DMA support!

7 files changed, 324 insertions, 35 deletions

diff --git a/ARM.cpp b/ARM.cpp
index 4b63c04..865be59 100644
--- a/ARM.cpp
+++ b/ARM.cpp
@@ -291,43 +291,15 @@ s32 ARM::Execute(s32 cycles)
             }
         }
 
-        // zorp
+        // TODO optimize this shit!!!
         if (NDS::HaltInterrupted(Num))
         {
             if (NDS::IME[Num]&1)
                 TriggerIRQ();
         }
 
-        //if (R[15]==0x2DD0)
-        //    printf("-> %08X %08X\n", R[13], Read32(0x0380FF7C));
-
-        //if (R[15]==0x37FEC7A)
-         //   printf("!!!!!!!! %08X %08X\n", R[14], CPSR);
-
-        //if (R[15] == 0x037FF102+4 && CPSR&0x20) printf("!!!!! %08X -> %08X, %08X -> %08X, %08X, %08X/%08X\n",
-        //                                               addr, R[15], cpsr, CPSR, CurInstr, R_SVC[2], R_IRQ[2]);
-
-        if (addr >= 0x37FAD68 && addr < 0x37FAE40)
-        {
-            if (addr==0x037FAE2A) debug=2;
-            //printf("!!! @ %08X %08X / %08X %08X/%08X\n", addr, R[15], Read32(0x03FFFFFC), Read32(0x04000210), Read32(0x04000214));
-        }
-        else if (debug==2)// && (CPSR&0x1F)==0x12)
-        {
-            //printf("[%08X|%08X] IRQ RET VAL = %08X | %d\n", R[15], CPSR, Read32(0x0380FF7C), cyclesrun);
-        }
-
-        /*if (R[15] == 0x03802D54+2)
-            printf("%08X -> %08X\n", addr, R[15]);
-
-            // 37FE284(ARM) -> 37FE256 -> 37FECA0 !!!
-        if (R[15] == 0x037FE256+4)
-            printf("ROYAL %08X %s -> %08X %s | %08X\n",
-                   addr, (cpsr&0x20)?"THUMB":"ARM", R[15], (CPSR&0x20)?"THUMB":"ARM",
-                   R[0]);
-        if (R[15] == 0x37FE27C)
-            printf("NOTROYAL %08X %08X %08X/%08X\n", R[0], Read32(0x0380593C + 0x3C), R_SVC[2], Read32(0x0380593C-4));*/
-        // R0 = 0380593C
+        //if (R[15] >= 0x3800170+4 && R[15]<=0x0380017A+4)
+        //    printf("!! %08X: %08X | %08X %08X\n", R[15]-4, R[13], R[0], R[1]);
 
