Further work on SPU implementation

src/core/host_interface.cpp

@@ -1,6 +1,7 @@
 #include "host_interface.h"
 #include "YBaseLib/ByteStream.h"
 #include "YBaseLib/Log.h"
+#include "common/audio_stream.h"
 #include "system.h"
 Log_SetChannel(HostInterface);
 

src/core/host_interface.h

@@ -2,6 +2,8 @@
 #include "types.h"
 #include <memory>
 
+class AudioStream;
+
 namespace GL {
 class Texture;
 }
@@ -14,6 +16,8 @@ public:
   HostInterface();
   virtual ~HostInterface();
 
+  AudioStream* GetAudioStream() const { return m_audio_stream.get(); }
+
   bool InitializeSystem(const char* filename, const char* exp1_filename);
 
   virtual void SetDisplayTexture(GL::Texture* texture, u32 offset_x, u32 offset_y, u32 width, u32 height, float aspect_ratio) = 0;
@@ -26,6 +30,8 @@ public:
   bool SaveState(const char* filename);
 
 protected:
+  std::unique_ptr<AudioStream> m_audio_stream;
+
   std::unique_ptr<System> m_system;
   bool m_running = false;
 };

src/core/spu.cpp

@@ -1,17 +1,26 @@
 #include "spu.h"
 #include "YBaseLib/Log.h"
+#include "common/audio_stream.h"
 #include "common/state_wrapper.h"
 #include "dma.h"
+#include "host_interface.h"
 #include "interrupt_controller.h"
 #include "system.h"
+#include <imgui.h>
 Log_SetChannel(SPU);
 
+static s16 Clamp16(s32 value)
+{
+  return static_cast<s16>(std::clamp<s32>(value, -32768, 32767));
+}
+
 SPU::SPU() = default;
 
 SPU::~SPU() = default;
 
 bool SPU::Initialize(System* system, DMA* dma, InterruptController* interrupt_controller)
 {
+  m_audio_stream = system->GetHostInterface()->GetAudioStream();
   m_system = system;
   m_dma = dma;
   m_interrupt_controller = interrupt_controller;
@@ -37,6 +46,9 @@ bool SPU::DoState(StateWrapper& sw)
 
 u16 SPU::ReadRegister(u32 offset)
 {
+  if (offset < (0x1F801D80 - SPU_BASE))
+    return ReadVoiceRegister(offset);
+
   switch (offset)
   {
     case 0x1F801DA6 - SPU_BASE:
@@ -52,9 +64,21 @@ u16 SPU::ReadRegister(u32 offset)
       return m_SPUCNT.bits;
 
     case 0x1F801DAE - SPU_BASE:
-      Log_DebugPrintf("SPU status register -> 0x%04X", ZeroExtend32(m_SPUCNT.bits));
+      // Log_DebugPrintf("SPU status register -> 0x%04X", ZeroExtend32(m_SPUCNT.bits));
       return m_SPUSTAT.bits;
 
+    case 0x1F801D88 - SPU_BASE:
+      return Truncate16(m_key_on_register);
+
+    case 0x1F801D8A - SPU_BASE:
+      return Truncate16(m_key_on_register >> 16);
+
+    case 0x1F801D8C - SPU_BASE:
+      return Truncate16(m_key_off_register);
+
+    case 0x1F801D8E - SPU_BASE:
+      return Truncate16(m_key_off_register >> 16);
+
     default:
       Log_ErrorPrintf("Unknown SPU register read: offset 0x%X (address 0x%08X)", offset, offset | SPU_BASE);
       return UINT16_C(0xFFFF);
@@ -63,13 +87,19 @@ u16 SPU::ReadRegister(u32 offset)
 
 void SPU::WriteRegister(u32 offset, u16 value)
 {
+  if (offset < (0x1F801D80 - SPU_BASE))
+  {
+    WriteVoiceRegister(offset, value);
+    return;
+  }
+
   switch (offset)
   {
     case 0x1F801DA6 - SPU_BASE:
     {
       Log_DebugPrintf("SPU transfer address register <- 0x%04X", ZeroExtend32(value));
       m_transfer_address_reg = value;
-      m_transfer_address = ZeroExtend32(value) * 8;
+      m_transfer_address = (ZeroExtend32(value) << VOICE_ADDRESS_SHIFT) & RAM_MASK;
       return;
     }
 
@@ -89,6 +119,82 @@ void SPU::WriteRegister(u32 offset, u16 value)
       return;
     }
 
