spu: add asm mixing code for ARMv5 too

[pcsx_rearmed.git] / plugins / dfsound / spu.c

diff --git a/plugins/dfsound/spu.c b/plugins/dfsound/spu.c
index b866cff..d26fa96 100644 (file)
--- a/plugins/dfsound/spu.c
+++ b/plugins/dfsound/spu.c
@@ -26,7 +26,6 @@
 #include "registers.h"
 #include "cfg.h"
 #include "dsoundoss.h"
-#include "regs.h"
 
 #ifdef ENABLE_NLS
 #include <libintl.h>
@@ -38,7 +37,7 @@
 #define N_(x) (x)
 #endif
 
-#ifdef __arm__
+#ifdef __ARM_ARCH_7A__
  #define ssat32_to_16(v) \
   asm("ssat %0,#16,%1" : "=r" (v) : "r" (v))
 #else
@@ -133,6 +132,8 @@ int lastch=-1;             // last channel processed on spu irq in timer mode
 static int lastns=0;       // last ns pos
 static int iSecureStart=0; // secure start counter
 
+#define CDDA_BUFFER_SIZE (16384 * sizeof(uint32_t)) // must be power of 2
+
 ////////////////////////////////////////////////////////////////////////
 // CODE AREA
 ////////////////////////////////////////////////////////////////////////
@@ -221,9 +222,10 @@ INLINE void InterpolateUp(int ch)
    s_chan[ch].SB[32]=0;
 
    s_chan[ch].SB[28]=(s_chan[ch].SB[28]*s_chan[ch].sinc)/0x20000L;
-   if(s_chan[ch].sinc<=0x8000)
-        s_chan[ch].SB[29]=s_chan[ch].SB[30]-(s_chan[ch].SB[28]*((0x10000/s_chan[ch].sinc)-1));
-   else s_chan[ch].SB[29]+=s_chan[ch].SB[28];
+   //if(s_chan[ch].sinc<=0x8000)
+   //     s_chan[ch].SB[29]=s_chan[ch].SB[30]-(s_chan[ch].SB[28]*((0x10000/s_chan[ch].sinc)-1));
+   //else
+   s_chan[ch].SB[29]+=s_chan[ch].SB[28];
   }
  else                                                  // no flags? add bigger val (if possible), calc smaller step, set flag1
   s_chan[ch].SB[29]+=s_chan[ch].SB[28];
@@ -287,19 +289,9 @@ INLINE void StartSound(int ch)
 // ALL KIND OF HELPERS
 ////////////////////////////////////////////////////////////////////////
 
-INLINE void VoiceChangeFrequency(int ch)
-{
- s_chan[ch].iUsedFreq=s_chan[ch].iActFreq;             // -> take it and calc steps
- s_chan[ch].sinc=s_chan[ch].iRawPitch<<4;
- if(!s_chan[ch].sinc) s_chan[ch].sinc=1;
- if(iUseInterpolation==1) s_chan[ch].SB[32]=1;         // -> freq change in simle imterpolation mode: set flag
-}
-
-////////////////////////////////////////////////////////////////////////
-
 INLINE int FModChangeFrequency(int ch,int ns)
 {
- int NP=s_chan[ch].iRawPitch;
+ unsigned int NP=s_chan[ch].iRawPitch;
  int sinc;
 
  NP=((32768L+iFMod[ns])*NP)/32768L;
@@ -307,12 +299,7 @@ INLINE int FModChangeFrequency(int ch,int ns)
  if(NP>0x3fff) NP=0x3fff;
  if(NP<0x1)    NP=0x1;
 
- NP=(44100L*NP)/(4096L);                               // calc frequency
-
- s_chan[ch].iActFreq=NP;
- s_chan[ch].iUsedFreq=NP;
- sinc=(((NP/10)<<16)/4410);
- if(!sinc) sinc=1;
+ sinc=NP<<4;                                           // calc frequency
  if(iUseInterpolation==1)                              // freq change in simple interpolation mode
   s_chan[ch].SB[32]=1;
  iFMod[ns]=0;
@@ -352,7 +339,7 @@ INLINE void StoreInterpolationVal(int ch,int fa)
 
 ////////////////////////////////////////////////////////////////////////
 
-INLINE int iGetInterpolationVal(int ch)
+INLINE int iGetInterpolationVal(int ch, int spos)
 {
  int fa;
 
@@ -364,7 +351,7 @@ INLINE int iGetInterpolationVal(int ch)
    case 3:                                             // cubic interpolation
     {
      long xd;int gpos;
-     xd = ((s_chan[ch].spos) >> 1)+1;
+     xd = (spos >> 1)+1;
      gpos = s_chan[ch].SB[28];
 
      fa  = gval(3) - 3*gval(2) + 3*gval(1) - gval0;
@@ -383,7 +370,7 @@ INLINE int iGetInterpolationVal(int ch)
    case 2:                                             // gauss interpolation
     {
      int vl, vr;int gpos;
-     vl = (s_chan[ch].spos >> 6) & ~3;
+     vl = (spos >> 6) & ~3;
      gpos = s_chan[ch].SB[28];
      vr=(gauss[vl]*gval0)&~2047;
      vr+=(gauss[vl+1]*gval(1))&~2047;
@@ -501,6 +488,7 @@ static int decode_block(int ch)
  }
 
  s_chan[ch].pCurr = start;                 // store values for next cycle
+ s_chan[ch].bJump = flags & 1;
 
  return ret;
 }
@@ -512,9 +500,6 @@ static int skip_block(int ch)
  int flags = start[1];
  int ret = 0;
 
