spu: support multiple pending buffers

for threaded implementation

10 files changed, 561 insertions, 323 deletions

diff --git a/plugins/dfsound/adsr.c b/plugins/dfsound/adsr.c
index 836fdbb..69e55bd 100644
--- a/plugins/dfsound/adsr.c
+++ b/plugins/dfsound/adsr.c
@@ -217,6 +217,131 @@ done:
  return ns;

 }

 

+static int SkipADSR(ADSRInfoEx *adsr, int ns_to)

+{

+ int EnvelopeVol = adsr->EnvelopeVol;

+ int ns = 0, val, rto, level;

+

+ if (adsr->State == ADSR_RELEASE)

+ {

+   val = RateTableSub[adsr->ReleaseRate * 4];

+   if (adsr->ReleaseModeExp)

+   {

+     for (; ns < ns_to; ns++)

+     {

+       EnvelopeVol += ((long long)val * EnvelopeVol) >> (15+16);

+       if (EnvelopeVol <= 0)

+         break;

+     }

+   }

+   else

+   {

+     EnvelopeVol += val * ns_to;

+     if (EnvelopeVol > 0)

+       ns = ns_to;

+   }

+   goto done;

+ }

+

+ switch (adsr->State)

+ {

+   case ADSR_ATTACK:                                   // -> attack

+     rto = 0;

+     if (adsr->AttackModeExp && EnvelopeVol >= 0x60000000)

+       rto = 8;

+     val = RateTableAdd[adsr->AttackRate + rto];

+

+     for (; ns < ns_to; ns++)

+     {

+       EnvelopeVol += val;

+       if (EnvelopeVol < 0)

+        break;

+     }

+     if (EnvelopeVol < 0) // overflow

+     {

+       EnvelopeVol = 0x7fffffff;

+       adsr->State = ADSR_DECAY;

+       ns++;

+       goto decay;

+     }

+     break;

+

+   //--------------------------------------------------//

+   decay:

+   case ADSR_DECAY:                                    // -> decay

+     val = RateTableSub[adsr->DecayRate * 4];

+     level = adsr->SustainLevel;

+

+     for (; ns < ns_to; )

+     {

+       EnvelopeVol += ((long long)val * EnvelopeVol) >> (15+16);

+       if (EnvelopeVol < 0)

+         EnvelopeVol = 0;

+

+       ns++;

+

+       if (((EnvelopeVol >> 27) & 0xf) <= level)

+       {

+         adsr->State = ADSR_SUSTAIN;

+         goto sustain;

+       }

+     }

+     break;

+

+   //--------------------------------------------------//

+   sustain:

+   case ADSR_SUSTAIN:                                  // -> sustain

+     if (adsr->SustainIncrease)

+     {

+       if (EnvelopeVol >= 0x7fff0000)

+       {

+         ns = ns_to;

+         break;

+       }

+

+       rto = 0;

+       if (adsr->SustainModeExp && EnvelopeVol >= 0x60000000)

+         rto = 8;

+       val = RateTableAdd[adsr->SustainRate + rto];

+

+       EnvelopeVol += val * (ns_to - ns);

+       if ((unsigned int)EnvelopeVol >= 0x7fe00000)

+       {

+         EnvelopeVol = 0x7fffffff;

+         ns = ns_to;

+         break;

+       }

+     }

+     else

+     {

+       val = RateTableSub[adsr->SustainRate];

+       if (adsr->SustainModeExp)

+       {

+         for (; ns < ns_to; ns++)

+         {

+           EnvelopeVol += ((long long)val * EnvelopeVol) >> (15+16);

+           if (EnvelopeVol < 0)

+             break;

+         }

+       }

+       else

+       {

+         EnvelopeVol += val * (ns_to - ns);

+         if (EnvelopeVol > 0)

+         {

+           ns = ns_to;

+           break;

+         }

+       }

+     }

+     break;

+ }

+

+done:

+ adsr->EnvelopeVol = EnvelopeVol;

+ return ns;

+}

+

 #endif

 

 /*

diff --git a/plugins/dfsound/arm_utils.S b/plugins/dfsound/arm_utils.S
index 2511bb3..9652313 100644
--- a/plugins/dfsound/arm_utils.S
+++ b/plugins/dfsound/arm_utils.S
@@ -14,7 +14,6 @@
 .data
 .align 2
 ptr_ChanBuf:   .word ESYM(ChanBuf)
-ptr_SSumLR:    .word ESYM(SSumLR)
 #endif
 
 .text
@@ -36,15 +35,11 @@ ptr_SSumLR:    .word ESYM(SSumLR)
 
 #ifdef __ARM_NEON__
 
-FUNCTION(mix_chan): @ (int start, int count, int lv, int rv)
+FUNCTION(mix_chan): @ (int *SSumLR, int count, int lv, int rv)
     vmov.32     d14[0], r2
     vmov.32     d14[1], r3             @ multipliers
-    load_varadr r2, SSumLR
-    mov         r12, r0
+    mov         r2, r0
     load_varadr r0, ChanBuf
-    ldr         r2, [r2]
-    add         r0, r12, lsl #2
-    add         r2, r12, lsl #3
 0:
     vldmia      r0!, {d0-d1}
     vldmia      r2, {d2-d5}
@@ -70,17 +65,12 @@ mc_finish:
     bx          lr
 
 
-FUNCTION(mix_chan_rvb): @ (int start, int count, int lv, int rv, int *rvb)
+FUNCTION(mix_chan_rvb): @ (int *SSumLR, int count, int lv, int rv, int *rvb)
     vmov.32     d14[0], r2
     vmov.32     d14[1], r3             @ multipliers
-    load_varadr r2, SSumLR
-    mov         r12, r0
+    mov         r2, r0
     load_varadr r0, ChanBuf
     ldr         r3, [sp]               @ rvb
-    ldr         r2, [r2]
-    add         r0, r12, lsl #2
-    add         r2, r12, lsl #3
-    add         r3, r12, lsl #3
 0:
     vldmia      r0!, {d0-d1}
     vldmia      r2, {d2-d5}
@@ -114,16 +104,12 @@ mcr_finish:
 
 #elif defined(HAVE_ARMV5)
 
-FUNCTION(mix_chan): @ (int start, int count, int lv, int rv)
+FUNCTION(mix_chan): @ (int *SSumLR, int count, int lv, int rv)
     stmfd       sp!, {r4-r8,lr}
     orr         r3, r2, r3, lsl #16
     lsl         r3, #1                 @ packed multipliers << 1
-    load_varadr r2, SSumLR
-    mov         r12, r0
+    mov         r2, r0
     load_varadr r0, ChanBuf
-    ldr         r2, [r2]
-    add         r0, r12, lsl #2
-    add         r2, r12, lsl #3
 0:
     ldmia       r0!, {r4,r5}
     ldmia       r2, {r6-r8,lr}
@@ -144,17 +130,13 @@ mc_finish:
     ldmfd       sp!, {r4-r8,pc}
 
