Darwin: Implement Mach exceptions ports instead of signals

Credit goes to the Julia language and Dolphin emulator for their free and open source implementations

bin/resources/GameIndex.yaml

@@ -17830,6 +17830,8 @@ SLES-51393:
 SLES-51397:
   name: "IndyCar Series"
   region: "PAL-M5"
+  gsHWFixes:
+    halfPixelOffset: 5 # Aligns post-processing and fixes depth line.
 SLES-51398:
   name: "World Championship Snooker 2003"
   region: "PAL-E"
@@ -19953,6 +19955,8 @@ SLES-52298:
   name: "IndyCar Series 2005"
   region: "PAL-M5"
   compat: 5
+  gsHWFixes:
+    halfPixelOffset: 5 # Aligns post-processing and fixes depth line.
 SLES-52308:
   name: "Karaoke Stage"
   region: "PAL-M5"
@@ -20123,6 +20127,9 @@ SLES-52378:
   name: "Euro Rally Champion"
   region: "PAL-M5"
   compat: 5
+  gsHWFixes:
+    halfPixelOffset: 5 # Aligns post-processing and fixes depth line.
+    textureInsideRT: 1 # Fixes broken fog rendering.
 SLES-52379:
   name: "Shrek 2"
   region: "PAL-E"
@@ -25057,6 +25064,8 @@ SLES-53957:
 SLES-53958:
   name: "Noble Racing"
   region: "PAL-E"
+  gsHWFixes:
+    halfPixelOffset: 5 # Aligns post-processing and fixes depth line.
 SLES-53959:
   name: "Pac-Man World 3"
   region: "PAL-M5"
@@ -66984,6 +66993,8 @@ SLUS-20641:
   name: "IndyCar Series"
   region: "NTSC-U"
   compat: 5
+  gsHWFixes:
+    halfPixelOffset: 5 # Aligns post-processing and fixes depth line.
 SLUS-20642:
   name: "Auto Modellista"
   region: "NTSC-U"

bin/resources/game_controller_db.txt

@@ -850,6 +850,7 @@ xinput,XInput Controller,a:b0,b:b1,back:b6,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,
 
