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 "libavutil/mem.h"
42#include "libavutil/x86/asm.h"
43#include "idct_xvid.h"
44#include "idctdsp.h"
45
46#if HAVE_SSE2_INLINE
47
48/**
49 * @file
50 * @brief SSE2 idct compatible with xvidmmx
51 */
52
53#define X8(x)     x,x,x,x,x,x,x,x
54
55#define ROW_SHIFT 11
56#define COL_SHIFT 6
57
58DECLARE_ASM_CONST(16, int16_t, tan1)[] = {X8(13036)}; // tan( pi/16)
59DECLARE_ASM_CONST(16, int16_t, tan2)[] = {X8(27146)}; // tan(2pi/16) = sqrt(2)-1
60DECLARE_ASM_CONST(16, int16_t, tan3)[] = {X8(43790)}; // tan(3pi/16)-1
61DECLARE_ASM_CONST(16, int16_t, sqrt2)[]= {X8(23170)}; // 0.5/sqrt(2)
62DECLARE_ASM_CONST(8,  uint8_t, m127)[] = {X8(127)};
63
64DECLARE_ASM_CONST(16, int16_t, iTab1)[] = {
65 0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d,
66 0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61,
67 0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7,
68 0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b
69};
70
71DECLARE_ASM_CONST(16, int16_t, iTab2)[] = {
72 0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5,
73 0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04,
74 0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41,
75 0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df
76};
77
78DECLARE_ASM_CONST(16, int16_t, iTab3)[] = {
79 0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf,
80 0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf,
81 0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d,
82 0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04
83};
84
85DECLARE_ASM_CONST(16, int16_t, iTab4)[] = {
86 0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746,
87 0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac,
88 0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df,
89 0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e
90};
91
92DECLARE_ASM_CONST(16, int32_t, walkenIdctRounders)[] = {
93 65536, 65536, 65536, 65536,
94  3597,  3597,  3597,  3597,
95  2260,  2260,  2260,  2260,
96  1203,  1203,  1203,  1203,
97   120,   120,   120,   120,
98   512,   512,   512,   512
99};
100
101// Temporary storage before the column pass
102#define ROW1 "%%xmm6"
103#define ROW3 "%%xmm4"
104#define ROW5 "%%xmm5"
105#define ROW7 "%%xmm7"
106
107#define CLEAR_ODD(r) "pxor  "r","r" \n\t"
108#define PUT_ODD(dst) "pshufhw  $0x1B, %%xmm2, "dst"   \n\t"
109
110#if ARCH_X86_64
111
112# define ROW0 "%%xmm8"
113# define REG0 ROW0
114# define ROW2 "%%xmm9"
115# define REG2 ROW2
116# define ROW4 "%%xmm10"
117# define REG4 ROW4
118# define ROW6 "%%xmm11"
119# define REG6 ROW6
120# define CLEAR_EVEN(r) CLEAR_ODD(r)
121# define PUT_EVEN(dst) PUT_ODD(dst)
122# define XMMS "%%xmm12"
123# define MOV_32_ONLY "#"
124# define SREG2 REG2
125# define TAN3 "%%xmm13"
126# define TAN1 "%%xmm14"
127
128#else
129
130# define ROW0 "(%0)"
131# define REG0 "%%xmm4"
132# define ROW2 "2*16(%0)"
133# define REG2 "%%xmm4"
