Create SIMD-accelerated version of compute_gru function

README

@@ -1,4 +1,12 @@
 RNNoise is a noise suppression library based on a recurrent neural network.
+A description of the algorithm is provided in the following paper:
+
+J.-M. Valin, A Hybrid DSP/Deep Learning Approach to Real-Time Full-Band Speech
+Enhancement, Proceedings of IEEE Multimedia Signal Processing (MMSP) Workshop,
+arXiv:1709.08243, 2018.
+https://arxiv.org/pdf/1709.08243.pdf
+
+An interactive demo is available at: https://jmvalin.ca/demo/rnnoise/
 
 To compile, just type:
 % ./autogen.sh

configure.ac

@@ -47,7 +47,7 @@ AC_SUBST(OP_LT_REVISION)
 AC_SUBST(OP_LT_AGE)
 
 CC_CHECK_CFLAGS_APPEND(
-  [-pedantic -Wall -Wextra -Wno-sign-compare -Wno-parentheses -Wno-long-long])
+  [-O3 -march=native -pedantic -Wall -Wextra -Wno-sign-compare -Wno-parentheses -Wno-long-long])
 
 # Platform-specific tweaks
 case $host in

src/compile.sh

@@ -1,3 +1,3 @@
 #!/bin/sh
 
-gcc -DTRAINING=1 -Wall -W -O3 -g -I../include denoise.c kiss_fft.c pitch.c celt_lpc.c rnn.c rnn_data.c -o denoise_training -lm
+gcc -DTRAINING=1 -march=native -Wall -W -O3 -g -I../include denoise.c kiss_fft.c pitch.c celt_lpc.c rnn.c rnn_data.c -o denoise_training -lm

src/denoise.c

@@ -270,6 +270,14 @@ int rnnoise_init(DenoiseState *st, RNNModel *model) {
   st->rnn.vad_gru_state = calloc(sizeof(float), st->rnn.model->vad_gru_size);
   st->rnn.noise_gru_state = calloc(sizeof(float), st->rnn.model->noise_gru_size);
   st->rnn.denoise_gru_state = calloc(sizeof(float), st->rnn.model->denoise_gru_size);
+  st->rnn.compute_gru_fct = &compute_gru;
+
+#if defined(__AVX2__)
+  if(is_avx2_supported() == 1) {
+    st->rnn.compute_gru_fct = &compute_gru_avx2;
+  }
+#endif
+
   return 0;
 }
 
@@ -408,13 +416,11 @@ static void frame_synthesis(DenoiseState *st, float *out, const kiss_fft_cpx *y)
 }
 
 static void biquad(float *y, float mem[2], const float *x, const float *b, const float *a, int N) {
-  int i;
-  for (i=0;i<N;i++) {
-    float xi, yi;
-    xi = x[i];
-    yi = x[i] + mem[0];
-    mem[0] = mem[1] + (b[0]*(double)xi - a[0]*(double)yi);
-    mem[1] = (b[1]*(double)xi - a[1]*(double)yi);
+  for (int i=0;i<N;i++) {
+    float xi = x[i];
+    float yi = xi + mem[0];
+    mem[0] = mem[1] + (b[0] * xi - a[0] * yi);
+    mem[1] = (b[1] * xi - a[1] * yi);
     y[i] = yi;
   }
 }