int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (*function)(), void *sa, void *sb, BLASLONG nthreads) {
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range[MAX_CPU_NUMBER + 1];
BLASLONG width, i, num_cpu;
if (!range_n) {
range[0] = 0;
i = arg -> n;
} else {
range[0] = range_n[0];
i = range_n[1] - range_n[0];
}
num_cpu = 0;
while (i > 0){
width = blas_quickdivide(i + nthreads - num_cpu - 1, nthreads - num_cpu);
i -= width;
if (i < 0) width = width + i;
range[num_cpu + 1] = range[num_cpu] + width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = function;
queue[num_cpu].args = arg;
queue[num_cpu].range_m = range_m;
queue[num_cpu].range_n = &range[num_cpu];
#if 0 //defined(LOONGSON3A)
queue[num_cpu].sa = sa + GEMM_OFFSET_A1 * num_cpu;
queue[num_cpu].sb = queue[num_cpu].sa + GEMM_OFFSET_A1 * 5;
#else
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
#endif
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
}
if (num_cpu) {
#if 0 //defined(LOONGSON3A)
queue[0].sa = sa;
queue[0].sb = sa + GEMM_OFFSET_A1 * 5;
#else
queue[0].sa = sa;
queue[0].sb = sb;
#endif
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu,
queue);
}
return 0;
}
static void _init_thread_memory(void *buffer) {
blas_queue_t queue[MAX_CPU_NUMBER];
int num_cpu;
for (num_cpu = 0; num_cpu < blas_num_threads; num_cpu++) {
blas_queue_init(&queue[num_cpu]);
queue[num_cpu].mode = BLAS_DOUBLE | BLAS_REAL;
queue[num_cpu].routine = &_touch_memory;
queue[num_cpu].args = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
}
queue[num_cpu - 1].next = NULL;
queue[0].sa = buffer;
exec_blas(num_cpu, queue);
}
int CNAME(BLASLONG m, FLOAT alpha, FLOAT *x, BLASLONG incx, FLOAT *y, BLASLONG incy, FLOAT *a, FLOAT *buffer, int nthreads){
#else
int CNAME(BLASLONG m, FLOAT *alpha, FLOAT *x, BLASLONG incx, FLOAT *y, BLASLONG incy, FLOAT *a, FLOAT *buffer, int nthreads){
#endif
blas_arg_t args;
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range_m[MAX_CPU_NUMBER + 1];
BLASLONG width, i, num_cpu;
double dnum;
int mask = 7;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
args.m = m;
args.a = (void *)x;
args.b = (void *)y;
args.c = (void *)a;
args.lda = incx;
args.ldb = incy;
#ifndef COMPLEX
args.alpha = (void *)α
#else
args.alpha = (void *)alpha;
#endif
dnum = (double)m * (double)m / (double)nthreads;
num_cpu = 0;
#ifndef LOWER
range_m[MAX_CPU_NUMBER] = m;
i = 0;
while (i < m){
if (nthreads - num_cpu > 1) {
double di = (double)(m - i);
if (di * di - dnum > 0) {
width = ((BLASLONG)(-sqrt(di * di - dnum) + di) + mask) & ~mask;
} else {
width = m - i;
}
if (width < 16) width = 16;
if (width > m - i) width = m - i;
} else {
width = m - i;
}
range_m[MAX_CPU_NUMBER - num_cpu - 1] = range_m[MAX_CPU_NUMBER - num_cpu] - width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = syr_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[MAX_CPU_NUMBER - num_cpu - 1];
queue[num_cpu].range_n = NULL;
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#else
range_m[0] = 0;
i = 0;
while (i < m){
if (nthreads - num_cpu > 1) {
double di = (double)(m - i);
if (di * di - dnum > 0) {
width = ((BLASLONG)(-sqrt(di * di - dnum) + di) + mask) & ~mask;
} else {
width = m - i;
}
if (width < 16) width = 16;
if (width > m - i) width = m - i;
} else {
width = m - i;
}
range_m[num_cpu + 1] = range_m[num_cpu] + width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = syr_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[num_cpu];
queue[num_cpu].range_n = NULL;
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#endif
if (num_cpu) {
queue[0].sa = NULL;
queue[0].sb = buffer;
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
return 0;
}
static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
*range_n, FLOAT *sa, FLOAT *sb, BLASLONG mypos){
blas_arg_t newarg;
job_t job[MAX_CPU_NUMBER];
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range_M[MAX_CPU_NUMBER + 1];
BLASLONG range_N[MAX_CPU_NUMBER + 1];
BLASLONG num_cpu_m, num_cpu_n;
BLASLONG nthreads = args -> nthreads;
