core, implement GG stereo

pico/pico.h

@@ -267,7 +267,7 @@ void Pico32xSetClocks(int msh2_hz, int ssh2_hz);
 #define PICO_SSH2_HZ ((int)(7670442.0 * 2.4))
 
 // sound.c
-extern void (*PsndMix_32_to_16l)(s16 *dest, s32 *src, int count);
+extern void (*PsndMix_32_to_16)(s16 *dest, s32 *src, int count);
 void PsndRerate(int preserve_state);
 
 // media.c

pico/sms.c

@@ -241,6 +241,8 @@ static void z80_sms_out(unsigned short a, unsigned char d)
       case 0x00:
         if ((PicoIn.AHW & PAHW_GG) && a < 0x8)   // GG I/O area
           Pico.ms.io_gg[a] = d;
+        if ((PicoIn.AHW & PAHW_GG) && a == 0x6)
+          SN76496Config(d);
         break;
       case 0x01:
         if ((PicoIn.AHW & PAHW_GG) && a < 0x8) { // GG I/O area

pico/sound/mix.c

@@ -18,7 +18,7 @@
 	val -= val >> 3; /* reduce level to avoid clipping */	\
 	if ((s16)val != val) val = (val < 0 ? MINOUT : MAXOUT)
 
-int mix_32_to_16l_level;
+int mix_32_to_16_level;
 
 static struct iir {
 	int	alpha;		// alpha for EMA low pass
@@ -63,33 +63,34 @@ static inline int filter_null(struct iir *fi2, int x)
 
 #define filter	filter_band
 
-#define mix_32_to_16l_stereo_core(dest, src, count, lv, fl) {	\
+#define mix_32_to_16_stereo_core(dest, src, count, lv, fl) {	\
 	int l, r;						\
 	struct iir lf = lfi2, rf = rfi2;			\
 								\
 	for (; count > 0; count--)				\
 	{							\
-		l = r = *dest;					\
+		l = *dest;					\
 		l += *src++ >> lv;				\
-		r += *src++ >> lv;				\
 		l = fl(&lf, l);					\
-		r = fl(&rf, r);					\
 		Limit16(l);					\
-		Limit16(r);					\
 		*dest++ = l;					\
+		r = *dest;					\
+		r += *src++ >> lv;				\
+		r = fl(&rf, r);					\
+		Limit16(r);					\
 		*dest++ = r;					\
 	}							\
 	lfi2 = lf, rfi2 = rf;					\
 }
 
-void mix_32_to_16l_stereo_lvl(s16 *dest, s32 *src, int count)
+void mix_32_to_16_stereo_lvl(s16 *dest, s32 *src, int count)
 {
-	mix_32_to_16l_stereo_core(dest, src, count, mix_32_to_16l_level, filter);
+	mix_32_to_16_stereo_core(dest, src, count, mix_32_to_16_level, filter);
 }
 
-void mix_32_to_16l_stereo(s16 *dest, s32 *src, int count)
+void mix_32_to_16_stereo(s16 *dest, s32 *src, int count)
 {
-	mix_32_to_16l_stereo_core(dest, src, count, 0, filter);
+	mix_32_to_16_stereo_core(dest, src, count, 0, filter);
 }
 
 void mix_32_to_16_mono(s16 *dest, s32 *src, int count)

pico/sound/mix.h

@@ -6,9 +6,9 @@ void mix_16h_to_32_s2(s32 *dest, s16 *src, int count);
 
 void mix_16h_to_32_resample_stereo(s32 *dest, s16 *src, int count, int fac16);
 void mix_16h_to_32_resample_mono(s32 *dest, s16 *src, int count, int fac16);
-void mix_32_to_16l_stereo(s16 *dest, s32 *src, int count);
+void mix_32_to_16_stereo(s16 *dest, s32 *src, int count);
 void mix_32_to_16_mono(s16 *dest, s32 *src, int count);
 
-extern int mix_32_to_16l_level;
+extern int mix_32_to_16_level;
-void mix_32_to_16l_stereo_lvl(s16 *dest, s32 *src, int count);
+void mix_32_to_16_stereo_lvl(s16 *dest, s32 *src, int count);
 void mix_reset(int alpha_q16);

pico/sound/mix_arm.S

@@ -282,29 +282,29 @@ m16_32_rsm_end:
 
 @ mix 32bit audio (with 16bits really used, upper bits indicate overflow) with normal 16 bit audio with left channel only
 @ warning: this function assumes dest is word aligned
-.global mix_32_to_16l_stereo @ short *dest, int *src, int count
+.global mix_32_to_16_stereo @ short *dest, int *src, int count
 
-mix_32_to_16l_stereo:
+mix_32_to_16_stereo:
     stmfd   sp!, {r4-r8,r10-r11,lr}
 
     mov     r2, r2, lsl #1
     subs    r2, r2, #4
-    bmi     m32_16l_st_end
+    bmi     m32_16_st_end
 
     ldr     r12, =filter
     ldr     r8, [r12], #4
     ldmia   r12, {r3,r10-r11,lr}
     str     r8, [sp, #-4]!
 
