annotate ps2/idct_mmi.c @ 1991:fa09319b80af libavcodec

10l
author michael
date Fri, 30 Apr 2004 01:35:45 +0000
parents 72ac356803ea
children ef2149182f1c
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
1 /*
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
2 Originally provided by Intel at AP-922
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
3 http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
4 (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm)
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
5 but in a limited edition.
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
6
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
7 column code adapted from peter gubanov
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
8 Copyright (c) 2000-2001 Peter Gubanov <peter@elecard.net.ru>
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
9 http://www.elecard.com/peter/idct.shtml
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
10 Rounding trick Copyright (c) 2000 Michel Lespinasse <walken@zoy.org>
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
11
1876
72ac356803ea credit update
melanson
parents: 1064
diff changeset
12 MMI port by Leon van Stuivenberg
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
13 */
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
14 #include "../common.h"
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
15 #include "../dsputil.h"
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
16 #include "mmi.h"
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
17
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
18 #define BITS_INV_ACC 5 // 4 or 5 for IEEE
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
19 #define SHIFT_INV_ROW (16 - BITS_INV_ACC)
696
477bcb3b2f0a ps2 idct bugfix patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
20 #define SHIFT_INV_COL (1 + BITS_INV_ACC)
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
21
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
22 #define TG1 6518
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
23 #define TG2 13573
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
24 #define TG3 21895
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
25 #define CS4 23170
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
26
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
27 #define ROUNDER_0 0
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
28 #define ROUNDER_1 16
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
29
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
30 #define TAB_i_04 (32+0)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
31 #define TAB_i_17 (32+64)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
32 #define TAB_i_26 (32+128)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
33 #define TAB_i_35 (32+192)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
34
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
35 #define TG_1_16 (32+256+0)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
36 #define TG_2_16 (32+256+16)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
37 #define TG_3_16 (32+256+32)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
38 #define COS_4_16 (32+256+48)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
39
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
40 #define CLIPMAX (32+256+64+0)
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
41
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
42 static short consttable[] align16 = {
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
43 /* rounder 0*/ // assume SHIFT_INV_ROW == 11
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
44 0x3ff, 1, 0x3ff, 1, 0x3ff, 1, 0x3ff, 1,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
45 /* rounder 1*/
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
46 0x3ff, 0, 0x3ff, 0, 0x3ff, 0, 0x3ff, 0,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
47 /* row 0/4*/
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
48 16384, 21407, -16384, -21407, 22725, 19266, -22725, -12873,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
49 8867, 16384, 8867, 16384, 4520, 12873, -4520, 19266,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
50 16384, -8867, 16384, -8867, 12873, -22725, 19266, -22725,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
51 21407, -16384, -21407, 16384, 19266, 4520, -12873, 4520,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
52 /* row 1/7*/
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
53 22725, 29692, -22725, -29692, 31521, 26722, -31521, -17855,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
54 12299, 22725, 12299, 22725, 6270, 17855, -6270, 26722,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
55 22725, -12299, 22725, -12299, 17855, -31521, 26722, -31521,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
56 29692, -22725, -29692, 22725, 26722, 6270, -17855, 6270,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
57 /* row 2/6*/
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
58 21407, 27969, -21407, -27969, 29692, 25172, -29692, -16819,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
59 11585, 21407, 11585, 21407, 5906, 16819, -5906, 25172,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
60 21407, -11585, 21407, -11585, 16819, -29692, 25172, -29692,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
61 27969, -21407, -27969, 21407, 25172, 5906, -16819, 5906,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
62 /*row 3/5*/
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
63 19266, 25172, -19266, -25172, 26722, 22654, -26722, -15137,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
64 10426, 19266, 10426, 19266, 5315, 15137, -5315, 22654,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
65 19266, -10426, 19266, -10426, 15137, -26722, 22654, -26722,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
66 25172, -19266, -25172, 19266, 22654, 5315, -15137, 5315,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
67 /*column constants*/
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
68 TG1, TG1, TG1, TG1, TG1, TG1, TG1, TG1,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
69 TG2, TG2, TG2, TG2, TG2, TG2, TG2, TG2,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
70 TG3, TG3, TG3, TG3, TG3, TG3, TG3, TG3,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
71 CS4, CS4, CS4, CS4, CS4, CS4, CS4, CS4,
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
72 /* clamp */
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
73 255, 255, 255, 255, 255, 255, 255, 255
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
74 };
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
75
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
76
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
77 #define DCT_8_INV_ROW1(blk, rowoff, taboff, rnd, outreg) { \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
78 lq(blk, rowoff, $16); /* r16 = x7 x5 x3 x1 x6 x4 x2 x0 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
79 /*slot*/ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
80 lq($24, 0+taboff, $17); /* r17 = w */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
81 /*delay slot $16*/ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
