f318e959995383d8fbf7ce7912396b8e9499b179
[libav.git] / libavcodec / x86 / xvididct_sse2.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
45 #include "idctdsp.h"
46 #include "xvididct.h"
47
48 #if HAVE_SSE2_INLINE
49
50 /**
51 * @file
52 * @brief SSE2 IDCT compatible with the Xvid IDCT
53 */
54
55 #define X8(x) x, x, x, x, x, x, x, x
56
57 DECLARE_ASM_CONST(16, int16_t, tan1)[] = { X8(13036) }; // tan( pi/16)
58 DECLARE_ASM_CONST(16, int16_t, tan2)[] = { X8(27146) }; // tan(2pi/16) = sqrt(2)-1
59 DECLARE_ASM_CONST(16, int16_t, tan3)[] = { X8(43790) }; // tan(3pi/16)-1
60 DECLARE_ASM_CONST(16, int16_t, sqrt2)[] = { X8(23170) }; // 0.5/sqrt(2)
61 DECLARE_ASM_CONST(8, uint8_t, m127)[] = { X8(127) };
62
63 DECLARE_ASM_CONST(16, int16_t, iTab1)[] = {
64 0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d,
65 0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61,
66 0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7,
67 0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b
68 };
69
70 DECLARE_ASM_CONST(16, int16_t, iTab2)[] = {
71 0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5,
72 0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04,
73 0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41,
74 0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df
75 };
76
77 DECLARE_ASM_CONST(16, int16_t, iTab3)[] = {
78 0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf,
79 0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf,
80 0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d,
81 0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04
82 };
83
84 DECLARE_ASM_CONST(16, int16_t, iTab4)[] = {
85 0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746,
86 0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac,
87 0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df,
88 0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e
89 };
90
91 DECLARE_ASM_CONST(16, int32_t, walkenIdctRounders)[] = {
92 65536, 65536, 65536, 65536,
93 3597, 3597, 3597, 3597,
94 2260, 2260, 2260, 2260,
95 1203, 1203, 1203, 1203,
96 120, 120, 120, 120,
97 512, 512, 512, 512
98 };
99
100 // Temporary storage before the column pass
101 #define ROW1 "%%xmm6"
102 #define ROW3 "%%xmm4"
103 #define ROW5 "%%xmm5"
104 #define ROW7 "%%xmm7"
105
106 #define CLEAR_ODD(r) "pxor "r","r" \n\t"
107 #define PUT_ODD(dst) "pshufhw $0x1B, %%xmm2, "dst" \n\t"
108
109 #if ARCH_X86_64
110
111 # define ROW0 "%%xmm8"
112 # define REG0 ROW0
113 # define ROW2 "%%xmm9"
114 # define REG2 ROW2
115 # define ROW4 "%%xmm10"