-m32_16l_st_loop:
+m32_16_st_loop:
     ldmia   r0,  {r8,r12}
     ldmia   r1!, {r4-r7}
+    add     r5, r5, r8, asr #16
+    add     r7, r7, r12,asr #16
     mov     r8, r8, lsl #16
     mov     r12,r12,lsl #16
     add     r4, r4, r8, asr #16
-    add     r5, r5, r8, asr #16
     add     r6, r6, r12,asr #16
-    add     r7, r7, r12,asr #16
     ldr     r12,[sp]
     LPfilt  r4, r3
     LPfilt  r5, lr
@@ -323,16 +323,17 @@ m32_16l_st_loop:
     orr     r4, r5, r4, lsr #16
     orr     r5, r7, r6, lsr #16
     stmia   r0!, {r4,r5}
-    bpl     m32_16l_st_loop
+    bpl     m32_16_st_loop
 
-m32_16l_st_end:
+m32_16_st_end:
     @ check for remaining bytes to convert
     tst     r2, #2
-    beq     m32_16l_st_no_unal2
+    beq     m32_16_st_no_unal2
-    ldrsh   r6, [r0]
+    ldr     r6, [r0]
     ldmia   r1!,{r4,r5}
-    add     r4, r4, r6
+    add     r5, r5, r6, asr #16
-    add     r5, r5, r6
+    mov     r6, r6, lsl #16
+    add     r4, r4, r6, asr #16
     ldr     r12,[sp]
     LPfilt  r4, r3
     LPfilt  r5, lr
@@ -344,7 +345,7 @@ m32_16l_st_end:
     orr     r4, r5, r4, lsr #16
     str     r4, [r0], #4
 
-m32_16l_st_no_unal2:
+m32_16_st_no_unal2:
     ldr     r12, =filter
     add     r12,r12, #4
     stmia   r12, {r3,r10-r11,lr}
@@ -386,10 +387,10 @@ m32_16_mo_loop:
     ldmia   r0,  {r8,r12}
     ldmia   r1!, {r4-r7}
     add     r5, r5, r8, asr #16
-    mov     r8, r8, lsl #16
-    add     r4, r4, r8, asr #16
     add     r7, r7, r12,asr #16
+    mov     r8, r8, lsl #16
     mov     r12,r12,lsl #16
+    add     r4, r4, r8, asr #16
     add     r6, r6, r12,asr #16
     ldr     r12,[sp]
     LPfilt  r4, r11
@@ -458,20 +459,20 @@ m32_16_mo_no_unal:
 .data
 .align 4
 
-.global mix_32_to_16l_level
+.global mix_32_to_16_level
-mix_32_to_16l_level:
+mix_32_to_16_level:
     .word   0
 
 .text
 .align 4
 
-@ same as mix_32_to_16l_stereo, but with additional shift
+@ same as mix_32_to_16_stereo, but with additional shift
-.global mix_32_to_16l_stereo_lvl @ short *dest, int *src, int count
+.global mix_32_to_16_stereo_lvl @ short *dest, int *src, int count
 
-mix_32_to_16l_stereo_lvl:
+mix_32_to_16_stereo_lvl:
     stmfd   sp!, {r4-r11,lr}
 
-    ldr     r9, =mix_32_to_16l_level
+    ldr     r9, =mix_32_to_16_level
     mov     lr, #1
     ldr     r9, [r9]
     ldr     r12, =filter
@@ -481,17 +482,17 @@ mix_32_to_16l_stereo_lvl:
 
     mov     r2, r2, lsl #1
     subs    r2, r2, #4
-    bmi     m32_16l_st_l_end
+    bmi     m32_16_st_l_end
 
-m32_16l_st_l_loop:
+m32_16_st_l_loop:
     ldmia   r0,  {r8,r12}
     ldmia   r1!, {r4-r7}
+    add     r5, r5, r8, asr #16
+    add     r7, r7, r12,asr #16
     mov     r8, r8, lsl #16
     mov     r12,r12,lsl #16
     add     r4, r4, r8, asr #16
-    add     r5, r5, r8, asr #16
     add     r6, r6, r12,asr #16
-    add     r7, r7, r12,asr #16
     mov     r4, r4, asr r9
     mov     r5, r5, asr r9
     mov     r6, r6, asr r9
@@ -514,16 +515,17 @@ m32_16l_st_l_loop:
     orr     r4, r5, r4, lsr #16
     orr     r5, r7, r6, lsr #16
     stmia   r0!, {r4,r5}
-    bpl     m32_16l_st_l_loop
+    bpl     m32_16_st_l_loop
 
