cv::Mat HogIntegralImageComputer::compute_block_features(const cv::Point &start, const cv::Point &end) const
{
// sanity checks
assert(end.y < _integral_image.rows-1 && end.x < _integral_image.cols-1 &&
end.x >= 0 && end.y >= 0);
assert(start.y < _integral_image.rows-1 && start.x < _integral_image.cols-1 &&
start.x >= 0 && start.y >= 0);
// We have top, middle and bottom blocks
// The border of blocks might overlap (for simplicity reasons)
int h = end.y - start.y + 1;
if (h < 3) {
//return cv::Mat::zeros(1, 8*3, CV_32FC1);
return cv::Mat::zeros(1, CC_HOG_CHANS*3, CV_32FC1);
}
int h_top = h * 0.2; // the first border
int h_bottom = h * 0.8; // the second border
h_top = std::min(h_top, h - 3);
h_bottom = std::min(h_bottom, h - 2);
cv::Point start1 = start;
cv::Point end1 = cv::Point(end.x, start.y + h_top);
cv::Point start2 = cv::Point(start.x, start.y + h_top + 1);
cv::Point end2 = cv::Point(end.x, start.y + h_bottom);
cv::Point start3 = cv::Point(start.x, start.y + h_bottom + 1);
cv::Point end3 = cv::Point(end.x, end.y);
cv::VecHogf v1 = vec_max(query_ii(start1, end1), cv::VecHogf(0.0f));
cv::VecHogf v2 = vec_max(query_ii(start2, end2), cv::VecHogf(0.0f));
cv::VecHogf v3 = vec_max(query_ii(start3, end3), cv::VecHogf(0.0f));
cv::VecHogf v_norm = v1 + v2 + v3;
assert(v_norm[4] + 1e-5 > 0);
v1 = v1 / (v_norm[CC_HOG_CHANS] + 1e-5);
v2 = v2 / (v_norm[CC_HOG_CHANS] + 1e-5);
v3 = v3 / (v_norm[CC_HOG_CHANS] + 1e-5);
#if CC_HOG_CHANS == 4
cv::Mat result = (cv::Mat_<float>(1,CC_HOG_CHANS*3) <<
v1[0], v1[1], v1[2], v1[3],
v2[0], v2[1], v2[2], v2[3],
v3[0], v3[1], v3[2], v3[3]);
#else
cv::Mat result = (cv::Mat_<float>(1,CC_HOG_CHANS*3) <<
v1[0], v1[1], v1[2], v1[3], v1[4], v1[5], v1[6], v1[7],
v2[0], v2[1], v2[2], v2[3], v2[4], v2[5], v2[6], v2[7],
v3[0], v3[1], v3[2], v3[3], v3[4], v3[5], v3[6], v3[7]);
#endif
cv::sqrt(result, result);
return result;
}
SIMD_INLINE void PartialSort5(v128_u8 a[5])
{
SortU8(a[2], a[3]);
SortU8(a[1], a[2]);
SortU8(a[2], a[3]);
a[4] = vec_max(a[1], a[4]);
a[0] = vec_min(a[0], a[3]);
SortU8(a[2], a[0]);
a[2] = vec_max(a[4], a[2]);
a[2] = vec_min(a[2], a[0]);
}
void DP_alpha_prey_w(PARAM *param, PRIOR *prior, DATA *data, const gsl_rng *r, int pid) {
int i,j,id;
float cur_alpha_prey, tmp_lambda, maxl, tmp;
float prob[_MAX_COMP_];
for(i=0;i<_MAX_COMP_;i++) prob[i] = log(prior->gamma_alpha_prey[i]);
cur_alpha_prey = param->alpha_prey[pid];
for(i=0;i<_MAX_COMP_;i++) {
for(j=0;j<data->preyNinter[pid];j++) {
id = data->p2i[pid][j];
if(param->Z[data->a2u[id]]) {
tmp_lambda = param->lambda_true[id] + prior->theta_alpha_prey[i] - cur_alpha_prey;
tmp = GSL_MIN(50.0, data->d[id]);
if(tmp > 0.0) prob[i] += log_poisson_g_prop(tmp, exp(tmp_lambda), param->eta[pid]);
}
}
}
maxl = vec_max(prob, _MAX_COMP_);
for(i=0;i<_MAX_COMP_;i++) prob[i] -= maxl;
for(i=0;i<_MAX_COMP_;i++) prob[i] = exp(prob[i]);
prior->w_alpha_prey[pid] = ranMultinom(r, prob, _MAX_COMP_);
param->alpha_prey[pid] = prior->theta_alpha_prey[prior->w_alpha_prey[pid]];
for(j=0;j<data->preyNinter[pid];j++) {
id = data->p2i[pid][j];
param->lambda_true[id] += param->alpha_prey[pid] - cur_alpha_prey;
}
}
// out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
static inline vec_u8_t h264_deblock_q1(register vec_u8_t p0,
register vec_u8_t p1,
register vec_u8_t p2,
register vec_u8_t q0,
register vec_u8_t tc0) {
register vec_u8_t average = vec_avg(p0, q0);
register vec_u8_t temp;
register vec_u8_t uncliped;
register vec_u8_t ones;
register vec_u8_t max;
register vec_u8_t min;
register vec_u8_t newp1;
temp = vec_xor(average, p2);
average = vec_avg(average, p2); /*avg(p2, avg(p0, q0)) */
ones = vec_splat_u8(1);
temp = vec_and(temp, ones); /*(p2^avg(p0, q0)) & 1 */
uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
max = vec_adds(p1, tc0);
min = vec_subs(p1, tc0);
newp1 = vec_max(min, uncliped);
newp1 = vec_min(max, newp1);
return newp1;
}
void DP_eta_w(PARAM *param, PRIOR *prior, DATA *data, const gsl_rng *r, int pid) {
int i,j,id;
float cur_eta, tmp_lambda, maxl, tmp;
float prob[_MAX_COMP_];
for(i=0;i<_MAX_COMP_;i++) prob[i] = log(prior->gamma_eta[i]);
cur_eta = param->eta[pid];
for(i=0;i<_MAX_COMP_;i++) {
for(j=0;j<data->preyNinter[pid];j++) {
id = data->p2i[pid][j];
if(param->Z[data->a2u[id]] && data->d[id] > 0.0) {
tmp_lambda = param->lambda_true[id];
//else tmp_lambda = param->lambda_false[id];
if(lowMode) tmp = GSL_MIN(_LM_, data->d[id]);
else tmp = data->d[id];
if(data->d[id] > 0.0) prob[i] += log_poisson_g_prop(tmp, exp(tmp_lambda), prior->theta_eta[i]);
}
}
}
maxl = vec_max(prob, _MAX_COMP_);
for(i=0;i<_MAX_COMP_;i++) prob[i] -= maxl;
for(i=0;i<_MAX_COMP_;i++) prob[i] = exp(prob[i]);
prior->w_eta[pid] = ranMultinom(r, prob, _MAX_COMP_);
param->eta[pid] = prior->theta_eta[prior->w_eta[pid]];
}
// out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
static inline vector unsigned char h264_deblock_q1(register vector unsigned char p0,
register vector unsigned char p1,
register vector unsigned char p2,
register vector unsigned char q0,
register vector unsigned char tc0) {
register vector unsigned char average = vec_avg(p0, q0);
register vector unsigned char temp;
register vector unsigned char uncliped;
register vector unsigned char ones;
register vector unsigned char max;
register vector unsigned char min;
register vector unsigned char newp1;
temp = vec_xor(average, p2);
average = vec_avg(average, p2); /*avg(p2, avg(p0, q0)) */
ones = vec_splat_u8(1);
temp = vec_and(temp, ones); /*(p2^avg(p0, q0)) & 1 */
uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
max = vec_adds(p1, tc0);
min = vec_subs(p1, tc0);
newp1 = vec_max(min, uncliped);
newp1 = vec_min(max, newp1);
return newp1;
}
void DP_mu_w(PARAM *param, PRIOR *prior, DATA *data, const gsl_rng *r, int pid) {
int i,j,id;
float cur_mu, tmp_lambda, maxl, tmp;
float prob[_MAX_COMP_];
for(i=0;i<_MAX_COMP_;i++) prob[i] = log(prior->gamma_mu[i]);
cur_mu = param->mu[pid];
for(i=0;i<_MAX_COMP_;i++) {
for(j=0;j<data->preyNinter[pid];j++) {
id = data->p2i[pid][j];
if(data->ctrl[data->i2IP[id]]) {
tmp_lambda = param->lambda_false[id] + prior->theta_mu[i] - cur_mu;
tmp = data->d[id];
prob[i] += log_poisson_g_prop(tmp, exp(tmp_lambda), param->eta0[data->i2p[id]]);
}
}
}
maxl = vec_max(prob, _MAX_COMP_);
for(i=0;i<_MAX_COMP_;i++) prob[i] -= maxl;
for(i=0;i<_MAX_COMP_;i++) prob[i] = exp(prob[i]);
prior->w_mu[pid] = ranMultinom(r, prob, _MAX_COMP_);
param->mu[pid] = prior->theta_mu[prior->w_mu[pid]];
for(j=0;j<data->preyNinter[pid];j++) {
id = data->p2i[pid][j];
param->lambda_false[id] += param->mu[pid] - cur_mu;
}
}
SIMD_INLINE void PartialSort9(v128_u8 a[9])
{
SortU8(a[1], a[2]); SortU8(a[4], a[5]); SortU8(a[7], a[8]);
SortU8(a[0], a[1]); SortU8(a[3], a[4]); SortU8(a[6], a[7]);
SortU8(a[1], a[2]); SortU8(a[4], a[5]); SortU8(a[7], a[8]);
a[3] = vec_max(a[0], a[3]);
a[5] = vec_min(a[5], a[8]);
SortU8(a[4], a[7]);
a[6] = vec_max(a[3], a[6]);
a[4] = vec_max(a[1], a[4]);
a[2] = vec_min(a[2], a[5]);
a[4] = vec_min(a[4], a[7]);
SortU8(a[4], a[2]);
a[4] = vec_max(a[6], a[4]);
a[4] = vec_min(a[4], a[2]);
}
void DP_alpha_IP_w(PARAM *param, PRIOR *prior, DATA *data, const gsl_rng *r, int pid) {
int i,j,id;
float cur_alpha_IP, tmp_lambda, maxl;
float prob[_MAX_COMP_];
for(i=0;i<_MAX_COMP_;i++) prob[i] = log(prior->gamma_alpha_IP[i]);
cur_alpha_IP = param->alpha_IP[pid];
for(i=0;i<_MAX_COMP_;i++) {
for(j=0;j<data->IPNinter[pid];j++) {
id = data->IP2i[pid][j];
if(param->Z[data->a2u[id]]) {
tmp_lambda = param->lambda_true[id] + prior->theta_alpha_IP[i] - cur_alpha_IP;
prob[i] += log_gaussian(data->d[id], (tmp_lambda), param->eta[data->i2p[id]]);
}
}
}
maxl = vec_max(prob, _MAX_COMP_);
for(i=0;i<_MAX_COMP_;i++) prob[i] -= maxl;
for(i=0;i<_MAX_COMP_;i++) prob[i] = exp(prob[i]);
prior->w_alpha_IP[pid] = ranMultinom(r, prob, _MAX_COMP_);
param->alpha_IP[pid] = prior->theta_alpha_IP[prior->w_alpha_IP[pid]];
for(j=0;j<data->IPNinter[pid];j++) {
id = data->IP2i[pid][j];
param->lambda_true[id] += param->alpha_IP[pid] - cur_alpha_IP;
}
}
int pix_abs16x16_y2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
{
int i;
int s __attribute__((aligned(16)));
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
vector unsigned char *tv;
vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
vector unsigned int sad;
vector signed int sumdiffs;
uint8_t *pix3 = pix2 + line_size;
s = 0;
sad = (vector unsigned int)vec_splat_u32(0);
/*
Due to the fact that pix3 = pix2 + line_size, the pix3 of one
iteration becomes pix2 in the next iteration. We can use this
fact to avoid a potentially expensive unaligned read, each
time around the loop.
