Bill Yi | 4e213d5 | 2015-06-23 13:53:11 -0700 | [diff] [blame] | 1 | /* Return the largest absolute value of a vector of signed shorts |
| 2 | |
| 3 | * This is the Altivec SIMD version. |
| 4 | |
| 5 | * Copyright 2004 Phil Karn, KA9Q |
| 6 | * May be used under the terms of the GNU Lesser General Public License (LGPL) |
| 7 | */ |
| 8 | |
| 9 | #include "fec.h" |
| 10 | |
| 11 | signed short peakval_av(signed short *in,int cnt){ |
| 12 | vector signed short x; |
| 13 | int pad; |
| 14 | union { vector signed char cv; vector signed short hv; signed short s[8]; signed char c[16];} s; |
| 15 | vector signed short smallest,largest; |
| 16 | |
| 17 | smallest = (vector signed short)(0); |
| 18 | largest = (vector signed short)(0); |
| 19 | if((pad = (int)in & 15)!=0){ |
| 20 | /* Load unaligned leading word */ |
| 21 | x = vec_perm(vec_ld(0,in),(vector signed short)(0),vec_lvsl(0,in)); |
| 22 | if(cnt < 8){ /* Shift right to chop stuff beyond end of short block */ |
| 23 | s.c[15] = (8-cnt)<<4; |
| 24 | x = vec_sro(x,s.cv); |
| 25 | } |
| 26 | smallest = vec_min(smallest,x); |
| 27 | largest = vec_max(largest,x); |
| 28 | in += 8-pad/2; |
| 29 | cnt -= 8-pad/2; |
| 30 | } |
| 31 | /* Everything is now aligned, rip through most of the block */ |
| 32 | while(cnt >= 8){ |
| 33 | x = vec_ld(0,in); |
| 34 | smallest = vec_min(smallest,x); |
| 35 | largest = vec_max(largest,x); |
| 36 | in += 8; |
| 37 | cnt -= 8; |
| 38 | } |
| 39 | /* Handle trailing fragment, if any */ |
| 40 | if(cnt > 0){ |
| 41 | x = vec_ld(0,in); |
| 42 | s.c[15] = (8-cnt)<<4; |
| 43 | x = vec_sro(x,s.cv); |
| 44 | smallest = vec_min(smallest,x); |
| 45 | largest = vec_max(largest,x); |
| 46 | } |
| 47 | /* Combine and extract result */ |
| 48 | largest = vec_max(largest,vec_abs(smallest)); |
| 49 | |
| 50 | s.c[15] = 64; /* Shift right four 16-bit words */ |
| 51 | largest = vec_max(largest,vec_sro(largest,s.cv)); |
| 52 | |
| 53 | s.c[15] = 32; /* Shift right two 16-bit words */ |
| 54 | largest = vec_max(largest,vec_sro(largest,s.cv)); |
| 55 | |
| 56 | s.c[15] = 16; /* Shift right one 16-bit word */ |
| 57 | largest = vec_max(largest,vec_sro(largest,s.cv)); |
| 58 | |
| 59 | s.hv = largest; |
| 60 | return s.s[7]; |
| 61 | } |