7 files changed, 76 insertions, 14 deletions

diff --git a/src/DMA.cpp b/src/DMA.cpp
index 5a7ca1c..14fe791 100644
--- a/src/DMA.cpp
+++ b/src/DMA.cpp
@@ -20,7 +20,7 @@
 #include "NDS.h"
 #include "DMA.h"
 #include "NDSCart.h"
-#include "GPU3D.h"
+#include "GPU.h"
 
 
 // NOTES ON DMA SHIT
@@ -153,8 +153,8 @@ void DMA::WriteCnt(u32 val)
         else if (StartMode == 0x07)
             GPU3D::CheckFIFODMA();
 
-        if (StartMode==0x04 || StartMode==0x06 || StartMode==0x13)
-            printf("UNIMPLEMENTED ARM%d DMA%d START MODE %02X\n", CPU?7:9, Num, StartMode);
+        if (StartMode==0x06 || StartMode==0x13)
+            printf("UNIMPLEMENTED ARM%d DMA%d START MODE %02X, %08X->%08X\n", CPU?7:9, Num, StartMode, SrcAddr, DstAddr);
     }
 }
 
@@ -194,6 +194,20 @@ void DMA::Start()
         return;
     }
 
+    // special path for the display FIFO. another gross hack.
+    // the display FIFO seems to be more like a circular buffer that holds 16 pixels
+    // from which the display controller reads. DMA is triggered every 8 pixels to fill it
+    // as it is being read from. emulating it properly would be resource intensive.
+    // proper emulation would only matter if something is trying to feed the FIFO manually
+    // instead of using the DMA. which is probably never happening. the only thing I know of
+    // that even uses the display FIFO is the aging cart.
+    if (StartMode == 0x04)
+    {
+        GPU::GPU2D_A->FIFODMA(CurSrcAddr);
+        CurSrcAddr += 256*2;
+        return;
+    }
+
     // TODO eventually: not stop if we're running code in ITCM
 
     Running = true;
diff --git a/src/DMA.h b/src/DMA.h
index 59a7f03..1b8bbb2 100644
--- a/src/DMA.h
+++ b/src/DMA.h
@@ -40,6 +40,12 @@ public:
             Start();
     }
 
+    void StopIfNeeded(u32 mode)
+    {
+        if (mode == StartMode)
+            Cnt &= ~0x80000000;
+    }
+
     u32 SrcAddr;
     u32 DstAddr;
     u32 Cnt;
diff --git a/src/GPU.cpp b/src/GPU.cpp
index aaef43f..5b9bf33 100644
--- a/src/GPU.cpp
+++ b/src/GPU.cpp
@@ -678,13 +678,23 @@ void StartScanline(u32 line)
     else
         DispStat[1] &= ~(1<<2);
 
+    if (line >= 2 && line < 194)
+        NDS::CheckDMAs(0, 0x03);
+    else if (line == 194)
+        NDS::StopDMAs(0, 0x03);
+
     if (line < 192)
     {
+        // fill a line from the display FIFO if needed.
+        // this isn't how the real thing works, but emulating it
+        // properly would be too much trouble given barely anything
+        // uses FIFO display
+        // (TODO, eventually: emulate it properly)
+        NDS::CheckDMAs(0, 0x04);
+
         // draw
         GPU2D_A->DrawScanline(line);
         GPU2D_B->DrawScanline(line);
-
-        //NDS::ScheduleEvent(LINE_CYCLES, StartScanline, line+1);
     }
     else if (line == 262)
     {
@@ -701,6 +711,8 @@ void StartScanline(u32 line)
             DispStat[0] |= (1<<0);
             DispStat[1] |= (1<<0);
 
+            NDS::StopDMAs(0, 0x04);
+
             NDS::CheckDMAs(0, 0x01);
             NDS::CheckDMAs(1, 0x11);
 
@@ -711,14 +723,8 @@ void StartScanline(u32 line)
             GPU2D_B->VBlank();
             GPU3D::VBlank();
         }
-
-        //NDS::ScheduleEvent(LINE_CYCLES, StartScanline, line+1);
-        //NDS::ScheduleEvent(NDS::Event_LCD, true, LINE_CYCLES, StartScanline, line+1);
     }
 
