2 * Copyright (C) 2004 the ffmpeg project
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * SSE2-optimized functions cribbed from the original VP3 source code.
26 #include "libavcodec/dsputil.h"
27 #include "dsputil_mmx.h"
30 static DECLARE_ALIGNED_16(const unsigned short, SSE2_idct_data[7 * 8]) =
32 64277,64277,64277,64277,64277,64277,64277,64277,
33 60547,60547,60547,60547,60547,60547,60547,60547,
34 54491,54491,54491,54491,54491,54491,54491,54491,
35 46341,46341,46341,46341,46341,46341,46341,46341,
36 36410,36410,36410,36410,36410,36410,36410,36410,
37 25080,25080,25080,25080,25080,25080,25080,25080,
38 12785,12785,12785,12785,12785,12785,12785,12785
42 #define SSE2_Column_IDCT() { \
44 movdqa_m2r(*I(3), xmm2); /* xmm2 = i3 */ \
45 movdqa_m2r(*C(3), xmm6); /* xmm6 = c3 */ \
47 movdqa_r2r(xmm2, xmm4); /* xmm4 = i3 */ \
48 movdqa_m2r(*I(5), xmm7); /* xmm7 = i5 */ \
50 pmulhw_r2r(xmm6, xmm4); /* xmm4 = c3 * i3 - i3 */ \
51 movdqa_m2r(*C(5), xmm1); /* xmm1 = c5 */ \
53 pmulhw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 - i5 */ \
54 movdqa_r2r(xmm1, xmm5); /* xmm5 = c5 */ \
56 pmulhw_r2r(xmm2, xmm1); /* xmm1 = c5 * i3 - i3 */ \
57 movdqa_m2r(*I(1), xmm3); /* xmm3 = i1 */ \
59 pmulhw_r2r(xmm7, xmm5); /* xmm5 = c5 * i5 - i5 */ \
60 movdqa_m2r(*C(1), xmm0); /* xmm0 = c1 */ \
62 /* all registers are in use */ \
64 paddw_r2r(xmm2, xmm4); /* xmm4 = c3 * i3 */ \
65 paddw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 */ \
67 paddw_r2r(xmm1, xmm2); /* xmm2 = c5 * i3 */ \
68 movdqa_m2r(*I(7), xmm1); /* xmm1 = i7 */ \
70 paddw_r2r(xmm5, xmm7); /* xmm7 = c5 * i5 */ \
71 movdqa_r2r(xmm0, xmm5); /* xmm5 = c1 */ \
73 pmulhw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 - i1 */ \
74 paddsw_r2r(xmm7, xmm4); /* xmm4 = c3 * i3 + c5 * i5 = C */ \
76 pmulhw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 - i7 */ \
77 movdqa_m2r(*C(7), xmm7); /* xmm7 = c7 */ \
79 psubsw_r2r(xmm2, xmm6); /* xmm6 = c3 * i5 - c5 * i3 = D */ \
80 paddw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 */ \
82 pmulhw_r2r(xmm7, xmm3); /* xmm3 = c7 * i1 */ \
83 movdqa_m2r(*I(2), xmm2); /* xmm2 = i2 */ \
85 pmulhw_r2r(xmm1, xmm7); /* xmm7 = c7 * i7 */ \
86 paddw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 */ \
88 movdqa_r2r(xmm2, xmm1); /* xmm1 = i2 */ \