116 # define REG4 ROW4
117 # define ROW6 "%%xmm11"
118 # define REG6 ROW6
119 # define CLEAR_EVEN(r) CLEAR_ODD(r)
120 # define PUT_EVEN(dst) PUT_ODD(dst)
121 # define XMMS "%%xmm12"
122 # define MOV_32_ONLY "#"
123 # define SREG2 REG2
124 # define TAN3 "%%xmm13"
125 # define TAN1 "%%xmm14"
126
127 #else
128
129 # define ROW0 "(%0)"
130 # define REG0 "%%xmm4"
131 # define ROW2 "2*16(%0)"
132 # define REG2 "%%xmm4"
133 # define ROW4 "4*16(%0)"
134 # define REG4 "%%xmm6"
135 # define ROW6 "6*16(%0)"
136 # define REG6 "%%xmm6"
137 # define CLEAR_EVEN(r)
138 # define PUT_EVEN(dst) \
139 "pshufhw $0x1B, %%xmm2, %%xmm2 \n\t" \
140 "movdqa %%xmm2, "dst" \n\t"
141 # define XMMS "%%xmm2"
142 # define MOV_32_ONLY "movdqa "
143 # define SREG2 "%%xmm7"
144 # define TAN3 "%%xmm0"
145 # define TAN1 "%%xmm2"
146
147 #endif
148
149 #define ROUND(x) "paddd "MANGLE(x)
150
151 #define JZ(reg, to) \
152 "testl "reg","reg" \n\t" \
153 "jz "to" \n\t"
154
155 #define JNZ(reg, to) \
156 "testl "reg","reg" \n\t" \
157 "jnz "to" \n\t"
158
159 #define TEST_ONE_ROW(src, reg, clear) \
160 clear \
161 "movq "src", %%mm1 \n\t" \
162 "por 8+"src", %%mm1 \n\t" \
163 "paddusb %%mm0, %%mm1 \n\t" \
164 "pmovmskb %%mm1, "reg" \n\t"
165
166 #define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2) \
167 clear1 \
168 clear2 \
169 "movq "row1", %%mm1 \n\t" \
170 "por 8+"row1", %%mm1 \n\t" \
171 "movq "row2", %%mm2 \n\t" \
172 "por 8+"row2", %%mm2 \n\t" \
173 "paddusb %%mm0, %%mm1 \n\t" \
174 "paddusb %%mm0, %%mm2 \n\t" \
175 "pmovmskb %%mm1, "reg1" \n\t" \
176 "pmovmskb %%mm2, "reg2" \n\t"
177
178 /// IDCT pass on rows.
179 #define iMTX_MULT(src, table, rounder, put) \
180 "movdqa "src", %%xmm3 \n\t" \
181 "movdqa %%xmm3, %%xmm0 \n\t" \
182 "pshufd $0x11, %%xmm3, %%xmm1 \n\t" /* 4602 */ \
183 "punpcklqdq %%xmm0, %%xmm0 \n\t" /* 0246 */ \
184 "pmaddwd "table", %%xmm0 \n\t" \
185 "pmaddwd 16+"table", %%xmm1 \n\t" \
186 "pshufd $0xBB, %%xmm3, %%xmm2 \n\t" /* 5713 */ \
187 "punpckhqdq %%xmm3, %%xmm3 \n\t" /* 1357 */ \
188 "pmaddwd 32+"table", %%xmm2 \n\t" \
189 "pmaddwd 48+"table", %%xmm3 \n\t" \
190 "paddd %%xmm1, %%xmm0 \n\t" \
191 "paddd %%xmm3, %%xmm2 \n\t" \
192 rounder", %%xmm0 \n\t" \
193 "movdqa %%xmm2, %%xmm3 \n\t" \
194 "paddd %%xmm0, %%xmm2 \n\t" \
195 "psubd %%xmm3, %%xmm0 \n\t" \
196 "psrad $11, %%xmm2 \n\t" \
197 "psrad $11, %%xmm0 \n\t" \
198 "packssdw %%xmm0, %%xmm2 \n\t" \
199 put \
200 "1: \n\t"
201
202 #define iLLM_HEAD \
203 "movdqa "MANGLE(tan3)", "TAN3" \n\t" \
204 "movdqa "MANGLE(tan1)", "TAN1" \n\t" \
205
206 /// IDCT pass on columns.
