Mercurial > libavcodec.hg
annotate i386/fdct_mmx.c @ 1922:0ed2d7ecd1e9 libavcodec
fix obnoxious ogg_packet passing from encoder to muxer
author | michael |
---|---|
date | Sun, 04 Apr 2004 14:39:20 +0000 |
parents | e31754bc5b65 |
children | 12408a3bf741 |
rev | line source |
---|---|
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
1 /* |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
2 * MMX optimized forward DCT |
429 | 3 * The gcc porting is Copyright (c) 2001 Fabrice Bellard. |
1739
07a484280a82
copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents:
1575
diff
changeset
|
4 * cleanup/optimizations are Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at> |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
5 * SSE2 optimization is Copyright (c) 2004 Denes Balatoni. |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
6 * |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
7 * from fdctam32.c - AP922 MMX(3D-Now) forward-DCT |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
8 * |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
9 * Intel Application Note AP-922 - fast, precise implementation of DCT |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
10 * http://developer.intel.com/vtune/cbts/appnotes.htm |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
11 * |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
12 * Also of inspiration: |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
13 * a page about fdct at http://www.geocities.com/ssavekar/dct.htm |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
14 * Skal's fdct at http://skal.planet-d.net/coding/dct.html |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
15 */ |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
16 #include "../common.h" |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
17 #include "mmx.h" |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
18 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
19 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align))) |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
20 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
21 ////////////////////////////////////////////////////////////////////// |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
22 // |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
23 // constants for the forward DCT |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
24 // ----------------------------- |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
25 // |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
26 // Be sure to check that your compiler is aligning all constants to QWORD |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
27 // (8-byte) memory boundaries! Otherwise the unaligned memory access will |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
28 // severely stall MMX execution. |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
29 // |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
30 ////////////////////////////////////////////////////////////////////// |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
31 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
32 #define BITS_FRW_ACC 3 //; 2 or 3 for accuracy |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
33 #define SHIFT_FRW_COL BITS_FRW_ACC |
635 | 34 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17 - 3) |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
35 #define RND_FRW_ROW (1 << (SHIFT_FRW_ROW-1)) |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
36 //#define RND_FRW_COL (1 << (SHIFT_FRW_COL-1)) |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
37 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
38 //concatenated table, for forward DCT transformation |
839 | 39 static const int16_t fdct_tg_all_16[] ATTR_ALIGN(8) = { |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
40 13036, 13036, 13036, 13036, // tg * (2<<16) + 0.5 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
41 27146, 27146, 27146, 27146, // tg * (2<<16) + 0.5 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
42 -21746, -21746, -21746, -21746, // tg * (2<<16) + 0.5 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
43 }; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
44 |
839 | 45 static const int16_t ocos_4_16[4] ATTR_ALIGN(8) = { |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
46 23170, 23170, 23170, 23170, //cos * (2<<15) + 0.5 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
47 }; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
48 |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
49 static const long long fdct_one_corr ATTR_ALIGN(8) = 0x0001000100010001LL; |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
50 |
839 | 51 static const long fdct_r_row[2] ATTR_ALIGN(8) = {RND_FRW_ROW, RND_FRW_ROW }; |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
52 |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
53 static const long fdct_r_row_sse2[4] ATTR_ALIGN(16) = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW}; |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
54 |
839 | 55 static const int16_t tab_frw_01234567[] ATTR_ALIGN(8) = { // forward_dct coeff table |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
56 16384, 16384, -8867, -21407, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
57 16384, 16384, 21407, 8867, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
58 16384, -16384, 21407, -8867, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
59 -16384, 16384, 8867, -21407, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
60 22725, 19266, -22725, -12873, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
61 12873, 4520, 19266, -4520, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
62 12873, -22725, 19266, -22725, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
63 4520, 19266, 4520, -12873, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
64 |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
65 22725, 22725, -12299, -29692, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
66 22725, 22725, 29692, 12299, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
67 22725, -22725, 29692, -12299, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
68 -22725, 22725, 12299, -29692, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
69 31521, 26722, -31521, -17855, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
70 17855, 6270, 26722, -6270, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
71 17855, -31521, 26722, -31521, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
72 6270, 26722, 6270, -17855, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
73 |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
74 21407, 21407, -11585, -27969, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
75 21407, 21407, 27969, 11585, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
76 21407, -21407, 27969, -11585, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
