spu: fix some thread safety issues

reverb and stuff

5 files changed, 91 insertions, 98 deletions

diff --git a/plugins/dfsound/externals.h b/plugins/dfsound/externals.h
index f6fc440..8bbf4e0 100644
--- a/plugins/dfsound/externals.h
+++ b/plugins/dfsound/externals.h
@@ -162,14 +162,8 @@ typedef struct
 

  int dirty;          // registers changed

 

- // normalized offsets

- int nIIR_DEST_A0, nIIR_DEST_A1, nIIR_DEST_B0, nIIR_DEST_B1,

- 	nACC_SRC_A0, nACC_SRC_A1, nACC_SRC_B0, nACC_SRC_B1, 

-	nIIR_SRC_A0, nIIR_SRC_A1, nIIR_SRC_B0, nIIR_SRC_B1,

-	nACC_SRC_C0, nACC_SRC_C1, nACC_SRC_D0, nACC_SRC_D1,

-	nMIX_DEST_A0, nMIX_DEST_A1, nMIX_DEST_B0, nMIX_DEST_B1;

  // MIX_DEST_xx - FB_SRC_x

- int nFB_SRC_A0, nFB_SRC_A1, nFB_SRC_B0, nFB_SRC_B1;

+ int FB_SRC_A0, FB_SRC_A1, FB_SRC_B0, FB_SRC_B1;

 } REVERBInfo;

 

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

diff --git a/plugins/dfsound/freeze.c b/plugins/dfsound/freeze.c
index 5b4267b..72c9d22 100644
--- a/plugins/dfsound/freeze.c
+++ b/plugins/dfsound/freeze.c
@@ -231,6 +231,8 @@ long CALLBACK SPUfreeze(uint32_t ulFreezeMode, SPUFreeze_t * pF,
 

  if(!pF) return 0;                                     // first check

 

+ do_samples(cycles, 1);

+

  if(ulFreezeMode)                                      // info or save?

   {//--------------------------------------------------//

    if(ulFreezeMode==1)                                 

@@ -242,8 +244,6 @@ long CALLBACK SPUfreeze(uint32_t ulFreezeMode, SPUFreeze_t * pF,
 

    if(ulFreezeMode==2) return 1;                       // info mode? ok, bye

                                                        // save mode:

-   do_samples(cycles, 1);

-

    memcpy(pF->cSPURam,spu.spuMem,0x80000);             // copy common infos

    memcpy(pF->cSPUPort,spu.regArea,0x200);

 

diff --git a/plugins/dfsound/registers.c b/plugins/dfsound/registers.c
index 4588fa7..91bcaf8 100644
--- a/plugins/dfsound/registers.c
+++ b/plugins/dfsound/registers.c
@@ -145,17 +145,6 @@ void CALLBACK SPUwriteRegister(unsigned long reg, unsigned short val,
       break;

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

     case H_SPUReverbAddr:

-      if(val==0xFFFF || val<=0x200)

-       {spu.rvb->StartAddr=spu.rvb->CurrAddr=0;}

-      else

-       {

-        const long iv=(unsigned long)val<<2;

-        if(spu.rvb->StartAddr!=iv)

-         {

-          spu.rvb->StartAddr=(unsigned long)val<<2;

-          spu.rvb->CurrAddr=spu.rvb->StartAddr;

-         }

-       }

       goto rvbd;

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

     case H_SPUirqAddr:

@@ -246,38 +235,38 @@ void CALLBACK SPUwriteRegister(unsigned long reg, unsigned short val,
       ReverbOn(16,24,val);

       break;

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

-    case H_Reverb+0   : spu.rvb->FB_SRC_A=val*4;         goto rvbd;

-    case H_Reverb+2   : spu.rvb->FB_SRC_B=val*4;         goto rvbd;

-    case H_Reverb+4   : spu.rvb->IIR_ALPHA=(short)val;   goto rvbd;

-    case H_Reverb+6   : spu.rvb->ACC_COEF_A=(short)val;  goto rvbd;

-    case H_Reverb+8   : spu.rvb->ACC_COEF_B=(short)val;  goto rvbd;

-    case H_Reverb+10  : spu.rvb->ACC_COEF_C=(short)val;  goto rvbd;

-    case H_Reverb+12  : spu.rvb->ACC_COEF_D=(short)val;  goto rvbd;

