avcodec: Rename xvidmmx IDCT to xvid
[libav.git] / libavcodec / x86 / idct_sse2_xvid.c
1 /*
2 * XVID MPEG-4 VIDEO CODEC
3 * - SSE2 inverse discrete cosine transform -
4 *
5 * Copyright(C) 2003 Pascal Massimino <skal@planet-d.net>
6 *
7 * Conversion to gcc syntax with modifications
8 * by Alexander Strange <astrange@ithinksw.com>
9 *
10 * Originally from dct/x86_asm/fdct_sse2_skal.asm in Xvid.
11 *
12 * This file is part of Libav.
13 *
14 * Vertical pass is an implementation of the scheme:
15 * Loeffler C., Ligtenberg A., and Moschytz C.S.:
16 * Practical Fast 1D DCT Algorithm with Eleven Multiplications,
17 * Proc. ICASSP 1989, 988-991.
18 *
19 * Horizontal pass is a double 4x4 vector/matrix multiplication,
20 * (see also Intel's Application Note 922:
21 * http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
22 * Copyright (C) 1999 Intel Corporation)
23 *
24 * More details at http://skal.planet-d.net/coding/dct.html
25 *
26 * Libav is free software; you can redistribute it and/or
27 * modify it under the terms of the GNU Lesser General Public
28 * License as published by the Free Software Foundation; either
29 * version 2.1 of the License, or (at your option) any later version.
30 *
31 * Libav is distributed in the hope that it will be useful,
32 * but WITHOUT ANY WARRANTY; without even the implied warranty of
33 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
34 * Lesser General Public License for more details.
35 *
36 * You should have received a copy of the GNU Lesser General Public License
37 * along with Libav; if not, write to the Free Software Foundation,
38 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
39 */
40
41 #include "libavutil/internal.h"
42 #include "libavutil/mem.h"
43 #include "libavutil/x86/asm.h"
44 #include "idct_xvid.h"
45 #include "idctdsp.h"
46
47 #if HAVE_SSE2_INLINE
48
49 /**
50 * @file
51 * @brief SSE2 IDCT compatible with the Xvid IDCT
52 */
53
54 #define X8(x) x,x,x,x,x,x,x,x
55
56 #define ROW_SHIFT 11
57 #define COL_SHIFT 6
58
59 DECLARE_ASM_CONST(16, int16_t, tan1)[] = {X8(13036)}; // tan( pi/16)
60 DECLARE_ASM_CONST(16, int16_t, tan2)[] = {X8(27146)}; // tan(2pi/16) = sqrt(2)-1
61 DECLARE_ASM_CONST(16, int16_t, tan3)[] = {X8(43790)}; // tan(3pi/16)-1
62 DECLARE_ASM_CONST(16, int16_t, sqrt2)[]= {X8(23170)}; // 0.5/sqrt(2)
63 DECLARE_ASM_CONST(8, uint8_t, m127)[] = {X8(127)};
64
65 DECLARE_ASM_CONST(16, int16_t, iTab1)[] = {
66 0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d,
67 0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61,
68 0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7,
69 0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b
70 };
71
72 DECLARE_ASM_CONST(16, int16_t, iTab2)[] = {
73 0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5,
74 0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04,
75 0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41,
76 0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df
77 };
78
79 DECLARE_ASM_CONST(16, int16_t, iTab3)[] = {
80 0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf,
81 0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf,
82 0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d,
83 0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04
84 };
85
86 DECLARE_ASM_CONST(16, int16_t, iTab4)[] = {
87 0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746,
