annotate mp3lib/dct64_k7.s @ 781:ee303142c2e0

improvements.
author nickols_k
date Sun, 13 May 2001 14:36:02 +0000
parents 59b0a9ec8604
children 3c53cbf53e7e
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
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
12 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
13 /// note: because K7 processors are an aggresive out-of-order three-way
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
14 /// superscalar ones instruction order is not significand for them.
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
15 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
16 /// Modified by Nick Kurshev <nickols_k@mail.ru>
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
17 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
18 /// The author of this program disclaim whole expressed or implied
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
19 /// warranties with regard to this program, and in no event shall the
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
20 /// author of this program liable to whatever resulted from the use of
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
21 /// this program. Use it at your own risk.
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
22 ///
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
23
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
24 .globl dct64_3dnowex
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
25 .type dct64_3dnowex,@function
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
26
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
27 /* Discrete Cosine Tansform (DCT) for subband synthesis */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
28 /* void dct64(real *a,real *b,real *c) */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
29 dct64_3dnowex:
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
30 subl $256,%esp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
31 pushl %ebp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
32 pushl %edi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
33 pushl %esi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
34 pushl %ebx
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
35 leal 16(%esp),%ebx /* ebx -> real tmp1[32] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
36 movl 284(%esp),%edi /* edi -> c */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
37 movl 276(%esp),%ebp /* ebp -> a */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
38 movl 280(%esp),%edx /* edx -> b */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
39 leal 128(%ebx),%esi /* esi -> real tmp2[32] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
40
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
41 / femms
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
42
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
43 // 1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
44 movl pnts,%eax
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
45
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
46 movq 0(%edi),%mm0 /* mm0 = c[0x00] | c[0x01]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
47 movq %mm0,%mm1 /* mm1 = mm0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
48 movd 124(%edi),%mm2 /* mm2 = c[0x1f] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
49 punpckldq 120(%edi),%mm2 /* mm2 = c[0x1f] | c[0x1E] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
50 pfadd %mm2,%mm0 /* mm0 = c[0x00]+c[0x1F] | c[0x1E]+c[0x01] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
51 movq %mm0,0(%ebx) /* tmp[0, 1] = mm0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
52 pfsub %mm2,%mm1 /* c[0x00]-c[0x1f] | c[0x01]-c[0x1e] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
53 pfmul 0(%eax),%mm1 /* (c[0x00]-c[0x1f])*pnts[0]|(c[0x01]-c[0x1e])*pnts[1]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
54 pswapd %mm1, %mm1 /* (c[0x01]-c[0x1e])*pnts[1]|(c[0x00]-c[0x1f])*pnts[0]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
55 movq %mm1, 120(%ebx) /* tmp1[30, 31]=mm1 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
56
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
57 movq 8(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
58 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
59 movd 116(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
60 punpckldq 112(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
61 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
62 movq %mm4,8(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
63 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
64 pfmul 8(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
65 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
66 movq %mm5, 112(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
67
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
68 movq 16(%edi),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
69 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
70 movd 108(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
71 punpckldq 104(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
72 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
73 movq %mm0,16(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
74 pfsub %mm2,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
75 pfmul 16(%eax),%mm1
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
76 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
77 movq %mm1, 104(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
78
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
79 movq 24(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
80 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
81 movd 100(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
82 punpckldq 96(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
83 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
84 movq %mm4,24(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
85 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
86 pfmul 24(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
87 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
88 movq %mm5, 96(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
89
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
90 movq 32(%edi),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
91 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
92 movd 92(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
93 punpckldq 88(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
94 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
95 movq %mm0,32(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
96 pfsub %mm2,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
97 pfmul 32(%eax),%mm1
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
98 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
99 movq %mm1, 88(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
100
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
101 movq 40(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
102 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
103 movd 84(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
104 punpckldq 80(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
105 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
106 movq %mm4,40(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
107 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
108 pfmul 40(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
109 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
110 movq %mm5, 80(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
111
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
112 movq 48(%edi),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
113 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
114 movd 76(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
115 punpckldq 72(%edi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
116 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
117 movq %mm0,48(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
118 pfsub %mm2,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
119 pfmul 48(%eax),%mm1
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
120 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
121 movq %mm1, 72(%ebx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
122
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
123 movq 56(%edi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
124 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
125 movd 68(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
126 punpckldq 64(%edi),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
127 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
128 movq %mm4,56(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
129 pfsub %mm6,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
130 pfmul 56(%eax),%mm5
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
131 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
132 movq %mm5, 64(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
133
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
134 // 2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