BLASLONG width, i, j, k, js;
BLASLONG m, n, n_from, n_to;
int mode;
#ifndef COMPLEX
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_REAL | BLAS_NODE;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_REAL | BLAS_NODE;
#else
mode = BLAS_SINGLE | BLAS_REAL | BLAS_NODE;
#endif
#else
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_COMPLEX | BLAS_NODE;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_COMPLEX | BLAS_NODE;
#else
mode = BLAS_SINGLE | BLAS_COMPLEX | BLAS_NODE;
#endif
#endif
newarg.m = args -> m;
newarg.n = args -> n;
newarg.k = args -> k;
newarg.a = args -> a;
newarg.b = args -> b;
newarg.c = args -> c;
newarg.lda = args -> lda;
newarg.ldb = args -> ldb;
newarg.ldc = args -> ldc;
newarg.alpha = args -> alpha;
newarg.beta = args -> beta;
newarg.nthreads = args -> nthreads;
newarg.common = (void *)job;
#ifdef PARAMTEST
newarg.gemm_p = args -> gemm_p;
newarg.gemm_q = args -> gemm_q;
newarg.gemm_r = args -> gemm_r;
#endif
if (!range_m) {
range_M[0] = 0;
m = args -> m;
} else {
range_M[0] = range_m[0];
m = range_m[1] - range_m[0];
}
num_cpu_m = 0;
while (m > 0){
width = blas_quickdivide(m + nthreads - num_cpu_m - 1, nthreads - num_cpu_m);
m -= width;
if (m < 0) width = width + m;
range_M[num_cpu_m + 1] = range_M[num_cpu_m] + width;
num_cpu_m ++;
}
for (i = 0; i < num_cpu_m; i++) {
queue[i].mode = mode;
queue[i].routine = inner_thread;
queue[i].args = &newarg;
queue[i].range_m = &range_M[i];
queue[i].range_n = &range_N[0];
queue[i].sa = NULL;
queue[i].sb = NULL;
queue[i].next = &queue[i + 1];
}
queue[0].sa = sa;
queue[0].sb = sb;
if (!range_n) {
n_from = 0;
n_to = args -> n;
} else {
n_from = range_n[0];
n_to = range_n[1];
}
for(js = n_from; js < n_to; js += GEMM_R * nthreads){
n = n_to - js;
if (n > GEMM_R * nthreads) n = GEMM_R * nthreads;
range_N[0] = js;
num_cpu_n = 0;
while (n > 0){
width = blas_quickdivide(n + nthreads - num_cpu_n - 1, nthreads - num_cpu_n);
n -= width;
if (n < 0) width = width + n;
range_N[num_cpu_n + 1] = range_N[num_cpu_n] + width;
num_cpu_n ++;
}
for (j = 0; j < num_cpu_m; j++) {
for (i = 0; i < num_cpu_m; i++) {
for (k = 0; k < DIVIDE_RATE; k++) {
job[j].working[i][CACHE_LINE_SIZE * k] = 0;
}
}
}
queue[num_cpu_m - 1].next = NULL;
exec_blas(num_cpu_m, queue);
}
return 0;
}
int CNAME(BLASLONG m, FLOAT *a, FLOAT *x, BLASLONG incx, FLOAT *buffer, int nthreads){
#else
int CNAME(BLASLONG m, FLOAT *a, FLOAT *x, BLASLONG incx, FLOAT *buffer, int nthreads){
#endif
blas_arg_t args;
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range_m[MAX_CPU_NUMBER + 1];
BLASLONG range_n[MAX_CPU_NUMBER];
BLASLONG width, i, num_cpu;
double dnum;
int mask = 7;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
args.m = m;
args.a = (void *)a;
args.b = (void *)x;
args.c = (void *)(buffer);
args.ldb = incx;
args.ldc = incx;
dnum = (double)m * (double)m / (double)nthreads;
num_cpu = 0;
#ifndef LOWER
range_m[MAX_CPU_NUMBER] = m;
i = 0;
while (i < m){
if (nthreads - num_cpu > 1) {
double di = (double)(m - i);
if (di * di - dnum > 0) {
width = ((BLASLONG)(-sqrt(di * di - dnum) + di) + mask) & ~mask;
} else {
width = m - i;
}
if (width < 16) width = 16;
if (width > m - i) width = m - i;
} else {
width = m - i;
}
range_m[MAX_CPU_NUMBER - num_cpu - 1] = range_m[MAX_CPU_NUMBER - num_cpu] - width;
range_n[num_cpu] = num_cpu * (((m + 15) & ~15) + 16);
queue[num_cpu].mode = mode;
queue[num_cpu].routine = tpmv_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[MAX_CPU_NUMBER - num_cpu - 1];
queue[num_cpu].range_n = &range_n[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#else
range_m[0] = 0;
i = 0;
while (i < m){
if (nthreads - num_cpu > 1) {
double di = (double)(m - i);
if (di * di - dnum > 0) {
width = ((BLASLONG)(-sqrt(di * di - dnum) + di) + mask) & ~mask;
} else {
width = m - i;
}
if (width < 16) width = 16;
if (width > m - i) width = m - i;
} else {
width = m - i;
}
range_m[num_cpu + 1] = range_m[num_cpu] + width;
range_n[num_cpu] = num_cpu * (((m + 15) & ~15) + 16);
queue[num_cpu].mode = mode;
