x86: allow using add_hfyu_median_prediction_cmov on any cpu with cmov
[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 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 "libavcodec/dsputil.h"
42 #include "libavutil/internal.h"
43 #include "libavutil/mem.h"
44 #include "libavutil/x86/asm.h"
45 #include "idct_xvid.h"
46 #include "dsputil_mmx.h"
47
48 #if HAVE_INLINE_ASM
49
50 /**
51  * @file
52  * @brief SSE2 idct compatible with xvidmmx
53  */
54
55 #define X8(x)     x,x,x,x,x,x,x,x
56
57 #define ROW_SHIFT 11
58 #define COL_SHIFT 6
59
60 DECLARE_ASM_CONST(16, int16_t, tan1)[] = {X8(13036)}; // tan( pi/16)
61 DECLARE_ASM_CONST(16, int16_t, tan2)[] = {X8(27146)}; // tan(2pi/16) = sqrt(2)-1
62 DECLARE_ASM_CONST(16, int16_t, tan3)[] = {X8(43790)}; // tan(3pi/16)-1
63 DECLARE_ASM_CONST(16, int16_t, sqrt2)[]= {X8(23170)}; // 0.5/sqrt(2)
64 DECLARE_ASM_CONST(8,  uint8_t, m127)[] = {X8(127)};
65
66 DECLARE_ASM_CONST(16, int16_t, iTab1)[] = {
67  0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d,
68  0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61,
69  0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7,
70  0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b
71 };
72
73 DECLARE_ASM_CONST(16, int16_t, iTab2)[] = {
74  0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5,
75  0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04,
76  0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41,
77  0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df
78 };
79
80 DECLARE_ASM_CONST(16, int16_t, iTab3)[] = {
81  0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf,
82  0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf,
83  0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d,
84  0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04
85 };
86
87 DECLARE_ASM_CONST(16, int16_t, iTab4)[] = {
88  0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746,
89  0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac,
90  0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df,
91  0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e
92 };
93
94 DECLARE_ASM_CONST(16, int32_t, walkenIdctRounders)[] = {
95  65536, 65536, 65536, 65536,
96   3597,  3597,  3597,  3597,
97   2260,  2260,  2260,  2260,
98   1203,  1203,  1203,  1203,
99    120,   120,   120,   120,
100    512,   512,   512,   512
101 };
102
103 // Temporary storage before the column pass
104 #define ROW1 "%%xmm6"
105 #define ROW3 "%%xmm4"
106 #define ROW5 "%%xmm5"
107 #define ROW7 "%%xmm7"
108
109 #define CLEAR_ODD(r) "pxor  "r","r" \n\t"
110 #define PUT_ODD(dst) "pshufhw  $0x1B, %%xmm2, "dst"   \n\t"
111
112 #if ARCH_X86_64
113
114 # define ROW0 "%%xmm8"
115 # define REG0 ROW0
116 # define ROW2 "%%xmm9"
117 # define REG2 ROW2
118 # define ROW4 "%%xmm10"
119 # define REG4 ROW4
120 # define ROW6 "%%xmm11"
121 # define REG6 ROW6
122 # define CLEAR_EVEN(r) CLEAR_ODD(r)
123 # define PUT_EVEN(dst) PUT_ODD(dst)
124 # define XMMS "%%xmm12"
125 # define MOV_32_ONLY "#"
126 # define SREG2 REG2
127 # define TAN3 "%%xmm13"
128 # define TAN1 "%%xmm14"
129
130 #else
131