         addr = R[15] - (CPSR&0x20 ? 4:8);
         cpsr = CPSR;
diff --git a/DMA.cpp b/DMA.cpp
new file mode 100644
index 0000000..1b04c3a
--- /dev/null
+++ b/DMA.cpp
@@ -0,0 +1,144 @@
+/*
+    Copyright 2016-2017 StapleButter
+
+    This file is part of melonDS.
+
+    melonDS is free software: you can redistribute it and/or modify it under
+    the terms of the GNU General Public License as published by the Free
+    Software Foundation, either version 3 of the License, or (at your option)
+    any later version.
+
+    melonDS is distributed in the hope that it will be useful, but WITHOUT ANY
+    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+    FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with melonDS. If not, see http://www.gnu.org/licenses/.
+*/
+
+#include <stdio.h>
+#include "NDS.h"
+#include "DMA.h"
+
+
+// NOTES ON DMA SHIT
+//
+// * could use optimized code paths for common types of DMA transfers. for example, VRAM
+// * needs to eventually be made more accurate anyway. DMA isn't instant.
+
+
+DMA::DMA(u32 cpu, u32 num)
+{
+    CPU = cpu;
+    Num = num;
+
+    Reset();
+}
+
+DMA::~DMA()
+{
+}
+
+void DMA::Reset()
+{
+    SrcAddr = 0;
+    DstAddr = 0;
+    Cnt = 0;
+
+    StartMode = 0;
+    CurSrcAddr = 0;
+    CurDstAddr = 0;
+    RemCount = 0;
+    SrcAddrInc = 0;
+    DstAddrInc = 0;
+}
+
+void DMA::WriteCnt(u32 val)
+{
+    u32 oldcnt = Cnt;
+    Cnt = val;
+
+    if ((!(oldcnt & 0x80000000)) && (val & 0x80000000))
+    {
+        CurSrcAddr = SrcAddr;
+        CurDstAddr = DstAddr;
+
+        switch (Cnt & 0x00060000)
+        {
+        case 0x00000000: DstAddrInc = 1; break;
+        case 0x00020000: DstAddrInc = -1; break;
+        case 0x00040000: DstAddrInc = 0; break;
+        case 0x00060000: DstAddrInc = 1; break;
+        }
+
+        switch (Cnt & 0x00180000)
+        {
+        case 0x00000000: SrcAddrInc = 1; break;
+        case 0x00080000: SrcAddrInc = -1; break;
+        case 0x00100000: SrcAddrInc = 0; break;
+        case 0x00180000: SrcAddrInc = 1; printf("BAD DMA SRC INC MODE 3\n"); break;
+        }
+
+        if (CPU == 0)
+            StartMode = (Cnt >> 27) & 0x7;
+        else
+            StartMode = ((Cnt >> 28) & 0x3) | 0x8;
+
+        if ((StartMode & 0x7) == 0)
+            Start();
+    }
+}
+
+void DMA::Start()
+{
+    u32 countmask;
+    if (CPU == 0)
+        countmask = 0x001FFFFF;
+    else
+        countmask = (Num==3 ? 0x0000FFFF : 0x00003FFF);
+
+    RemCount = Cnt & countmask;
+    if (!RemCount)
+        RemCount = countmask+1;
+
+    if ((Cnt & 0x00060000) == 0x00060000)
+        CurDstAddr = DstAddr;
+
+    printf("ARM%d DMA%d %08X->%08X %d units %dbit\n", CPU?7:9, Num, CurSrcAddr, CurDstAddr, RemCount, (Cnt & 0x04000000)?16:32);
+
+    // TODO: NOT MAKE THE DMA INSTANT!!
+    if (Cnt & 0x04000000)
+    {
+        u16 (*readfn)(u32) = CPU ? NDS::ARM7Read16 : NDS::ARM9Read16;
+        void (*writefn)(u32,u16) = CPU ? NDS::ARM7Write16 : NDS::ARM9Write16;
+
+        while (RemCount > 0)
+        {
+            writefn(CurDstAddr, readfn(CurSrcAddr));
+
+            CurSrcAddr += SrcAddrInc<<1;
+            CurDstAddr += DstAddrInc<<1;
+            RemCount--;
+        }
+    }
+    else
+    {
+        u32 (*readfn)(u32) = CPU ? NDS::ARM7Read32 : NDS::ARM9Read32;
+        void (*writefn)(u32,u32) = CPU ? NDS::ARM7Write32 : NDS::ARM9Write32;
+
+        while (RemCount > 0)
+        {
+            writefn(CurDstAddr, readfn(CurSrcAddr));
+
+            CurSrcAddr += SrcAddrInc<<2;
+            CurDstAddr += DstAddrInc<<2;
+            RemCount--;
+        }
+    }
+
+    if (!(Cnt & 0x02000000))
+        Cnt &= ~0x80000000;
+
+    if (Cnt & 0x40000000)
+        NDS::TriggerIRQ(CPU, NDS::IRQ_DMA0 + Num);
+}
diff --git a/DMA.h b/DMA.h
new file mode 100644
index 0000000..619b163
--- /dev/null
+++ b/DMA.h
@@ -0,0 +1,56 @@
+/*
+    Copyright 2016-2017 StapleButter
+
+    This file is part of melonDS.
+
+    melonDS is free software: you can redistribute it and/or modify it under
+    the terms of the GNU General Public License as published by the Free
+    Software Foundation, either version 3 of the License, or (at your option)
+    any later version.
+
+    melonDS is distributed in the hope that it will be useful, but WITHOUT ANY
+    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+    FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with melonDS. If not, see http://www.gnu.org/licenses/.
+*/
+
+#ifndef DMA_H
+#define DMA_H
+
+#include "types.h"
+
+class DMA
+{
+public:
+    DMA(u32 cpu, u32 num);
+    ~DMA();
+
+    void Reset();
+
+    void WriteCnt(u32 val);
+    void Start();
+
+    void StartIfNeeded(u32 mode)
+    {
+        if ((mode == StartMode) && (Cnt & 0x80000000))
+            Start();
+    }
+
+    u32 SrcAddr;
+    u32 DstAddr;
+    u32 Cnt;
+
+private:
+    u32 CPU, Num;
+
+    u32 StartMode;
+    u32 CurSrcAddr;
+    u32 CurDstAddr;
+    u32 RemCount;
+    u32 SrcAddrInc;
+    u32 DstAddrInc;
+};
+
+#endif
diff --git a/NDS.cpp b/NDS.cpp
index ba45c78..522c51e 100644
--- a/NDS.cpp
+++ b/NDS.cpp
@@ -21,6 +21,7 @@
 #include "NDS.h"
 #include "ARM.h"
 #include "CP15.h"
+#include "DMA.h"
 #include "FIFO.h"
 #include "GPU2D.h"
 #include "SPI.h"
@@ -79,6 +80,9 @@ u16 PowerControl7;
 