queue[num_cpu].routine = tpmv_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[num_cpu];
queue[num_cpu].range_n = &range_n[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#endif
if (num_cpu) {
queue[0].sa = NULL;
queue[0].sb = buffer + num_cpu * (((m + 255) & ~255) + 16) * COMPSIZE;
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
#ifndef TRANS
for (i = 1; i < num_cpu; i ++) {
#ifndef LOWER
AXPYU_K(range_m[MAX_CPU_NUMBER - i], 0, 0, ONE,
#ifdef COMPLEX
ZERO,
#endif
buffer + range_n[i] * COMPSIZE, 1, buffer, 1, NULL, 0);
#else
AXPYU_K(m - range_m[i], 0, 0, ONE,
#ifdef COMPLEX
ZERO,
#endif
buffer + (range_n[i] + range_m[i]) * COMPSIZE, 1, buffer + range_m[i] * COMPSIZE, 1, NULL, 0);
#endif
}
#endif
COPY_K(m, buffer, 1, x, incx);
return 0;
}
int blas_level1_thread(int mode, BLASLONG m, BLASLONG n, BLASLONG k, void *alpha,
void *a, BLASLONG lda,
void *b, BLASLONG ldb,
void *c, BLASLONG ldc, int (*function)(), int nthreads){
blas_queue_t queue[MAX_CPU_NUMBER];
blas_arg_t args [MAX_CPU_NUMBER];
BLASLONG i, width, astride, bstride;
int num_cpu, calc_type;
calc_type = (mode & BLAS_PREC) + ((mode & BLAS_COMPLEX) != 0) + 2;
mode |= BLAS_LEGACY;
for (i = 0; i < nthreads; i++) blas_queue_init(&queue[i]);
num_cpu = 0;
i = m;
while (i > 0){
/* Adjust Parameters */
width = blas_quickdivide(i + nthreads - num_cpu - 1,
nthreads - num_cpu);
i -= width;
if (i < 0) width = width + i;
astride = width * lda;
if (!(mode & BLAS_TRANSB_T)) {
bstride = width * ldb;
} else {
bstride = width;
}
astride <<= calc_type;
bstride <<= calc_type;
args[num_cpu].m = width;
args[num_cpu].n = n;
args[num_cpu].k = k;
args[num_cpu].a = (void *)a;
args[num_cpu].b = (void *)b;
args[num_cpu].c = (void *)c;
args[num_cpu].lda = lda;
args[num_cpu].ldb = ldb;
args[num_cpu].ldc = ldc;
args[num_cpu].alpha = alpha;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = function;
queue[num_cpu].args = &args[num_cpu];
queue[num_cpu].next = &queue[num_cpu + 1];
a = (void *)((BLASULONG)a + astride);
b = (void *)((BLASULONG)b + bstride);
num_cpu ++;
}
if (num_cpu) {
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
return 0;
}
int CNAME(BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG incx, FLOAT *y, BLASLONG incy, FLOAT *buffer, int nthreads){
#else
int CNAME(BLASLONG n, BLASLONG k, FLOAT *alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG incx, FLOAT *y, BLASLONG incy, FLOAT *buffer, int nthreads){
#endif
blas_arg_t args;
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range_m[MAX_CPU_NUMBER + 1];
BLASLONG range_n[MAX_CPU_NUMBER];
BLASLONG width, i, num_cpu;
double dnum;
int mask = 7;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
args.n = n;
args.k = k;
args.a = (void *)a;
args.b = (void *)x;
args.c = (void *)buffer;
args.lda = lda;
args.ldb = incx;
args.ldc = incy;
dnum = (double)n * (double)n / (double)nthreads;
num_cpu = 0;
if (n < 2 * k) {
#ifndef LOWER
range_m[MAX_CPU_NUMBER] = n;
i = 0;
while (i < n){
if (nthreads - num_cpu > 1) {
double di = (double)(n - i);
if (di * di - dnum > 0) {
width = ((BLASLONG)(-sqrt(di * di - dnum) + di) + mask) & ~mask;
} else {
width = n - i;
}
if (width < 16) width = 16;
if (width > n - i) width = n - i;
} else {
width = n - i;
}
range_m[MAX_CPU_NUMBER - num_cpu - 1] = range_m[MAX_CPU_NUMBER - num_cpu] - width;
range_n[num_cpu] = num_cpu * (((n + 15) & ~15) + 16);
queue[num_cpu].mode = mode;
queue[num_cpu].routine = sbmv_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[MAX_CPU_NUMBER - num_cpu - 1];
queue[num_cpu].range_n = &range_n[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#else
range_m[0] = 0;
i = 0;
while (i < n){
if (nthreads - num_cpu > 1) {
double di = (double)(n - i);
if (di * di - dnum > 0) {
width = ((BLASLONG)(-sqrt(di * di - dnum) + di) + mask) & ~mask;
} else {
width = n - i;
}
if (width < 16) width = 16;
if (width > n - i) width = n - i;
} else {
width = n - i;
}
range_m[num_cpu + 1] = range_m[num_cpu] + width;
range_n[num_cpu] = num_cpu * (((n + 15) & ~15) + 16);
queue[num_cpu].mode = mode;