89 pmulhw_m2r(*C(2), xmm2); /* xmm2 = i2 * c2 -i2 */ \
91 psubsw_r2r(xmm5, xmm3); /* xmm3 = c7 * i1 - c1 * i7 = B */ \
92 movdqa_m2r(*I(6), xmm5); /* xmm5 = i6 */ \
94 paddsw_r2r(xmm7, xmm0); /* xmm0 = c1 * i1 + c7 * i7 = A */ \
95 movdqa_r2r(xmm5, xmm7); /* xmm7 = i6 */ \
97 psubsw_r2r(xmm4, xmm0); /* xmm0 = A - C */ \
98 pmulhw_m2r(*C(2), xmm5); /* xmm5 = c2 * i6 - i6 */ \
100 paddw_r2r(xmm1, xmm2); /* xmm2 = i2 * c2 */ \
101 pmulhw_m2r(*C(6), xmm1); /* xmm1 = c6 * i2 */ \
103 paddsw_r2r(xmm4, xmm4); /* xmm4 = C + C */ \
104 paddsw_r2r(xmm0, xmm4); /* xmm4 = A + C = C. */ \
106 psubsw_r2r(xmm6, xmm3); /* xmm3 = B - D */ \
107 paddw_r2r(xmm7, xmm5); /* xmm5 = c2 * i6 */ \
109 paddsw_r2r(xmm6, xmm6); /* xmm6 = D + D */ \
110 pmulhw_m2r(*C(6), xmm7); /* xmm7 = c6 * i6 */ \
112 paddsw_r2r(xmm3, xmm6); /* xmm6 = B + D = D. */ \
113 movdqa_r2m(xmm4, *I(1)); /* Save C. at I(1) */ \
115 psubsw_r2r(xmm5, xmm1); /* xmm1 = c6 * i2 - c2 * i6 = H */ \
116 movdqa_m2r(*C(4), xmm4); /* xmm4 = c4 */ \
118 movdqa_r2r(xmm3, xmm5); /* xmm5 = B - D */ \
119 pmulhw_r2r(xmm4, xmm3); /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
121 paddsw_r2r(xmm2, xmm7); /* xmm7 = c2 * i2 + c6 * i6 = G */ \
122 movdqa_r2m(xmm6, *I(2)); /* Save D. at I(2) */ \
124 movdqa_r2r(xmm0, xmm2); /* xmm2 = A - C */ \
125 movdqa_m2r(*I(0), xmm6); /* xmm6 = i0 */ \
127 pmulhw_r2r(xmm4, xmm0); /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
128 paddw_r2r(xmm3, xmm5); /* xmm5 = c4 * ( B - D ) = B. */ \
130 movdqa_m2r(*I(4), xmm3); /* xmm3 = i4 */ \
131 psubsw_r2r(xmm1, xmm5); /* xmm5 = B. - H = B.. */ \
133 paddw_r2r(xmm0, xmm2); /* xmm2 = c4 * ( A - C) = A. */ \
134 psubsw_r2r(xmm3, xmm6); /* xmm6 = i0 - i4 */ \
136 movdqa_r2r(xmm6, xmm0); /* xmm0 = i0 - i4 */ \
137 pmulhw_r2r(xmm4, xmm6); /* xmm6 = (c4 - 1) * (i0 - i4) = F */ \
139 paddsw_r2r(xmm3, xmm3); /* xmm3 = i4 + i4 */ \
140 paddsw_r2r(xmm1, xmm1); /* xmm1 = H + H */ \
142 paddsw_r2r(xmm0, xmm3); /* xmm3 = i0 + i4 */ \
143 paddsw_r2r(xmm5, xmm1); /* xmm1 = B. + H = H. */ \
145 pmulhw_r2r(xmm3, xmm4); /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
146 paddw_r2r(xmm0, xmm6); /* xmm6 = c4 * ( i0 - i4 ) */ \
148 psubsw_r2r(xmm2, xmm6); /* xmm6 = F - A. = F. */ \
149 paddsw_r2r(xmm2, xmm2); /* xmm2 = A. + A. */ \
