From 9f5f953a95c8617cd2fdd24f7402af632c1b07c2 Mon Sep 17 00:00:00 2001 From: tslil Date: Sat, 19 Sep 2026 20:24:08 +0100 Subject: transposing the weights gives ~2.5x speedup --- include/weights_6.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'include/weights_6.h') diff --git a/include/weights_6.h b/include/weights_6.h index be64253..2a6ef2f 100644 --- a/include/weights_6.h +++ b/include/weights_6.h @@ -21,10 +21,10 @@ #define SIX_INP_NUM (2 + 6 * 6) #define SIX_DENSE_NUM 64 -#define NN_ALIGNED __attribute__((aligned(16))) +#define NN_ALIGNED __attribute__((aligned(32))) extern const float SIX_embedding[7]; -extern const float SIX_dense_weights[SIX_DENSE_NUM][SIX_INP_NUM] NN_ALIGNED; +extern const float SIX_dense_weights[SIX_INP_NUM][SIX_DENSE_NUM] NN_ALIGNED; extern const float SIX_dense_biases[SIX_DENSE_NUM]; extern const float SIX_logit_weights[SIX_DENSE_NUM] NN_ALIGNED; extern const float SIX_logit_bias; -- cgit v1.3.1