use SIMD to speed up nr_fo_compensation()

openair1/PHY/nr_phy_common/src/nr_phy_common.c

@@ -459,10 +459,34 @@ void nr_fo_compensation(double fo_Hz, int samples_per_ms, int sample_offset, c16
   const double phase_inc = -fo_Hz / (samples_per_ms * 1000);
   double phase = sample_offset * phase_inc;
   phase -= (int)phase;
-  for (int i = 0; i < size; i++) {
+#if 1
-    c16_t rot = get_sin_cos(phase);
+  // The bottleneck is the calculation of the complex rotation values using get_sin_cos().
-    *rxdata_ptr = c16mulShift(*rxdata_ptr, rot, 14);
+  // This code path does not compute these values for the complete OFDM symbol, but only for a smaller CHUNK size.
-    rxdata_ptr++;
+  // After applying the rotation to a CHUNK size of the output, these rotation values are efficiently rotated further by `rot_vec`.
+  // Unfortunately, this propagates small errors from one chunk to the next.
+  // Therefore, there is a tradeoff between speed (better with small CHUNK sizes) and accuracy (better with large CHUNK sizes).
+#define CHUNK 128
+  c16_t rot[CHUNK] __attribute__((aligned(32)));
+  for (int i = 0; i < CHUNK; i++) {
+    rot[i] = get_sin_cos(phase);
     phase += phase_inc;
   }
+  const c16_t rot_vec = get_sin_cos(CHUNK * phase_inc);
+  while (size > CHUNK) {
+    mult_complex_vectors(rxdata_ptr, rot, rxdata_ptr, CHUNK, 14);
+    rxdata_ptr += CHUNK;
+    rotate_cpx_vector(rot, &rot_vec, rot, CHUNK, 14);
+    size -= CHUNK;
+  }
+  mult_complex_vectors(rxdata_ptr, rot, rxdata_ptr, size, 14);
+#else
+  // This code path computes the complex rotation values for the complete OFDM symbol using get_sin_cos().
+  // This is more accurate, but also slower than the code path above.
+  c16_t rot[size] __attribute__((aligned(32)));
+  for (int i = 0; i < size; i++) {
+    rot[i] = get_sin_cos(phase);
+    phase += phase_inc;
+  }
+  mult_complex_vectors(rxdata_ptr, rot, rxdata_ptr, size, 14);
+#endif
 }