77 -21407, 21407, 11585, -27969, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
78 29692, 25172, -29692, -16819, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
79 16819, 5906, 25172, -5906, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
80 16819, -29692, 25172, -29692, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
81 5906, 25172, 5906, -16819, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
82 |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
83 19266, 19266, -10426, -25172, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
84 19266, 19266, 25172, 10426, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
85 19266, -19266, 25172, -10426, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
86 -19266, 19266, 10426, -25172, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
87 26722, 22654, -26722, -15137, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
88 15137, 5315, 22654, -5315, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
89 15137, -26722, 22654, -26722, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
90 5315, 22654, 5315, -15137, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
91 |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
92 16384, 16384, -8867, -21407, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
93 16384, 16384, 21407, 8867, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
94 16384, -16384, 21407, -8867, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
95 -16384, 16384, 8867, -21407, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
96 22725, 19266, -22725, -12873, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
97 12873, 4520, 19266, -4520, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
98 12873, -22725, 19266, -22725, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
99 4520, 19266, 4520, -12873, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
100 |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
101 19266, 19266, -10426, -25172, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
102 19266, 19266, 25172, 10426, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
103 19266, -19266, 25172, -10426, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
104 -19266, 19266, 10426, -25172, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
105 26722, 22654, -26722, -15137, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
106 15137, 5315, 22654, -5315, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
107 15137, -26722, 22654, -26722, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
108 5315, 22654, 5315, -15137, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
109 |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
110 21407, 21407, -11585, -27969, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
111 21407, 21407, 27969, 11585, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
112 21407, -21407, 27969, -11585, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
113 -21407, 21407, 11585, -27969, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
114 29692, 25172, -29692, -16819, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
115 16819, 5906, 25172, -5906, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
116 16819, -29692, 25172, -29692, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
117 5906, 25172, 5906, -16819, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
118 |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
119 22725, 22725, -12299, -29692, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
120 22725, 22725, 29692, 12299, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
121 22725, -22725, 29692, -12299, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
122 -22725, 22725, 12299, -29692, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
123 31521, 26722, -31521, -17855, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
124 17855, 6270, 26722, -6270, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
125 17855, -31521, 26722, -31521, |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
126 6270, 26722, 6270, -17855, |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
127 }; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
128 |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
129 static const int16_t tab_frw_01234567_sse2[] ATTR_ALIGN(16) = { // forward_dct coeff table |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
130 #define TABLE_SSE2 C4, C4, C1, C3, -C6, -C2, -C1, -C5, \ |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
131 C4, C4, C5, C7, C2, C6, C3, -C7, \ |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
132 -C4, C4, C7, C3, C6, -C2, C7, -C5, \ |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
133 C4, -C4, C5, -C1, C2, -C6, C3, -C1, |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
134 // c1..c7 * cos(pi/4) * 2^15 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
135 #define C1 22725 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
136 #define C2 21407 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
137 #define C3 19266 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
138 #define C4 16384 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
139 #define C5 12873 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
140 #define C6 8867 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
141 #define C7 4520 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
142 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
143 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
144 #undef C1 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
145 #undef C2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
146 #undef C3 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
147 #undef C4 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
148 #undef C5 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
149 #undef C6 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
150 #undef C7 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
151 #define C1 31521 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
152 #define C2 29692 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
153 #define C3 26722 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
154 #define C4 22725 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
155 #define C5 17855 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
156 #define C6 12299 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
157 #define C7 6270 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
158 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
159 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
160 #undef C1 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
161 #undef C2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
162 #undef C3 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
163 #undef C4 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
164 #undef C5 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