-    case H_Reverb+14  : spu.rvb->IIR_COEF=(short)val;    goto rvbd;

-    case H_Reverb+16  : spu.rvb->FB_ALPHA=(short)val;    goto rvbd;

-    case H_Reverb+18  : spu.rvb->FB_X=(short)val;        goto rvbd;

-    case H_Reverb+20  : spu.rvb->IIR_DEST_A0=val*4;      goto rvbd;

-    case H_Reverb+22  : spu.rvb->IIR_DEST_A1=val*4;      goto rvbd;

-    case H_Reverb+24  : spu.rvb->ACC_SRC_A0=val*4;       goto rvbd;

-    case H_Reverb+26  : spu.rvb->ACC_SRC_A1=val*4;       goto rvbd;

-    case H_Reverb+28  : spu.rvb->ACC_SRC_B0=val*4;       goto rvbd;

-    case H_Reverb+30  : spu.rvb->ACC_SRC_B1=val*4;       goto rvbd;

-    case H_Reverb+32  : spu.rvb->IIR_SRC_A0=val*4;       goto rvbd;

-    case H_Reverb+34  : spu.rvb->IIR_SRC_A1=val*4;       goto rvbd;

-    case H_Reverb+36  : spu.rvb->IIR_DEST_B0=val*4;      goto rvbd;

-    case H_Reverb+38  : spu.rvb->IIR_DEST_B1=val*4;      goto rvbd;

-    case H_Reverb+40  : spu.rvb->ACC_SRC_C0=val*4;       goto rvbd;

-    case H_Reverb+42  : spu.rvb->ACC_SRC_C1=val*4;       goto rvbd;

-    case H_Reverb+44  : spu.rvb->ACC_SRC_D0=val*4;       goto rvbd;

-    case H_Reverb+46  : spu.rvb->ACC_SRC_D1=val*4;       goto rvbd;

-    case H_Reverb+48  : spu.rvb->IIR_SRC_B1=val*4;       goto rvbd;

-    case H_Reverb+50  : spu.rvb->IIR_SRC_B0=val*4;       goto rvbd;

-    case H_Reverb+52  : spu.rvb->MIX_DEST_A0=val*4;      goto rvbd;

-    case H_Reverb+54  : spu.rvb->MIX_DEST_A1=val*4;      goto rvbd;

-    case H_Reverb+56  : spu.rvb->MIX_DEST_B0=val*4;      goto rvbd;

-    case H_Reverb+58  : spu.rvb->MIX_DEST_B1=val*4;      goto rvbd;

-    case H_Reverb+60  : spu.rvb->IN_COEF_L=(short)val;   goto rvbd;

-    case H_Reverb+62  : spu.rvb->IN_COEF_R=(short)val;   goto rvbd;

+    case H_Reverb+0   : goto rvbd;

+    case H_Reverb+2   : goto rvbd;

+    case H_Reverb+4   : spu.rvb->IIR_ALPHA=(short)val;   break;

+    case H_Reverb+6   : spu.rvb->ACC_COEF_A=(short)val;  break;

+    case H_Reverb+8   : spu.rvb->ACC_COEF_B=(short)val;  break;

+    case H_Reverb+10  : spu.rvb->ACC_COEF_C=(short)val;  break;

+    case H_Reverb+12  : spu.rvb->ACC_COEF_D=(short)val;  break;

+    case H_Reverb+14  : spu.rvb->IIR_COEF=(short)val;    break;

+    case H_Reverb+16  : spu.rvb->FB_ALPHA=(short)val;    break;

+    case H_Reverb+18  : spu.rvb->FB_X=(short)val;        break;

+    case H_Reverb+20  : goto rvbd;

+    case H_Reverb+22  : goto rvbd;

+    case H_Reverb+24  : goto rvbd;

+    case H_Reverb+26  : goto rvbd;

+    case H_Reverb+28  : goto rvbd;

+    case H_Reverb+30  : goto rvbd;

+    case H_Reverb+32  : goto rvbd;

+    case H_Reverb+34  : goto rvbd;

+    case H_Reverb+36  : goto rvbd;

+    case H_Reverb+38  : goto rvbd;

+    case H_Reverb+40  : goto rvbd;

+    case H_Reverb+42  : goto rvbd;

+    case H_Reverb+44  : goto rvbd;

