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"
29 DECLARE_ALIGNED_16(const uint16_t, ff_vp3_idct_data[7 * 8]) =
31 64277,64277,64277,64277,64277,64277,64277,64277,
32 60547,60547,60547,60547,60547,60547,60547,60547,
33 54491,54491,54491,54491,54491,54491,54491,54491,
34 46341,46341,46341,46341,46341,46341,46341,46341,
35 36410,36410,36410,36410,36410,36410,36410,36410,
36 25080,25080,25080,25080,25080,25080,25080,25080,
37 12785,12785,12785,12785,12785,12785,12785,12785
41 #define SSE2_Column_IDCT() \
43 "movdqa "I(3)", %%xmm2 \n\t" /* xmm2 = i3 */ \
44 "movdqa "C(3)", %%xmm6 \n\t" /* xmm6 = c3 */ \
46 "movdqa %%xmm2, %%xmm4 \n\t" /* xmm4 = i3 */ \
47 "movdqa "I(5)", %%xmm7 \n\t" /* xmm7 = i5 */ \
49 "pmulhw %%xmm6, %%xmm4 \n\t" /* xmm4 = c3 * i3 - i3 */ \
50 "movdqa "C(5)", %%xmm1 \n\t" /* xmm1 = c5 */ \
52 "pmulhw %%xmm7, %%xmm6 \n\t" /* xmm6 = c3 * i5 - i5 */ \
53 "movdqa %%xmm1, %%xmm5 \n\t" /* xmm5 = c5 */ \
55 "pmulhw %%xmm2, %%xmm1 \n\t" /* xmm1 = c5 * i3 - i3 */ \
56 "movdqa "I(1)", %%xmm3 \n\t" /* xmm3 = i1 */ \
58 "pmulhw %%xmm7, %%xmm5 \n\t" /* xmm5 = c5 * i5 - i5 */ \
59 "movdqa "C(1)", %%xmm0 \n\t" /* xmm0 = c1 */ \
61 /* all registers are in use */ \
63 "paddw %%xmm2, %%xmm4 \n\t" /* xmm4 = c3 * i3 */ \
64 "paddw %%xmm7, %%xmm6 \n\t" /* xmm6 = c3 * i5 */ \
66 "paddw %%xmm1, %%xmm2 \n\t" /* xmm2 = c5 * i3 */ \
67 "movdqa "I(7)", %%xmm1 \n\t" /* xmm1 = i7 */ \
69 "paddw %%xmm5, %%xmm7 \n\t" /* xmm7 = c5 * i5 */ \
70 "movdqa %%xmm0, %%xmm5 \n\t" /* xmm5 = c1 */ \
72 "pmulhw %%xmm3, %%xmm0 \n\t" /* xmm0 = c1 * i1 - i1 */ \
73 "paddsw %%xmm7, %%xmm4 \n\t" /* xmm4 = c3 * i3 + c5 * i5 = C */ \
75 "pmulhw %%xmm1, %%xmm5 \n\t" /* xmm5 = c1 * i7 - i7 */ \
76 "movdqa "C(7)", %%xmm7 \n\t" /* xmm7 = c7 */ \
78 "psubsw %%xmm2, %%xmm6 \n\t" /* xmm6 = c3 * i5 - c5 * i3 = D */ \
79 "paddw %%xmm3, %%xmm0 \n\t" /* xmm0 = c1 * i1 */ \
81 "pmulhw %%xmm7, %%xmm3 \n\t" /* xmm3 = c7 * i1 */ \
82 "movdqa "I(2)", %%xmm2 \n\t" /* xmm2 = i2 */ \
84 "pmulhw %%xmm1, %%xmm7 \n\t" /* xmm7 = c7 * i7 */ \
85 "paddw %%xmm1, %%xmm5 \n\t" /* xmm5 = c1 * i7 */ \
87 "movdqa %%xmm2, %%xmm1 \n\t" /* xmm1 = i2 */ \
88 "pmulhw "C(2)", %%xmm2 \n\t" /* xmm2 = i2 * c2 -i2 */ \
90 "psubsw %%xmm5, %%xmm3 \n\t" /* xmm3 = c7 * i1 - c1 * i7 = B */ \
91 "movdqa "I(6)", %%xmm5 \n\t" /* xmm5 = i6 */ \