165 #undef C6 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
166 #undef C7 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
167 #define C1 29692 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
168 #define C2 27969 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
169 #define C3 25172 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
170 #define C4 21407 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
171 #define C5 16819 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
172 #define C6 11585 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
173 #define C7 5906 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
174 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
175 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
176 #undef C1 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
177 #undef C2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
178 #undef C3 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
179 #undef C4 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
180 #undef C5 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
181 #undef C6 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
182 #undef C7 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
183 #define C1 26722 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
184 #define C2 25172 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
185 #define C3 22654 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
186 #define C4 19266 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
187 #define C5 15137 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
188 #define C6 10426 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
189 #define C7 5315 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
190 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
191 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
192 #undef C1 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
193 #undef C2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
194 #undef C3 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
195 #undef C4 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
196 #undef C5 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
197 #undef C6 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
198 #undef C7 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
199 #define C1 22725 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
200 #define C2 21407 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
201 #define C3 19266 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
202 #define C4 16384 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
203 #define C5 12873 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
204 #define C6 8867 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
205 #define C7 4520 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
206 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
207 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
208 #undef C1 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
209 #undef C2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
210 #undef C3 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
211 #undef C4 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
212 #undef C5 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
213 #undef C6 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
214 #undef C7 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
215 #define C1 26722 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
216 #define C2 25172 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
217 #define C3 22654 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
218 #define C4 19266 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
219 #define C5 15137 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
220 #define C6 10426 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
221 #define C7 5315 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
222 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
223 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
224 #undef C1 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
225 #undef C2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
226 #undef C3 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
227 #undef C4 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
228 #undef C5 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
229 #undef C6 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
230 #undef C7 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
231 #define C1 29692 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
232 #define C2 27969 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
233 #define C3 25172 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
234 #define C4 21407 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
235 #define C5 16819 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
236 #define C6 11585 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
237 #define C7 5906 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
238 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
239 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
240 #undef C1 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
241 #undef C2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
242 #undef C3 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
243 #undef C4 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
244 #undef C5 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
245 #undef C6 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
246 #undef C7 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
247 #define C1 31521 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
248 #define C2 29692 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
249 #define C3 26722 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
250 #define C4 22725 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
251 #define C5 17855 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
252 #define C6 12299 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
253 #define C7 6270 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
254 TABLE_SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
255 }; |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
256 |
1564
b6b7d080f1a1
inline -> always_inline (842 -> 690 cpu cycles for dct_quantize() difference for the dct itself should be even bigger)
michael
parents:
839
diff
changeset
|
257 static always_inline void fdct_col(const int16_t *in, int16_t *out, int offset) |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
258 { |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
259 movq_m2r(*(in + offset + 1 * 8), mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
260 movq_m2r(*(in + offset + 6 * 8), mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
261 movq_r2r(mm0, mm2); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
