int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG dummy) {
BLASLONG m, n, lda, ldb;
FLOAT *beta, *a, *b;
BLASLONG ls, is, js;
BLASLONG min_l, min_i, min_j;
BLASLONG jjs, min_jj;
m = args -> m;
n = args -> n;
a = (FLOAT *)args -> a;
b = (FLOAT *)args -> b;
lda = args -> lda;
ldb = args -> ldb;
beta = (FLOAT *)args -> beta;
if (range_n) {
BLASLONG n_from = *(((BLASLONG *)range_n) + 0);
BLASLONG n_to = *(((BLASLONG *)range_n) + 1);
n = n_to - n_from;
b += n_from * ldb * COMPSIZE;
}
if (beta) {
#ifndef COMPLEX
if (beta[0] != ONE)
GEMM_BETA(m, n, 0, beta[0], NULL, 0, NULL, 0, b, ldb);
if (beta[0] == ZERO) return 0;
#else
if ((beta[0] != ONE) || (beta[1] != ZERO))
GEMM_BETA(m, n, 0, beta[0], beta[1], NULL, 0, NULL, 0, b, ldb);
if ((beta[0] == ZERO) && (beta[1] == ZERO)) return 0;
#endif
}
for(js = 0; js < n; js += GEMM_R){
min_j = n - js;
if (min_j > GEMM_R) min_j = GEMM_R;
#if (!defined(UPPER) && !defined(TRANSA)) || (defined(UPPER) && defined(TRANSA))
for(ls = 0; ls < m; ls += GEMM_Q){
min_l = m - ls;
if (min_l > GEMM_Q) min_l = GEMM_Q;
min_i = min_l;
if (min_i > GEMM_P) min_i = GEMM_P;
#ifndef TRANSA
TRSM_ILTCOPY(min_l, min_i, a + (ls + ls * lda) * COMPSIZE, lda, 0, sa);
#else
TRSM_IUNCOPY(min_l, min_i, a + (ls + ls * lda) * COMPSIZE, lda, 0, sa);
#endif
for(jjs = js; jjs < js + min_j; jjs += min_jj){
min_jj = min_j + js - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
GEMM_ONCOPY(min_l, min_jj, b + (ls + jjs * ldb) * COMPSIZE, ldb, sb + min_l * (jjs - js) * COMPSIZE);
TRSM_KERNEL(min_i, min_jj, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb + min_l * (jjs - js) * COMPSIZE,
b + (ls + jjs * ldb) * COMPSIZE, ldb, 0);
}
for(is = ls + min_i; is < ls + min_l; is += GEMM_P){
min_i = ls + min_l - is;
if (min_i > GEMM_P) min_i = GEMM_P;
#ifndef TRANSA
TRSM_ILTCOPY(min_l, min_i, a + (is + ls * lda) * COMPSIZE, lda, is - ls, sa);
#else
TRSM_IUNCOPY(min_l, min_i, a + (ls + is * lda) * COMPSIZE, lda, is - ls, sa);
#endif
TRSM_KERNEL(min_i, min_j, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb, b + (is + js * ldb) * COMPSIZE, ldb, is - ls);
}
for(is = ls + min_l; is < m; is += GEMM_P){
min_i = m - is;
if (min_i > GEMM_P) min_i = GEMM_P;
#ifndef TRANSA
GEMM_ITCOPY(min_l, min_i, a + (is + ls * lda) * COMPSIZE, lda, sa);
#else
GEMM_INCOPY(min_l, min_i, a + (ls + is * lda) * COMPSIZE, lda, sa);
#endif
GEMM_KERNEL(min_i, min_j, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb, b + (is + js * ldb) * COMPSIZE, ldb);
}
}
#else
BLASLONG start_is;
for(ls = m; ls > 0; ls -= GEMM_Q){
min_l = ls;
if (min_l > GEMM_Q) min_l = GEMM_Q;
start_is = ls - min_l;
while (start_is + GEMM_P < ls) start_is += GEMM_P;
min_i = ls - start_is;
if (min_i > GEMM_P) min_i = GEMM_P;
#ifndef TRANSA
TRSM_IUTCOPY(min_l, min_i, a + (start_is + (ls - min_l) * lda) * COMPSIZE, lda, start_is - (ls - min_l), sa);
#else
TRSM_ILNCOPY(min_l, min_i, a + ((ls - min_l) + start_is * lda) * COMPSIZE, lda, start_is - (ls - min_l), sa);
#endif
for(jjs = js; jjs < js + min_j; jjs += min_jj){
min_jj = min_j + js - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
GEMM_ONCOPY(min_l, min_jj, b + (ls - min_l + jjs * ldb) * COMPSIZE, ldb, sb + min_l * (jjs - js) * COMPSIZE);
TRSM_KERNEL(min_i, min_jj, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb + min_l * (jjs - js) * COMPSIZE,
b + (start_is + jjs * ldb) * COMPSIZE, ldb, start_is - ls + min_l);
}
for(is = start_is - GEMM_P; is >= ls - min_l; is -= GEMM_P){
min_i = ls - is;
if (min_i > GEMM_P) min_i = GEMM_P;
#ifndef TRANSA
TRSM_IUTCOPY(min_l, min_i, a + (is + (ls - min_l) * lda) * COMPSIZE, lda, is - (ls - min_l), sa);