+    case H_Reverb+46  : goto rvbd;

+    case H_Reverb+48  : goto rvbd;

+    case H_Reverb+50  : goto rvbd;

+    case H_Reverb+52  : goto rvbd;

+    case H_Reverb+54  : goto rvbd;

+    case H_Reverb+56  : goto rvbd;

+    case H_Reverb+58  : goto rvbd;

+    case H_Reverb+60  : spu.rvb->IN_COEF_L=(short)val;   break;

+    case H_Reverb+62  : spu.rvb->IN_COEF_R=(short)val;   break;

    }

  return;

 

diff --git a/plugins/dfsound/reverb.c b/plugins/dfsound/reverb.c
index 2ff6edc..ec570fb 100644
--- a/plugins/dfsound/reverb.c
+++ b/plugins/dfsound/reverb.c
@@ -50,16 +50,16 @@ INLINE int rvb2ram_offs(int curr, int space, int iOff)
 

 // get_buffer content helper: takes care about wraps

 #define g_buffer(var) \

- ((int)(signed short)spu.spuMem[rvb2ram_offs(curr_addr, space, rvb->n##var)])

+ ((int)(signed short)spu.spuMem[rvb2ram_offs(curr_addr, space, rvb->var)])

 

 // saturate iVal and store it as var

 #define s_buffer(var, iVal) \

  ssat32_to_16(iVal); \

- spu.spuMem[rvb2ram_offs(curr_addr, space, rvb->n##var)] = iVal

+ spu.spuMem[rvb2ram_offs(curr_addr, space, rvb->var)] = iVal

 

 #define s_buffer1(var, iVal) \

  ssat32_to_16(iVal); \

- spu.spuMem[rvb2ram_offs(curr_addr, space, rvb->n##var + 1)] = iVal

+ spu.spuMem[rvb2ram_offs(curr_addr, space, rvb->var + 1)] = iVal

 

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

 

@@ -174,43 +174,54 @@ static void MixREVERB_off(int *SSumLR, int ns_to, int curr_addr)
 static void REVERBPrep(void)

 {

  REVERBInfo *rvb = spu.rvb;

- int space = 0x40000 - rvb->StartAddr;

- int t;

- #define prep_offs(v) \

-   t = rvb->v; \

+ int space, t;

+

+ t = spu.regArea[(H_SPUReverbAddr - 0xc00) >> 1];

+ if (t == 0xFFFF || t <= 0x200)

+  spu.rvb->StartAddr = spu.rvb->CurrAddr = 0;

+ else if (spu.rvb->StartAddr != (t << 2))

+  spu.rvb->StartAddr = spu.rvb->CurrAddr = t << 2;

+

+ space = 0x40000 - rvb->StartAddr;

+

+ #define prep_offs(v, r) \

+   t = spu.regArea[(0x1c0 + r) >> 1] * 4; \

    while (t >= space) \

      t -= space; \

-   rvb->n##v = t

- #define prep_offs2(d, v1, v2) \

-   t = rvb->v1 - rvb->v2; \

+   rvb->v = t

+ #define prep_offs2(d, r1, r2) \

+   t = spu.regArea[(0x1c0 + r1) >> 1] * 4; \

+   t -= spu.regArea[(0x1c0 + r2) >> 1] * 4; \

+   while (t < 0) \

+     t += space; \

    while (t >= space) \

      t -= space; \

-   rvb->n##d = t

-

- prep_offs(IIR_SRC_A0);

- prep_offs(IIR_SRC_A1);

- prep_offs(IIR_SRC_B0);

- prep_offs(IIR_SRC_B1);

- prep_offs(IIR_DEST_A0);

- prep_offs(IIR_DEST_A1);

- prep_offs(IIR_DEST_B0);

- prep_offs(IIR_DEST_B1);

- prep_offs(ACC_SRC_A0);

- prep_offs(ACC_SRC_A1);

- prep_offs(ACC_SRC_B0);

- prep_offs(ACC_SRC_B1);

- prep_offs(ACC_SRC_C0);

- prep_offs(ACC_SRC_C1);

- prep_offs(ACC_SRC_D0);

- prep_offs(ACC_SRC_D1);

- prep_offs(MIX_DEST_A0);

- prep_offs(MIX_DEST_A1);

- prep_offs(MIX_DEST_B0);