 
-FUNCTION(mix_chan_rvb): @ (int start, int count, int lv, int rv, int *rvb)
+FUNCTION(mix_chan_rvb): @ (int *SSumLR, int count, int lv, int rv, int *rvb)
     stmfd       sp!, {r4-r8,lr}
     orr         lr, r2, r3, lsl #16
     lsl         lr, #1
-    load_varadr r2, SSumLR
-    ldr         r3, [sp]               @ rvb
-    ldr         r2, [r2]
-    load_varadr r4, ChanBuf
-    add         r2, r2, r0, lsl #3
-    add         r3, r3, r0, lsl #3
-    add         r0, r4, r0, lsl #2
+    mov         r2, r0
+    load_varadr r0, ChanBuf
+    ldr         r3, [sp, #6*4]         @ rvb
 0:
     ldr         r4, [r0], #4
     ldmia       r2, {r6,r7}
diff --git a/plugins/dfsound/externals.h b/plugins/dfsound/externals.h
index d3bcbc6..4832fac 100644
--- a/plugins/dfsound/externals.h
+++ b/plugins/dfsound/externals.h
@@ -206,8 +206,6 @@ typedef struct
  void (CALLBACK *cddavCallback)(unsigned short,unsigned short);

  void (CALLBACK *scheduleCallback)(unsigned int);

 

- int           * sRVBStart;

-

  xa_decode_t   * xapGlobal;

  unsigned int  * XAFeed;

  unsigned int  * XAPlay;

@@ -228,7 +226,10 @@ typedef struct
  SPUCHAN       * s_chan;

  int           * SB;

 

- int             pad[30];

+ int           * RVB;

+ int           * SSumLR;

+

+ int             pad[29];

  unsigned short  regArea[0x400];

 } SPUInfo;

 

diff --git a/plugins/dfsound/reverb.c b/plugins/dfsound/reverb.c
index b7bcf12..7e32b8e 100644
--- a/plugins/dfsound/reverb.c
+++ b/plugins/dfsound/reverb.c
@@ -40,15 +40,6 @@ INLINE void StartREVERB(int ch)
 }

 

 ////////////////////////////////////////////////////////////////////////

-// HELPER FOR NEILL'S REVERB: re-inits our reverb mixing buf

-////////////////////////////////////////////////////////////////////////

-

-INLINE void InitREVERB(int ns_to)

-{

- memset(spu.sRVBStart,0,ns_to*sizeof(spu.sRVBStart[0])*2);

-}

-

-////////////////////////////////////////////////////////////////////////

 

 INLINE int rvb2ram_offs(int curr, int space, int iOff)

 {

@@ -73,7 +64,7 @@ INLINE int rvb2ram_offs(int curr, int space, int iOff)
 ////////////////////////////////////////////////////////////////////////

 

 // portions based on spu2-x from PCSX2

-static void MixREVERB(int ns_to)

+static void MixREVERB(int *SSumLR, int *RVB, int ns_to)

 {

  int l_old = rvb.iRVBLeft;

  int r_old = rvb.iRVBRight;

@@ -87,8 +78,8 @@ static void MixREVERB(int ns_to)
    int ACC0, ACC1, FB_A0, FB_A1, FB_B0, FB_B1;

    int mix_dest_a0, mix_dest_a1, mix_dest_b0, mix_dest_b1;

 

-   int input_L = spu.sRVBStart[ns]   * rvb.IN_COEF_L;

-   int input_R = spu.sRVBStart[ns+1] * rvb.IN_COEF_R;

+   int input_L = RVB[ns]   * rvb.IN_COEF_L;

+   int input_R = RVB[ns+1] * rvb.IN_COEF_R;

 

    int IIR_INPUT_A0 = ((g_buffer(IIR_SRC_A0) * rvb.IIR_COEF) + input_L) >> 15;

    int IIR_INPUT_A1 = ((g_buffer(IIR_SRC_A1) * rvb.IIR_COEF) + input_R) >> 15;

@@ -158,7 +149,7 @@ static void MixREVERB(int ns_to)
  rvb.CurrAddr = curr_addr;

 }

 

-static void MixREVERB_off(int ns_to)

+static void MixREVERB_off(int *SSumLR, int ns_to)

 {

  int l_old = rvb.iRVBLeft;

  int r_old = rvb.iRVBRight;

@@ -236,7 +227,7 @@ static void prepare_offsets(void)
  rvb.dirty = 0;

 }

 

-INLINE void REVERBDo(int ns_to)

+INLINE void REVERBDo(int *SSumLR, int *RVB, int ns_to)

 {

  if (!rvb.StartAddr)                                   // reverb is off

  {

@@ -249,14 +240,14 @@ INLINE void REVERBDo(int ns_to)
   if (unlikely(rvb.dirty))

    prepare_offsets();

 

-  MixREVERB(ns_to);

+  MixREVERB(SSumLR, RVB, ns_to);

  }

  else if (rvb.VolLeft || rvb.VolRight)

  {

   if (unlikely(rvb.dirty))

    prepare_offsets();

 

-  MixREVERB_off(ns_to);

+  MixREVERB_off(SSumLR, ns_to);

  }

  else                                                  // -> reverb off

  {

diff --git a/plugins/dfsound/spu.c b/plugins/dfsound/spu.c
index c6a06fd..f5edd3a 100644
--- a/plugins/dfsound/spu.c
+++ b/plugins/dfsound/spu.c
@@ -77,47 +77,8 @@ SPUConfig       spu_config;
 
 REVERBInfo      rvb;
 
-#if defined(THREAD_ENABLED) || defined(WANT_THREAD_CODE)
-
-// worker thread state
-static struct spu_worker {
- unsigned int pending:1;
- unsigned int exit_thread:1;
- unsigned int stale_cache:1;
- int ns_to;
- int ctrl;
- int decode_pos;
- int silentch;
- unsigned int chmask;
- struct {
-  int spos;
-  int sbpos;
-  int sinc;
-  int start;
-  int loop;
-  int ns_to;
-  ADSRInfoEx adsr;
-  // might want to add vol and fmod flags..
- } ch[24];
- struct {
-  struct {
-   int adsrState;
-   int adsrEnvelopeVol;
-  } ch[24];
-  unsigned int chan_end;
-  unsigned int decode_dirty;
- } r;
-} *worker;
-
-#else
-static const void * const worker = NULL;
-#endif
-
-// certain globals (were local before, but with the new timeproc I need em global)
-
 static int iFMod[NSSIZE];
 int ChanBuf[NSSIZE];
-int *SSumLR;
 
 #define CDDA_BUFFER_SIZE (16384 * sizeof(uint32_t)) // must be power of 2
 
@@ -267,17 +228,8 @@ static int check_irq(int ch, unsigned char *pos)
 // START SOUND... called by main thread to setup a new sound on a channel
 ////////////////////////////////////////////////////////////////////////
 
-INLINE void StartSound(int ch)
+static void StartSoundSB(int *SB)
 {
- SPUCHAN *s_chan = &spu.s_chan[ch];
- int *SB = spu.SB + ch * SB_SIZE;
-
- StartADSR(ch);
- StartREVERB(ch);
-
- s_chan->prevflags=2;
-
- s_chan->iSBPos=27;
  SB[26]=0;                                             // init mixing vars
  SB[27]=0;
 
@@ -285,6 +237,17 @@ INLINE void StartSound(int ch)
  SB[29]=0;                                             // init our interpolation helpers
  SB[30]=0;
  SB[31]=0;
+}
+
+static void StartSoundMain(int ch)
+{
+ SPUCHAN *s_chan = &spu.s_chan[ch];
+
+ StartADSR(ch);
+ StartREVERB(ch);
+
+ s_chan->prevflags=2;
+ s_chan->iSBPos=27;
  s_chan->spos=0;
 
  spu.dwNewChannel&=~(1<<ch);                           // clear new channel bit
@@ -292,6 +255,12 @@ INLINE void StartSound(int ch)
  spu.dwChannelDead&=~(1<<ch);
 }
 