 Timer Timers[8];
 
+DMA* DMAs[8];
+u32 DMA9Fill[4];
+
 u16 IPCSync9, IPCSync7;
 u16 IPCFIFOCnt9, IPCFIFOCnt7;
 FIFO* IPCFIFO9; // FIFO in which the ARM9 writes
@@ -102,6 +106,15 @@ void Init()
     ARM9 = new ARM(0);
     ARM7 = new ARM(1);
 
+    DMAs[0] = new DMA(0, 0);
+    DMAs[1] = new DMA(0, 1);
+    DMAs[2] = new DMA(0, 2);
+    DMAs[3] = new DMA(0, 3);
+    DMAs[4] = new DMA(1, 0);
+    DMAs[5] = new DMA(1, 1);
+    DMAs[6] = new DMA(1, 2);
+    DMAs[7] = new DMA(1, 3);
+
     IPCFIFO9 = new FIFO(16);
     IPCFIFO7 = new FIFO(16);
 
@@ -156,6 +169,7 @@ void LoadROM()
 void Reset()
 {
     FILE* f;
+    u32 i;
 
     f = fopen("bios9.bin", "rb");
     if (!f)
@@ -216,6 +230,9 @@ void Reset()
 
     memset(Timers, 0, 8*sizeof(Timer));
 
+    for (i = 0; i < 8; i++) DMAs[i]->Reset();
+    memset(DMA9Fill, 0, 4*4);
+
     GPU2D::Reset();
     SPI::Reset();
     Wifi::Reset();
@@ -991,7 +1008,7 @@ u8 ARM7Read8(u32 addr)
         return 0;
     }
 
-    printf("unknown arm7 read8 %08X\n", addr);
+    printf("unknown arm7 read8 %08X %08X %08X/%08X\n", addr, ARM7->R[15], ARM7->R[0], ARM7->R[1]);
     return 0;
 }
 
@@ -1229,6 +1246,15 @@ u16 ARM9IORead16(u32 addr)
     case 0x04000004: return GPU2D::DispStat[0];
     case 0x04000006: return GPU2D::VCount;
 
+    case 0x040000E0: return ((u16*)DMA9Fill)[0];
+    case 0x040000E2: return ((u16*)DMA9Fill)[1];
+    case 0x040000E4: return ((u16*)DMA9Fill)[2];
+    case 0x040000E6: return ((u16*)DMA9Fill)[3];
+    case 0x040000E8: return ((u16*)DMA9Fill)[4];
+    case 0x040000EA: return ((u16*)DMA9Fill)[5];
+    case 0x040000EC: return ((u16*)DMA9Fill)[6];
+    case 0x040000EE: return ((u16*)DMA9Fill)[7];
+
     case 0x04000100: return Timers[0].Counter;
     case 0x04000102: return Timers[0].Control;
     case 0x04000104: return Timers[1].Counter;
@@ -1268,6 +1294,24 @@ u32 ARM9IORead32(u32 addr)
     {
     case 0x04000004: return GPU2D::DispStat[0] | (GPU2D::VCount << 16);
 
