cleaning up viterbi neon

master
yagoda 7 years ago
parent f273ec959b
commit 8751645a48

@ -13,7 +13,6 @@
#include "parity.h"
//#define DEBUG
//#define HAVE_NEON
#ifdef HAVE_NEON
#include <arm_neon.h>
@ -22,8 +21,6 @@ typedef union {
unsigned char c[64];
uint8x16_t v[4];
} metric_t;
typedef union {
unsigned long w[2];
unsigned char c[8];
@ -31,18 +28,16 @@ typedef union {
uint8x8_t v[1];
} decision_t;
union branchtab27{
unsigned char c[32];
unsigned char c[32];
uint8x16_t v[2];
} Branchtab37_neon[3];
int8_t __attribute__((aligned(16))) xr[8];
uint8x8_t mask_and;
int8x8_t mask_shift;
int8_t __attribute__((aligned(16))) xr[8];
uint8x8_t mask_and;
int8x8_t mask_shift;
int firstGo;
/* State info for instance of Viterbi decoder */
struct v37 {
metric_t metrics1; /* path metric buffer 1 */
@ -77,7 +72,6 @@ void clear_v37_neon(struct v37 *vp) {
int init_viterbi37_neon(void *p, int starting_state) {
struct v37 *vp = p;
uint32_t i;
firstGo = 1;
for(i=0;i<64;i++)
vp->metrics1.c[i] = 63;
@ -88,7 +82,6 @@ int init_viterbi37_neon(void *p, int starting_state) {
mask_and = vdup_n_u8(0x80);
mask_shift = vld1_s8(xr);
vp->old_metrics = &vp->metrics1;
vp->new_metrics = &vp->metrics2;
vp->dp = vp->decisions;
@ -176,32 +169,31 @@ void print_uint8x16_t(char *s, uint8x16_t val) {
printf("\n");
}
static inline int movemask_neon(uint8x16_t movemask_low_in)
{
uint8x8_t lo = vget_low_u8(movemask_low_in);
uint8x8_t hi = vget_high_u8(movemask_low_in);
lo = vand_u8(lo, mask_and);
lo = vshl_u8(lo, mask_shift);
hi = vand_u8(hi, mask_and);
hi = vshl_u8(hi, mask_shift);
static inline int movemask_neon(uint8x16_t movemask_low_in) {
uint8x8_t lo = vget_low_u8(movemask_low_in);
uint8x8_t hi = vget_high_u8(movemask_low_in);
lo = vand_u8(lo, mask_and);
lo = vshl_u8(lo, mask_shift);
hi = vand_u8(hi, mask_and);
hi = vshl_u8(hi, mask_shift);
lo = vpadd_u8(lo, lo);
lo = vpadd_u8(lo, lo);
lo = vpadd_u8(lo, lo);
lo = vpadd_u8(lo, lo);
lo = vpadd_u8(lo, lo);
lo = vpadd_u8(lo, lo);
hi = vpadd_u8(hi, hi);
hi = vpadd_u8(hi, hi);
hi = vpadd_u8(hi, hi);
hi = vpadd_u8(hi, hi);
hi = vpadd_u8(hi, hi);
hi = vpadd_u8(hi, hi);
return ((hi[0] << 8) | (lo[0] & 0xFF));
return ((hi[0] << 8) | (lo[0] & 0xFF));
}
void update_viterbi37_blk_neon(void *p,unsigned char *syms,int nbits, uint32_t *best_state) {
struct v37 *vp = p;
decision_t *d;
uint8_t thirtyone;
thirtyone = 31;
uint8_t thirtyone;
thirtyone = 31;
if(p == NULL)
return;
@ -217,19 +209,18 @@ void update_viterbi37_blk_neon(void *p,unsigned char *syms,int nbits, uint32_t *
while(nbits--) {
uint8x16_t sym0v,sym1v,sym2v;
void *tmp;
int i;
// printf("nbits=%d, syms=%d,%d,%d\n", nbits, syms[0], syms[1], syms[2]);fflush(stdout);
/* Splat the 0th symbol across sym0v, the 1st symbol across sym1v, etc */
sym0v = vld1q_dup_u8(syms); // passing a char as opposed to a pointer to a char
sym0v = vld1q_dup_u8(syms);
sym1v = vld1q_dup_u8(syms+1);
sym2v = vld1q_dup_u8(syms+2);
syms += 3;
for(i=0;i<2;i++){
for(i=0;i<2;i++) {