#else
TRSM_ILNCOPY(min_l, min_i, a + ((ls - min_l) + is * lda) * COMPSIZE, lda, is - (ls - min_l), sa);
#endif
TRSM_KERNEL(min_i, min_j, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb,
b + (is + js * ldb) * COMPSIZE, ldb, + is - (ls - min_l) );
}
for(is = 0; is < ls - min_l; is += GEMM_P){
min_i = ls - min_l - is;
if (min_i > GEMM_P) min_i = GEMM_P;
#ifndef TRANSA
GEMM_ITCOPY(min_l, min_i, a + (is + (ls - min_l) * lda) * COMPSIZE, lda, sa);
#else
GEMM_INCOPY(min_l, min_i, a + ((ls - min_l) + is * lda) * COMPSIZE, lda, sa);
#endif
GEMM_KERNEL(min_i, min_j, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb, b + (is + js * ldb) * COMPSIZE, ldb);
}
}
#endif
}
return 0;
}
blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG myid) {
BLASLONG n, lda;
FLOAT *a;
BLASLONG info;
BLASLONG bk, blocking;
BLASLONG is, min_i;
BLASLONG jjs, min_jj;
BLASLONG range_N[2];
BLASLONG j, js, min_j;
#ifdef SHARED_ARRAY
FLOAT *aa;
#endif
FLOAT *sb2 = (FLOAT *)((((BLASLONG)sb
+ GEMM_PQ * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)
+ GEMM_OFFSET_B);
n = args -> n;
a = (FLOAT *)args -> a;
lda = args -> lda;
if (range_n) {
n = range_n[1] - range_n[0];
a += range_n[0] * (lda + 1) * COMPSIZE;
}
if (n <= DTB_ENTRIES / 2) {
info = POTF2_U(args, NULL, range_n, sa, sb, 0);
return info;
}
blocking = GEMM_Q;
if (n <= 4 * GEMM_Q) blocking = (n + 3) / 4;
for (j = 0; j < n; j += blocking) {
bk = n - j;
if (bk > blocking) bk = blocking;
if (!range_n) {
range_N[0] = j;
range_N[1] = j + bk;
} else {
range_N[0] = range_n[0] + j;
range_N[1] = range_n[0] + j + bk;
}
info = CNAME(args, NULL, range_N, sa, sb, 0);
if (info) return info + j;
if (n - j - bk > 0) {
TRSM_IUNCOPY(bk, bk, a + (j + j * lda) * COMPSIZE, lda, 0, sb);
for(js = j + bk; js < n; js += REAL_GEMM_R) {
min_j = n - js;
if (min_j > REAL_GEMM_R) min_j = REAL_GEMM_R;
for(jjs = js; jjs < js + min_j; jjs += GEMM_UNROLL_N){
min_jj = min_j + js - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
GEMM_ONCOPY(bk, min_jj, a + (j + jjs * lda) * COMPSIZE, lda, sb2 + bk * (jjs - js) * COMPSIZE);
for (is = 0; is < bk; is += GEMM_P) {
min_i = bk - is;
if (min_i > GEMM_P) min_i = GEMM_P;
TRSM_KERNEL (min_i, min_jj, bk, dm1,
#ifdef COMPLEX
ZERO,
#endif
sb + bk * is * COMPSIZE,
sb2 + bk * (jjs - js) * COMPSIZE,
a + (j + is + jjs * lda) * COMPSIZE, lda, is);
}
}
for (is = j + bk; is < js + min_j; is += min_i) {
min_i = js + min_j - is;
if (min_i >= GEMM_P * 2) {
min_i = GEMM_P;
} else
if (min_i > GEMM_P) {
min_i = (min_i / 2 + GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
}
#ifdef SHARED_ARRAY
if ((is >= js) && (is + min_i <= js + min_j)) {
aa = sb2 + bk * (is - js) * COMPSIZE;
} else {
GEMM_INCOPY(bk, min_i, a + (j + is * lda) * COMPSIZE, lda, sa);
aa = sa;
}
#else
GEMM_INCOPY(bk, min_i, a + (j + is * lda) * COMPSIZE, lda, sa);
#endif
SYRK_KERNEL_U(min_i, min_j, bk,
dm1,
SA, sb2,
a + (is + js * lda) * COMPSIZE, lda,
is - js);
}
}
}
}
return 0;
}
static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG mypos){
FLOAT *buffer[DIVIDE_RATE];
BLASLONG k, lda;
BLASLONG m_from, m_to;
FLOAT *alpha;
FLOAT *a, *c;
job_t *job = (job_t *)args -> common;
BLASLONG xxx, bufferside;
BLASLONG jjs, min_jj;
BLASLONG is, min_i, div_n;
BLASLONG i, current;
k = K;
a = (FLOAT *)A;
c = (FLOAT *)C;
lda = LDA;
alpha = (FLOAT *)args -> alpha;
m_from = range_n[mypos + 0];
m_to = range_n[mypos + 1];
#if 0
fprintf(stderr, "Thread[%ld] m_from : %ld m_to : %ld\n", mypos, m_from, m_to);
#endif
div_n = ((m_to - m_from + DIVIDE_RATE - 1) / DIVIDE_RATE + GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