93 "paddsw %%xmm7, %%xmm0 \n\t" /* xmm0 = c1 * i1 + c7 * i7 = A */ \
94 "movdqa %%xmm5, %%xmm7 \n\t" /* xmm7 = i6 */ \
96 "psubsw %%xmm4, %%xmm0 \n\t" /* xmm0 = A - C */ \
97 "pmulhw "C(2)", %%xmm5 \n\t" /* xmm5 = c2 * i6 - i6 */ \
99 "paddw %%xmm1, %%xmm2 \n\t" /* xmm2 = i2 * c2 */ \
100 "pmulhw "C(6)", %%xmm1 \n\t" /* xmm1 = c6 * i2 */ \
102 "paddsw %%xmm4, %%xmm4 \n\t" /* xmm4 = C + C */ \
103 "paddsw %%xmm0, %%xmm4 \n\t" /* xmm4 = A + C = C. */ \
105 "psubsw %%xmm6, %%xmm3 \n\t" /* xmm3 = B - D */ \
106 "paddw %%xmm7, %%xmm5 \n\t" /* xmm5 = c2 * i6 */ \
108 "paddsw %%xmm6, %%xmm6 \n\t" /* xmm6 = D + D */ \
109 "pmulhw "C(6)", %%xmm7 \n\t" /* xmm7 = c6 * i6 */ \
111 "paddsw %%xmm3, %%xmm6 \n\t" /* xmm6 = B + D = D. */ \
112 "movdqa %%xmm4, "I(1)" \n\t" /* Save C. at I(1) */ \
114 "psubsw %%xmm5, %%xmm1 \n\t" /* xmm1 = c6 * i2 - c2 * i6 = H */ \
115 "movdqa "C(4)", %%xmm4 \n\t" /* xmm4 = c4 */ \
117 "movdqa %%xmm3, %%xmm5 \n\t" /* xmm5 = B - D */ \
118 "pmulhw %%xmm4, %%xmm3 \n\t" /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
120 "paddsw %%xmm2, %%xmm7 \n\t" /* xmm7 = c2 * i2 + c6 * i6 = G */ \
121 "movdqa %%xmm6, "I(2)" \n\t" /* Save D. at I(2) */ \
123 "movdqa %%xmm0, %%xmm2 \n\t" /* xmm2 = A - C */ \
124 "movdqa "I(0)", %%xmm6 \n\t" /* xmm6 = i0 */ \
126 "pmulhw %%xmm4, %%xmm0 \n\t" /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
127 "paddw %%xmm3, %%xmm5 \n\t" /* xmm5 = c4 * ( B - D ) = B. */ \
129 "movdqa "I(4)", %%xmm3 \n\t" /* xmm3 = i4 */ \
130 "psubsw %%xmm1, %%xmm5 \n\t" /* xmm5 = B. - H = B.. */ \
132 "paddw %%xmm0, %%xmm2 \n\t" /* xmm2 = c4 * ( A - C) = A. */ \
133 "psubsw %%xmm3, %%xmm6 \n\t" /* xmm6 = i0 - i4 */ \
135 "movdqa %%xmm6, %%xmm0 \n\t" /* xmm0 = i0 - i4 */ \
136 "pmulhw %%xmm4, %%xmm6 \n\t" /* xmm6 = (c4 - 1) * (i0 - i4) = F */ \
138 "paddsw %%xmm3, %%xmm3 \n\t" /* xmm3 = i4 + i4 */ \
139 "paddsw %%xmm1, %%xmm1 \n\t" /* xmm1 = H + H */ \
141 "paddsw %%xmm0, %%xmm3 \n\t" /* xmm3 = i0 + i4 */ \
142 "paddsw %%xmm5, %%xmm1 \n\t" /* xmm1 = B. + H = H. */ \
144 "pmulhw %%xmm3, %%xmm4 \n\t" /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
145 "paddw %%xmm0, %%xmm6 \n\t" /* xmm6 = c4 * ( i0 - i4 ) */ \
147 "psubsw %%xmm2, %%xmm6 \n\t" /* xmm6 = F - A. = F. */ \
148 "paddsw %%xmm2, %%xmm2 \n\t" /* xmm2 = A. + A. */ \
150 "movdqa "I(1)", %%xmm0 \n\t" /* Load C. from I(1) */ \
