Merge remote-tracking branch 'qatar/master'
[ffmpeg.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 FFmpeg.
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  * FFmpeg 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  * FFmpeg 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 FFmpeg; if not, write to the Free Software Foundation,
38  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
39  */
40
41 #include "libavcodec/dsputil.h"
42 #include "libavutil/x86_cpu.h"
43 #include "idct_xvid.h"
44 #include "dsputil_mmx.h"
45
46 /**
47  * @file
48  * @brief SSE2 idct compatible with xvidmmx
49  */
50
51 #define X8(x)     x,x,x,x,x,x,x,x
52
53 #define ROW_SHIFT 11
54 #define COL_SHIFT 6
55
56 DECLARE_ASM_CONST(16, int16_t, tan1)[] = {X8(13036)}; // tan( pi/16)
57 DECLARE_ASM_CONST(16, int16_t, tan2)[] = {X8(27146)}; // tan(2pi/16) = sqrt(2)-1
58 DECLARE_ASM_CONST(16, int16_t, tan3)[] = {X8(43790)}; // tan(3pi/16)-1
59 DECLARE_ASM_CONST(16, int16_t, sqrt2)[]= {X8(23170)}; // 0.5/sqrt(2)
60 DECLARE_ASM_CONST(8,  uint8_t, m127)[] = {X8(127)};
61
62 DECLARE_ASM_CONST(16, int16_t, iTab1)[] = {
63  0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d,
64  0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61,
65  0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7,
66  0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b
67 };
68
69 DECLARE_ASM_CONST(16, int16_t, iTab2)[] = {
70  0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5,
71  0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04,
72  0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41,
73  0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df
74 };
75
76 DECLARE_ASM_CONST(16, int16_t, iTab3)[] = {
77  0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf,
78  0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf,
79  0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d,
80  0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04
81 };
82
83 DECLARE_ASM_CONST(16, int16_t, iTab4)[] = {
84  0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746,
85  0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac,
86  0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df,
87  0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e
88 };
89
90 DECLARE_ASM_CONST(16, int32_t, walkenIdctRounders)[] = {
91  65536, 65536, 65536, 65536,
92   3597,  3597,  3597,  3597,
93   2260,  2260,  2260,  2260,
94   1203,  1203,  1203,  1203,
95    120,   120,   120,   120,
96    512,   512,   512,   512
97 };
98
99 // Temporary storage before the column pass
100 #define ROW1 "%%xmm6"
101 #define ROW3 "%%xmm4"
102 #define ROW5 "%%xmm5"
103 #define ROW7 "%%xmm7"
104
105 #define CLEAR_ODD(r) "pxor  "r","r" \n\t"
106 #define PUT_ODD(dst) "pshufhw  $0x1B, %%xmm2, "dst"   \n\t"
107
108 #if ARCH_X86_64
109
110 # define ROW0 "%%xmm8"
111 # define REG0 ROW0
112 # define ROW2 "%%xmm9"
113 # define REG2 ROW2
114 # define ROW4 "%%xmm10"
115 # define REG4 ROW4
116 # define ROW6 "%%xmm11"
117 # define REG6 ROW6
118 # define CLEAR_EVEN(r) CLEAR_ODD(r)
119 # define PUT_EVEN(dst) PUT_ODD(dst)
120 # define XMMS "%%xmm12"
121 # define MOV_32_ONLY "#"
122 # define SREG2 REG2
