From 73be2f3e01e22c7a695edf98b6d511b1f0a3eba7 Mon Sep 17 00:00:00 2001
From: Jaklyy <102590697+Jaklyy@users.noreply.github.com>
Date: Fri, 13 Dec 2024 13:09:42 -0500
Subject: [PATCH] tweak dmas to be more accurate (actually less?)

---
 src/CP15.cpp | 33 +++++++++++++++++++++++++++------
 src/DMA.cpp  | 17 ++++++++++-------
 src/NDS.cpp  | 20 +++++++-------------
 src/NDS.h    |  2 +-
 4 files changed, 45 insertions(+), 27 deletions(-)

diff --git a/src/CP15.cpp b/src/CP15.cpp
index 747f2955..e89275c8 100644
--- a/src/CP15.cpp
+++ b/src/CP15.cpp
@@ -454,7 +454,8 @@ bool ARMv5::ICacheLookup(const u32 addr)
     // BIST test State register (See arm946e-s Rev 1 technical manual, 2.3.15 "Register 15, test State Register")
     if (CP15BISTTestStateRegister & CP15_BIST_TR_DISABLE_ICACHE_LINEFILL) [[unlikely]]
         return false;
-
+        
+    if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
     WriteBufferDrain();
     FetchAddr[16] = addr;
     QueueFunction(&ARMv5::ICacheLookup_2);
@@ -688,6 +689,8 @@ bool ARMv5::DCacheLookup(const u32 addr)
     // BIST test State register (See arm946e-s Rev 1 technical manual, 2.3.15 "Register 15, test State Register")
     if (CP15BISTTestStateRegister & CP15_BIST_TR_DISABLE_DCACHE_LINEFILL) [[unlikely]]
         return false;
+
+    if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
     WriteBufferDrain(); // checkme?
 
     FetchAddr[16] = addr;
@@ -2214,10 +2217,12 @@ void ARMv5::CodeRead32(u32 addr)
     // bus reads can only overlap with dcache streaming by 6 cycles
     if (DCacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = DCacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
     
+    if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
     if (PU_Map[addr>>12] & 0x30)
         WriteBufferDrain();
     else
@@ -2346,10 +2351,12 @@ void ARMv5::DRead8_2()
     // checkme: does dcache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
-
+    
+    if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
     if (PU_Map[addr>>12] & 0x30)
         WriteBufferDrain();
     else
@@ -2460,10 +2467,12 @@ void ARMv5::DRead16_2()
     // checkme: does cache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
-
+    
+    if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
     if (PU_Map[addr>>12] & 0x30)
         WriteBufferDrain();
     else
@@ -2575,10 +2584,12 @@ void ARMv5::DRead32_2()
     // checkme: does cache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
-
+    
+    if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
     if (PU_Map[addr>>12] & 0x30)
         WriteBufferDrain();
     else
@@ -2676,15 +2687,17 @@ void ARMv5::DRead32S_2()
     // checkme: does cache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
-
+    
+    if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
     if (PU_Map[addr>>12] & 0x30) // checkme
         WriteBufferDrain();
     else
         WriteBufferCheck<1>();
-
+        
     QueueFunction(&ARMv5::DRead32S_3);
 }
 
@@ -2809,12 +2822,14 @@ void ARMv5::DWrite8_2()
     // checkme: does cache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
 
     if (!(PU_Map[addr>>12] & (0x30)))
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         WriteBufferCheck<2>();
         QueueFunction(&ARMv5::DWrite8_3);
     }
@@ -2922,12 +2937,14 @@ void ARMv5::DWrite16_2()
     // checkme: does cache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
 
     if (!(PU_Map[addr>>12] & 0x30))
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         WriteBufferCheck<2>();
         QueueFunction(&ARMv5::DWrite16_3);
     }
@@ -3040,12 +3057,14 @@ void ARMv5::DWrite32_2()
     // checkme: does cache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
 
     if (!(PU_Map[addr>>12] & 0x30))
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         WriteBufferCheck<2>();
         QueueFunction(&ARMv5::DWrite32_3);
     }
@@ -3153,12 +3172,14 @@ void ARMv5::DWrite32S_2()
     // checkme: does cache trigger this?
     if (ICacheStreamPtr < 7)
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         u64 time = ICacheStreamTimes[6] - 6; // checkme: minus 6?
         if (NDS.ARM9Timestamp < time) NDS.ARM9Timestamp = time;
     }
 