+    case 0x1F801D88 - SPU_BASE:
+    {
+      Log_DebugPrintf("SPU key on low <- 0x%04X", ZeroExtend32(value));
+      m_system->Synchronize();
+      m_key_on_register = (m_key_on_register & 0xFFFF0000) | ZeroExtend32(value);
+
+      u16 bits = value;
+      for (u32 i = 0; i < 16; i++)
+      {
+        if (bits & 0x01)
+        {
+          Log_DebugPrintf("Voice %u key on", i);
+          m_voices[i].KeyOn();
+        }
+        bits >>= 1;
+      }
+    }
+    break;
+
+    case 0x1F801D8A - SPU_BASE:
+    {
+      Log_DebugPrintf("SPU key on high <- 0x%04X", ZeroExtend32(value));
+      m_system->Synchronize();
+      m_key_on_register = (m_key_on_register & 0x0000FFFF) | (ZeroExtend32(value) << 16);
+
+      u16 bits = value;
+      for (u32 i = 16; i < NUM_VOICES; i++)
+      {
+        if (bits & 0x01)
+        {
+          Log_DebugPrintf("Voice %u key on", i);
+          m_voices[i].KeyOn();
+        }
+        bits >>= 1;
+      }
+    }
+    break;
+
+    case 0x1F801D8C - SPU_BASE:
+    {
+      Log_DebugPrintf("SPU key off low <- 0x%04X", ZeroExtend32(value));
+      m_system->Synchronize();
+      m_key_on_register = (m_key_on_register & 0xFFFF0000) | ZeroExtend32(value);
+
+      u16 bits = value;
+      for (u32 i = 0; i < 16; i++)
+      {
+        if (bits & 0x01)
+        {
+          Log_DebugPrintf("Voice %u key off", i);
+          m_voices[i].KeyOff();
+        }
+        bits >>= 1;
+      }
+    }
+    break;
+
+    case 0x1F801D8E - SPU_BASE:
+    {
+      Log_DebugPrintf("SPU key off high <- 0x%04X", ZeroExtend32(value));
+      m_system->Synchronize();
+      m_key_on_register = (m_key_on_register & 0x0000FFFF) | (ZeroExtend32(value) << 16);
+
+      u16 bits = value;
+      for (u32 i = 16; i < NUM_VOICES; i++)
+      {
+        if (bits & 0x01)
+        {
+          Log_DebugPrintf("Voice %u key off", i);
+          m_voices[i].KeyOff();
+        }
+        bits >>= 1;
+      }
+    }
+    break;
+
       // read-only registers
     case 0x1F801DAE - SPU_BASE:
     {
@@ -104,6 +210,89 @@ void SPU::WriteRegister(u32 offset, u16 value)
   }
 }
 
+u16 SPU::ReadVoiceRegister(u32 offset)
+{
+  const u32 reg_index = (offset & 0x0F) / 2;
+  const u32 voice_index = ((offset >> 4) & 0x1F);
+  Assert(voice_index < 24);
+
+  return m_voices[voice_index].regs.index[reg_index];
+}
+
+void SPU::WriteVoiceRegister(u32 offset, u16 value)
+{
+  // per-voice registers
+  const u32 reg_index = (offset & 0x0F);
+  const u32 voice_index = ((offset >> 4) & 0x1F);
+  Assert(voice_index < 24);
+
+  switch (reg_index)
+  {
+    case 0x00: // volume left
+    {
+      Log_DebugPrintf("SPU voice %u volume left <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.volume_left.bits = value;
+    }
+    break;
+
+    case 0x02: // volume right
+    {
+      Log_DebugPrintf("SPU voice %u volume right <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.volume_right.bits = value;
+    }
+    break;
+
+    case 0x04: // sample rate
+    {
+      Log_DebugPrintf("SPU voice %u ADPCM sample rate <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.adpcm_sample_rate = value;
+    }
+    break;
+
+    case 0x06: // start address
+    {
+      Log_DebugPrintf("SPU voice %u ADPCM start address <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.adpcm_start_address = value;
+    }
+    break;
+
+    case 0x08: // adsr low
+    {
+      Log_WarningPrintf("SPU voice %u ADSR low <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.adsr.bits_low = value;
+    }
+    break;
+
+    case 0x0A: // adsr high
+    {
+      Log_WarningPrintf("SPU voice %u ADSR high <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.adsr.bits_high = value;
+    }
+    break;
+
+    case 0x0C: // adsr volume
+    {
+      Log_DebugPrintf("SPU voice %u ADSR volume <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.adsr_volume = value;
+    }
+    break;
+
+    case 0x0E: // repeat address
+    {
+      Log_DebugPrintf("SPU voice %u ADPCM repeat address <- 0x%04X", voice_index, value);
+      m_voices[voice_index].regs.adpcm_repeat_address = value;
+    }
+    break;
+
+    default:
+    {
+      Log_ErrorPrintf("Unknown SPU voice %u register write: offset 0x%X (address 0x%08X) value 0x%04X", offset,
+                      voice_index, offset | SPU_BASE, ZeroExtend32(value));
+    }
+    break;
+  }
+}
+
 u32 SPU::DMARead()
 {
   const u16 lsb = RAMTransferRead();
@@ -135,6 +324,333 @@ u16 SPU::RAMTransferRead()
 