+static void StartSound(int ch)
+{
+ StartSoundMain(ch);
+ StartSoundSB(spu.SB + ch * SB_SIZE);
+}
+
 ////////////////////////////////////////////////////////////////////////
 // ALL KIND OF HELPERS
 ////////////////////////////////////////////////////////////////////////
@@ -439,7 +408,7 @@ static void decode_block_data(int *dest, const unsigned char *src, int predict_n
  }
 }
 
-static int decode_block(int ch, int *SB)
+static int decode_block(void *unused, int ch, int *SB)
 {
  SPUCHAN *s_chan = &spu.s_chan[ch];
  unsigned char *start;
@@ -520,38 +489,6 @@ static int skip_block(int ch)
  return ret;
 }
 
-#if defined(THREAD_ENABLED) || defined(WANT_THREAD_CODE)
-
-static int decode_block_work(int ch, int *SB)
-{
- const unsigned char *ram = spu.spuMemC;
- int predict_nr, shift_factor, flags;
- int start = worker->ch[ch].start;
- int loop = worker->ch[ch].loop;
-
- predict_nr = ram[start];
- shift_factor = predict_nr & 0xf;
- predict_nr >>= 4;
-
- decode_block_data(SB, ram + start + 2, predict_nr, shift_factor);
-
- flags = ram[start + 1];
- if (flags & 4)
-  loop = start;                            // loop adress
-
- start += 16;
-
- if (flags & 1)                            // 1: stop/loop
-  start = loop;
-
- worker->ch[ch].start = start & 0x7ffff;
- worker->ch[ch].loop = loop;
-
- return 0;
-}
-
-#endif
-
 // if irq is going to trigger sooner than in upd_samples, set upd_samples
 static void scan_for_irq(int ch, unsigned int *upd_samples)
 {
@@ -594,8 +531,9 @@ static void scan_for_irq(int ch, unsigned int *upd_samples)
 }
 
 #define make_do_samples(name, fmod_code, interp_start, interp1_code, interp2_code, interp_end) \
-static noinline int do_samples_##name(int (*decode_f)(int ch, int *SB), int ch, \
- int ns_to, int *SB, int sinc, int *spos, int *sbpos) \
+static noinline int do_samples_##name( \
+ int (*decode_f)(void *context, int ch, int *SB), void *ctx, \
+ int ch, int ns_to, int *SB, int sinc, int *spos, int *sbpos) \
 {                                            \
  int ns, d, fa;                              \
  int ret = ns_to;                            \
@@ -612,7 +550,7 @@ static noinline int do_samples_##name(int (*decode_f)(int ch, int *SB), int ch,
    if (*sbpos >= 28)                         \
    {                                         \
     *sbpos = 0;                              \
-    d = decode_f(ch, SB);                    \
+    d = decode_f(ctx, ch, SB);               \
     if (d && ns < ret)                       \
      ret = ns;                               \
    }                                         \
@@ -723,13 +661,12 @@ static int do_samples_noise(int ch, int ns_to)
 
 #ifdef HAVE_ARMV5
 // 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, int *rvb);
+extern void mix_chan(int *SSumLR, int count, int lv, int rv);
+extern void mix_chan_rvb(int *SSumLR, int count, int lv, int rv, int *rvb);
 #else
-static void mix_chan(int start, int count, int lv, int rv)
+static void mix_chan(int *SSumLR, int count, int lv, int rv)
 {
- int *dst = SSumLR + start * 2;
- const int *src = ChanBuf + start;
+ const int *src = ChanBuf;
  int l, r;
 
  while (count--)
@@ -738,16 +675,16 @@ static void mix_chan(int start, int count, int lv, int rv)
 
    l = (sval * lv) >> 14;
    r = (sval * rv) >> 14;
-   *dst++ += l;
-   *dst++ += r;
+   *SSumLR++ += l;
+   *SSumLR++ += r;
   }
 }
 
-static void mix_chan_rvb(int start, int count, int lv, int rv, int *rvb)
+static void mix_chan_rvb(int *SSumLR, int count, int lv, int rv, int *rvb)
 {
- int *dst = SSumLR + start * 2;
- int *drvb = rvb + start * 2;
- const int *src = ChanBuf + start;
+ const int *src = ChanBuf;
+ int *dst = SSumLR;
+ int *drvb = rvb;
  int l, r;
 
  while (count--)
@@ -828,7 +765,13 @@ static void do_channels(int ns_to)
  int *SB, sinc;
  int ch, d;
 
- InitREVERB(ns_to);
+ memset(spu.RVB, 0, ns_to * sizeof(spu.RVB[0]) * 2);
+
+ mask = spu.dwNewChannel & 0xffffff;
+ for (ch = 0; mask != 0; ch++, mask >>= 1) {
+  if (mask & 1)
+   StartSound(ch);
+ }
 
  mask = spu.dwChannelOn & 0xffffff;
  for (ch = 0; mask != 0; ch++, mask >>= 1)         // loop em all...
@@ -843,13 +786,13 @@ static void do_channels(int ns_to)
     d = do_samples_noise(ch, ns_to);
    else if (s_chan->bFMod == 2
          || (s_chan->bFMod == 0 && spu_config.iUseInterpolation == 0))
-    d = do_samples_noint(decode_block, ch, ns_to,
+    d = do_samples_noint(decode_block, NULL, ch, ns_to,
           SB, sinc, &s_chan->spos, &s_chan->iSBPos);
    else if (s_chan->bFMod == 0 && spu_config.iUseInterpolation == 1)
-    d = do_samples_simple(decode_block, ch, ns_to,
+    d = do_samples_simple(decode_block, NULL, ch, ns_to,
           SB, sinc, &s_chan->spos, &s_chan->iSBPos);
    else
-    d = do_samples_default(decode_block, ch, ns_to,
+    d = do_samples_default(decode_block, NULL, ch, ns_to,
           SB, sinc, &s_chan->spos, &s_chan->iSBPos);
 
    d = MixADSR(&s_chan->ADSRX, d);
@@ -868,150 +811,240 @@ static void do_channels(int ns_to)
    if (s_chan->bFMod == 2)                         // fmod freq channel
     memcpy(iFMod, &ChanBuf, ns_to * sizeof(iFMod[0]));
    if (s_chan->bRVBActive)
-    mix_chan_rvb(0, ns_to, s_chan->iLeftVolume, s_chan->iRightVolume, spu.sRVBStart);
+    mix_chan_rvb(spu.SSumLR, ns_to, s_chan->iLeftVolume, s_chan->iRightVolume, spu.RVB);
    else
-    mix_chan(0, ns_to, s_chan->iLeftVolume, s_chan->iRightVolume);
+    mix_chan(spu.SSumLR, ns_to, s_chan->iLeftVolume, s_chan->iRightVolume);
   }
 }
 
-static void do_samples_finish(int ns_to, int silentch, int decode_pos);
+static void do_samples_finish(int *SSumLR, int *RVB, int ns_to,
+ int silentch, int decode_pos);
 
 // optional worker thread handling
 
 #if defined(THREAD_ENABLED) || defined(WANT_THREAD_CODE)
 
