annotate mp3lib/dct64_k7.s @ 1173:3c53cbf53e7e

Better 3dnow! optimization
author nickols_k
date Wed, 20 Jun 2001 07:54:19 +0000
parents ee303142c2e0
children 03b7e2955a20
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
1 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
2 /// Replacement of dct64() with AMD's 3DNowEx(DSP)! SIMD operations support
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
3 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
4 /// This code based 'dct64_3dnow.s' by Syuuhei Kashiyama
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
5 /// <squash@mb.kcom.ne.jp>,only some types of changes have been made:
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
6 ///
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
7 /// - added new opcodes PSWAPD, PFPNACC
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
8 /// - decreased number of opcodes (as it was suggested by k7 manual)
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
9 /// (using memory reference as operand of instructions)
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
10 /// - Phase 6 is rewritten with mixing of cpu and mmx opcodes
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
11 /// - change function name for support 3DNowEx! automatic detect
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
12 /// - negation of 3dnow reg was replaced with PXOR 0x800000000, MMi instead
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
13 /// of PFMUL as it was suggested by athlon manual. (Two not separated PFMUL
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
14 /// can not be paired, but PXOR can be).
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
15 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
16 /// note: because K7 processors are an aggresive out-of-order three-way
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
17 /// superscalar ones instruction order is not significand for them.
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
18 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
19 /// Modified by Nick Kurshev <nickols_k@mail.ru>
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
20 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
21 /// The author of this program disclaim whole expressed or implied
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
22 /// warranties with regard to this program, and in no event shall the
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
23 /// author of this program liable to whatever resulted from the use of
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
24 /// this program. Use it at your own risk.
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
25 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
26
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
27 .data
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
28 .align 8
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
29 plus_minus_3dnow: .long 0x00000000, 0x80000000
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
30
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
31 .text
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
32 .globl dct64_3dnowex
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
33 .type dct64_3dnowex,@function
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
34
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
35 /* Discrete Cosine Tansform (DCT) for subband synthesis */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
36 /* void dct64(real *a,real *b,real *c) */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
37 dct64_3dnowex:
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
38 subl $256,%esp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
39 pushl %ebp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
40 pushl %edi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
41 pushl %esi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
42 pushl %ebx
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
43 leal 16(%esp),%ebx /* ebx -> real tmp1[32] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
44 movl 284(%esp),%edi /* edi -> c */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
45 movl 276(%esp),%ebp /* ebp -> a */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
46 movl 280(%esp),%edx /* edx -> b */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
47 leal 128(%ebx),%esi /* esi -> real tmp2[32] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
48
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
49 / femms
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
50
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
51 // 1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
52 movl pnts,%eax
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
53
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
54 movq 0(%edi),%mm0 /* mm0 = c[0x00] | c[0x01]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
55 movq %mm0,%mm1 /* mm1 = mm0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
56 movd 124(%edi),%mm2 /* mm2 = c[0x1f] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
57 punpckldq 120(%edi),%mm2 /* mm2 = c[0x1f] | c[0x1E] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
58 pfadd %mm2,%mm0 /* mm0 = c[0x00]+c[0x1F] | c[0x1E]+c[0x01] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
59 movq %mm0,0(%ebx) /* tmp[0, 1] = mm0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
60 pfsub %mm2,%mm1 /* c[0x00]-c[0x1f] | c[0x01]-c[0x1e] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
61 pfmul 0(%eax),%mm1 /* (c[0x00]-c[0x1f])*pnts[0]|(c[0x01]-c[0x1e])*pnts[1]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
62 pswapd %mm1, %mm1 /* (c[0x01]-c[0x1e])*pnts[1]|(c[0x00]-c[0x1f])*pnts[0]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
63 movq %mm1, 120(%ebx) /* tmp1[30, 31]=mm1 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
64
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
65 movq 8(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
66 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
67 movd 116(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
68 punpckldq 112(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
69 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
70 movq %mm4,8(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
71 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
72 pfmul 8(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
73 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
74 movq %mm5, 112(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
75
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
76 movq 16(%edi),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
77 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
78 movd 108(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
79 punpckldq 104(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
80 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
81 movq %mm0,16(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
82 pfsub %mm2,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
83 pfmul 16(%eax),%mm1
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
84 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
85 movq %mm1, 104(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
86
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
87 movq 24(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
88 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
89 movd 100(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
90 punpckldq 96(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
91 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
92 movq %mm4,24(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
93 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
94 pfmul 24(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
95 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
96 movq %mm5, 96(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
97
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
98 movq 32(%edi),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
99 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
100 movd 92(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
101 punpckldq 88(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
102 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
103 movq %mm0,32(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
104 pfsub %mm2,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
105 pfmul 32(%eax),%mm1
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
106 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
107 movq %mm1, 88(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
108
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
109 movq 40(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
110 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
111 movd 84(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
112 punpckldq 80(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
113 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
114 movq %mm4,40(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
115 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
116 pfmul 40(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
117 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
118 movq %mm5, 80(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
119
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