151 movdqa_m2r(*I(1), xmm0); /* Load C. from I(1) */ \
152 paddsw_r2r(xmm6, xmm2); /* xmm2 = F + A. = A.. */ \
154 paddw_r2r(xmm3, xmm4); /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
155 psubsw_r2r(xmm1, xmm2); /* xmm2 = A.. - H. = R2 */ \
157 paddsw_m2r(*Eight, xmm2); /* Adjust R2 and R1 before shifting */ \
158 paddsw_r2r(xmm1, xmm1); /* xmm1 = H. + H. */ \
160 paddsw_r2r(xmm2, xmm1); /* xmm1 = A.. + H. = R1 */ \
161 psraw_i2r(4, xmm2); /* xmm2 = op2 */ \
163 psubsw_r2r(xmm7, xmm4); /* xmm4 = E - G = E. */ \
164 psraw_i2r(4, xmm1); /* xmm1 = op1 */ \
166 movdqa_m2r(*I(2), xmm3); /* Load D. from I(2) */ \
167 paddsw_r2r(xmm7, xmm7); /* xmm7 = G + G */ \
169 movdqa_r2m(xmm2, *O(2)); /* Write out op2 */ \
170 paddsw_r2r(xmm4, xmm7); /* xmm7 = E + G = G. */ \
172 movdqa_r2m(xmm1, *O(1)); /* Write out op1 */ \
173 psubsw_r2r(xmm3, xmm4); /* xmm4 = E. - D. = R4 */ \
175 paddsw_m2r(*Eight, xmm4); /* Adjust R4 and R3 before shifting */ \
176 paddsw_r2r(xmm3, xmm3); /* xmm3 = D. + D. */ \
178 paddsw_r2r(xmm4, xmm3); /* xmm3 = E. + D. = R3 */ \
179 psraw_i2r(4, xmm4); /* xmm4 = op4 */ \
181 psubsw_r2r(xmm5, xmm6); /* xmm6 = F. - B..= R6 */ \
182 psraw_i2r(4, xmm3); /* xmm3 = op3 */ \
184 paddsw_m2r(*Eight, xmm6); /* Adjust R6 and R5 before shifting */ \
185 paddsw_r2r(xmm5, xmm5); /* xmm5 = B.. + B.. */ \
187 paddsw_r2r(xmm6, xmm5); /* xmm5 = F. + B.. = R5 */ \
188 psraw_i2r(4, xmm6); /* xmm6 = op6 */ \
190 movdqa_r2m(xmm4, *O(4)); /* Write out op4 */ \
191 psraw_i2r(4, xmm5); /* xmm5 = op5 */ \
193 movdqa_r2m(xmm3, *O(3)); /* Write out op3 */ \
194 psubsw_r2r(xmm0, xmm7); /* xmm7 = G. - C. = R7 */ \
196 paddsw_m2r(*Eight, xmm7); /* Adjust R7 and R0 before shifting */ \
197 paddsw_r2r(xmm0, xmm0); /* xmm0 = C. + C. */ \
199 paddsw_r2r(xmm7, xmm0); /* xmm0 = G. + C. */ \
200 psraw_i2r(4, xmm7); /* xmm7 = op7 */ \
202 movdqa_r2m(xmm6, *O(6)); /* Write out op6 */ \
203 psraw_i2r(4, xmm0); /* xmm0 = op0 */ \
205 movdqa_r2m(xmm5, *O(5)); /* Write out op5 */ \
206 movdqa_r2m(xmm7, *O(7)); /* Write out op7 */ \
208 movdqa_r2m(xmm0, *O(0)); /* Write out op0 */ \
210 } /* End of SSE2_Column_IDCT macro */
213 #define SSE2_Row_IDCT() { \
215 movdqa_m2r(*I(3), xmm2); /* xmm2 = i3 */ \
216 movdqa_m2r(*C(3), xmm6); /* xmm6 = c3 */ \
218 movdqa_r2r(xmm2, xmm4); /* xmm4 = i3 */ \