Read unaligned pixels into our vectors. The vectors are as follows:
pix2v: pix2[0]-pix2[15]
Split the pixel vectors into shorts
*/
tv = (vector unsigned char *) &pix2[0];
pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
for(i=0;i<16;i++) {
/*
Read unaligned pixels into our vectors. The vectors are as follows:
pix1v: pix1[0]-pix1[15]
pix3v: pix3[0]-pix3[15]
*/
tv = (vector unsigned char *) pix1;
pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
tv = (vector unsigned char *) &pix3[0];
pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
/* Calculate the average vector */
avgv = vec_avg(pix2v, pix3v);
/* Calculate a sum of abs differences vector */
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
/* Add each 4 pixel group together and put 4 results into sad */
sad = vec_sum4s(t5, sad);
pix1 += line_size;
pix2v = pix3v;
pix3 += line_size;
}
/* Sum up the four partial sums, and put the result into s */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
void
gimp_composite_lighten_rgba8_rgba8_rgba8_altivec (GimpCompositeContext *ctx)
{
const guchar *A = ctx->A;
const guchar *B = ctx->B;
guchar *D = ctx->D;
guint length = ctx->n_pixels;
vector unsigned char a,b,d,alpha_a,alpha_b;
while (length >= 4)
{
a=LoadUnaligned(A);
b=LoadUnaligned(B);
alpha_a=vec_and(a, alphamask);
alpha_b=vec_and(b, alphamask);
d=vec_min(alpha_a, alpha_b);
a=vec_andc(a, alphamask);
a=vec_adds(a, d);
b=vec_andc(b, alphamask);
d=vec_max(a, b);
StoreUnaligned(d, D);
A+=16;
B+=16;
D+=16;
length-=4;
}
/* process last pixels */
length = length*4;
a=LoadUnalignedLess(A, length);
b=LoadUnalignedLess(B, length);
alpha_a=vec_and(a,alphamask);
alpha_b=vec_and(b,alphamask);
d=vec_min(alpha_a,alpha_b);
a=vec_andc(a,alphamask);
a=vec_adds(a,d);
b=vec_andc(b,alphamask);
d=vec_max(a, b);
StoreUnalignedLess(d, D, length);
}
double log_sum(double *x, int len) {
int i;
double res, sum;
sum = vec_max(x, len);
for(i=0;i<len;i++) x[i] -= sum;
for(i=0;i<len;i++) x[i] = exp(x[i]);
res = vec_sum(x, len);
return res;
}
__SIMDd _SIMD_max_pd(__SIMDd a, __SIMDd b)
{
#ifdef USE_SSE
return _mm_max_pd(a,b);
#elif defined USE_AVX
return _mm256_max_pd(a,b);
#elif defined USE_IBM
return vec_max(a,b);
#endif
}
folgen_vektor_p
folgen_vektor_add(folgen_vektor_p f, folgen_vektor_p g) {
folgen_vektor_p back;
int k, d, size, dim, a, b, test_f, test_g;
vec_p max, vec_1, n, n_f, n_g, r;
ASSERT( f->grad->dim == g->grad->dim );
dim = f->grad->dim;
max = vec_max( f->grad, g->grad );
vec_1 = vec_one( dim );
n = vec_add( vec_1, max );
n_f = vec_add( vec_1, f->grad );
n_g = vec_add( vec_1, g->grad );
size = vec_size( n );
back = folgen_vektor_new( max );
for(k=0;k<size;k++) {
r = entry_one2d( k, n );
test_f = 0;
test_g = 0;
for(d=0;d<dim;d++) {
if( r->array[d] > f->grad->array[d] ) {
test_f = test_f + 1;
}
if( r->array[d] > g->grad->array[d] ) {
test_g = test_g + 1;
}
}
if( (test_f == 0) && (test_g == 0) ) {
a = entry_d2one( r, n_f );
b = entry_d2one( r, n_g );
folge_del( back->vektor[k] );
back->vektor[k] = folge_add( f->vektor[a], g->vektor[b] );
}
if( (test_f != 0) && (test_g == 0) ) {
b = entry_d2one( r, n_g );
folge_del( back->vektor[k] );
back->vektor[k] = folge_copy( g->vektor[b] );
}
if( (test_f == 0) && (test_g != 0) ) {
a = entry_d2one( r, n_f );
folge_del( back->vektor[k] );
back->vektor[k] = folge_copy( f->vektor[a] );
}
vec_del( r );
}
vec_del( n );
vec_del( n_f );
vec_del( n_g );
vec_del( vec_1 );
return back;
}
/* Assumes input points to the beginning of the SSS symbol. The SSS symbol start is
* given by SSS_SYMBOL_ST() macro in sss.h.
* Estimates the m0 and m1 values and saves in m0_value and m1_value
* the resulted correlation (higher is more likely)
*
*/
void sss_synch_m0m1(sss_synch_t *q, cf_t *input, int *m0, float *m0_value,
int *m1, float *m1_value) {
/* This is aprox 3-4 kbytes of stack. Consider moving to sss_synch_t?? */
cf_t zdelay[N_SSS+1],zconj[N_SSS+1],zprod[N_SSS+1];
cf_t y[2][N_SSS+1], z[N_SSS+1], tmp[N_SSS+1];
float tmp_real[N_SSS+1];
cf_t input_fft[SSS_DFT_LEN];
int i;
dft_run_c2c(&q->dftp_input, input, input_fft);
for (i = 0; i < N_SSS; i++) {
y[0][i] = input_fft[SSS_POS_SYMBOL + 2 * i];
y[1][i] = input_fft[SSS_POS_SYMBOL + 2 * i + 1];
}
vec_dot_prod(y[0], q->fc_tables.c[0], z, N_SSS);
memcpy(zdelay, &z[1], (N_SSS - 1) * sizeof(cf_t));
vec_conj(z, zconj, N_SSS - 1);
vec_dot_prod(zdelay, zconj, zprod, N_SSS - 1);
corr_all_zs(zprod, q->fc_tables.s, tmp);
vec_abs(tmp, tmp_real, N_SSS);
vec_max(tmp_real, m0_value, m0, N_SSS);
vec_dot_prod(y[1], q->fc_tables.c[1], tmp, N_SSS);
vec_dot_prod(tmp, q->fc_tables.z1[*m0], z, N_SSS);
memcpy(zdelay, &z[1], (N_SSS - 1) * sizeof(cf_t));
vec_conj(z, zconj, N_SSS - 1);
vec_dot_prod(zdelay, zconj, zprod, N_SSS - 1);
corr_all_zs(zprod, q->fc_tables.s, tmp);
vec_abs(tmp, tmp_real, N_SSS);
vec_max(tmp_real, m1_value, m1, N_SSS);
}
/**
* Sum of Squared Errors for a 8x8 block.
* AltiVec-enhanced.
* It's the pix_abs8x8_altivec code above w/ squaring added.
*/
int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size)
{
int i;
int s __attribute__((aligned(16)));
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sum;
vector signed int sumsqr;
sum = (vector unsigned int)vec_splat_u32(0);
permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
for(i=0;i<8;i++) {
/* Read potentially unaligned pixels into t1 and t2
Since we're reading 16 pixels, and actually only want 8,
mask out the last 8 pixels. The 0s don't change the sum. */
perm1 = vec_lvsl(0, pix1);
pix1v = (vector unsigned char *) pix1;
perm2 = vec_lvsl(0, pix2);
pix2v = (vector unsigned char *) pix2;
t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
/*
Since we want to use unsigned chars, we can take advantage
of the fact that abs(a-b)^2 = (a-b)^2.