queue[num_cpu].routine = sbmv_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[num_cpu];
queue[num_cpu].range_n = &range_n[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#endif
} else {
range_m[0] = 0;
i = n;
while (i > 0){
width = blas_quickdivide(i + nthreads - num_cpu - 1, nthreads - num_cpu);
if (width < 4) width = 4;
if (i < width) width = i;
range_m[num_cpu + 1] = range_m[num_cpu] + width;
range_n[num_cpu] = num_cpu * ((n + 15) & ~15);
queue[num_cpu].mode = mode;
queue[num_cpu].routine = sbmv_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[num_cpu];
queue[num_cpu].range_n = &range_n[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i -= width;
}
}
if (num_cpu) {
queue[0].sa = NULL;
queue[0].sb = buffer;
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
for (i = 1; i < num_cpu; i ++) {
AXPYU_K(n, 0, 0,
#ifndef COMPLEX
ONE,
#else
ONE, ZERO,
#endif
(FLOAT*)(queue[i].sb), 1, buffer, 1, NULL, 0);
}
AXPYU_K(n, 0, 0,
#ifndef COMPLEX
alpha,
#else
alpha[0], alpha[1],
#endif
buffer, 1, y, incy, NULL, 0);
return 0;
}
int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (*function)(), void *sa, void *sb, BLASLONG nthreads) {
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range_M[MAX_CPU_NUMBER + 1], range_N[MAX_CPU_NUMBER + 1];
BLASLONG procs, total_procs, num_cpu_m, num_cpu_n;
BLASLONG width, i, j;
BLASLONG divM, divN;
divM = divide_rule[nthreads][0];
divN = divide_rule[nthreads][1];
if (!range_m) {
range_M[0] = 0;
i = arg -> m;
} else {
range_M[0] = range_M[0];
i = range_M[1] - range_M[0];
}
num_cpu_m = 0;
while (i > 0){
width = blas_quickdivide(i + divM - num_cpu_m - 1, divM - num_cpu_m);
i -= width;
if (i < 0) width = width + i;
range_M[num_cpu_m + 1] = range_M[num_cpu_m] + width;
num_cpu_m ++;
}
if (!range_n) {
range_N[0] = 0;
i = arg -> n;
} else {
range_N[0] = range_n[0];
i = range_n[1] - range_n[0];
}
num_cpu_n = 0;
while (i > 0){
width = blas_quickdivide(i + divN - num_cpu_n - 1, divN - num_cpu_n);
i -= width;
if (i < 0) width = width + i;
range_N[num_cpu_n + 1] = range_N[num_cpu_n] + width;
num_cpu_n ++;
}
procs = 0;
for (j = 0; j < num_cpu_n; j++) {
for (i = 0; i < num_cpu_m; i++) {
queue[procs].mode = mode;
queue[procs].routine = function;
queue[procs].args = arg;
queue[procs].range_m = &range_M[i];
queue[procs].range_n = &range_N[j];
queue[procs].sa = NULL;
queue[procs].sb = NULL;
queue[procs].next = &queue[procs + 1];
procs ++;
}
}
if (procs) {
queue[0].sa = sa;
queue[0].sb = sb;
queue[procs - 1].next = NULL;
exec_blas(procs, queue);
}
return 0;
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG ku, BLASLONG kl, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG incx, FLOAT *y, BLASLONG incy, FLOAT *buffer, int nthreads){
#else
int CNAME(BLASLONG m, BLASLONG n, BLASLONG ku, BLASLONG kl, FLOAT *alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG incx, FLOAT *y, BLASLONG incy, FLOAT *buffer, int nthreads){
#endif
blas_arg_t args;
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range_m[MAX_CPU_NUMBER];
BLASLONG range_n[MAX_CPU_NUMBER + 1];
BLASLONG width, i, num_cpu;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
args.m = m;
args.n = n;
args.a = (void *)a;
args.b = (void *)x;
args.c = (void *)buffer;
args.lda = lda;
args.ldb = incx;
args.ldc = ku;
args.ldd = kl;
num_cpu = 0;
range_n[0] = 0;
i = n;
while (i > 0){
width = blas_quickdivide(i + nthreads - num_cpu - 1, nthreads - num_cpu);
if (width < 4) width = 4;
if (i < width) width = i;
range_n[num_cpu + 1] = range_n[num_cpu] + width;
#ifndef TRANSA
range_m[num_cpu] = num_cpu * ((m + 15) & ~15);
#else
range_m[num_cpu] = num_cpu * ((n + 15) & ~15);
#endif
queue[num_cpu].mode = mode;
queue[num_cpu].routine = gbmv_kernel;
queue[num_cpu].args = &args;
queue[num_cpu].range_m = &range_m[num_cpu];
queue[num_cpu].range_n = &range_n[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i -= width;
}
if (num_cpu) {
queue[0].sa = NULL;
#ifndef TRANSA
queue[0].sb = buffer + num_cpu * (((m + 255) & ~255) + 16) * COMPSIZE;
#else