132 # define ROW0 "(%0)"
133 # define REG0 "%%xmm4"
134 # define ROW2 "2*16(%0)"
135 # define REG2 "%%xmm4"
136 # define ROW4 "4*16(%0)"
137 # define REG4 "%%xmm6"
138 # define ROW6 "6*16(%0)"
139 # define REG6 "%%xmm6"
140 # define CLEAR_EVEN(r)
141 # define PUT_EVEN(dst) \
142     "pshufhw  $0x1B, %%xmm2, %%xmm2   \n\t" \
143     "movdqa          %%xmm2, "dst"    \n\t"
144 # define XMMS "%%xmm2"
145 # define MOV_32_ONLY "movdqa "
146 # define SREG2 "%%xmm7"
147 # define TAN3 "%%xmm0"
148 # define TAN1 "%%xmm2"
149
150 #endif
151
152 #define ROUND(x) "paddd   "MANGLE(x)
153
154 #define JZ(reg, to)                         \
155     "testl     "reg","reg"            \n\t" \
156     "jz        "to"                   \n\t"
157
158 #define JNZ(reg, to)                        \
159     "testl     "reg","reg"            \n\t" \
160     "jnz       "to"                   \n\t"
161
162 #define TEST_ONE_ROW(src, reg, clear)       \
163     clear                                   \
164     "movq     "src", %%mm1            \n\t" \
165     "por    8+"src", %%mm1            \n\t" \
166     "paddusb  %%mm0, %%mm1            \n\t" \
167     "pmovmskb %%mm1, "reg"            \n\t"
168
169 #define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2) \
170     clear1                                  \
171     clear2                                  \
172     "movq     "row1", %%mm1           \n\t" \
173     "por    8+"row1", %%mm1           \n\t" \
174     "movq     "row2", %%mm2           \n\t" \
175     "por    8+"row2", %%mm2           \n\t" \
176     "paddusb   %%mm0, %%mm1           \n\t" \
177     "paddusb   %%mm0, %%mm2           \n\t" \
178     "pmovmskb  %%mm1, "reg1"          \n\t" \
179     "pmovmskb  %%mm2, "reg2"          \n\t"
180
181 ///IDCT pass on rows.
182 #define iMTX_MULT(src, table, rounder, put) \
183     "movdqa        "src", %%xmm3      \n\t" \
184     "movdqa       %%xmm3, %%xmm0      \n\t" \
185     "pshufd   $0x11, %%xmm3, %%xmm1   \n\t" /* 4602 */ \
186     "punpcklqdq   %%xmm0, %%xmm0      \n\t" /* 0246 */ \
187     "pmaddwd     "table", %%xmm0      \n\t" \
188     "pmaddwd  16+"table", %%xmm1      \n\t" \
189     "pshufd   $0xBB, %%xmm3, %%xmm2   \n\t" /* 5713 */ \
190     "punpckhqdq   %%xmm3, %%xmm3      \n\t" /* 1357 */ \
191     "pmaddwd  32+"table", %%xmm2      \n\t" \
192     "pmaddwd  48+"table", %%xmm3      \n\t" \
193     "paddd        %%xmm1, %%xmm0      \n\t" \
194     "paddd        %%xmm3, %%xmm2      \n\t" \
195     rounder",     %%xmm0              \n\t" \
196     "movdqa       %%xmm2, %%xmm3      \n\t" \
197     "paddd        %%xmm0, %%xmm2      \n\t" \
198     "psubd        %%xmm3, %%xmm0      \n\t" \
199     "psrad           $11, %%xmm2      \n\t" \
200     "psrad           $11, %%xmm0      \n\t" \
201     "packssdw     %%xmm0, %%xmm2      \n\t" \
202     put                                     \
203     "1:                               \n\t"
204
205 #define iLLM_HEAD                           \
206     "movdqa   "MANGLE(tan3)", "TAN3"  \n\t" \
207     "movdqa   "MANGLE(tan1)", "TAN1"  \n\t" \
208
209 ///IDCT pass on columns.