219 movdqa_m2r(*I(5), xmm7); /* xmm7 = i5 */ \
221 pmulhw_r2r(xmm6, xmm4); /* xmm4 = c3 * i3 - i3 */ \
222 movdqa_m2r(*C(5), xmm1); /* xmm1 = c5 */ \
224 pmulhw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 - i5 */ \
225 movdqa_r2r(xmm1, xmm5); /* xmm5 = c5 */ \
227 pmulhw_r2r(xmm2, xmm1); /* xmm1 = c5 * i3 - i3 */ \
228 movdqa_m2r(*I(1), xmm3); /* xmm3 = i1 */ \
230 pmulhw_r2r(xmm7, xmm5); /* xmm5 = c5 * i5 - i5 */ \
231 movdqa_m2r(*C(1), xmm0); /* xmm0 = c1 */ \
233 /* all registers are in use */ \
235 paddw_r2r(xmm2, xmm4); /* xmm4 = c3 * i3 */ \
236 paddw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 */ \
238 paddw_r2r(xmm1, xmm2); /* xmm2 = c5 * i3 */ \
239 movdqa_m2r(*I(7), xmm1); /* xmm1 = i7 */ \
241 paddw_r2r(xmm5, xmm7); /* xmm7 = c5 * i5 */ \
242 movdqa_r2r(xmm0, xmm5); /* xmm5 = c1 */ \
244 pmulhw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 - i1 */ \
245 paddsw_r2r(xmm7, xmm4); /* xmm4 = c3 * i3 + c5 * i5 = C */ \
247 pmulhw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 - i7 */ \
248 movdqa_m2r(*C(7), xmm7); /* xmm7 = c7 */ \
250 psubsw_r2r(xmm2, xmm6); /* xmm6 = c3 * i5 - c5 * i3 = D */ \
251 paddw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 */ \
253 pmulhw_r2r(xmm7, xmm3); /* xmm3 = c7 * i1 */ \
254 movdqa_m2r(*I(2), xmm2); /* xmm2 = i2 */ \
256 pmulhw_r2r(xmm1, xmm7); /* xmm7 = c7 * i7 */ \
257 paddw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 */ \
259 movdqa_r2r(xmm2, xmm1); /* xmm1 = i2 */ \
260 pmulhw_m2r(*C(2), xmm2); /* xmm2 = i2 * c2 -i2 */ \
262 psubsw_r2r(xmm5, xmm3); /* xmm3 = c7 * i1 - c1 * i7 = B */ \
263 movdqa_m2r(*I(6), xmm5); /* xmm5 = i6 */ \
265 paddsw_r2r(xmm7, xmm0); /* xmm0 = c1 * i1 + c7 * i7 = A */ \
266 movdqa_r2r(xmm5, xmm7); /* xmm7 = i6 */ \
268 psubsw_r2r(xmm4, xmm0); /* xmm0 = A - C */ \
269 pmulhw_m2r(*C(2), xmm5); /* xmm5 = c2 * i6 - i6 */ \
271 paddw_r2r(xmm1, xmm2); /* xmm2 = i2 * c2 */ \
272 pmulhw_m2r(*C(6), xmm1); /* xmm1 = c6 * i2 */ \
274 paddsw_r2r(xmm4, xmm4); /* xmm4 = C + C */ \
275 paddsw_r2r(xmm0, xmm4); /* xmm4 = A + C = C. */ \
277 psubsw_r2r(xmm6, xmm3); /* xmm3 = B - D */ \
278 paddw_r2r(xmm7, xmm5); /* xmm5 = c2 * i6 */ \
280 paddsw_r2r(xmm6, xmm6); /* xmm6 = D + D */ \
281 pmulhw_m2r(*C(6), xmm7); /* xmm7 = c6 * i6 */ \
283 paddsw_r2r(xmm3, xmm6); /* xmm6 = B + D = D. */ \
