Mixer: Switch back to f32 to get better saturation behavior

But use arm CMSIS functions this time to keep it fast!

software/src/Audio/AudioMixer.h

@@ -12,54 +12,44 @@ class AudioMixer : public AudioStream {
 
   void gain(size_t channel, float gain) {
     if (channel >= NumChannels) return;
-    int8_t shift = 0;
-    while (gain < -1.0f || gain > 0.999969f) {
-      gain /= 2.0f;
-      shift++;
-    }
-    scale_fracts[channel] = float_to_q15(gain);
-    shifts[channel] = shift;
+    gains[channel] = gain;
   }
 
   virtual void update(void) {
-    audio_block_t* out = NULL;
-    q15_t aux[AUDIO_BLOCK_SAMPLES];
+    float aux[AUDIO_BLOCK_SAMPLES];
+    float out[AUDIO_BLOCK_SAMPLES];
+    bool out_filled = false;
 
+    // Originally I had this scaling and mixing all with q15 but that has much
+    // less desirable saturation behavior with phase correlated material.
     for (size_t channel = 0; channel < NumChannels; channel++) {
-      q15_t scalar = scale_fracts[channel];
-      int8_t shift = shifts[channel];
       auto* in = receiveReadOnly(channel);
       if (in) {
-        if (!out) {
-          out = allocate();
-          if (!out) {
-            release(in);
-            return;
-          }
-          arm_fill_q15(0, out->data, AUDIO_BLOCK_SAMPLES);
+        float gain = gains[channel];
+        if (!out_filled) {
+          arm_fill_f32(0.0f, out, AUDIO_BLOCK_SAMPLES);
+          out_filled = true;
         }
-        if (scalar != 0) {
-          // Don't scale for unity gain; that's represented by 0.5 in q15 with a
-          // shift of 1
-          if (scalar != 16384 || shift != 1) {
-            arm_scale_q15(in->data, scalar, shift, aux, AUDIO_BLOCK_SAMPLES);
-            arm_add_q15(aux, out->data, out->data, AUDIO_BLOCK_SAMPLES);
-          } else {
-            arm_add_q15(in->data, out->data, out->data, AUDIO_BLOCK_SAMPLES);
+        if (gain != 0.0f) {
+          arm_q15_to_float(in->data, aux, AUDIO_BLOCK_SAMPLES);
+          if (gain != 1.0f) {
+            arm_scale_f32(aux, gain, aux, AUDIO_BLOCK_SAMPLES);
           }
+          arm_add_f32(aux, out, out, AUDIO_BLOCK_SAMPLES);
         }
         release(in);
       }
     }
 
-    if (out != NULL) {
-      transmit(out);
-      release(out);
+    if (out_filled) {
+      audio_block_t* out_block = allocate();
+      arm_float_to_q15(out, out_block->data, AUDIO_BLOCK_SAMPLES);
+      transmit(out_block);
+      release(out_block);
     }
   }
 
 private:
   audio_block_t* inputQueueArray[NumChannels];
-  std::array<q15_t, NumChannels> scale_fracts;
-  std::array<int8_t, NumChannels> shifts;
+  std::array<float, NumChannels> gains = {0.0f};
 };