134# define ROW4 "4*16(%0)"
135# define REG4 "%%xmm6"
136# define ROW6 "6*16(%0)"
137# define REG6 "%%xmm6"
138# define CLEAR_EVEN(r)
139# define PUT_EVEN(dst) \
140    "pshufhw  $0x1B, %%xmm2, %%xmm2   \n\t" \
141    "movdqa          %%xmm2, "dst"    \n\t"
142# define XMMS "%%xmm2"
143# define MOV_32_ONLY "movdqa "
144# define SREG2 "%%xmm7"
145# define TAN3 "%%xmm0"
146# define TAN1 "%%xmm2"
147
148#endif
149
150#define ROUND(x) "paddd   "x
151
152#define JZ(reg, to)                         \
153    "testl     "reg","reg"            \n\t" \
154    "jz        "to"                   \n\t"
155
156#define JNZ(reg, to)                        \
157    "testl     "reg","reg"            \n\t" \
158    "jnz       "to"                   \n\t"
159
160#define TEST_ONE_ROW(src, reg, clear)       \
161    clear                                   \
162    "movq     "src", %%mm1            \n\t" \
163    "por    8+"src", %%mm1            \n\t" \
164    "paddusb  %%mm0, %%mm1            \n\t" \
165    "pmovmskb %%mm1, "reg"            \n\t"
166
167#define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2) \
168    clear1                                  \
169    clear2                                  \
170    "movq     "row1", %%mm1           \n\t" \
171    "por    8+"row1", %%mm1           \n\t" \
172    "movq     "row2", %%mm2           \n\t" \
173    "por    8+"row2", %%mm2           \n\t" \
174    "paddusb   %%mm0, %%mm1           \n\t" \
175    "paddusb   %%mm0, %%mm2           \n\t" \
176    "pmovmskb  %%mm1, "reg1"          \n\t" \
177    "pmovmskb  %%mm2, "reg2"          \n\t"
178
179///IDCT pass on rows.
180#define iMTX_MULT(src, table, rounder, put) \
181    "movdqa        "src", %%xmm3      \n\t" \
182    "movdqa       %%xmm3, %%xmm0      \n\t" \
183    "pshufd   $0x11, %%xmm3, %%xmm1   \n\t" /* 4602 */ \
184    "punpcklqdq   %%xmm0, %%xmm0      \n\t" /* 0246 */ \
185    "pmaddwd     "table", %%xmm0      \n\t" \
186    "pmaddwd  16+"table", %%xmm1      \n\t" \
187    "pshufd   $0xBB, %%xmm3, %%xmm2   \n\t" /* 5713 */ \
188    "punpckhqdq   %%xmm3, %%xmm3      \n\t" /* 1357 */ \
189    "pmaddwd  32+"table", %%xmm2      \n\t" \
190    "pmaddwd  48+"table", %%xmm3      \n\t" \
191    "paddd        %%xmm1, %%xmm0      \n\t" \
192    "paddd        %%xmm3, %%xmm2      \n\t" \
193    rounder",     %%xmm0              \n\t" \
194    "movdqa       %%xmm2, %%xmm3      \n\t" \
195    "paddd        %%xmm0, %%xmm2      \n\t" \
196    "psubd        %%xmm3, %%xmm0      \n\t" \
197    "psrad           $11, %%xmm2      \n\t" \
198    "psrad           $11, %%xmm0      \n\t" \
199    "packssdw     %%xmm0, %%xmm2      \n\t" \
200    put                                     \
201    "1:                               \n\t"
202
203#define iLLM_HEAD                           \
204    "movdqa   "MANGLE(tan3)", "TAN3"  \n\t" \
205    "movdqa   "MANGLE(tan1)", "TAN1"  \n\t" \
206
207///IDCT pass on columns.