284 movdqa_r2m(xmm4, *I(1)); /* Save C. at I(1) */ \
286 psubsw_r2r(xmm5, xmm1); /* xmm1 = c6 * i2 - c2 * i6 = H */ \
287 movdqa_m2r(*C(4), xmm4); /* xmm4 = c4 */ \
289 movdqa_r2r(xmm3, xmm5); /* xmm5 = B - D */ \
290 pmulhw_r2r(xmm4, xmm3); /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
292 paddsw_r2r(xmm2, xmm7); /* xmm7 = c2 * i2 + c6 * i6 = G */ \
293 movdqa_r2m(xmm6, *I(2)); /* Save D. at I(2) */ \
295 movdqa_r2r(xmm0, xmm2); /* xmm2 = A - C */ \
296 movdqa_m2r(*I(0), xmm6); /* xmm6 = i0 */ \
298 pmulhw_r2r(xmm4, xmm0); /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
299 paddw_r2r(xmm3, xmm5); /* xmm5 = c4 * ( B - D ) = B. */ \
301 movdqa_m2r(*I(4), xmm3); /* xmm3 = i4 */ \
302 psubsw_r2r(xmm1, xmm5); /* xmm5 = B. - H = B.. */ \
304 paddw_r2r(xmm0, xmm2); /* xmm2 = c4 * ( A - C) = A. */ \
305 psubsw_r2r(xmm3, xmm6); /* xmm6 = i0 - i4 */ \
307 movdqa_r2r(xmm6, xmm0); /* xmm0 = i0 - i4 */ \
308 pmulhw_r2r(xmm4, xmm6); /* xmm6 = ( c4 - 1 ) * ( i0 - i4 ) = F */ \
310 paddsw_r2r(xmm3, xmm3); /* xmm3 = i4 + i4 */ \
311 paddsw_r2r(xmm1, xmm1); /* xmm1 = H + H */ \
313 paddsw_r2r(xmm0, xmm3); /* xmm3 = i0 + i4 */ \
314 paddsw_r2r(xmm5, xmm1); /* xmm1 = B. + H = H. */ \
316 pmulhw_r2r(xmm3, xmm4); /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
317 paddw_r2r(xmm0, xmm6); /* xmm6 = c4 * ( i0 - i4 ) */ \
319 psubsw_r2r(xmm2, xmm6); /* xmm6 = F - A. = F. */ \
320 paddsw_r2r(xmm2, xmm2); /* xmm2 = A. + A. */ \
322 movdqa_m2r(*I(1), xmm0); /* Load C. from I(1) */ \
323 paddsw_r2r(xmm6, xmm2); /* xmm2 = F + A. = A.. */ \
325 paddw_r2r(xmm3, xmm4); /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
326 psubsw_r2r(xmm1, xmm2); /* xmm2 = A.. - H. = R2 */ \
328 paddsw_r2r(xmm1, xmm1); /* xmm1 = H. + H. */ \
329 paddsw_r2r(xmm2, xmm1); /* xmm1 = A.. + H. = R1 */ \
331 psubsw_r2r(xmm7, xmm4); /* xmm4 = E - G = E. */ \
333 movdqa_m2r(*I(2), xmm3); /* Load D. from I(2) */ \
334 paddsw_r2r(xmm7, xmm7); /* xmm7 = G + G */ \
336 movdqa_r2m(xmm2, *I(2)); /* Write out op2 */ \
337 paddsw_r2r(xmm4, xmm7); /* xmm7 = E + G = G. */ \
339 movdqa_r2m(xmm1, *I(1)); /* Write out op1 */ \
340 psubsw_r2r(xmm3, xmm4); /* xmm4 = E. - D. = R4 */ \
342 paddsw_r2r(xmm3, xmm3); /* xmm3 = D. + D. */ \
344 paddsw_r2r(xmm4, xmm3); /* xmm3 = E. + D. = R3 */ \