- prep_offs(MIX_DEST_B1);

- prep_offs2(FB_SRC_A0, MIX_DEST_A0, FB_SRC_A);

- prep_offs2(FB_SRC_A1, MIX_DEST_A1, FB_SRC_A);

- prep_offs2(FB_SRC_B0, MIX_DEST_B0, FB_SRC_B);

- prep_offs2(FB_SRC_B1, MIX_DEST_B1, FB_SRC_B);

+   rvb->d = t

+

+ prep_offs(IIR_SRC_A0, 32);

+ prep_offs(IIR_SRC_A1, 34);

+ prep_offs(IIR_SRC_B0, 36);

+ prep_offs(IIR_SRC_B1, 38);

+ prep_offs(IIR_DEST_A0, 20);

+ prep_offs(IIR_DEST_A1, 22);

+ prep_offs(IIR_DEST_B0, 36);

+ prep_offs(IIR_DEST_B1, 38);

+ prep_offs(ACC_SRC_A0, 24);

+ prep_offs(ACC_SRC_A1, 26);

+ prep_offs(ACC_SRC_B0, 28);

+ prep_offs(ACC_SRC_B1, 30);

+ prep_offs(ACC_SRC_C0, 40);

+ prep_offs(ACC_SRC_C1, 42);

+ prep_offs(ACC_SRC_D0, 44);

+ prep_offs(ACC_SRC_D1, 46);

+ prep_offs(MIX_DEST_A0, 52);

+ prep_offs(MIX_DEST_A1, 54);

+ prep_offs(MIX_DEST_B0, 56);

+ prep_offs(MIX_DEST_B1, 58);

+ prep_offs2(FB_SRC_A0, 52, 0);

+ prep_offs2(FB_SRC_A1, 54, 0);

+ prep_offs2(FB_SRC_B0, 56, 2);

+ prep_offs2(FB_SRC_B1, 58, 2);

 

 #undef prep_offs

 #undef prep_offs2

diff --git a/plugins/dfsound/spu.c b/plugins/dfsound/spu.c
index 8681d35..193a5d0 100644
--- a/plugins/dfsound/spu.c
+++ b/plugins/dfsound/spu.c
@@ -818,12 +818,8 @@ static void do_channels(int ns_to)
   }
 
   if (spu.rvb->StartAddr) {
-   if (do_rvb) {
-    if (unlikely(spu.rvb->dirty))
-     REVERBPrep();
-
+   if (do_rvb)
     REVERBDo(spu.SSumLR, RVB, ns_to, spu.rvb->CurrAddr);
-   }
 
    spu.rvb->CurrAddr += ns_to / 2;
    while (spu.rvb->CurrAddr >= 0x40000)
@@ -970,11 +966,8 @@ static void queue_channel_work(int ns_to, unsigned int silentch)
 
  work->rvb_addr = 0;
  if (spu.rvb->StartAddr) {
-  if (spu_config.iUseReverb) {
-   if (unlikely(spu.rvb->dirty))
-    REVERBPrep();
+  if (spu_config.iUseReverb)
    work->rvb_addr = spu.rvb->CurrAddr;
-  }
 
   spu.rvb->CurrAddr += ns_to / 2;
   while (spu.rvb->CurrAddr >= 0x40000)
@@ -989,9 +982,9 @@ static void do_channel_work(struct work_item *work)
 {
  unsigned int mask;
  unsigned int decode_dirty_ch = 0;
+ const SPUCHAN *s_chan;
  int *SB, sinc, spos, sbpos;
  int d, ch, ns_to;
- SPUCHAN *s_chan;
 
  ns_to = work->ns_to;
 
@@ -1057,6 +1050,9 @@ static void sync_worker_thread(int force)
  struct work_item *work;
  int done, used_space;
 
+ // rvb offsets will change, thread may be using them
+ force |= spu.rvb->dirty && spu.rvb->StartAddr;
+
  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,
@@ -1150,6 +1146,9 @@ void do_samples(unsigned int cycles_to, int do_direct)
      }
    }
 
+  if (unlikely(spu.rvb->dirty))
+   REVERBPrep();
+
   if (do_direct || worker == NULL || !spu_config.iUseThread) {
    do_channels(ns_to);
    do_samples_finish(spu.SSumLR, ns_to, silentch, spu.decode_pos);