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 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;
}