     if (!(PU_Map[addr>>12] & 0x30)) // non-bufferable
     {
+        if (NDS.ARM9Timestamp < NDS.DMA9Timestamp) NDS.ARM9Timestamp = NDS.DMA9Timestamp;
         WriteBufferCheck<2>();
         QueueFunction(&ARMv5::DWrite32S_3);
     }
diff --git a/src/DMA.cpp b/src/DMA.cpp
index 5269a974..4ea122f3 100644
--- a/src/DMA.cpp
+++ b/src/DMA.cpp
@@ -565,14 +565,15 @@ u32 DMA::UnitTimings7_32(bool burststart)
 
 void DMA::Run9()
 {
-    if (NDS.ARM9Timestamp >= NDS.ARM9Target) return;
+    NDS.DMA9Timestamp = std::max(NDS.DMA9Timestamp, NDS.ARM9Timestamp);
+    NDS.DMA9Timestamp = (NDS.DMA9Timestamp + ((1<<NDS.ARM9ClockShift)-1)) & ~((1<<NDS.ARM9ClockShift)-1);
+
+    if (NDS.DMA9Timestamp-1 >= NDS.ARM9Target) return;
 
     Executing = true;
 
     // add NS penalty for first accesses in burst
     int burststart = Running-1;
-
-    NDS.ARM9Timestamp = (NDS.ARM9Timestamp + ((1<<NDS.ARM9ClockShift)-1)) & ~((1<<NDS.ARM9ClockShift)-1);
     
     if (!(Cnt & (1<<26)))
     {
@@ -587,7 +588,7 @@ void DMA::Run9()
             }
             Running = 2;
 
-            NDS.ARM9Timestamp += (UnitTimings9_16(burststart) << NDS.ARM9ClockShift);
+            NDS.DMA9Timestamp += (UnitTimings9_16(burststart) << NDS.ARM9ClockShift);
             burststart -= 1;
 
             NDS.ARM9Write16(CurDstAddr, NDS.ARM9Read16(CurSrcAddr));
@@ -597,7 +598,7 @@ void DMA::Run9()
             IterCount--;
             RemCount--;
 
-            if (NDS.ARM9Timestamp >= NDS.ARM9Target) break;
+            if (NDS.DMA9Timestamp-1 >= NDS.ARM9Target) break;
         }
     }
     else
@@ -613,7 +614,7 @@ void DMA::Run9()
             }
             Running = 2;
 
-            NDS.ARM9Timestamp += (UnitTimings9_32(burststart) << NDS.ARM9ClockShift);
+            NDS.DMA9Timestamp += (UnitTimings9_32(burststart) << NDS.ARM9ClockShift);
             burststart -= 1;
 
             NDS.ARM9Write32(CurDstAddr, NDS.ARM9Read32(CurSrcAddr));
@@ -623,10 +624,12 @@ void DMA::Run9()
             IterCount--;
             RemCount--;
 
-            if (NDS.ARM9Timestamp >= NDS.ARM9Target) break;
+            if (NDS.DMA9Timestamp-1 >= NDS.ARM9Target) break;
         }
     }
 
+    NDS.DMA9Timestamp -= 1;
+
     if (burststart == 0) Running = 1;  
 
     Executing = false;
diff --git a/src/NDS.cpp b/src/NDS.cpp
index 2f97a9db..5d6badef 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -470,7 +470,7 @@ void NDS::Reset()
     // unitialised on the first run
     ARM9.CP15Reset();
 