135 movl pnts+4,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
136 / 0, 14
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
137 movq 0(%ebx),%mm0 /* mm0 = tmp1[0] | tmp1[1] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
138 movq %mm0,%mm1
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
139 movd 60(%ebx),%mm2 /* mm2 = tmp1[0x0F] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
140 punpckldq 56(%ebx),%mm2 /* mm2 = tmp1[0x0E] | tmp1[0x0F] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
141 movq 0(%eax),%mm3 /* mm3 = pnts[0] | pnts[1] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
142 pfadd %mm2,%mm0 /* mm0 = tmp1[0]+tmp1[0x0F]|tmp1[1]+tmp1[0x0E]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
143 movq %mm0,0(%esi) /* tmp2[0, 1] = mm0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
144 pfsub %mm2,%mm1 /* mm1 = tmp1[0]-tmp1[0x0F]|tmp1[1]-tmp1[0x0E]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
145 pfmul %mm3,%mm1 /* mm1 = (tmp1[0]-tmp1[0x0F])*pnts[0]|(tmp1[1]-tmp1[0x0E])*pnts[1]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
146 pswapd %mm1, %mm1 /* mm1 = (tmp1[1]-tmp1[0x0E])*pnts[1]|(tmp1[0]-tmp1[0x0F])*pnts[0]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
147 movq %mm1, 56(%esi) /* tmp2[0x0E, 0x0F] = mm1 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
148 / 16, 30
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
149 movq 64(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
150 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
151 movd 124(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
152 punpckldq 120(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
153 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
154 movq %mm0,64(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
155 pfsubr %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
156 pfmul %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
157 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
158 movq %mm1, 120(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
159 movq 8(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
160 / 2, 12
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
161 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
162 movd 52(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
163 punpckldq 48(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
164 movq 8(%eax),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
165 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
166 movq %mm4,8(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
167 pfsub %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
168 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
169 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
170 movq %mm5, 48(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
171 movq 72(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
172 / 18, 28
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
173 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
174 movd 116(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
175 punpckldq 112(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
176 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
177 movq %mm4,72(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
178 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
179 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
180 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
181 movq %mm5, 112(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
182 movq 16(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
183 / 4, 10
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
184 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
185 movd 44(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
186 punpckldq 40(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
187 movq 16(%eax),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
188 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
189 movq %mm0,16(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
190 pfsub %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
191 pfmul %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
192 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
193 movq %mm1, 40(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
194 movq 80(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
195 / 20, 26
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
196 movq %mm0,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
197 movd 108(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
198 punpckldq 104(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
199 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
200 movq %mm0,80(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
201 pfsubr %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
202 pfmul %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
203 pswapd %mm1, %mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
204 movq %mm1, 104(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
205 movq 24(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
206 / 6, 8
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
207 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
208 movd 36(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
209 punpckldq 32(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
210 movq 24(%eax),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
211 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
212 movq %mm4,24(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
213 pfsub %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
214 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
215 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
216 movq %mm5, 32(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
217 movq 88(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
218 / 22, 24
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
219 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
220 movd 100(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
221 punpckldq 96(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
222 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
223 movq %mm4,88(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
224 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
225 pfmul %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
226 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
227 movq %mm5, 96(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
228
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
229 // 3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
230 movl pnts+8,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
231 movq 0(%eax),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
232 movq 8(%eax),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
233 movq 0(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
234 / 0, 6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
235 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
236 movd 28(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
237 punpckldq 24(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
238 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
239 pfsub %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
240 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
241 movq %mm2,0(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
242 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
243 movq %mm3, 24(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
244 movq 8(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
245 / 2, 4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
246 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
247 movd 20(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
248 punpckldq 16(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
249 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
250 pfsub %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
251 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
252 movq %mm5,8(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
253 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
254 movq %mm6, 16(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
255 movq 32(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
256 / 8, 14
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
257 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
258 movd 60(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
259 punpckldq 56(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
260 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
261 pfsubr %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
262 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
263 movq %mm2,32(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
264 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
265 movq %mm3, 56(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
266 movq 40(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
267 / 10, 12
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
268 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
269 movd 52(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
270 punpckldq 48(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