 void SPU::RAMTransferWrite(u16 value)
 {
+  Log_TracePrintf("SPU RAM @ 0x%08X (voice 0x%04X) <- 0x%04X", m_transfer_address,
+                  m_transfer_address >> VOICE_ADDRESS_SHIFT, ZeroExtend32(value));
   std::memcpy(&m_ram[m_transfer_address], &value, sizeof(value));
   m_transfer_address = (m_transfer_address + sizeof(value)) & RAM_MASK;
 }
+
+void SPU::Execute(TickCount ticks)
+{
+  TickCount num_samples = (ticks + m_ticks_carry) / SYSCLK_TICKS_PER_SPU_TICK;
+  m_ticks_carry = (ticks + m_ticks_carry) % SYSCLK_TICKS_PER_SPU_TICK;
+  if (num_samples == 0 || !m_SPUCNT.enable)
+    return;
+
+  for (TickCount i = 0; i < num_samples; i++)
+    GenerateSample();
+}
+
+void SPU::Voice::KeyOn()
+{
+  current_address = regs.adpcm_start_address;
+  has_samples = false;
+  key_on = true;
+}
+
+void SPU::Voice::KeyOff()
+{
+  has_samples = false;
+  key_on = false;
+}
+
+void SPU::Voice::DecodeBlock()
+{
+  previous_block_last_samples[2] = current_block_samples[NUM_SAMPLES_PER_ADPCM_BLOCK - 1];
+  previous_block_last_samples[1] = current_block_samples[NUM_SAMPLES_PER_ADPCM_BLOCK - 2];
+  previous_block_last_samples[0] = current_block_samples[NUM_SAMPLES_PER_ADPCM_BLOCK - 3];
+  DecodeADPCMBlock(current_block, current_block_samples.data(), adpcm_state.data());
+}
+
+SPU::SampleFormat SPU::Voice::SampleBlock(s32 index) const
+{
+  if (index < 0)
+  {
+    DebugAssert(index >= -3);
+    return previous_block_last_samples[index + 3];
+  }
+
+  return current_block_samples[index];
+}
+
+s32 SPU::Voice::Interpolate() const
+{
+  static constexpr std::array<s32, 0x200> gauss = {{
+    -0x001, -0x001, -0x001, -0x001, -0x001, -0x001, -0x001, -0x001, //
+    -0x001, -0x001, -0x001, -0x001, -0x001, -0x001, -0x001, -0x001, //
+    0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0001, //
+    0x0001, 0x0001, 0x0001, 0x0002, 0x0002, 0x0002, 0x0003, 0x0003, //
+    0x0003, 0x0004, 0x0004, 0x0005, 0x0005, 0x0006, 0x0007, 0x0007, //
+    0x0008, 0x0009, 0x0009, 0x000A, 0x000B, 0x000C, 0x000D, 0x000E, //
+    0x000F, 0x0010, 0x0011, 0x0012, 0x0013, 0x0015, 0x0016, 0x0018, // entry
+    0x0019, 0x001B, 0x001C, 0x001E, 0x0020, 0x0021, 0x0023, 0x0025, // 000..07F
+    0x0027, 0x0029, 0x002C, 0x002E, 0x0030, 0x0033, 0x0035, 0x0038, //
+    0x003A, 0x003D, 0x0040, 0x0043, 0x0046, 0x0049, 0x004D, 0x0050, //
+    0x0054, 0x0057, 0x005B, 0x005F, 0x0063, 0x0067, 0x006B, 0x006F, //
+    0x0074, 0x0078, 0x007D, 0x0082, 0x0087, 0x008C, 0x0091, 0x0096, //
+    0x009C, 0x00A1, 0x00A7, 0x00AD, 0x00B3, 0x00BA, 0x00C0, 0x00C7, //
+    0x00CD, 0x00D4, 0x00DB, 0x00E3, 0x00EA, 0x00F2, 0x00FA, 0x0101, //
+    0x010A, 0x0112, 0x011B, 0x0123, 0x012C, 0x0135, 0x013F, 0x0148, //
+    0x0152, 0x015C, 0x0166, 0x0171, 0x017B, 0x0186, 0x0191, 0x019C, //
+    0x01A8, 0x01B4, 0x01C0, 0x01CC, 0x01D9, 0x01E5, 0x01F2, 0x0200, //
+    0x020D, 0x021B, 0x0229, 0x0237, 0x0246, 0x0255, 0x0264, 0x0273, //
+    0x0283, 0x0293, 0x02A3, 0x02B4, 0x02C4, 0x02D6, 0x02E7, 0x02F9, //
+    0x030B, 0x031D, 0x0330, 0x0343, 0x0356, 0x036A, 0x037E, 0x0392, //
+    0x03A7, 0x03BC, 0x03D1, 0x03E7, 0x03FC, 0x0413, 0x042A, 0x0441, //
+    0x0458, 0x0470, 0x0488, 0x04A0, 0x04B9, 0x04D2, 0x04EC, 0x0506, //
+    0x0520, 0x053B, 0x0556, 0x0572, 0x058E, 0x05AA, 0x05C7, 0x05E4, // entry
+    0x0601, 0x061F, 0x063E, 0x065C, 0x067C, 0x069B, 0x06BB, 0x06DC, // 080..0FF
+    0x06FD, 0x071E, 0x0740, 0x0762, 0x0784, 0x07A7, 0x07CB, 0x07EF, //
+    0x0813, 0x0838, 0x085D, 0x0883, 0x08A9, 0x08D0, 0x08F7, 0x091E, //
+    0x0946, 0x096F, 0x0998, 0x09C1, 0x09EB, 0x0A16, 0x0A40, 0x0A6C, //
+    0x0A98, 0x0AC4, 0x0AF1, 0x0B1E, 0x0B4C, 0x0B7A, 0x0BA9, 0x0BD8, //
+    0x0C07, 0x0C38, 0x0C68, 0x0C99, 0x0CCB, 0x0CFD, 0x0D30, 0x0D63, //
+    0x0D97, 0x0DCB, 0x0E00, 0x0E35, 0x0E6B, 0x0EA1, 0x0ED7, 0x0F0F, //
+    0x0F46, 0x0F7F, 0x0FB7, 0x0FF1, 0x102A, 0x1065, 0x109F, 0x10DB, //
+    0x1116, 0x1153, 0x118F, 0x11CD, 0x120B, 0x1249, 0x1288, 0x12C7, //
+    0x1307, 0x1347, 0x1388, 0x13C9, 0x140B, 0x144D, 0x1490, 0x14D4, //
+    0x1517, 0x155C, 0x15A0, 0x15E6, 0x162C, 0x1672, 0x16B9, 0x1700, //
+    0x1747, 0x1790, 0x17D8, 0x1821, 0x186B, 0x18B5, 0x1900, 0x194B, //
+    0x1996, 0x19E2, 0x1A2E, 0x1A7B, 0x1AC8, 0x1B16, 0x1B64, 0x1BB3, //
+    0x1C02, 0x1C51, 0x1CA1, 0x1CF1, 0x1D42, 0x1D93, 0x1DE5, 0x1E37, //
+    0x1E89, 0x1EDC, 0x1F2F, 0x1F82, 0x1FD6, 0x202A, 0x207F, 0x20D4, //
