \n\t"/* x5*w22+x1*w20 x5*w18+x1*w16*/\ | |
251 "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\ | |
252 "pmaddwd %%mm6,%%mm7 \n\t"/* x7*w23+x3*w21 x7*w19+x3*w17*/\ | |
253 "pmaddwd 24+" #A3 ",%%mm2 \n\t"/* x6*w15+x2*w13 x6*w11+x2*w09*/\ | |
254 "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\ | |
255 "pmaddwd 48+" #A3 ",%%mm5 \n\t"/* x5*w30+x1*w28 x5*w26+x1*w24*/\ | |
256 "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\ | |
257 "pmaddwd 56+" #A3 ",%%mm6 \n\t"/* x7*w31+x3*w29 x7*w27+x3*w25*/\ | |
258 "paddd %%mm7,%%mm1 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\ | |
259 "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\ | |
260 "psubd %%mm1,%%mm3 \n\t"/* a1-b1 a0-b0*/\ | |
261 "psrad $11,%%mm3 \n\t"/* y6=a1-b1 y7=a0-b0*/\ | |
262 "paddd %%mm4,%%mm1 \n\t"/* 4 ; a1+b1 a0+b0*/\ | |
263 "paddd %%mm2,%%mm0 \n\t"/* 2 ; a3=sum(even3) a2=sum(even2)*/\ | |
264 "psrad $11,%%mm1 \n\t"/* y1=a1+b1 y0=a0+b0*/\ | |
265 "paddd %%mm6,%%mm5 \n\t"/* 6 ; b3=sum(odd3) b2=sum(odd2)*/\ | |
266 "movq %%mm0,%%mm4 \n\t"/* 4 ; a3 a2*/\ | |
267 "paddd %%mm5,%%mm0 \n\t"/* a3+b3 a2+b2*/\ | |
268 "psubd %%mm5,%%mm4 \n\t"/* 5 ; a3-b3 a2-b2*/\ | |
269 "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\ | |
270 "psrad $11,%%mm4 \n\t"/* y4=a3-b3 y5=a2-b2*/\ | |
271 "packssdw %%mm0,%%mm1 \n\t"/* 0 ; y3 y2 y1 y0*/\ | |
272 "packssdw %%mm3,%%mm4 \n\t"/* 3 ; y6 y7 y4 y5*/\ | |
273 "movq %%mm4,%%mm7 \n\t"/* 7 ; y6 y7 y4 y5*/\ | |
274 "psrld $16,%%mm4 \n\t"/* 0 y6 0 y4*/\ | |
275 "pslld $16,%%mm7 \n\t"/* y7 0 y5 0*/\ | |
276 "movq %%mm1," #A2 " \n\t"/* 1 ; save y3 y2 y1 y0*/\ | |
277 "por %%mm4,%%mm7 \n\t"/* 4 ; y7 y6 y5 y4*/\ | |
278 "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\ | |
279 | |
280 | |
281 //----------------------------------------------------------------------------- | |
282 // DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER | |
283 //----------------------------------------------------------------------------- | |
284 | |
285 #define DCT_8_INV_ROW_XMM(A1,A2,A3,A4)\ | |
286 "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\ | |
287 "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\ | |
288 "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\ | |
289 "movq " #A3 ",%%mm3 \n\t"/* 3 ; w05 w04 w01 w00*/\ | |
290 "pshufw $0x88,%%mm0,%%mm0 \n\t"/* x2 x0 x2 x0*/\ | |
291 "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w06 w03 w02*/\ | |
292 "movq %%mm1,%%mm5 \n\t"/* 5 ; x7 x6 x5 x4*/\ | |
293 "pmaddwd %%mm0,%%mm3 \n\t"/* x2*w05+x0*w04 x2*w01+x0*w00*/\ | |
294 "movq 32+" #A3 ",%%mm6 \n\t"/* 6 ; w21 w20 w17 w16*/\ | |
295 "pshufw $0x88,%%mm1,%%mm1 \n\t"/* x6 x4 x6 x4*/\ | |