-m32_16l_st_l_end:
+m32_16_st_l_end:
     @ check for remaining bytes to convert
     tst     r2, #2
-    beq     m32_16l_st_l_no_unal2
+    beq     m32_16_st_l_no_unal2
-    ldrsh   r6, [r0]
+    ldr     r6, [r0]
     ldmia   r1!,{r4,r5}
-    add     r4, r4, r6
+    add     r5, r5, r6, asr #16
-    add     r5, r5, r6
+    mov     r6, r6, lsl #16
+    add     r4, r4, r6, asr #16
     mov     r4, r4, asr r9
     mov     r5, r5, asr r9
     ldr     r12,[sp]
@@ -537,7 +539,7 @@ m32_16l_st_l_end:
     orr     r4, r5, r4, lsr #16
     str     r4, [r0], #4
 
-m32_16l_st_l_no_unal2:
+m32_16_st_l_no_unal2:
     ldr     r12, =filter
     add     r12,r12, #4
     stmia   r12, {r3,r10-r11,lr}

pico/sound/sn76496.c

@@ -56,7 +56,7 @@ struct SN76496
 	int Period[4];
 	int Count[4];
 	int Output[4];
-	int pad[1];
+	int Panning;
 };
 
 static struct SN76496 ono_sn; // one and only SN76496
@@ -135,7 +135,6 @@ void SN76496Update(short *buffer, int length, int stereo)
 	while (length > 0)
 	{
 		int vol[4];
-		unsigned int out;
 		int left;
 
 
@@ -203,19 +202,43 @@ void SN76496Update(short *buffer, int length, int stereo)
 		} while (left > 0 && R->Volume[3]);
 		if (R->Output[3]) vol[3] -= R->Count[3];
 
-		out = vol[0] * R->Volume[0] + vol[1] * R->Volume[1] +
+		length--;
+		if (R->Panning == 0xff || !stereo) {
+			unsigned int out =
+				vol[0] * R->Volume[0] + vol[1] * R->Volume[1] +
 				vol[2] * R->Volume[2] + vol[3] * R->Volume[3];
 
-		if (out > MAX_OUTPUT * STEP) out = MAX_OUTPUT * STEP;
+			if (out > MAX_OUTPUT * STEP) out = MAX_OUTPUT * STEP;
 
-		out /= STEP; // will be optimized to shift; max 0x4800 = 18432
+			out /= STEP; // will be optimized to shift; max 0x4800 = 18432
-		*buffer++ += out;
+			*buffer++ += out;
-		if (stereo) *buffer++ += out;
+			if (stereo) *buffer++ += out;
+		} else {
+#define P(n) !!(R->Panning & (1<<(n)))
+			unsigned int outl =
+				vol[0] * R->Volume[0] * P(4) + vol[1] * R->Volume[1] * P(5) +
+				vol[2] * R->Volume[2] * P(6) + vol[3] * R->Volume[3] * P(7);
+			unsigned int outr =
+				vol[0] * R->Volume[0] * P(0) + vol[1] * R->Volume[1] * P(1) +
+				vol[2] * R->Volume[2] * P(2) + vol[3] * R->Volume[3] * P(3);
+#undef P
+			if (outl > MAX_OUTPUT * STEP) outl = MAX_OUTPUT * STEP;
+			if (outr > MAX_OUTPUT * STEP) outr = MAX_OUTPUT * STEP;
 
-		length--;
+			outl /= STEP; // will be optimized to shift; max 0x4800 = 18432
+			outr /= STEP; // will be optimized to shift; max 0x4800 = 18432
+			*buffer++ += outl;
+			*buffer++ += outr;
+		}
 	}
 }
 
+void SN76496Config(int panning)
+{
+	struct SN76496 *R = &ono_sn;
+	R->Panning = panning & 0xff;
+}
+
 
 static void SN76496_set_clock(struct SN76496 *R,int clock)
 {
@@ -286,6 +309,7 @@ int SN76496_init(int clock,int sample_rate)
 
 	// added
 	SN76496_set_gain(R, 0);
+	R->Panning = 0xff;
 
 	return 0;
 }

pico/sound/sn76496.h

@@ -3,6 +3,7 @@
 
 void SN76496Write(int data);
 void SN76496Update(short *buffer,int length,int stereo);
+void SN76496Config(int panning);
 int  SN76496_init(int clock,int sample_rate);
 
 #endif

pico/sound/sound.c

@@ -15,7 +15,7 @@
 #include "resampler.h"
 #include "mix.h"
 
-void (*PsndMix_32_to_16l)(s16 *dest, s32 *src, int count) = mix_32_to_16l_stereo;
+void (*PsndMix_32_to_16)(s16 *dest, s32 *src, int count) = mix_32_to_16_stereo;
 