123 # define TAN3 "%%xmm13"
124 # define TAN1 "%%xmm14"
125
126 #else
127
128 # define ROW0 "(%0)"
129 # define REG0 "%%xmm4"
130 # define ROW2 "2*16(%0)"
131 # define REG2 "%%xmm4"
132 # define ROW4 "4*16(%0)"
133 # define REG4 "%%xmm6"
134 # define ROW6 "6*16(%0)"
135 # define REG6 "%%xmm6"
136 # define CLEAR_EVEN(r)
137 # define PUT_EVEN(dst) \
138     "pshufhw  $0x1B, %%xmm2, %%xmm2   \n\t" \
139     "movdqa          %%xmm2, "dst"    \n\t"
140 # define XMMS "%%xmm2"
141 # define MOV_32_ONLY "movdqa "
142 # define SREG2 "%%xmm7"
143 # define TAN3 "%%xmm0"
144 # define TAN1 "%%xmm2"
145
146 #endif
147
148 #define ROUND(x) "paddd   "MANGLE(x)
149
150 #define JZ(reg, to)                         \
151     "testl     "reg","reg"            \n\t" \
152     "jz        "to"                   \n\t"
153
154 #define JNZ(reg, to)                        \
155     "testl     "reg","reg"            \n\t" \
156     "jnz       "to"                   \n\t"
157
158 #define TEST_ONE_ROW(src, reg, clear)       \
159     clear                                   \
160     "movq     "src", %%mm1            \n\t" \
161     "por    8+"src", %%mm1            \n\t" \
162     "paddusb  %%mm0, %%mm1            \n\t" \
163     "pmovmskb %%mm1, "reg"            \n\t"
164
165 #define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2) \
166     clear1                                  \
167     clear2                                  \
168     "movq     "row1", %%mm1           \n\t" \
169     "por    8+"row1", %%mm1           \n\t" \
170     "movq     "row2", %%mm2           \n\t" \
171     "por    8+"row2", %%mm2           \n\t" \
172     "paddusb   %%mm0, %%mm1           \n\t" \
173     "paddusb   %%mm0, %%mm2           \n\t" \
174     "pmovmskb  %%mm1, "reg1"          \n\t" \
175     "pmovmskb  %%mm2, "reg2"          \n\t"
176
177 ///IDCT pass on rows.
178 #define iMTX_MULT(src, table, rounder, put) \
179     "movdqa        "src", %%xmm3      \n\t" \
180     "movdqa       %%xmm3, %%xmm0      \n\t" \
181     "pshufd   $0x11, %%xmm3, %%xmm1   \n\t" /* 4602 */ \
182     "punpcklqdq   %%xmm0, %%xmm0      \n\t" /* 0246 */ \
183     "pmaddwd     "table", %%xmm0      \n\t" \
184     "pmaddwd  16+"table", %%xmm1      \n\t" \
185     "pshufd   $0xBB, %%xmm3, %%xmm2   \n\t" /* 5713 */ \
186     "punpckhqdq   %%xmm3, %%xmm3      \n\t" /* 1357 */ \
187     "pmaddwd  32+"table", %%xmm2      \n\t" \
188     "pmaddwd  48+"table", %%xmm3      \n\t" \
189     "paddd        %%xmm1, %%xmm0      \n\t" \
190     "paddd        %%xmm3, %%xmm2      \n\t" \
191     rounder",     %%xmm0              \n\t" \
192     "movdqa       %%xmm2, %%xmm3      \n\t" \
193     "paddd        %%xmm0, %%xmm2      \n\t" \
194     "psubd        %%xmm3, %%xmm0      \n\t" \
195     "psrad           $11, %%xmm2      \n\t" \
196     "psrad           $11, %%xmm0      \n\t" \
197     "packssdw     %%xmm0, %%xmm2      \n\t" \
198     put                                     \
199     "1:                               \n\t"
200
201 #define iLLM_HEAD                           \
202     "movdqa   "MANGLE(tan3)", "TAN3"  \n\t" \
203     "movdqa   "MANGLE(tan1)", "TAN1"  \n\t" \
204
205 ///IDCT pass on columns.