88 0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac,
89 0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df,
90 0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e
91 };
92
93 DECLARE_ASM_CONST(16, int32_t, walkenIdctRounders)[] = {
94 65536, 65536, 65536, 65536,
95 3597, 3597, 3597, 3597,
96 2260, 2260, 2260, 2260,
97 1203, 1203, 1203, 1203,
98 120, 120, 120, 120,
99 512, 512, 512, 512
100 };
101
102 // Temporary storage before the column pass
103 #define ROW1 "%%xmm6"
104 #define ROW3 "%%xmm4"
105 #define ROW5 "%%xmm5"
106 #define ROW7 "%%xmm7"
107
108 #define CLEAR_ODD(r) "pxor "r","r" \n\t"
109 #define PUT_ODD(dst) "pshufhw $0x1B, %%xmm2, "dst" \n\t"
110
111 #if ARCH_X86_64
112
113 # define ROW0 "%%xmm8"
114 # define REG0 ROW0
115 # define ROW2 "%%xmm9"
116 # define REG2 ROW2
117 # define ROW4 "%%xmm10"
118 # define REG4 ROW4
119 # define ROW6 "%%xmm11"
120 # define REG6 ROW6
121 # define CLEAR_EVEN(r) CLEAR_ODD(r)
122 # define PUT_EVEN(dst) PUT_ODD(dst)
123 # define XMMS "%%xmm12"
124 # define MOV_32_ONLY "#"
125 # define SREG2 REG2
126 # define TAN3 "%%xmm13"
127 # define TAN1 "%%xmm14"
128
129 #else
130
131 # define ROW0 "(%0)"
132 # define REG0 "%%xmm4"
133 # define ROW2 "2*16(%0)"
134 # define REG2 "%%xmm4"
135 # define ROW4 "4*16(%0)"
136 # define REG4 "%%xmm6"
137 # define ROW6 "6*16(%0)"
138 # define REG6 "%%xmm6"
139 # define CLEAR_EVEN(r)
140 # define PUT_EVEN(dst) \
141 "pshufhw $0x1B, %%xmm2, %%xmm2 \n\t" \
142 "movdqa %%xmm2, "dst" \n\t"
143 # define XMMS "%%xmm2"
144 # define MOV_32_ONLY "movdqa "
145 # define SREG2 "%%xmm7"
146 # define TAN3 "%%xmm0"
147 # define TAN1 "%%xmm2"
148
149 #endif
150
151 #define ROUND(x) "paddd "MANGLE(x)
152
153 #define JZ(reg, to) \
154 "testl "reg","reg" \n\t" \
155 "jz "to" \n\t"
156
157 #define JNZ(reg, to) \
158 "testl "reg","reg" \n\t" \
159 "jnz "to" \n\t"
160
161 #define TEST_ONE_ROW(src, reg, clear) \
162 clear \
163 "movq "src", %%mm1 \n\t" \
164 "por 8+"src", %%mm1 \n\t" \
165 "paddusb %%mm0, %%mm1 \n\t" \
166 "pmovmskb %%mm1, "reg" \n\t"
167
168 #define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2) \
169 clear1 \
170 clear2 \
171 "movq "row1", %%mm1 \n\t" \
172 "por 8+"row1", %%mm1 \n\t" \
173 "movq "row2", %%mm2 \n\t" \
174 "por 8+"row2", %%mm2 \n\t" \
175 "paddusb %%mm0, %%mm1 \n\t" \
176 "paddusb %%mm0, %%mm2 \n\t" \
177 "pmovmskb %%mm1, "reg1" \n\t" \
178 "pmovmskb %%mm2, "reg2" \n\t"
179
180 ///IDCT pass on rows.
181 #define iMTX_MULT(src, table, rounder, put) \
182 "movdqa "src", %%xmm3 \n\t" \
183 "movdqa %%xmm3, %%xmm0 \n\t" \
184 "pshufd $0x11, %%xmm3, %%xmm1 \n\t" /* 4602 */ \
185 "punpcklqdq %%xmm0, %%xmm0 \n\t" /* 0246 */ \
186 "pmaddwd "table", %%xmm0 \n\t" \
187 "pmaddwd 16+"table", %%xmm1 \n\t" \
188 "pshufd $0xBB, %%xmm3, %%xmm2 \n\t" /* 5713 */ \
189 "punpckhqdq %%xmm3, %%xmm3 \n\t" /* 1357 */ \
190 "pmaddwd 32+"table", %%xmm2 \n\t" \
191 "pmaddwd 48+"table", %%xmm3 \n\t" \
192 "paddd %%xmm1, %%xmm0 \n\t" \
193 "paddd %%xmm3, %%xmm2 \n\t" \
194 rounder", %%xmm0 \n\t" \
195 "movdqa %%xmm2, %%xmm3 \n\t" \
196 "paddd %%xmm0, %%xmm2 \n\t" \
197 "psubd %%xmm3, %%xmm0 \n\t" \
198 "psrad $11, %%xmm2 \n\t" \
199 "psrad $11, %%xmm0 \n\t" \
200 "packssdw %%xmm0, %%xmm2 \n\t" \
201 put \
202 "1: \n\t"
203
204 #define iLLM_HEAD \
205 "movdqa "MANGLE(tan3)", "TAN3" \n\t" \
206 "movdqa "MANGLE(tan1)", "TAN1" \n\t" \
207
208 ///IDCT pass on columns.
