diff options
| author | tslil <tslil@posteo.de> | 2026-09-19 20:24:08 +0100 |
|---|---|---|
| committer | tslil <tslil@posteo.de> | 2026-09-19 20:32:46 +0100 |
| commit | 9f5f953a95c8617cd2fdd24f7402af632c1b07c2 (patch) | |
| tree | d78abd447065baf3d100935a2cfbb181d50b0fb6 /include/weights_5.h | |
| parent | fe7268110d93e2abe4d4b51ac64e83521a8d6382 (diff) | |
transposing the weights gives ~2.5x speedup
Diffstat (limited to 'include/weights_5.h')
| -rw-r--r-- | include/weights_5.h | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/include/weights_5.h b/include/weights_5.h index d48a05a..74a4366 100644 --- a/include/weights_5.h +++ b/include/weights_5.h @@ -22,10 +22,10 @@ #define FIVE_DENSE_NUM 64 // this does some kind of magic vectorisation -#define NN_ALIGNED __attribute__((aligned(16))) +#define NN_ALIGNED __attribute__((aligned(32))) extern const float FIVE_embedding[7]; -extern const float FIVE_dense_weights[FIVE_DENSE_NUM][FIVE_INP_NUM] NN_ALIGNED; +extern const float FIVE_dense_weights[FIVE_INP_NUM][FIVE_DENSE_NUM] NN_ALIGNED; extern const float FIVE_dense_biases[FIVE_DENSE_NUM]; extern const float FIVE_logit_weights[FIVE_DENSE_NUM] NN_ALIGNED; extern const float FIVE_logit_bias; |