296 "pmaddwd %%mm1,%%mm4 \n\t"/* x6*w07+x4*w06 x6*w03+x4*w02*/\ | |
297 "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w22 w19 w18*/\ | |
298 "pshufw $0xdd,%%mm2,%%mm2 \n\t"/* x3 x1 x3 x1*/\ | |
299 "pmaddwd %%mm2,%%mm6 \n\t"/* x3*w21+x1*w20 x3*w17+x1*w16*/\ | |
300 "pshufw $0xdd,%%mm5,%%mm5 \n\t"/* x7 x5 x7 x5*/\ | |
301 "pmaddwd %%mm5,%%mm7 \n\t"/* x7*w23+x5*w22 x7*w19+x5*w18*/\ | |
302 "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\ | |
303 "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x2*w13+x0*w12 x2*w09+x0*w08*/\ | |
304 "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\ | |
305 "pmaddwd 24+" #A3 ",%%mm1 \n\t"/* x6*w15+x4*w14 x6*w11+x4*w10*/\ | |
306 "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\ | |
307 "pmaddwd 48+" #A3 ",%%mm2 \n\t"/* x3*w29+x1*w28 x3*w25+x1*w24*/\ | |
308 "paddd %%mm7,%%mm6 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\ | |
309 "pmaddwd 56+" #A3 ",%%mm5 \n\t"/* x7*w31+x5*w30 x7*w27+x5*w26*/\ | |
310 "paddd %%mm6,%%mm3 \n\t"/* a1+b1 a0+b0*/\ | |
311 "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\ | |
312 "psrad $11,%%mm3 \n\t"/* y1=a1+b1 y0=a0+b0*/\ | |
313 "paddd %%mm1,%%mm0 \n\t"/* 1 ; a3=sum(even3) a2=sum(even2)*/\ | |
314 "psubd %%mm6,%%mm4 \n\t"/* 6 ; a1-b1 a0-b0*/\ | |
315 "movq %%mm0,%%mm7 \n\t"/* 7 ; a3 a2*/\ | |
316 "paddd %%mm5,%%mm2 \n\t"/* 5 ; b3=sum(odd3) b2=sum(odd2)*/\ | |
317 "paddd %%mm2,%%mm0 \n\t"/* a3+b3 a2+b2*/\ | |
318 "psrad $11,%%mm4 \n\t"/* y6=a1-b1 y7=a0-b0*/\ | |
319 "psubd %%mm2,%%mm7 \n\t"/* 2 ; a3-b3 a2-b2*/\ | |
320 "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\ | |
321 "psrad $11,%%mm7 \n\t"/* y4=a3-b3 y5=a2-b2*/\ | |
322 "packssdw %%mm0,%%mm3 \n\t"/* 0 ; y3 y2 y1 y0*/\ | |
323 "packssdw %%mm4,%%mm7 \n\t"/* 4 ; y6 y7 y4 y5*/\ | |
324 "movq %%mm3, " #A2 " \n\t"/* 3 ; save y3 y2 y1 y0*/\ | |
325 "pshufw $0xb1,%%mm7,%%mm7 \n\t"/* y7 y6 y5 y4*/\ | |
326 "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\ | |
327 | |
328 | |
329 //----------------------------------------------------------------------------- | |
330 // | |
331 // The first stage DCT 8x8 - forward DCTs of columns | |
332 // | |
333 // The %2puts are multiplied | |
334 // for rows 0,4 - on cos_4_16, | |
335 // for rows 1,7 - on cos_1_16, | |
336 // for rows 2,6 - on cos_2_16, | |
337 // for rows 3,5 - on cos_3_16 | |
338 // and are shifted to the left for rise of accuracy | |
339 // | |
340 //----------------------------------------------------------------------------- | |
341 // | |
342 // The 8-point scaled forward DCT algorithm (26a8m) | |
343 // | |
344 //----------------------------------------------------------------------------- | |
345 // | |
346 // #define DCT_8_FRW_COL(x, y) | |