208#define iLLM_PASS(dct)                      \
209    "movdqa   "TAN3", %%xmm1          \n\t" \
210    "movdqa   "TAN1", %%xmm3          \n\t" \
211    "pmulhw   %%xmm4, "TAN3"          \n\t" \
212    "pmulhw   %%xmm5, %%xmm1          \n\t" \
213    "paddsw   %%xmm4, "TAN3"          \n\t" \
214    "paddsw   %%xmm5, %%xmm1          \n\t" \
215    "psubsw   %%xmm5, "TAN3"          \n\t" \
216    "paddsw   %%xmm4, %%xmm1          \n\t" \
217    "pmulhw   %%xmm7, %%xmm3          \n\t" \
218    "pmulhw   %%xmm6, "TAN1"          \n\t" \
219    "paddsw   %%xmm6, %%xmm3          \n\t" \
220    "psubsw   %%xmm7, "TAN1"          \n\t" \
221    "movdqa   %%xmm3, %%xmm7          \n\t" \
222    "movdqa   "TAN1", %%xmm6          \n\t" \
223    "psubsw   %%xmm1, %%xmm3          \n\t" \
224    "psubsw   "TAN3", "TAN1"          \n\t" \
225    "paddsw   %%xmm7, %%xmm1          \n\t" \
226    "paddsw   %%xmm6, "TAN3"          \n\t" \
227    "movdqa   %%xmm3, %%xmm6          \n\t" \
228    "psubsw   "TAN3", %%xmm3          \n\t" \
229    "paddsw   %%xmm6, "TAN3"          \n\t" \
230    "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
231    "pmulhw   %%xmm4, %%xmm3          \n\t" \
232    "pmulhw   %%xmm4, "TAN3"          \n\t" \
233    "paddsw   "TAN3", "TAN3"          \n\t" \
234    "paddsw   %%xmm3, %%xmm3          \n\t" \
235    "movdqa   "MANGLE(tan2)", %%xmm7  \n\t" \
236    MOV_32_ONLY ROW2", "REG2"         \n\t" \
237    MOV_32_ONLY ROW6", "REG6"         \n\t" \
238    "movdqa   %%xmm7, %%xmm5          \n\t" \
239    "pmulhw   "REG6", %%xmm7          \n\t" \
240    "pmulhw   "REG2", %%xmm5          \n\t" \
241    "paddsw   "REG2", %%xmm7          \n\t" \
242    "psubsw   "REG6", %%xmm5          \n\t" \
243    MOV_32_ONLY ROW0", "REG0"         \n\t" \
244    MOV_32_ONLY ROW4", "REG4"         \n\t" \
245    MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
246    "movdqa   "REG0", "XMMS"          \n\t" \
247    "psubsw   "REG4", "REG0"          \n\t" \
248    "paddsw   "XMMS", "REG4"          \n\t" \
249    "movdqa   "REG4", "XMMS"          \n\t" \
250    "psubsw   %%xmm7, "REG4"          \n\t" \
251    "paddsw   "XMMS", %%xmm7          \n\t" \
252    "movdqa   "REG0", "XMMS"          \n\t" \
253    "psubsw   %%xmm5, "REG0"          \n\t" \
254    "paddsw   "XMMS", %%xmm5          \n\t" \
255    "movdqa   %%xmm5, "XMMS"          \n\t" \
256    "psubsw   "TAN3", %%xmm5          \n\t" \
257    "paddsw   "XMMS", "TAN3"          \n\t" \
258    "movdqa   "REG0", "XMMS"          \n\t" \
259    "psubsw   %%xmm3, "REG0"          \n\t" \
260    "paddsw   "XMMS", %%xmm3          \n\t" \
261    MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
262    "psraw        $6, %%xmm5          \n\t" \
263    "psraw        $6, "REG0"          \n\t" \
264    "psraw        $6, "TAN3"          \n\t" \
265    "psraw        $6, %%xmm3          \n\t" \
266    "movdqa   "TAN3", 1*16("dct")     \n\t" \
267    "movdqa   %%xmm3, 2*16("dct")     \n\t" \
268    "movdqa   "REG0", 5*16("dct")     \n\t" \
269    "movdqa   %%xmm5, 6*16("dct")     \n\t" \
270    "movdqa   %%xmm7, %%xmm0          \n\t" \
271    "movdqa   "REG4", %%xmm4          \n\t" \
272    "psubsw   %%xmm1, %%xmm7          \n\t" \
273    "psubsw   "TAN1", "REG4"          \n\t" \
274    "paddsw   %%xmm0, %%xmm1          \n\t" \
275    "paddsw   %%xmm4, "TAN1"          \n\t" \
276    "psraw        $6, %%xmm1          \n\t" \
277    "psraw        $6, %%xmm7          \n\t" \
278    "psraw        $6, "TAN1"          \n\t" \
279    "psraw        $6, "REG4"          \n\t" \
280    "movdqa   %%xmm1, ("dct")         \n\t" \
281    "movdqa   "TAN1", 3*16("dct")     \n\t" \
282    "movdqa   "REG4", 4*16("dct")     \n\t" \
283    "movdqa   %%xmm7, 7*16("dct")     \n\t"
284
285///IDCT pass on columns, assuming rows 4-7 are zero.