206 #define iLLM_PASS(dct)                      \
207     "movdqa   "TAN3", %%xmm1          \n\t" \
208     "movdqa   "TAN1", %%xmm3          \n\t" \
209     "pmulhw   %%xmm4, "TAN3"          \n\t" \
210     "pmulhw   %%xmm5, %%xmm1          \n\t" \
211     "paddsw   %%xmm4, "TAN3"          \n\t" \
212     "paddsw   %%xmm5, %%xmm1          \n\t" \
213     "psubsw   %%xmm5, "TAN3"          \n\t" \
214     "paddsw   %%xmm4, %%xmm1          \n\t" \
215     "pmulhw   %%xmm7, %%xmm3          \n\t" \
216     "pmulhw   %%xmm6, "TAN1"          \n\t" \
217     "paddsw   %%xmm6, %%xmm3          \n\t" \
218     "psubsw   %%xmm7, "TAN1"          \n\t" \
219     "movdqa   %%xmm3, %%xmm7          \n\t" \
220     "movdqa   "TAN1", %%xmm6          \n\t" \
221     "psubsw   %%xmm1, %%xmm3          \n\t" \
222     "psubsw   "TAN3", "TAN1"          \n\t" \
223     "paddsw   %%xmm7, %%xmm1          \n\t" \
224     "paddsw   %%xmm6, "TAN3"          \n\t" \
225     "movdqa   %%xmm3, %%xmm6          \n\t" \
226     "psubsw   "TAN3", %%xmm3          \n\t" \
227     "paddsw   %%xmm6, "TAN3"          \n\t" \
228     "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
229     "pmulhw   %%xmm4, %%xmm3          \n\t" \
230     "pmulhw   %%xmm4, "TAN3"          \n\t" \
231     "paddsw   "TAN3", "TAN3"          \n\t" \
232     "paddsw   %%xmm3, %%xmm3          \n\t" \
233     "movdqa   "MANGLE(tan2)", %%xmm7  \n\t" \
234     MOV_32_ONLY ROW2", "REG2"         \n\t" \
235     MOV_32_ONLY ROW6", "REG6"         \n\t" \
236     "movdqa   %%xmm7, %%xmm5          \n\t" \
237     "pmulhw   "REG6", %%xmm7          \n\t" \
238     "pmulhw   "REG2", %%xmm5          \n\t" \
239     "paddsw   "REG2", %%xmm7          \n\t" \
240     "psubsw   "REG6", %%xmm5          \n\t" \
241     MOV_32_ONLY ROW0", "REG0"         \n\t" \
242     MOV_32_ONLY ROW4", "REG4"         \n\t" \
243     MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
244     "movdqa   "REG0", "XMMS"          \n\t" \
245     "psubsw   "REG4", "REG0"          \n\t" \
246     "paddsw   "XMMS", "REG4"          \n\t" \
247     "movdqa   "REG4", "XMMS"          \n\t" \
248     "psubsw   %%xmm7, "REG4"          \n\t" \
249     "paddsw   "XMMS", %%xmm7          \n\t" \
250     "movdqa   "REG0", "XMMS"          \n\t" \
251     "psubsw   %%xmm5, "REG0"          \n\t" \
252     "paddsw   "XMMS", %%xmm5          \n\t" \
253     "movdqa   %%xmm5, "XMMS"          \n\t" \
254     "psubsw   "TAN3", %%xmm5          \n\t" \
255     "paddsw   "XMMS", "TAN3"          \n\t" \
256     "movdqa   "REG0", "XMMS"          \n\t" \
257     "psubsw   %%xmm3, "REG0"          \n\t" \
258     "paddsw   "XMMS", %%xmm3          \n\t" \
259     MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
260     "psraw        $6, %%xmm5          \n\t" \
261     "psraw        $6, "REG0"          \n\t" \
262     "psraw        $6, "TAN3"          \n\t" \
263     "psraw        $6, %%xmm3          \n\t" \
264     "movdqa   "TAN3", 1*16("dct")     \n\t" \
265     "movdqa   %%xmm3, 2*16("dct")     \n\t" \
266     "movdqa   "REG0", 5*16("dct")     \n\t" \
267     "movdqa   %%xmm5, 6*16("dct")     \n\t" \
268     "movdqa   %%xmm7, %%xmm0          \n\t" \
269     "movdqa   "REG4", %%xmm4          \n\t" \
270     "psubsw   %%xmm1, %%xmm7          \n\t" \
271     "psubsw   "TAN1", "REG4"          \n\t" \
272     "paddsw   %%xmm0, %%xmm1          \n\t" \
273     "paddsw   %%xmm4, "TAN1"          \n\t" \
274     "psraw        $6, %%xmm1          \n\t" \
275     "psraw        $6, %%xmm7          \n\t" \
276     "psraw        $6, "TAN1"          \n\t" \
277     "psraw        $6, "REG4"          \n\t" \
278     "movdqa   %%xmm1, ("dct")         \n\t" \
279     "movdqa   "TAN1", 3*16("dct")     \n\t" \
280     "movdqa   "REG4", 4*16("dct")     \n\t" \
281     "movdqa   %%xmm7, 7*16("dct")     \n\t"
282
283 ///IDCT pass on columns, assuming rows 4-7 are zero.