151 "paddsw %%xmm6, %%xmm2 \n\t" /* xmm2 = F + A. = A.. */ \
153 "paddw %%xmm3, %%xmm4 \n\t" /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
154 "psubsw %%xmm1, %%xmm2 \n\t" /* xmm2 = A.. - H. = R2 */ \
156 "paddsw "OC_8", %%xmm2 \n\t" /* Adjust R2 and R1 before shifting */ \
157 "paddsw %%xmm1, %%xmm1 \n\t" /* xmm1 = H. + H. */ \
159 "paddsw %%xmm2, %%xmm1 \n\t" /* xmm1 = A.. + H. = R1 */ \
160 "psraw $4, %%xmm2 \n\t" /* xmm2 = op2 */ \
162 "psubsw %%xmm7, %%xmm4 \n\t" /* xmm4 = E - G = E. */ \
163 "psraw $4, %%xmm1 \n\t" /* xmm1 = op1 */ \
165 "movdqa "I(2)", %%xmm3 \n\t" /* Load D. from I(2) */ \
166 "paddsw %%xmm7, %%xmm7 \n\t" /* xmm7 = G + G */ \
168 "movdqa %%xmm2, "O(2)" \n\t" /* Write out op2 */ \
169 "paddsw %%xmm4, %%xmm7 \n\t" /* xmm7 = E + G = G. */ \
171 "movdqa %%xmm1, "O(1)" \n\t" /* Write out op1 */ \
172 "psubsw %%xmm3, %%xmm4 \n\t" /* xmm4 = E. - D. = R4 */ \
174 "paddsw "OC_8", %%xmm4 \n\t" /* Adjust R4 and R3 before shifting */ \
175 "paddsw %%xmm3, %%xmm3 \n\t" /* xmm3 = D. + D. */ \
177 "paddsw %%xmm4, %%xmm3 \n\t" /* xmm3 = E. + D. = R3 */ \
178 "psraw $4, %%xmm4 \n\t" /* xmm4 = op4 */ \
180 "psubsw %%xmm5, %%xmm6 \n\t" /* xmm6 = F. - B..= R6 */ \
181 "psraw $4, %%xmm3 \n\t" /* xmm3 = op3 */ \
183 "paddsw "OC_8", %%xmm6 \n\t" /* Adjust R6 and R5 before shifting */ \
184 "paddsw %%xmm5, %%xmm5 \n\t" /* xmm5 = B.. + B.. */ \
186 "paddsw %%xmm6, %%xmm5 \n\t" /* xmm5 = F. + B.. = R5 */ \
187 "psraw $4, %%xmm6 \n\t" /* xmm6 = op6 */ \
189 "movdqa %%xmm4, "O(4)" \n\t" /* Write out op4 */ \
190 "psraw $4, %%xmm5 \n\t" /* xmm5 = op5 */ \
192 "movdqa %%xmm3, "O(3)" \n\t" /* Write out op3 */ \
193 "psubsw %%xmm0, %%xmm7 \n\t" /* xmm7 = G. - C. = R7 */ \
195 "paddsw "OC_8", %%xmm7 \n\t" /* Adjust R7 and R0 before shifting */ \
196 "paddsw %%xmm0, %%xmm0 \n\t" /* xmm0 = C. + C. */ \
198 "paddsw %%xmm7, %%xmm0 \n\t" /* xmm0 = G. + C. */ \
199 "psraw $4, %%xmm7 \n\t" /* xmm7 = op7 */ \
201 "movdqa %%xmm6, "O(6)" \n\t" /* Write out op6 */ \
202 "psraw $4, %%xmm0 \n\t" /* xmm0 = op0 */ \
204 "movdqa %%xmm5, "O(5)" \n\t" /* Write out op5 */ \
205 "movdqa %%xmm7, "O(7)" \n\t" /* Write out op7 */ \
207 "movdqa %%xmm0, "O(0)" \n\t" /* Write out op0 */
210 #define SSE2_Row_IDCT() \
212 "movdqa "I(3)", %%xmm2 \n\t" /* xmm2 = i3 */ \
213 "movdqa "C(3)", %%xmm6 \n\t" /* xmm6 = c3 */ \
215 "movdqa %%xmm2, %%xmm4 \n\t" /* xmm4 = i3 */ \