+    0x2129, 0x217F, 0x21D5, 0x222C, 0x2282, 0x22DA, 0x2331, 0x2389, // entry
+    0x23E1, 0x2439, 0x2492, 0x24EB, 0x2545, 0x259E, 0x25F8, 0x2653, // 100..17F
+    0x26AD, 0x2708, 0x2763, 0x27BE, 0x281A, 0x2876, 0x28D2, 0x292E, //
+    0x298B, 0x29E7, 0x2A44, 0x2AA1, 0x2AFF, 0x2B5C, 0x2BBA, 0x2C18, //
+    0x2C76, 0x2CD4, 0x2D33, 0x2D91, 0x2DF0, 0x2E4F, 0x2EAE, 0x2F0D, //
+    0x2F6C, 0x2FCC, 0x302B, 0x308B, 0x30EA, 0x314A, 0x31AA, 0x3209, //
+    0x3269, 0x32C9, 0x3329, 0x3389, 0x33E9, 0x3449, 0x34A9, 0x3509, //
+    0x3569, 0x35C9, 0x3629, 0x3689, 0x36E8, 0x3748, 0x37A8, 0x3807, //
+    0x3867, 0x38C6, 0x3926, 0x3985, 0x39E4, 0x3A43, 0x3AA2, 0x3B00, //
+    0x3B5F, 0x3BBD, 0x3C1B, 0x3C79, 0x3CD7, 0x3D35, 0x3D92, 0x3DEF, //
+    0x3E4C, 0x3EA9, 0x3F05, 0x3F62, 0x3FBD, 0x4019, 0x4074, 0x40D0, //
+    0x412A, 0x4185, 0x41DF, 0x4239, 0x4292, 0x42EB, 0x4344, 0x439C, //
+    0x43F4, 0x444C, 0x44A3, 0x44FA, 0x4550, 0x45A6, 0x45FC, 0x4651, //
+    0x46A6, 0x46FA, 0x474E, 0x47A1, 0x47F4, 0x4846, 0x4898, 0x48E9, //
+    0x493A, 0x498A, 0x49D9, 0x4A29, 0x4A77, 0x4AC5, 0x4B13, 0x4B5F, //
+    0x4BAC, 0x4BF7, 0x4C42, 0x4C8D, 0x4CD7, 0x4D20, 0x4D68, 0x4DB0, //
+    0x4DF7, 0x4E3E, 0x4E84, 0x4EC9, 0x4F0E, 0x4F52, 0x4F95, 0x4FD7, // entry
+    0x5019, 0x505A, 0x509A, 0x50DA, 0x5118, 0x5156, 0x5194, 0x51D0, // 180..1FF
+    0x520C, 0x5247, 0x5281, 0x52BA, 0x52F3, 0x532A, 0x5361, 0x5397, //
+    0x53CC, 0x5401, 0x5434, 0x5467, 0x5499, 0x54CA, 0x54FA, 0x5529, //
+    0x5558, 0x5585, 0x55B2, 0x55DE, 0x5609, 0x5632, 0x565B, 0x5684, //
+    0x56AB, 0x56D1, 0x56F6, 0x571B, 0x573E, 0x5761, 0x5782, 0x57A3, //
+    0x57C3, 0x57E2, 0x57FF, 0x581C, 0x5838, 0x5853, 0x586D, 0x5886, //
+    0x589E, 0x58B5, 0x58CB, 0x58E0, 0x58F4, 0x5907, 0x5919, 0x592A, //
+    0x593A, 0x5949, 0x5958, 0x5965, 0x5971, 0x597C, 0x5986, 0x598F, //
+    0x5997, 0x599E, 0x59A4, 0x59A9, 0x59AD, 0x59B0, 0x59B2, 0x59B3  //
+  }};
+
+  const u8 i = counter.interpolation_index;
+  const s32 s = static_cast<s32>(ZeroExtend32(counter.sample_index.GetValue()));
+
+  s32 out = gauss[0x0FF - i] * s32(SampleBlock(s - 3));
+  out += gauss[0x1FF - i] * s32(SampleBlock(s - 2));
+  out += gauss[0x100 + i] * s32(SampleBlock(s - 1));
+  out += gauss[0x000 + i] * s32(SampleBlock(s - 0));
+  return out;
+}
+
+void SPU::ReadADPCMBlock(u16 address, ADPCMBlock* block)
+{
+  u32 ram_address = (ZeroExtend32(address) * 8) & RAM_MASK;
+
+  // 16 bytes, so 2 8-byte blocks for the interrupt check
+  if (m_SPUCNT.irq9_enable)
+  {
+    if (m_irq_address == address || m_irq_address == (address + 1))
+    {
+      Log_DebugPrintf("SPU IRQ at address 0x%08X", ram_address);
+      m_SPUSTAT.irq9_flag = true;
+      m_interrupt_controller->InterruptRequest(InterruptController::IRQ::SPU);
+    }
+  }
+
+  // fast path - no wrap-around
+  if ((ram_address + sizeof(ADPCMBlock)) <= RAM_SIZE)
+  {
+    std::memcpy(block, &m_ram[ram_address], sizeof(ADPCMBlock));
+    return;
+  }
+
+  block->shift_filter.bits = m_ram[ram_address];
+  ram_address = (ram_address + 1) & RAM_MASK;
+  block->flags.bits = m_ram[ram_address];
+  ram_address = (ram_address + 1) & RAM_MASK;
+  for (u32 i = 0; i < 14; i++)
+  {
+    block->data[i] = m_ram[ram_address];
+    ram_address = (ram_address + 1) & RAM_MASK;
+  }
+}
+
+void SPU::DecodeADPCMBlock(const ADPCMBlock& block, SampleFormat out_samples[NUM_SAMPLES_PER_ADPCM_BLOCK], s32 state[2])
+{
+  static constexpr std::array<s32, 5> filter_table_pos = {{0, 60, 115, 98, 122}};
+  static constexpr std::array<s32, 5> filter_table_neg = {{0, 0, -52, -55, -60}};
+
+  // pre-lookup
+  const u8 shift = block.shift_filter.shift;
+  const u8 filter_index = std::min<u8>(block.shift_filter.filter, 4);
+  const s32 filter_pos = filter_table_pos[filter_index];
+  const s32 filter_neg = filter_table_neg[filter_index];
+  s32 last_samples[2] = {state[0], state[1]};
+
+  // samples
+  for (u32 i = 0; i < NUM_SAMPLES_PER_ADPCM_BLOCK; i++)
+  {
+    const u8 nibble = (block.data[i / 2] >> (4 * (i % 2))) & 0x0F;
+    s32 sample = SignExtendN<4, s32>(static_cast<s32>(ZeroExtend32(nibble)));
+    sample >>= shift;
+
+    sample += ((last_samples[0] * filter_pos) + (last_samples[1] * filter_neg) + 32) / 64;
+
+    out_samples[i] =
+      static_cast<s16>(std::clamp<s32>(sample, std::numeric_limits<s16>::min(), std::numeric_limits<s16>::max()));
+
+    state[1] = state[0];
+    state[0] = sample;
+  }
+}