queue[0].sb = buffer + num_cpu * (((n + 255) & ~255) + 16) * COMPSIZE;
#endif
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
for (i = 1; i < num_cpu; i ++) {
AXPYU_K(
#ifndef TRANSA
m,
#else
n,
#endif
0, 0,
#ifndef COMPLEX
ONE,
#else
ONE, ZERO,
#endif
buffer + range_m[i] * COMPSIZE, 1, buffer, 1, NULL, 0);
}
AXPYU_K(
#ifndef TRANSA
m,
#else
n,
#endif
0, 0,
#ifndef COMPLEX
alpha,
#else
alpha[0], alpha[1],
#endif
buffer, 1, y, incy, NULL, 0);
return 0;
}
static int thread_driver(blas_arg_t *args, FLOAT *sa, FLOAT *sb){
blas_arg_t newarg;
#ifndef USE_ALLOC_HEAP
job_t job[MAX_CPU_NUMBER];
#else
job_t * job = NULL;
#endif
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range[MAX_CPU_NUMBER + 100];
BLASLONG num_cpu;
BLASLONG nthreads = args -> nthreads;
BLASLONG width, i, j, k;
BLASLONG n, n_from, n_to;
int mode, mask;
double dnum;
#ifndef COMPLEX
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_REAL;
mask = MAX(QGEMM_UNROLL_M, QGEMM_UNROLL_N) - 1;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_REAL;
mask = MAX(DGEMM_UNROLL_M, DGEMM_UNROLL_N) - 1;
#else
mode = BLAS_SINGLE | BLAS_REAL;
mask = MAX(SGEMM_UNROLL_M, SGEMM_UNROLL_N) - 1;
#endif
#else
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_COMPLEX;
mask = MAX(XGEMM_UNROLL_M, XGEMM_UNROLL_N) - 1;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_COMPLEX;
mask = MAX(ZGEMM_UNROLL_M, ZGEMM_UNROLL_N) - 1;
#else
mode = BLAS_SINGLE | BLAS_COMPLEX;
mask = MAX(CGEMM_UNROLL_M, CGEMM_UNROLL_N) - 1;
#endif
#endif
newarg.m = args -> m;
newarg.k = args -> k;
newarg.a = args -> a;
newarg.b = args -> b;
newarg.c = args -> c;
newarg.lda = args -> lda;
newarg.alpha = args -> alpha;
#ifdef USE_ALLOC_HEAP
job = (job_t*)malloc(MAX_CPU_NUMBER * sizeof(job_t));
if(job==NULL){
fprintf(stderr, "OpenBLAS: malloc failed in %s\n", __func__);
exit(1);
}
#endif
newarg.common = (void *)job;
n_from = 0;
n_to = args -> m;
#ifndef LOWER
range[MAX_CPU_NUMBER] = n_to - n_from;
range[0] = 0;
num_cpu = 0;
i = 0;
n = n_to - n_from;
dnum = (double)n * (double)n /(double)nthreads;
while (i < n){
if (nthreads - num_cpu > 1) {
double di = (double)i;
width = (((BLASLONG)(sqrt(di * di + dnum) - di) + mask) & ~mask);
if (num_cpu == 0) width = n - ((n - width) & ~mask);
if ((width > n - i) || (width < mask)) width = n - i;
} else {
width = n - i;
}
range[MAX_CPU_NUMBER - num_cpu - 1] = range[MAX_CPU_NUMBER - num_cpu] - width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = inner_thread;
queue[num_cpu].args = &newarg;
queue[num_cpu].range_m = NULL;
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
for (i = 0; i < num_cpu; i ++) queue[i].range_n = &range[MAX_CPU_NUMBER - num_cpu];
#else
range[0] = 0;
num_cpu = 0;
i = 0;
n = n_to - n_from;
dnum = (double)n * (double)n /(double)nthreads;
while (i < n){
if (nthreads - num_cpu > 1) {
double di = (double)i;
width = (((BLASLONG)(sqrt(di * di + dnum) - di) + mask) & ~mask);
if ((width > n - i) || (width < mask)) width = n - i;
} else {
width = n - i;
}
range[num_cpu + 1] = range[num_cpu] + width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = inner_thread;
queue[num_cpu].args = &newarg;
queue[num_cpu].range_m = NULL;
queue[num_cpu].range_n = range;
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#endif
newarg.nthreads = num_cpu;
if (num_cpu) {
for (j = 0; j < num_cpu; j++) {
for (i = 0; i < num_cpu; i++) {
for (k = 0; k < DIVIDE_RATE; k++) {
job[j].working[i][CACHE_LINE_SIZE * k] = 0;
}
}
}
queue[0].sa = sa;
queue[0].sb = sb;
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
#ifdef USE_ALLOC_HEAP
free(job);
#endif
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, ldc;
BLASLONG m_from, m_to, n_from, n_to;
FLOAT *alpha, *beta;
FLOAT *a, *c;
job_t *job = (job_t *)args -> common;
BLASLONG xxx, bufferside;
BLASLONG ls, min_l, jjs, min_jj;
BLASLONG is, min_i, div_n;
BLASLONG i, current;
#ifdef LOWER
BLASLONG start_i;
#endif
#ifdef TIMING
BLASLONG rpcc_counter;
BLASLONG copy_A = 0;
BLASLONG copy_B = 0;
BLASLONG kernel = 0;
BLASLONG waiting1 = 0;
BLASLONG waiting2 = 0;