120 movq 48(%edi),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
121 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
122 movd 76(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
123 punpckldq 72(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
124 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
125 movq %mm0,48(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
126 pfsub %mm2,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
127 pfmul 48(%eax),%mm1
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
128 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
129 movq %mm1, 72(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
130
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
131 movq 56(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
132 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
133 movd 68(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
134 punpckldq 64(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
135 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
136 movq %mm4,56(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
137 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
138 pfmul 56(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
139 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
140 movq %mm5, 64(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
141
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
142 // 2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
143 movl pnts+4,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
144 / 0, 14
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
145 movq 0(%ebx),%mm0 /* mm0 = tmp1[0] | tmp1[1] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
146 movq %mm0,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
147 movd 60(%ebx),%mm2 /* mm2 = tmp1[0x0F] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
148 punpckldq 56(%ebx),%mm2 /* mm2 = tmp1[0x0E] | tmp1[0x0F] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
149 movq 0(%eax),%mm3 /* mm3 = pnts[0] | pnts[1] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
150 pfadd %mm2,%mm0 /* mm0 = tmp1[0]+tmp1[0x0F]|tmp1[1]+tmp1[0x0E]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
151 movq %mm0,0(%esi) /* tmp2[0, 1] = mm0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
152 pfsub %mm2,%mm1 /* mm1 = tmp1[0]-tmp1[0x0F]|tmp1[1]-tmp1[0x0E]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
153 pfmul %mm3,%mm1 /* mm1 = (tmp1[0]-tmp1[0x0F])*pnts[0]|(tmp1[1]-tmp1[0x0E])*pnts[1]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
154 pswapd %mm1, %mm1 /* mm1 = (tmp1[1]-tmp1[0x0E])*pnts[1]|(tmp1[0]-tmp1[0x0F])*pnts[0]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
155 movq %mm1, 56(%esi) /* tmp2[0x0E, 0x0F] = mm1 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
156 / 16, 30
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
157 movq 64(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
158 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
159 movd 124(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
160 punpckldq 120(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
161 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
162 movq %mm0,64(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
163 pfsubr %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
164 pfmul %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
165 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
166 movq %mm1, 120(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
167 movq 8(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
168 / 2, 12
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
169 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
170 movd 52(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
171 punpckldq 48(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
172 movq 8(%eax),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
173 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
174 movq %mm4,8(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
175 pfsub %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
176 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
177 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
178 movq %mm5, 48(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
179 movq 72(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
180 / 18, 28
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
181 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
182 movd 116(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
183 punpckldq 112(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
184 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
185 movq %mm4,72(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
186 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
187 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
188 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
189 movq %mm5, 112(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
190 movq 16(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
191 / 4, 10
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
192 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
193 movd 44(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
194 punpckldq 40(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
195 movq 16(%eax),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
196 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
197 movq %mm0,16(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
198 pfsub %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
199 pfmul %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
200 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
201 movq %mm1, 40(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
202 movq 80(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
203 / 20, 26
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
204 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
205 movd 108(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
206 punpckldq 104(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
207 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
208 movq %mm0,80(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
209 pfsubr %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
210 pfmul %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
211 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
212 movq %mm1, 104(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
213 movq 24(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
214 / 6, 8
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
215 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
216 movd 36(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
217 punpckldq 32(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
218 movq 24(%eax),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
219 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
220 movq %mm4,24(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
221 pfsub %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
222 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
223 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
224 movq %mm5, 32(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
225 movq 88(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
226 / 22, 24
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
227 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
228 movd 100(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
229 punpckldq 96(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
230 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
231 movq %mm4,88(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
232 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
233 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
234 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
235 movq %mm5, 96(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
236
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
237 // 3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
238 movl pnts+8,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
239 movq 0(%eax),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
240 movq 8(%eax),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
241 movq 0(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
242 / 0, 6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
243 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
244 movd 28(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
245 punpckldq 24(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
246 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
247 pfsub %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
248 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
249 movq %mm2,0(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
250 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
251 movq %mm3, 24(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
252 movq 8(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
253 / 2, 4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
254 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
255 movd 20(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
256 punpckldq 16(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
257 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