216 "movdqa "I(5)", %%xmm7 \n\t" /* xmm7 = i5 */ \
218 "pmulhw %%xmm6, %%xmm4 \n\t" /* xmm4 = c3 * i3 - i3 */ \
219 "movdqa "C(5)", %%xmm1 \n\t" /* xmm1 = c5 */ \
221 "pmulhw %%xmm7, %%xmm6 \n\t" /* xmm6 = c3 * i5 - i5 */ \
222 "movdqa %%xmm1, %%xmm5 \n\t" /* xmm5 = c5 */ \
224 "pmulhw %%xmm2, %%xmm1 \n\t" /* xmm1 = c5 * i3 - i3 */ \
225 "movdqa "I(1)", %%xmm3 \n\t" /* xmm3 = i1 */ \
227 "pmulhw %%xmm7, %%xmm5 \n\t" /* xmm5 = c5 * i5 - i5 */ \
228 "movdqa "C(1)", %%xmm0 \n\t" /* xmm0 = c1 */ \
230 /* all registers are in use */ \
232 "paddw %%xmm2, %%xmm4 \n\t" /* xmm4 = c3 * i3 */ \
233 "paddw %%xmm7, %%xmm6 \n\t" /* xmm6 = c3 * i5 */ \
235 "paddw %%xmm1, %%xmm2 \n\t" /* xmm2 = c5 * i3 */ \
236 "movdqa "I(7)", %%xmm1 \n\t" /* xmm1 = i7 */ \
238 "paddw %%xmm5, %%xmm7 \n\t" /* xmm7 = c5 * i5 */ \
239 "movdqa %%xmm0, %%xmm5 \n\t" /* xmm5 = c1 */ \
241 "pmulhw %%xmm3, %%xmm0 \n\t" /* xmm0 = c1 * i1 - i1 */ \
242 "paddsw %%xmm7, %%xmm4 \n\t" /* xmm4 = c3 * i3 + c5 * i5 = C */ \
244 "pmulhw %%xmm1, %%xmm5 \n\t" /* xmm5 = c1 * i7 - i7 */ \
245 "movdqa "C(7)", %%xmm7 \n\t" /* xmm7 = c7 */ \
247 "psubsw %%xmm2, %%xmm6 \n\t" /* xmm6 = c3 * i5 - c5 * i3 = D */ \
248 "paddw %%xmm3, %%xmm0 \n\t" /* xmm0 = c1 * i1 */ \
250 "pmulhw %%xmm7, %%xmm3 \n\t" /* xmm3 = c7 * i1 */ \
251 "movdqa "I(2)", %%xmm2 \n\t" /* xmm2 = i2 */ \
253 "pmulhw %%xmm1, %%xmm7 \n\t" /* xmm7 = c7 * i7 */ \
254 "paddw %%xmm1, %%xmm5 \n\t" /* xmm5 = c1 * i7 */ \
256 "movdqa %%xmm2, %%xmm1 \n\t" /* xmm1 = i2 */ \
257 "pmulhw "C(2)", %%xmm2 \n\t" /* xmm2 = i2 * c2 -i2 */ \
259 "psubsw %%xmm5, %%xmm3 \n\t" /* xmm3 = c7 * i1 - c1 * i7 = B */ \
260 "movdqa "I(6)", %%xmm5 \n\t" /* xmm5 = i6 */ \
262 "paddsw %%xmm7, %%xmm0 \n\t" /* xmm0 = c1 * i1 + c7 * i7 = A */ \
263 "movdqa %%xmm5, %%xmm7 \n\t" /* xmm7 = i6 */ \
265 "psubsw %%xmm4, %%xmm0 \n\t" /* xmm0 = A - C */ \
266 "pmulhw "C(2)", %%xmm5 \n\t" /* xmm5 = c2 * i6 - i6 */ \
268 "paddw %%xmm1, %%xmm2 \n\t" /* xmm2 = i2 * c2 */ \
269 "pmulhw "C(6)", %%xmm1 \n\t" /* xmm1 = c6 * i2 */ \
271 "paddsw %%xmm4, %%xmm4 \n\t" /* xmm4 = C + C */ \
272 "paddsw %%xmm0, %%xmm4 \n\t" /* xmm4 = A + C = C. */ \
274 "psubsw %%xmm6, %%xmm3 \n\t" /* xmm3 = B - D */ \
275 "paddw %%xmm7, %%xmm5 \n\t" /* xmm5 = c2 * i6 */ \
277 "paddsw %%xmm6, %%xmm6 \n\t" /* xmm6 = D + D */ \
278 "pmulhw "C(6)", %%xmm7 \n\t" /* xmm7 = c6 * i6 */ \
