Bill Yi | 4e213d5 | 2015-06-23 13:53:11 -0700 | [diff] [blame] | 1 | /* Compute the sum of the squares of a vector of signed shorts |
| 2 | |
| 3 | * This is the Altivec SIMD version. It's a little hairy because Altivec |
| 4 | * does not do 64-bit operations directly, so we have to accumulate separate |
| 5 | * 32-bit sums and carries |
| 6 | |
| 7 | * Copyright 2004 Phil Karn, KA9Q |
| 8 | * May be used under the terms of the GNU Lesser General Public License (LGPL) |
| 9 | */ |
| 10 | |
| 11 | #include "fec.h" |
| 12 | |
| 13 | unsigned long long sumsq_av(signed short *in,int cnt){ |
| 14 | long long sum; |
| 15 | vector signed short x; |
| 16 | vector unsigned int sums,carries,s1,s2; |
| 17 | int pad; |
| 18 | union { vector unsigned char cv; vector unsigned int iv; unsigned int w[4]; unsigned char c[16];} s; |
| 19 | |
| 20 | carries = sums = (vector unsigned int)(0); |
| 21 | if((pad = (int)in & 15)!=0){ |
| 22 | /* Load unaligned leading word */ |
| 23 | x = vec_perm(vec_ld(0,in),(vector signed short)(0),vec_lvsl(0,in)); |
| 24 | if(cnt < 8){ /* Shift right to chop stuff beyond end of short block */ |
| 25 | s.c[15] = (8-cnt)<<4; |
| 26 | x = vec_sro(x,s.cv); |
| 27 | } |
| 28 | sums = (vector unsigned int)vec_msum(x,x,(vector signed int)(0)); |
| 29 | in += 8-pad/2; |
| 30 | cnt -= 8-pad/2; |
| 31 | } |
| 32 | /* Everything is now aligned, rip through most of the block */ |
| 33 | while(cnt >= 8){ |
| 34 | x = vec_ld(0,in); |
| 35 | /* A single vec_msum cannot overflow, but we have to sum it with |
| 36 | * the earlier terms separately to handle the carries |
| 37 | * The cast to unsigned is OK because squares are always positive |
| 38 | */ |
| 39 | s1 = (vector unsigned int)vec_msum(x,x,(vector signed int)(0)); |
| 40 | carries = vec_add(carries,vec_addc(sums,s1)); |
| 41 | sums = vec_add(sums,s1); |
| 42 | in += 8; |
| 43 | cnt -= 8; |
| 44 | } |
| 45 | /* Handle trailing fragment, if any */ |
| 46 | if(cnt > 0){ |
| 47 | x = vec_ld(0,in); |
| 48 | s.c[15] = (8-cnt)<<4; |
| 49 | x = vec_sro(x,s.cv); |
| 50 | s1 = (vector unsigned int)vec_msum(x,x,(vector signed int)(0)); |
| 51 | carries = vec_add(carries,vec_addc(sums,s1)); |
| 52 | sums = vec_add(sums,s1); |
| 53 | } |
| 54 | /* Combine 4 sub-sums and carries */ |
| 55 | s.c[15] = 64; /* Shift right two 32-bit words */ |
| 56 | s1 = vec_sro(sums,s.cv); |
| 57 | s2 = vec_sro(carries,s.cv); |
| 58 | carries = vec_add(carries,vec_addc(sums,s1)); |
| 59 | sums = vec_add(sums,s1); |
| 60 | carries = vec_add(carries,s2); |
| 61 | |
| 62 | s.c[15] = 32; /* Shift right one 32-bit word */ |
| 63 | s1 = vec_sro(sums,s.cv); |
| 64 | s2 = vec_sro(carries,s.cv); |
| 65 | carries = vec_add(carries,vec_addc(sums,s1)); |
| 66 | sums = vec_add(sums,s1); |
| 67 | carries = vec_add(carries,s2); |
| 68 | |
| 69 | /* Extract sum and carries from right-hand words and combine into result */ |
| 70 | s.iv = sums; |
| 71 | sum = s.w[3]; |
| 72 | |
| 73 | s.iv = carries; |
| 74 | sum += (long long)s.w[3] << 32; |
| 75 | |
| 76 | return sum; |
| 77 | } |
| 78 | |