blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG myid) {
BLASLONG m, n, lda, offset;
BLASLONG j, js, jmin, is, imin, jc, jcmin;
BLASLONG jjs, min_jj;
blasint *ipiv, iinfo, info;
BLASLONG jb, mn, blocking;
FLOAT *a, *offsetA, *offsetB;
BLASLONG range_N[2];
FLOAT *sbb;
m = args -> m;
n = args -> n;
a = (FLOAT *)args -> a;
lda = args -> lda;
ipiv = (blasint *)args -> c;
offset = 0;
if (range_n) {
m -= range_n[0];
n = range_n[1] - range_n[0];
offset = range_n[0];
a += range_n[0] * (lda + 1) * COMPSIZE;
}
if (m <= 0 || n <= 0) return 0;
mn = MIN(m, n);
blocking = (mn / 2 + GEMM_UNROLL_N - 1) & ~(GEMM_UNROLL_N - 1);
if (blocking > GEMM_Q) blocking = GEMM_Q;
if (blocking <= GEMM_UNROLL_N * 2) {
info = GETF2(args, NULL, range_n, sa, sb, 0);
return info;
}
sbb = (FLOAT *)((((long)(sb + blocking * blocking * COMPSIZE) + GEMM_ALIGN) & ~GEMM_ALIGN) + GEMM_OFFSET_B);
info = 0;
for (j = 0; j < mn; j += blocking) {
jb = mn - j;
if (jb > blocking) jb = blocking;
offsetA = a + j * lda * COMPSIZE;
offsetB = a + (j + jb) * lda * COMPSIZE;
range_N[0] = offset + j;
range_N[1] = offset + j + jb;
iinfo = CNAME(args, NULL, range_N, sa, sb, 0);
if (iinfo && !info) info = iinfo + j;
if (j + jb < n) {
TRSM_ILTCOPY(jb, jb, offsetA + j * COMPSIZE, lda, 0, sb);
for (js = j + jb; js < n; js += REAL_GEMM_R){
jmin = n - js;
if (jmin > REAL_GEMM_R) jmin = REAL_GEMM_R;
for (jjs = js; jjs < js + jmin; jjs += GEMM_UNROLL_N){
min_jj = js + jmin - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
#if 0
LASWP_PLUS(min_jj, j + offset + 1, j + jb + offset, ZERO,
#ifdef COMPLEX
ZERO,
#endif
a + (- offset + jjs * lda) * COMPSIZE, lda, NULL, 0 , ipiv, 1);
GEMM_ONCOPY (jb, min_jj, a + (j + jjs * lda) * COMPSIZE, lda, sbb + jb * (jjs - js) * COMPSIZE);
#else
LASWP_NCOPY(min_jj, j + offset + 1, j + jb + offset,
a + (- offset + jjs * lda) * COMPSIZE, lda, ipiv, sbb + jb * (jjs - js) * COMPSIZE);
#endif
for (jc = 0; jc < jb; jc += GEMM_P) {
jcmin = jb - jc;
if (jcmin > GEMM_P) jcmin = GEMM_P;
TRSM_KERNEL_LT(jcmin, min_jj, jb, dm1,
#ifdef COMPLEX
ZERO,
#endif
sb + jb * jc * COMPSIZE,
sbb + jb * (jjs - js) * COMPSIZE,
a + (j + jc + jjs * lda) * COMPSIZE, lda, jc);
}
}
for (is = j + jb; is < m; is += GEMM_P){
imin = m - is;
if (imin > GEMM_P) imin = GEMM_P;
GEMM_ITCOPY (jb, imin, offsetA + is * COMPSIZE, lda, sa);
GEMM_KERNEL_N(imin, jmin, jb, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sbb, a + (is + js * lda) * COMPSIZE, lda);
}
}
}
}
for (j = 0; j < mn; j += jb) {
jb = MIN(mn - j, blocking);
LASWP_PLUS(jb, j + jb + offset + 1, mn + offset, ZERO,
#ifdef COMPLEX
ZERO,
#endif
a - (offset - j * lda) * COMPSIZE, lda, NULL, 0 , ipiv, 1);
}
return info;
}
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_m) {
BLASLONG m_from = *(((BLASLONG *)range_m) + 0);
BLASLONG m_to = *(((BLASLONG *)range_m) + 1);
m = m_to - m_from;
b += m_from * 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
}
#if (defined(UPPER) && !defined(TRANSA)) || (!defined(UPPER) && defined(TRANSA))
for(js = 0; js < n; js += GEMM_R){
min_j = n - js;
if (min_j > GEMM_R) min_j = GEMM_R;
for(ls = 0; ls < js; ls += GEMM_Q){
min_l = js - ls;
if (min_l > GEMM_Q) min_l = GEMM_Q;
min_i = m;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (ls * ldb) * COMPSIZE, ldb, sa);
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;
#ifndef TRANSA
GEMM_ONCOPY(min_l, min_jj, a + (ls + jjs * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);
