ref: 3eac8c12e476f68eaa7608cb078379e980d3a7ab
parent: 83e95a5ce6619720131e13fa3f39bb8ab3119ecf
author: Jean-Marc Valin <jmvalin@amazon.com>
date: Tue Aug 1 19:48:34 EDT 2023
Avoid sin()/cos() calls for pitch reference
--- a/dnn/fwgan.c
+++ b/dnn/fwgan.c
@@ -48,15 +48,34 @@
#define FWGAN_FEATURES (NB_FEATURES-1)
-static void pitch_embeddings(float *pembed, double *phase, double w0) {
+static void pitch_embeddings(float *pembed, float *phase, double w0) {
int i;
- /* FIXME: This could be speeded up by making phase a unit-norm complex value, rotating it
- by exp(-i*w0) each sample, and renormalizing once in while. */
+ float wreal, wimag;
+#if 1
+ /* This Taylor expansion should be good enough since w0 is always small. */
+ float w2 = w0*w0;
+ wreal = 1 - .5*w2*(1.f - 0.083333333f*w2);
+ wimag = w0*(1 - 0.166666667f*w2*(1.f - 0.05f*w2));
+#else
+ wreal = cos(w0);
+ wimag = sin(w0);
+#endif
+ /* Speed-up phase reference by making phase a unit-norm complex value and rotating it
+ by exp(-i*w0) each sample. */
for (i=0;i<SUBFRAME_SIZE;i++) {
- *phase += w0;
- pembed[i] = sin(*phase);
- pembed[SUBFRAME_SIZE+i] = cos(*phase);
+ float tmp;
+ tmp = phase[0]*wreal - phase[1]*wimag;
+ phase[1] = phase[0]*wimag + phase[1]*wreal;
+ phase[0] = tmp;
+ pembed[i] = phase[1];
+ pembed[SUBFRAME_SIZE+i] = phase[0];
}
+ /* Renormalize once per sub-frame, though we could probably do it even less frequently. */
+ {
+ float r = 1.f/sqrt(phase[0]*phase[0] + phase[1]*phase[1]);
+ phase[0] *= r;
+ phase[1] *= r;
+ }
}
static void run_fwgan_upsampler(FWGANState *st, float *cond, const float *features)
@@ -76,6 +95,7 @@
float tmp1[MAX_CONT_SIZE];
float tmp2[MAX_CONT_SIZE];
FWGAN *model;
+ st->embed_phase[0] = 1;
model = &st->model;
norm2 = celt_inner_prod(pcm0, pcm0, CONT_PCM_INPUTS, st->arch);
norm_1 = 1.f/sqrt(1e-8f + norm2);
@@ -158,7 +178,7 @@
FWGAN *model;
model = &st->model;
- pitch_embeddings(pembed, &st->embed_phase, w0);
+ pitch_embeddings(pembed, st->embed_phase, w0);
/* Interleave bfcc_cond and pembed for each subframe in feat_in. */
OPUS_COPY(&feat_in[BFCC_WITH_CORR_UPSAMPLER_FC_OUT_SIZE/4], &cond[0], BFCC_WITH_CORR_UPSAMPLER_FC_OUT_SIZE/4);
OPUS_COPY(&feat_in[0], &pembed[0], FWGAN_FRAME_SIZE/2);
--- a/dnn/fwgan.h
+++ b/dnn/fwgan.h
@@ -45,7 +45,7 @@
FWGAN model;
int arch;
int cont_initialized;
- double embed_phase;
+ float embed_phase[2];
float last_gain;
float last_lpc[LPC_ORDER];
float syn_mem[LPC_ORDER];
--
⑨