Add optional support for Tracy

CMakeLists.txt

@@ -19,6 +19,7 @@ include(CheckSymbolExists)
 include(CheckLibraryExists)
 include(CMakeDependentOption)
 include(CheckIPOSupported)
+include(cmake/FetchDependencies.cmake)
 
 set(CMAKE_OSX_DEPLOYMENT_TARGET "10.15" CACHE STRING "Minimum OS X deployment version")
 

cmake/FetchDependencies.cmake

@@ -0,0 +1,50 @@
+include(FetchContent)
+
+macro(define_git_dependency_vars name default_url default_tag)
+    string(TOUPPER ${name} VAR_NAME)
+    string(MAKE_C_IDENTIFIER ${VAR_NAME} VAR_NAME)
+
+    if (NOT ${VAR_NAME}_REPOSITORY_URL)
+        set(
+                "${VAR_NAME}_REPOSITORY_URL"
+                "${default_url}"
+                CACHE STRING
+                "${name} repository URL. Set this to use a fork."
+                FORCE
+        )
+    endif ()
+
+    if (NOT ${VAR_NAME}_REPOSITORY_TAG)
+        set(
+                "${VAR_NAME}_REPOSITORY_TAG"
+                "${default_tag}"
+                CACHE STRING
+                "${name} repository commit hash or tag. Set this when using a new version or a custom branch."
+                FORCE
+        )
+    endif ()
+endmacro()
+
+function(fetch_dependency name default_url default_tag)
+    define_git_dependency_vars(${name} ${default_url} ${default_tag})
+
+    message(STATUS "Using ${name}: ${${VAR_NAME}_REPOSITORY_URL} (ref ${${VAR_NAME}_REPOSITORY_TAG})")
+
+    FetchContent_Declare(
+            ${name}
+            GIT_REPOSITORY "${${VAR_NAME}_REPOSITORY_URL}"
+            GIT_TAG "${${VAR_NAME}_REPOSITORY_TAG}"
+    )
+
+    FetchContent_GetProperties(${name})
+endfunction()
+
+option(TRACY_ENABLE "Build with Tracy support." OFF)
+
+if (TRACY_ENABLE)
+    fetch_dependency(tracy "https://github.com/wolfpld/tracy.git" "v0.10")
+    FetchContent_MakeAvailable(tracy)
+    set(BUILD_SHARED_LIBS OFF)
+    set(TRACY_STATIC ON)
+    set(TRACY_ON_DEMAND ON)
+endif()

src/ARM.cpp

@@ -26,6 +26,7 @@
 #include "ARMJIT.h"
 #include "Platform.h"
 #include "GPU.h"
+#include "Tracy.h"
 
 #ifdef JIT_ENABLED
 #include "ARMJIT.h"
@@ -114,6 +115,7 @@ ARM::ARM(u32 num)
     : GdbStub(this, Platform::GetConfigInt(num ? Platform::GdbPortARM7 : Platform::GdbPortARM9))
 #endif
 {
+    ZoneScopedN(TracyFunction);
     // well uh
     Num = num;
 
@@ -135,6 +137,7 @@ ARM::~ARM()
 
 ARMv5::ARMv5() : ARM(0)
 {
+    ZoneScopedN(TracyFunction);
 #ifndef JIT_ENABLED
     DTCM = new u8[DTCMPhysicalSize];
 #endif
@@ -258,6 +261,7 @@ void ARMv4::Reset()
 
 void ARM::DoSavestate(Savestate* file)
 {
+    ZoneScopedN(TracyFunction);
     file->Section((char*)(Num ? "ARM7" : "ARM9"));
 
     file->Var32((u32*)&Cycles);
@@ -318,6 +322,7 @@ void ARM::DoSavestate(Savestate* file)
 
 void ARMv5::DoSavestate(Savestate* file)
 {
+    ZoneScopedN(TracyFunction);
     ARM::DoSavestate(file);
     CP15DoSavestate(file);
 }

src/ARMJIT.cpp

@@ -43,6 +43,7 @@
 #include "Wifi.h"
 #include "NDSCart.h"
 #include "Platform.h"
+#include "Tracy.h"
 
 using Platform::Log;
 using Platform::LogLevel;
@@ -1179,6 +1180,7 @@ template void CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_C>(u3
 
 void ResetBlockCache()
 {
+    ZoneScopedN(TracyFunction);
     Log(LogLevel::Debug, "Resetting JIT block cache...\n");
 
     // could be replace through a function which only resets

src/ARMJIT_Memory.cpp

@@ -45,6 +45,7 @@
 #include "Wifi.h"
 #include "NDSCart.h"
 #include "SPU.h"
+#include "Tracy.h"
 