271 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
272 pfsubr %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
273 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
274 movq %mm5,40(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
275 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
276 movq %mm6, 48(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
277 movq 64(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
278 / 16, 22
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
279 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
280 movd 92(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
281 punpckldq 88(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
282 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
283 pfsub %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
284 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
285 movq %mm2,64(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
286 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
287 movq %mm3, 88(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
288 movq 72(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
289 / 18, 20
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
290 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
291 movd 84(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
292 punpckldq 80(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
293 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
294 pfsub %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
295 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
296 movq %mm5,72(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
297 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
298 movq %mm6, 80(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
299 movq 96(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
300 / 24, 30
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
301 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
302 movd 124(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
303 punpckldq 120(%esi),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
304 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
305 pfsubr %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
306 pfmul %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
307 movq %mm2,96(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
308 pswapd %mm3, %mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
309 movq %mm3, 120(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
310 movq 104(%esi),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
311 / 26, 28
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
312 movq %mm5,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
313 movd 116(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
314 punpckldq 112(%esi),%mm7
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
315 pfadd %mm7,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
316 pfsubr %mm7,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
317 pfmul %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
318 movq %mm5,104(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
319 pswapd %mm6, %mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
320 movq %mm6, 112(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
321
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
322 // 4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
323 movl pnts+12,%eax
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
324 movq 0(%eax),%mm0 /* mm0 = pnts[3] | pnts[4] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
325 movq 0(%ebx),%mm1 /* mm1 = tmp1[0] | tmp1[1] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
326 / 0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
327 movq %mm1,%mm2
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
328 movd 12(%ebx),%mm3 /* mm3 = tmp1[3] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
329 punpckldq 8(%ebx),%mm3 /* mm3 = tmp1[3] | tmp1[2] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
330 pfadd %mm3,%mm1 /* mm1 = tmp1[0]+tmp1[3] | tmp1[1]+tmp1[2]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
331 pfsub %mm3,%mm2 /* mm2 = tmp1[0]-tmp1[3] | tmp1[0]-tmp1[2]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
332 pfmul %mm0,%mm2 /* mm2 = tmp1[0]-tmp1[3]*pnts[3]|tmp1[0]-tmp1[2]*pnts[4]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
333 movq %mm1,0(%esi) /* tmp2[0, 1] = mm1 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
334 pswapd %mm2, %mm2 /* mm2 = tmp1[0]-tmp1[2]*pnts[4]|tmp1[0]-tmp1[3]*pnts[3] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
335 movq %mm2, 8(%esi) /* tmp2[2, 3] = mm2 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
336 movq 16(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
337 / 4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
338 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
339 movd 28(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
340 punpckldq 24(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
341 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
342 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
343 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
344 movq %mm4,16(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
345 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
346 movq %mm5, 24(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
347 movq 32(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
348 / 8
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
349 movq %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
350 movd 44(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
351 punpckldq 40(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
352 pfadd %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
353 pfsub %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
354 pfmul %mm0,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
355 movq %mm1,32(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
356 pswapd %mm2, %mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
357 movq %mm2, 40(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
358 movq 48(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
359 / 12
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
360 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
361 movd 60(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
362 punpckldq 56(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
363 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
364 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
365 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
366 movq %mm4,48(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
367 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
368 movq %mm5, 56(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
369 movq 64(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
370 / 16
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
371 movq %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
372 movd 76(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
373 punpckldq 72(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
374 pfadd %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
375 pfsub %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
376 pfmul %mm0,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
377 movq %mm1,64(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
378 pswapd %mm2, %mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
379 movq %mm2, 72(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
380 movq 80(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
381 / 20
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
382 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
383 movd 92(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
384 punpckldq 88(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
385 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
386 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
387 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
388 movq %mm4,80(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
389 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
390 movq %mm5, 88(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
391 movq 96(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
392 / 24
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
393 movq %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
394 movd 108(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
395 punpckldq 104(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
396 pfadd %mm3,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
397 pfsub %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
398 pfmul %mm0,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
399 movq %mm1,96(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
400 pswapd %mm2, %mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
401 movq %mm2, 104(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
402 movq 112(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
403 / 28
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
404 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