*/
/* Calculate abs differences vector */
t3 = vec_max(t1, t2);
t4 = vec_min(t1, t2);
t5 = vec_sub(t3, t4);
/* Square the values and add them to our sum */
sum = vec_msum(t5, t5, sum);
pix1 += line_size;
pix2 += line_size;
}
/* Sum up the four partial sums, and put the result into s */
sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
sumsqr = vec_splat(sumsqr, 3);
vec_ste(sumsqr, 0, &s);
return s;
}
/*! <em>Pointwise Maximum</em> of vA and vB. */
inline vec4i vec4i_max(vec4i vA, vec4i vB) {
#if defined(__ALTIVEC__) /* AltiVec */
return vec_max((vector int)vA, (vector int)vB);
#else /* Scalar */
int * s1 = (int*)&vA;
int * s2 = (int*)&vB;
vec4i max;
int * smax = (int*)&max;
smax[0] = MAX(s1[0], s2[0]);
smax[1] = MAX(s1[1], s2[1]);
smax[2] = MAX(s1[2], s2[2]);
smax[3] = MAX(s1[3], s2[3]);
return max;
#endif
}
int pix_abs16x16_x2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
{
int i;
int s __attribute__((aligned(16)));
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
vector unsigned char *tv;
vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
vector unsigned int sad;
vector signed int sumdiffs;
s = 0;
sad = (vector unsigned int)vec_splat_u32(0);
for(i=0;i<16;i++) {
/*
Read unaligned pixels into our vectors. The vectors are as follows:
pix1v: pix1[0]-pix1[15]
pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
*/
tv = (vector unsigned char *) pix1;
pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
tv = (vector unsigned char *) &pix2[0];
pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
tv = (vector unsigned char *) &pix2[1];
pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
/* Calculate the average vector */
avgv = vec_avg(pix2v, pix2iv);
/* Calculate a sum of abs differences vector */
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
/* Add each 4 pixel group together and put 4 results into sad */
sad = vec_sum4s(t5, sad);
pix1 += line_size;
pix2 += line_size;
}
/* Sum up the four partial sums, and put the result into s */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
{
int i;
int s __attribute__((aligned(16)));
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sad;
vector signed int sumdiffs;
sad = (vector unsigned int)vec_splat_u32(0);
permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
for(i=0;i<8;i++) {
/* Read potentially unaligned pixels into t1 and t2
Since we're reading 16 pixels, and actually only want 8,
mask out the last 8 pixels. The 0s don't change the sum. */
perm1 = vec_lvsl(0, pix1);
pix1v = (vector unsigned char *) pix1;
perm2 = vec_lvsl(0, pix2);
pix2v = (vector unsigned char *) pix2;
t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
/* Calculate a sum of abs differences vector */
t3 = vec_max(t1, t2);
t4 = vec_min(t1, t2);
t5 = vec_sub(t3, t4);
/* Add each 4 pixel group together and put 4 results into sad */
sad = vec_sum4s(t5, sad);
pix1 += line_size;
pix2 += line_size;
}
/* Sum up the four partial sums, and put the result into s */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
static int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
int s;
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
vector unsigned char perm1 = vec_lvsl(0, pix2);
vector unsigned char perm2 = vec_add(perm1, vec_splat_u8(1));
vector unsigned char pix2l, pix2r;
vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
vector unsigned int sad;
vector signed int sumdiffs;
s = 0;
sad = (vector unsigned int)vec_splat_u32(0);
for (i = 0; i < h; i++) {
/* Read unaligned pixels into our vectors. The vectors are as follows:
pix1v: pix1[0]-pix1[15]
pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] */
pix1v = vec_ld( 0, pix1);
pix2l = vec_ld( 0, pix2);
pix2r = vec_ld(16, pix2);
pix2v = vec_perm(pix2l, pix2r, perm1);
pix2iv = vec_perm(pix2l, pix2r, perm2);
/* Calculate the average vector */
avgv = vec_avg(pix2v, pix2iv);
/* Calculate a sum of abs differences vector */
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
/* Add each 4 pixel group together and put 4 results into sad */
sad = vec_sum4s(t5, sad);
pix1 += line_size;
pix2 += line_size;
}
/* Sum up the four partial sums, and put the result into s */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
folge_p
folge_add(folge_p f, folge_p g) {
folge_p back;
int size, k, size_g, size_f;
fepc_real_t x, y;
vec_p temp1, temp2, max, lang, min, r;
size_f = vec_size( f->lang );
size_g = vec_size( g->lang );
if(size_f == 0) {
back = folge_copy( g );
return back;
}
if(size_g == 0) {
back = folge_copy( f );
return back;
}
if( (size_g!=0) && (size_f!=0) ) {
min = vec_min( f->start, g->start );
temp1 = vec_add( f->start, f->lang );
temp2 = vec_add( g->start, g->lang );
max = vec_max( temp1, temp2 );
vec_del( temp1 );
vec_del( temp2 );
lang = vec_op( 1, max, -1, min);
vec_del( max );
back = folge_new( min, lang );
size = vec_size( lang );
for(k=0;k<size;k++) {
temp1 = entry_one2d( k, lang );
r = vec_add( min, temp1 );
vec_del( temp1 );
x = folge_glied( r, f );
y = folge_glied( r, g );
vec_del( r );
back->glied[k] = x + y;
}
return back;
}
}
static unsigned reg_sad_altivec(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int width, const int height, const unsigned stride1, const unsigned stride2)
{
vector unsigned int vsad = {0,0,0,0}, vzero = {0,0,0,0};
vector signed int sumdiffs;
int tmpsad, sad = 0;
int y, x;
for (y = 0; y < height; ++y) {
vector unsigned char perm1, perm2;
perm1 = vec_lvsl(0, &data1[y * stride1]);
perm2 = vec_lvsl(0, &data2[y * stride2]);
for (x = 0; x <= width-16; x+=16) {
vector unsigned char t1, t2, t3, t4, t5;
vector unsigned char *current, *previous;
current = (vector unsigned char *) &data1[y * stride1 + x];
previous = (vector unsigned char *) &data2[y * stride2 + x];
t1 = vec_perm(current[0], current[1], perm1 ); /* align current vector */
t2 = vec_perm(previous[0], previous[1], perm2 );/* align previous vector */
t3 = vec_max(t1, t2 ); /* find largest of two */
t4 = vec_min(t1, t2 ); /* find smaller of two */
t5 = vec_sub(t3, t4); /* find absolute difference */
vsad = vec_sum4s(t5, vsad); /* accumulate sum of differences */
}
for (; x < width; ++x) {
sad += abs(data1[y * stride1 + x] - data2[y * stride2 + x]);
}
}
sumdiffs = vec_sums((vector signed int) vsad, (vector signed int) vzero);
/* copy vector sum into unaligned result */
sumdiffs = vec_splat( sumdiffs, 3);
vec_ste( sumdiffs, 0, &tmpsad );
sad += tmpsad;
return sad;
}
void DP_eta_w(PARAM *param, PRIOR *prior, DATA *data, const gsl_rng *r, int pid) {
int i,j,id;
float cur_eta, tmp_lambda, maxl;
float prob[_MAX_COMP_];
for(i=0;i<_MAX_COMP_;i++) prob[i] = log(prior->gamma_eta[i]);
cur_eta = param->eta[pid];
for(i=0;i<_MAX_COMP_;i++) {
for(j=0;j<data->preyNinter[pid];j++) {
id = data->p2i[pid][j];
if(param->Z[data->a2u[id]]) tmp_lambda = param->lambda_true[id];
else tmp_lambda = param->lambda_false[id];
prob[i] += log_gaussian(data->d[id], (tmp_lambda), prior->theta_eta[i]);
}
}
maxl = vec_max(prob, _MAX_COMP_);
for(i=0;i<_MAX_COMP_;i++) prob[i] -= maxl;
for(i=0;i<_MAX_COMP_;i++) prob[i] = exp(prob[i]);
prior->w_eta[pid] = ranMultinom(r, prob, _MAX_COMP_);
param->eta[pid] = prior->theta_eta[prior->w_eta[pid]];
}
int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
{
int i;
int s __attribute__((aligned(16)));
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm1, perm2, *pix1v, *pix2v;
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sad;
vector signed int sumdiffs;
sad = (vector unsigned int)vec_splat_u32(0);
for(i=0;i<16;i++) {
/* Read potentially unaligned pixels into t1 and t2 */
perm1 = vec_lvsl(0, pix1);
pix1v = (vector unsigned char *) pix1;
perm2 = vec_lvsl(0, pix2);
pix2v = (vector unsigned char *) pix2;
t1 = vec_perm(pix1v[0], pix1v[1], perm1);
t2 = vec_perm(pix2v[0], pix2v[1], perm2);
/* Calculate a sum of abs differences vector */
t3 = vec_max(t1, t2);
t4 = vec_min(t1, t2);
t5 = vec_sub(t3, t4);
/* Add each 4 pixel group together and put 4 results into sad */
sad = vec_sum4s(t5, sad);
pix1 += line_size;
pix2 += line_size;
}
/* Sum up the four partial sums, and put the result into s */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
static int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
int s;
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm = vec_lvsl(0, pix2);
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sad;
vector signed int sumdiffs;
sad = (vector unsigned int)vec_splat_u32(0);
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2 */
vector unsigned char pix2l = vec_ld( 0, pix2);
vector unsigned char pix2r = vec_ld(15, pix2);
t1 = vec_ld(0, pix1);
t2 = vec_perm(pix2l, pix2r, perm);
/* Calculate a sum of abs differences vector */
t3 = vec_max(t1, t2);
t4 = vec_min(t1, t2);
t5 = vec_sub(t3, t4);
/* Add each 4 pixel group together and put 4 results into sad */
sad = vec_sum4s(t5, sad);
pix1 += line_size;
pix2 += line_size;
}
/* Sum up the four partial sums, and put the result into s */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
fepc_real_t
folge_norm(folge_p f, folge_p g) {
vec_p temp, temp1, temp2, min, max, lang, vec_1;
int k, dim, size;
fepc_real_t norm, diff;
ASSERT(f->start->dim == g->start->dim);
dim = f->start->dim;
vec_1 = vec_one( dim );
min = vec_min( f->start, g->start );
temp1 = vec_add( f->start, f->lang );
temp2 = vec_add( g->start, g->lang );
temp = vec_max( temp1, temp2 );
vec_del( temp1 );
vec_del( temp2 );
