b17618c0 |
1 | /* |
2 | * (C) GraÅžvydas "notaz" Ignotas, 2011 |
3 | * |
4 | * This work is licensed under the terms of any of these licenses |
5 | * (at your option): |
6 | * - GNU GPL, version 2 or later. |
7 | * - GNU LGPL, version 2.1 or later. |
8 | * See the COPYING file in the top-level directory. |
9 | */ |
10 | |
665f33e1 |
11 | #include "arm_features.h" |
b17618c0 |
12 | |
13 | .text |
14 | .align 2 |
15 | |
c67af2ac |
16 | .macro load_varadr reg var |
17 | #if defined(__ARM_ARCH_7A__) && !defined(__PIC__) |
5c6457c3 |
18 | movw \reg, #:lower16:EVAR(\var) |
19 | movt \reg, #:upper16:EVAR(\var) |
c67af2ac |
20 | #else |
5c6457c3 |
21 | ldr \reg, =EVAR(\var) |
c67af2ac |
22 | #endif |
23 | .endm |
24 | |
25 | #ifdef __ARM_NEON__ |
b17618c0 |
26 | |
5c6457c3 |
27 | FUNCTION(mix_chan): @ (int start, int count, int lv, int rv) |
b17618c0 |
28 | vmov.32 d14[0], r2 |
29 | vmov.32 d14[1], r3 @ multipliers |
30 | mov r12, r0 |
c67af2ac |
31 | load_varadr r0, ChanBuf |
32 | load_varadr r2, SSumLR |
b17618c0 |
33 | add r0, r12, lsl #2 |
34 | add r2, r12, lsl #3 |
35 | 0: |
36 | vldmia r0!, {d0-d1} |
37 | vldmia r2, {d2-d5} |
38 | vmul.s32 d10, d14, d0[0] |
39 | vmul.s32 d11, d14, d0[1] |
40 | vmul.s32 d12, d14, d1[0] |
41 | vmul.s32 d13, d14, d1[1] |
42 | vsra.s32 q1, q5, #14 |
43 | vsra.s32 q2, q6, #14 |
44 | subs r1, #4 |
45 | blt mc_finish |
46 | vstmia r2!, {d2-d5} |
47 | bgt 0b |
48 | nop |
49 | bxeq lr |
50 | |
51 | mc_finish: |
52 | vstmia r2!, {d2} |
587fa7de |
53 | cmp r1, #-2 |
b17618c0 |
54 | vstmiage r2!, {d3} |
587fa7de |
55 | cmp r1, #-1 |
b17618c0 |
56 | vstmiage r2!, {d4} |
57 | bx lr |
58 | |
59 | |
5c6457c3 |
60 | FUNCTION(mix_chan_rvb): @ (int start, int count, int lv, int rv) |
b17618c0 |
61 | vmov.32 d14[0], r2 |
62 | vmov.32 d14[1], r3 @ multipliers |
63 | mov r12, r0 |
c67af2ac |
64 | load_varadr r0, ChanBuf |
65 | load_varadr r3, sRVBStart |
66 | load_varadr r2, SSumLR |
b17618c0 |
67 | ldr r3, [r3] |
68 | add r0, r12, lsl #2 |
69 | add r2, r12, lsl #3 |
70 | add r3, r12, lsl #3 |
71 | 0: |
72 | vldmia r0!, {d0-d1} |
73 | vldmia r2, {d2-d5} |
74 | vldmia r3, {d6-d9} |
75 | vmul.s32 d10, d14, d0[0] |
76 | vmul.s32 d11, d14, d0[1] |
77 | vmul.s32 d12, d14, d1[0] |
78 | vmul.s32 d13, d14, d1[1] |
79 | vsra.s32 q1, q5, #14 |
80 | vsra.s32 q2, q6, #14 |
81 | vsra.s32 q3, q5, #14 |
82 | vsra.s32 q4, q6, #14 |
83 | subs r1, #4 |
84 | blt mcr_finish |
85 | vstmia r2!, {d2-d5} |
86 | vstmia r3!, {d6-d9} |
87 | bgt 0b |
88 | nop |
89 | bxeq lr |
90 | |
91 | mcr_finish: |
92 | vstmia r2!, {d2} |
93 | vstmia r3!, {d6} |
587fa7de |
94 | cmp r1, #-2 |
b17618c0 |
95 | vstmiage r2!, {d3} |
96 | vstmiage r3!, {d7} |
587fa7de |
97 | cmp r1, #-1 |
b17618c0 |
98 | vstmiage r2!, {d4} |
99 | vstmiage r3!, {d8} |
100 | bx lr |
101 | |
665f33e1 |
102 | #elif defined(HAVE_ARMV5) |
3a721c1f |
103 | |
5c6457c3 |
104 | FUNCTION(mix_chan): @ (int start, int count, int lv, int rv) |
3a721c1f |
105 | stmfd sp!, {r4-r8,lr} |
106 | orr r3, r2, r3, lsl #16 |
107 | lsl r3, #1 @ packed multipliers << 1 |
108 | mov r12, r0 |
c67af2ac |
109 | load_varadr r0, ChanBuf |
110 | load_varadr r2, SSumLR |
3a721c1f |
111 | add r0, r12, lsl #2 |
112 | add r2, r12, lsl #3 |
113 | 0: |
114 | ldmia r0!, {r4,r5} |
115 | ldmia r2, {r6-r8,lr} |
116 | lsl r4, #1 @ adjust for mul |
117 | lsl r5, #1 |
118 | smlawb r6, r4, r3, r6 |
119 | smlawt r7, r4, r3, r7 |
120 | smlawb r8, r5, r3, r8 |
121 | smlawt lr, r5, r3, lr |
122 | subs r1, #2 |
123 | blt mc_finish |
124 | stmia r2!, {r6-r8,lr} |
125 | bgt 0b |
126 | ldmeqfd sp!, {r4-r8,pc} |
127 | |
128 | mc_finish: |
129 | stmia r2!, {r6,r7} |
130 | ldmfd sp!, {r4-r8,pc} |
131 | |
132 | |
5c6457c3 |
133 | FUNCTION(mix_chan_rvb): @ (int start, int count, int lv, int rv) |
3a721c1f |
134 | stmfd sp!, {r4-r8,lr} |
135 | orr lr, r2, r3, lsl #16 |
136 | lsl lr, #1 |
c67af2ac |
137 | load_varadr r3, sRVBStart |
138 | load_varadr r2, SSumLR |
139 | load_varadr r4, ChanBuf |
3a721c1f |
140 | ldr r3, [r3] |
141 | add r2, r2, r0, lsl #3 |
142 | add r3, r3, r0, lsl #3 |
143 | add r0, r4, r0, lsl #2 |
144 | 0: |
145 | ldr r4, [r0], #4 |
146 | ldmia r2, {r6,r7} |
147 | ldmia r3, {r8,r12} |
148 | lsl r4, #1 |
149 | smlawb r6, r4, lr, r6 @ supposedly takes single cycle? |
150 | smlawt r7, r4, lr, r7 |
151 | smlawb r8, r4, lr, r8 |
152 | smlawt r12,r4, lr, r12 |
153 | subs r1, #1 |
154 | stmia r2!, {r6,r7} |
155 | stmia r3!, {r8,r12} |
156 | bgt 0b |
157 | ldmfd sp!, {r4-r8,pc} |
158 | |
c67af2ac |
159 | #endif |
3a721c1f |
160 | |
b17618c0 |
161 | @ vim:filetype=armasm |