262 movq_m2r(*(in + offset + 2 * 8), mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
263 paddsw_r2r(mm1, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
264 movq_m2r(*(in + offset + 5 * 8), mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
265 psllw_i2r(SHIFT_FRW_COL, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
266 movq_m2r(*(in + offset + 0 * 8), mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
267 paddsw_r2r(mm3, mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
268 paddsw_m2r(*(in + offset + 7 * 8), mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
269 psllw_i2r(SHIFT_FRW_COL, mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
270 movq_r2r(mm0, mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
271 psubsw_r2r(mm1, mm2); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
272 movq_m2r(*(fdct_tg_all_16 + 4), mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
273 psubsw_r2r(mm4, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
274 movq_m2r(*(in + offset + 3 * 8), mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
275 pmulhw_r2r(mm0, mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
276 paddsw_m2r(*(in + offset + 4 * 8), mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
277 psllw_i2r(SHIFT_FRW_COL, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
278 paddsw_r2r(mm4, mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
279 psllw_i2r(SHIFT_FRW_COL, mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
280 movq_r2r(mm5, mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
281 psubsw_r2r(mm7, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
282 paddsw_r2r(mm5, mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
283 paddsw_r2r(mm7, mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
284 por_m2r(fdct_one_corr, mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
285 psllw_i2r(SHIFT_FRW_COL + 1, mm2); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
286 pmulhw_m2r(*(fdct_tg_all_16 + 4), mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
287 movq_r2r(mm4, mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
288 psubsw_m2r(*(in + offset + 5 * 8), mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
289 psubsw_r2r(mm6, mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
290 movq_r2m(mm1, *(out + offset + 2 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
291 paddsw_r2r(mm6, mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
292 movq_m2r(*(in + offset + 3 * 8), mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
293 psllw_i2r(SHIFT_FRW_COL + 1, mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
294 psubsw_m2r(*(in + offset + 4 * 8), mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
295 movq_r2r(mm2, mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
296 movq_r2m(mm4, *(out + offset + 4 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
297 paddsw_r2r(mm3, mm2); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
298 pmulhw_m2r(*ocos_4_16, mm2); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
299 psubsw_r2r(mm3, mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
300 pmulhw_m2r(*ocos_4_16, mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
301 psubsw_r2r(mm0, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
302 por_m2r(fdct_one_corr, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
303 psllw_i2r(SHIFT_FRW_COL, mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
304 por_m2r(fdct_one_corr, mm2); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
305 movq_r2r(mm1, mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
306 movq_m2r(*(in + offset + 0 * 8), mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
307 paddsw_r2r(mm6, mm1); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
308 psubsw_m2r(*(in + offset + 7 * 8), mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
309 psubsw_r2r(mm6, mm4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
310 movq_m2r(*(fdct_tg_all_16 + 0), mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
311 psllw_i2r(SHIFT_FRW_COL, mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
312 movq_m2r(*(fdct_tg_all_16 + 8), mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
313 pmulhw_r2r(mm1, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
314 movq_r2m(mm7, *(out + offset + 0 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
315 pmulhw_r2r(mm4, mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
316 movq_r2m(mm5, *(out + offset + 6 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
317 movq_r2r(mm3, mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
318 movq_m2r(*(fdct_tg_all_16 + 8), mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
319 psubsw_r2r(mm2, mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
320 paddsw_r2r(mm2, mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
321 pmulhw_r2r(mm7, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
322 paddsw_r2r(mm3, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
323 paddsw_r2r(mm4, mm6); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
324 pmulhw_m2r(*(fdct_tg_all_16 + 0), mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
325 por_m2r(fdct_one_corr, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
326 paddsw_r2r(mm7, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
327 psubsw_r2r(mm6, mm7); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
328 movq_r2m(mm0, *(out + offset + 1 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
329 paddsw_r2r(mm4, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
330 movq_r2m(mm7, *(out + offset + 3 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
331 psubsw_r2r(mm1, mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
332 movq_r2m(mm5, *(out + offset + 5 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
333 movq_r2m(mm3, *(out + offset + 7 * 8)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
334 } |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
335 |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
336 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
337 static always_inline void fdct_row_sse2(const int16_t *in, int16_t *out) |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
338 { |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
339 asm volatile( |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
340 ".macro FDCT_ROW_SSE2_H1 i t \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
341 "movq \\i(%0), %%xmm2 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
342 "movq \\i+8(%0), %%xmm0 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
343 "movdqa \\t+32(%1), %%xmm3 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
344 "movdqa \\t+48(%1), %%xmm7 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
345 "movdqa \\t(%1), %%xmm4 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
346 "movdqa \\t+16(%1), %%xmm5 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
347 ".endm \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