BLASLONG waiting3 = 0;
BLASLONG waiting6[MAX_CPU_NUMBER];
BLASLONG ops = 0;
for (i = 0; i < args -> nthreads; i++) waiting6[i] = 0;
#endif
k = K;
a = (FLOAT *)A;
c = (FLOAT *)C;
lda = LDA;
ldc = LDC;
alpha = (FLOAT *)args -> alpha;
beta = (FLOAT *)args -> beta;
m_from = 0;
m_to = N;
/* Global Range */
n_from = 0;
n_to = N;
if (range_n) {
m_from = range_n[mypos + 0];
m_to = range_n[mypos + 1];
n_from = range_n[0];
n_to = range_n[args -> nthreads];
}
if (beta) {
#if !defined(COMPLEX) || defined(HERK)
if (beta[0] != ONE)
#else
if ((beta[0] != ONE) || (beta[1] != ZERO))
#endif
syrk_beta(m_from, m_to, n_from, n_to, beta, c, ldc);
}
if ((k == 0) || (alpha == NULL)) return 0;
if ((alpha[0] == ZERO)
#if defined(COMPLEX) && !defined(HERK)
&& (alpha[1] == ZERO)
#endif
) return 0;
#if 0
fprintf(stderr, "Thread[%ld] m_from : %ld m_to : %ld n_from : %ld n_to : %ld\n", mypos, m_from, m_to, n_from, n_to);
#endif
div_n = ((m_to - m_from + DIVIDE_RATE - 1) / DIVIDE_RATE
+ GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
buffer[0] = sb;
for (i = 1; i < DIVIDE_RATE; i++) {
buffer[i] = buffer[i - 1] + GEMM_Q * div_n * COMPSIZE;
}
for(ls = 0; ls < k; ls += min_l){
min_l = k - ls;
if (min_l >= GEMM_Q * 2) {
min_l = GEMM_Q;
} else {
if (min_l > GEMM_Q) min_l = (min_l + 1) / 2;
}
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 / 2 + GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
}
}
#ifdef LOWER
xxx = (m_to - m_from - min_i) % GEMM_P;
if (xxx) min_i -= GEMM_P - xxx;
#endif
START_RPCC();
#ifndef LOWER
ICOPY_OPERATION(min_l, min_i, a, lda, ls, m_from, sa);
#else
ICOPY_OPERATION(min_l, min_i, a, lda, ls, m_to - min_i, sa);
#endif
STOP_RPCC(copy_A);
div_n = ((m_to - m_from + DIVIDE_RATE - 1) / DIVIDE_RATE
+ GEMM_UNROLL_MN - 1) & ~(GEMM_UNROLL_MN - 1);
for (xxx = m_from, bufferside = 0; xxx < m_to; xxx += div_n, bufferside ++) {
START_RPCC();
/* Make sure if no one is using buffer */
#ifndef LOWER
for (i = 0; i < mypos; i++)
#else
for (i = mypos + 1; i < args -> nthreads; i++)
#endif
while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;};
STOP_RPCC(waiting1);
#ifndef LOWER
for(jjs = xxx; jjs < MIN(m_to, xxx + div_n); jjs += min_jj){
min_jj = MIN(m_to, xxx + div_n) - jjs;
if (xxx == m_from) {
if (min_jj > min_i) min_jj = min_i;
} else {
if (min_jj > GEMM_UNROLL_MN) min_jj = GEMM_UNROLL_MN;
}
START_RPCC();
OCOPY_OPERATION(min_l, min_jj, a, lda, ls, jjs,
buffer[bufferside] + min_l * (jjs - xxx) * COMPSIZE);
STOP_RPCC(copy_B);
START_RPCC();
KERNEL_OPERATION(min_i, min_jj, min_l, alpha,
sa, buffer[bufferside] + min_l * (jjs - xxx) * COMPSIZE,
c, ldc, m_from, jjs);
STOP_RPCC(kernel);
#ifdef TIMING
ops += 2 * min_i * min_jj * min_l;
#endif
}
#else
for(jjs = xxx; jjs < MIN(m_to, xxx + div_n); jjs += min_jj){
min_jj = MIN(m_to, xxx + div_n) - jjs;
if (min_jj > GEMM_UNROLL_MN) min_jj = GEMM_UNROLL_MN;
START_RPCC();
OCOPY_OPERATION(min_l, min_jj, a, lda, ls, jjs,
buffer[bufferside] + min_l * (jjs - xxx) * COMPSIZE);
STOP_RPCC(copy_B);
START_RPCC();
KERNEL_OPERATION(min_i, min_jj, min_l, alpha,
sa, buffer[bufferside] + min_l * (jjs - xxx) * COMPSIZE,
c, ldc, m_to - min_i, jjs);
STOP_RPCC(kernel);
#ifdef TIMING
ops += 2 * min_i * min_jj * min_l;
#endif
}
#endif
#ifndef LOWER
for (i = 0; i <= mypos; i++)
#else
for (i = mypos; i < args -> nthreads; i++)
#endif
job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside];
WMB;
}
#ifndef LOWER
current = mypos + 1;
while (current < args -> nthreads) {
#else
current = mypos - 1;
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 ++) {
START_RPCC();
/* thread has to wait */
while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;};
STOP_RPCC(waiting2);
START_RPCC();
#ifndef LOWER
KERNEL_OPERATION(min_i, MIN(range_n[current + 1] - xxx, div_n), min_l, alpha,
sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
c, ldc,
m_from,
xxx);
#else
KERNEL_OPERATION(min_i, MIN(range_n[current + 1] - xxx, div_n), min_l, alpha,
sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
c, ldc,
m_to - min_i,
xxx);