lang = vec_op( 1, temp, -1, min );
vec_del( temp );
size = vec_size( lang );
norm = 0.0;
for(k=0;k<size;k++) {
temp = entry_one2d( k, lang );
temp1 = vec_add( temp, min );
diff = folge_glied( temp1, f ) - folge_glied( temp1, g );
norm = norm + ( diff * diff );
vec_del( temp );
vec_del( temp1 );
}
vec_del( vec_1 );
vec_del( min );
vec_del( lang );
norm = sqrt(norm);
return norm;
}
template<> SIMD_INLINE v128_s16 InterferenceChange<false>(v128_s16 statistic, v128_s16 value, v128_s16 saturation)
{
return vec_max(vec_sub(statistic, value), saturation);
}
void test1() {
// CHECK-LABEL: define void @test1
// CHECK-LE-LABEL: define void @test1
res_vf = vec_abs(vf);
// CHECK: call <4 x float> @llvm.fabs.v4f32(<4 x float> %{{[0-9]*}})
// CHECK-LE: call <4 x float> @llvm.fabs.v4f32(<4 x float> %{{[0-9]*}})
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_add(vd, vd);
// CHECK: fadd <2 x double>
// CHECK-LE: fadd <2 x double>
res_vd = vec_and(vbll, vd);
// CHECK: and <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
// CHECK-LE: and <2 x i64>
// CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
res_vd = vec_and(vd, vbll);
// CHECK: and <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
// CHECK-LE: and <2 x i64>
// CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
res_vd = vec_and(vd, vd);
// CHECK: and <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
// CHECK-LE: and <2 x i64>
// CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_andc(vbll, vd);
// CHECK: bitcast <2 x double> %{{[0-9]*}} to <2 x i64>
// CHECK: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1>
// CHECK: and <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
// CHECK-LE: bitcast <2 x double> %{{[0-9]*}} to <2 x i64>
// CHECK-LE: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1>
// CHECK-LE: and <2 x i64>
// CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_andc(vd, vbll);
// CHECK: bitcast <2 x double> %{{[0-9]*}} to <2 x i64>
// CHECK: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1>
// CHECK: and <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
// CHECK-LE: bitcast <2 x double> %{{[0-9]*}} to <2 x i64>
// CHECK-LE: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1>
// CHECK-LE: and <2 x i64>
// CHECK-LE: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
dummy();
// CHECK: call void @dummy()
res_vd = vec_andc(vd, vd);
// CHECK: bitcast <2 x double> %{{[0-9]*}} to <2 x i64>
// CHECK: xor <2 x i64> %{{[0-9]*}}, <i64 -1, i64 -1>
// CHECK: and <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]*}} to <2 x double>
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_ceil(vd);
// CHECK: call <2 x double> @llvm.ceil.v2f64(<2 x double> %{{[0-9]*}})
// CHECK-LE: call <2 x double> @llvm.ceil.v2f64(<2 x double> %{{[0-9]*}})
res_vf = vec_ceil(vf);
// CHECK: call <4 x float> @llvm.ceil.v4f32(<4 x float> %{{[0-9]*}})
// CHECK-LE: call <4 x float> @llvm.ceil.v4f32(<4 x float> %{{[0-9]*}})
res_vbll = vec_cmpeq(vd, vd);
// CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpeqdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
// CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpeqdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
res_vbi = vec_cmpeq(vf, vf);
// CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpeqsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
// CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpeqsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
res_vbll = vec_cmpge(vd, vd);
// CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
// CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
res_vbi = vec_cmpge(vf, vf);
// CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
// CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
res_vbll = vec_cmpgt(vd, vd);
// CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
// CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
res_vbi = vec_cmpgt(vf, vf);
// CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
// CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
res_vbll = vec_cmple(vd, vd);
// CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
// CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgedp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
res_vbi = vec_cmple(vf, vf);
// CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
// CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgesp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
res_vbll = vec_cmplt(vd, vd);
// CHECK: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
// CHECK-LE: call <2 x i64> @llvm.ppc.vsx.xvcmpgtdp(<2 x double> %{{[0-9]*}}, <2 x double> %{{[0-9]*}})
res_vbi = vec_cmplt(vf, vf);
// CHECK: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
// CHECK-LE: call <4 x i32> @llvm.ppc.vsx.xvcmpgtsp(<4 x float> %{{[0-9]*}}, <4 x float> %{{[0-9]*}})
/* vec_cpsgn */
res_vf = vec_cpsgn(vf, vf);
// CHECK: call <4 x float> @llvm.copysign.v4f32(<4 x float> %{{.+}}, <4 x float> %{{.+}})
// CHECK-LE: call <4 x float> @llvm.copysign.v4f32(<4 x float> %{{.+}}, <4 x float> %{{.+}})
res_vd = vec_cpsgn(vd, vd);
// CHECK: call <2 x double> @llvm.copysign.v2f64(<2 x double> %{{.+}}, <2 x double> %{{.+}})
// CHECK-LE: call <2 x double> @llvm.copysign.v2f64(<2 x double> %{{.+}}, <2 x double> %{{.+}})
/* vec_div */
res_vsll = vec_div(vsll, vsll);
// CHECK: sdiv <2 x i64>
// CHECK-LE: sdiv <2 x i64>
res_vull = vec_div(vull, vull);
// CHECK: udiv <2 x i64>
// CHECK-LE: udiv <2 x i64>
res_vf = vec_div(vf, vf);
// CHECK: fdiv <4 x float>
// CHECK-LE: fdiv <4 x float>
res_vd = vec_div(vd, vd);
// CHECK: fdiv <2 x double>
// CHECK-LE: fdiv <2 x double>
/* vec_max */
res_vf = vec_max(vf, vf);
// CHECK: @llvm.ppc.vsx.xvmaxsp
// CHECK-LE: @llvm.ppc.vsx.xvmaxsp
res_vd = vec_max(vd, vd);
// CHECK: @llvm.ppc.vsx.xvmaxdp
// CHECK-LE: @llvm.ppc.vsx.xvmaxdp
res_vf = vec_vmaxfp(vf, vf);
// CHECK: @llvm.ppc.vsx.xvmaxsp
// CHECK-LE: @llvm.ppc.vsx.xvmaxsp
/* vec_min */
res_vf = vec_min(vf, vf);
// CHECK: @llvm.ppc.vsx.xvminsp
// CHECK-LE: @llvm.ppc.vsx.xvminsp
res_vd = vec_min(vd, vd);
// CHECK: @llvm.ppc.vsx.xvmindp
// CHECK-LE: @llvm.ppc.vsx.xvmindp
res_vf = vec_vminfp(vf, vf);
// CHECK: @llvm.ppc.vsx.xvminsp
// CHECK-LE: @llvm.ppc.vsx.xvminsp
res_d = __builtin_vsx_xsmaxdp(d, d);
// CHECK: @llvm.ppc.vsx.xsmaxdp
// CHECK-LE: @llvm.ppc.vsx.xsmaxdp
res_d = __builtin_vsx_xsmindp(d, d);
// CHECK: @llvm.ppc.vsx.xsmindp
// CHECK-LE: @llvm.ppc.vsx.xsmindp
/* vec_perm */
res_vsll = vec_perm(vsll, vsll, vuc);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_perm(vull, vull, vuc);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vbll = vec_perm(vbll, vbll, vuc);
// CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
// CHECK-LE: xor <16 x i8>
// CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
res_vf = vec_round(vf);
// CHECK: call <4 x float> @llvm.round.v4f32(<4 x float>
// CHECK-LE: call <4 x float> @llvm.round.v4f32(<4 x float>
res_vd = vec_round(vd);
// CHECK: call <2 x double> @llvm.round.v2f64(<2 x double>
// CHECK-LE: call <2 x double> @llvm.round.v2f64(<2 x double>
res_vd = vec_perm(vd, vd, vuc);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vd = vec_splat(vd, 1);
// CHECK: [[T1:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32>
// CHECK: [[T2:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32>
// CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
// CHECK-LE: xor <16 x i8>
// CHECK-LE: [[T1:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32>
// CHECK-LE: [[T2:%.+]] = bitcast <2 x double> {{.+}} to <4 x i32>
// CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
res_vbll = vec_splat(vbll, 1);
// CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
// CHECK-LE: xor <16 x i8>
// CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
res_vsll = vec_splat(vsll, 1);
// CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
// CHECK-LE: xor <16 x i8>
// CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
res_vull = vec_splat(vull, 1);
// CHECK: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
// CHECK-LE: xor <16 x i8>
// CHECK-LE: [[T1:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: [[T2:%.+]] = bitcast <2 x i64> {{.+}} to <4 x i32>
// CHECK-LE: call <4 x i32> @llvm.ppc.altivec.vperm(<4 x i32> [[T1]], <4 x i32> [[T2]], <16 x i8>
res_vsi = vec_pack(vsll, vsll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vui = vec_pack(vull, vull);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vbi = vec_pack(vbll, vbll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vsll = vec_vperm(vsll, vsll, vuc);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_vperm(vull, vull, vuc);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vd = vec_vperm(vd, vd, vuc);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
/* vec_vsx_ld */
res_vsi = vec_vsx_ld(0, &vsi);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vui = vec_vsx_ld(0, &vui);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vf = vec_vsx_ld (0, &vf);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vsll = vec_vsx_ld(0, &vsll);
// CHECK: @llvm.ppc.vsx.lxvd2x
// CHECK-LE: @llvm.ppc.vsx.lxvd2x
res_vull = vec_vsx_ld(0, &vull);
// CHECK: @llvm.ppc.vsx.lxvd2x
// CHECK-LE: @llvm.ppc.vsx.lxvd2x
res_vd = vec_vsx_ld(0, &vd);
// CHECK: @llvm.ppc.vsx.lxvd2x
// CHECK-LE: @llvm.ppc.vsx.lxvd2x
res_vull = vec_vsx_ld(0, &vull);