209 #define iLLM_PASS(dct) \
210 "movdqa "TAN3", %%xmm1 \n\t" \
211 "movdqa "TAN1", %%xmm3 \n\t" \
212 "pmulhw %%xmm4, "TAN3" \n\t" \
213 "pmulhw %%xmm5, %%xmm1 \n\t" \
214 "paddsw %%xmm4, "TAN3" \n\t" \
215 "paddsw %%xmm5, %%xmm1 \n\t" \
216 "psubsw %%xmm5, "TAN3" \n\t" \
217 "paddsw %%xmm4, %%xmm1 \n\t" \
218 "pmulhw %%xmm7, %%xmm3 \n\t" \
219 "pmulhw %%xmm6, "TAN1" \n\t" \
220 "paddsw %%xmm6, %%xmm3 \n\t" \
221 "psubsw %%xmm7, "TAN1" \n\t" \
222 "movdqa %%xmm3, %%xmm7 \n\t" \
223 "movdqa "TAN1", %%xmm6 \n\t" \
224 "psubsw %%xmm1, %%xmm3 \n\t" \
225 "psubsw "TAN3", "TAN1" \n\t" \
226 "paddsw %%xmm7, %%xmm1 \n\t" \
227 "paddsw %%xmm6, "TAN3" \n\t" \
228 "movdqa %%xmm3, %%xmm6 \n\t" \
229 "psubsw "TAN3", %%xmm3 \n\t" \
230 "paddsw %%xmm6, "TAN3" \n\t" \
231 "movdqa "MANGLE(sqrt2)", %%xmm4 \n\t" \
232 "pmulhw %%xmm4, %%xmm3 \n\t" \
233 "pmulhw %%xmm4, "TAN3" \n\t" \
234 "paddsw "TAN3", "TAN3" \n\t" \
235 "paddsw %%xmm3, %%xmm3 \n\t" \
236 "movdqa "MANGLE(tan2)", %%xmm7 \n\t" \
237 MOV_32_ONLY ROW2", "REG2" \n\t" \
238 MOV_32_ONLY ROW6", "REG6" \n\t" \
239 "movdqa %%xmm7, %%xmm5 \n\t" \
240 "pmulhw "REG6", %%xmm7 \n\t" \
241 "pmulhw "REG2", %%xmm5 \n\t" \
242 "paddsw "REG2", %%xmm7 \n\t" \
243 "psubsw "REG6", %%xmm5 \n\t" \
244 MOV_32_ONLY ROW0", "REG0" \n\t" \
245 MOV_32_ONLY ROW4", "REG4" \n\t" \
246 MOV_32_ONLY" "TAN1", (%0) \n\t" \
247 "movdqa "REG0", "XMMS" \n\t" \
248 "psubsw "REG4", "REG0" \n\t" \
249 "paddsw "XMMS", "REG4" \n\t" \
250 "movdqa "REG4", "XMMS" \n\t" \
251 "psubsw %%xmm7, "REG4" \n\t" \
252 "paddsw "XMMS", %%xmm7 \n\t" \
253 "movdqa "REG0", "XMMS" \n\t" \
254 "psubsw %%xmm5, "REG0" \n\t" \
255 "paddsw "XMMS", %%xmm5 \n\t" \
256 "movdqa %%xmm5, "XMMS" \n\t" \
257 "psubsw "TAN3", %%xmm5 \n\t" \
258 "paddsw "XMMS", "TAN3" \n\t" \
259 "movdqa "REG0", "XMMS" \n\t" \
260 "psubsw %%xmm3, "REG0" \n\t" \
261 "paddsw "XMMS", %%xmm3 \n\t" \
262 MOV_32_ONLY" (%0), "TAN1" \n\t" \
263 "psraw $6, %%xmm5 \n\t" \
264 "psraw $6, "REG0" \n\t" \
265 "psraw $6, "TAN3" \n\t" \
266 "psraw $6, %%xmm3 \n\t" \
267 "movdqa "TAN3", 1*16("dct") \n\t" \
268 "movdqa %%xmm3, 2*16("dct") \n\t" \
269 "movdqa "REG0", 5*16("dct") \n\t" \
270 "movdqa %%xmm5, 6*16("dct") \n\t" \
271 "movdqa %%xmm7, %%xmm0 \n\t" \
272 "movdqa "REG4", %%xmm4 \n\t" \
273 "psubsw %%xmm1, %%xmm7 \n\t" \
274 "psubsw "TAN1", "REG4" \n\t" \
