diff --git a/src/GPU3D.cpp b/src/GPU3D.cpp
index b98c4fcc..fd1b98f5 100644
--- a/src/GPU3D.cpp
+++ b/src/GPU3D.cpp
@@ -153,7 +153,8 @@ FIFO<CmdFIFOEntry>* CmdPIPE;
 u32 NumCommands, CurCommand, ParamCount, TotalParams;
 
 u32 DispCnt;
-u32 AlphaRef;
+u8 AlphaRefVal;
+u8 AlphaRef;
 
 u16 ToonTable[32];
 u16 EdgeTable[8];
@@ -1932,7 +1933,8 @@ void Write8(u32 addr, u8 val)
     switch (addr)
     {
     case 0x04000340:
-        AlphaRef = val & 0x1F;
+        AlphaRefVal = val & 0x1F;
+        AlphaRef = (DispCnt & (1<<2)) ? AlphaRefVal : 0;
         return;
     }
 
@@ -1953,10 +1955,12 @@ void Write16(u32 addr, u16 val)
         DispCnt = (val & 0x4FFF) | (DispCnt & 0x3000);
         if (val & (1<<12)) DispCnt &= ~(1<<12);
         if (val & (1<<13)) DispCnt &= ~(1<<13);
+        AlphaRef = (DispCnt & (1<<2)) ? AlphaRefVal : 0;
         return;
 
     case 0x04000340:
-        AlphaRef = val & 0x1F;
+        AlphaRefVal = val & 0x1F;
+        AlphaRef = (DispCnt & (1<<2)) ? AlphaRefVal : 0;
         return;
 
     case 0x04000350:
@@ -2014,10 +2018,12 @@ void Write32(u32 addr, u32 val)
         DispCnt = (val & 0x4FFF) | (DispCnt & 0x3000);
         if (val & (1<<12)) DispCnt &= ~(1<<12);
         if (val & (1<<13)) DispCnt &= ~(1<<13);
+        AlphaRef = (DispCnt & (1<<2)) ? AlphaRefVal : 0;
         return;
 
     case 0x04000340:
-        AlphaRef = val & 0x1F;
+        AlphaRefVal = val & 0x1F;
+        AlphaRef = (DispCnt & (1<<2)) ? AlphaRefVal : 0;
         return;
 
     case 0x04000350:
diff --git a/src/GPU3D.h b/src/GPU3D.h
index 81856a75..79287f7c 100644
--- a/src/GPU3D.h
+++ b/src/GPU3D.h
@@ -65,7 +65,7 @@ typedef struct
 } Polygon;
 
 extern u32 DispCnt;
-extern u32 AlphaRef;
+extern u8 AlphaRef;
 extern s32 Viewport[4];
 extern u32 RenderClearAttr1, RenderClearAttr2;
 
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index eb38ee79..649f01f0 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -894,6 +894,10 @@ void RenderPolygon(Polygon* polygon)
             r_filledge = slope_end->Increment==0;
         }
 
+        int yedge = 0;
+        if (y == ytop)        yedge = 0x4;
+        else if (y == ybot-1) yedge = 0x8;
+
         Interpolator interpX(xstart, xend+1, wl, wr, 8);
 
         for (s32 x = xstart; x <= xend; x++)
@@ -901,9 +905,7 @@ void RenderPolygon(Polygon* polygon)
             if (x < 0) continue;
             if (x > 255) break;
 
-            int edge = 0;
-            if (y == ytop)            edge |= 0x4;
-            else if (y == ybot-1)     edge |= 0x8;
+            int edge = yedge;
             if (x < l_edgeend)        edge |= 0x1;
             else if (x > r_edgestart) edge |= 0x2;
 
@@ -966,14 +968,7 @@ void RenderPolygon(Polygon* polygon)
 
             // alpha test
             // TODO: check alpha test when blending is disabled
-            if (DispCnt & (1<<2))
-            {
-                if (alpha <= AlphaRef) continue;
-            }
-            else
-            {
-                if (alpha == 0) continue;
-            }
+            if (alpha <= AlphaRef) continue;
 
             if (alpha == 31)
             {
diff --git a/src/NDS.cpp b/src/NDS.cpp
index 629c3474..62b89c6d 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -83,6 +83,7 @@ u16 PowerControl7;
 u16 ARM7BIOSProt;
 
 Timer Timers[8];
+u8 TimerCheckMask[2];
 
 DMA* DMAs[8];
 u32 DMA9Fill[4];
@@ -294,6 +295,8 @@ void Reset()
     CPUStop = 0;
 
     memset(Timers, 0, 8*sizeof(Timer));
+    TimerCheckMask[0] = 0;
+    TimerCheckMask[1] = 0;
 
     for (i = 0; i < 8; i++) DMAs[i]->Reset();
     memset(DMA9Fill, 0, 4*4);
@@ -614,8 +617,8 @@ void HandleTimerOverflow(u32 tid)
 void RunTimer(u32 tid, s32 cycles)
 {
     Timer* timer = &Timers[tid];
-    if ((timer->Cnt & 0x84) != 0x80)
-        return;
+    //if ((timer->Cnt & 0x84) != 0x80)
+    //    return;
 
     u32 oldcount = timer->Counter;
     timer->Counter += (cycles << timer->CycleShift);
@@ -625,10 +628,12 @@ void RunTimer(u32 tid, s32 cycles)
 
 void RunTimingCriticalDevices(u32 cpu, s32 cycles)
 {
-    RunTimer((cpu<<2)+0, cycles);
-    RunTimer((cpu<<2)+1, cycles);
-    RunTimer((cpu<<2)+2, cycles);
-    RunTimer((cpu<<2)+3, cycles);
+    register u32 timermask = TimerCheckMask[cpu];
+
+    if (timermask & 0x1) RunTimer((cpu<<2)+0, cycles);
+    if (timermask & 0x2) RunTimer((cpu<<2)+1, cycles);
+    if (timermask & 0x4) RunTimer((cpu<<2)+2, cycles);
+    if (timermask & 0x8) RunTimer((cpu<<2)+3, cycles);
 
     if (cpu == 0)
     {
@@ -681,6 +686,11 @@ void TimerStart(u32 id, u16 cnt)
     {
         timer->Counter = timer->Reload << 16;
     }
+
+    if ((cnt & 0x84) == 0x80)
+        TimerCheckMask[id>>2] |= (1<<(id&0x3));
+    else
+        TimerCheckMask[id>>2] &= ~(1<<(id&0x3));
 }