210 #define iLLM_PASS(dct)                      \
211     "movdqa   "TAN3", %%xmm1          \n\t" \
212     "movdqa   "TAN1", %%xmm3          \n\t" \
213     "pmulhw   %%xmm4, "TAN3"          \n\t" \
214     "pmulhw   %%xmm5, %%xmm1          \n\t" \
215     "paddsw   %%xmm4, "TAN3"          \n\t" \
216     "paddsw   %%xmm5, %%xmm1          \n\t" \
217     "psubsw   %%xmm5, "TAN3"          \n\t" \
218     "paddsw   %%xmm4, %%xmm1          \n\t" \
219     "pmulhw   %%xmm7, %%xmm3          \n\t" \
220     "pmulhw   %%xmm6, "TAN1"          \n\t" \
221     "paddsw   %%xmm6, %%xmm3          \n\t" \
222     "psubsw   %%xmm7, "TAN1"          \n\t" \
223     "movdqa   %%xmm3, %%xmm7          \n\t" \
224     "movdqa   "TAN1", %%xmm6          \n\t" \
225     "psubsw   %%xmm1, %%xmm3          \n\t" \
226     "psubsw   "TAN3", "TAN1"          \n\t" \
227     "paddsw   %%xmm7, %%xmm1          \n\t" \
228     "paddsw   %%xmm6, "TAN3"          \n\t" \
229     "movdqa   %%xmm3, %%xmm6          \n\t" \
230     "psubsw   "TAN3", %%xmm3          \n\t" \
231     "paddsw   %%xmm6, "TAN3"          \n\t" \
232     "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
233     "pmulhw   %%xmm4, %%xmm3          \n\t" \
234     "pmulhw   %%xmm4, "TAN3"          \n\t" \
235     "paddsw   "TAN3", "TAN3"          \n\t" \
236     "paddsw   %%xmm3, %%xmm3          \n\t" \
237     "movdqa   "MANGLE(tan2)", %%xmm7  \n\t" \
238     MOV_32_ONLY ROW2", "REG2"         \n\t" \
239     MOV_32_ONLY ROW6", "REG6"         \n\t" \
240     "movdqa   %%xmm7, %%xmm5          \n\t" \
241     "pmulhw   "REG6", %%xmm7          \n\t" \
242     "pmulhw   "REG2", %%xmm5          \n\t" \
243     "paddsw   "REG2", %%xmm7          \n\t" \
244     "psubsw   "REG6", %%xmm5          \n\t" \
245     MOV_32_ONLY ROW0", "REG0"         \n\t" \
246     MOV_32_ONLY ROW4", "REG4"         \n\t" \
247     MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
248     "movdqa   "REG0", "XMMS"          \n\t" \
249     "psubsw   "REG4", "REG0"          \n\t" \
250     "paddsw   "XMMS", "REG4"          \n\t" \
251     "movdqa   "REG4", "XMMS"          \n\t" \
252     "psubsw   %%xmm7, "REG4"          \n\t" \
253     "paddsw   "XMMS", %%xmm7          \n\t" \
254     "movdqa   "REG0", "XMMS"          \n\t" \
255     "psubsw   %%xmm5, "REG0"          \n\t" \
256     "paddsw   "XMMS", %%xmm5          \n\t" \
257     "movdqa   %%xmm5, "XMMS"          \n\t" \
258     "psubsw   "TAN3", %%xmm5          \n\t" \
259     "paddsw   "XMMS", "TAN3"          \n\t" \
260     "movdqa   "REG0", "XMMS"          \n\t" \
261     "psubsw   %%xmm3, "REG0"          \n\t" \
262     "paddsw   "XMMS", %%xmm3          \n\t" \
263     MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
264     "psraw        $6, %%xmm5          \n\t" \
265     "psraw        $6, "REG0"          \n\t" \
266     "psraw        $6, "TAN3"          \n\t" \
267     "psraw        $6, %%xmm3          \n\t" \
268     "movdqa   "TAN3", 1*16("dct")     \n\t" \
269     "movdqa   %%xmm3, 2*16("dct")     \n\t" \
270     "movdqa   "REG0", 5*16("dct")     \n\t" \
271     "movdqa   %%xmm5, 6*16("dct")     \n\t" \
272     "movdqa   %%xmm7, %%xmm0          \n\t" \
273     "movdqa   "REG4", %%xmm4          \n\t" \
274     "psubsw   %%xmm1, %%xmm7          \n\t" \
275     "psubsw   "TAN1", "REG4"          \n\t" \
276     "paddsw   %%xmm0, %%xmm1          \n\t" \
277     "paddsw   %%xmm4, "TAN1"          \n\t" \
278     "psraw        $6, %%xmm1          \n\t" \
279     "psraw        $6, %%xmm7          \n\t" \
280     "psraw        $6, "TAN1"          \n\t" \
281     "psraw        $6, "REG4"          \n\t" \
282     "movdqa   %%xmm1, ("dct")         \n\t" \
283     "movdqa   "TAN1", 3*16("dct")     \n\t" \
284     "movdqa   "REG4", 4*16("dct")     \n\t" \
285     "movdqa   %%xmm7, 7*16("dct")     \n\t"
286
287 ///IDCT pass on columns, assuming rows 4-7 are zero.