258 pfsub %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
259 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
260 movq %mm5,8(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
261 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
262 movq %mm6, 16(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
263 movq 32(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
264 / 8, 14
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
265 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
266 movd 60(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
267 punpckldq 56(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
268 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
269 pfsubr %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
270 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
271 movq %mm2,32(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
272 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
273 movq %mm3, 56(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
274 movq 40(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
275 / 10, 12
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
276 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
277 movd 52(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
278 punpckldq 48(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
279 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
280 pfsubr %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
281 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
282 movq %mm5,40(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
283 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
284 movq %mm6, 48(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
285 movq 64(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
286 / 16, 22
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
287 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
288 movd 92(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
289 punpckldq 88(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
290 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
291 pfsub %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
292 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
293 movq %mm2,64(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
294 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
295 movq %mm3, 88(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
296 movq 72(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
297 / 18, 20
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
298 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
299 movd 84(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
300 punpckldq 80(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
301 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
302 pfsub %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
303 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
304 movq %mm5,72(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
305 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
306 movq %mm6, 80(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
307 movq 96(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
308 / 24, 30
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
309 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
310 movd 124(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
311 punpckldq 120(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
312 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
313 pfsubr %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
314 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
315 movq %mm2,96(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
316 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
317 movq %mm3, 120(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
318 movq 104(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
319 / 26, 28
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
320 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
321 movd 116(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
322 punpckldq 112(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
323 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
324 pfsubr %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
325 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
326 movq %mm5,104(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
327 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
328 movq %mm6, 112(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
329
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
330 // 4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
331 movl pnts+12,%eax
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
332 movq 0(%eax),%mm0 /* mm0 = pnts[3] | pnts[4] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
333 movq 0(%ebx),%mm1 /* mm1 = tmp1[0] | tmp1[1] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
334 / 0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
335 movq %mm1,%mm2
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
336 movd 12(%ebx),%mm3 /* mm3 = tmp1[3] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
337 punpckldq 8(%ebx),%mm3 /* mm3 = tmp1[3] | tmp1[2] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
338 pfadd %mm3,%mm1 /* mm1 = tmp1[0]+tmp1[3] | tmp1[1]+tmp1[2]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
339 pfsub %mm3,%mm2 /* mm2 = tmp1[0]-tmp1[3] | tmp1[0]-tmp1[2]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
340 pfmul %mm0,%mm2 /* mm2 = tmp1[0]-tmp1[3]*pnts[3]|tmp1[0]-tmp1[2]*pnts[4]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
341 movq %mm1,0(%esi) /* tmp2[0, 1] = mm1 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
342 pswapd %mm2, %mm2 /* mm2 = tmp1[0]-tmp1[2]*pnts[4]|tmp1[0]-tmp1[3]*pnts[3] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
343 movq %mm2, 8(%esi) /* tmp2[2, 3] = mm2 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
344 movq 16(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
345 / 4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
346 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
347 movd 28(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
348 punpckldq 24(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
349 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
350 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
351 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
352 movq %mm4,16(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
353 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
354 movq %mm5, 24(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
355 movq 32(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
356 / 8
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
357 movq %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
358 movd 44(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
359 punpckldq 40(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
360 pfadd %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
361 pfsub %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
362 pfmul %mm0,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
363 movq %mm1,32(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
364 pswapd %mm2, %mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
365 movq %mm2, 40(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
366 movq 48(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
367 / 12
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
368 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
369 movd 60(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
370 punpckldq 56(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
371 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
372 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
373 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
374 movq %mm4,48(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
375 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
376 movq %mm5, 56(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
377 movq 64(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
378 / 16
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
379 movq %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
380 movd 76(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
381 punpckldq 72(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
382 pfadd %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
383 pfsub %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
384 pfmul %mm0,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
385 movq %mm1,64(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
386 pswapd %mm2, %mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
387 movq %mm2, 72(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
388 movq 80(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
389 / 20
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
390 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
391 movd 92(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
392 punpckldq 88(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
393 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
394 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
395 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
396 movq %mm4,80(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
397 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