// CHECK: @llvm.ppc.vsx.lxvd2x
// CHECK-LE: @llvm.ppc.vsx.lxvd2x
res_vd = vec_vsx_ld(0, &vd);
// CHECK: @llvm.ppc.vsx.lxvd2x
// CHECK-LE: @llvm.ppc.vsx.lxvd2x
res_vss = vec_vsx_ld(0, &vss);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vss = vec_vsx_ld(0, &ss);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vus = vec_vsx_ld(0, &vus);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vus = vec_vsx_ld(0, &us);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vbc = vec_vsx_ld(0, &vbc);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vsc = vec_vsx_ld(0, &vsc);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vuc = vec_vsx_ld(0, &vuc);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vsc = vec_vsx_ld(0, &sc);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
res_vuc = vec_vsx_ld(0, &uc);
// CHECK: @llvm.ppc.vsx.lxvw4x
// CHECK-LE: @llvm.ppc.vsx.lxvw4x
/* vec_vsx_st */
vec_vsx_st(vsi, 0, &res_vsi);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vsi, 0, &res_si);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vui, 0, &res_vui);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vui, 0, &res_ui);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vf, 0, &res_vf);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vsll, 0, &res_vsll);
// CHECK: @llvm.ppc.vsx.stxvd2x
// CHECK-LE: @llvm.ppc.vsx.stxvd2x
vec_vsx_st(vull, 0, &res_vull);
// CHECK: @llvm.ppc.vsx.stxvd2x
// CHECK-LE: @llvm.ppc.vsx.stxvd2x
vec_vsx_st(vd, 0, &res_vd);
// CHECK: @llvm.ppc.vsx.stxvd2x
// CHECK-LE: @llvm.ppc.vsx.stxvd2x
vec_vsx_st(vss, 0, &res_vss);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vss, 0, &res_ss);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vus, 0, &res_vus);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vus, 0, &res_us);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vsc, 0, &res_vsc);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vsc, 0, &res_sc);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vuc, 0, &res_vuc);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vuc, 0, &res_uc);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vbc, 0, &res_vbc);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vbc, 0, &res_sc);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
vec_vsx_st(vbc, 0, &res_uc);
// CHECK: @llvm.ppc.vsx.stxvw4x
// CHECK-LE: @llvm.ppc.vsx.stxvw4x
/* vec_and */
res_vsll = vec_and(vsll, vsll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vsll = vec_and(vbll, vsll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vsll = vec_and(vsll, vbll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_and(vull, vull);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_and(vbll, vull);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_and(vull, vbll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vbll = vec_and(vbll, vbll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
/* vec_vand */
res_vsll = vec_vand(vsll, vsll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vsll = vec_vand(vbll, vsll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vsll = vec_vand(vsll, vbll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_vand(vull, vull);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_vand(vbll, vull);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_vand(vull, vbll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
res_vbll = vec_vand(vbll, vbll);
// CHECK: and <2 x i64>
// CHECK-LE: and <2 x i64>
/* vec_andc */
res_vsll = vec_andc(vsll, vsll);
// CHECK: xor <2 x i64>
// CHECK: and <2 x i64>
// CHECK-LE: xor <2 x i64>
// CHECK-LE: and <2 x i64>
res_vsll = vec_andc(vbll, vsll);
// CHECK: xor <2 x i64>
// CHECK: and <2 x i64>
// CHECK-LE: xor <2 x i64>
// CHECK-LE: and <2 x i64>
res_vsll = vec_andc(vsll, vbll);
// CHECK: xor <2 x i64>
// CHECK: and <2 x i64>
// CHECK-LE: xor <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_andc(vull, vull);
// CHECK: xor <2 x i64>
// CHECK: and <2 x i64>
// CHECK-LE: xor <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_andc(vbll, vull);
// CHECK: xor <2 x i64>
// CHECK: and <2 x i64>
// CHECK-LE: xor <2 x i64>
// CHECK-LE: and <2 x i64>
res_vull = vec_andc(vull, vbll);
// CHECK: xor <2 x i64>
// CHECK: and <2 x i64>
// CHECK-LE: xor <2 x i64>
// CHECK-LE: and <2 x i64>
res_vbll = vec_andc(vbll, vbll);
// CHECK: xor <2 x i64>
// CHECK: and <2 x i64>
// CHECK-LE: xor <2 x i64>
// CHECK-LE: and <2 x i64>
res_vf = vec_floor(vf);
// CHECK: call <4 x float> @llvm.floor.v4f32(<4 x float> %{{[0-9]+}})
// CHECK-LE: call <4 x float> @llvm.floor.v4f32(<4 x float> %{{[0-9]+}})
res_vd = vec_floor(vd);
// CHECK: call <2 x double> @llvm.floor.v2f64(<2 x double> %{{[0-9]+}})
// CHECK-LE: call <2 x double> @llvm.floor.v2f64(<2 x double> %{{[0-9]+}})
res_vf = vec_madd(vf, vf, vf);
// CHECK: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}})
// CHECK-LE: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}})
res_vd = vec_madd(vd, vd, vd);
// CHECK: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}})
// CHECK-LE: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}})
/* vec_mergeh */
res_vsll = vec_mergeh(vsll, vsll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vsll = vec_mergeh(vsll, vbll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vsll = vec_mergeh(vbll, vsll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_mergeh(vull, vull);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_mergeh(vull, vbll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_mergeh(vbll, vull);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
/* vec_mergel */
res_vsll = vec_mergel(vsll, vsll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vsll = vec_mergel(vsll, vbll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vsll = vec_mergel(vbll, vsll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_mergel(vull, vull);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_mergel(vull, vbll);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
res_vull = vec_mergel(vbll, vull);
// CHECK: @llvm.ppc.altivec.vperm
// CHECK-LE: @llvm.ppc.altivec.vperm
/* vec_msub */
res_vf = vec_msub(vf, vf, vf);
// CHECK: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}}
// CHECK-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float>
// CHECK-LE: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}}
// CHECK-LE-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float>
res_vd = vec_msub(vd, vd, vd);
// CHECK: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}}
// CHECK-NEXT: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double>
// CHECK-LE: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}}
// CHECK-LE-NEXT: call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double>
res_vsll = vec_mul(vsll, vsll);
// CHECK: mul <2 x i64>
// CHECK-LE: mul <2 x i64>
res_vull = vec_mul(vull, vull);
// CHECK: mul <2 x i64>
// CHECK-LE: mul <2 x i64>
res_vf = vec_mul(vf, vf);
// CHECK: fmul <4 x float> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: fmul <4 x float> %{{[0-9]+}}, %{{[0-9]+}}
res_vd = vec_mul(vd, vd);
// CHECK: fmul <2 x double> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: fmul <2 x double> %{{[0-9]+}}, %{{[0-9]+}}
res_vf = vec_nearbyint(vf);
// CHECK: call <4 x float> @llvm.round.v4f32(<4 x float> %{{[0-9]+}})
// CHECK-LE: call <4 x float> @llvm.round.v4f32(<4 x float> %{{[0-9]+}})
res_vd = vec_nearbyint(vd);
// CHECK: call <2 x double> @llvm.round.v2f64(<2 x double> %{{[0-9]+}})
// CHECK-LE: call <2 x double> @llvm.round.v2f64(<2 x double> %{{[0-9]+}})
res_vf = vec_nmadd(vf, vf, vf);
// CHECK: [[FM:[0-9]+]] = call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}})
// CHECK-NEXT: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %[[FM]]
// CHECK-LE: [[FM:[0-9]+]] = call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}})
// CHECK-LE-NEXT: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %[[FM]]
res_vd = vec_nmadd(vd, vd, vd);
// CHECK: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}})
// CHECK-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]]
// CHECK-LE: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}})
// CHECK-LE-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]]
res_vf = vec_nmsub(vf, vf, vf);
// CHECK: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}}
// CHECK-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float>
// CHECK: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}}
// CHECK-LE: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}}
// CHECK-LE-NEXT: call <4 x float> @llvm.fma.v4f32(<4 x float> %{{[0-9]+}}, <4 x float> %{{[0-9]+}}, <4 x float>
// CHECK-LE: fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{[0-9]+}}
res_vd = vec_nmsub(vd, vd, vd);
// CHECK: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}}
// CHECK-NEXT: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double>
// CHECK-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]]
// CHECK-LE: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %{{[0-9]+}}
// CHECK-LE-NEXT: [[FM:[0-9]+]] = call <2 x double> @llvm.fma.v2f64(<2 x double> %{{[0-9]+}}, <2 x double> %{{[0-9]+}}, <2 x double>
// CHECK-LE-NEXT: fsub <2 x double> <double -0.000000e+00, double -0.000000e+00>, %[[FM]]
/* vec_nor */
res_vsll = vec_nor(vsll, vsll);
// CHECK: or <2 x i64>
// CHECK: xor <2 x i64>
// CHECK-LE: or <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_nor(vull, vull);