-    // checkme
-    if (line == 0) NDS::CheckDMAs(0, 0x03);
-
     NDS::ScheduleEvent(NDS::Event_LCD, true, HBLANK_CYCLES, StartHBlank, line);
 }
 
diff --git a/src/GPU2D.cpp b/src/GPU2D.cpp
index cedfe1e..3717cf5 100644
--- a/src/GPU2D.cpp
+++ b/src/GPU2D.cpp
@@ -91,6 +91,8 @@ void GPU2D::Reset()
     EVB = 0;
     EVY = 0;
 
+    memset(DispFIFOBuffer, 0, 256*2);
+
     CaptureCnt = 0;
 
     MasterBrightness = 0;
@@ -337,9 +339,17 @@ void GPU2D::DrawScanline(u32 line)
         }
         break;
 
-    case 3: // FIFO display
+    case 3: // FIFO display (grossly inaccurate)
         {
-            // TODO
+            for (int i = 0; i < 256; i++)
+            {
+                u16 color = DispFIFOBuffer[i];
+                u8 r = (color & 0x001F) << 1;
+                u8 g = (color & 0x03E0) >> 4;
+                u8 b = (color & 0x7C00) >> 9;
+
+                dst[i] = r | (g << 8) | (b << 16);
+            }
         }
         break;
     }
@@ -445,7 +455,8 @@ void GPU2D::DoCapture(u32 line, u32 width, u32* src)
 
     if (CaptureCnt & (1<<25))
     {
-        // TODO: FIFO mode
+        srcB = &DispFIFOBuffer[0];
+        srcBaddr = 0;
     }
     else
     {
@@ -570,6 +581,17 @@ void GPU2D::DoCapture(u32 line, u32 width, u32* src)
     }
 }
 
+void GPU2D::FIFODMA(u32 addr)
+{
+    for (int i = 0; i < 256; i += 2)
+    {
+        u32 val = NDS::ARM9Read32(addr);
+        addr += 4;
+        DispFIFOBuffer[i] = val & 0xFFFF;
+        DispFIFOBuffer[i+1] = val >> 16;
+    }
+}
+
 
 void GPU2D::BGExtPalDirty(u32 base)
 {
diff --git a/src/GPU2D.h b/src/GPU2D.h
index 4136440..acbe967 100644
--- a/src/GPU2D.h
+++ b/src/GPU2D.h
@@ -45,10 +45,14 @@ public:
     u16* GetBGExtPal(u32 slot, u32 pal);
     u16* GetOBJExtPal(u32 pal);
 
+    void FIFODMA(u32 addr);
+
 private:
     u32 Num;
     u32* Framebuffer;
 
+    u16 DispFIFOBuffer[256];
+
     u32 DispCnt;
     u16 BGCnt[4];
 
diff --git a/src/NDS.cpp b/src/NDS.cpp
index 3ab543e..d2fa59c 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -622,6 +622,15 @@ void CheckDMAs(u32 cpu, u32 mode)
     DMAs[cpu+3]->StartIfNeeded(mode);
 }
 
+void StopDMAs(u32 cpu, u32 mode)
+{
+    cpu <<= 2;
+    DMAs[cpu+0]->StopIfNeeded(mode);
+    DMAs[cpu+1]->StopIfNeeded(mode);
+    DMAs[cpu+2]->StopIfNeeded(mode);
+    DMAs[cpu+3]->StopIfNeeded(mode);
+}
+
 
 
 
diff --git a/src/NDS.h b/src/NDS.h
index 28f777e..9721c84 100644
--- a/src/NDS.h
+++ b/src/NDS.h
@@ -148,6 +148,7 @@ void StopCPU(u32 cpu, u32 mask);
 void ResumeCPU(u32 cpu, u32 mask);
 
 void CheckDMAs(u32 cpu, u32 mode);
+void StopDMAs(u32 cpu, u32 mode);
 
 void RunTimingCriticalDevices(u32 cpu, s32 cycles);