-    ARM9Timestamp = 0; ARM9Target = 0;
+    ARM9Timestamp = 0; DMA9Timestamp = 0; ARM9Target = 0;
     ARM7Timestamp = 0; ARM7Target = 0;
     MainRAMTimestamp = 0;
     A9ContentionTS = 0; ConTSLock = false;
@@ -1127,7 +1127,7 @@ void NDS::MainRAMHandleARM9()
                 burststart -= 1;
                 if (burststart <= 0) dma->Running = 1;
                 else dma->Running = 2;
-                ARM9Timestamp = A9ContentionTS << ARM9ClockShift;
+                DMA9Timestamp = (A9ContentionTS << ARM9ClockShift) - 1;
                 memset(&ARM9.MRTrack, 0, sizeof(ARM9.MRTrack));
                 ConTSLock = false;
                 if (dma->RemCount)
@@ -1229,7 +1229,7 @@ void NDS::MainRAMHandleARM9()
                 burststart -= 1;
                 if (burststart <= 0) Running = 1;
                 else dma->Running = 2;
-                ARM9Timestamp = A9ContentionTS << ARM9ClockShift;
+                DMA9Timestamp = (A9ContentionTS << ARM9ClockShift) - 1;
                 memset(&ARM9.MRTrack, 0, sizeof(ARM9.MRTrack));
                 ConTSLock = false;
                 if (dma->RemCount)
@@ -1416,6 +1416,8 @@ bool NDS::MainRAMHandle()
 
         if (ARM9.MRTrack.Type > MainRAMType::WriteBufferCmds)
             A9ContentionTS = (ARM9.WBTimestamp + ((1<<ARM9ClockShift)-1)) >> ARM9ClockShift;
+        else if (ARM9.MRTrack.Type == MainRAMType::DMA16 || ARM9.MRTrack.Type == MainRAMType::DMA32)
+            A9ContentionTS = (DMA9Timestamp + ((1<<ARM9ClockShift)-1)) >> ARM9ClockShift;
         else
             A9ContentionTS = (ARM9Timestamp + ((1<<ARM9ClockShift)-1)) >> ARM9ClockShift;
     }
@@ -1523,7 +1525,7 @@ u32 NDS::RunFrame()
                 ARM9Target = target << ARM9ClockShift;
                 CurCPU = 0;
 
-                while (ARM9Timestamp < ARM9Target)
+                while (std::max(ARM9Timestamp, DMA9Timestamp) < ARM9Target)
                 {
                     if (ARM9.MRTrack.Type == MainRAMType::Null)
                     {
@@ -1532,11 +1534,10 @@ u32 NDS::RunFrame()
                             // GXFIFO stall
                             s32 cycles = GPU.GPU3D.CyclesToRunFor();
 
-                            ARM9Timestamp = std::min(ARM9Target, ARM9Timestamp+(cycles<<ARM9ClockShift));
+                            DMA9Timestamp = std::min(ARM9Target, std::min(ARM9Timestamp+(cycles<<ARM9ClockShift), DMA9Timestamp+(cycles<<ARM9ClockShift)));
                         }
                         else if (CPUStop & CPUStop_DMA9)
                         {
-                            u64 ts = ARM9Timestamp;
                             DMAs[0].Run();
                             if (!(CPUStop & CPUStop_GXStall) && (ARM9.MRTrack.Type == MainRAMType::Null)) DMAs[1].Run();
                             if (!(CPUStop & CPUStop_GXStall) && (ARM9.MRTrack.Type == MainRAMType::Null)) DMAs[2].Run();
@@ -1546,13 +1547,6 @@ u32 NDS::RunFrame()
                                 auto& dsi = dynamic_cast<melonDS::DSi&>(*this);
                                 dsi.RunNDMAs(0);
                             }
-                            ts = ARM9Timestamp - ts;
-                            for (int i = 0; i < 7; i++)
-                            {
-                                ARM9.ICacheStreamTimes[i] += ts;
-                                ARM9.DCacheStreamTimes[i] += ts;
-                            }
-                            ARM9.WBTimestamp += ts;
                         }
                         else
                         {
diff --git a/src/NDS.h b/src/NDS.h
index b0b5a911..939a7a67 100644
--- a/src/NDS.h
+++ b/src/NDS.h
@@ -255,7 +255,7 @@ public: // TODO: Encapsulate the rest of these members
     bool LagFrameFlag;
 
     // no need to worry about those overflowing, they can keep going for atleast 4350 years
-    u64 ARM9Timestamp, ARM9Target;
+    u64 ARM9Timestamp, DMA9Timestamp, ARM9Target;
     u64 ARM7Timestamp, ARM7Target;
     u64 MainRAMTimestamp;
     u64 A9ContentionTS; bool ConTSLock;