 #include <stdlib.h>
 
@@ -877,6 +878,7 @@ void DeInit()
 
 void Reset()
 {
+    ZoneScopedN(TracyFunction);
     for (int region = 0; region < memregions_Count; region++)
     {
         for (int i = 0; i < Mappings[region].Length; i++)

src/CMakeLists.txt

@@ -140,6 +140,12 @@ if (ENABLE_JIT)
     endif()
 endif()
 
+if (TRACY_ENABLE)
+    target_link_libraries(core PUBLIC TracyClient)
+    target_include_directories(core SYSTEM PUBLIC TracyClient)
+    target_compile_definitions(core PUBLIC HAVE_TRACY TRACY_ENABLE TRACY_ENABLED)
+endif()
+
 if (WIN32)
     target_link_libraries(core PRIVATE ole32 comctl32 wsock32 ws2_32)
 elseif(NOT APPLE)

src/CP15.cpp

@@ -22,6 +22,7 @@
 #include "DSi.h"
 #include "ARM.h"
 #include "Platform.h"
+#include "Tracy.h"
 
 #ifdef JIT_ENABLED
 #include "ARMJIT.h"
@@ -75,6 +76,7 @@ void ARMv5::CP15Reset()
 
 void ARMv5::CP15DoSavestate(Savestate* file)
 {
+    ZoneScopedN(TracyFunction);
     file->Section("CP15");
 
     file->Var32(&CP15Control);
@@ -105,6 +107,7 @@ void ARMv5::CP15DoSavestate(Savestate* file)
 
 void ARMv5::UpdateDTCMSetting()
 {
+    ZoneScopedN(TracyFunction);
     u32 newDTCMBase;
     u32 newDTCMMask;
     u32 newDTCMSize;
@@ -135,6 +138,7 @@ void ARMv5::UpdateDTCMSetting()
 
 void ARMv5::UpdateITCMSetting()
 {
+    ZoneScopedN(TracyFunction);
     if (CP15Control & (1<<18))
     {
         ITCMSize = 0x200 << ((ITCMSetting >> 1) & 0x1F);
@@ -260,6 +264,7 @@ void ARMv5::UpdatePURegion(u32 n)
 
 void ARMv5::UpdatePURegions(bool update_all)
 {
+    ZoneScopedN(TracyFunction);
     if (!(CP15Control & (1<<0)))
     {
         // PU disabled
@@ -295,6 +300,7 @@ void ARMv5::UpdatePURegions(bool update_all)
 
 void ARMv5::UpdateRegionTimings(u32 addrstart, u32 addrend)
 {
+    ZoneScopedN(TracyFunction);
     for (u32 i = addrstart; i < addrend; i++)
     {
         u8 pu = PU_Map[i];

src/DMA.cpp

@@ -23,6 +23,7 @@
 #include "GPU.h"
 #include "DMA_Timings.h"
 #include "Platform.h"
+#include "Tracy.h"
 
 using Platform::Log;
 using Platform::LogLevel;
@@ -82,6 +83,7 @@ void DMA::Reset()
 
 void DMA::DoSavestate(Savestate* file)
 {
+    ZoneScopedN(TracyFunction);
     char magic[5] = "DMAx";
     magic[3] = '0' + Num + (CPU*4);
     file->Section(magic);

src/GPU.cpp

@@ -25,6 +25,7 @@
 #endif
 
 #include "GPU2D_Soft.h"
+#include "Tracy.h"
 
 using Platform::Log;
 using Platform::LogLevel;
@@ -184,6 +185,7 @@ void DeInit()
 
 void ResetVRAMCache()
 {
+    ZoneScopedN(TracyFunction);
     for (int i = 0; i < 9; i++)
         VRAMDirty[i] = NonStupidBitField<128*1024/VRAMDirtyGranularity>();
 
@@ -212,6 +214,7 @@ void ResetVRAMCache()
 
 void Reset()
 {
+    ZoneScopedN(TracyFunction);
     VCount = 0;
     NextVCount = -1;
     TotalScanlines = 0;
@@ -315,6 +318,7 @@ void Stop()
 
 void DoSavestate(Savestate* file)
 {
+    ZoneScopedN(TracyFunction);
     file->Section("GPUG");
 
     file->Var16(&VCount);
@@ -1030,6 +1034,7 @@ void DisplayFIFO(u32 x)
 
 void StartFrame()
 {
+    ZoneScopedN(TracyFunction);
     // only run the display FIFO if needed:
     // * if it is used for display or capture
     // * if we have display FIFO DMA
@@ -1041,6 +1046,8 @@ void StartFrame()
 
 void StartHBlank(u32 line)
 {
+    ZoneScopedN(TracyFunction);
+    ZoneValue(line);
     DispStat[0] |= (1<<1);
     DispStat[1] |= (1<<1);
 
@@ -1084,6 +1091,8 @@ void StartHBlank(u32 line)
 
 void FinishFrame(u32 lines)
 {
+    ZoneScopedN(TracyFunction);
+    ZoneValue(lines);
     FrontBuffer = FrontBuffer ? 0 : 1;
     AssignFramebuffers();
 