405 movd 124(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
406 punpckldq 120(%ebx),%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
407 pfadd %mm6,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
408 pfsubr %mm6,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
409 pfmul %mm0,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
410 movq %mm4,112(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
411 pswapd %mm5, %mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
412 movq %mm5, 120(%esi)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
413
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
414 // 5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
415 movl $-1,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
416 movd %eax,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
417 movl $1,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
418 movd %eax,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
419 / L | H
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
420 punpckldq %mm1,%mm0
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
421 pi2fd %mm0,%mm0 /* mm0 = 1.0 | -1.0 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
422 movd %eax,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
423 pi2fd %mm1,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
424 movl pnts+16,%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
425 movd 0(%eax),%mm2
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
426 punpckldq %mm2,%mm1 /* mm1 = 1.0 | cos0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
427 movq 0(%esi),%mm2 /* mm2 = tmp2[0] | tmp2[1] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
428 / 0
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
429 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
430 pswapd %mm2, %mm2 /* mm2 = tmp2[0]+tmp2[1]|tmp2[0]-tmp2[1]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
431 pfmul %mm1,%mm2 /* mm2 = tmp2[0]+tmp2[1]|(tmp2[0]-tmp2[1])*cos0*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
432 movq %mm2,0(%ebx) /* tmp1[0, 1] = mm2 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
433 movq 8(%esi),%mm4 /* mm4 = tmp2[2] | tmp2[3]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
434 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
435 pswapd %mm4, %mm4 /* mm4 = tmp2[2]+tmp2[3]|tmp2[2]-tmp2[3]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
436 pfmul %mm0,%mm4 /* mm4 = tmp2[2]+tmp2[3]|tmp2[3]-tmp2[2]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
437 pfmul %mm1,%mm4 /* mm4 = tmp2[2]+tmp2[3]|(tmp2[3]-tmp2[2])*cos0*/
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
438 movq %mm4,%mm5
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
439 psrlq $32,%mm5 /* mm5 = (tmp2[3]-tmp2[2])*cos0 */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
440 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
441 movq %mm4,8(%ebx) /* tmp1[2, 3] = mm4 */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
442 movq 16(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
443 / 4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
444 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
445 pswapd %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
446
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
447 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
448 movq 24(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
449 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
450 pswapd %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
451
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
452 pfmul %mm0,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
453 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
454 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
455 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
456 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
457 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
458 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
459 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
460 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
461 movq %mm2,16(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
462 movq %mm4,24(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
463 movq 32(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
464 / 8
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
465 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
466 pswapd %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
467
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
468 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
469 movq %mm2,32(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
470 movq 40(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
471 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
472 pswapd %mm4, %mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
473 pfmul %mm0,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
474 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
475 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
476 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
477 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
478 movq %mm4,40(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
479 movq 48(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
480 / 12
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
481 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
482 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
483 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
484 movq 56(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
485 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
486 pswapd %mm4, %mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
487 pfmul %mm0,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
488 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
489 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
490 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
491 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
492 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
493 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
494 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
495 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
496 movq %mm2,48(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
497 movq %mm4,56(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
498 movq 64(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
499 / 16
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
500 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
501 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
502 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
503 movq %mm2,64(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
504 movq 72(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
505 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
506 pswapd %mm4, %mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
507 pfmul %mm0,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
508 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
509 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
510 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
511 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
512 movq %mm4,72(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
513 movq 80(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
514 / 20
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
515 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
516 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
517 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
518 movq 88(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
519 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
520 pswapd %mm4, %mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
521 pfmul %mm0,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
522 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
523 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
524 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
525 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
526 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
527 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
528 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
529 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
530 movq %mm2,80(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
531 movq %mm4,88(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
532 movq 96(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
533 / 24
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
534 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
535 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
536 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
537 movq %mm2,96(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
538 movq 104(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
539 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
540 pswapd %mm4, %mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
541 pfmul %mm0,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
542 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