+// worker thread state
+static struct spu_worker {
+ union {
+  struct {
+   unsigned int exit_thread;
+   unsigned int i_ready;
+   unsigned int i_reaped;
+   unsigned int req_sent; // dsp
+   unsigned int last_boot_cnt;
+  };
+  // aligning for C64X_DSP
+  unsigned int _pad0[128/4];
+ };
+ union {
+  struct {
+   unsigned int i_done;
+   unsigned int active; // dsp
+   unsigned int boot_cnt;
+  };
+  unsigned int _pad1[128/4];
+ };
+ struct work_item {
+  int ns_to;
+  int ctrl;
+  int decode_pos;
+  unsigned int channels_new;
+  unsigned int channels_on;
+  unsigned int channels_silent;
+  struct {
+   int spos;
+   int sbpos;
+   int sinc;
+   int start;
+   int loop;
+   int ns_to;
+   ADSRInfoEx adsr;
+   // might want to add vol and fmod flags..
+  } ch[24];
+  int RVB[NSSIZE * 2];
+  int SSumLR[NSSIZE * 2];
+ } i[4];
+} *worker;
+
+#define WORK_MAXCNT (sizeof(worker->i) / sizeof(worker->i[0]))
+#define WORK_I_MASK (WORK_MAXCNT - 1)
+
 static void thread_work_start(void);
-static void thread_work_wait_sync(void);
-static void thread_sync_caches(void);
+static void thread_work_wait_sync(struct work_item *work, int force);
+static int  thread_get_i_done(void);
+
+static int decode_block_work(void *context, int ch, int *SB)
+{
+ const unsigned char *ram = spu.spuMemC;
+ int predict_nr, shift_factor, flags;
+ struct work_item *work = context;
+ int start = work->ch[ch].start;
+ int loop = work->ch[ch].loop;
+
+ predict_nr = ram[start];
+ shift_factor = predict_nr & 0xf;
+ predict_nr >>= 4;
+
+ decode_block_data(SB, ram + start + 2, predict_nr, shift_factor);
+
+ flags = ram[start + 1];
+ if (flags & 4)
+  loop = start;                            // loop adress
+
+ start += 16;
+
+ if (flags & 1)                            // 1: stop/loop
+  start = loop;
+
+ work->ch[ch].start = start & 0x7ffff;
+ work->ch[ch].loop = loop;
+
+ return 0;
+}
 
-static void queue_channel_work(int ns_to, int silentch)
+static void queue_channel_work(int ns_to, unsigned int silentch)
 {
- const SPUCHAN *s_chan;
+ struct work_item *work;
+ SPUCHAN *s_chan;
  unsigned int mask;
- int ch;
+ int ch, d;
 
- worker->ns_to = ns_to;
- worker->ctrl = spu.spuCtrl;
- worker->decode_pos = spu.decode_pos;
- worker->silentch = silentch;
+ work = &worker->i[worker->i_ready & WORK_I_MASK];
+ work->ns_to = ns_to;
+ work->ctrl = spu.spuCtrl;
+ work->decode_pos = spu.decode_pos;
+ work->channels_silent = silentch;
+
+ mask = work->channels_new = spu.dwNewChannel & 0xffffff;
+ for (ch = 0; mask != 0; ch++, mask >>= 1) {
+  if (mask & 1)
+   StartSoundMain(ch);
+ }
+
+ mask = work->channels_on = spu.dwChannelOn & 0xffffff;
+ spu.decode_dirty_ch |= mask & 0x0a;
 
- mask = worker->chmask = spu.dwChannelOn & 0xffffff;
  for (ch = 0; mask != 0; ch++, mask >>= 1)
   {
    if (!(mask & 1)) continue;
 
    s_chan = &spu.s_chan[ch];
-   worker->ch[ch].spos = s_chan->spos;
-   worker->ch[ch].sbpos = s_chan->iSBPos;
-   worker->ch[ch].sinc = s_chan->sinc;
-   worker->ch[ch].adsr = s_chan->ADSRX;
-   worker->ch[ch].start = s_chan->pCurr - spu.spuMemC;
-   worker->ch[ch].loop = s_chan->pLoop - spu.spuMemC;
+   work->ch[ch].spos = s_chan->spos;
+   work->ch[ch].sbpos = s_chan->iSBPos;
+   work->ch[ch].sinc = s_chan->sinc;
+   work->ch[ch].adsr = s_chan->ADSRX;
+   work->ch[ch].start = s_chan->pCurr - spu.spuMemC;
+   work->ch[ch].loop = s_chan->pLoop - spu.spuMemC;
    if (s_chan->prevflags & 1)
-    worker->ch[ch].start = worker->ch[ch].loop;
+    work->ch[ch].start = work->ch[ch].loop;
 
-   worker->ch[ch].ns_to = do_samples_skip(ch, ns_to);
+   d = do_samples_skip(ch, ns_to);
+   work->ch[ch].ns_to = d;
+
+   // note: d is not accurate on skip
+   d = SkipADSR(&s_chan->ADSRX, d);
+   if (d < ns_to) {
+    spu.dwChannelOn &= ~(1 << ch);
+    s_chan->ADSRX.EnvelopeVol = 0;
+   }
   }
 
- worker->pending = 1;
+ worker->i_ready++;
  thread_work_start();
 }
 
-static void do_channel_work(void)
+static void do_channel_work(struct work_item *work)
 {
- unsigned int mask, endmask = 0;
+ unsigned int mask;
  unsigned int decode_dirty_ch = 0;
  int *SB, sinc, spos, sbpos;
  int d, ch, ns_to;
  SPUCHAN *s_chan;
 
- ns_to = worker->ns_to;
- memset(spu.sRVBStart, 0, ns_to * sizeof(spu.sRVBStart[0]) * 2);
+ ns_to = work->ns_to;
+ memset(work->RVB, 0, ns_to * sizeof(work->RVB[0]) * 2);
 
- mask = worker->chmask;
+ mask = work->channels_new;
+ for (ch = 0; mask != 0; ch++, mask >>= 1) {
+  if (mask & 1)
+   StartSoundSB(spu.SB + ch * SB_SIZE);
+ }
+
+ mask = work->channels_on;
  for (ch = 0; mask != 0; ch++, mask >>= 1)
   {
    if (!(mask & 1)) continue;
 
-   d = worker->ch[ch].ns_to;
-   spos = worker->ch[ch].spos;
-   sbpos = worker->ch[ch].sbpos;
-   sinc = worker->ch[ch].sinc;
+   d = work->ch[ch].ns_to;
+   spos = work->ch[ch].spos;
+   sbpos = work->ch[ch].sbpos;
+   sinc = work->ch[ch].sinc;
 
    s_chan = &spu.s_chan[ch];
    SB = spu.SB + ch * SB_SIZE;
 
    if (s_chan->bNoise)
-    do_lsfr_samples(d, worker->ctrl, &spu.dwNoiseCount, &spu.dwNoiseVal);
+    do_lsfr_samples(d, work->ctrl, &spu.dwNoiseCount, &spu.dwNoiseVal);
    else if (s_chan->bFMod == 2
          || (s_chan->bFMod == 0 && spu_config.iUseInterpolation == 0))
-    do_samples_noint(decode_block_work, ch, d, SB, sinc, &spos, &sbpos);
+    do_samples_noint(decode_block_work, work, ch, d, SB, sinc, &spos, &sbpos);
    else if (s_chan->bFMod == 0 && spu_config.iUseInterpolation == 1)
-    do_samples_simple(decode_block_work, ch, d, SB, sinc, &spos, &sbpos);
+    do_samples_simple(decode_block_work, work, ch, d, SB, sinc, &spos, &sbpos);
    else
-    do_samples_default(decode_block_work, ch, d, SB, sinc, &spos, &sbpos);
+    do_samples_default(decode_block_work, work, ch, d, SB, sinc, &spos, &sbpos);
 
