Merge pull request #5997 from leoetlino/red-steel

DSP: Fix missing masking for accelerator registers

Source/Core/Core/DSP/DSPAccelerator.cpp

@@ -9,6 +9,7 @@
 #include "Common/MathUtil.h"
 
 #include "Core/DSP/DSPCore.h"
+#include "Core/DSP/DSPHWInterface.h"
 #include "Core/DSP/DSPHost.h"
 
 namespace DSP
@@ -177,8 +178,8 @@ u16 dsp_read_accelerator()
     DSPCore_SetException(EXP_ACCOV);
   }
 
-  g_dsp.ifx_regs[DSP_ACCAH] = Address >> 16;
+  gdsp_ifx_write(DSP_ACCAH, Address >> 16);
-  g_dsp.ifx_regs[DSP_ACCAL] = Address & 0xffff;
+  gdsp_ifx_write(DSP_ACCAL, Address & 0xffff);
   return val;
 }
 }  // namespace DSP

Source/Core/Core/DSP/DSPHWInterface.cpp

@@ -156,6 +156,18 @@ void gdsp_ifx_write(u32 addr, u32 val)
         dsp_step_accelerator();
         break;
   */
+
+  // Masking occurs for the start and end addresses as soon as the registers are written to.
+  case DSP_ACSAH:
+  case DSP_ACEAH:
+    g_dsp.ifx_regs[addr & 0xff] = val & 0x3fff;
+    break;
+
+  // This also happens for the current address, but with a different mask.
+  case DSP_ACCAH:
+    g_dsp.ifx_regs[addr & 0xff] = val & 0xbfff;
+    break;
+
   default:
     if ((addr & 0xff) >= 0xa0)
     {

Source/Core/Core/HW/DSPHLE/UCodes/AXVoice.h

@@ -171,9 +171,12 @@ static bool acc_end_reached;
 void AcceleratorSetup(PB_TYPE* pb, u32* cur_addr)
 {
   acc_pb = pb;
-  acc_loop_addr = HILO_TO_32(pb->audio_addr.loop_addr);
+  // Masking occurs for the start and end addresses as soon as the registers are written to.
-  acc_end_addr = HILO_TO_32(pb->audio_addr.end_addr);
+  acc_loop_addr = HILO_TO_32(pb->audio_addr.loop_addr) & 0x3fffffff;
+  acc_end_addr = HILO_TO_32(pb->audio_addr.end_addr) & 0x3fffffff;
   acc_cur_addr = cur_addr;
+  // It also happens for the current address, but with a different mask.
+  *acc_cur_addr &= 0xbfffffff;
   acc_end_reached = false;
 }
 
@@ -456,8 +459,8 @@ void GetInputSamples(PB_TYPE& pb, s16* samples, u16 count, const s16* coeffs)
   pb.src.cur_addr_frac = (curr_pos & 0xFFFF);
 
   // Update current position in the PB.
-  pb.audio_addr.cur_addr_hi = (u16)(cur_addr >> 16);
+  pb.audio_addr.cur_addr_hi = static_cast<u16>(cur_addr >> 16) & 0xbfff;
-  pb.audio_addr.cur_addr_lo = (u16)(cur_addr & 0xFFFF);
+  pb.audio_addr.cur_addr_lo = static_cast<u16>(cur_addr);
 }
 
 // Add samples to an output buffer, with optional volume ramping.

Source/DSPSpy/tests/accelerator_test.ds

@@ -0,0 +1,153 @@
+; Test for the DSP accelerator, in particular to check its masking behaviour.
+; See https://github.com/dolphin-emu/dolphin/pull/5997
+incdir  "tests"
+include "dsp_base.inc"
+jmp test_main
+
+; Writes the passed format, start and end addresses to the accelerator registers,
+; then reads them back to registers.
+; The current address is initialised to the start.
+; Parameters:
+;   AC0.H: sample format
+;   AC0.M/L: start address
+;   AC1.M/L: end address
+test_accelerator_addrs_ex:
+  ; Set the sample format
+  sr @0xffd1, $AC0.H
+
+  ; Set the accelerator start and current address.
+  srs @ACSAH, $AC0.M
+  srs @ACCAH, $AC0.M
+  srs @ACSAL, $AC0.L
+  srs @ACCAL, $AC0.L
+  ; Set the accelerator end address.
+  srs @ACEAH, $AC1.M
+  srs @ACEAL, $AC1.L
+
+  ; Move the values back to registers that can be printed by dspspy.
+  ; AC0 -> start, AC1 -> end, AX0 -> current
+  lri $AC0.H, #0
+  lrs $AC0.M, @ACSAH
+  lrs $AC0.L, @ACSAL
+  lri $AC1.H, #0
+  lrs $AC1.M, @ACEAH
+  lrs $AC1.L, @ACEAL
+  lrs $AX0.H, @ACCAH
+  lrs $AX0.L, @ACCAL
+
+  ; Make the accelerator read memory
+  lrs $AX1.H, @ARAM
+  lrs $AX1.H, @ARAM
+  ; AX1 -> new current position after read
+  lrs $AX1.H, @ACCAH
+  lrs $AX1.L, @ACCAL
+
+  call send_back
+  ret
+
+; Same as test_accelerator_addrs_ex, but with the end address set to start + 0x1000.
+test_accelerator_addrs:
+  lri $AC1.H, #0
+  lri $AC1.M, #0
+  lri $AC1.L, #0x1000
+  add $ACC1, $ACC0
+  jmp test_accelerator_addrs_ex
+
+test_main:
+  ; Test 1
+  lri $AC0.H, #0x19 ; 8-bit PCM
+  lri $AC0.M, #0x0000 ; start
+  lri $AC0.L, #0x1000 ; start
+  call test_accelerator_addrs
+
+  ; Test 2
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x0000 ; start
+  lri $AC0.L, #0x1000 ; start
+  call test_accelerator_addrs
+
+  ; Test 3
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x1000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 4
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x2000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 5
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x3000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 6
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x3fff ; start
+  lri $AC0.L, #0xffff ; start
+  call test_accelerator_addrs
+
+  ; Test 7
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x4000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 8
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x5000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 9
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x6000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 10
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x7000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 11
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x7fff ; start
+  lri $AC0.L, #0xffff ; start
+  call test_accelerator_addrs
+
+  ; Test 12
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x8000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 13
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0x9000 ; start
+  lri $AC0.L, #0x0000 ; start
+  call test_accelerator_addrs
+
+  ; Test 14
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0xc000 ; start
+  lri $AC0.L, #0x1000 ; start
+  call test_accelerator_addrs
+
+  ; Test 15
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0xd000 ; start
+  lri $AC0.L, #0x1000 ; start
+  call test_accelerator_addrs
+
+  ; Test 16
+  lri $AC0.H, #0xa ; 16-bit PCM
+  lri $AC0.M, #0xffff ; start
+  lri $AC0.L, #0xffff ; start
+  call test_accelerator_addrs
+
+  jmp end_of_test