82 lq($24, 16+taboff, $18);/* r18 = w */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
83 prevh($16, $2); /* r2 = x1 x3 x5 x7 x0 x2 x4 x6 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
84 lq($24, 32+taboff, $19);/* r19 = w */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
85 phmadh($17, $16, $17); /* r17 = b1"b0'a1"a0' */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
86 lq($24, 48+taboff, $20);/* r20 = w */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
87 phmadh($18, $2, $18); /* r18 = b1'b0"a1'a0" */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
88 phmadh($19, $16, $19); /* r19 = b3"b2'a3"a2' */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
89 phmadh($20, $2, $20); /* r20 = b3'b2"a3'a2" */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
90 paddw($17, $18, $17); /* r17 = (b1)(b0)(a1)(a0) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
91 paddw($19, $20, $19); /* r19 = (b3)(b2)(a3)(a2) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
92 pcpyld($19, $17, $18); /* r18 = (a3)(a2)(a1)(a0) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
93 pcpyud($17, $19, $20); /* r20 = (b3)(b2)(b1)(b0) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
94 paddw($18, rnd, $18); /* r18 = (a3)(a2)(a1)(a0) */\
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
95 paddw($18, $20, $17); /* r17 = ()()()(a0+b0) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
96 psubw($18, $20, $20); /* r20 = ()()()(a0-b0) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
97 psraw($17, SHIFT_INV_ROW, $17); /* r17 = (y3 y2 y1 y0) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
98 psraw($20, SHIFT_INV_ROW, $20); /* r20 = (y4 y5 y6 y7) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
99 ppach($20, $17, outreg);/* out = y4 y5 y6 y7 y3 y2 y1 y0 Note order */ \
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
100 \
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
101 prevh(outreg, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
102 pcpyud($2, $2, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
103 pcpyld($2, outreg, outreg); \
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
104 }
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
105
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
106
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
107 #define DCT_8_INV_COL8() \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
108 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
109 lq($24, TG_3_16, $2); /* r2 = tn3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
110 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
111 pmulth($11, $2, $17); /* r17 = x3 * tn3 (6420) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
112 psraw($17, 15, $17); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
113 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
114 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
115 pinteh($3, $17, $17); /* r17 = x3 * tn3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
116 psubh($17, $13, $17); /* r17 = tm35 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
117 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
118 pmulth($13, $2, $18); /* r18 = x5 * tn3 (6420) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
119 psraw($18, 15, $18); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
120 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
121 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
122 pinteh($3, $18, $18); /* r18 = x5 * tn3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
123 paddh($18, $11, $18); /* r18 = tp35 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
124 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
125 lq($24, TG_1_16, $2); /* r2 = tn1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
126 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
127 pmulth($15, $2, $19); /* r19 = x7 * tn1 (6420) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
128 psraw($19, 15, $19); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
129 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
130 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
131 pinteh($3, $19, $19); /* r19 = x7 * tn1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
132 paddh($19, $9, $19); /* r19 = tp17 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
133 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
134 pmulth($9, $2, $20); /* r20 = x1 * tn1 (6420) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
135 psraw($20, 15, $20); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
136 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
137 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
138 pinteh($3, $20, $20); /* r20 = x1 * tn1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
139 psubh($20, $15, $20); /* r20 = tm17 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
140 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
141 psubh($19, $18, $3); /* r3 = t1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
142 paddh($20, $17, $16); /* r16 = t2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
143 psubh($20, $17, $23); /* r23 = b3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
144 paddh($19, $18, $20); /* r20 = b0 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
145 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
146 lq($24, COS_4_16, $2); /* r2 = cs4 */ \
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
147 \
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
148 paddh($3, $16, $21); /* r21 = t1+t2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
149 psubh($3, $16, $22); /* r22 = t1-t2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
150 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
151 pmulth($21, $2, $21); /* r21 = cs4 * (t1+t2) 6420 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
152 psraw($21, 15, $21); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
153 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
154 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
155 pinteh($3, $21, $21); /* r21 = b1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
156 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
157 pmulth($22, $2, $22); /* r22 = cs4 * (t1-t2) 6420 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
158 psraw($22, 15, $22); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
159 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
160 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
161 pinteh($3, $22, $22); /* r22 = b2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
162 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
163 lq($24, TG_2_16, $2); /* r2 = tn2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
164 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
165 pmulth($10, $2, $17); /* r17 = x2 * tn2 (6420) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
166 psraw($17, 15, $17); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
167 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
168 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
169 pinteh($3, $17, $17); /* r17 = x3 * tn3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
