shithub: opus

Info  •  Files  •  Log  •  Branches

Minor optimization: merging all 3 embeddings

--- a/dnn/lpcnet.c

+++ b/dnn/lpcnet.c

@@ -81,10 +81,14 @@

 {

     float gru_a_input[3*GRU_A_STATE_SIZE];

     float in_b[GRU_A_STATE_SIZE+FEATURE_DENSE2_OUT_SIZE];

+#if 1

+    compute_gru_a_input(gru_a_input, gru_a_condition, GRU_A_STATE_SIZE, &gru_a_embed_sig, last_sig, &gru_a_embed_pred, pred, &gru_a_embed_exc, last_exc);

+#else

     RNN_COPY(gru_a_input, gru_a_condition, 3*GRU_A_STATE_SIZE);

     accum_embedding(&gru_a_embed_sig, gru_a_input, last_sig);

     accum_embedding(&gru_a_embed_pred, gru_a_input, pred);

     accum_embedding(&gru_a_embed_exc, gru_a_input, last_exc);

+#endif

     /*compute_gru3(&gru_a, net->gru_a_state, gru_a_input);*/

     compute_sparse_gru(&sparse_gru_a, net->gru_a_state, gru_a_input);

     RNN_COPY(in_b, net->gru_a_state, GRU_A_STATE_SIZE);

--- a/dnn/nnet.c

+++ b/dnn/nnet.c

@@ -395,6 +395,15 @@

    }    

 }

+void compute_gru_a_input(float *output, const float *input, int N, const EmbeddingLayer *layer1, int val1, const EmbeddingLayer *layer2, int val2, const EmbeddingLayer *layer3, int val3) {

+   int i;

+   for (i=0;i<3*N;i++) {

+      output[i] = input[i] + layer1->embedding_weights[val1*layer1->dim + i]

+                           + layer2->embedding_weights[val2*layer2->dim + i]

+                           + layer3->embedding_weights[val3*layer3->dim + i];

+   }

+}

+

 void accum_embedding(const EmbeddingLayer *layer, float *output, int input)

 {

    int i;

--- a/dnn/nnet.h

+++ b/dnn/nnet.h

@@ -113,6 +113,8 @@

 void accum_embedding(const EmbeddingLayer *layer, float *output, int input);

+void compute_gru_a_input(float *output, const float *input, int N, const EmbeddingLayer *layer1, int val1, const EmbeddingLayer *layer2, int val2, const EmbeddingLayer *layer3, int val3);

+

 int sample_from_pdf(const float *pdf, int N, float exp_boost, float pdf_floor);

 #endif /* _MLP_H_ */

-- 

⑨