207 #define iLLM_PASS(dct) \
208 "movdqa "TAN3", %%xmm1 \n\t" \
209 "movdqa "TAN1", %%xmm3 \n\t" \
210 "pmulhw %%xmm4, "TAN3" \n\t" \
211 "pmulhw %%xmm5, %%xmm1 \n\t" \
212 "paddsw %%xmm4, "TAN3" \n\t" \
213 "paddsw %%xmm5, %%xmm1 \n\t" \
214 "psubsw %%xmm5, "TAN3" \n\t" \
215 "paddsw %%xmm4, %%xmm1 \n\t" \
216 "pmulhw %%xmm7, %%xmm3 \n\t" \
217 "pmulhw %%xmm6, "TAN1" \n\t" \
218 "paddsw %%xmm6, %%xmm3 \n\t" \
219 "psubsw %%xmm7, "TAN1" \n\t" \
220 "movdqa %%xmm3, %%xmm7 \n\t" \
221 "movdqa "TAN1", %%xmm6 \n\t" \
222 "psubsw %%xmm1, %%xmm3 \n\t" \
223 "psubsw "TAN3", "TAN1" \n\t" \
224 "paddsw %%xmm7, %%xmm1 \n\t" \
225 "paddsw %%xmm6, "TAN3" \n\t" \
226 "movdqa %%xmm3, %%xmm6 \n\t" \
227 "psubsw "TAN3", %%xmm3 \n\t" \
228 "paddsw %%xmm6, "TAN3" \n\t" \
229 "movdqa "MANGLE(sqrt2)", %%xmm4 \n\t" \
230 "pmulhw %%xmm4, %%xmm3 \n\t" \
231 "pmulhw %%xmm4, "TAN3" \n\t" \
232 "paddsw "TAN3", "TAN3" \n\t" \
233 "paddsw %%xmm3, %%xmm3 \n\t" \
234 "movdqa "MANGLE(tan2)", %%xmm7 \n\t" \
235 MOV_32_ONLY ROW2", "REG2" \n\t" \
236 MOV_32_ONLY ROW6", "REG6" \n\t" \
237 "movdqa %%xmm7, %%xmm5 \n\t" \
238 "pmulhw "REG6", %%xmm7 \n\t" \
239 "pmulhw "REG2", %%xmm5 \n\t" \
240 "paddsw "REG2", %%xmm7 \n\t" \
241 "psubsw "REG6", %%xmm5 \n\t" \
242 MOV_32_ONLY ROW0", "REG0" \n\t" \
243 MOV_32_ONLY ROW4", "REG4" \n\t" \
244 MOV_32_ONLY" "TAN1", (%0) \n\t" \
245 "movdqa "REG0", "XMMS" \n\t" \
246 "psubsw "REG4", "REG0" \n\t" \
247 "paddsw "XMMS", "REG4" \n\t" \
248 "movdqa "REG4", "XMMS" \n\t" \
249 "psubsw %%xmm7, "REG4" \n\t" \
250 "paddsw "XMMS", %%xmm7 \n\t" \
251 "movdqa "REG0", "XMMS" \n\t" \
252 "psubsw %%xmm5, "REG0" \n\t" \
253 "paddsw "XMMS", %%xmm5 \n\t" \
254 "movdqa %%xmm5, "XMMS" \n\t" \
255 "psubsw "TAN3", %%xmm5 \n\t" \
256 "paddsw "XMMS", "TAN3" \n\t" \
257 "movdqa "REG0", "XMMS" \n\t" \
258 "psubsw %%xmm3, "REG0" \n\t" \
259 "paddsw "XMMS", %%xmm3 \n\t" \
260 MOV_32_ONLY" (%0), "TAN1" \n\t" \
261 "psraw $6, %%xmm5 \n\t" \
262 "psraw $6, "REG0" \n\t" \
263 "psraw $6, "TAN3" \n\t" \
264 "psraw $6, %%xmm3 \n\t" \
265 "movdqa "TAN3", 1*16("dct") \n\t" \
266 "movdqa %%xmm3, 2*16("dct") \n\t" \
267 "movdqa "REG0", 5*16("dct") \n\t" \
268 "movdqa %%xmm5, 6*16("dct") \n\t" \
269 "movdqa %%xmm7, %%xmm0 \n\t" \
270 "movdqa "REG4", %%xmm4 \n\t" \
271 "psubsw %%xmm1, %%xmm7 \n\t" \
272 "psubsw "TAN1", "REG4" \n\t" \
273 "paddsw %%xmm0, %%xmm1 \n\t" \
274 "paddsw %%xmm4, "TAN1" \n\t" \
275 "psraw $6, %%xmm1 \n\t" \
276 "psraw $6, %%xmm7 \n\t" \
277 "psraw $6, "TAN1" \n\t" \
278 "psraw $6, "REG4" \n\t" \
279 "movdqa %%xmm1, ("dct") \n\t" \
280 "movdqa "TAN1", 3*16("dct") \n\t" \
281 "movdqa "REG4", 4*16("dct") \n\t" \
282 "movdqa %%xmm7, 7*16("dct") \n\t"
283
284 /// IDCT pass on columns, assuming rows 4-7 are zero.