288 #define iLLM_PASS_SPARSE(dct)               \
289     "pmulhw   %%xmm4, "TAN3"          \n\t" \
290     "paddsw   %%xmm4, "TAN3"          \n\t" \
291     "movdqa   %%xmm6, %%xmm3          \n\t" \
292     "pmulhw   %%xmm6, "TAN1"          \n\t" \
293     "movdqa   %%xmm4, %%xmm1          \n\t" \
294     "psubsw   %%xmm1, %%xmm3          \n\t" \
295     "paddsw   %%xmm6, %%xmm1          \n\t" \
296     "movdqa   "TAN1", %%xmm6          \n\t" \
297     "psubsw   "TAN3", "TAN1"          \n\t" \
298     "paddsw   %%xmm6, "TAN3"          \n\t" \
299     "movdqa   %%xmm3, %%xmm6          \n\t" \
300     "psubsw   "TAN3", %%xmm3          \n\t" \
301     "paddsw   %%xmm6, "TAN3"          \n\t" \
302     "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
303     "pmulhw   %%xmm4, %%xmm3          \n\t" \
304     "pmulhw   %%xmm4, "TAN3"          \n\t" \
305     "paddsw   "TAN3", "TAN3"          \n\t" \
306     "paddsw   %%xmm3, %%xmm3          \n\t" \
307     "movdqa   "MANGLE(tan2)", %%xmm5  \n\t" \
308     MOV_32_ONLY ROW2", "SREG2"        \n\t" \
309     "pmulhw   "SREG2", %%xmm5         \n\t" \
310     MOV_32_ONLY ROW0", "REG0"         \n\t" \
311     "movdqa   "REG0", %%xmm6          \n\t" \
312     "psubsw   "SREG2", %%xmm6         \n\t" \
313     "paddsw   "REG0", "SREG2"         \n\t" \
314     MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
315     "movdqa   "REG0", "XMMS"          \n\t" \
316     "psubsw   %%xmm5, "REG0"          \n\t" \
317     "paddsw   "XMMS", %%xmm5          \n\t" \
318     "movdqa   %%xmm5, "XMMS"          \n\t" \
319     "psubsw   "TAN3", %%xmm5          \n\t" \
320     "paddsw   "XMMS", "TAN3"          \n\t" \
321     "movdqa   "REG0", "XMMS"          \n\t" \
322     "psubsw   %%xmm3, "REG0"          \n\t" \
323     "paddsw   "XMMS", %%xmm3          \n\t" \
324     MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
325     "psraw        $6, %%xmm5          \n\t" \
326     "psraw        $6, "REG0"          \n\t" \
327     "psraw        $6, "TAN3"          \n\t" \
328     "psraw        $6, %%xmm3          \n\t" \
329     "movdqa   "TAN3", 1*16("dct")     \n\t" \
330     "movdqa   %%xmm3, 2*16("dct")     \n\t" \
331     "movdqa   "REG0", 5*16("dct")     \n\t" \
332     "movdqa   %%xmm5, 6*16("dct")     \n\t" \
333     "movdqa   "SREG2", %%xmm0         \n\t" \
334     "movdqa   %%xmm6, %%xmm4          \n\t" \
335     "psubsw   %%xmm1, "SREG2"         \n\t" \
336     "psubsw   "TAN1", %%xmm6          \n\t" \
337     "paddsw   %%xmm0, %%xmm1          \n\t" \
338     "paddsw   %%xmm4, "TAN1"          \n\t" \
339     "psraw        $6, %%xmm1          \n\t" \
340     "psraw        $6, "SREG2"         \n\t" \
341     "psraw        $6, "TAN1"          \n\t" \
342     "psraw        $6, %%xmm6          \n\t" \
343     "movdqa   %%xmm1, ("dct")         \n\t" \