uint8x16_t decision0, decision1, metric, m_metric, m0, m1, m2, m3, survivor0, survivor1;
/* Form branch metrics */
@ -246,48 +237,35 @@ void update_viterbi37_blk_neon(void *p,unsigned char *syms,int nbits, uint32_t *
m2 = vaddq_u8(vp->old_metrics->v[i],m_metric);
/* Compare and select, using modulo arithmetic */
decision0 = (uint8x16_t)vcgtq_s8(vsubq_s8((int8x16_t)m0,(int8x16_t)m1),vdupq_n_s8(0));
decision1 = (uint8x16_t)vcgtq_s8(vsubq_s8((int8x16_t)m2,(int8x16_t)m3),vdupq_n_s8(0));
survivor0 = vorrq_u8(vandq_u8(decision0,m1),vandq_u8(vmvnq_u8(decision0),m0));
survivor1 = vorrq_u8 (vandq_u8(decision1,m3),vandq_u8(vmvnq_u8(decision1),m2) );
////// equal to _mm_unpacklo_epi8 //////////
uint8x8_t a1 = vget_low_u8(decision0);
uint8x8_t b1 = vget_low_u8(decision1);
uint8x8x2_t result = vzip_u8(a1, b1);
uint8x16_t movemask_low_in = vcombine_u8(result.val[0], result.val[1]);
/////////////////////////////////////////
/* Pack each set of decisions into 16 bits */
uint8x8_t a1 = vget_low_u8(decision0);
uint8x8_t b1 = vget_low_u8(decision1);
uint8x8x2_t result = vzip_u8(a1, b1);
uint8x16_t movemask_low_in = vcombine_u8(result.val[0], result.val[1]);
////////equal to _mm_movemask_epi8 ////////
d->s[2*i] = movemask_neon(movemask_low_in);
d->s[2*i] = movemask_neon(movemask_low_in);
///////equal to _mm_unpackhi_epi8////////////
a1 = vget_high_u8(decision0);
b1 = vget_high_u8(decision1);
result = vzip_u8(a1, b1);
uint8x16_t movemask_hi_in = vcombine_u8(result.val[0], result.val[1]);
a1 = vget_high_u8(decision0);
b1 = vget_high_u8(decision1);
result = vzip_u8(a1, b1);
uint8x16_t movemask_hi_in = vcombine_u8(result.val[0], result.val[1]);
d->s[2*i+1] = movemask_neon(movemask_hi_in);
a1 = vget_low_u8(survivor0);
b1 = vget_low_u8(survivor1);
result = vzip_u8(a1, b1);
vp->new_metrics->v[2*i] = vcombine_u8(result.val[0], result.val[1]);
////////equal to _mm_movemask//////////////
d->s[2*i+1] = movemask_neon(movemask_hi_in);
a1 = vget_low_u8(survivor0);
b1 = vget_low_u8(survivor1);
result = vzip_u8(a1, b1);
vp->new_metrics->v[2*i] = vcombine_u8(result.val[0], result.val[1]);
a1 = vget_high_u8(survivor0);
b1 = vget_high_u8(survivor1);
result = vzip_u8(a1, b1);
vp->new_metrics->v[2*i+1] = vcombine_u8(result.val[0], result.val[1]);
a1 = vget_high_u8(survivor0);
b1 = vget_high_u8(survivor1);
result = vzip_u8(a1, b1);
vp->new_metrics->v[2*i+1] = vcombine_u8(result.val[0], result.val[1]);
}
@ -315,8 +293,7 @@ void update_viterbi37_blk_neon(void *p,unsigned char *syms,int nbits, uint32_t *
/* We cannot use a saturated subtract, because we often have to adjust by more than SHRT_MAX
* This is okay since it can't overflow anyway
*/
for(i=0;i<4;i++)
{
for(i=0;i<4;i++) {
vp->new_metrics->v[i] = vsubq_u8(vp->new_metrics->v[i],adjustv);
}
@ -326,7 +303,6 @@ void update_viterbi37_blk_neon(void *p,unsigned char *syms,int nbits, uint32_t *
tmp = vp->old_metrics;
vp->old_metrics = vp->new_metrics;
vp->new_metrics = tmp;
//firstGo = 0;
}
if (best_state) {

Loading…
Cancel
Save