280 "paddsw %%xmm3, %%xmm6 \n\t" /* xmm6 = B + D = D. */ \
281 "movdqa %%xmm4, "I(1)" \n\t" /* Save C. at I(1) */ \
283 "psubsw %%xmm5, %%xmm1 \n\t" /* xmm1 = c6 * i2 - c2 * i6 = H */ \
284 "movdqa "C(4)", %%xmm4 \n\t" /* xmm4 = c4 */ \
286 "movdqa %%xmm3, %%xmm5 \n\t" /* xmm5 = B - D */ \
287 "pmulhw %%xmm4, %%xmm3 \n\t" /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
289 "paddsw %%xmm2, %%xmm7 \n\t" /* xmm7 = c2 * i2 + c6 * i6 = G */ \
290 "movdqa %%xmm6, "I(2)" \n\t" /* Save D. at I(2) */ \
292 "movdqa %%xmm0, %%xmm2 \n\t" /* xmm2 = A - C */ \
293 "movdqa "I(0)", %%xmm6 \n\t" /* xmm6 = i0 */ \
295 "pmulhw %%xmm4, %%xmm0 \n\t" /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
296 "paddw %%xmm3, %%xmm5 \n\t" /* xmm5 = c4 * ( B - D ) = B. */ \
298 "movdqa "I(4)", %%xmm3 \n\t" /* xmm3 = i4 */ \
299 "psubsw %%xmm1, %%xmm5 \n\t" /* xmm5 = B. - H = B.. */ \
301 "paddw %%xmm0, %%xmm2 \n\t" /* xmm2 = c4 * ( A - C) = A. */ \
302 "psubsw %%xmm3, %%xmm6 \n\t" /* xmm6 = i0 - i4 */ \
304 "movdqa %%xmm6, %%xmm0 \n\t" /* xmm0 = i0 - i4 */ \
305 "pmulhw %%xmm4, %%xmm6 \n\t" /* xmm6 = ( c4 - 1 ) * ( i0 - i4 ) = F */ \
307 "paddsw %%xmm3, %%xmm3 \n\t" /* xmm3 = i4 + i4 */ \
308 "paddsw %%xmm1, %%xmm1 \n\t" /* xmm1 = H + H */ \
310 "paddsw %%xmm0, %%xmm3 \n\t" /* xmm3 = i0 + i4 */ \
311 "paddsw %%xmm5, %%xmm1 \n\t" /* xmm1 = B. + H = H. */ \
313 "pmulhw %%xmm3, %%xmm4 \n\t" /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
314 "paddw %%xmm0, %%xmm6 \n\t" /* xmm6 = c4 * ( i0 - i4 ) */ \
316 "psubsw %%xmm2, %%xmm6 \n\t" /* xmm6 = F - A. = F. */ \
317 "paddsw %%xmm2, %%xmm2 \n\t" /* xmm2 = A. + A. */ \
319 "movdqa "I(1)", %%xmm0 \n\t" /* Load C. from I(1) */ \
320 "paddsw %%xmm6, %%xmm2 \n\t" /* xmm2 = F + A. = A.. */ \
322 "paddw %%xmm3, %%xmm4 \n\t" /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
323 "psubsw %%xmm1, %%xmm2 \n\t" /* xmm2 = A.. - H. = R2 */ \
325 "paddsw %%xmm1, %%xmm1 \n\t" /* xmm1 = H. + H. */ \
326 "paddsw %%xmm2, %%xmm1 \n\t" /* xmm1 = A.. + H. = R1 */ \
328 "psubsw %%xmm7, %%xmm4 \n\t" /* xmm4 = E - G = E. */ \
330 "movdqa "I(2)", %%xmm3 \n\t" /* Load D. from I(2) */ \
331 "paddsw %%xmm7, %%xmm7 \n\t" /* xmm7 = G + G */ \
333 "movdqa %%xmm2, "I(2)" \n\t" /* Write out op2 */ \
334 "paddsw %%xmm4, %%xmm7 \n\t" /* xmm7 = E + G = G. */ \