347 //{ | |
348 // short t0, t1, t2, t3, t4, t5, t6, t7; | |
349 // short tp03, tm03, tp12, tm12, tp65, tm65; | |
350 // short tp465, tm465, tp765, tm765; | |
351 // | |
352 // t0 = LEFT_SHIFT ( x[0] + x[7] ); | |
353 // t1 = LEFT_SHIFT ( x[1] + x[6] ); | |
354 // t2 = LEFT_SHIFT ( x[2] + x[5] ); | |
355 // t3 = LEFT_SHIFT ( x[3] + x[4] ); | |
356 // t4 = LEFT_SHIFT ( x[3] - x[4] ); | |
357 // t5 = LEFT_SHIFT ( x[2] - x[5] ); | |
358 // t6 = LEFT_SHIFT ( x[1] - x[6] ); | |
359 // t7 = LEFT_SHIFT ( x[0] - x[7] ); | |
360 // | |
361 // tp03 = t0 + t3; | |
362 // tm03 = t0 - t3; | |
363 // tp12 = t1 + t2; | |
364 // tm12 = t1 - t2; | |
365 // | |
366 // y[0] = tp03 + tp12; | |
367 // y[4] = tp03 - tp12; | |
368 // | |
369 // y[2] = tm03 + tm12 * tg_2_16; | |
370 // y[6] = tm03 * tg_2_16 - tm12; | |
371 // | |
372 // tp65 =(t6 +t5 )*cos_4_16; | |
373 // tm65 =(t6 -t5 )*cos_4_16; | |
374 // | |
375 // tp765 = t7 + tp65; | |
376 // tm765 = t7 - tp65; | |
377 // tp465 = t4 + tm65; | |
378 // tm465 = t4 - tm65; | |
379 // | |
380 // y[1] = tp765 + tp465 * tg_1_16; | |
381 // y[7] = tp765 * tg_1_16 - tp465; | |
382 // y[5] = tm765 * tg_3_16 + tm465; | |
383 // y[3] = tm765 - tm465 * tg_3_16; | |
384 //} | |
385 // | |
386 //----------------------------------------------------------------------------- | |
387 | |
388 //----------------------------------------------------------------------------- | |
389 // DCT_8_INV_COL_4 INP,OUT | |
390 //----------------------------------------------------------------------------- | |
391 | |
392 #define DCT_8_INV_COL(A1,A2)\ | |
393 "movq 2*8(%3),%%mm0\n\t"\ | |
394 "movq 16*3+" #A1 ",%%mm3\n\t"\ | |
395 "movq %%mm0,%%mm1 \n\t"/* tg_3_16*/\ | |
396 "movq 16*5+" #A1 ",%%mm5\n\t"\ | |
397 "pmulhw %%mm3,%%mm0 \n\t"/* x3*(tg_3_16-1)*/\ | |
398 "movq (%3),%%mm4\n\t"\ | |
399 "pmulhw %%mm5,%%mm1 \n\t"/* x5*(tg_3_16-1)*/\ | |
400 "movq 16*7+" #A1 ",%%mm7\n\t"\ | |
401 "movq %%mm4,%%mm2 \n\t"/* tg_1_16*/\ | |
402 "movq 16*1+" #A1 ",%%mm6\n\t"\ | |
403 "pmulhw %%mm7,%%mm4 \n\t"/* x7*tg_1_16*/\ | |
404 "paddsw %%mm3,%%mm0 \n\t"/* x3*tg_3_16*/\ | |
405 "pmulhw %%mm6,%%mm2 \n\t"/* x1*tg_1_16*/\ | |
406 "paddsw %%mm3,%%mm1 \n\t"/* x3+x5*(tg_3_16-1)*/\ | |
407 "psubsw %%mm5,%%mm0 \n\t"/* x3*tg_3_16-x5 = tm35*/\ | |
408 "movq 3*8(%3),%%mm3\n\t"\ | |
409 "paddsw %%mm5,%%mm1 \n\t"/* x3+x5*tg_3_16 = tp35*/\ | |
410 "paddsw %%mm6,%%mm4 \n\t"/* x1+tg_1_16*x7 = tp17*/\ | |
411 "psubsw %%mm7,%%mm2 \n\t"/* x1*tg_1_16-x7 = tm17*/\ | |
412 "movq %%mm4,%%mm5 \n\t"/* tp17*/\ | |
413 "movq %%mm2,%%mm6 \n\t"/* tm17*/\ | |
414 "paddsw %%mm1,%%mm5 \n\t"/* tp17+tp35 = b0*/\ | |
415 "psubsw %%mm0,%%mm6 \n\t"/* tm17-tm35 = b3*/\ | |