+
+std::tuple<SPU::SampleFormat, SPU::SampleFormat> SPU::SampleVoice(u32 voice_index)
+{
+  Voice& voice = m_voices[voice_index];
+  if (!voice.key_on)
+    return std::make_tuple<s16, s16>(0, 0);
+
+  if (!voice.has_samples)
+  {
+    ReadADPCMBlock(voice.current_address, &voice.current_block);
+    voice.DecodeBlock();
+    voice.has_samples = true;
+
+    if (voice.current_block.flags.loop_start)
+    {
+      Log_DebugPrintf("Voice %u loop start @ 0x%08X", voice_index, ZeroExtend32(voice.current_address));
+      voice.regs.adpcm_repeat_address = voice.current_address;
+    }
+  }
+
+  // TODO: Pulse modulation
+  u16 step = voice.regs.adpcm_sample_rate;
+  step = std::min<u16>(step, 0x4000);
+  voice.counter.bits += step;
+
+  if (voice.counter.sample_index >= NUM_SAMPLES_PER_ADPCM_BLOCK)
+  {
+    // next block
+    voice.counter.sample_index -= NUM_SAMPLES_PER_ADPCM_BLOCK;
+    voice.has_samples = false;
+
+    // handle flags
+    if (voice.current_block.flags.loop_end)
+    {
+      if (!voice.current_block.flags.loop_repeat)
+      {
+        Log_DebugPrintf("Voice %u loop end+mute @ 0x%08X", voice_index, ZeroExtend32(voice.current_address));
+        m_endx_register |= (u32(1) << voice_index);
+        voice.KeyOff();
+      }
+      else
+      {
+        Log_DebugPrintf("Voice %u loop end+repeat @ 0x%08X", voice_index, ZeroExtend32(voice.current_address));
+        voice.current_address = voice.regs.adpcm_repeat_address;
+      }
+    }
+    else
+    {
+      voice.current_address += 2;
+    }
+  }
+
+  // TODO: Volume
+  s32 sample = voice.Interpolate();
+  return std::make_tuple(Clamp16(sample), Clamp16(sample));
+}
+
+void SPU::GenerateSample()
+{
+  s32 left_sum = 0;
+  s32 right_sum = 0;
+  for (u32 i = 0; i < NUM_VOICES; i++)
+  {
+    const auto [left, right] = SampleVoice(i);
+    left_sum += left;
+    right_sum += right;
+  }
+
+  Log_DebugPrintf("SPU sample %d %d", left_sum, right_sum);
+  AudioStream::SampleType samples[2] = {Clamp16(left_sum), Clamp16(right_sum)};
+  m_audio_stream->WriteSamples(samples, countof(samples));
+}
+
+void SPU::DrawDebugWindow()
+{
+  if (!m_debug_window_open)
+    return;
+
+  ImGui::SetNextWindowSize(ImVec2(700, 400), ImGuiCond_FirstUseEver);
+  if (!ImGui::Begin("SPU State", &m_debug_window_open))
+  {
+    ImGui::End();
+    return;
+  }
+
+  // draw voice states
+  if (ImGui::CollapsingHeader("Voice State", ImGuiTreeNodeFlags_DefaultOpen))
+  {
+    static constexpr u32 NUM_COLUMNS = 11;
+
+    ImGui::Columns(NUM_COLUMNS);
+
+    // headers
+    static constexpr std::array<const char*, NUM_COLUMNS> column_titles = {
+      {"#", "InterpIndex", "SampleIndex", "CurAddr", "StartAddr", "RepeatAddr", "SampleRate", "VolLeft", "VolRight",
+       "ADSR", "ADSRVol"}};
+    for (u32 i = 0; i < NUM_COLUMNS; i++)
+    {
+      ImGui::TextUnformatted(column_titles[i]);
+      ImGui::NextColumn();
+    }
+
+    // states
+    for (u32 voice_index = 0; voice_index < NUM_VOICES; voice_index++)
+    {
+      const Voice& v = m_voices[voice_index];
+      ImVec4 color = v.key_on ? ImVec4(1.0f, 1.0f, 1.0f, 1.0f) : ImVec4(0.5f, 0.5f, 0.5f, 1.0f);
+      ImGui::TextColored(color, "%u", ZeroExtend32(voice_index));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%u", ZeroExtend32(v.counter.interpolation_index.GetValue()));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%u", ZeroExtend32(v.counter.sample_index.GetValue()));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%04X", ZeroExtend32(v.current_address));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%04X", ZeroExtend32(v.regs.adpcm_start_address));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%04X", ZeroExtend32(v.regs.adpcm_repeat_address));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%.2f", (float(v.regs.adpcm_sample_rate) / 16383.0f) * 44100.0f);
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%04X", ZeroExtend32(v.regs.volume_left.bits));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%04X", ZeroExtend32(v.regs.volume_right.bits));
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%08X", v.regs.adsr.bits);
+      ImGui::NextColumn();
+      ImGui::TextColored(color, "%04X", ZeroExtend32(v.regs.adsr_volume));
+      ImGui::NextColumn();
+    }
+
+    ImGui::Columns(1);
+  }
+
+  ImGui::End();
+}
+
+void SPU::DrawDebugMenu()
+{
+  // TODO: Show RAM, etc.
+}