buffer[0] = (FLOAT *)((((BLASULONG)(sb + k * k * COMPSIZE) + GEMM_ALIGN) & ~GEMM_ALIGN) + GEMM_OFFSET_B);
for (i = 1; i < DIVIDE_RATE; i++) {
buffer[i] = buffer[i - 1] + GEMM_Q * div_n * COMPSIZE;
}
#ifndef LOWER
TRSM_IUNCOPY(k, k, (FLOAT *)S, lda, 0, sb);
#else
TRSM_OLTCOPY(k, k, (FLOAT *)S, lda, 0, sb);
#endif
for (xxx = m_from, bufferside = 0; xxx < m_to; xxx += div_n, bufferside ++) {
for(jjs = xxx; jjs < MIN(m_to, xxx + div_n); jjs += min_jj){
min_jj = MIN(m_to, xxx + div_n) - jjs;
#ifndef LOWER
if (min_jj > GEMM_UNROLL_MN) min_jj = GEMM_UNROLL_MN;
#else
if (min_jj > GEMM_P) min_jj = GEMM_P;
#endif
#ifndef LOWER
OCOPY_OPERATION (k, min_jj, a, lda, 0, jjs, buffer[bufferside] + k * (jjs - xxx) * COMPSIZE);
TRSM_KERNEL (k, min_jj, k, dm1,
#ifdef COMPLEX
ZERO,
#endif
sb,
buffer[bufferside] + k * (jjs - xxx) * COMPSIZE,
a + jjs * lda * COMPSIZE, lda, 0);
#else
ICOPY_OPERATION (k, min_jj, a, lda, 0, jjs, buffer[bufferside] + k * (jjs - xxx) * COMPSIZE);
TRSM_KERNEL (min_jj, k, k, dm1,
#ifdef COMPLEX
ZERO,
#endif
buffer[bufferside] + k * (jjs - xxx) * COMPSIZE,
sb,
a + jjs * COMPSIZE, lda, 0);
#endif
}
#ifndef LOWER
for (i = 0; i <= mypos; i++)
job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside];
#else
for (i = mypos; i < args -> nthreads; i++)
job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside];
#endif
WMB;
}
min_i = m_to - m_from;
if (min_i >= GEMM_P * 2) {
min_i = GEMM_P;
} else
if (min_i > GEMM_P) {
min_i = ((min_i + 1) / 2 + GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
}
#ifndef LOWER
ICOPY_OPERATION(k, min_i, a, lda, 0, m_from, sa);
#else
OCOPY_OPERATION(k, min_i, a, lda, 0, m_from, sa);
#endif
current = mypos;
#ifndef LOWER
while (current < args -> nthreads)
#else
while (current >= 0)
#endif
{
div_n = ((range_n[current + 1] - range_n[current] + DIVIDE_RATE - 1) / DIVIDE_RATE + GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
for (xxx = range_n[current], bufferside = 0; xxx < range_n[current + 1]; xxx += div_n, bufferside ++) {
/* thread has to wait */
if (current != mypos) while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;};
KERNEL_OPERATION(min_i, MIN(range_n[current + 1] - xxx, div_n), k, alpha,
sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
c, lda, m_from, xxx);
if (m_from + min_i >= m_to) {
job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
WMB;
}
}
#ifndef LOWER
current ++;
#else
current --;
#endif
}
for(is = m_from + min_i; is < m_to; is += min_i){
min_i = m_to - is;
if (min_i >= GEMM_P * 2) {
min_i = GEMM_P;
} else
if (min_i > GEMM_P) {
min_i = ((min_i + 1) / 2 + GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
}
#ifndef LOWER
ICOPY_OPERATION(k, min_i, a, lda, 0, is, sa);
#else
OCOPY_OPERATION(k, min_i, a, lda, 0, is, sa);
#endif
current = mypos;
#ifndef LOWER
while (current < args -> nthreads)
#else
while (current >= 0)
#endif
{
div_n = ((range_n[current + 1] - range_n[current] + DIVIDE_RATE - 1) / DIVIDE_RATE + GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
for (xxx = range_n[current], bufferside = 0; xxx < range_n[current + 1]; xxx += div_n, bufferside ++) {
KERNEL_OPERATION(min_i, MIN(range_n[current + 1] - xxx, div_n), k, alpha,
sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
c, lda, is, xxx);
if (is + min_i >= m_to) {
job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
WMB;
}
}
#ifndef LOWER
current ++;
#else
current --;
#endif
}
}
for (i = 0; i < args -> nthreads; i++) {
if (i != mypos) {
for (xxx = 0; xxx < DIVIDE_RATE; xxx++) {
while (job[mypos].working[i][CACHE_LINE_SIZE * xxx] ) {YIELDING;};
}
}
}
return 0;
}