284 #define iLLM_PASS_SPARSE(dct)               \
285     "pmulhw   %%xmm4, "TAN3"          \n\t" \
286     "paddsw   %%xmm4, "TAN3"          \n\t" \
287     "movdqa   %%xmm6, %%xmm3          \n\t" \
288     "pmulhw   %%xmm6, "TAN1"          \n\t" \
289     "movdqa   %%xmm4, %%xmm1          \n\t" \
290     "psubsw   %%xmm1, %%xmm3          \n\t" \
291     "paddsw   %%xmm6, %%xmm1          \n\t" \
292     "movdqa   "TAN1", %%xmm6          \n\t" \
293     "psubsw   "TAN3", "TAN1"          \n\t" \
294     "paddsw   %%xmm6, "TAN3"          \n\t" \
295     "movdqa   %%xmm3, %%xmm6          \n\t" \
296     "psubsw   "TAN3", %%xmm3          \n\t" \
297     "paddsw   %%xmm6, "TAN3"          \n\t" \
298     "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
299     "pmulhw   %%xmm4, %%xmm3          \n\t" \
300     "pmulhw   %%xmm4, "TAN3"          \n\t" \
301     "paddsw   "TAN3", "TAN3"          \n\t" \
302     "paddsw   %%xmm3, %%xmm3          \n\t" \
303     "movdqa   "MANGLE(tan2)", %%xmm5  \n\t" \
304     MOV_32_ONLY ROW2", "SREG2"        \n\t" \
305     "pmulhw   "SREG2", %%xmm5         \n\t" \
306     MOV_32_ONLY ROW0", "REG0"         \n\t" \
307     "movdqa   "REG0", %%xmm6          \n\t" \
308     "psubsw   "SREG2", %%xmm6         \n\t" \
309     "paddsw   "REG0", "SREG2"         \n\t" \
310     MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
311     "movdqa   "REG0", "XMMS"          \n\t" \
312     "psubsw   %%xmm5, "REG0"          \n\t" \
313     "paddsw   "XMMS", %%xmm5          \n\t" \
314     "movdqa   %%xmm5, "XMMS"          \n\t" \
315     "psubsw   "TAN3", %%xmm5          \n\t" \
316     "paddsw   "XMMS", "TAN3"          \n\t" \
317     "movdqa   "REG0", "XMMS"          \n\t" \
318     "psubsw   %%xmm3, "REG0"          \n\t" \
319     "paddsw   "XMMS", %%xmm3          \n\t" \
320     MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
321     "psraw        $6, %%xmm5          \n\t" \
322     "psraw        $6, "REG0"          \n\t" \
323     "psraw        $6, "TAN3"          \n\t" \
324     "psraw        $6, %%xmm3          \n\t" \
325     "movdqa   "TAN3", 1*16("dct")     \n\t" \
326     "movdqa   %%xmm3, 2*16("dct")     \n\t" \
327     "movdqa   "REG0", 5*16("dct")     \n\t" \
328     "movdqa   %%xmm5, 6*16("dct")     \n\t" \
329     "movdqa   "SREG2", %%xmm0         \n\t" \
330     "movdqa   %%xmm6, %%xmm4          \n\t" \
331     "psubsw   %%xmm1, "SREG2"         \n\t" \
332     "psubsw   "TAN1", %%xmm6          \n\t" \
333     "paddsw   %%xmm0, %%xmm1          \n\t" \
334     "paddsw   %%xmm4, "TAN1"          \n\t" \
335     "psraw        $6, %%xmm1          \n\t" \
336     "psraw        $6, "SREG2"         \n\t" \
337     "psraw        $6, "TAN1"          \n\t" \
338     "psraw        $6, %%xmm6          \n\t" \
339     "movdqa   %%xmm1, ("dct")         \n\t" \