src/core/spu.h

@@ -3,6 +3,7 @@
 #include "types.h"
 #include <array>
 
+class AudioStream;
 class StateWrapper;
 
 class System;
@@ -12,6 +13,8 @@ class InterruptController;
 class SPU
 {
 public:
+  using SampleFormat = s16;
+
   SPU();
   ~SPU();
 
@@ -25,14 +28,28 @@ public:
   u32 DMARead();
   void DMAWrite(u32 value);
 
+  void Execute(TickCount ticks);
+
+  // Render statistics debug window.
+  void DrawDebugWindow();
+
+  // Manipulating debug options.
+  void DrawDebugMenu();
+
 private:
   static constexpr u32 RAM_SIZE = 512 * 1024;
   static constexpr u32 RAM_MASK = RAM_SIZE - 1;
   static constexpr u32 SPU_BASE = 0x1F801C00;
+  static constexpr u32 NUM_VOICES = 24;
+  static constexpr u32 NUM_VOICE_REGISTERS = 8;
+  static constexpr u32 VOICE_ADDRESS_SHIFT = 3;
+  static constexpr u32 NUM_SAMPLES_PER_ADPCM_BLOCK = 28;
+  static constexpr u32 SAMPLE_RATE = 44100;
+  static constexpr u32 SYSCLK_TICKS_PER_SPU_TICK = MASTER_CLOCK / SAMPLE_RATE; // 0x300
 
   enum class RAMTransferMode : u8
   {
-    Stopped =0,
+    Stopped = 0,
     ManualWrite = 1,
     DMAWrite = 2,
     DMARead = 3
@@ -70,24 +87,133 @@ private:
     BitField<u16, u8, 0, 6> mode;
   };
 