-   d = MixADSR(&worker->ch[ch].adsr, d);
+   d = MixADSR(&work->ch[ch].adsr, d);
    if (d < ns_to) {
-    endmask |= 1 << ch;
-    worker->ch[ch].adsr.EnvelopeVol = 0;
+    work->ch[ch].adsr.EnvelopeVol = 0;
     memset(&ChanBuf[d], 0, (ns_to - d) * sizeof(ChanBuf[0]));
    }
-   worker->r.ch[ch].adsrState = worker->ch[ch].adsr.State;
-   worker->r.ch[ch].adsrEnvelopeVol = worker->ch[ch].adsr.EnvelopeVol;
 
    if (ch == 1 || ch == 3)
     {
-     do_decode_bufs(spu.spuMem, ch/2, ns_to, worker->decode_pos);
+     do_decode_bufs(spu.spuMem, ch/2, ns_to, work->decode_pos);
      decode_dirty_ch |= 1 << ch;
     }
 
    if (s_chan->bFMod == 2)                         // fmod freq channel
     memcpy(iFMod, &ChanBuf, ns_to * sizeof(iFMod[0]));
    if (s_chan->bRVBActive)
-    mix_chan_rvb(0, ns_to, s_chan->iLeftVolume, s_chan->iRightVolume, spu.sRVBStart);
+    mix_chan_rvb(work->SSumLR, ns_to,
+      s_chan->iLeftVolume, s_chan->iRightVolume, work->RVB);
    else
-    mix_chan(0, ns_to, s_chan->iLeftVolume, s_chan->iRightVolume);
+    mix_chan(work->SSumLR, ns_to, s_chan->iLeftVolume, s_chan->iRightVolume);
   }
-
-  worker->r.chan_end = endmask;
-  worker->r.decode_dirty = decode_dirty_ch;
 }
 
-static void sync_worker_thread(int do_direct)
+static void sync_worker_thread(int force)
 {
- unsigned int mask;
- int ch;
+ struct work_item *work;
+ int done, used_space;
 
- if (do_direct)
-  thread_sync_caches();
- if (!worker->pending)
-  return;
+ done = thread_get_i_done() - worker->i_reaped;
+ used_space = worker->i_ready - worker->i_reaped;
+ //printf("done: %d use: %d dsp: %u/%u\n", done, used_space,
+ //  worker->boot_cnt, worker->i_done);
 
- thread_work_wait_sync();
- worker->pending = 0;
+ while ((force && used_space > 0) || used_space >= WORK_MAXCNT || done > 0) {
+  work = &worker->i[worker->i_reaped & WORK_I_MASK];
+  thread_work_wait_sync(work, force);
 
- mask = worker->chmask;
- for (ch = 0; mask != 0; ch++, mask >>= 1) {
-  if (!(mask & 1)) continue;
+  do_samples_finish(work->SSumLR, work->RVB, work->ns_to,
+   work->channels_silent, work->decode_pos);
 
-  // be sure there was no keyoff while thread was working
-  if (spu.s_chan[ch].ADSRX.State != ADSR_RELEASE)
-    spu.s_chan[ch].ADSRX.State = worker->r.ch[ch].adsrState;
-  spu.s_chan[ch].ADSRX.EnvelopeVol = worker->r.ch[ch].adsrEnvelopeVol;
+  worker->i_reaped++;
+  done = thread_get_i_done() - worker->i_reaped;
+  used_space = worker->i_ready - worker->i_reaped;
  }
-
- spu.dwChannelOn &= ~worker->r.chan_end;
- spu.decode_dirty_ch |= worker->r.decode_dirty;
-
- do_samples_finish(worker->ns_to, worker->silentch,
-  worker->decode_pos);
 }
 
 #else
 
 static void queue_channel_work(int ns_to, int silentch) {}
-static void sync_worker_thread(int do_direct) {}
+static void sync_worker_thread(int force) {}
+
+static const void * const worker = NULL;
 
 #endif // THREAD_ENABLED
 
@@ -1022,10 +1055,9 @@ static void sync_worker_thread(int do_direct) {}
 
 void do_samples(unsigned int cycles_to, int do_direct)
 {
- unsigned int mask;
- int ch, ns_to;
- int silentch;
+ unsigned int silentch;
  int cycle_diff;
+ int ns_to;
 
  cycle_diff = cycles_to - spu.cycles_played;
  if (cycle_diff < -2*1048576 || cycle_diff > 2*1048576)
@@ -1035,7 +1067,9 @@ void do_samples(unsigned int cycles_to, int do_direct)
    return;
   }
 
- do_direct |= (cycle_diff < 64 * 768);
+ silentch = ~(spu.dwChannelOn | spu.dwNewChannel) & 0xffffff;
+
+ do_direct |= (silentch == 0xffffff);
  if (worker != NULL)
   sync_worker_thread(do_direct);
 
@@ -1078,26 +1112,12 @@ void do_samples(unsigned int cycles_to, int do_direct)
      }
    }
 
-  mask = spu.dwNewChannel & 0xffffff;
-  for (ch = 0; mask != 0; ch++, mask >>= 1) {
-   if (mask & 1)
-    StartSound(ch);
-  }
-
-  silentch = ~spu.dwChannelOn & 0xffffff;
-
-  if (spu.dwChannelOn == 0) {
-   InitREVERB(ns_to);
-   do_samples_finish(ns_to, silentch, spu.decode_pos);
+  if (do_direct || worker == NULL || !spu_config.iUseThread) {
+   do_channels(ns_to);
+   do_samples_finish(spu.SSumLR, spu.RVB, ns_to, silentch, spu.decode_pos);
   }
   else {
-   if (do_direct || worker == NULL || !spu_config.iUseThread) {
-    do_channels(ns_to);
-    do_samples_finish(ns_to, silentch, spu.decode_pos);
-   }
-   else {
-    queue_channel_work(ns_to, silentch);
-   }
+   queue_channel_work(ns_to, silentch);
   }
 
   // advance "stopped" channels that can cause irqs
@@ -1109,13 +1129,15 @@ void do_samples(unsigned int cycles_to, int do_direct)
   spu.decode_pos = (spu.decode_pos + ns_to) & 0x1ff;
 }
 
-static void do_samples_finish(int ns_to, int silentch, int decode_pos)
+static void do_samples_finish(int *SSumLR, int *RVB, int ns_to,
+ int silentch, int decode_pos)
 {
   int volmult = spu_config.iVolume;
   int ns;
   int d;
 
-  if(unlikely(silentch & spu.decode_dirty_ch & (1<<1))) // must clear silent channel decode buffers
+  // must clear silent channel decode buffers
+  if(unlikely(silentch & spu.decode_dirty_ch & (1<<1)))
    {
     memset(&spu.spuMem[0x800/2], 0, 0x400);
     spu.decode_dirty_ch &= ~(1<<1);
@@ -1129,13 +1151,13 @@ static void do_samples_finish(int ns_to, int silentch, int decode_pos)
   //---------------------------------------------------//
   // mix XA infos (if any)
 