336 "movdqa %%xmm1, "I(1)" \n\t" /* Write out op1 */ \
337 "psubsw %%xmm3, %%xmm4 \n\t" /* xmm4 = E. - D. = R4 */ \
339 "paddsw %%xmm3, %%xmm3 \n\t" /* xmm3 = D. + D. */ \
341 "paddsw %%xmm4, %%xmm3 \n\t" /* xmm3 = E. + D. = R3 */ \
343 "psubsw %%xmm5, %%xmm6 \n\t" /* xmm6 = F. - B..= R6 */ \
345 "paddsw %%xmm5, %%xmm5 \n\t" /* xmm5 = B.. + B.. */ \
347 "paddsw %%xmm6, %%xmm5 \n\t" /* xmm5 = F. + B.. = R5 */ \
349 "movdqa %%xmm4, "I(4)" \n\t" /* Write out op4 */ \
351 "movdqa %%xmm3, "I(3)" \n\t" /* Write out op3 */ \
352 "psubsw %%xmm0, %%xmm7 \n\t" /* xmm7 = G. - C. = R7 */ \
354 "paddsw %%xmm0, %%xmm0 \n\t" /* xmm0 = C. + C. */ \
356 "paddsw %%xmm7, %%xmm0 \n\t" /* xmm0 = G. + C. */ \
358 "movdqa %%xmm6, "I(6)" \n\t" /* Write out op6 */ \
360 "movdqa %%xmm5, "I(5)" \n\t" /* Write out op5 */ \
361 "movdqa %%xmm7, "I(7)" \n\t" /* Write out op7 */ \
364 "movdqa %%xmm0, "I(0)" \n\t" /* Write out op0 */
367 #define SSE2_Transpose() \
369 "movdqa "I(4)", %%xmm4 \n\t" /* xmm4=e7e6e5e4e3e2e1e0 */ \
370 "movdqa "I(5)", %%xmm0 \n\t" /* xmm4=f7f6f5f4f3f2f1f0 */ \
372 "movdqa %%xmm4, %%xmm5 \n\t" /* make a copy */ \
373 "punpcklwd %%xmm0, %%xmm4 \n\t" /* xmm4=f3e3f2e2f1e1f0e0 */ \
375 "punpckhwd %%xmm0, %%xmm5 \n\t" /* xmm5=f7e7f6e6f5e5f4e4 */ \
376 "movdqa "I(6)", %%xmm6 \n\t" /* xmm6=g7g6g5g4g3g2g1g0 */ \
378 "movdqa "I(7)", %%xmm0 \n\t" /* xmm0=h7h6h5h4h3h2h1h0 */ \
379 "movdqa %%xmm6, %%xmm7 \n\t" /* make a copy */ \
381 "punpcklwd %%xmm0, %%xmm6 \n\t" /* xmm6=h3g3h3g2h1g1h0g0 */ \
382 "punpckhwd %%xmm0, %%xmm7 \n\t" /* xmm7=h7g7h6g6h5g5h4g4 */ \
384 "movdqa %%xmm4, %%xmm3 \n\t" /* make a copy */ \
385 "punpckldq %%xmm6, %%xmm4 \n\t" /* xmm4=h1g1f1e1h0g0f0e0 */ \
387 "punpckhdq %%xmm6, %%xmm3 \n\t" /* xmm3=h3g3g3e3h2g2f2e2 */ \
388 "movdqa %%xmm3, "I(6)" \n\t" /* save h3g3g3e3h2g2f2e2 */ \
390 "movdqa %%xmm5, %%xmm6 \n\t" /* make a copy */ \
391 "punpckldq %%xmm7, %%xmm5 \n\t" /* xmm5=h5g5f5e5h4g4f4e4 */ \
393 "punpckhdq %%xmm7, %%xmm6 \n\t" /* xmm6=h7g7f7e7h6g6f6e6 */ \
394 "movdqa "I(0)", %%xmm0 \n\t" /* xmm0=a7a6a5a4a3a2a1a0 */ \
396 "movdqa "I(1)", %%xmm1 \n\t" /* xmm1=b7b6b5b4b3b2b1b0 */ \
397 "movdqa %%xmm0, %%xmm7 \n\t" /* make a copy */ \
399 "punpcklwd %%xmm1, %%xmm0 \n\t" /* xmm0=b3a3b2a2b1a1b0a0 */ \
400 "punpckhwd %%xmm1, %%xmm7 \n\t" /* xmm7=b7a7b6a6b5a5b4a4 */ \
402 "movdqa "I(2)", %%xmm2 \n\t" /* xmm2=c7c6c5c4c3c2c1c0 */ \