#else
GEMM_OTCOPY(min_l, min_jj, a + (jjs + ls * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);
#endif
GEMM_KERNEL(min_i, min_jj, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb + min_l * (jjs - js) * COMPSIZE,
b + (jjs * ldb) * COMPSIZE, ldb);
}
for(is = min_i; is < m; is += GEMM_P){
min_i = m - is;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (is + ls * ldb) * COMPSIZE, ldb, sa);
GEMM_KERNEL(min_i, min_j, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb, b + (is + js * ldb) * COMPSIZE, ldb);
}
}
for(ls = js; ls < js + min_j; ls += GEMM_Q){
min_l = js + min_j - ls;
if (min_l > GEMM_Q) min_l = GEMM_Q;
min_i = m;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (ls * ldb) * COMPSIZE, ldb, sa);
#ifndef TRANSA
TRSM_OUNCOPY(min_l, min_l, a + (ls + ls * lda) * COMPSIZE, lda, 0, sb);
#else
TRSM_OLTCOPY(min_l, min_l, a + (ls + ls * lda) * COMPSIZE, lda, 0, sb);
#endif
TRSM_KERNEL(min_i, min_l, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb,
b + (ls * ldb) * COMPSIZE, ldb, 0);
for(jjs = 0; jjs < min_j - min_l - ls + js; jjs += min_jj){
min_jj = min_j - min_l - ls + js - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
#ifndef TRANSA
GEMM_ONCOPY (min_l, min_jj, a + (ls + (ls + min_l + jjs) * lda) * COMPSIZE, lda,
sb + min_l * (min_l + jjs) * COMPSIZE);
#else
GEMM_OTCOPY (min_l, min_jj, a + ((ls + min_l + jjs) + ls * lda) * COMPSIZE, lda,
sb + min_l * (min_l + jjs) * COMPSIZE);
#endif
GEMM_KERNEL(min_i, min_jj, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb + min_l * (min_l + jjs) * COMPSIZE,
b + (min_l + ls + jjs) * ldb * COMPSIZE, ldb);
}
for(is = min_i; is < m; is += GEMM_P){
min_i = m - is;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (is + ls * ldb) * COMPSIZE, ldb, sa);
TRSM_KERNEL(min_i, min_l, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb,
b + (is + ls * ldb) * COMPSIZE, ldb, 0);
GEMM_KERNEL(min_i, min_j - min_l + js - ls, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb + min_l * min_l * COMPSIZE,
b + (is + ( min_l + ls) * ldb) * COMPSIZE, ldb);
}
}
}
#else
BLASLONG start_ls;
for(js = n; js > 0; js -= GEMM_R){
min_j = js;
if (min_j > GEMM_R) min_j = GEMM_R;
for (ls = js; ls < n; ls += GEMM_Q) {
min_l = n - ls;
if (min_l > GEMM_Q) min_l = GEMM_Q;
min_i = m;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (ls * ldb) * COMPSIZE, ldb, sa);
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;
#ifndef TRANSA
GEMM_ONCOPY(min_l, min_jj, a + (ls + (jjs - min_j) * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);
#else
GEMM_OTCOPY(min_l, min_jj, a + ((jjs - min_j) + ls * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);
#endif
GEMM_KERNEL(min_i, min_jj, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb + min_l * (jjs - js) * COMPSIZE,
b + (jjs - min_j) * ldb * COMPSIZE, ldb);
}
for(is = min_i; is < m; is += GEMM_P){
min_i = m - is;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (is + ls * ldb) * COMPSIZE, ldb, sa);
GEMM_KERNEL(min_i, min_j, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sb, b + (is + (js - min_j) * ldb) * COMPSIZE, ldb);
}
}
start_ls = js - min_j;
while (start_ls + GEMM_Q < js) start_ls += GEMM_Q;
for(ls = start_ls; ls >= js - min_j; ls -= GEMM_Q){
min_l = js - ls;
if (min_l > GEMM_Q) min_l = GEMM_Q;
min_i = m;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (ls * ldb) * COMPSIZE, ldb, sa);
#ifndef TRANSA
TRSM_OLNCOPY(min_l, min_l, a + (ls + ls * lda) * COMPSIZE, lda,
