GBA: Fix ADPCM encoding issues (#474)

* ADPCM: Add 8bit debugging, fix limit cycles

* ADPCM: Increase output gain

src/platform/gba/packer/ad4/AD4.h

@@ -1,10 +1,13 @@
 #pragma once
 
+//! Uncommenting this line writes the output to Debug.sb as 8bit signed PCM
+//#define DEBUG_OUTPUT
+
 #include <stdint.h>
+#include <stdio.h>
 
 struct AD4State_t {
 	 int32_t zM1, zM2;
-	 int32_t Tap;
 	 int32_t Quant;
 	 int32_t Output;
 	uint32_t MaxOutputLevel;
@@ -13,7 +16,6 @@ struct AD4State_t {
 void AD4_Init(struct AD4State_t *State) {
 	State->zM1    = 0;
 	State->zM2    = 0;
-	State->Tap    = 0;
 	State->Quant  = 0x0800;
 	State->Output = 0;
 	State->MaxOutputLevel = 0;
@@ -28,39 +30,58 @@ uint32_t AD4_EncodeFrame(struct AD4State_t *State, const int16_t *Data) {
 	uint8_t  n;
 	 int32_t zM1    = State->zM1;
 	 int32_t zM2    = State->zM2;
-	 int32_t Tap    = State->Tap;
 	 int32_t Quant  = State->Quant;
 	 int32_t Output = State->Output;
 	uint32_t MaxOutputLevel = State->MaxOutputLevel;
 	uint32_t FrameData = 0;
+#ifdef DEBUG_OUTPUT
+	static FILE *DebugFile = NULL;
+	if(!DebugFile) DebugFile = fopen("Debug.sb", "ab");
+#endif
 	for(n=0;n<8;n++) {
 		//! Get input, compute prediction, and quantize residue
 		//! Note that we minimize error of Output rather than Y, which implies
 		//! applying the post-filter in the analysis equation to get the residue.
 		int32_t X = Data[n];
 		int32_t P = zM1 - zM2;
-		int32_t R = X - (P + (Tap - (Tap >> 3))); {
+		int32_t R = X - (P + Output - (Output >> 3)); {
 #if 0 //! Lower RMSE, but sounds noisier
 			R = (2*R + ((R < 0) ? (-Quant) : (+Quant))) / (2*Quant); //! (R + Sign[R]*(Quant/2)) / Quant
 #else
-			R /= Quant;
+			//! Round positive residues up, and negative residues towards 0.
+			//! I have no idea why, but this fixes limit cycles.
+			if(R > 0) R = (R + (Quant-1)) / Quant;
+			else      R = (R +         0) / Quant;
 #endif
 			if(R < -8) R = -8;
 			if(R > +7) R = +7;
 		}
-		int32_t Y = P + R*Quant;
 
-		//! Calculate output value and update maximum level
+		//! Calculate output value and apply limiting
 		//! Post-filter: Hpost(z) = 1 / Hpre(z) = 1 / (1 - (7/8)z^-1)
-		Output = Y + Output - (Output >> 3);
+		int32_t Y;
+		Output = Output - (Output >> 3);
+		for(;;) {
+			Y = P + R*Quant;
+			if(Output+Y < -32768) {
+				if(R < +7) R++;
+				else break;
+			} else if(Output+Y > +32767) {
+				if(R > -8) R--;
+				else break;
+			} else break;
+		}
+		Output += Y;
+
+		//! Update maximum level after attempted clipping
 		uint32_t Level = (uint32_t)((Output < 0) ? (-Output) : (+Output));
 		if(Level > MaxOutputLevel) MaxOutputLevel = Level;
 
-		//! Do the same, but for the encoding tap. This is needed to
-		//! avoid a limit oscillation on silence from round-off error.
-		//! Technically, it does mean a different output, but it should
-		//! be close enough to what we want that it shouldn't matter.
-		Tap = Y + Tap - ((Tap + 4 - (Tap < 0)) >> 3); //! Y + Round[Tap*7/8]
+		//! Write debug output
+#ifdef DEBUG_OUTPUT
+		int8_t DebugSample = Output >> 8;
+		fwrite(&DebugSample, sizeof(int8_t), 1, DebugFile);
+#endif
 
 		//! Update taps and push residue to frame
 		zM2 = zM1;
@@ -73,7 +94,6 @@ uint32_t AD4_EncodeFrame(struct AD4State_t *State, const int16_t *Data) {
 	}
 	State->zM1    = zM1;
 	State->zM2    = zM2;
-	State->Tap    = Tap;
 	State->Quant  = Quant;
 	State->Output = Output;
 	State->MaxOutputLevel = MaxOutputLevel;

src/platform/gba/packer/tracks/convert.bat

@@ -6,7 +6,7 @@ set DEMO_FOLDER=conv_demo\
 
 for /r %IN_FOLDER% %%i in (*.wav) do (
 	ffmpeg -v panic -y -i %%i -ar 10512 -f s16le -acodec pcm_s16le temp.raw
-	ad4 temp.raw %OUT_FOLDER%%%~ni.ad4 -0.8
+	ad4 temp.raw %OUT_FOLDER%%%~ni.ad4 -0.5
 )
 del /f temp.raw