#endif
STOP_RPCC(kernel);
#ifdef TIMING
ops += 2 * min_i * MIN(range_n[current + 1] - xxx, div_n) * min_l;
#endif
if (m_to - m_from == min_i) {
job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
}
}
#ifndef LOWER
current ++;
#else
current --;
#endif
}
#ifndef LOWER
for(is = m_from + min_i; is < m_to; is += min_i){
min_i = m_to - is;
#else
start_i = min_i;
for(is = m_from; is < m_to - start_i; is += min_i){
min_i = m_to - start_i - is;
#endif
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);
}
START_RPCC();
ICOPY_OPERATION(min_l, min_i, a, lda, ls, is, sa);
STOP_RPCC(copy_A);
current = mypos;
do {
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 ++) {
START_RPCC();
KERNEL_OPERATION(min_i, MIN(range_n[current + 1] - xxx, div_n), min_l, alpha,
sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
c, ldc, is, xxx);
STOP_RPCC(kernel);
#ifdef TIMING
ops += 2 * min_i * MIN(range_n[current + 1] - xxx, div_n) * min_l;
#endif
#ifndef LOWER
if (is + min_i >= m_to) {
#else
if (is + min_i >= m_to - start_i) {
#endif
/* Thread doesn't need this buffer any more */
job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
WMB;
}
}
#ifndef LOWER
current ++;
} while (current != args -> nthreads);
#else
current --;
} while (current >= 0);
#endif
}
}
START_RPCC();
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;};
}
}
}
STOP_RPCC(waiting3);
#ifdef TIMING
BLASLONG waiting = waiting1 + waiting2 + waiting3;
BLASLONG total = copy_A + copy_B + kernel + waiting;
fprintf(stderr, "GEMM [%2ld] Copy_A : %6.2f Copy_B : %6.2f Wait1 : %6.2f Wait2 : %6.2f Wait3 : %6.2f Kernel : %6.2f",
mypos, (double)copy_A /(double)total * 100., (double)copy_B /(double)total * 100.,
(double)waiting1 /(double)total * 100.,
(double)waiting2 /(double)total * 100.,
(double)waiting3 /(double)total * 100.,
(double)ops/(double)kernel / 4. * 100.);
#if 0
fprintf(stderr, "GEMM [%2ld] Copy_A : %6.2ld Copy_B : %6.2ld Wait : %6.2ld\n",
mypos, copy_A, copy_B, waiting);
fprintf(stderr, "Waiting[%2ld] %6.2f %6.2f %6.2f\n",
mypos,
(double)waiting1/(double)waiting * 100.,
(double)waiting2/(double)waiting * 100.,
(double)waiting3/(double)waiting * 100.);
#endif
fprintf(stderr, "\n");
#endif
return 0;
}
int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLOAT *sb, BLASLONG mypos){
blas_arg_t newarg;
#ifndef USE_ALLOC_HEAP
job_t job[MAX_CPU_NUMBER];
#else
job_t * job = NULL;
#endif
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range[MAX_CPU_NUMBER + 100];
BLASLONG num_cpu;
BLASLONG nthreads = args -> nthreads;
BLASLONG width, i, j, k;
BLASLONG n, n_from, n_to;
int mode, mask;
double dnum;
if ((nthreads == 1) || (args -> n < nthreads * SWITCH_RATIO)) {
SYRK_LOCAL(args, range_m, range_n, sa, sb, 0);
return 0;
}
#ifndef COMPLEX
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_REAL;
mask = MAX(QGEMM_UNROLL_M, QGEMM_UNROLL_N) - 1;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_REAL;
mask = MAX(DGEMM_UNROLL_M, DGEMM_UNROLL_N) - 1;
#else
mode = BLAS_SINGLE | BLAS_REAL;
mask = MAX(SGEMM_UNROLL_M, SGEMM_UNROLL_N) - 1;
#endif
#else
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_COMPLEX;
mask = MAX(XGEMM_UNROLL_M, XGEMM_UNROLL_N) - 1;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_COMPLEX;
mask = MAX(ZGEMM_UNROLL_M, ZGEMM_UNROLL_N) - 1;
#else
mode = BLAS_SINGLE | BLAS_COMPLEX;
mask = MAX(CGEMM_UNROLL_M, CGEMM_UNROLL_N) - 1;
#endif
#endif
newarg.m = args -> m;
newarg.n = args -> n;
newarg.k = args -> k;
newarg.a = args -> a;
newarg.b = args -> b;
newarg.c = args -> c;
newarg.lda = args -> lda;
newarg.ldb = args -> ldb;
newarg.ldc = args -> ldc;
newarg.alpha = args -> alpha;
newarg.beta = args -> beta;
#ifdef USE_ALLOC_HEAP
job = (job_t*)malloc(MAX_CPU_NUMBER * sizeof(job_t));
if(job==NULL){
fprintf(stderr, "OpenBLAS: malloc failed in %s\n", __func__);
exit(1);
}
#endif
newarg.common = (void *)job;
if (!range_n) {
n_from = 0;
n_to = args -> n;
} else {
n_from = range_n[0];
n_to = range_n[1] - range_n[0];
}
#ifndef LOWER
range[MAX_CPU_NUMBER] = n_to - n_from;