346 psubsw_r2r(xmm5, xmm6); /* xmm6 = F. - B..= R6 */ \
348 paddsw_r2r(xmm5, xmm5); /* xmm5 = B.. + B.. */ \
350 paddsw_r2r(xmm6, xmm5); /* xmm5 = F. + B.. = R5 */ \
352 movdqa_r2m(xmm4, *I(4)); /* Write out op4 */ \
354 movdqa_r2m(xmm3, *I(3)); /* Write out op3 */ \
355 psubsw_r2r(xmm0, xmm7); /* xmm7 = G. - C. = R7 */ \
357 paddsw_r2r(xmm0, xmm0); /* xmm0 = C. + C. */ \
359 paddsw_r2r(xmm7, xmm0); /* xmm0 = G. + C. */ \
361 movdqa_r2m(xmm6, *I(6)); /* Write out op6 */ \
363 movdqa_r2m(xmm5, *I(5)); /* Write out op5 */ \
364 movdqa_r2m(xmm7, *I(7)); /* Write out op7 */ \
366 movdqa_r2m(xmm0, *I(0)); /* Write out op0 */ \
368 } /* End of SSE2_Row_IDCT macro */
371 #define SSE2_Transpose() { \
373 movdqa_m2r(*I(4), xmm4); /* xmm4=e7e6e5e4e3e2e1e0 */ \
374 movdqa_m2r(*I(5), xmm0); /* xmm4=f7f6f5f4f3f2f1f0 */ \
376 movdqa_r2r(xmm4, xmm5); /* make a copy */ \
377 punpcklwd_r2r(xmm0, xmm4); /* xmm4=f3e3f2e2f1e1f0e0 */ \
379 punpckhwd_r2r(xmm0, xmm5); /* xmm5=f7e7f6e6f5e5f4e4 */ \
380 movdqa_m2r(*I(6), xmm6); /* xmm6=g7g6g5g4g3g2g1g0 */ \
382 movdqa_m2r(*I(7), xmm0); /* xmm0=h7h6h5h4h3h2h1h0 */ \
383 movdqa_r2r(xmm6, xmm7); /* make a copy */ \
385 punpcklwd_r2r(xmm0, xmm6); /* xmm6=h3g3h3g2h1g1h0g0 */ \
386 punpckhwd_r2r(xmm0, xmm7); /* xmm7=h7g7h6g6h5g5h4g4 */ \
388 movdqa_r2r(xmm4, xmm3); /* make a copy */ \
389 punpckldq_r2r(xmm6, xmm4); /* xmm4=h1g1f1e1h0g0f0e0 */ \
391 punpckhdq_r2r(xmm6, xmm3); /* xmm3=h3g3g3e3h2g2f2e2 */ \
392 movdqa_r2m(xmm3, *I(6)); /* save h3g3g3e3h2g2f2e2 */ \
394 movdqa_r2r(xmm5, xmm6); /* make a copy */ \
395 punpckldq_r2r(xmm7, xmm5); /* xmm5=h5g5f5e5h4g4f4e4 */ \
397 punpckhdq_r2r(xmm7, xmm6); /* xmm6=h7g7f7e7h6g6f6e6 */ \
398 movdqa_m2r(*I(0), xmm0); /* xmm0=a7a6a5a4a3a2a1a0 */ \
400 movdqa_m2r(*I(1), xmm1); /* xmm1=b7b6b5b4b3b2b1b0 */ \
401 movdqa_r2r(xmm0, xmm7); /* make a copy */ \
403 punpcklwd_r2r(xmm1, xmm0); /* xmm0=b3a3b2a2b1a1b0a0 */ \
404 punpckhwd_r2r(xmm1, xmm7); /* xmm7=b7a7b6a6b5a5b4a4 */ \
406 movdqa_m2r(*I(2), xmm2); /* xmm2=c7c6c5c4c3c2c1c0 */ \
407 movdqa_m2r(*I(3), xmm3); /* xmm3=d7d6d5d4d3d2d1d0 */ \
409 movdqa_r2r(xmm2, xmm1); /* make a copy */ \
410 punpcklwd_r2r(xmm3, xmm2); /* xmm2=d3c3d2c2d1c1d0c0 */ \
412 punpckhwd_r2r(xmm3, xmm1); /* xmm1=d7c7d6c6d5c5d4c4 */ \
413 movdqa_r2r(xmm0, xmm3); /* make a copy */ \