275 "paddsw %%xmm0, %%xmm1 \n\t" \
276 "paddsw %%xmm4, "TAN1" \n\t" \
277 "psraw $6, %%xmm1 \n\t" \
278 "psraw $6, %%xmm7 \n\t" \
279 "psraw $6, "TAN1" \n\t" \
280 "psraw $6, "REG4" \n\t" \
281 "movdqa %%xmm1, ("dct") \n\t" \
282 "movdqa "TAN1", 3*16("dct") \n\t" \
283 "movdqa "REG4", 4*16("dct") \n\t" \
284 "movdqa %%xmm7, 7*16("dct") \n\t"
285
286 ///IDCT pass on columns, assuming rows 4-7 are zero.
287 #define iLLM_PASS_SPARSE(dct) \
288 "pmulhw %%xmm4, "TAN3" \n\t" \
289 "paddsw %%xmm4, "TAN3" \n\t" \
290 "movdqa %%xmm6, %%xmm3 \n\t" \
291 "pmulhw %%xmm6, "TAN1" \n\t" \
292 "movdqa %%xmm4, %%xmm1 \n\t" \
293 "psubsw %%xmm1, %%xmm3 \n\t" \
294 "paddsw %%xmm6, %%xmm1 \n\t" \
295 "movdqa "TAN1", %%xmm6 \n\t" \
296 "psubsw "TAN3", "TAN1" \n\t" \
297 "paddsw %%xmm6, "TAN3" \n\t" \
298 "movdqa %%xmm3, %%xmm6 \n\t" \
299 "psubsw "TAN3", %%xmm3 \n\t" \
300 "paddsw %%xmm6, "TAN3" \n\t" \
301 "movdqa "MANGLE(sqrt2)", %%xmm4 \n\t" \
302 "pmulhw %%xmm4, %%xmm3 \n\t" \
303 "pmulhw %%xmm4, "TAN3" \n\t" \
304 "paddsw "TAN3", "TAN3" \n\t" \
305 "paddsw %%xmm3, %%xmm3 \n\t" \
306 "movdqa "MANGLE(tan2)", %%xmm5 \n\t" \
307 MOV_32_ONLY ROW2", "SREG2" \n\t" \
308 "pmulhw "SREG2", %%xmm5 \n\t" \
309 MOV_32_ONLY ROW0", "REG0" \n\t" \
310 "movdqa "REG0", %%xmm6 \n\t" \
311 "psubsw "SREG2", %%xmm6 \n\t" \
312 "paddsw "REG0", "SREG2" \n\t" \
313 MOV_32_ONLY" "TAN1", (%0) \n\t" \
314 "movdqa "REG0", "XMMS" \n\t" \
315 "psubsw %%xmm5, "REG0" \n\t" \
316 "paddsw "XMMS", %%xmm5 \n\t" \
317 "movdqa %%xmm5, "XMMS" \n\t" \
318 "psubsw "TAN3", %%xmm5 \n\t" \
319 "paddsw "XMMS", "TAN3" \n\t" \
320 "movdqa "REG0", "XMMS" \n\t" \
321 "psubsw %%xmm3, "REG0" \n\t" \
322 "paddsw "XMMS", %%xmm3 \n\t" \
323 MOV_32_ONLY" (%0), "TAN1" \n\t" \
324 "psraw $6, %%xmm5 \n\t" \
325 "psraw $6, "REG0" \n\t" \
326 "psraw $6, "TAN3" \n\t" \
327 "psraw $6, %%xmm3 \n\t" \
328 "movdqa "TAN3", 1*16("dct") \n\t" \
329 "movdqa %%xmm3, 2*16("dct") \n\t" \
330 "movdqa "REG0", 5*16("dct") \n\t" \
331 "movdqa %%xmm5, 6*16("dct") \n\t" \
332 "movdqa "SREG2", %%xmm0 \n\t" \
333 "movdqa %%xmm6, %%xmm4 \n\t" \
334 "psubsw %%xmm1, "SREG2" \n\t" \
335 "psubsw "TAN1", %%xmm6 \n\t" \
336 "paddsw %%xmm0, %%xmm1 \n\t" \
337 "paddsw %%xmm4, "TAN1" \n\t" \
338 "psraw $6, %%xmm1 \n\t" \
339 "psraw $6, "SREG2" \n\t" \
340 "psraw $6, "TAN1" \n\t" \
