Refactor conformal factor evaluation.

1 files changed, 29 insertions, 0 deletions

diff --git a/init.c b/init.c
index 72dabed..0cc1f32 100644
--- a/init.c
+++ b/init.c
@@ -23,14 +23,41 @@
 
 #include "brill_data.h"
 #include "basis.h"
+#include "cpu.h"
 #include "internal.h"
 #include "qfunc.h"
+#include "threadpool.h"
+
+double bdi_scalarproduct_metric_fma3(size_t len1, size_t len2, double *mat,
+                                     double *vec1, double *vec2);
+double bdi_scalarproduct_metric_avx(size_t len1, size_t len2, double *mat,
+                                    double *vec1, double *vec2);
+double bdi_scalarproduct_metric_sse3(size_t len1, size_t len2, double *mat,
+                                     double *vec1, double *vec2);
+double bdi_scalarproduct_metric_c(size_t len1, size_t len2, double *mat,
+                                  double *vec1, double *vec2);
 
 static int brill_init_check_options(BDContext *bd)
 {
     BDPriv *s = bd->priv;
     int ret;
 
+    bdi_init_cpu_flags(bd);
+
+    s->scalarproduct_metric = bdi_scalarproduct_metric_c;
+    if (EXTERNAL_SSE3(s->cpu_flags))
+        s->scalarproduct_metric = bdi_scalarproduct_metric_sse3;
+    if (EXTERNAL_AVX(s->cpu_flags))
+        s->scalarproduct_metric = bdi_scalarproduct_metric_avx;
+    if (EXTERNAL_FMA3(s->cpu_flags))
+        s->scalarproduct_metric = bdi_scalarproduct_metric_fma3;
+
+    if (!bd->nb_threads) {
+        bd->nb_threads = bdi_cpu_count();
+        if (!bd->nb_threads)
+            bd->nb_threads = 1;
+    }
+
     ret = bdi_qfunc_init(bd, &s->qfunc, bd->q_func_type,
                          bd->amplitude, bd->eppley_n);
     if (!s->qfunc)
@@ -87,6 +114,8 @@ BDContext *bd_context_alloc(void)
     if (!bd)
         return NULL;
 
+    bd->nb_threads = 1;
+
     bd->q_func_type = BD_Q_FUNC_GUNDLACH;
     bd->amplitude   = 1.0;
     bd->eppley_n    = 5;