398 movq %mm5, 88(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
399 movq 96(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
400 / 24
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
401 movq %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
402 movd 108(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
403 punpckldq 104(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
404 pfadd %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
405 pfsub %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
406 pfmul %mm0,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
407 movq %mm1,96(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
408 pswapd %mm2, %mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
409 movq %mm2, 104(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
410 movq 112(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
411 / 28
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
412 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
413 movd 124(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
414 punpckldq 120(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
415 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
416 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
417 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
418 movq %mm4,112(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
419 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
420 movq %mm5, 120(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
421
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
422 // 5
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
423 movq plus_minus_3dnow, %mm0 /* mm0 = 1.0 | -1.0 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
424 movl $1,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
425 movd %eax,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
426 pi2fd %mm1,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
427 movl pnts+16,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
428 movd 0(%eax),%mm2
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
429 punpckldq %mm2,%mm1 /* mm1 = 1.0 | cos0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
430 movq 0(%esi),%mm2 /* mm2 = tmp2[0] | tmp2[1] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
431 / 0
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
432 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
433 pswapd %mm2, %mm2 /* mm2 = tmp2[0]+tmp2[1]|tmp2[0]-tmp2[1]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
434 pfmul %mm1,%mm2 /* mm2 = tmp2[0]+tmp2[1]|(tmp2[0]-tmp2[1])*cos0*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
435 movq %mm2,0(%ebx) /* tmp1[0, 1] = mm2 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
436 movq 8(%esi),%mm4 /* mm4 = tmp2[2] | tmp2[3]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
437 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
438 pswapd %mm4, %mm4 /* mm4 = tmp2[2]+tmp2[3]|tmp2[2]-tmp2[3]*/
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
439 pxor %mm0,%mm4 /* mm4 = tmp2[2]+tmp2[3]|tmp2[3]-tmp2[2]*/
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
440 pfmul %mm1,%mm4 /* mm4 = tmp2[2]+tmp2[3]|(tmp2[3]-tmp2[2])*cos0*/
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
441 movq %mm4,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
442 psrlq $32,%mm5 /* mm5 = (tmp2[3]-tmp2[2])*cos0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
443 pfacc %mm5,%mm4 /* mm4 = tmp2[2]+tmp2[3]+(tmp2[3]-tmp2[2])*cos0|(tmp2[3]-tmp2[2])*cos0*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
444 movq %mm4,8(%ebx) /* tmp1[2, 3] = mm4 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
445 movq 16(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
446 / 4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
447 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
448 pswapd %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
449
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
450 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
451 movq 24(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
452 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
453 pswapd %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
454
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
455 pxor %mm0,%mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
456 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
457 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
458 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
459 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
460 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
461 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
462 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
463 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
464 movq %mm2,16(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
465 movq %mm4,24(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
466 movq 32(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
467 / 8
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
468 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
469 pswapd %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
470
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
471 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
472 movq %mm2,32(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
473 movq 40(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
474 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
475 pswapd %mm4, %mm4
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
476 pxor %mm0,%mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
477 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
478 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
479 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
480 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
481 movq %mm4,40(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
482 movq 48(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
483 / 12
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
484 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
485 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
486 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
487 movq 56(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
488 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
489 pswapd %mm4, %mm4
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
490 pxor %mm0,%mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
491 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
492 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
493 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
494 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
495 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
496 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
497 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
498 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
499 movq %mm2,48(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
500 movq %mm4,56(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
501 movq 64(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
502 / 16
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
503 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
504 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
505 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
506 movq %mm2,64(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
507 movq 72(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
508 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
509 pswapd %mm4, %mm4
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
510 pxor %mm0,%mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
511 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
512 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
513 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
514 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
515 movq %mm4,72(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
516 movq 80(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
517 / 20
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
518 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
519 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
520 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
521 movq 88(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
522 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
523 pswapd %mm4, %mm4
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
524 pxor %mm0,%mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
525 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
526 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
527 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
528 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
529 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
530 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
531 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
532 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
533 movq %mm2,80(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
534 movq %mm4,88(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
535 movq 96(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
536 / 24
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
537 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