range[0] = 0;
num_cpu = 0;
i = 0;
n = n_to - n_from;
dnum = (double)n * (double)n /(double)nthreads;
while (i < n){
if (nthreads - num_cpu > 1) {
double di = (double)i;
width = (((BLASLONG)(sqrt(di * di + dnum) - di) + mask) & ~mask);
if (num_cpu == 0) width = n - ((n - width) & ~mask);
if ((width > n - i) || (width < mask)) width = n - i;
} else {
width = n - i;
}
range[MAX_CPU_NUMBER - num_cpu - 1] = range[MAX_CPU_NUMBER - num_cpu] - width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = inner_thread;
queue[num_cpu].args = &newarg;
queue[num_cpu].range_m = range_m;
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
for (i = 0; i < num_cpu; i ++) queue[i].range_n = &range[MAX_CPU_NUMBER - num_cpu];
#else
range[0] = 0;
num_cpu = 0;
i = 0;
n = n_to - n_from;
dnum = (double)n * (double)n /(double)nthreads;
while (i < n){
if (nthreads - num_cpu > 1) {
double di = (double)i;
width = (((BLASLONG)(sqrt(di * di + dnum) - di) + mask) & ~mask);
if ((width > n - i) || (width < mask)) width = n - i;
} else {
width = n - i;
}
range[num_cpu + 1] = range[num_cpu] + width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = inner_thread;
queue[num_cpu].args = &newarg;
queue[num_cpu].range_m = range_m;
queue[num_cpu].range_n = range;
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
#endif
newarg.nthreads = num_cpu;
if (num_cpu) {
for (j = 0; j < num_cpu; j++) {
for (i = 0; i < num_cpu; i++) {
for (k = 0; k < DIVIDE_RATE; k++) {
job[j].working[i][CACHE_LINE_SIZE * k] = 0;
}
}
}
queue[0].sa = sa;
queue[0].sb = sb;
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
#ifdef USE_ALLOC_HEAP
free(job);
#endif
return 0;
}
int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (*function)(), void *sa, void *sb, BLASLONG nthreads) {
blas_queue_t queue[MAX_CPU_NUMBER];
BLASLONG range[MAX_CPU_NUMBER + 1];
BLASLONG width, i;
BLASLONG n_from, n_to;
double dnum, nf, nt, di;
int num_cpu;
int mask = 0;
if (!(mode & BLAS_COMPLEX)) {
switch (mode & BLAS_PREC) {
case BLAS_SINGLE:
mask = SGEMM_UNROLL_MN - 1;
break;
case BLAS_DOUBLE:
mask = DGEMM_UNROLL_MN - 1;
break;
#ifdef EXPRECISION
case BLAS_XDOUBLE:
mask = MAX(QGEMM_UNROLL_M, QGEMM_UNROLL_N) - 1;
break;
#endif
}
} else {
switch (mode & BLAS_PREC) {
case BLAS_SINGLE:
mask = CGEMM_UNROLL_MN - 1;
break;
case BLAS_DOUBLE:
mask = ZGEMM_UNROLL_MN - 1;
break;
#ifdef EXPRECISION
case BLAS_XDOUBLE:
mask = MAX(XGEMM_UNROLL_M, XGEMM_UNROLL_N) - 1;
break;
#endif
}
}
n_from = 0;
n_to = arg -> n;
if (range_n) {
n_from = *(range_n + 0);
n_to = *(range_n + 1);
}
if (!(mode & BLAS_UPLO)) {
nf = (double)(n_from);
nt = (double)(n_to);
dnum = (nt * nt - nf * nf) / (double)nthreads;
num_cpu = 0;
range[0] = n_from;
i = n_from;
while (i < n_to){
if (nthreads - num_cpu > 1) {
di = (double)i;
width = ((BLASLONG)( sqrt(di * di + dnum) - di) + mask) & ~mask;
if ((width <= 0) || (width > n_to - i)) width = n_to - i;
} else {
width = n_to - i;
}
range[num_cpu + 1] = range[num_cpu] + width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = function;
queue[num_cpu].args = arg;
queue[num_cpu].range_m = range_m;
queue[num_cpu].range_n = &range[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
} else {
nf = (double)(arg -> n - n_from);
nt = (double)(arg -> n - n_to);
dnum = (nt * nt - nf * nf) / (double)nthreads;
num_cpu = 0;
range[0] = n_from;
i = n_from;
while (i < n_to){
if (nthreads - num_cpu > 1) {
di = (double)(arg -> n - i);
width = ((BLASLONG)(-sqrt(di * di + dnum) + di) + mask) & ~mask;
if ((width <= 0) || (width > n_to - i)) width = n_to - i;
} else {
width = n_to - i;
}
range[num_cpu + 1] = range[num_cpu] + width;
queue[num_cpu].mode = mode;
queue[num_cpu].routine = function;
queue[num_cpu].args = arg;
queue[num_cpu].range_m = range_m;
queue[num_cpu].range_n = &range[num_cpu];
queue[num_cpu].sa = NULL;
queue[num_cpu].sb = NULL;
queue[num_cpu].next = &queue[num_cpu + 1];
num_cpu ++;
i += width;
}
}
if (num_cpu) {
queue[0].sa = sa;
queue[0].sb = sb;
queue[num_cpu - 1].next = NULL;
exec_blas(num_cpu, queue);
}
return 0;
}