344     "movdqa   "TAN1", 3*16("dct")     \n\t" \
345     "movdqa   %%xmm6, 4*16("dct")     \n\t" \
346     "movdqa   "SREG2", 7*16("dct")    \n\t"
347
348 inline void ff_idct_xvid_sse2(short *block)
349 {
350     __asm__ volatile(
351     "movq     "MANGLE(m127)", %%mm0                              \n\t"
352     iMTX_MULT("(%0)",     MANGLE(iTab1), ROUND(walkenIdctRounders),      PUT_EVEN(ROW0))
353     iMTX_MULT("1*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+1*16), PUT_ODD(ROW1))
354     iMTX_MULT("2*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+2*16), PUT_EVEN(ROW2))
355
356     TEST_TWO_ROWS("3*16(%0)", "4*16(%0)", "%%eax", "%%ecx", CLEAR_ODD(ROW3), CLEAR_EVEN(ROW4))
357     JZ("%%eax", "1f")
358     iMTX_MULT("3*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+3*16), PUT_ODD(ROW3))
359
360     TEST_TWO_ROWS("5*16(%0)", "6*16(%0)", "%%eax", "%%edx", CLEAR_ODD(ROW5), CLEAR_EVEN(ROW6))
361     TEST_ONE_ROW("7*16(%0)", "%%esi", CLEAR_ODD(ROW7))
362     iLLM_HEAD
363     ".p2align 4 \n\t"
364     JNZ("%%ecx", "2f")
365     JNZ("%%eax", "3f")
366     JNZ("%%edx", "4f")
367     JNZ("%%esi", "5f")
368     iLLM_PASS_SPARSE("%0")
369     "jmp 6f                                                      \n\t"
370     "2:                                                          \n\t"
371     iMTX_MULT("4*16(%0)", MANGLE(iTab1), "#", PUT_EVEN(ROW4))
372     "3:                                                          \n\t"
373     iMTX_MULT("5*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+4*16), PUT_ODD(ROW5))
374     JZ("%%edx", "1f")
375     "4:                                                          \n\t"
376     iMTX_MULT("6*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+5*16), PUT_EVEN(ROW6))
377     JZ("%%esi", "1f")
378     "5:                                                          \n\t"
379     iMTX_MULT("7*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+5*16), PUT_ODD(ROW7))
380 #if !ARCH_X86_64
381     iLLM_HEAD
382 #endif
383     iLLM_PASS("%0")
384     "6:                                                          \n\t"
385     : "+r"(block)
386     :
387     : XMM_CLOBBERS("%xmm0" , "%xmm1" , "%xmm2" , "%xmm3" ,
388                    "%xmm4" , "%xmm5" , "%xmm6" , "%xmm7" ,)
389 #if ARCH_X86_64
390       XMM_CLOBBERS("%xmm8" , "%xmm9" , "%xmm10", "%xmm11",
391                    "%xmm12", "%xmm13", "%xmm14",)
392 #endif
393       "%eax", "%ecx", "%edx", "%esi", "memory"
394     );
395 }
396
397 void ff_idct_xvid_sse2_put(uint8_t *dest, int line_size, short *block)
398 {
399     ff_idct_xvid_sse2(block);
400     ff_put_pixels_clamped_mmx(block, dest, line_size);
401 }
402
403 void ff_idct_xvid_sse2_add(uint8_t *dest, int line_size, short *block)
404 {
405     ff_idct_xvid_sse2(block);
406     ff_add_pixels_clamped_mmx(block, dest, line_size);
407 }
408
409 #endif /* HAVE_INLINE_ASM */