0, sb + min_l * (min_j - js + ls) * COMPSIZE);
#else
TRSM_OUTCOPY(min_l, min_l, a + (ls + ls * lda) * COMPSIZE, lda,
0, sb + min_l * (min_j - js + ls) * COMPSIZE);
#endif
TRSM_KERNEL(min_i, min_l, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb + min_l * (min_j - js + ls) * COMPSIZE,
b + (ls * ldb) * COMPSIZE, ldb, 0);
for(jjs = 0; jjs < min_j - js + ls; jjs += min_jj){
min_jj = min_j - js + ls - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
#ifndef TRANSA
GEMM_ONCOPY (min_l, min_jj, a + (ls + (js - min_j + jjs) * lda) * COMPSIZE, lda,
sb + min_l * jjs * COMPSIZE);
#else
GEMM_OTCOPY (min_l, min_jj, a + ((js - min_j + jjs) + ls * lda) * COMPSIZE, lda,
sb + min_l * jjs * COMPSIZE);
#endif
GEMM_KERNEL(min_i, min_jj, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb + min_l * jjs * COMPSIZE,
b + (js - min_j + jjs) * ldb * COMPSIZE, ldb);
}
for(is = min_i; is < m; is += GEMM_P){
min_i = m - is;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY(min_l, min_i, b + (is + ls * ldb) * COMPSIZE, ldb, sa);
TRSM_KERNEL(min_i, min_l, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb + min_l * (min_j - js + ls) * COMPSIZE,
b + (is + ls * ldb) * COMPSIZE, ldb, 0);
GEMM_KERNEL(min_i, min_j - js + ls, min_l, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa,
sb,
b + (is + (js - min_j) * ldb) * COMPSIZE, ldb);
}
}
}
#endif
return 0;
}
static void inner_basic_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG mypos){
BLASLONG is, min_i;
BLASLONG js, min_j;
BLASLONG jjs, min_jj;
BLASLONG m = args -> m;
BLASLONG n = args -> n;
BLASLONG k = args -> k;
BLASLONG lda = args -> lda;
BLASLONG off = args -> ldb;
FLOAT *b = (FLOAT *)args -> b + (k ) * COMPSIZE;
FLOAT *c = (FLOAT *)args -> b + ( k * lda) * COMPSIZE;
FLOAT *d = (FLOAT *)args -> b + (k + k * lda) * COMPSIZE;
FLOAT *sbb = sb;
volatile BLASLONG *flag = (volatile BLASLONG *)args -> d;
blasint *ipiv = (blasint *)args -> c;
if (range_n) {
n = range_n[1] - range_n[0];
c += range_n[0] * lda * COMPSIZE;
d += range_n[0] * lda * COMPSIZE;
}
if (args -> a == NULL) {
TRSM_ILTCOPY(k, k, (FLOAT *)args -> b, lda, 0, sb);
sbb = (FLOAT *)((((long)(sb + k * k * COMPSIZE) + GEMM_ALIGN) & ~GEMM_ALIGN) + GEMM_OFFSET_B);
} else {
sb = (FLOAT *)args -> a;
}
for (js = 0; 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 = js + min_j - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
if (GEMM_UNROLL_N <= 8) {
LASWP_NCOPY(min_jj, off + 1, off + k,
c + (- off + jjs * lda) * COMPSIZE, lda,
ipiv, sbb + k * (jjs - js) * COMPSIZE);
} else {
LASWP_PLUS(min_jj, off + 1, off + k, ZERO,
#ifdef COMPLEX
ZERO,
#endif
c + (- off + jjs * lda) * COMPSIZE, lda, NULL, 0, ipiv, 1);
GEMM_ONCOPY (k, min_jj, c + jjs * lda * COMPSIZE, lda, sbb + (jjs - js) * k * COMPSIZE);
}
for (is = 0; is < k; is += GEMM_P) {
min_i = k - is;
if (min_i > GEMM_P) min_i = GEMM_P;
TRSM_KERNEL_LT(min_i, min_jj, k, dm1,
#ifdef COMPLEX
ZERO,
#endif
sb + k * is * COMPSIZE,
sbb + (jjs - js) * k * COMPSIZE,
c + (is + jjs * lda) * COMPSIZE, lda, is);
}
}
if ((js + REAL_GEMM_R >= n) && (mypos >= 0)) flag[mypos * CACHE_LINE_SIZE] = 0;
for (is = 0; is < m; is += GEMM_P){
min_i = m - is;
if (min_i > GEMM_P) min_i = GEMM_P;
GEMM_ITCOPY (k, min_i, b + is * COMPSIZE, lda, sa);
GEMM_KERNEL_N(min_i, min_j, k, dm1,
#ifdef COMPLEX
ZERO,
#endif
sa, sbb, d + (is + js * lda) * COMPSIZE, lda);
}
}
}
static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG mypos){
