Force vectorization for DNN primitives

Avoids having to write intrinsics for simple loops

1 files changed, 11 insertions, 0 deletions

diff --git a/dnn/nnet_arch.h b/dnn/nnet_arch.h
index 3c53e619..b61ebf54 100644
--- a/dnn/nnet_arch.h
+++ b/dnn/nnet_arch.h
@@ -38,6 +38,13 @@
 
 #define RTCD_SUF(name) CAT_SUFFIX(name, RTCD_ARCH)
 
+/* Force vectorization on for DNN code because some of the loops rely on
+   compiler vectorization rather than explicitly using intrinsics. */
+#ifdef __GNUC__
+#pragma GCC push_options
+#pragma GCC optimize("tree-vectorize")
+#endif
+
 
 #define MAX_ACTIVATIONS (4096)
 
@@ -216,4 +223,8 @@ void RTCD_SUF(compute_conv2d_)(const Conv2dLayer *conv, float *out, float *mem,
    }
 }
 
+#ifdef __GNUC__
+#pragma GCC pop_options
+#endif
+
 #endif