416 "psubsw %%mm1,%%mm4 \n\t"/* tp17-tp35 = t1*/\ | |
417 "paddsw %%mm0,%%mm2 \n\t"/* tm17+tm35 = t2*/\ | |
418 "movq 1*8(%3),%%mm7\n\t"\ | |
419 "movq %%mm4,%%mm1 \n\t"/* t1*/\ | |
420 "movq %%mm5,3*16 +" #A2 "\n\t"/* save b0*/\ | |
421 "paddsw %%mm2,%%mm1 \n\t"/* t1+t2*/\ | |
422 "movq %%mm6,5*16 +" #A2 "\n\t"/* save b3*/\ | |
423 "psubsw %%mm2,%%mm4 \n\t"/* t1-t2*/\ | |
424 "movq 2*16+" #A1 ",%%mm5\n\t"\ | |
425 "movq %%mm7,%%mm0 \n\t"/* tg_2_16*/\ | |
426 "movq 6*16+" #A1 ",%%mm6\n\t"\ | |
427 "pmulhw %%mm5,%%mm0 \n\t"/* x2*tg_2_16*/\ | |
428 "pmulhw %%mm6,%%mm7 \n\t"/* x6*tg_2_16*/\ | |
429 "pmulhw %%mm3,%%mm1 \n\t"/* ocos_4_16*(t1+t2) = b1/2*/\ | |
430 "movq 0*16+" #A1 ",%%mm2\n\t"\ | |
431 "pmulhw %%mm3,%%mm4 \n\t"/* ocos_4_16*(t1-t2) = b2/2*/\ | |
432 "psubsw %%mm6,%%mm0 \n\t"/* t2*tg_2_16-x6 = tm26*/\ | |
433 "movq %%mm2,%%mm3 \n\t"/* x0*/\ | |
434 "movq 4*16+" #A1 ",%%mm6\n\t"\ | |
435 "paddsw %%mm5,%%mm7 \n\t"/* x2+x6*tg_2_16 = tp26*/\ | |
436 "paddsw %%mm6,%%mm2 \n\t"/* x0+x4 = tp04*/\ | |
437 "psubsw %%mm6,%%mm3 \n\t"/* x0-x4 = tm04*/\ | |
438 "movq %%mm2,%%mm5 \n\t"/* tp04*/\ | |
439 "movq %%mm3,%%mm6 \n\t"/* tm04*/\ | |
440 "psubsw %%mm7,%%mm2 \n\t"/* tp04-tp26 = a3*/\ | |
441 "paddsw %%mm0,%%mm3 \n\t"/* tm04+tm26 = a1*/\ | |
442 "paddsw %%mm1,%%mm1 \n\t"/* b1*/\ | |
443 "paddsw %%mm4,%%mm4 \n\t"/* b2*/\ | |
444 "paddsw %%mm7,%%mm5 \n\t"/* tp04+tp26 = a0*/\ | |
445 "psubsw %%mm0,%%mm6 \n\t"/* tm04-tm26 = a2*/\ | |
446 "movq %%mm3,%%mm7 \n\t"/* a1*/\ | |
447 "movq %%mm6,%%mm0 \n\t"/* a2*/\ | |
448 "paddsw %%mm1,%%mm3 \n\t"/* a1+b1*/\ | |
449 "paddsw %%mm4,%%mm6 \n\t"/* a2+b2*/\ | |
450 "psraw $6,%%mm3 \n\t"/* dst1*/\ | |
451 "psubsw %%mm1,%%mm7 \n\t"/* a1-b1*/\ | |
452 "psraw $6,%%mm6 \n\t"/* dst2*/\ | |
453 "psubsw %%mm4,%%mm0 \n\t"/* a2-b2*/\ | |
454 "movq 3*16+" #A2 ",%%mm1 \n\t"/* load b0*/\ | |
455 "psraw $6,%%mm7 \n\t"/* dst6*/\ | |
456 "movq %%mm5,%%mm4 \n\t"/* a0*/\ | |
457 "psraw $6,%%mm0 \n\t"/* dst5*/\ | |
458 "movq %%mm3,1*16+" #A2 "\n\t"\ | |
459 "paddsw %%mm1,%%mm5 \n\t"/* a0+b0*/\ | |
460 "movq %%mm6,2*16+" #A2 "\n\t"\ | |
461 "psubsw %%mm1,%%mm4 \n\t"/* a0-b0*/\ | |
462 "movq 5*16+" #A2 ",%%mm3 \n\t"/* load b3*/\ | |
463 "psraw $6,%%mm5 \n\t"/* dst0*/\ | |
464 "movq %%mm2,%%mm6 \n\t"/* a3*/\ | |
465 "psraw $6,%%mm4 \n\t"/* dst7*/\ | |
466 "movq %%mm0,5*16+" #A2 "\n\t"\ | |
467 "paddsw %%mm3,%%mm2 \n\t"/* a3+b3*/\ | |
468 "movq %%mm7,6*16+" #A2 "\n\t"\ | |
469 "psubsw %%mm3,%%mm6 \n\t"/* a3-b3*/\ | |
470 "movq %%mm5,0*16+" #A2 "\n\t"\ | |
471 "psraw $6,%%mm2 \n\t"/* dst3*/\ | |
472 "movq %%mm4,7*16+" #A2 "\n\t"\ | |
473 "psraw $6,%%mm6 \n\t"/* dst4*/\ | |
474 "movq %%mm2,3*16+" #A2 "\n\t"\ | |