-  MixXA(ns_to, decode_pos);
+  MixXA(SSumLR, ns_to, decode_pos);
   
   ///////////////////////////////////////////////////////
   // mix all channels (including reverb) into one buffer
 
   if(spu_config.iUseReverb)
-   REVERBDo(ns_to);
+   REVERBDo(SSumLR, RVB, ns_to);
 
   if((spu.spuCtrl&0x4000)==0) // muted? (rare, don't optimize for this)
    {
@@ -1254,7 +1276,6 @@ int CALLBACK SPUplayCDDAchannel(short *pcm, int nbytes)
 // to be called after state load
 void ClearWorkingState(void)
 {
- memset(SSumLR, 0, NSSIZE * 2 * 4);                    // init some mixing buffers
  memset(iFMod, 0, sizeof(iFMod));
  spu.pS=(short *)spu.pSpuBuffer;                       // setup soundbuffer pointer
 }
@@ -1265,8 +1286,8 @@ void SetupStreams(void)
  int i;
 
  spu.pSpuBuffer = (unsigned char *)malloc(32768);      // alloc mixing buffer
- spu.sRVBStart = calloc(NSSIZE * 2, sizeof(spu.sRVBStart[0]));
- SSumLR = calloc(NSSIZE * 2, sizeof(SSumLR[0]));
+ spu.RVB = calloc(NSSIZE * 2, sizeof(spu.RVB[0]));
+ spu.SSumLR = calloc(NSSIZE * 2, sizeof(spu.SSumLR[0]));
 
  spu.XAStart =                                         // alloc xa buffer
   (uint32_t *)malloc(44100 * sizeof(uint32_t));
@@ -1298,10 +1319,10 @@ void RemoveStreams(void)
 { 
  free(spu.pSpuBuffer);                                 // free mixing buffer
  spu.pSpuBuffer = NULL;
- free(spu.sRVBStart);                                  // free reverb buffer
- spu.sRVBStart = NULL;
- free(SSumLR);
- SSumLR = NULL;
+ free(spu.RVB);                                        // free reverb buffer
+ spu.RVB = NULL;
+ free(spu.SSumLR);
+ spu.SSumLR = NULL;
  free(spu.XAStart);                                    // free XA buffer
  spu.XAStart = NULL;
  free(spu.CDDAStart);                                  // free CDDA buffer
@@ -1332,23 +1353,28 @@ static void thread_work_start(void)
  sem_post(&t.sem_avail);
 }
 
-static void thread_work_wait_sync(void)
+static void thread_work_wait_sync(struct work_item *work, int force)
 {
  sem_wait(&t.sem_done);
 }
 
-static void thread_sync_caches(void)
+static int thread_get_i_done(void)
 {
+ return worker->i_done;
 }
 
 static void *spu_worker_thread(void *unused)
 {
+ struct work_item *work;
+
  while (1) {
   sem_wait(&t.sem_avail);
   if (worker->exit_thread)
    break;
 
-  do_channel_work();
+  work = &worker->i[worker->i_done & WORK_I_MASK];
+  do_channel_work(work);
+  worker->i_done++;
 
   sem_post(&t.sem_done);
  }
@@ -1377,6 +1403,7 @@ static void init_spu_thread(void)
  if (ret != 0)
   goto fail_thread;
 
+ spu_config.iThreadAvail = 1;
  return;
 
 fail_thread:
@@ -1386,6 +1413,7 @@ fail_sem_done:
 fail_sem_avail:
  free(worker);
  worker = NULL;
+ spu_config.iThreadAvail = 0;
 }
 
 static void exit_spu_thread(void)
diff --git a/plugins/dfsound/spu_c64x.c b/plugins/dfsound/spu_c64x.c
index d829d29..200ab38 100644
--- a/plugins/dfsound/spu_c64x.c
+++ b/plugins/dfsound/spu_c64x.c
@@ -23,13 +23,11 @@
 
 #include <dlfcn.h>
 #include <stddef.h>
+#include <unistd.h>
 
 #include <inc_libc64_mini.h>
 #include "spu_c64x.h"
 
-static dsp_mem_region_t region;
-static dsp_component_id_t compid;
-
 static struct {
  void *handle;
  int  (*dsp_open)(void);
@@ -42,49 +40,114 @@ static struct {
  int  (*dsp_rpc_recv)(dsp_msg_t *_msgFrom);
  int  (*dsp_rpc)(const dsp_msg_t *_msgTo, dsp_msg_t *_msgFrom);
  void (*dsp_logbuf_print)(void);
+
+ dsp_mem_region_t region;
+ dsp_component_id_t compid;
 } f;
 
 static void thread_work_start(void)
 {
+ struct region_mem *mem;
  dsp_msg_t msg;
  int ret;
 
- DSP_MSG_INIT(&msg, compid, CCMD_DOIT, 0, 0);
+ // make sure new work is written out
+ __sync_synchronize();
+
+ // this should be safe, as dsp checks for new work even
+ // after it decrements ->active
+ // cacheline: i_done, active
+ f.dsp_cache_inv_virt(&worker->i_done, 64);
+ if (worker->active == ACTIVE_CNT)
+  return;
+
+ // to start the DSP, dsp_rpc_send() must be used,
+ // but before that, previous request must be finished
+ if (worker->req_sent) {
+  if (worker->boot_cnt == worker->last_boot_cnt) {
+   // hopefully still booting
+   //printf("booting?\n");
+   return;
+  }
+
+  ret = f.dsp_rpc_recv(&msg);
+  if (ret != 0) {
+   fprintf(stderr, "dsp_rpc_recv failed: %d\n", ret);
+   f.dsp_logbuf_print();
+   worker->req_sent = 0;
+   spu_config.iUseThread = 0;
+   return;
+  }
+ }
+
+ f.dsp_cache_inv_virt(&worker->i_done, 64);
+ worker->last_boot_cnt = worker->boot_cnt;
+
+ mem = (void *)f.region.virt_addr;
+ memcpy(&mem->spu_config, &spu_config, sizeof(mem->spu_config));
+
+ DSP_MSG_INIT(&msg, f.compid, CCMD_DOIT, f.region.phys_addr, 0);
  ret = f.dsp_rpc_send(&msg);
  if (ret != 0) {
   fprintf(stderr, "dsp_rpc_send failed: %d\n", ret);
   f.dsp_logbuf_print();
-  // maybe stop using the DSP?
+  spu_config.iUseThread = 0;
+  return;
  }
+ worker->req_sent = 1;
 }
 
-static void thread_work_wait_sync(void)
+static int thread_get_i_done(void)
 {
- dsp_msg_t msg;
+ f.dsp_cache_inv_virt(&worker->i_done, sizeof(worker->i_done));
+ return worker->i_done;
+}
+
+static void thread_work_wait_sync(struct work_item *work, int force)
+{
+ int limit = 1000;
  int ns_to;
- int ret;
 