348 ".macro FDCT_ROW_SSE2_H2 i t \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
349 "movq \\i(%0), %%xmm2 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
350 "movq \\i+8(%0), %%xmm0 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
351 "movdqa \\t+32(%1), %%xmm3 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
352 "movdqa \\t+48(%1), %%xmm7 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
353 ".endm \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
354 ".macro FDCT_ROW_SSE2 i \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
355 "movq %%xmm2, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
356 "pshuflw $27, %%xmm0, %%xmm0 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
357 "paddsw %%xmm0, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
358 "psubsw %%xmm0, %%xmm2 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
359 "punpckldq %%xmm2, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
360 "pshufd $78, %%xmm1, %%xmm2 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
361 "pmaddwd %%xmm2, %%xmm3 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
362 "pmaddwd %%xmm1, %%xmm7 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
363 "pmaddwd %%xmm5, %%xmm2 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
364 "pmaddwd %%xmm4, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
365 "paddd %%xmm7, %%xmm3 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
366 "paddd %%xmm2, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
367 "paddd %%xmm6, %%xmm3 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
368 "paddd %%xmm6, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
369 "psrad %3, %%xmm3 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
370 "psrad %3, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
371 "packssdw %%xmm3, %%xmm1 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
372 "movdqa %%xmm1, \\i(%4) \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
373 ".endm \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
374 "movdqa (%2), %%xmm6 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
375 "FDCT_ROW_SSE2_H1 0 0 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
376 "FDCT_ROW_SSE2 0 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
377 "FDCT_ROW_SSE2_H2 64 0 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
378 "FDCT_ROW_SSE2 64 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
379 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
380 "FDCT_ROW_SSE2_H1 16 64 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
381 "FDCT_ROW_SSE2 16 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
382 "FDCT_ROW_SSE2_H2 112 64 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
383 "FDCT_ROW_SSE2 112 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
384 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
385 "FDCT_ROW_SSE2_H1 32 128 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
386 "FDCT_ROW_SSE2 32 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
387 "FDCT_ROW_SSE2_H2 96 128 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
388 "FDCT_ROW_SSE2 96 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
389 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
390 "FDCT_ROW_SSE2_H1 48 192 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
391 "FDCT_ROW_SSE2 48 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
392 "FDCT_ROW_SSE2_H2 80 192 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
393 "FDCT_ROW_SSE2 80 \n\t" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
394 : |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
395 : "r" (in), "r" (tab_frw_01234567_sse2), "r" (fdct_r_row_sse2), "i" (SHIFT_FRW_ROW), "r" (out) |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
396 ); |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
397 } |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
398 |
1570 | 399 static always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table) |
1564
b6b7d080f1a1
inline -> always_inline (842 -> 690 cpu cycles for dct_quantize() difference for the dct itself should be even bigger)
michael
parents:
839
diff
changeset
|
400 { |
1565 | 401 pshufw_m2r(*(in + 4), mm5, 0x1B); |
402 movq_m2r(*(in + 0), mm0); | |
1570 | 403 movq_r2r(mm0, mm1); |
404 paddsw_r2r(mm5, mm0); | |
405 psubsw_r2r(mm5, mm1); | |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
406 pshufw_r2r(mm0, mm2, 0x4E); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
407 pshufw_r2r(mm1, mm3, 0x4E); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
408 movq_m2r(*(table + 0), mm4); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
409 movq_m2r(*(table + 4), mm6); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
410 movq_m2r(*(table + 16), mm5); |
1570 | 411 movq_m2r(*(table + 20), mm7); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
412 pmaddwd_r2r(mm0, mm4); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
413 pmaddwd_r2r(mm1, mm5); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
414 pmaddwd_r2r(mm2, mm6); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
415 pmaddwd_r2r(mm3, mm7); |
1570 | 416 pmaddwd_m2r(*(table + 8), mm0); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
417 pmaddwd_m2r(*(table + 12), mm2); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
418 pmaddwd_m2r(*(table + 24), mm1); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
419 pmaddwd_m2r(*(table + 28), mm3); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
420 paddd_r2r(mm6, mm4); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
421 paddd_r2r(mm7, mm5); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
422 paddd_r2r(mm2, mm0); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
423 paddd_r2r(mm3, mm1); |
1570 | 424 movq_m2r(*fdct_r_row, mm7); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
425 paddd_r2r(mm7, mm4); |
1570 | 426 paddd_r2r(mm7, mm5); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
427 paddd_r2r(mm7, mm0); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
428 paddd_r2r(mm7, mm1); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
429 psrad_i2r(SHIFT_FRW_ROW, mm4); |
1570 | 430 psrad_i2r(SHIFT_FRW_ROW, mm5); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
431 psrad_i2r(SHIFT_FRW_ROW, mm0); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
432 psrad_i2r(SHIFT_FRW_ROW, mm1); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
433 packssdw_r2r(mm0, mm4); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
434 packssdw_r2r(mm1, mm5); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
435 movq_r2r(mm4, mm2); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
436 punpcklwd_r2r(mm5, mm4); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
437 punpckhwd_r2r(mm5, mm2); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
438 movq_r2m(mm4, *(out + 0)); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
439 movq_r2m(mm2, *(out + 4)); |
1570 | 440 } |
441 | |