-#if 0
+  union ADSRRegister
+  {
+    u32 bits;
+    struct
+    {
+      u16 bits_low;
+      u16 bits_high;
+    };
+
+    BitField<u32, u8, 0, 4> sustain_level;
+    BitField<u32, u8, 4, 4> decay_shift;
+    BitField<u32, u8, 8, 2> attack_step;
+    BitField<u32, u8, 10, 6> attack_shift;
+    BitField<u32, bool, 15, 1> attack_exponential;
+
+    BitField<u32, u8, 16, 5> release_shift;
+    BitField<u32, bool, 21, 1> release_exponential;
+    BitField<u32, u8, 22, 2> sustain_step;
+    BitField<u32, u8, 24, 5> sustain_shift;
+    BitField<u32, bool, 30, 1> sustain_direction_decrease;
+    BitField<u32, bool, 31, 1> sustain_exponential;
+  };
+
+  union VolumeRegister
+  {
+    u16 bits;
+
+    BitField<u16, bool, 15, 1> sweep_mode;
+    BitField<u16, u16, 0, 15> fixed_volume; // divided by 2
+
+    BitField<u16, bool, 14, 1> sweep_exponential;
+    BitField<u16, bool, 13, 1> sweep_direction_decrease;
+    BitField<u16, bool, 12, 1> sweep_phase_negative;
+    BitField<u16, u8, 2, 5> sweep_shift;
+    BitField<u16, u8, 0, 2> sweep_step;
+  };
+
+  // organized so we can replace this with a u16 array in the future
+  union VoiceRegisters
+  {
+    u16 index[NUM_VOICE_REGISTERS];
+
+    struct
+    {
+      VolumeRegister volume_left;
+      VolumeRegister volume_right;
+
+      u16 adpcm_sample_rate;   // VxPitch
+      u16 adpcm_start_address; // multiply by 8
+
+      ADSRRegister adsr;
+      u16 adsr_volume;
+
+      u16 adpcm_repeat_address; // multiply by 8
+    };
+  };
+
+  union VoiceCounter
+  {
+    // promoted to u32 because of overflow
+    u32 bits;
+
+    BitField<u32, u8, 4, 8> interpolation_index;
+    BitField<u32, u8, 12, 5> sample_index;
+  };
+
+  struct ADPCMBlock
+  {
+    union
+    {
+      u8 bits;
+
+      BitField<u8, u8, 0, 4> shift;
+      BitField<u8, u8, 4, 3> filter;
+    } shift_filter;
+    union
+    {
+      u8 bits;
+
+      BitField<u8, bool, 0, 1> loop_end;
+      BitField<u8, bool, 1, 1> loop_repeat;
+      BitField<u8, bool, 2, 1> loop_start;
+    } flags;
+
+    u8 data[NUM_SAMPLES_PER_ADPCM_BLOCK / 2];
+
+    u8 ReadSample(u32 index) const { return (data[index / 2] >> ((index % 2) * 4)) & 0x0F; }
+  };
+
   struct Voice
   {
-    static constexpr u32 NUM_REGS = 8;
+    u16 current_address;
-    static constexpr u32 NUM_FLAGS = 6;
+    VoiceRegisters regs;
+    VoiceCounter counter;
+    ADPCMBlock current_block;   // TODO Drop this after decoding
+    std::array<SampleFormat, NUM_SAMPLES_PER_ADPCM_BLOCK> current_block_samples;
+    std::array<SampleFormat, 3> previous_block_last_samples;
+    std::array<s32, 2> adpcm_state;
 
-    std::array<u16, NUM_REGS> regs;
+    bool has_samples;
+    bool key_on;
 
+    void KeyOn();
+    void KeyOff();
+
+    void DecodeBlock();
+    SampleFormat SampleBlock(s32 index) const;
+    s32 Interpolate() const;
   };
-#endif
+
+  u16 ReadVoiceRegister(u32 offset);
+  void WriteVoiceRegister(u32 offset, u16 value);
 
   void UpdateDMARequest();
   u16 RAMTransferRead();
   void RAMTransferWrite(u16 value);
 
+  void ReadADPCMBlock(u16 address, ADPCMBlock* block);
+  static void DecodeADPCMBlock(const ADPCMBlock& block, SampleFormat* out_samples, s32* state);
+  std::tuple<SampleFormat, SampleFormat> SampleVoice(u32 voice_index);
+  void GenerateSample();
+
   System* m_system = nullptr;
   DMA* m_dma = nullptr;
   InterruptController* m_interrupt_controller = nullptr;
+  AudioStream* m_audio_stream = nullptr;
+  bool m_debug_window_open = true;
 
   SPUCNT m_SPUCNT = {};
   SPUSTAT m_SPUSTAT = {};
@@ -95,5 +221,14 @@ private:
   u16 m_transfer_address_reg = 0;
   u32 m_transfer_address = 0;
 