- ns_to = worker->ns_to;
- f.dsp_cache_inv_virt(spu.sRVBStart, sizeof(spu.sRVBStart[0]) * 2 * ns_to);
- f.dsp_cache_inv_virt(SSumLR, sizeof(SSumLR[0]) * 2 * ns_to);
- f.dsp_cache_inv_virt(&worker->r, sizeof(worker->r));
- worker->stale_cache = 1; // SB, ram
+ ns_to = work->ns_to;
+ f.dsp_cache_inv_virt(work->RVB, sizeof(work->RVB[0]) * 2 * ns_to);
+ f.dsp_cache_inv_virt(work->SSumLR, sizeof(work->SSumLR[0]) * 2 * ns_to);
+ __builtin_prefetch(work->RVB);
+ __builtin_prefetch(work->SSumLR);
 
- ret = f.dsp_rpc_recv(&msg);
- if (ret != 0) {
-  fprintf(stderr, "dsp_rpc_recv failed: %d\n", ret);
-  f.dsp_logbuf_print();
+ while (worker->i_done == worker->i_reaped && limit-- > 0) {
+  if (!worker->active) {
+   printf("dsp: broken sync\n");
+   worker->last_boot_cnt = ~0;
+   break;
+  }
+
+  usleep(500);
+  f.dsp_cache_inv_virt(&worker->i_done, 64);
  }
- //f.dsp_logbuf_print();
-}
 
-// called before ARM decides to do SPU mixing itself
-static void thread_sync_caches(void)
-{
- if (worker->stale_cache) {
+ if (limit == 0)
+  printf("dsp: wait timeout\n");
+
+ // still in results loop?
+ if (worker->i_reaped != worker->i_done - 1)
+  return;
+
+ if (worker->req_sent && (force || worker->i_done == worker->i_ready)) {
+  dsp_msg_t msg;
+  int ret;
+
+  ret = f.dsp_rpc_recv(&msg);
+  if (ret != 0) {
+   fprintf(stderr, "dsp_rpc_recv failed: %d\n", ret);
+   f.dsp_logbuf_print();
+   spu_config.iUseThread = 0;
+  }
+  worker->req_sent = 0;
+ }
+
+ if (force) {
   f.dsp_cache_inv_virt(spu.SB, sizeof(spu.SB[0]) * SB_SIZE * 24);
   f.dsp_cache_inv_virt(spu.spuMemC + 0x800, 0x800);
-  worker->stale_cache = 0;
  }
 }
 
@@ -101,7 +164,7 @@ static void init_spu_thread(void)
   f.handle = dlopen(lib, RTLD_NOW);
   if (f.handle == NULL) {
    fprintf(stderr, "can't load %s: %s\n", lib, dlerror());
-   return;
+   goto fail_open;
   }
   #define LDS(name) \
     failed |= (f.name = dlsym(f.handle, #name)) == NULL
@@ -120,32 +183,32 @@ static void init_spu_thread(void)
    fprintf(stderr, "missing symbol(s) in %s\n", lib);
    dlclose(f.handle);
    f.handle = NULL;
-   return;
+   goto fail_open;
   }
  }
 
  ret = f.dsp_open();
  if (ret != 0) {
   fprintf(stderr, "dsp_open failed: %d\n", ret);
-  return;
+  goto fail_open;
  }
 
- ret = f.dsp_component_load(NULL, COMPONENT_NAME, &compid);
+ ret = f.dsp_component_load(NULL, COMPONENT_NAME, &f.compid);
  if (ret != 0) {
   fprintf(stderr, "dsp_component_load failed: %d\n", ret);
   goto fail_cload;
  }
 
- region = f.dsp_shm_alloc(DSP_CACHE_R, sizeof(*mem)); // writethrough
- if (region.size < sizeof(*mem) || region.virt_addr == 0) {
+ f.region = f.dsp_shm_alloc(DSP_CACHE_R, sizeof(*mem)); // writethrough
+ if (f.region.size < sizeof(*mem) || f.region.virt_addr == 0) {
   fprintf(stderr, "dsp_shm_alloc failed\n");
   goto fail_mem;
  }
- mem = (void *)region.virt_addr;
+ mem = (void *)f.region.virt_addr;
 
  memcpy(&mem->spu_config, &spu_config, sizeof(mem->spu_config));
 
- DSP_MSG_INIT(&init_msg, compid, CCMD_INIT, region.phys_addr, 0);
+ DSP_MSG_INIT(&init_msg, f.compid, CCMD_INIT, f.region.phys_addr, 0);
  ret = f.dsp_rpc(&init_msg, &msg_in);
  if (ret != 0) {
   fprintf(stderr, "dsp_rpc failed: %d\n", ret);
@@ -162,56 +225,56 @@ static void init_spu_thread(void)
     mem->offsetof_s_chan1, offsetof(typeof(*mem), s_chan[1]));
   goto fail_init;
  }
- if (mem->offsetof_worker_ram != offsetof(typeof(*mem), worker.ch[1])) {
+ if (mem->offsetof_spos_3_20 != offsetof(typeof(*mem), worker.i[3].ch[20])) {
   fprintf(stderr, "error: size mismatch 3: %d vs %zd\n",
-    mem->offsetof_worker_ram, offsetof(typeof(*mem), worker.ch[1]));
+    mem->offsetof_spos_3_20, offsetof(typeof(*mem), worker.i[3].ch[20]));
   goto fail_init;
  }
 
  // override default allocations
  free(spu.spuMemC);
  spu.spuMemC = mem->spu_ram;
- free(spu.sRVBStart);
- spu.sRVBStart = mem->RVB;
- free(SSumLR);
- SSumLR = mem->SSumLR;
  free(spu.SB);
  spu.SB = mem->SB;
  free(spu.s_chan);
  spu.s_chan = mem->s_chan;
  worker = &mem->worker;
 
- printf("spu: C64x DSP ready (id=%d).\n", (int)compid);
+ printf("spu: C64x DSP ready (id=%d).\n", (int)f.compid);
  f.dsp_logbuf_print();
 
-pcnt_init();
+ spu_config.iThreadAvail = 1;
  (void)do_channel_work; // used by DSP instead
  return;
 
 fail_init:
- f.dsp_shm_free(region);
+ f.dsp_shm_free(f.region);
 fail_mem:
  // no component unload func?
 fail_cload:
- printf("spu: C64x DSP init failed.\n");
  f.dsp_logbuf_print();
  f.dsp_close();
+fail_open:
+ printf("spu: C64x DSP init failed.\n");
+ spu_config.iUseThread = spu_config.iThreadAvail = 0;
  worker = NULL;
 }
 
 static void exit_spu_thread(void)
 {
+ dsp_msg_t msg;
+
  if (worker == NULL)
   return;
 
- if (worker->pending)
-  thread_work_wait_sync();
- f.dsp_shm_free(region);
+ if (worker->req_sent)
+  f.dsp_rpc_recv(&msg);
+
+ f.dsp_logbuf_print();
+ f.dsp_shm_free(f.region);
  f.dsp_close();
 
  spu.spuMemC = NULL;
- spu.sRVBStart = NULL;
- SSumLR = NULL;
  spu.SB = NULL;
  spu.s_chan = NULL;
  worker = NULL;
diff --git a/plugins/dfsound/spu_c64x.h b/plugins/dfsound/spu_c64x.h
index d4e73e9..bb20cc3 100644
--- a/plugins/dfsound/spu_c64x.h
+++ b/plugins/dfsound/spu_c64x.h
@@ -7,12 +7,10 @@ enum {
 
 struct region_mem {
  unsigned char spu_ram[512 * 1024];
- int RVB[NSSIZE * 2];
- int SSumLR[NSSIZE * 2];
  int SB[SB_SIZE * 24];
  // careful not to lose ARM writes by DSP overwriting
  // with old data when it's writing out neighbor cachelines
- int _pad1[128/4 - ((NSSIZE * 4 + SB_SIZE * 24) & (128/4 - 1))];
+ int _pad1[128/4 - ((SB_SIZE * 24) & (128/4 - 1))];
  SPUCHAN s_chan[24 + 1];
  int _pad2[128/4 - ((sizeof(SPUCHAN) * 25 / 4) & (128/4 - 1))];
  struct spu_worker worker;
@@ -20,6 +18,7 @@ struct region_mem {
  // init/debug
  int sizeof_region_mem;
  int offsetof_s_chan1;
- int offsetof_worker_ram;
+ int offsetof_spos_3_20;
 };
 