286#define iLLM_PASS_SPARSE(dct)               \
287    "pmulhw   %%xmm4, "TAN3"          \n\t" \
288    "paddsw   %%xmm4, "TAN3"          \n\t" \
289    "movdqa   %%xmm6, %%xmm3          \n\t" \
290    "pmulhw   %%xmm6, "TAN1"          \n\t" \
291    "movdqa   %%xmm4, %%xmm1          \n\t" \
292    "psubsw   %%xmm1, %%xmm3          \n\t" \
293    "paddsw   %%xmm6, %%xmm1          \n\t" \
294    "movdqa   "TAN1", %%xmm6          \n\t" \
295    "psubsw   "TAN3", "TAN1"          \n\t" \
296    "paddsw   %%xmm6, "TAN3"          \n\t" \
297    "movdqa   %%xmm3, %%xmm6          \n\t" \
298    "psubsw   "TAN3", %%xmm3          \n\t" \
299    "paddsw   %%xmm6, "TAN3"          \n\t" \
300    "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
301    "pmulhw   %%xmm4, %%xmm3          \n\t" \
302    "pmulhw   %%xmm4, "TAN3"          \n\t" \
303    "paddsw   "TAN3", "TAN3"          \n\t" \
304    "paddsw   %%xmm3, %%xmm3          \n\t" \
305    "movdqa   "MANGLE(tan2)", %%xmm5  \n\t" \
306    MOV_32_ONLY ROW2", "SREG2"        \n\t" \
307    "pmulhw   "SREG2", %%xmm5         \n\t" \
308    MOV_32_ONLY ROW0", "REG0"         \n\t" \
309    "movdqa   "REG0", %%xmm6          \n\t" \
310    "psubsw   "SREG2", %%xmm6         \n\t" \
311    "paddsw   "REG0", "SREG2"         \n\t" \
312    MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
313    "movdqa   "REG0", "XMMS"          \n\t" \
314    "psubsw   %%xmm5, "REG0"          \n\t" \
315    "paddsw   "XMMS", %%xmm5          \n\t" \
316    "movdqa   %%xmm5, "XMMS"          \n\t" \
317    "psubsw   "TAN3", %%xmm5          \n\t" \
318    "paddsw   "XMMS", "TAN3"          \n\t" \
319    "movdqa   "REG0", "XMMS"          \n\t" \
320    "psubsw   %%xmm3, "REG0"          \n\t" \
321    "paddsw   "XMMS", %%xmm3          \n\t" \
322    MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
323    "psraw        $6, %%xmm5          \n\t" \
324    "psraw        $6, "REG0"          \n\t" \
325    "psraw        $6, "TAN3"          \n\t" \
326    "psraw        $6, %%xmm3          \n\t" \
327    "movdqa   "TAN3", 1*16("dct")     \n\t" \
328    "movdqa   %%xmm3, 2*16("dct")     \n\t" \
329    "movdqa   "REG0", 5*16("dct")     \n\t" \
330    "movdqa   %%xmm5, 6*16("dct")     \n\t" \
331    "movdqa   "SREG2", %%xmm0         \n\t" \
332    "movdqa   %%xmm6, %%xmm4          \n\t" \
333    "psubsw   %%xmm1, "SREG2"         \n\t" \
334    "psubsw   "TAN1", %%xmm6          \n\t" \
335    "paddsw   %%xmm0, %%xmm1          \n\t" \
336    "paddsw   %%xmm4, "TAN1"          \n\t" \
337    "psraw        $6, %%xmm1          \n\t" \
338    "psraw        $6, "SREG2"         \n\t" \
339    "psraw        $6, "TAN1"          \n\t" \
340    "psraw        $6, %%xmm6          \n\t" \
341    "movdqa   %%xmm1, ("dct")         \n\t" \
342    "movdqa   "TAN1", 3*16("dct")     \n\t" \
343    "movdqa   %%xmm6, 4*16("dct")     \n\t" \