285 #define iLLM_PASS_SPARSE(dct) \
286 "pmulhw %%xmm4, "TAN3" \n\t" \
287 "paddsw %%xmm4, "TAN3" \n\t" \
288 "movdqa %%xmm6, %%xmm3 \n\t" \
289 "pmulhw %%xmm6, "TAN1" \n\t" \
290 "movdqa %%xmm4, %%xmm1 \n\t" \
291 "psubsw %%xmm1, %%xmm3 \n\t" \
292 "paddsw %%xmm6, %%xmm1 \n\t" \
293 "movdqa "TAN1", %%xmm6 \n\t" \
294 "psubsw "TAN3", "TAN1" \n\t" \
295 "paddsw %%xmm6, "TAN3" \n\t" \
296 "movdqa %%xmm3, %%xmm6 \n\t" \
297 "psubsw "TAN3", %%xmm3 \n\t" \
298 "paddsw %%xmm6, "TAN3" \n\t" \
299 "movdqa "MANGLE(sqrt2)", %%xmm4 \n\t" \
300 "pmulhw %%xmm4, %%xmm3 \n\t" \
301 "pmulhw %%xmm4, "TAN3" \n\t" \
302 "paddsw "TAN3", "TAN3" \n\t" \
303 "paddsw %%xmm3, %%xmm3 \n\t" \
304 "movdqa "MANGLE(tan2)", %%xmm5 \n\t" \
305 MOV_32_ONLY ROW2", "SREG2" \n\t" \
306 "pmulhw "SREG2", %%xmm5 \n\t" \
307 MOV_32_ONLY ROW0", "REG0" \n\t" \
308 "movdqa "REG0", %%xmm6 \n\t" \
309 "psubsw "SREG2", %%xmm6 \n\t" \
310 "paddsw "REG0", "SREG2" \n\t" \
311 MOV_32_ONLY" "TAN1", (%0) \n\t" \
312 "movdqa "REG0", "XMMS" \n\t" \
313 "psubsw %%xmm5, "REG0" \n\t" \
314 "paddsw "XMMS", %%xmm5 \n\t" \
315 "movdqa %%xmm5, "XMMS" \n\t" \
316 "psubsw "TAN3", %%xmm5 \n\t" \
317 "paddsw "XMMS", "TAN3" \n\t" \
318 "movdqa "REG0", "XMMS" \n\t" \
319 "psubsw %%xmm3, "REG0" \n\t" \
320 "paddsw "XMMS", %%xmm3 \n\t" \
321 MOV_32_ONLY" (%0), "TAN1" \n\t" \
322 "psraw $6, %%xmm5 \n\t" \
323 "psraw $6, "REG0" \n\t" \
324 "psraw $6, "TAN3" \n\t" \
325 "psraw $6, %%xmm3 \n\t" \
326 "movdqa "TAN3", 1*16("dct") \n\t" \
327 "movdqa %%xmm3, 2*16("dct") \n\t" \
328 "movdqa "REG0", 5*16("dct") \n\t" \
329 "movdqa %%xmm5, 6*16("dct") \n\t" \
330 "movdqa "SREG2", %%xmm0 \n\t" \
331 "movdqa %%xmm6, %%xmm4 \n\t" \
332 "psubsw %%xmm1, "SREG2" \n\t" \
333 "psubsw "TAN1", %%xmm6 \n\t" \
334 "paddsw %%xmm0, %%xmm1 \n\t" \
335 "paddsw %%xmm4, "TAN1" \n\t" \
336 "psraw $6, %%xmm1 \n\t" \
337 "psraw $6, "SREG2" \n\t" \
338 "psraw $6, "TAN1" \n\t" \
339 "psraw $6, %%xmm6 \n\t" \
340 "movdqa %%xmm1, ("dct") \n\t" \
341 "movdqa "TAN1", 3*16("dct") \n\t" \
342 "movdqa %%xmm6, 4*16("dct") \n\t" \
343 "movdqa "SREG2", 7*16("dct") \n\t"
344
345 inline void ff_xvid_idct_sse2(short *block)