538 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
539 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
540 movq %mm2,96(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
541 movq 104(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
542 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
543 pswapd %mm4, %mm4
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
544 pxor %mm0,%mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
545 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
546 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
547 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
548 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
549 movq %mm4,104(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
550 movq 112(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
551 / 28
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
552 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
553 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
554 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
555 movq 120(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
556 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
557 pswapd %mm4, %mm4
1173
3c53cbf53e7e Better 3dnow! optimization
nickols_k
parents: 781
diff changeset
558 pxor %mm0,%mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
559 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
560 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
561 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
562 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
563 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
564 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
565 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
566 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
567 movq %mm2,112(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
568 movq %mm4,120(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
569
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
570 // Phase6
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
571 movd 0(%ebx),%mm0
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
572 movd %mm0,1024(%ebp)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
573 movl 4(%ebx),%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
574 movl %eax,0(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
575 movl %eax,0(%edx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
576 movd 8(%ebx),%mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
577 movd %mm2,512(%ebp)
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
578 movd 12(%ebx),%mm3
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
579 movd %mm3,512(%edx)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
580
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
581 movl 16(%ebx),%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
582 movl %eax,768(%ebp)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
583 movd 20(%ebx),%mm5
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
584 movd %mm5,256(%edx)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
585
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
586 movd 24(%ebx),%mm6
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
587 movd %mm6,256(%ebp)
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
588 movd 28(%ebx),%mm7
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
589 movd %mm7,768(%edx)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
590
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
591 movq 32(%ebx),%mm0 /* mm0 = tmp1[8] | tmp1[9] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
592 movq 48(%ebx),%mm1 /* mm1 = tmp1[12] | tmp1[13] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
593 pfadd %mm1,%mm0 /* mm0 = tmp1[8]+tmp1[12]| tmp1[9]+tmp1[13]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
594 movd %mm0,896(%ebp) /* a[0xE0] = tmp1[8]+tmp1[12] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
595 psrlq $32,%mm0
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
596 movd %mm0,128(%edx) /* a[0x20] = tmp1[9]+tmp1[13] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
597 movq 40(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
598 pfadd %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
599 movd %mm1,640(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
600 psrlq $32,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
601 movd %mm1,384(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
602
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
603 movq 56(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
604 pfadd %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
605 movd %mm2,384(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
606 psrlq $32,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
607 movd %mm2,640(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
608
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
609 movd 36(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
610 pfadd %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
611 movd %mm3,128(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
612 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
613 movd %mm3,896(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
614 movq 96(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
615 movq 64(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
616
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
617 movq 112(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
618 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
619 movq %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
620 pfadd %mm1,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
621 movd %mm3,960(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
622 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
623 movd %mm3,64(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
624 movq 80(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
625 pfadd %mm1,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
626 movd %mm0,832(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
627 psrlq $32,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
628 movd %mm0,192(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
629 movq 104(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
630 pfadd %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
631 movq %mm2,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
632 pfadd %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
633 movd %mm4,704(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
634 psrlq $32,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
635 movd %mm4,320(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
636 movq 72(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
637 pfadd %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
638 movd %mm2,576(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
639 psrlq $32,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
640 movd %mm2,448(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
641
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
642 movq 120(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
643 pfadd %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
644 movq %mm3,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
645 pfadd %mm1,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
646 movd %mm5,448(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
647 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
648 movd %mm5,576(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
649 movq 88(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
650 pfadd %mm1,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
651 movd %mm3,320(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
652 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
653 movd %mm3,704(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
654
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
655 movd 100(%ebx),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
656 pfadd %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
657 movq %mm4,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
658 pfadd %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
659 movd %mm6,192(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
660 psrlq $32,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
661 movd %mm6,832(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
662 movd 68(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
663 pfadd %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
664 movd %mm4,64(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
665 psrlq $32,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
666 movd %mm4,960(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
667
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
668 / femms
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
669
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
670 popl %ebx
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
671 popl %esi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
672 popl %edi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
673 popl %ebp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
674 addl $256,%esp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
675
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
676 ret $12
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
677