341 "psraw $6, %%xmm6 \n\t" \
342 "movdqa %%xmm1, ("dct") \n\t" \
343 "movdqa "TAN1", 3*16("dct") \n\t" \
344 "movdqa %%xmm6, 4*16("dct") \n\t" \
345 "movdqa "SREG2", 7*16("dct") \n\t"
346
347 inline void ff_idct_xvid_sse2(short *block)
348 {
349 __asm__ volatile(
350 "movq "MANGLE(m127)", %%mm0 \n\t"
351 iMTX_MULT("(%0)", MANGLE(iTab1), ROUND(walkenIdctRounders), PUT_EVEN(ROW0))
352 iMTX_MULT("1*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+1*16), PUT_ODD(ROW1))
353 iMTX_MULT("2*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+2*16), PUT_EVEN(ROW2))
354
355 TEST_TWO_ROWS("3*16(%0)", "4*16(%0)", "%%eax", "%%ecx", CLEAR_ODD(ROW3), CLEAR_EVEN(ROW4))
356 JZ("%%eax", "1f")
357 iMTX_MULT("3*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+3*16), PUT_ODD(ROW3))
358
359 TEST_TWO_ROWS("5*16(%0)", "6*16(%0)", "%%eax", "%%edx", CLEAR_ODD(ROW5), CLEAR_EVEN(ROW6))
360 TEST_ONE_ROW("7*16(%0)", "%%esi", CLEAR_ODD(ROW7))
361 iLLM_HEAD
362 ".p2align 4 \n\t"
363 JNZ("%%ecx", "2f")
364 JNZ("%%eax", "3f")
365 JNZ("%%edx", "4f")
366 JNZ("%%esi", "5f")
367 iLLM_PASS_SPARSE("%0")
368 "jmp 6f \n\t"
369 "2: \n\t"
370 iMTX_MULT("4*16(%0)", MANGLE(iTab1), "#", PUT_EVEN(ROW4))
371 "3: \n\t"
372 iMTX_MULT("5*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+4*16), PUT_ODD(ROW5))
373 JZ("%%edx", "1f")
374 "4: \n\t"
375 iMTX_MULT("6*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+5*16), PUT_EVEN(ROW6))
376 JZ("%%esi", "1f")
377 "5: \n\t"
378 iMTX_MULT("7*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+5*16), PUT_ODD(ROW7))
379 #if ARCH_X86_32
380 iLLM_HEAD
381 #endif
382 iLLM_PASS("%0")
383 "6: \n\t"
384 : "+r"(block)
385 :
386 : XMM_CLOBBERS("%xmm0" , "%xmm1" , "%xmm2" , "%xmm3" ,
387 "%xmm4" , "%xmm5" , "%xmm6" , "%xmm7" ,)
388 #if ARCH_X86_64
389 XMM_CLOBBERS("%xmm8" , "%xmm9" , "%xmm10", "%xmm11",
390 "%xmm12", "%xmm13", "%xmm14",)
391 #endif
392 "%eax", "%ecx", "%edx", "%esi", "memory"
393 );
394 }
395
396 void ff_idct_xvid_sse2_put(uint8_t *dest, int line_size, short *block)
397 {
398 ff_idct_xvid_sse2(block);
399 ff_put_pixels_clamped_mmx(block, dest, line_size);
400 }
401
402 void ff_idct_xvid_sse2_add(uint8_t *dest, int line_size, short *block)
403 {
404 ff_idct_xvid_sse2(block);
405 ff_add_pixels_clamped_mmx(block, dest, line_size);
406 }
407
408 #endif /* HAVE_SSE2_INLINE */