415 punpckldq_r2r(xmm2, xmm0); /* xmm0=d1c1b1a1d0c0b0a0 */ \
416 punpckhdq_r2r(xmm2, xmm3); /* xmm3=d3c3b3a3d2c2b2a2 */ \
418 movdqa_r2r(xmm7, xmm2); /* make a copy */ \
419 punpckldq_r2r(xmm1, xmm2); /* xmm2=d5c5b5a5d4c4b4a4 */ \
421 punpckhdq_r2r(xmm1, xmm7); /* xmm7=d7c7b7a7d6c6b6a6 */ \
422 movdqa_r2r(xmm0, xmm1); /* make a copy */ \
424 punpcklqdq_r2r(xmm4, xmm0); /* xmm0=h0g0f0e0d0c0b0a0 */ \
425 punpckhqdq_r2r(xmm4, xmm1); /* xmm1=h1g1g1e1d1c1b1a1 */ \
427 movdqa_r2m(xmm0, *I(0)); /* save I(0) */ \
428 movdqa_r2m(xmm1, *I(1)); /* save I(1) */ \
430 movdqa_m2r(*I(6), xmm0); /* load h3g3g3e3h2g2f2e2 */ \
431 movdqa_r2r(xmm3, xmm1); /* make a copy */ \
433 punpcklqdq_r2r(xmm0, xmm1); /* xmm1=h2g2f2e2d2c2b2a2 */ \
434 punpckhqdq_r2r(xmm0, xmm3); /* xmm3=h3g3f3e3d3c3b3a3 */ \
436 movdqa_r2r(xmm2, xmm4); /* make a copy */ \
437 punpcklqdq_r2r(xmm5, xmm4); /* xmm4=h4g4f4e4d4c4b4a4 */ \
439 punpckhqdq_r2r(xmm5, xmm2); /* xmm2=h5g5f5e5d5c5b5a5 */ \
440 movdqa_r2m(xmm1, *I(2)); /* save I(2) */ \
442 movdqa_r2m(xmm3, *I(3)); /* save I(3) */ \
443 movdqa_r2m(xmm4, *I(4)); /* save I(4) */ \
445 movdqa_r2m(xmm2, *I(5)); /* save I(5) */ \
446 movdqa_r2r(xmm7, xmm5); /* make a copy */ \
448 punpcklqdq_r2r(xmm6, xmm5); /* xmm5=h6g6f6e6d6c6b6a6 */ \
449 punpckhqdq_r2r(xmm6, xmm7); /* xmm7=h7g7f7e7d7c7b7a7 */ \
451 movdqa_r2m(xmm5, *I(6)); /* save I(6) */ \
452 movdqa_r2m(xmm7, *I(7)); /* save I(7) */ \
454 } /* End of Transpose Macro */
456 void ff_vp3_idct_sse2(int16_t *input_data)
458 unsigned char *input_bytes = (unsigned char *)input_data;
459 unsigned char *output_data_bytes = (unsigned char *)input_data;
460 const unsigned char *idct_data_bytes = (const unsigned char *)SSE2_idct_data;
461 const unsigned char *Eight = (const unsigned char *)&ff_pw_8;
463 #define eax input_bytes
464 #define edx idct_data_bytes
466 #define I(i) (eax + 16 * i)
467 #define O(i) (ebx + 16 * i)
468 #define C(i) (edx + 16 * (i-1))
470 #define ebx output_data_bytes
479 void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block)
481 ff_vp3_idct_sse2(block);
482 put_signed_pixels_clamped_mmx(block, dest, line_size);
485 void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block)
487 ff_vp3_idct_sse2(block);
488 add_pixels_clamped_mmx(block, dest, line_size);