170 psubh($17, $14, $17); /* r17 = tm26 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
171 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
172 pmulth($14, $2, $18); /* r18 = x6 * tn2 (6420) */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
173 psraw($18, 15, $18); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
174 pmfhl_uw($3); /* r3 = 7531 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
175 psraw($3, 15, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
176 pinteh($3, $18, $18); /* r18 = x6 * tn2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
177 paddh($18, $10, $18); /* r18 = tp26 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
178 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
179 paddh($8, $12, $2); /* r2 = tp04 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
180 psubh($8, $12, $3); /* r3 = tm04 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
181 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
182 paddh($2, $18, $16); /* r16 = a0 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
183 psubh($2, $18, $19); /* r19 = a3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
184 psubh($3, $17, $18); /* r18 = a2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
185 paddh($3, $17, $17); /* r17 = a1 */
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
186
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
187
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
188 #define DCT_8_INV_COL8_STORE(blk) \
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
189 \
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
190 paddh($16, $20, $2); /* y0 a0+b0 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
191 psubh($16, $20, $16); /* y7 a0-b0 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
192 psrah($2, SHIFT_INV_COL, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
193 psrah($16, SHIFT_INV_COL, $16); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
194 sq($2, 0, blk); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
195 sq($16, 112, blk); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
196 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
197 paddh($17, $21, $3); /* y1 a1+b1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
198 psubh($17, $21, $17); /* y6 a1-b1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
199 psrah($3, SHIFT_INV_COL, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
200 psrah($17, SHIFT_INV_COL, $17); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
201 sq($3, 16, blk); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
202 sq($17, 96, blk); \
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
203 \
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
204 paddh($18, $22, $2); /* y2 a2+b2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
205 psubh($18, $22, $18); /* y5 a2-b2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
206 psrah($2, SHIFT_INV_COL, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
207 psrah($18, SHIFT_INV_COL, $18); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
208 sq($2, 32, blk); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
209 sq($18, 80, blk); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
210 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
211 paddh($19, $23, $3); /* y3 a3+b3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
212 psubh($19, $23, $19); /* y4 a3-b3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
213 psrah($3, SHIFT_INV_COL, $3); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
214 psrah($19, SHIFT_INV_COL, $19); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
215 sq($3, 48, blk); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
216 sq($19, 64, blk);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
217
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
218
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
219
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
220 #define DCT_8_INV_COL8_PMS() \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
221 paddh($16, $20, $2); /* y0 a0+b0 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
222 psubh($16, $20, $20); /* y7 a0-b0 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
223 psrah($2, SHIFT_INV_COL, $16); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
224 psrah($20, SHIFT_INV_COL, $20); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
225 \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
226 paddh($17, $21, $3); /* y1 a1+b1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
227 psubh($17, $21, $21); /* y6 a1-b1 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
228 psrah($3, SHIFT_INV_COL, $17); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
229 psrah($21, SHIFT_INV_COL, $21); \
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
230 \
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
231 paddh($18, $22, $2); /* y2 a2+b2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
232 psubh($18, $22, $22); /* y5 a2-b2 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
233 psrah($2, SHIFT_INV_COL, $18); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
234 psrah($22, SHIFT_INV_COL, $22); \
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
235 \
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
236 paddh($19, $23, $3); /* y3 a3+b3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
237 psubh($19, $23, $23); /* y4 a3-b3 */ \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
238 psrah($3, SHIFT_INV_COL, $19); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
239 psrah($23, SHIFT_INV_COL, $23);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
240
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
241 #define PUT(rs) \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
242 pminh(rs, $11, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
243 pmaxh($2, $0, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
244 ppacb($0, $2, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
245 sd3(2, 0, 4); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
246 __asm__ __volatile__ ("add $4, $5, $4");
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
247
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
248 #define DCT_8_INV_COL8_PUT() \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
249 PUT($16); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
250 PUT($17); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
251 PUT($18); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
252 PUT($19); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
253 PUT($23); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
254 PUT($22); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
255 PUT($21); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
256 PUT($20);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
257
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
258 #define ADD(rs) \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
259 ld3(4, 0, 2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
260 pextlb($0, $2, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
261 paddh($2, rs, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
262 pminh($2, $11, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