// CHECK: or <2 x i64>
// CHECK: xor <2 x i64>
// CHECK-LE: or <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_nor(vbll, vbll);
// CHECK: or <2 x i64>
// CHECK: xor <2 x i64>
// CHECK-LE: or <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vd = vec_nor(vd, vd);
// CHECK: bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK: [[OR:%.+]] = or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-NEXT: xor <2 x i64> [[OR]], <i64 -1, i64 -1>
// CHECK-LE: bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK-LE: [[OR:%.+]] = or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE-NEXT: xor <2 x i64> [[OR]], <i64 -1, i64 -1>
/* vec_or */
res_vsll = vec_or(vsll, vsll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vsll = vec_or(vbll, vsll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vsll = vec_or(vsll, vbll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vull = vec_or(vull, vull);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vull = vec_or(vbll, vull);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vull = vec_or(vull, vbll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vbll = vec_or(vbll, vbll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vd = vec_or(vd, vd);
// CHECK: bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK: or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK-LE: or <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
res_vd = vec_or(vbll, vd);
// CHECK: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK: [[T2:%.+]] = or <2 x i64> %{{[0-9]+}}, [[T1]]
// CHECK: bitcast <2 x i64> [[T2]] to <2 x double>
// CHECK-LE: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK-LE: [[T2:%.+]] = or <2 x i64> %{{[0-9]+}}, [[T1]]
// CHECK-LE: bitcast <2 x i64> [[T2]] to <2 x double>
res_vd = vec_or(vd, vbll);
// CHECK: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK: [[T2:%.+]] = or <2 x i64> [[T1]], %{{[0-9]+}}
// CHECK: bitcast <2 x i64> [[T2]] to <2 x double>
// CHECK-LE: [[T1:%.+]] = bitcast <2 x double> %{{[0-9]+}} to <2 x i64>
// CHECK-LE: [[T2:%.+]] = or <2 x i64> [[T1]], %{{[0-9]+}}
// CHECK-LE: bitcast <2 x i64> [[T2]] to <2 x double>
res_vf = vec_re(vf);
// CHECK: call <4 x float> @llvm.ppc.vsx.xvresp(<4 x float>
// CHECK-LE: call <4 x float> @llvm.ppc.vsx.xvresp(<4 x float>
res_vd = vec_re(vd);
// CHECK: call <2 x double> @llvm.ppc.vsx.xvredp(<2 x double>
// CHECK-LE: call <2 x double> @llvm.ppc.vsx.xvredp(<2 x double>
res_vf = vec_rint(vf);
// CHECK: call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %{{[0-9]+}})
// CHECK-LE: call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %{{[0-9]+}})
res_vd = vec_rint(vd);
// CHECK: call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %{{[0-9]+}})
// CHECK-LE: call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %{{[0-9]+}})
res_vf = vec_rsqrte(vf);
// CHECK: call <4 x float> @llvm.ppc.vsx.xvrsqrtesp(<4 x float> %{{[0-9]+}})
// CHECK-LE: call <4 x float> @llvm.ppc.vsx.xvrsqrtesp(<4 x float> %{{[0-9]+}})
res_vd = vec_rsqrte(vd);
// CHECK: call <2 x double> @llvm.ppc.vsx.xvrsqrtedp(<2 x double> %{{[0-9]+}})
// CHECK-LE: call <2 x double> @llvm.ppc.vsx.xvrsqrtedp(<2 x double> %{{[0-9]+}})
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vf = vec_sel(vd, vd, vbll);
// CHECK: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1>
// CHECK: and <2 x i64> %{{[0-9]+}},
// CHECK: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK: or <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]+}} to <2 x double>
// CHECK-LE: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1>
// CHECK-LE: and <2 x i64> %{{[0-9]+}},
// CHECK-LE: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: or <2 x i64>
// CHECK-LE: bitcast <2 x i64> %{{[0-9]+}} to <2 x double>
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_sel(vd, vd, vull);
// CHECK: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1>
// CHECK: and <2 x i64> %{{[0-9]+}},
// CHECK: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK: or <2 x i64>
// CHECK: bitcast <2 x i64> %{{[0-9]+}} to <2 x double>
// CHECK-LE: xor <2 x i64> %{{[0-9]+}}, <i64 -1, i64 -1>
// CHECK-LE: and <2 x i64> %{{[0-9]+}},
// CHECK-LE: and <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: or <2 x i64>
// CHECK-LE: bitcast <2 x i64> %{{[0-9]+}} to <2 x double>
res_vf = vec_sqrt(vf);
// CHECK: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{[0-9]+}})
// CHECK-LE: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{[0-9]+}})
res_vd = vec_sqrt(vd);
// CHECK: call <2 x double> @llvm.sqrt.v2f64(<2 x double> %{{[0-9]+}})
// CHECK-LE: call <2 x double> @llvm.sqrt.v2f64(<2 x double> %{{[0-9]+}})
res_vd = vec_sub(vd, vd);
// CHECK: fsub <2 x double> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: fsub <2 x double> %{{[0-9]+}}, %{{[0-9]+}}
res_vf = vec_trunc(vf);
// CHECK: call <4 x float> @llvm.trunc.v4f32(<4 x float> %{{[0-9]+}})
// CHECK-LE: call <4 x float> @llvm.trunc.v4f32(<4 x float> %{{[0-9]+}})
res_vd = vec_trunc(vd);
// CHECK: call <2 x double> @llvm.trunc.v2f64(<2 x double> %{{[0-9]+}})
// CHECK-LE: call <2 x double> @llvm.trunc.v2f64(<2 x double> %{{[0-9]+}})
/* vec_vor */
res_vsll = vec_vor(vsll, vsll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vsll = vec_vor(vbll, vsll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vsll = vec_vor(vsll, vbll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vull = vec_vor(vull, vull);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vull = vec_vor(vbll, vull);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vull = vec_vor(vull, vbll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
res_vbll = vec_vor(vbll, vbll);
// CHECK: or <2 x i64>
// CHECK-LE: or <2 x i64>
/* vec_xor */
res_vsll = vec_xor(vsll, vsll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vsll = vec_xor(vbll, vsll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vsll = vec_xor(vsll, vbll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_xor(vull, vull);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_xor(vbll, vull);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_xor(vull, vbll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vbll = vec_xor(vbll, vbll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_xor(vd, vd);
// CHECK: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK: bitcast <2 x i64> [[X1]] to <2 x double>
// CHECK-LE: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: bitcast <2 x i64> [[X1]] to <2 x double>
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_xor(vd, vbll);
// CHECK: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK: bitcast <2 x i64> [[X1]] to <2 x double>
// CHECK-LE: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: bitcast <2 x i64> [[X1]] to <2 x double>
dummy();
// CHECK: call void @dummy()
// CHECK-LE: call void @dummy()
res_vd = vec_xor(vbll, vd);
// CHECK: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK: bitcast <2 x i64> [[X1]] to <2 x double>
// CHECK-LE: [[X1:%.+]] = xor <2 x i64> %{{[0-9]+}}, %{{[0-9]+}}
// CHECK-LE: bitcast <2 x i64> [[X1]] to <2 x double>
/* vec_vxor */
res_vsll = vec_vxor(vsll, vsll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vsll = vec_vxor(vbll, vsll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vsll = vec_vxor(vsll, vbll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_vxor(vull, vull);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_vxor(vbll, vull);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vull = vec_vxor(vull, vbll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vbll = vec_vxor(vbll, vbll);
// CHECK: xor <2 x i64>
// CHECK-LE: xor <2 x i64>
res_vsll = vec_cts(vd, 0);
// CHECK: fmul <2 x double>
// CHECK: fptosi <2 x double> %{{.*}} to <2 x i64>
// CHECK-LE: fmul <2 x double>
// CHECK-LE: fptosi <2 x double> %{{.*}} to <2 x i64>
res_vsll = vec_cts(vd, 31);
// CHECK: fmul <2 x double>
// CHECK: fptosi <2 x double> %{{.*}} to <2 x i64>
// CHECK-LE: fmul <2 x double>
// CHECK-LE: fptosi <2 x double> %{{.*}} to <2 x i64>
res_vsll = vec_ctu(vd, 0);
// CHECK: fmul <2 x double>
// CHECK: fptoui <2 x double> %{{.*}} to <2 x i64>
// CHECK-LE: fmul <2 x double>
// CHECK-LE: fptoui <2 x double> %{{.*}} to <2 x i64>
res_vsll = vec_ctu(vd, 31);
// CHECK: fmul <2 x double>
// CHECK: fptoui <2 x double> %{{.*}} to <2 x i64>
// CHECK-LE: fmul <2 x double>
// CHECK-LE: fptoui <2 x double> %{{.*}} to <2 x i64>
res_vd = vec_ctf(vsll, 0);
// CHECK: sitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK: fmul <2 x double>
// CHECK-LE: sitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK-LE: fmul <2 x double>
res_vd = vec_ctf(vsll, 31);
// CHECK: sitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK: fmul <2 x double>
// CHECK-LE: sitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK-LE: fmul <2 x double>
res_vd = vec_ctf(vull, 0);
// CHECK: uitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK: fmul <2 x double>
// CHECK-LE: uitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK-LE: fmul <2 x double>
res_vd = vec_ctf(vull, 31);
// CHECK: uitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK: fmul <2 x double>
// CHECK-LE: uitofp <2 x i64> %{{.*}} to <2 x double>
// CHECK-LE: fmul <2 x double>
}
/* Function: p7_ViterbiFilter()
* Synopsis: Calculates Viterbi score, vewy vewy fast, in limited precision.