346 {
347 __asm__ volatile (
348 "movq "MANGLE (m127) ", %%mm0 \n\t"
349 iMTX_MULT("(%0)", MANGLE(iTab1), ROUND(walkenIdctRounders), PUT_EVEN(ROW0))
350 iMTX_MULT("1*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders + 1 * 16), PUT_ODD(ROW1))
351 iMTX_MULT("2*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders + 2 * 16), PUT_EVEN(ROW2))
352
353 TEST_TWO_ROWS("3*16(%0)", "4*16(%0)", "%%eax", "%%ecx", CLEAR_ODD(ROW3), CLEAR_EVEN(ROW4))
354 JZ("%%eax", "1f")
355 iMTX_MULT("3*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders + 3 * 16), PUT_ODD(ROW3))
356
357 TEST_TWO_ROWS("5*16(%0)", "6*16(%0)", "%%eax", "%%edx", CLEAR_ODD(ROW5), CLEAR_EVEN(ROW6))
358 TEST_ONE_ROW("7*16(%0)", "%%esi", CLEAR_ODD(ROW7))
359 iLLM_HEAD
360 ".p2align 4 \n\t"
361 JNZ("%%ecx", "2f")
362 JNZ("%%eax", "3f")
363 JNZ("%%edx", "4f")
364 JNZ("%%esi", "5f")
365 iLLM_PASS_SPARSE("%0")
366 "jmp 6f \n\t"
367 "2: \n\t"
368 iMTX_MULT("4*16(%0)", MANGLE(iTab1), "#", PUT_EVEN(ROW4))
369 "3: \n\t"
370 iMTX_MULT("5*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders + 4 * 16), PUT_ODD(ROW5))
371 JZ("%%edx", "1f")
372 "4: \n\t"
373 iMTX_MULT("6*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders + 5 * 16), PUT_EVEN(ROW6))
374 JZ("%%esi", "1f")
375 "5: \n\t"
376 iMTX_MULT("7*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders + 5 * 16), PUT_ODD(ROW7))
377 #if ARCH_X86_32
378 iLLM_HEAD
379 #endif
380 iLLM_PASS("%0")
381 "6: \n\t"
382 : "+r" (block)
383 :
384 : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3",
385 "%xmm4", "%xmm5", "%xmm6", "%xmm7", )
386 #if ARCH_X86_64
387 XMM_CLOBBERS("%xmm8", "%xmm9", "%xmm10", "%xmm11",
388 "%xmm12", "%xmm13", "%xmm14", )
389 #endif
390 "%eax", "%ecx", "%edx", "%esi", "memory");
391 }
392
393 void ff_xvid_idct_sse2_put(uint8_t *dest, ptrdiff_t line_size, short *block)
394 {
395 ff_xvid_idct_sse2(block);
396 ff_put_pixels_clamped_mmx(block, dest, line_size);
397 }
398
399 void ff_xvid_idct_sse2_add(uint8_t *dest, ptrdiff_t line_size, short *block)
400 {
401 ff_xvid_idct_sse2(block);
402 ff_add_pixels_clamped_mmx(block, dest, line_size);
403 }
404
405 #endif /* HAVE_SSE2_INLINE */