+#define ACTIVE_CNT 3
diff --git a/plugins/dfsound/spu_c64x_dspcode.c b/plugins/dfsound/spu_c64x_dspcode.c
index 117a296..97d3028 100644
--- a/plugins/dfsound/spu_c64x_dspcode.c
+++ b/plugins/dfsound/spu_c64x_dspcode.c
@@ -30,38 +30,80 @@
 
 /* dummy deps, some bloat but avoids ifdef hell in SPU code.. */
 static void thread_work_start(void) {}
-static void thread_work_wait_sync(void) {}
-static void thread_sync_caches(void) {}
+static void thread_work_wait_sync(struct work_item *work, int force) {}
+static int  thread_get_i_done(void) { return 0; }
 struct out_driver *out_current;
 void SetupSound(void) {}
 
-#if 0
-// no use, c64_tools does BCACHE_wbInvAll..
-static void sync_caches(void)
+
+static void invalidate_cache(struct work_item *work)
+{
+ syscalls.cache_inv(work, offsetof(typeof(*work), RVB), 1);
+ syscalls.cache_inv(spu.s_chan, sizeof(spu.s_chan[0]) * 24, 0);
+ syscalls.cache_inv(work->SSumLR,
+   sizeof(work->SSumLR[0]) * 2 * work->ns_to, 0);
+}
+
+static void writeout_cache(struct work_item *work)
 {
- int ns_to = worker->ns_to;
+ int ns_to = work->ns_to;
+
+ syscalls.cache_wb(work->RVB, sizeof(work->RVB[0]) * 2 * ns_to, 1);
+ syscalls.cache_wb(work->SSumLR, sizeof(work->SSumLR[0]) * 2 * ns_to, 1);
+}
 
- syscalls.cache_wb(spu.sRVBStart, sizeof(spu.sRVBStart[0]) * 2 * ns_to, 1);
- syscalls.cache_wb(SSumLR, sizeof(SSumLR[0]) * 2 * ns_to, 1);
+static void do_processing(void)
+{
+ struct work_item *work;
+ int left, dirty = 0;
 
- syscalls.cache_wbInv(worker, sizeof(*worker), 1);
+ while (worker->active)
+ {
+  // i_ready is in first cacheline
+  syscalls.cache_inv(worker, 64, 1);
+
+  left = worker->i_ready - worker->i_done;
+  if (left > 0) {
+   dirty = 1;
+   worker->active = ACTIVE_CNT;
+   syscalls.cache_wb(&worker->active, 4, 1);
+
+   work = &worker->i[worker->i_done & WORK_I_MASK];
+   invalidate_cache(work);
+   do_channel_work(work);
+   writeout_cache(work);
+
+   worker->i_done++;
+   syscalls.cache_wb(&worker->i_done, 4, 1);
+   continue;
+  }
+
+  // nothing to do? Write out non-critical caches
+  if (dirty) {
+   syscalls.cache_wb(spu.spuMemC + 0x800, 0x800, 1);
+   syscalls.cache_wb(spu.SB, sizeof(spu.SB[0]) * SB_SIZE * 24, 1);
+   dirty = 0;
+   continue;
+  }
+
+  // this ->active loop thing is to avoid a race where we miss
+  // new work and clear ->active just after ARM checks it
+  worker->active--;
+  syscalls.cache_wb(&worker->active, 4, 1);
+ }
 }
-#endif
 
 static unsigned int exec(dsp_component_cmd_t cmd,
   unsigned int arg1, unsigned int arg2,
   unsigned int *ret1, unsigned int *ret2)
 {
  struct region_mem *mem = (void *)arg1;
- int i;
 
  switch (cmd) {
   case CCMD_INIT:
    InitADSR();
 
    spu.spuMemC = mem->spu_ram;
-   spu.sRVBStart = mem->RVB;
-   SSumLR = mem->SSumLR;
    spu.SB = mem->SB;
    spu.s_chan = mem->s_chan;
    worker = &mem->worker;
@@ -69,18 +111,22 @@ static unsigned int exec(dsp_component_cmd_t cmd,
 
    mem->sizeof_region_mem = sizeof(*mem);
    mem->offsetof_s_chan1 = offsetof(typeof(*mem), s_chan[1]);
-   mem->offsetof_worker_ram = offsetof(typeof(*mem), worker.ch[1]);
+   mem->offsetof_spos_3_20 = offsetof(typeof(*mem), worker.i[3].ch[20]);
    // seems to be unneeded, no write-alloc? but just in case..
    syscalls.cache_wb(&mem->sizeof_region_mem, 3 * 4, 1);
    break;
 
   case CCMD_DOIT:
-   do_channel_work();
+   worker->active = ACTIVE_CNT;
+   worker->boot_cnt++;
+   syscalls.cache_wb(&worker->i_done, 64, 1);
+   memcpy(&spu_config, &mem->spu_config, sizeof(spu_config));
+
+   do_processing();
+
    // c64_tools lib does BCACHE_wbInvAll() when it receives mailbox irq,
-   // so there is no benefit of syncing only what's needed.
-   // But call wbInvAll() anyway in case c64_tools is ever fixed..
-   //sync_caches();
-   syscalls.cache_wbInvAll();
+   // but invalidate anyway in case c64_tools is ever fixed..
+   syscalls.cache_inv(mem, sizeof(mem->spu_ram) + sizeof(mem->SB), 0);
    break;
 
   default:
diff --git a/plugins/dfsound/spu_config.h b/plugins/dfsound/spu_config.h
index 9a139e1..fce1cda 100644
--- a/plugins/dfsound/spu_config.h
+++ b/plugins/dfsound/spu_config.h
@@ -8,6 +8,9 @@ typedef struct
  int        iUseInterpolation;
  int        iTempo;
  int        iUseThread;
+
+ // status
+ int        iThreadAvail;
 } SPUConfig;
 
 extern SPUConfig spu_config;
diff --git a/plugins/dfsound/xa.c b/plugins/dfsound/xa.c
index 6c0ce4b..ad7e824 100644
--- a/plugins/dfsound/xa.c
+++ b/plugins/dfsound/xa.c
@@ -38,7 +38,7 @@ static int gauss_window[8] = {0, 0, 0, 0, 0, 0, 0, 0};
 // MIX XA & CDDA
 ////////////////////////////////////////////////////////////////////////
 
-INLINE void MixXA(int ns_to, int decode_pos)
+INLINE void MixXA(int *SSumLR, int ns_to, int decode_pos)
 {
  int cursor = decode_pos;
  int ns;