* Incept: SRE, Tue Nov 27 09:15:24 2007 [Janelia]
*
* Purpose: Calculates an approximation of the Viterbi score for sequence
* <dsq> of length <L> residues, using optimized profile <om>,
* and a preallocated one-row DP matrix <ox>. Return the
* estimated Viterbi score (in nats) in <ret_sc>.
*
* Score may overflow (and will, on high-scoring
* sequences), but will not underflow.
*
* The model must be in a local alignment mode; other modes
* cannot provide the necessary guarantee of no underflow.
*
* This is a striped SIMD Viterbi implementation using Intel
* VMX integer intrinsics \citep{Farrar07}, in reduced
* precision (signed words, 16 bits).
*
* Args: dsq - digital target sequence, 1..L
* L - length of dsq in residues
* om - optimized profile
* ox - DP matrix
* ret_sc - RETURN: Viterbi score (in nats)
*
* Returns: <eslOK> on success;
* <eslERANGE> if the score overflows; in this case
* <*ret_sc> is <eslINFINITY>, and the sequence can
* be treated as a high-scoring hit.
*
* Throws: <eslEINVAL> if <ox> allocation is too small, or if
* profile isn't in a local alignment mode. (Must be in local
* alignment mode because that's what helps us guarantee
* limited dynamic range.)
*
* Xref: [Farrar07] for ideas behind striped SIMD DP.
* J2/46-47 for layout of HMMER's striped SIMD DP.
* J2/50 for single row DP.
* J2/60 for reduced precision (epu8)
* J2/65 for initial benchmarking
* J2/66 for precision maximization
* J4/138-140 for reimplementation in 16-bit precision
*/
int
p7_ViterbiFilter(const ESL_DSQ *dsq, int L, const P7_OPROFILE *om, P7_OMX *ox, float *ret_sc)
{
vector signed short mpv, dpv, ipv; /* previous row values */
vector signed short sv; /* temp storage of 1 curr row value in progress */
vector signed short dcv; /* delayed storage of D(i,q+1) */
vector signed short xEv; /* E state: keeps max for Mk->E as we go */
vector signed short xBv; /* B state: splatted vector of B[i-1] for B->Mk calculations */
vector signed short Dmaxv; /* keeps track of maximum D cell on row */
int16_t xE, xB, xC, xJ, xN; /* special states' scores */
int16_t Dmax; /* maximum D cell score on row */
int i; /* counter over sequence positions 1..L */
int q; /* counter over vectors 0..nq-1 */
int Q; /* segment length: # of vectors */
vector signed short *dp; /* using {MDI}MX(q) macro requires initialization of <dp> */
vector signed short *rsc; /* will point at om->ru[x] for residue x[i] */
vector signed short *tsc; /* will point into (and step thru) om->tu */
vector signed short negInfv;
Q = p7O_NQW(om->M);
dp = ox->dpw[0];
/* Check that the DP matrix is ok for us. */
if (Q > ox->allocQ8) ESL_EXCEPTION(eslEINVAL, "DP matrix allocated too small");
if (om->mode != p7_LOCAL && om->mode != p7_UNILOCAL) ESL_EXCEPTION(eslEINVAL, "Fast filter only works for local alignment");
ox->M = om->M;
negInfv = esl_vmx_set_s16((signed short)-32768);
/* Initialization. In unsigned arithmetic, -infinity is -32768
*/
for (q = 0; q < Q; q++)
MMXo(q) = IMXo(q) = DMXo(q) = negInfv;
xN = om->base_w;
xB = xN + om->xw[p7O_N][p7O_MOVE];
xJ = -32768;
xC = -32768;
xE = -32768;
#if p7_DEBUGGING
if (ox->debugging) p7_omx_DumpVFRow(ox, 0, xE, 0, xJ, xB, xC); /* first 0 is <rowi>: do header. second 0 is xN: always 0 here. */
#endif
for (i = 1; i <= L; i++)
{
rsc = om->rwv[dsq[i]];
tsc = om->twv;
dcv = negInfv; /* "-infinity" */
xEv = negInfv;
Dmaxv = negInfv;
xBv = esl_vmx_set_s16(xB);
/* Right shifts by 1 value (2 bytes). 4,8,12,x becomes x,4,8,12.
* Because ia32 is littlendian, this means a left bit shift.
* Zeros shift on automatically; replace it with -32768.
*/
mpv = MMXo(Q-1); mpv = vec_sld(negInfv, mpv, 14);
dpv = DMXo(Q-1); dpv = vec_sld(negInfv, dpv, 14);
ipv = IMXo(Q-1); ipv = vec_sld(negInfv, ipv, 14);
for (q = 0; q < Q; q++)
{
/* Calculate new MMXo(i,q); don't store it yet, hold it in sv. */
sv = vec_adds(xBv, *tsc); tsc++;
sv = vec_max (sv, vec_adds(mpv, *tsc)); tsc++;
sv = vec_max (sv, vec_adds(ipv, *tsc)); tsc++;
sv = vec_max (sv, vec_adds(dpv, *tsc)); tsc++;
sv = vec_adds(sv, *rsc); rsc++;
xEv = vec_max(xEv, sv);
/* Load {MDI}(i-1,q) into mpv, dpv, ipv;
* {MDI}MX(q) is then the current, not the prev row
*/
mpv = MMXo(q);
dpv = DMXo(q);
ipv = IMXo(q);
/* Do the delayed stores of {MD}(i,q) now that memory is usable */
MMXo(q) = sv;
DMXo(q) = dcv;
/* Calculate the next D(i,q+1) partially: M->D only;
* delay storage, holding it in dcv
*/
dcv = vec_adds(sv, *tsc); tsc++;
Dmaxv = vec_max(dcv, Dmaxv);
/* Calculate and store I(i,q) */
sv = vec_adds(mpv, *tsc); tsc++;
IMXo(q)= vec_max(sv, vec_adds(ipv, *tsc)); tsc++;
}
/* Now the "special" states, which start from Mk->E (->C, ->J->B) */
xE = esl_vmx_hmax_s16(xEv);
if (xE >= 32767) { *ret_sc = eslINFINITY; return eslERANGE; } /* immediately detect overflow */
xN = xN + om->xw[p7O_N][p7O_LOOP];
xC = ESL_MAX(xC + om->xw[p7O_C][p7O_LOOP], xE + om->xw[p7O_E][p7O_MOVE]);
xJ = ESL_MAX(xJ + om->xw[p7O_J][p7O_LOOP], xE + om->xw[p7O_E][p7O_LOOP]);
xB = ESL_MAX(xJ + om->xw[p7O_J][p7O_MOVE], xN + om->xw[p7O_N][p7O_MOVE]);
/* and now xB will carry over into next i, and xC carries over after i=L */
/* Finally the "lazy F" loop (sensu [Farrar07]). We can often
* prove that we don't need to evaluate any D->D paths at all.
*
* The observation is that if we can show that on the next row,
* B->M(i+1,k) paths always dominate M->D->...->D->M(i+1,k) paths
* for all k, then we don't need any D->D calculations.
*
* The test condition is:
* max_k D(i,k) + max_k ( TDD(k-2) + TDM(k-1) - TBM(k) ) < xB(i)
* So:
* max_k (TDD(k-2) + TDM(k-1) - TBM(k)) is precalc'ed in om->dd_bound;
* max_k D(i,k) is why we tracked Dmaxv;
* xB(i) was just calculated above.
*/
Dmax = esl_vmx_hmax_s16(Dmaxv);
if (Dmax + om->ddbound_w > xB)
{
/* Now we're obligated to do at least one complete DD path to be sure. */
/* dcv has carried through from end of q loop above */
dcv = vec_sld(negInfv, dcv, 14);
tsc = om->twv + 7*Q; /* set tsc to start of the DD's */
for (q = 0; q < Q; q++)
{
DMXo(q) = vec_max(dcv, DMXo(q));
dcv = vec_adds(DMXo(q), *tsc); tsc++;
}
/* We may have to do up to three more passes; the check
* is for whether crossing a segment boundary can improve
* our score.
*/
do {
dcv = vec_sld(negInfv, dcv, 14);
tsc = om->twv + 7*Q; /* set tsc to start of the DD's */
for (q = 0; q < Q; q++)
{
if (! vec_any_gt(dcv, DMXo(q))) break;
DMXo(q) = vec_max(dcv, DMXo(q));
dcv = vec_adds(DMXo(q), *tsc); tsc++;
}
} while (q == Q);
}
else /* not calculating DD? then just store the last M->D vector calc'ed.*/
DMXo(0) = vec_sld(negInfv, dcv, 14);
#if p7_DEBUGGING
if (ox->debugging) p7_omx_DumpVFRow(ox, i, xE, 0, xJ, xB, xC);
#endif
} /* end loop over sequence residues 1..L */
/* finally C->T */
if (xC > -32768)
{
*ret_sc = (float) xC + (float) om->xw[p7O_C][p7O_MOVE] - (float) om->base_w;
/* *ret_sc += L * om->ncj_roundoff; see J4/150 for rationale: superceded by -3.0nat approximation*/
*ret_sc /= om->scale_w;
*ret_sc -= 3.0; /* the NN/CC/JJ=0,-3nat approximation: see J5/36. That's ~ L \log \frac{L}{L+3}, for our NN,CC,JJ contrib */
}
else *ret_sc = -eslINFINITY;
return eslOK;
}
static int dct_quantize_altivec(MpegEncContext* s,
DCTELEM* data, int n,
int qscale, int* overflow)
{
int lastNonZero;
vector float row0, row1, row2, row3, row4, row5, row6, row7;
vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
const vector float zero = (const vector float)FOUROF(0.);
// used after quantize step
int oldBaseValue = 0;
// Load the data into the row/alt vectors
{
vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
data0 = vec_ld(0, data);
data1 = vec_ld(16, data);
data2 = vec_ld(32, data);
data3 = vec_ld(48, data);
data4 = vec_ld(64, data);
data5 = vec_ld(80, data);
data6 = vec_ld(96, data);
data7 = vec_ld(112, data);
// Transpose the data before we start
TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
// load the data into floating point vectors. We load
// the high half of each row into the main row vectors
// and the low half into the alt vectors.