340     "movdqa   "TAN1", 3*16("dct")     \n\t" \
341     "movdqa   %%xmm6, 4*16("dct")     \n\t" \
342     "movdqa   "SREG2", 7*16("dct")    \n\t"
343
344 inline void ff_idct_xvid_sse2(short *block)
345 {
346     __asm__ volatile(
347     "movq     "MANGLE(m127)", %%mm0                              \n\t"
348     iMTX_MULT("(%0)",     MANGLE(iTab1), ROUND(walkenIdctRounders),      PUT_EVEN(ROW0))
349     iMTX_MULT("1*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+1*16), PUT_ODD(ROW1))
350     iMTX_MULT("2*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+2*16), PUT_EVEN(ROW2))
351
352     TEST_TWO_ROWS("3*16(%0)", "4*16(%0)", "%%eax", "%%ecx", CLEAR_ODD(ROW3), CLEAR_EVEN(ROW4))
353     JZ("%%eax", "1f")
354     iMTX_MULT("3*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+3*16), PUT_ODD(ROW3))
355
356     TEST_TWO_ROWS("5*16(%0)", "6*16(%0)", "%%eax", "%%edx", CLEAR_ODD(ROW5), CLEAR_EVEN(ROW6))
357     TEST_ONE_ROW("7*16(%0)", "%%esi", CLEAR_ODD(ROW7))
358     iLLM_HEAD
359     ".p2align 4 \n\t"
360     JNZ("%%ecx", "2f")
361     JNZ("%%eax", "3f")
362     JNZ("%%edx", "4f")
363     JNZ("%%esi", "5f")
364     iLLM_PASS_SPARSE("%0")
365     "jmp 6f                                                      \n\t"
366     "2:                                                          \n\t"
367     iMTX_MULT("4*16(%0)", MANGLE(iTab1), "#", PUT_EVEN(ROW4))
368     "3:                                                          \n\t"
369     iMTX_MULT("5*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+4*16), PUT_ODD(ROW5))
370     JZ("%%edx", "1f")
371     "4:                                                          \n\t"
372     iMTX_MULT("6*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+5*16), PUT_EVEN(ROW6))
373     JZ("%%esi", "1f")
374     "5:                                                          \n\t"
375     iMTX_MULT("7*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+5*16), PUT_ODD(ROW7))
376 #if !ARCH_X86_64
377     iLLM_HEAD
378 #endif
379     iLLM_PASS("%0")
380     "6:                                                          \n\t"
381     : "+r"(block)
382     :
383     : XMM_CLOBBERS("%xmm0" , "%xmm1" , "%xmm2" , "%xmm3" ,
384                    "%xmm4" , "%xmm5" , "%xmm6" , "%xmm7" ,)
385 #if ARCH_X86_64
386       XMM_CLOBBERS("%xmm8" , "%xmm9" , "%xmm10", "%xmm11",
387                    "%xmm12", "%xmm13", "%xmm14",)
388 #endif
389       "%eax", "%ecx", "%edx", "%esi", "memory"
390     );
391 }
392
393 void ff_idct_xvid_sse2_put(uint8_t *dest, int line_size, short *block)
394 {
395     ff_idct_xvid_sse2(block);
396     ff_put_pixels_clamped_mmx(block, dest, line_size);
397 }
398
399 void ff_idct_xvid_sse2_add(uint8_t *dest, int line_size, short *block)
400 {
401     ff_idct_xvid_sse2(block);
402     ff_add_pixels_clamped_mmx(block, dest, line_size);
403 }