1 /* Copyright (c) 2022 Amazon
2 Written by Jan Buethe */
3 /*
4 Redistribution and use in source and binary forms, with or without
5 modification, are permitted provided that the following conditions
6 are met:
7
8 - Redistributions of source code must retain the above copyright
9 notice, this list of conditions and the following disclaimer.
10
11 - Redistributions in binary form must reproduce the above copyright
12 notice, this list of conditions and the following disclaimer in the
13 documentation and/or other materials provided with the distribution.
14
15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
16 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
17 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
18 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
19 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
20 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
21 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
22 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
23 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
24 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
25 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 */
27
28 #include <math.h>
29
30 #ifdef HAVE_CONFIG_H
31 #include "config.h"
32 #endif
33
34
35 #include "dred_rdovae_enc.h"
36 #include "os_support.h"
37 #include "dred_rdovae_constants.h"
38
conv1_cond_init(float * mem,int len,int dilation,int * init)39 static void conv1_cond_init(float *mem, int len, int dilation, int *init)
40 {
41 if (!*init) {
42 int i;
43 for (i=0;i<dilation;i++) OPUS_CLEAR(&mem[i*len], len);
44 }
45 *init = 1;
46 }
47
dred_rdovae_encode_dframe(RDOVAEEncState * enc_state,const RDOVAEEnc * model,float * latents,float * initial_state,const float * input,int arch)48 void dred_rdovae_encode_dframe(
49 RDOVAEEncState *enc_state, /* io: encoder state */
50 const RDOVAEEnc *model,
51 float *latents, /* o: latent vector */
52 float *initial_state, /* o: initial state */
53 const float *input, /* i: double feature frame (concatenated) */
54 int arch
55 )
56 {
57 float padded_latents[DRED_PADDED_LATENT_DIM];
58 float padded_state[DRED_PADDED_STATE_DIM];
59 float buffer[ENC_DENSE1_OUT_SIZE + ENC_GRU1_OUT_SIZE + ENC_GRU2_OUT_SIZE + ENC_GRU3_OUT_SIZE + ENC_GRU4_OUT_SIZE + ENC_GRU5_OUT_SIZE
60 + ENC_CONV1_OUT_SIZE + ENC_CONV2_OUT_SIZE + ENC_CONV3_OUT_SIZE + ENC_CONV4_OUT_SIZE + ENC_CONV5_OUT_SIZE];
61 float state_hidden[GDENSE1_OUT_SIZE];
62 int output_index = 0;
63
64 /* run encoder stack and concatenate output in buffer*/
65 compute_generic_dense(&model->enc_dense1, &buffer[output_index], input, ACTIVATION_TANH, arch);
66 output_index += ENC_DENSE1_OUT_SIZE;
67
68 compute_generic_gru(&model->enc_gru1_input, &model->enc_gru1_recurrent, enc_state->gru1_state, buffer, arch);
69 OPUS_COPY(&buffer[output_index], enc_state->gru1_state, ENC_GRU1_OUT_SIZE);
70 output_index += ENC_GRU1_OUT_SIZE;
71 conv1_cond_init(enc_state->conv1_state, output_index, 1, &enc_state->initialized);
72 compute_generic_conv1d(&model->enc_conv1, &buffer[output_index], enc_state->conv1_state, buffer, output_index, ACTIVATION_TANH, arch);
73 output_index += ENC_CONV1_OUT_SIZE;
74
75 compute_generic_gru(&model->enc_gru2_input, &model->enc_gru2_recurrent, enc_state->gru2_state, buffer, arch);
76 OPUS_COPY(&buffer[output_index], enc_state->gru2_state, ENC_GRU2_OUT_SIZE);
77 output_index += ENC_GRU2_OUT_SIZE;
78 conv1_cond_init(enc_state->conv2_state, output_index, 2, &enc_state->initialized);
79 compute_generic_conv1d_dilation(&model->enc_conv2, &buffer[output_index], enc_state->conv2_state, buffer, output_index, 2, ACTIVATION_TANH, arch);
80 output_index += ENC_CONV2_OUT_SIZE;
81
82 compute_generic_gru(&model->enc_gru3_input, &model->enc_gru3_recurrent, enc_state->gru3_state, buffer, arch);
83 OPUS_COPY(&buffer[output_index], enc_state->gru3_state, ENC_GRU3_OUT_SIZE);
84 output_index += ENC_GRU3_OUT_SIZE;
85 conv1_cond_init(enc_state->conv3_state, output_index, 2, &enc_state->initialized);
86 compute_generic_conv1d_dilation(&model->enc_conv3, &buffer[output_index], enc_state->conv3_state, buffer, output_index, 2, ACTIVATION_TANH, arch);
87 output_index += ENC_CONV3_OUT_SIZE;
88
89 compute_generic_gru(&model->enc_gru4_input, &model->enc_gru4_recurrent, enc_state->gru4_state, buffer, arch);
90 OPUS_COPY(&buffer[output_index], enc_state->gru4_state, ENC_GRU4_OUT_SIZE);
91 output_index += ENC_GRU4_OUT_SIZE;
92 conv1_cond_init(enc_state->conv4_state, output_index, 2, &enc_state->initialized);
93 compute_generic_conv1d_dilation(&model->enc_conv4, &buffer[output_index], enc_state->conv4_state, buffer, output_index, 2, ACTIVATION_TANH, arch);
94 output_index += ENC_CONV4_OUT_SIZE;
95
96 compute_generic_gru(&model->enc_gru5_input, &model->enc_gru5_recurrent, enc_state->gru5_state, buffer, arch);
97 OPUS_COPY(&buffer[output_index], enc_state->gru5_state, ENC_GRU5_OUT_SIZE);
98 output_index += ENC_GRU5_OUT_SIZE;
99 conv1_cond_init(enc_state->conv5_state, output_index, 2, &enc_state->initialized);
100 compute_generic_conv1d_dilation(&model->enc_conv5, &buffer[output_index], enc_state->conv5_state, buffer, output_index, 2, ACTIVATION_TANH, arch);
101 output_index += ENC_CONV5_OUT_SIZE;
102
103 compute_generic_dense(&model->enc_zdense, padded_latents, buffer, ACTIVATION_LINEAR, arch);
104 OPUS_COPY(latents, padded_latents, DRED_LATENT_DIM);
105
106 /* next, calculate initial state */
107 compute_generic_dense(&model->gdense1, state_hidden, buffer, ACTIVATION_TANH, arch);
108 compute_generic_dense(&model->gdense2, padded_state, state_hidden, ACTIVATION_LINEAR, arch);
109 OPUS_COPY(initial_state, padded_state, DRED_STATE_DIM);
110 }
111