543 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
544 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
545 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
546 movq %mm4,104(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
547 movq 112(%esi),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
548 / 28
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
549 pfpnacc %mm2, %mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
550 pswapd %mm2, %mm2
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
551 pfmul %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
552 movq 120(%esi),%mm4
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
553 pfpnacc %mm4, %mm4
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
554 pswapd %mm4, %mm4
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
555 pfmul %mm0,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
556 pfmul %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
557 movq %mm4,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
558 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
559 pfacc %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
560 movq %mm2,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
561 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
562 pfadd %mm4,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
563 pfadd %mm3,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
564 movq %mm2,112(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
565 movq %mm4,120(%ebx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
566
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
567 // Phase6
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
568 movd 0(%ebx),%mm0
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
569 movd %mm0,1024(%ebp)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
570 movl 4(%ebx),%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
571 movl %eax,0(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
572 movl %eax,0(%edx)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
573 movd 8(%ebx),%mm2
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
574 movd %mm2,512(%ebp)
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
575 movd 12(%ebx),%mm3
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
576 movd %mm3,512(%edx)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
577
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
578 movl 16(%ebx),%eax
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
579 movl %eax,768(%ebp)
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
580 movd 20(%ebx),%mm5
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
581 movd %mm5,256(%edx)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
582
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
583 movd 24(%ebx),%mm6
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
584 movd %mm6,256(%ebp)
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
585 movd 28(%ebx),%mm7
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
586 movd %mm7,768(%edx)
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
587
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
588 movq 32(%ebx),%mm0 /* mm0 = tmp1[8] | tmp1[9] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
589 movq 48(%ebx),%mm1 /* mm1 = tmp1[12] | tmp1[13] */
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
590 pfadd %mm1,%mm0 /* mm0 = tmp1[8]+tmp1[12]| tmp1[9]+tmp1[13]*/
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
591 movd %mm0,896(%ebp) /* a[0xE0] = tmp1[8]+tmp1[12] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
592 psrlq $32,%mm0
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
593 movd %mm0,128(%edx) /* a[0x20] = tmp1[9]+tmp1[13] */
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
594 movq 40(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
595 pfadd %mm2,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
596 movd %mm1,640(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
597 psrlq $32,%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
598 movd %mm1,384(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
599
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
600 movq 56(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
601 pfadd %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
602 movd %mm2,384(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
603 psrlq $32,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
604 movd %mm2,640(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
605
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
606 movd 36(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
607 pfadd %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
608 movd %mm3,128(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
609 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
610 movd %mm3,896(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
611 movq 96(%ebx),%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
612 movq 64(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
613
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
614 movq 112(%ebx),%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
615 pfadd %mm2,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
616 movq %mm0,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
617 pfadd %mm1,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
618 movd %mm3,960(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
619 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
620 movd %mm3,64(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
621 movq 80(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
622 pfadd %mm1,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
623 movd %mm0,832(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
624 psrlq $32,%mm0
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
625 movd %mm0,192(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
626 movq 104(%ebx),%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
627 pfadd %mm3,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
628 movq %mm2,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
629 pfadd %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
630 movd %mm4,704(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
631 psrlq $32,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
632 movd %mm4,320(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
633 movq 72(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
634 pfadd %mm1,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
635 movd %mm2,576(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
636 psrlq $32,%mm2
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
637 movd %mm2,448(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
638
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
639 movq 120(%ebx),%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
640 pfadd %mm4,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
641 movq %mm3,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
642 pfadd %mm1,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
643 movd %mm5,448(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
644 psrlq $32,%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
645 movd %mm5,576(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
646 movq 88(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
647 pfadd %mm1,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
648 movd %mm3,320(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
649 psrlq $32,%mm3
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
650 movd %mm3,704(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
651
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
652 movd 100(%ebx),%mm5
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
653 pfadd %mm5,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
654 movq %mm4,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
655 pfadd %mm1,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
656 movd %mm6,192(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
657 psrlq $32,%mm6
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
658 movd %mm6,832(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
659 movd 68(%ebx),%mm1
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
660 pfadd %mm1,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
661 movd %mm4,64(%ebp)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
662 psrlq $32,%mm4
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
663 movd %mm4,960(%edx)
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
664
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
665 / femms
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
666
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
667 popl %ebx
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
668 popl %esi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
669 popl %edi
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
670 popl %ebp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
671 addl $256,%esp
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
672
781
ee303142c2e0 improvements.
nickols_k
parents: 736
diff changeset
673 ret $12
736
59b0a9ec8604 K7 3dnow-dsp support
nickols_k
parents:
diff changeset
674