344    "movdqa   "SREG2", 7*16("dct")    \n\t"
345
346av_extern_inline void ff_idct_xvid_sse2(short *block)
347{
348    __asm__ volatile(
349    "movq     "MANGLE(m127)", %%mm0                              \n\t"
350    iMTX_MULT("(%0)",     MANGLE(iTab1), ROUND(MANGLE(walkenIdctRounders)),      PUT_EVEN(ROW0))
351    iMTX_MULT("1*16(%0)", MANGLE(iTab2), ROUND("1*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW1))
352    iMTX_MULT("2*16(%0)", MANGLE(iTab3), ROUND("2*16+"MANGLE(walkenIdctRounders)), PUT_EVEN(ROW2))
353
354    TEST_TWO_ROWS("3*16(%0)", "4*16(%0)", "%%eax", "%%ecx", CLEAR_ODD(ROW3), CLEAR_EVEN(ROW4))
355    JZ("%%eax", "1f")
356    iMTX_MULT("3*16(%0)", MANGLE(iTab4), ROUND("3*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW3))
357
358    TEST_TWO_ROWS("5*16(%0)", "6*16(%0)", "%%eax", "%%edx", CLEAR_ODD(ROW5), CLEAR_EVEN(ROW6))
359    TEST_ONE_ROW("7*16(%0)", "%%esi", CLEAR_ODD(ROW7))
360    iLLM_HEAD
361    ".p2align 4 \n\t"
362    JNZ("%%ecx", "2f")
363    JNZ("%%eax", "3f")
364    JNZ("%%edx", "4f")
365    JNZ("%%esi", "5f")
366    iLLM_PASS_SPARSE("%0")
367    "jmp 6f                                                      \n\t"
368    "2:                                                          \n\t"
369    iMTX_MULT("4*16(%0)", MANGLE(iTab1), "#", PUT_EVEN(ROW4))
370    "3:                                                          \n\t"
371    iMTX_MULT("5*16(%0)", MANGLE(iTab4), ROUND("4*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW5))
372    JZ("%%edx", "1f")
373    "4:                                                          \n\t"
374    iMTX_MULT("6*16(%0)", MANGLE(iTab3), ROUND("5*16+"MANGLE(walkenIdctRounders)), PUT_EVEN(ROW6))
375    JZ("%%esi", "1f")
376    "5:                                                          \n\t"
377    iMTX_MULT("7*16(%0)", MANGLE(iTab2), ROUND("5*16+"MANGLE(walkenIdctRounders)), PUT_ODD(ROW7))
378#if ARCH_X86_32
379    iLLM_HEAD
380#endif
381    iLLM_PASS("%0")
382    "6:                                                          \n\t"
383    : "+r"(block)
384    : NAMED_CONSTRAINTS_ARRAY(m127,iTab1,walkenIdctRounders,iTab2,iTab3,iTab4,tan3,tan1,tan2,sqrt2)
385    : XMM_CLOBBERS("%xmm0" , "%xmm1" , "%xmm2" , "%xmm3" ,
386                   "%xmm4" , "%xmm5" , "%xmm6" , "%xmm7" ,)
387#if ARCH_X86_64
388      XMM_CLOBBERS("%xmm8" , "%xmm9" , "%xmm10", "%xmm11",
389                   "%xmm12", "%xmm13", "%xmm14",)
390#endif
391      "%eax", "%ecx", "%edx", "%esi", "memory"
392    );
393}
394
395void ff_idct_xvid_sse2_put(uint8_t *dest, int line_size, short *block)
396{
397    ff_idct_xvid_sse2(block);
398    ff_put_pixels_clamped_mmx(block, dest, line_size);
399}
400
401void ff_idct_xvid_sse2_add(uint8_t *dest, int line_size, short *block)
402{
403    ff_idct_xvid_sse2(block);
404    ff_add_pixels_clamped_mmx(block, dest, line_size);
405}
406
407#endif /* HAVE_SSE2_INLINE */
408