@@ -1098,6 +1107,8 @@ void FinishFrame(u32 lines)
 
 void StartScanline(u32 line)
 {
+    ZoneScopedN(TracyFunction);
+    ZoneValue(line);
     if (line == 0)
         VCount = 0;
     else if (NextVCount != 0xFFFFFFFF)

src/GPU2D.cpp

@@ -20,6 +20,7 @@
 #include <string.h>
 #include "NDS.h"
 #include "GPU.h"
+#include "Tracy.h"
 
 using Platform::Log;
 using Platform::LogLevel;
@@ -142,6 +143,7 @@ void Unit::Reset()
 
 void Unit::DoSavestate(Savestate* file)
 {
+    ZoneScopedN(TracyFunction);
     file->Section((char*)(Num ? "GP2B" : "GP2A"));
 
     file->Var32(&DispCnt);

src/GPU3D.cpp

@@ -23,6 +23,7 @@
 #include "GPU.h"
 #include "FIFO.h"
 #include "Platform.h"
+#include "Tracy.h"
 
 using Platform::Log;
 using Platform::LogLevel;
@@ -310,6 +311,7 @@ void ResetRenderingState()
 
 void Reset()
 {
+    ZoneScopedN(TracyFunction);
     CmdFIFO.Clear();
     CmdPIPE.Clear();
 
@@ -391,6 +393,7 @@ void Reset()
 
 void DoSavestate(Savestate* file)
 {
+    ZoneScopedN(TracyFunction);
     file->Section("GP3D");
 
     CmdFIFO.DoSavestate(file);
@@ -811,6 +814,7 @@ void AddCycles(s32 num)
 
 void NextVertexSlot()
 {
+    ZoneScopedN(TracyFunction);
     s32 num = (9 - VertexSlotCounter) + 1;
 
     for (;;)
@@ -854,6 +858,7 @@ void NextVertexSlot()
 
 void StallPolygonPipeline(s32 delay, s32 nonstalldelay)
 {
+    ZoneScopedN(TracyFunction);
     if (PolygonPipeline > 0)
     {
         CycleCount += PolygonPipeline + delay;
@@ -880,6 +885,7 @@ void StallPolygonPipeline(s32 delay, s32 nonstalldelay)
 template<int comp, s32 plane, bool attribs>
 void ClipSegment(Vertex* outbuf, Vertex* vin, Vertex* vout)
 {
+    ZoneScopedN(TracyFunction);
     s64 factor_num = vin->Position[3] - (plane*vin->Position[comp]);
     s32 factor_den = factor_num - (vout->Position[3] - (plane*vout->Position[comp]));
 
@@ -909,6 +915,7 @@ void ClipSegment(Vertex* outbuf, Vertex* vin, Vertex* vout)
 template<int comp, bool attribs>
 int ClipAgainstPlane(Vertex* vertices, int nverts, int clipstart)
 {
+    ZoneScopedN(TracyFunction);
     Vertex temp[10];
     int prev, next;
     int c = clipstart;
@@ -990,6 +997,7 @@ int ClipAgainstPlane(Vertex* vertices, int nverts, int clipstart)
 template<bool attribs>
 int ClipPolygon(Vertex* vertices, int nverts, int clipstart)
 {
+    ZoneScopedN(TracyFunction);
     // clip.
     // for each vertex:
     // if it's outside, check if the previous and next vertices are inside
@@ -1022,6 +1030,7 @@ bool ClipCoordsEqual(Vertex* a, Vertex* b)
 
 void SubmitPolygon()
 {
+    ZoneScopedN(TracyFunction);
     Vertex clippedvertices[10];
     Vertex* reusedvertices[2];
     int clipstart = 0;
@@ -1834,6 +1843,7 @@ inline void VertexPipelineCmdDelayed4()
 
 void ExecuteCommand()
 {
+    ZoneScopedN(TracyFunction);
     CmdFIFOEntry entry = CmdFIFORead();
 
     //printf("FIFO: processing %02X %08X. Levels: FIFO=%d, PIPE=%d\n", entry.Command, entry.Param, CmdFIFO->Level(), CmdPIPE->Level());
@@ -2452,6 +2462,7 @@ void FinishWork(s32 cycles)
 
 void Run()
 {
+    ZoneScopedN(TracyFunction);
     if (!GeometryEnabled || FlushRequest ||
         (CmdPIPE.IsEmpty() && !(GXStat & (1<<27))))
     {
@@ -2529,6 +2540,7 @@ bool YSort(Polygon* a, Polygon* b)
 
 void VBlank()
 {
+    ZoneScopedN(TracyFunction);
     if (GeometryEnabled)
     {
         if (RenderingEnabled)