+    case 0x040000B0: return DMAs[0]->SrcAddr;
+    case 0x040000B4: return DMAs[0]->DstAddr;
+    case 0x040000B8: return DMAs[0]->Cnt;
+    case 0x040000BC: return DMAs[1]->SrcAddr;
+    case 0x040000C0: return DMAs[1]->DstAddr;
+    case 0x040000C4: return DMAs[1]->Cnt;
+    case 0x040000C8: return DMAs[2]->SrcAddr;
+    case 0x040000CC: return DMAs[2]->DstAddr;
+    case 0x040000D0: return DMAs[2]->Cnt;
+    case 0x040000D4: return DMAs[3]->SrcAddr;
+    case 0x040000D8: return DMAs[3]->DstAddr;
+    case 0x040000DC: return DMAs[3]->Cnt;
+
+    case 0x040000E0: return DMA9Fill[0];
+    case 0x040000E4: return DMA9Fill[1];
+    case 0x040000E8: return DMA9Fill[2];
+    case 0x040000EC: return DMA9Fill[3];
+
     case 0x04000100: return Timers[0].Counter | (Timers[0].Control << 16);
     case 0x04000104: return Timers[1].Counter | (Timers[1].Control << 16);
     case 0x04000108: return Timers[2].Counter | (Timers[2].Control << 16);
@@ -1409,6 +1453,24 @@ void ARM9IOWrite32(u32 addr, u32 val)
 {
     switch (addr)
     {
+    case 0x040000B0: DMAs[0]->SrcAddr = val; return;
+    case 0x040000B4: DMAs[0]->DstAddr = val; return;
+    case 0x040000B8: DMAs[0]->WriteCnt(val); return;
+    case 0x040000BC: DMAs[1]->SrcAddr = val; return;
+    case 0x040000C0: DMAs[1]->DstAddr = val; return;
+    case 0x040000C4: DMAs[1]->WriteCnt(val); return;
+    case 0x040000C8: DMAs[2]->SrcAddr = val; return;
+    case 0x040000CC: DMAs[2]->DstAddr = val; return;
+    case 0x040000D0: DMAs[2]->WriteCnt(val); return;
+    case 0x040000D4: DMAs[3]->SrcAddr = val; return;
+    case 0x040000D8: DMAs[3]->DstAddr = val; return;
+    case 0x040000DC: DMAs[3]->WriteCnt(val); return;
+
+    case 0x040000E0: DMA9Fill[0] = val; return;
+    case 0x040000E4: DMA9Fill[1] = val; return;
+    case 0x040000E8: DMA9Fill[2] = val; return;
+    case 0x040000EC: DMA9Fill[3] = val; return;
+
     case 0x04000100:
         Timers[0].Reload = val & 0xFFFF;
         TimerStart(0, val>>16);
@@ -1561,6 +1623,19 @@ u32 ARM7IORead32(u32 addr)
     {
     case 0x04000004: return GPU2D::DispStat[1] | (GPU2D::VCount << 16);
 
+    case 0x040000B0: return DMAs[4]->SrcAddr;
+    case 0x040000B4: return DMAs[4]->DstAddr;
+    case 0x040000B8: return DMAs[4]->Cnt;
+    case 0x040000BC: return DMAs[5]->SrcAddr;
+    case 0x040000C0: return DMAs[5]->DstAddr;
+    case 0x040000C4: return DMAs[5]->Cnt;
+    case 0x040000C8: return DMAs[6]->SrcAddr;
+    case 0x040000CC: return DMAs[6]->DstAddr;
+    case 0x040000D0: return DMAs[6]->Cnt;
+    case 0x040000D4: return DMAs[7]->SrcAddr;
+    case 0x040000D8: return DMAs[7]->DstAddr;
+    case 0x040000DC: return DMAs[7]->Cnt;
+
     case 0x04000100: return Timers[4].Counter | (Timers[4].Control << 16);
     case 0x04000104: return Timers[5].Counter | (Timers[5].Control << 16);
     case 0x04000108: return Timers[6].Counter | (Timers[6].Control << 16);
@@ -1713,6 +1788,19 @@ void ARM7IOWrite32(u32 addr, u32 val)
 {
     switch (addr)
     {
+    case 0x040000B0: DMAs[4]->SrcAddr = val; return;
+    case 0x040000B4: DMAs[4]->DstAddr = val; return;
+    case 0x040000B8: DMAs[4]->WriteCnt(val); return;
+    case 0x040000BC: DMAs[5]->SrcAddr = val; return;
+    case 0x040000C0: DMAs[5]->DstAddr = val; return;
+    case 0x040000C4: DMAs[5]->WriteCnt(val); return;
+    case 0x040000C8: DMAs[6]->SrcAddr = val; return;
+    case 0x040000CC: DMAs[6]->DstAddr = val; return;
+    case 0x040000D0: DMAs[6]->WriteCnt(val); return;
+    case 0x040000D4: DMAs[7]->SrcAddr = val; return;
+    case 0x040000D8: DMAs[7]->DstAddr = val; return;
+    case 0x040000DC: DMAs[7]->WriteCnt(val); return;
+
     case 0x04000100:
         Timers[4].Reload = val & 0xFFFF;
         TimerStart(4, val>>16);
diff --git a/main.cpp b/main.cpp
index 5ccf6c1..5b0edd7 100644
--- a/main.cpp
+++ b/main.cpp
@@ -128,6 +128,9 @@ int main()
 
