]>
Commit | Line | Data |
---|---|---|
81f80d8c MF |
1 | typedef long int int64_t; |
2 | typedef int64_t pa_reg_x86; | |
3 | void pa_volume_s16ne_mmx(short *samples, int *volumes, unsigned channels, | |
4 | unsigned length) | |
5 | { | |
6 | long int channel; | |
7 | long int temp; | |
8 | __asm__ __volatile__(" xor %3, %3 \n\t" | |
9 | " sar $1, %2 \n\t" | |
10 | " test $1, %2 \n\t" | |
11 | " je 2f \n\t" | |
12 | " movd (%1, %3, 4), %%mm0 \n\t" | |
13 | " movw (%0), %w4 \n\t" | |
14 | " movd %4, %%mm1 \n\t" | |
15 | " pxor %%mm4, %%mm4 \n\t" | |
16 | " punpcklwd %%mm4, " "%%mm1" " \n\t" | |
17 | " pcmpgtw " "%%mm0" ", %%mm4 \n\t" | |
18 | " pand " "%%mm1" ", %%mm4 \n\t" | |
19 | " movq " "%%mm1" ", %%mm5 \n\t" | |
20 | " pmulhw " "%%mm0" ", " "%%mm1" " \n\t" | |
21 | " paddw %%mm4, " "%%mm1" " \n\t" | |
22 | " psrld $16, " "%%mm0" " \n\t" | |
23 | " pmaddwd %%mm5, " "%%mm0" " \n\t" | |
24 | " paddd " "%%mm1" ", " "%%mm0" " \n\t" | |
25 | " packssdw " "%%mm0" ", " "%%mm0" " \n\t" | |
26 | " movd %%mm0, %4 \n\t" | |
27 | " movw %w4, (%0) \n\t" | |
28 | " add $2, %0 \n\t" " add " "$1" | |
29 | ", %3 \n\t" | |
30 | " mov %3, %4 \n\t" " sub " "%5" | |
31 | ", %4 \n\t" | |
32 | " cmovae %4, %3 \n\t" | |
33 | "2: \n\t" | |
34 | " sar $1, %2 \n\t" | |
35 | " test $1, %2 \n\t" | |
36 | " je 4f \n\t" | |
37 | "3: \n\t" | |
38 | " movq (%1, %3, 4), %%mm0 \n\t" | |
39 | " movd (%0), %%mm1 \n\t" | |
40 | " pxor %%mm4, %%mm4 \n\t" | |
41 | " punpcklwd %%mm4, " "%%mm1" " \n\t" | |
42 | " pcmpgtw " "%%mm0" ", %%mm4 \n\t" | |
43 | " pand " "%%mm1" ", %%mm4 \n\t" | |
44 | " movq " "%%mm1" ", %%mm5 \n\t" | |
45 | " pmulhw " "%%mm0" ", " "%%mm1" " \n\t" | |
46 | " paddw %%mm4, " "%%mm1" " \n\t" | |
47 | " psrld $16, " "%%mm0" " \n\t" | |
48 | " pmaddwd %%mm5, " "%%mm0" " \n\t" | |
49 | " paddd " "%%mm1" ", " "%%mm0" " \n\t" | |
50 | " packssdw " "%%mm0" ", " "%%mm0" " \n\t" | |
51 | " movd %%mm0, (%0) \n\t" | |
52 | " add $4, %0 \n\t" " add " "$2" | |
53 | ", %3 \n\t" | |
54 | " mov %3, %4 \n\t" " sub " "%5" | |
55 | ", %4 \n\t" | |
56 | " cmovae %4, %3 \n\t" | |
57 | "4: \n\t" | |
58 | " sar $1, %2 \n\t" | |
59 | " cmp $0, %2 \n\t" | |
60 | " je 6f \n\t" | |
61 | "5: \n\t" | |
62 | " movq (%1, %3, 4), %%mm0 \n\t" | |
63 | " movq 8(%1, %3, 4), %%mm2 \n\t" | |
64 | " movd (%0), %%mm1 \n\t" | |
65 | " movd 4(%0), %%mm3 \n\t" | |
66 | " pxor %%mm4, %%mm4 \n\t" | |
67 | " punpcklwd %%mm4, " "%%mm1" " \n\t" | |
68 | " pcmpgtw " "%%mm0" ", %%mm4 \n\t" | |
69 | " pand " "%%mm1" ", %%mm4 \n\t" | |
70 | " movq " "%%mm1" ", %%mm5 \n\t" | |
71 | " pmulhw " "%%mm0" ", " "%%mm1" " \n\t" | |
72 | " paddw %%mm4, " "%%mm1" " \n\t" | |
73 | " psrld $16, " "%%mm0" " \n\t" | |
74 | " pmaddwd %%mm5, " "%%mm0" " \n\t" | |
75 | " paddd " "%%mm1" ", " "%%mm0" " \n\t" | |
76 | " packssdw " "%%mm0" ", " "%%mm0" " \n\t" | |
77 | " pxor %%mm4, %%mm4 \n\t" | |
78 | " punpcklwd %%mm4, " "%%mm3" " \n\t" | |
79 | " pcmpgtw " "%%mm2" ", %%mm4 \n\t" | |
80 | " pand " "%%mm3" ", %%mm4 \n\t" | |
81 | " movq " "%%mm3" ", %%mm5 \n\t" | |
82 | " pmulhw " "%%mm2" ", " "%%mm3" " \n\t" | |
83 | " paddw %%mm4, " "%%mm3" " \n\t" | |
84 | " psrld $16, " "%%mm2" " \n\t" | |
85 | " pmaddwd %%mm5, " "%%mm2" " \n\t" | |
86 | " paddd " "%%mm3" ", " "%%mm2" " \n\t" | |
87 | " packssdw " "%%mm2" ", " "%%mm2" " \n\t" | |
88 | " movd %%mm0, (%0) \n\t" | |
89 | " movd %%mm2, 4(%0) \n\t" | |
90 | " add $8, %0 \n\t" " add " "$4" | |
91 | ", %3 \n\t" | |
92 | " mov %3, %4 \n\t" " sub " "%5" | |
93 | ", %4 \n\t" | |
94 | " cmovae %4, %3 \n\t" | |
95 | " dec %2 \n\t" | |
96 | " jne 5b \n\t" | |
97 | "6: \n\t" | |
98 | " emms \n\t":"+r" | |
99 | (samples), "+r"(volumes), "+r"(length), | |
100 | "=D"((pa_reg_x86) channel), | |
101 | "=&r"(temp):"X"((pa_reg_x86) channels):"cc"); | |
102 | } |