+  u16 m_irq_address = 0;
+
+  u32 m_key_on_register = 0;
+  u32 m_key_off_register = 0;
+  u32 m_endx_register = 0;
+
+  TickCount m_ticks_carry = 0;
+
+  std::array<Voice, NUM_VOICES> m_voices{};
   std::array<u8, RAM_SIZE> m_ram{};
 };

src/core/system.cpp

@@ -288,6 +288,7 @@ void System::Synchronize()
   m_cdrom->Execute(pending_ticks);
   m_pad->Execute(pending_ticks);
   m_dma->Execute(pending_ticks);
+  m_spu->Execute(pending_ticks);
 }
 
 void System::SetDowncount(TickCount downcount)

src/core/system.h

@@ -33,6 +33,7 @@ public:
   CPU::Core* GetCPU() const { return m_cpu.get(); }
   Bus* GetBus() const { return m_bus.get(); }
   GPU* GetGPU() const { return m_gpu.get(); }
+  SPU* GetSPU() const { return m_spu.get(); }
 
   u32 GetFrameNumber() const { return m_frame_number; }
   u32 GetInternalFrameNumber() const { return m_internal_frame_number; }

src/duckstation/main.cpp

@@ -91,7 +91,7 @@ int main(int argc, char* argv[])
   // const LOGLEVEL level = LOGLEVEL_DEV;
   // const LOGLEVEL level = LOGLEVEL_PROFILE;
   // g_pLog->SetConsoleOutputParams(true, nullptr, level);
-  g_pLog->SetConsoleOutputParams(true, "Pad SPU", level);
+  g_pLog->SetConsoleOutputParams(true, "Pad", level);
   g_pLog->SetFilterLevel(level);
 #else
   g_pLog->SetConsoleOutputParams(true, nullptr, LOGLEVEL_DEBUG);

src/duckstation/sdl_interface.cpp

@@ -5,8 +5,10 @@
 #include "core/digital_controller.h"
 #include "core/gpu.h"
 #include "core/memory_card.h"
+#include "core/spu.h"
 #include "core/system.h"
 #include "icon.h"
+#include "sdl_audio_stream.h"
 #include <cinttypes>
 #include <glad.h>
 #include <imgui.h>
@@ -173,11 +175,26 @@ void main()
   return true;
 }
 
+bool SDLInterface::CreateAudioStream()
+{
+  m_audio_stream = std::make_unique<SDLAudioStream>();
+  if (!m_audio_stream->Reconfigure(44100, 2))
+  {
+    Panic("Failed to open audio stream");
+    return false;
+  }
+
+  return true;
+}
+
 std::unique_ptr<SDLInterface> SDLInterface::Create()
 {
   std::unique_ptr<SDLInterface> intf = std::make_unique<SDLInterface>();
-  if (!intf->CreateSDLWindow() || !intf->CreateGLContext() || !intf->CreateImGuiContext() || !intf->CreateGLResources())
+  if (!intf->CreateSDLWindow() || !intf->CreateGLContext() || !intf->CreateImGuiContext() ||
+      !intf->CreateGLResources() || !intf->CreateAudioStream())
+  {
     return nullptr;
+  }
 
   return intf;
 }
@@ -312,15 +329,15 @@ bool SDLInterface::HandleSDLEvent(const SDL_Event* event)
         break;
 
         case SDL_SCANCODE_TAB:
-          {
+        {
-            // Window framebuffer has to be bound to call SetSwapInterval.
+          // Window framebuffer has to be bound to call SetSwapInterval.
-            GLint current_fbo = 0;
+          GLint current_fbo = 0;
-            glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &current_fbo);
+          glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &current_fbo);
-            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+          glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
-            SDL_GL_SetSwapInterval(pressed ? 0 : 1);
+          SDL_GL_SetSwapInterval(pressed ? 0 : 1);
-            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, current_fbo);
+          glBindFramebuffer(GL_DRAW_FRAMEBUFFER, current_fbo);
-          }
+        }
-          break;
+        break;
 
         default:
           break;
@@ -476,6 +493,8 @@ void SDLInterface::DrawImGui()
   if (m_show_gpu_statistics)
     m_system->GetGPU()->DrawStatistics();
 
+  m_system->GetSPU()->DrawDebugWindow();
+
   DrawOSDMessages();
 
   ImGui::Render();
@@ -569,6 +588,13 @@ void SDLInterface::DrawMainMenuBar()
       ImGui::EndMenu();
     }
 
+    if (ImGui::BeginMenu("SPU"))
+    {
+      m_system->GetSPU()->DrawDebugMenu();
+
+      ImGui::EndMenu();
+    }
+
     ImGui::EndMenu();
   }
 
@@ -689,6 +715,8 @@ void SDLInterface::ConnectDevices()
 
 void SDLInterface::Run()
 {
+  m_audio_stream->PauseOutput(false);
+
   while (m_running)
   {
     for (;;)

src/duckstation/sdl_interface.h

@@ -13,6 +13,7 @@
 class System;
 class DigitalController;
 class MemoryCard;
+class AudioStream;
 
 class SDLInterface : public HostInterface
 {
@@ -47,6 +48,7 @@ private:
   bool CreateGLContext();
   bool CreateImGuiContext();
   bool CreateGLResources();
+  bool CreateAudioStream();
 
   // We only pass mouse input through if it's grabbed
   bool IsWindowFullscreen() const;