row0 = vec_ctf(vec_unpackh(data0), 0);
alt0 = vec_ctf(vec_unpackl(data0), 0);
row1 = vec_ctf(vec_unpackh(data1), 0);
alt1 = vec_ctf(vec_unpackl(data1), 0);
row2 = vec_ctf(vec_unpackh(data2), 0);
alt2 = vec_ctf(vec_unpackl(data2), 0);
row3 = vec_ctf(vec_unpackh(data3), 0);
alt3 = vec_ctf(vec_unpackl(data3), 0);
row4 = vec_ctf(vec_unpackh(data4), 0);
alt4 = vec_ctf(vec_unpackl(data4), 0);
row5 = vec_ctf(vec_unpackh(data5), 0);
alt5 = vec_ctf(vec_unpackl(data5), 0);
row6 = vec_ctf(vec_unpackh(data6), 0);
alt6 = vec_ctf(vec_unpackl(data6), 0);
row7 = vec_ctf(vec_unpackh(data7), 0);
alt7 = vec_ctf(vec_unpackl(data7), 0);
}
// The following block could exist as a separate an altivec dct
// function. However, if we put it inline, the DCT data can remain
// in the vector local variables, as floats, which we'll use during the
// quantize step...
{
const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
int whichPass, whichHalf;
for(whichPass = 1; whichPass<=2; whichPass++) {
for(whichHalf = 1; whichHalf<=2; whichHalf++) {
vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
vector float tmp10, tmp11, tmp12, tmp13;
vector float z1, z2, z3, z4, z5;
tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
// dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
row0 = vec_add(tmp10, tmp11);
// dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
row4 = vec_sub(tmp10, tmp11);
// z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
// dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
// CONST_BITS-PASS1_BITS);
row2 = vec_madd(tmp13, vec_0_765366865, z1);
// dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
// CONST_BITS-PASS1_BITS);
row6 = vec_madd(tmp12, vec_1_847759065, z1);
z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
// z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
// z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
z3 = vec_madd(z3, vec_1_961570560, z5);
// z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
z4 = vec_madd(z4, vec_0_390180644, z5);
// The following adds are rolled into the multiplies above
// z3 = vec_add(z3, z5); // z3 += z5;
// z4 = vec_add(z4, z5); // z4 += z5;
// z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
// Wow! It's actually more efficient to roll this multiply
// into the adds below, even thought the multiply gets done twice!
// z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
// z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
// Same with this one...
// z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
// tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
// dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
// tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
// dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
// tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
// dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
// tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
// dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
// Swap the row values with the alts. If this is the first half,
// this sets up the low values to be acted on in the second half.
// If this is the second half, it puts the high values back in
// the row values where they are expected to be when we're done.
SWAP(row0, alt0);
SWAP(row1, alt1);
SWAP(row2, alt2);
SWAP(row3, alt3);
SWAP(row4, alt4);
SWAP(row5, alt5);
SWAP(row6, alt6);
SWAP(row7, alt7);
}
if (whichPass == 1) {
// transpose the data for the second pass
// First, block transpose the upper right with lower left.
SWAP(row4, alt0);
SWAP(row5, alt1);
SWAP(row6, alt2);
SWAP(row7, alt3);
// Now, transpose each block of four
TRANSPOSE4(row0, row1, row2, row3);
TRANSPOSE4(row4, row5, row6, row7);
TRANSPOSE4(alt0, alt1, alt2, alt3);
TRANSPOSE4(alt4, alt5, alt6, alt7);
}
}
}
// perform the quantize step, using the floating point data
// still in the row/alt registers
{
const int* biasAddr;
const vector signed int* qmat;
vector float bias, negBias;
if (s->mb_intra) {
vector signed int baseVector;
// We must cache element 0 in the intra case
// (it needs special handling).
baseVector = vec_cts(vec_splat(row0, 0), 0);
vec_ste(baseVector, 0, &oldBaseValue);
qmat = (vector signed int*)s->q_intra_matrix[qscale];
biasAddr = &(s->intra_quant_bias);
} else {
qmat = (vector signed int*)s->q_inter_matrix[qscale];
biasAddr = &(s->inter_quant_bias);
}
// Load the bias vector (We add 0.5 to the bias so that we're
// rounding when we convert to int, instead of flooring.)
{
vector signed int biasInt;
const vector float negOneFloat = (vector float)FOUROF(-1.0f);
LOAD4(biasInt, biasAddr);
bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
negBias = vec_madd(bias, negOneFloat, zero);
}
{
vector float q0, q1, q2, q3, q4, q5, q6, q7;
q0 = vec_ctf(qmat[0], QMAT_SHIFT);
q1 = vec_ctf(qmat[2], QMAT_SHIFT);
q2 = vec_ctf(qmat[4], QMAT_SHIFT);
q3 = vec_ctf(qmat[6], QMAT_SHIFT);
q4 = vec_ctf(qmat[8], QMAT_SHIFT);
q5 = vec_ctf(qmat[10], QMAT_SHIFT);
q6 = vec_ctf(qmat[12], QMAT_SHIFT);
q7 = vec_ctf(qmat[14], QMAT_SHIFT);
row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
vec_cmpgt(row0, zero));
row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
vec_cmpgt(row1, zero));
row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
vec_cmpgt(row2, zero));
row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
vec_cmpgt(row3, zero));
row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
vec_cmpgt(row4, zero));
row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
vec_cmpgt(row5, zero));
row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
vec_cmpgt(row6, zero));
row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
vec_cmpgt(row7, zero));
q0 = vec_ctf(qmat[1], QMAT_SHIFT);
q1 = vec_ctf(qmat[3], QMAT_SHIFT);
q2 = vec_ctf(qmat[5], QMAT_SHIFT);
q3 = vec_ctf(qmat[7], QMAT_SHIFT);
q4 = vec_ctf(qmat[9], QMAT_SHIFT);
q5 = vec_ctf(qmat[11], QMAT_SHIFT);
q6 = vec_ctf(qmat[13], QMAT_SHIFT);
q7 = vec_ctf(qmat[15], QMAT_SHIFT);
alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
vec_cmpgt(alt0, zero));
alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
vec_cmpgt(alt1, zero));
alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
vec_cmpgt(alt2, zero));
alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
vec_cmpgt(alt3, zero));
alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
vec_cmpgt(alt4, zero));
alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
vec_cmpgt(alt5, zero));
alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
vec_cmpgt(alt6, zero));
alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
vec_cmpgt(alt7, zero));
}
}
// Store the data back into the original block
{
vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
{
// Clamp for overflow
vector signed int max_q_int, min_q_int;
vector signed short max_q, min_q;
LOAD4(max_q_int, &(s->max_qcoeff));
LOAD4(min_q_int, &(s->min_qcoeff));
max_q = vec_pack(max_q_int, max_q_int);
min_q = vec_pack(min_q_int, min_q_int);
data0 = vec_max(vec_min(data0, max_q), min_q);
data1 = vec_max(vec_min(data1, max_q), min_q);
data2 = vec_max(vec_min(data2, max_q), min_q);
data4 = vec_max(vec_min(data4, max_q), min_q);
data5 = vec_max(vec_min(data5, max_q), min_q);
data6 = vec_max(vec_min(data6, max_q), min_q);
data7 = vec_max(vec_min(data7, max_q), min_q);
}
{
vector bool char zero_01, zero_23, zero_45, zero_67;
vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67;
vector signed char negOne = vec_splat_s8(-1);
vector signed char* scanPtr =
(vector signed char*)(s->intra_scantable.inverse);
signed char lastNonZeroChar;
// Determine the largest non-zero index.
zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
vec_cmpeq(data1, (vector signed short)zero));
zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
vec_cmpeq(data3, (vector signed short)zero));
zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
vec_cmpeq(data5, (vector signed short)zero));
zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
vec_cmpeq(data7, (vector signed short)zero));
// 64 biggest values
scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
// 32 largest values
scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67);
// 16 largest values
scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45);
// 8 largest values
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
// 4 largest values
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
// 2 largest values
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
// largest value
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
scanIndexes_01 = vec_splat(scanIndexes_01, 0);
vec_ste(scanIndexes_01, 0, &lastNonZeroChar);
lastNonZero = lastNonZeroChar;
// While the data is still in vectors we check for the transpose IDCT permute
// and handle it using the vector unit if we can. This is the permute used
// by the altivec idct, so it is common when using the altivec dct.
if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
}
vec_st(data0, 0, data);
vec_st(data1, 16, data);
vec_st(data2, 32, data);
vec_st(data3, 48, data);
vec_st(data4, 64, data);
vec_st(data5, 80, data);
vec_st(data6, 96, data);
vec_st(data7, 112, data);
}
}
// special handling of block[0]
if (s->mb_intra) {
if (!s->h263_aic) {
if (n < 4)
oldBaseValue /= s->y_dc_scale;
else
oldBaseValue /= s->c_dc_scale;
}
// Divide by 8, rounding the result
data[0] = (oldBaseValue + 4) >> 3;
}
// We handled the transpose permutation above and we don't
// need to permute the "no" permutation case.
if ((lastNonZero > 0) &&
(s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
(s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) {
ff_block_permute(data, s->dsp.idct_permutation,
s->intra_scantable.scantable, lastNonZero);
}
return lastNonZero;
}