     NDS::Init();
 
+    u32 nframes = 0;
+    u32 lasttick = GetTickCount();
+
     for (;;)
     {
         MSG msg;
@@ -149,6 +152,21 @@ int main()
         HDC dc = GetDC(melon);
         SetDIBitsToDevice(dc, 0, 0, 256, 384, 0, 0, 0, 384, GPU2D::Framebuffer, (BITMAPINFO*)&bmp, DIB_RGB_COLORS);
         UpdateWindow(melon);
+
+        nframes++;
+        if (nframes >= 30)
+        {
+            u32 tick = GetTickCount();
+            u32 diff = tick - lasttick;
+            lasttick = tick;
+
+            u32 fps = (nframes * 1000) / diff;
+            nframes = 0;
+
+            char melontitle[100];
+            sprintf(melontitle, "melonDS | %d FPS", fps);
+            SetWindowText(melon, melontitle);
+        }
     }
 
     return 0;
diff --git a/melonDS.cbp b/melonDS.cbp
index 744fbeb..107ed3d 100644
--- a/melonDS.cbp
+++ b/melonDS.cbp
@@ -48,6 +48,8 @@
 		<Unit filename="ARM_InstrTable.h" />
 		<Unit filename="CP15.cpp" />
 		<Unit filename="CP15.h" />
+		<Unit filename="DMA.cpp" />
+		<Unit filename="DMA.h" />
 		<Unit filename="FIFO.cpp" />
 		<Unit filename="FIFO.h" />
 		<Unit filename="GPU2D.cpp" />
diff --git a/melonDS.depend b/melonDS.depend
index 32c870b..7b48b42 100644
--- a/melonDS.depend
+++ b/melonDS.depend
@@ -1,5 +1,5 @@
 # depslib dependency file v1.0
-1481167563 source:c:\documents\sources\melonds\main.cpp
+1484695094 source:c:\documents\sources\melonds\main.cpp
 	<stdio.h>
 	<windows.h>
 	"NDS.h"
@@ -10,18 +10,19 @@
 
 1481161027 c:\documents\sources\melonds\types.h
 
-1484616465 source:c:\documents\sources\melonds\nds.cpp
+1484699425 source:c:\documents\sources\melonds\nds.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
 	"ARM.h"
 	"CP15.h"
+	"DMA.h"
 	"FIFO.h"
 	"GPU2D.h"
 	"SPI.h"
 	"Wifi.h"
 
-1484538131 source:c:\documents\sources\melonds\arm.cpp
+1484693558 source:c:\documents\sources\melonds\arm.cpp
 	<stdio.h>
 	"NDS.h"
 	"ARM.h"
@@ -102,3 +103,11 @@
 1484612398 c:\documents\sources\melonds\fifo.h
 	"types.h"
 
+1484699433 source:c:\documents\sources\melonds\dma.cpp
+	<stdio.h>
+	"NDS.h"
+	"DMA.h"
+
+1484698068 c:\documents\sources\melonds\dma.h
+	"types.h"
+