475 "movq %%mm6,4*16+" #A2 "\n\t" | |
476 | |
477 //============================================================================= | |
478 // Code | |
479 //============================================================================= | |
480 | |
481 //----------------------------------------------------------------------------- | |
482 // void idct_mmx(uint16_t block[64]); | |
483 //----------------------------------------------------------------------------- | |
484 | |
485 | |
486 void ff_idct_xvid_mmx(short *block){ | |
487 __asm__ volatile( | |
488 //# Process each row | |
489 DCT_8_INV_ROW_MMX(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1)) | |
490 DCT_8_INV_ROW_MMX(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1)) | |
491 DCT_8_INV_ROW_MMX(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1)) | |
492 DCT_8_INV_ROW_MMX(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1)) | |
493 DCT_8_INV_ROW_MMX(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1)) | |
494 DCT_8_INV_ROW_MMX(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1)) | |
495 DCT_8_INV_ROW_MMX(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1)) | |
496 DCT_8_INV_ROW_MMX(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1)) | |
497 | |
498 //# Process the columns (4 at a time) | |
499 DCT_8_INV_COL(0(%0), 0(%0)) | |
500 DCT_8_INV_COL(8(%0), 8(%0)) | |
501 :: "r"(block), "r"(rounder_0), "r"(tab_i_04_mmx), "r"(tg_1_16)); | |
502 } | |
503 | |
504 //----------------------------------------------------------------------------- | |
505 // void idct_xmm(uint16_t block[64]); | |
506 //----------------------------------------------------------------------------- | |
507 | |
508 | |
509 void ff_idct_xvid_mmx2(short *block){ | |
510 __asm__ volatile( | |
511 //# Process each row | |
512 DCT_8_INV_ROW_XMM(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1)) | |
513 DCT_8_INV_ROW_XMM(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1)) | |
514 DCT_8_INV_ROW_XMM(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1)) | |
515 DCT_8_INV_ROW_XMM(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1)) | |
516 DCT_8_INV_ROW_XMM(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1)) | |
517 DCT_8_INV_ROW_XMM(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1)) | |
518 DCT_8_INV_ROW_XMM(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1)) | |
519 DCT_8_INV_ROW_XMM(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1)) | |
520 | |
521 //# Process the columns (4 at a time) | |
522 DCT_8_INV_COL(0(%0), 0(%0)) | |
523 DCT_8_INV_COL(8(%0), 8(%0)) | |
524 :: "r"(block), "r"(rounder_0), "r"(tab_i_04_xmm), "r"(tg_1_16)); | |
525 } | |
526 |