job_t *job = (job_t *)args -> common;
BLASLONG xxx, bufferside;
FLOAT *buffer[DIVIDE_RATE];
BLASLONG jjs, min_jj, div_n;
BLASLONG i, current;
BLASLONG is, min_i;
BLASLONG m, n_from, n_to;
BLASLONG k = args -> k;
BLASLONG lda = args -> lda;
BLASLONG off = args -> ldb;
FLOAT *a = (FLOAT *)args -> b + (k ) * COMPSIZE;
FLOAT *b = (FLOAT *)args -> b + ( k * lda) * COMPSIZE;
FLOAT *c = (FLOAT *)args -> b + (k + k * lda) * COMPSIZE;
FLOAT *sbb= sb;
blasint *ipiv = (blasint *)args -> c;
volatile BLASLONG *flag = (volatile BLASLONG *)args -> d;
if (args -> a == NULL) {
TRSM_ILTCOPY(k, k, (FLOAT *)args -> b, lda, 0, sb);
sbb = (FLOAT *)((((long)(sb + k * k * COMPSIZE) + GEMM_ALIGN) & ~GEMM_ALIGN) + GEMM_OFFSET_B);
} else {
sb = (FLOAT *)args -> a;
}
m = range_m[1] - range_m[0];
n_from = range_n[mypos + 0];
n_to = range_n[mypos + 1];
a += range_m[0] * COMPSIZE;
c += range_m[0] * COMPSIZE;
div_n = (n_to - n_from + DIVIDE_RATE - 1) / DIVIDE_RATE;
buffer[0] = sbb;
for (i = 1; i < DIVIDE_RATE; i++) {
buffer[i] = buffer[i - 1] + GEMM_Q * ((div_n + GEMM_UNROLL_N - 1) & ~(GEMM_UNROLL_N - 1)) * COMPSIZE;
}
for (xxx = n_from, bufferside = 0; xxx < n_to; xxx += div_n, bufferside ++) {
for (i = 0; i < args -> nthreads; i++)
while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {};
for(jjs = xxx; jjs < MIN(n_to, xxx + div_n); jjs += min_jj){
min_jj = MIN(n_to, xxx + div_n) - jjs;
if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
if (GEMM_UNROLL_N <= 8) {
LASWP_NCOPY(min_jj, off + 1, off + k,
b + (- off + jjs * lda) * COMPSIZE, lda,
ipiv, buffer[bufferside] + (jjs - xxx) * k * COMPSIZE);
} else {
LASWP_PLUS(min_jj, off + 1, off + k, ZERO,
#ifdef COMPLEX
ZERO,
#endif
b + (- off + jjs * lda) * COMPSIZE, lda, NULL, 0, ipiv, 1);
GEMM_ONCOPY (k, min_jj, b + jjs * lda * COMPSIZE, lda,
buffer[bufferside] + (jjs - xxx) * k * COMPSIZE);
}
for (is = 0; is < k; is += GEMM_P) {
min_i = k - is;
if (min_i > GEMM_P) min_i = GEMM_P;
TRSM_KERNEL_LT(min_i, min_jj, k, dm1,
#ifdef COMPLEX
ZERO,
#endif
sb + k * is * COMPSIZE,
buffer[bufferside] + (jjs - xxx) * k * COMPSIZE,
b + (is + jjs * lda) * COMPSIZE, lda, is);
}
}
for (i = 0; i < args -> nthreads; i++)
job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside];
}
flag[mypos * CACHE_LINE_SIZE] = 0;
if (m == 0) {
for (xxx = 0; xxx < DIVIDE_RATE; xxx++) {
job[mypos].working[mypos][CACHE_LINE_SIZE * xxx] = 0;
}
}
for(is = 0; is < m; is += min_i){
min_i = m - 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_M - 1) & ~(GEMM_UNROLL_M - 1);
}
ICOPY_OPERATION(k, min_i, a, lda, 0, is, sa);
current = mypos;
do {
div_n = (range_n[current + 1] - range_n[current] + DIVIDE_RATE - 1) / DIVIDE_RATE;
for (xxx = range_n[current], bufferside = 0; xxx < range_n[current + 1]; xxx += div_n, bufferside ++) {
if ((current != mypos) && (!is)) {
while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {};
}
KERNEL_OPERATION(min_i, MIN(range_n[current + 1] - xxx, div_n), k,
sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
c, lda, is, xxx);
if (is + min_i >= m) {
job[current].working[mypos][CACHE_LINE_SIZE * bufferside] = 0;
}
}
current ++;
if (current >= args -> nthreads) current = 0;
} while (current != mypos);
}
for (i = 0; i < args -> nthreads; i++) {
for (xxx = 0; xxx < DIVIDE_RATE; xxx++) {
while (job[mypos].working[i][CACHE_LINE_SIZE * xxx] ) {};
}
}
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;
}
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;
}