 // master int buffer to mix to
 // +1 for a fill triggered by an instruction overhanging into the next scanline
@@ -228,7 +228,7 @@ void PsndRerate(int preserve_state)
     PsndClear();
 
   // set mixer
-  PsndMix_32_to_16l = (PicoIn.opt & POPT_EN_STEREO) ? mix_32_to_16l_stereo : mix_32_to_16_mono;
+  PsndMix_32_to_16 = (PicoIn.opt & POPT_EN_STEREO) ? mix_32_to_16_stereo : mix_32_to_16_mono;
   mix_reset(PicoIn.opt & POPT_EN_SNDFILTER ? PicoIn.sndFilterAlpha : 0);
 
   if (PicoIn.AHW & PAHW_PICO)
@@ -275,8 +275,8 @@ PICO_INTERNAL void PsndDoDAC(int cyc_to)
   if (PicoIn.opt & POPT_EN_STEREO) {
     s16 *d = PicoIn.sndOut + pos*2;
     // left channel only, mixed ro right channel in mixing phase
-    *d++ += Pico.snd.dac_val2; d++;
+    *d++ += Pico.snd.dac_val2, *d++ += Pico.snd.dac_val2;
-    while (--len) *d++ += Pico.snd.dac_val, d++;
+    while (--len) *d++ += Pico.snd.dac_val, *d++ += Pico.snd.dac_val;
   } else {
     s16 *d = PicoIn.sndOut + pos;
     *d++ += Pico.snd.dac_val2;
@@ -345,10 +345,15 @@ PICO_INTERNAL void PsndDoSMSFM(int cyc_to)
   if (Pico.m.hardware & PMS_HW_FMUSED) {
     buf += pos;
     PsndFMUpdate(buf32, len, 0, 0);
-    while (len--) {
+    if (stereo) 
-      *buf++ += *buf32++;
+      while (len--) {
-      buf += stereo;
+        *buf++ += *buf32;
-    }
+        *buf++ += *buf32++;
+      }
+    else
+      while (len--) {
+        *buf++ += *buf32++;
+      }
   }
 }
 
@@ -506,10 +511,10 @@ static int PsndRender(int offset, int length)
     s16 *dacbuf = PicoIn.sndOut + (daclen << stereo);
     Pico.snd.dac_pos += (length-daclen) << 20;
     *dacbuf++ += Pico.snd.dac_val2;
-    if (stereo) dacbuf++;
+    if (stereo) *dacbuf++ += Pico.snd.dac_val2;
     for (daclen++; length-daclen > 0; daclen++) {
       *dacbuf++ += Pico.snd.dac_val;
-      if (stereo) dacbuf++;
+      if (stereo) *dacbuf++ += Pico.snd.dac_val;
     }
     Pico.snd.dac_val2 = Pico.snd.dac_val;
   }
@@ -544,7 +549,7 @@ static int PsndRender(int offset, int length)
 
   // convert + limit to normal 16bit output
   if (PicoIn.sndOut)
-    PsndMix_32_to_16l(PicoIn.sndOut+(offset<<stereo), buf32, length-offset);
+    PsndMix_32_to_16(PicoIn.sndOut+(offset<<stereo), buf32, length-offset);
 
   pprof_end(sound);
 
@@ -589,21 +594,18 @@ static int PsndRenderMS(int offset, int length)
     int len = (length-ym2413len);
     if (Pico.m.hardware & PMS_HW_FMUSED) {
       PsndFMUpdate(buf32, len, 0, 0);
-      while (len--) {
+      if (stereo)
-        *ym2413buf++ += *buf32++;
+        while (len--) {
-        ym2413buf += stereo;
+          *ym2413buf++ += *buf32;
-      }
+          *ym2413buf++ += *buf32++;
+        }
+      else
+        while (len--) {
+          *ym2413buf++ += *buf32++;
+        }
     }
   }
 
-  // upmix to "stereo" if needed
-  if (PicoIn.opt & POPT_EN_STEREO) {
-    int i;
-    s16 *p;
-    for (i = length, p = (s16 *)PicoIn.sndOut; i > 0; i--, p+=2)
-      *(p + 1) = *p;
-  }
-
   pprof_end(sound);
 
   return length;

platform/gp2x/emu.c

@@ -753,10 +753,10 @@ void plat_update_volume(int has_changed, int is_up)
 
 	/* set the right mixer func */
 	if (vol >= 5)
-		PsndMix_32_to_16l = mix_32_to_16l_stereo;
+		PsndMix_32_to_16 = mix_32_to_16_stereo;
 	else {
-		mix_32_to_16l_level = 5 - vol;
+		mix_32_to_16_level = 5 - vol;
-		PsndMix_32_to_16l = mix_32_to_16l_stereo_lvl;
+		PsndMix_32_to_16 = mix_32_to_16_stereo_lvl;
 	}
 }