- // Tron Bonne hack, probably wrong (could be wrong memory contents..)
- if(flags & ~7) flags = 0;
-
  if(start == pSpuIrq)
  {
   do_irq();
@@ -522,13 +507,14 @@ static int skip_block(int ch)
  }
 
  if(flags & 4)
-  s_chan[ch].pLoop=start;
+  s_chan[ch].pLoop = start;
 
  s_chan[ch].pCurr += 16;
 
  if(flags & 1)
   s_chan[ch].pCurr = s_chan[ch].pLoop;
 
+ s_chan[ch].bJump = flags & 1;
  return ret;
 }
 
@@ -584,7 +570,7 @@ out:                                         \
 
 make_do_samples(default, fmod_recv_check, ,
   StoreInterpolationVal(ch, fa),
-  ChanBuf[ns] = iGetInterpolationVal(ch), )
+  ChanBuf[ns] = iGetInterpolationVal(ch, spos), )
 make_do_samples(noint, , fa = s_chan[ch].SB[29], , ChanBuf[ns] = fa, s_chan[ch].SB[29] = fa)
 
 #define simple_interp_store \
@@ -638,7 +624,7 @@ static int do_samples_noise(int ch, int ns, int ns_to)
 }
 
 #ifdef __arm__
-// asm code
+// asm code; lv and rv must be 0-3fff
 extern void mix_chan(int start, int count, int lv, int rv);
 extern void mix_chan_rvb(int start, int count, int lv, int rv);
 #else
@@ -750,9 +736,6 @@ static void *MAINThread(void *arg)
        if(dwNewChannel&(1<<ch)) StartSound(ch);        // start new sound
        if(!(dwChannelOn&(1<<ch))) continue;            // channel not playing? next
 
-       if(s_chan[ch].iActFreq!=s_chan[ch].iUsedFreq)   // new psx frequency?
-        VoiceChangeFrequency(ch);
-
        if(s_chan[ch].bNoise)
         d=do_samples_noise(ch, ns_from, ns_to);
        else if(s_chan[ch].bFMod==2 || (s_chan[ch].bFMod==0 && iUseInterpolation==0))
@@ -791,20 +774,18 @@ static void *MAINThread(void *arg)
        if(s_chan[ch].pCurr > pSpuIrq && s_chan[ch].pLoop > pSpuIrq)
         continue;
 
-       if(s_chan[ch].iActFreq!=s_chan[ch].iUsedFreq)   // new psx frequency?
-         VoiceChangeFrequency(ch);
-
        s_chan[ch].spos += s_chan[ch].sinc * NSSIZE;
        while(s_chan[ch].spos >= 28 * 0x10000)
         {
-         unsigned char *start=s_chan[ch].pCurr;
+         unsigned char *start = s_chan[ch].pCurr;
 
          // no need for bIRQReturn since the channel is silent
          iSpuAsyncWait |= skip_block(ch);
          if(start == s_chan[ch].pCurr)
           {
            // looping on self
-           dwChannelDead|=1<<ch;
+           dwChannelDead |= 1<<ch;
+           s_chan[ch].spos = 0;
            break;
           }
 
@@ -923,12 +904,14 @@ static void *MAINThread(void *arg)
 // SPU ASYNC... even newer epsxe func
 //  1 time every 'cycle' cycles... harhar
 
+// rearmed: called every 2ms now
+
 void CALLBACK SPUasync(unsigned long cycle)
 {
  if(iSpuAsyncWait)
   {
    iSpuAsyncWait++;
-   if(iSpuAsyncWait<=16/FRAG_MSECS) return;
+   if(iSpuAsyncWait<=16/2) return;
    iSpuAsyncWait=0;
   }
 
@@ -970,12 +953,12 @@ void CALLBACK SPUplayADPCMchannel(xa_decode_t *xap)
 }
 
 // CDDA AUDIO
-void CALLBACK SPUplayCDDAchannel(short *pcm, int nbytes)
+int CALLBACK SPUplayCDDAchannel(short *pcm, int nbytes)
 {
- if (!pcm)      return;
- if (nbytes<=0) return;
+ if (!pcm)      return -1;
+ if (nbytes<=0) return -1;
 
- FeedCDDA((unsigned char *)pcm, nbytes);
+ return FeedCDDA((unsigned char *)pcm, nbytes);
 }
 
 // SETUPTIMER: init of certain buffers and threads/timers
@@ -1033,7 +1016,7 @@ void SetupStreams(void)
  XAFeed  = XAStart;
 
  CDDAStart =                                           // alloc cdda buffer
-  (uint32_t *)malloc(16384 * sizeof(uint32_t));
+  (uint32_t *)malloc(CDDA_BUFFER_SIZE);
  CDDAEnd   = CDDAStart + 16384;
  CDDAPlay  = CDDAStart;
  CDDAFeed  = CDDAStart;
@@ -1045,7 +1028,6 @@ void SetupStreams(void)
 //   s_chan[i].hMutex=CreateMutex(NULL,FALSE,NULL);
    s_chan[i].ADSRX.SustainLevel = 0xf;                 // -> init sustain
    s_chan[i].pLoop=spuMemC;
-   s_chan[i].pStart=spuMemC;
    s_chan[i].pCurr=spuMemC;
   }