403 "movdqa "I(3)", %%xmm3 \n\t" /* xmm3=d7d6d5d4d3d2d1d0 */ \
405 "movdqa %%xmm2, %%xmm1 \n\t" /* make a copy */ \
406 "punpcklwd %%xmm3, %%xmm2 \n\t" /* xmm2=d3c3d2c2d1c1d0c0 */ \
408 "punpckhwd %%xmm3, %%xmm1 \n\t" /* xmm1=d7c7d6c6d5c5d4c4 */ \
409 "movdqa %%xmm0, %%xmm3 \n\t" /* make a copy */ \
411 "punpckldq %%xmm2, %%xmm0 \n\t" /* xmm0=d1c1b1a1d0c0b0a0 */ \
412 "punpckhdq %%xmm2, %%xmm3 \n\t" /* xmm3=d3c3b3a3d2c2b2a2 */ \
414 "movdqa %%xmm7, %%xmm2 \n\t" /* make a copy */ \
415 "punpckldq %%xmm1, %%xmm2 \n\t" /* xmm2=d5c5b5a5d4c4b4a4 */ \
417 "punpckhdq %%xmm1, %%xmm7 \n\t" /* xmm7=d7c7b7a7d6c6b6a6 */ \
418 "movdqa %%xmm0, %%xmm1 \n\t" /* make a copy */ \
420 "punpcklqdq %%xmm4, %%xmm0 \n\t" /* xmm0=h0g0f0e0d0c0b0a0 */ \
421 "punpckhqdq %%xmm4, %%xmm1 \n\t" /* xmm1=h1g1g1e1d1c1b1a1 */ \
423 "movdqa %%xmm0, "I(0)" \n\t" /* save I(0) */ \
424 "movdqa %%xmm1, "I(1)" \n\t" /* save I(1) */ \
426 "movdqa "I(6)", %%xmm0 \n\t" /* load h3g3g3e3h2g2f2e2 */ \
427 "movdqa %%xmm3, %%xmm1 \n\t" /* make a copy */ \
429 "punpcklqdq %%xmm0, %%xmm1 \n\t" /* xmm1=h2g2f2e2d2c2b2a2 */ \
430 "punpckhqdq %%xmm0, %%xmm3 \n\t" /* xmm3=h3g3f3e3d3c3b3a3 */ \
432 "movdqa %%xmm2, %%xmm4 \n\t" /* make a copy */ \
433 "punpcklqdq %%xmm5, %%xmm4 \n\t" /* xmm4=h4g4f4e4d4c4b4a4 */ \
435 "punpckhqdq %%xmm5, %%xmm2 \n\t" /* xmm2=h5g5f5e5d5c5b5a5 */ \
436 "movdqa %%xmm1, "I(2)" \n\t" /* save I(2) */ \
438 "movdqa %%xmm3, "I(3)" \n\t" /* save I(3) */ \
439 "movdqa %%xmm4, "I(4)" \n\t" /* save I(4) */ \
441 "movdqa %%xmm2, "I(5)" \n\t" /* save I(5) */ \
442 "movdqa %%xmm7, %%xmm5 \n\t" /* make a copy */ \
444 "punpcklqdq %%xmm6, %%xmm5 \n\t" /* xmm5=h6g6f6e6d6c6b6a6 */ \
445 "punpckhqdq %%xmm6, %%xmm7 \n\t" /* xmm7=h7g7f7e7d7c7b7a7 */ \
448 "movdqa %%xmm5, "I(6)" \n\t" /* save I(6) */ \
449 "movdqa %%xmm7, "I(7)" \n\t" /* save I(7) */
451 void ff_vp3_idct_sse2(int16_t *input_data)
455 #define I(x) AV_STRINGIFY(16*x)"(%0)"
457 #define C(x) AV_STRINGIFY(16*(x-1))"(%1)"
465 :: "r"(input_data), "r"(ff_vp3_idct_data), "m"(ff_pw_8)
469 void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block)
471 ff_vp3_idct_sse2(block);
472 put_signed_pixels_clamped_mmx(block, dest, line_size);
475 void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block)
477 ff_vp3_idct_sse2(block);
478 add_pixels_clamped_mmx(block, dest, line_size);