442 static always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table) | |
443 { | |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
444 movd_m2r(*(in + 6), mm1); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
445 punpcklwd_m2r(*(in + 4), mm1); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
446 movq_r2r(mm1, mm2); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
447 psrlq_i2r(0x20, mm1); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
448 movq_m2r(*(in + 0), mm0); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
449 punpcklwd_r2r(mm2, mm1); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
450 movq_r2r(mm0, mm5); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
451 paddsw_r2r(mm1, mm0); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
452 psubsw_r2r(mm1, mm5); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
453 movq_r2r(mm0, mm1); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
454 movq_r2r(mm5, mm6); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
455 punpckldq_r2r(mm5, mm3); |
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
456 punpckhdq_r2r(mm3, mm6); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
457 movq_m2r(*(table + 0), mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
458 movq_m2r(*(table + 4), mm4); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
459 punpckldq_r2r(mm0, mm2); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
460 pmaddwd_r2r(mm0, mm3); |
1575
f16ae8e69bd9
reorder table instead of wasting instructions to reorder the input to match the table
michael
parents:
1574
diff
changeset
|
461 punpckhdq_r2r(mm2, mm1); |
1570 | 462 movq_m2r(*(table + 16), mm2); |
463 pmaddwd_r2r(mm1, mm4); | |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
464 pmaddwd_m2r(*(table + 8), mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
465 movq_m2r(*(table + 20), mm7); |
1570 | 466 pmaddwd_r2r(mm5, mm2); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
467 paddd_m2r(*fdct_r_row, mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
468 pmaddwd_r2r(mm6, mm7); |
1570 | 469 pmaddwd_m2r(*(table + 12), mm1); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
470 paddd_r2r(mm4, mm3); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
471 pmaddwd_m2r(*(table + 24), mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
472 pmaddwd_m2r(*(table + 28), mm6); |
1570 | 473 paddd_r2r(mm7, mm2); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
474 paddd_m2r(*fdct_r_row, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
475 psrad_i2r(SHIFT_FRW_ROW, mm3); |
1570 | 476 paddd_m2r(*fdct_r_row, mm2); |
477 paddd_r2r(mm1, mm0); | |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
478 paddd_m2r(*fdct_r_row, mm5); |
1570 | 479 psrad_i2r(SHIFT_FRW_ROW, mm2); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
480 paddd_r2r(mm6, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
481 psrad_i2r(SHIFT_FRW_ROW, mm0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
482 psrad_i2r(SHIFT_FRW_ROW, mm5); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
483 packssdw_r2r(mm0, mm3); |
1570 | 484 packssdw_r2r(mm5, mm2); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
485 movq_r2r(mm3, mm6); |
1570 | 486 punpcklwd_r2r(mm2, mm3); |
487 punpckhwd_r2r(mm2, mm6); | |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
488 movq_r2m(mm3, *(out + 0)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
489 movq_r2m(mm6, *(out + 4)); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
490 } |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
491 |
687
9abb13c21fbe
fdct_mmx -> ff_fdct_mmx (renamed to avoid namespace conflict with xvid)
arpi_esp
parents:
635
diff
changeset
|
492 void ff_fdct_mmx(int16_t *block) |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
493 { |
839 | 494 int64_t align_tmp[16] ATTR_ALIGN(8); |
495 int16_t * const block_tmp= (int16_t*)align_tmp; | |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
496 int16_t *block1, *out; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
497 const int16_t *table; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
498 int i; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
499 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
500 block1 = block_tmp; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
501 fdct_col(block, block1, 0); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
502 fdct_col(block, block1, 4); |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
503 |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
504 block1 = block_tmp; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
505 table = tab_frw_01234567; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
506 out = block; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
507 for(i=8;i>0;i--) { |
1570 | 508 fdct_row_mmx(block1, out, table); |
72
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
509 block1 += 8; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
510 table += 32; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
511 out += 8; |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
512 } |
3049d6d452a3
suppressed nasm dependancy - rewrote forward DCT and motion estimation code
glantau
parents:
diff
changeset
|
513 } |
1565 | 514 |
515 void ff_fdct_mmx2(int16_t *block) | |
516 { | |
517 int64_t align_tmp[16] ATTR_ALIGN(8); | |
518 int16_t * const block_tmp= (int16_t*)align_tmp; | |
519 int16_t *block1, *out; | |
520 const int16_t *table; | |
521 int i; | |
522 | |
523 block1 = block_tmp; | |
524 fdct_col(block, block1, 0); | |
525 fdct_col(block, block1, 4); | |
526 | |
527 block1 = block_tmp; | |
528 table = tab_frw_01234567; | |
529 out = block; | |
530 for(i=8;i>0;i--) { | |
1570 | 531 fdct_row_mmx2(block1, out, table); |
1565 | 532 block1 += 8; |
533 table += 32; | |
534 out += 8; | |
535 } | |
536 } | |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
537 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
538 void ff_fdct_sse2(int16_t *block) |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
539 { |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
540 int64_t align_tmp[16] ATTR_ALIGN(8); |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
541 int16_t * const block_tmp= (int16_t*)align_tmp; |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
542 int16_t *block1; |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
543 int i; |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
544 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
545 block1 = block_tmp; |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
546 fdct_col(block, block1, 0); |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
547 fdct_col(block, block1, 4); |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
548 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
549 fdct_row_sse2(block1, block); |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
550 } |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
551 |