263 pmaxh($2, $0, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
264 ppacb($0, $2, $2); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
265 sd3(2, 0, 4); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
266 __asm__ __volatile__ ("add $4, $5, $4");
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
267
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
268 /*fixme: schedule*/
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
269 #define DCT_8_INV_COL8_ADD() \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
270 ADD($16); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
271 ADD($17); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
272 ADD($18); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
273 ADD($19); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
274 ADD($23); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
275 ADD($22); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
276 ADD($21); \
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
277 ADD($20);
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
278
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
279
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
280 void ff_mmi_idct(int16_t * block)
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
281 {
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
282 /* $4 = block */
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
283 __asm__ __volatile__("la $24, %0"::"m"(consttable[0]));
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
284 lq($24, ROUNDER_0, $8);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
285 lq($24, ROUNDER_1, $7);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
286 DCT_8_INV_ROW1($4, 0, TAB_i_04, $8, $8);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
287 DCT_8_INV_ROW1($4, 16, TAB_i_17, $7, $9);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
288 DCT_8_INV_ROW1($4, 32, TAB_i_26, $7, $10);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
289 DCT_8_INV_ROW1($4, 48, TAB_i_35, $7, $11);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
290 DCT_8_INV_ROW1($4, 64, TAB_i_04, $7, $12);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
291 DCT_8_INV_ROW1($4, 80, TAB_i_35, $7, $13);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
292 DCT_8_INV_ROW1($4, 96, TAB_i_26, $7, $14);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
293 DCT_8_INV_ROW1($4, 112, TAB_i_17, $7, $15);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
294 DCT_8_INV_COL8();
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
295 DCT_8_INV_COL8_STORE($4);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
296
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
297 //let savedtemp regs be saved
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
298 __asm__ __volatile__(" ":::"$16", "$17", "$18", "$19", "$20", "$21", "$22", "$23");
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
299 }
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
300
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
301
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 721
diff changeset
302 void ff_mmi_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
303 {
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
304 /* $4 = dest, $5 = line_size, $6 = block */
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
305 __asm__ __volatile__("la $24, %0"::"m"(consttable[0]));
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
306 lq($24, ROUNDER_0, $8);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
307 lq($24, ROUNDER_1, $7);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
308 DCT_8_INV_ROW1($6, 0, TAB_i_04, $8, $8);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
309 DCT_8_INV_ROW1($6, 16, TAB_i_17, $7, $9);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
310 DCT_8_INV_ROW1($6, 32, TAB_i_26, $7, $10);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
311 DCT_8_INV_ROW1($6, 48, TAB_i_35, $7, $11);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
312 DCT_8_INV_ROW1($6, 64, TAB_i_04, $7, $12);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
313 DCT_8_INV_ROW1($6, 80, TAB_i_35, $7, $13);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
314 DCT_8_INV_ROW1($6, 96, TAB_i_26, $7, $14);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
315 DCT_8_INV_ROW1($6, 112, TAB_i_17, $7, $15);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
316 DCT_8_INV_COL8();
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
317 lq($24, CLIPMAX, $11);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
318 DCT_8_INV_COL8_PMS();
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
319 DCT_8_INV_COL8_PUT();
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
320
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
321 //let savedtemp regs be saved
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
322 __asm__ __volatile__(" ":::"$16", "$17", "$18", "$19", "$20", "$21", "$22", "$23");
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
323 }
696
477bcb3b2f0a ps2 idct bugfix patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
324
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
325
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 721
diff changeset
326 void ff_mmi_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
327 {
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
328 /* $4 = dest, $5 = line_size, $6 = block */
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
329 __asm__ __volatile__("la $24, %0"::"m"(consttable[0]));
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
330 lq($24, ROUNDER_0, $8);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
331 lq($24, ROUNDER_1, $7);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
332 DCT_8_INV_ROW1($6, 0, TAB_i_04, $8, $8);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
333 DCT_8_INV_ROW1($6, 16, TAB_i_17, $7, $9);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
334 DCT_8_INV_ROW1($6, 32, TAB_i_26, $7, $10);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
335 DCT_8_INV_ROW1($6, 48, TAB_i_35, $7, $11);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
336 DCT_8_INV_ROW1($6, 64, TAB_i_04, $7, $12);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
337 DCT_8_INV_ROW1($6, 80, TAB_i_35, $7, $13);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
338 DCT_8_INV_ROW1($6, 96, TAB_i_26, $7, $14);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
339 DCT_8_INV_ROW1($6, 112, TAB_i_17, $7, $15);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
340 DCT_8_INV_COL8();
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
341 lq($24, CLIPMAX, $11);
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
342 DCT_8_INV_COL8_PMS();
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
343 DCT_8_INV_COL8_ADD();
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
344
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
345 //let savedtemp regs be saved
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
346 __asm__ __volatile__(" ":::"$16", "$17", "$18", "$19", "$20", "$21", "$22", "$23");
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
347 }
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 696
diff changeset
348