 # Mac OS X
 030000008f0e00000300000009010000,2 In 1 Joystick,a:b2,b:b1,back:b8,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,leftshoulder:b4,leftstick:b10,lefttrigger:b6,leftx:a0,lefty:a1,rightshoulder:b5,rightstick:b11,righttrigger:b7,rightx:a2,righty:a3,start:b9,x:b3,y:b0,platform:Mac OS X,
+03000000c82d00001930000000000000,8BitDo 64,a:b0,b:b1,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,righttrigger:b9,rightx:a2,righty:a3,start:b11,platform:Mac OS X,
 03000000c82d00001930000000020000,8BitDo 64,a:b0,b:b1,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,righttrigger:b9,rightx:a2,righty:a3,start:b11,platform:Mac OS X,
 03000000c82d00001930000001000000,8BitDo 64,a:b0,b:b1,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,righttrigger:b9,rightx:a2,righty:a3,start:b11,platform:Mac OS X,
 03000000c82d00000031000001000000,8BitDo Adapter,a:b0,b:b1,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,rightstick:b14,righttrigger:b9,rightx:a2,righty:a3,start:b11,x:b3,y:b4,platform:Mac OS X,
@@ -895,7 +896,7 @@ xinput,XInput Controller,a:b0,b:b1,back:b6,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,
 03000000c82d00001290000001000000,8BitDo SN30,a:b1,b:b0,back:b10,dpdown:+a1,dpleft:-a0,dpright:+a0,dpup:-a1,leftshoulder:b6,rightshoulder:b7,start:b11,x:b4,y:b3,platform:Mac OS X,
 03000000c82d00004028000000010000,8BitDo SN30,a:b1,b:b0,back:b10,dpdown:+a1,dpleft:-a0,dpright:+a0,dpup:-a1,leftshoulder:b6,rightshoulder:b7,start:b11,x:b4,y:b3,platform:Mac OS X,
 03000000c82d00000160000001000000,8BitDo SN30 Pro,a:b1,b:b0,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,leftshoulder:b6,leftstick:b13,lefttrigger:a4,leftx:a0,lefty:a1,rightshoulder:b7,rightstick:b14,righttrigger:a5,rightx:a2,righty:a3,start:b11,x:b4,y:b3,platform:Mac OS X,
-03000000c82d00000161000000010000,8BitDo SN30 Pro,a:b1,b:b0,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b2,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,rightstick:b14,righttrigger:b9,rightx:a2,righty:a5,start:b11,x:b4,y:b3,platform:Mac OS X,
+03000000c82d00000161000000010000,8BitDo SN30 Pro,a:b1,b:b0,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b2,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,rightstick:b14,righttrigger:b9,rightx:a2,righty:a3,start:b11,x:b4,y:b3,platform:Mac OS X,
 03000000c82d00000260000001000000,8BitDo SN30 Pro Plus,a:b1,b:b0,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b2,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,rightstick:b14,righttrigger:b9,rightx:a2,righty:a3,start:b11,x:b4,y:b3,platform:Mac OS X,
 03000000c82d00000261000000010000,8BitDo SN30 Pro Plus,a:b1,b:b0,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b2,leftshoulder:b6,leftstick:b13,lefttrigger:b8,leftx:a0,lefty:a1,rightshoulder:b7,rightstick:b14,righttrigger:b9,rightx:a2,righty:a3,start:b11,x:b4,y:b3,platform:Mac OS X,
 03000000c82d00001230000000010000,8BitDo Ultimate,a:b0,b:b1,back:b10,dpdown:h0.4,dpleft:h0.8,dpright:h0.2,dpup:h0.1,guide:b12,leftshoulder:b6,leftstick:b13,lefttrigger:a5,leftx:a0,lefty:a1,paddle1:b2,paddle2:b5,rightshoulder:b7,rightstick:b14,righttrigger:a4,rightx:a2,righty:a3,start:b11,x:b3,y:b4,platform:Mac OS X,

common/Darwin/DarwinMisc.cpp

@@ -11,6 +11,7 @@
 #include "common/Threading.h"
 #include "common/WindowInfo.h"
 #include "common/HostSys.h"
+#include "fmt/format.h"
 
 #include <csignal>
 #include <cstring>
@@ -19,12 +20,15 @@
 #include <sys/mman.h>
 #include <sys/types.h>
 #include <sys/sysctl.h>
+#include <thread>
 #include <time.h>
 #include <mach/mach_init.h>
 #include <mach/mach_port.h>
 #include <mach/mach_time.h>
 #include <mach/mach_vm.h>
+#include <mach/message.h>
 #include <mach/task.h>
+#include <mach/thread_state.h>
 #include <mach/vm_map.h>
 #include <mutex>
 #include <ApplicationServices/ApplicationServices.h>
@@ -294,7 +298,7 @@ static CPUInfo CalcCPUInfo()
 	std::vector<DarwinMisc::CPUClass> classes = DarwinMisc::GetCPUClasses();
 	out.num_clusters = static_cast<u32>(classes.size());
 	out.num_big_cores = classes.empty() ? 0 : classes[0].num_physical;
-	out.num_threads   = classes.empty() ? 0 : classes[0].num_logical;
+	out.num_threads = classes.empty() ? 0 : classes[0].num_logical;
 	out.num_small_cores = 0;
 	for (std::size_t i = 1; i < classes.size(); i++)
 	{
@@ -568,15 +572,206 @@ void HostSys::EndCodeWrite()
 
 #endif // _M_ARM64
 
+#define USE_MACH_EXCEPTION_PORTS
+
 namespace PageFaultHandler
 {
+#ifdef USE_MACH_EXCEPTION_PORTS
+	static void SignalHandler(mach_port_t port);
+#else
 	static void SignalHandler(int sig, siginfo_t* info, void* ctx);
+#endif
 
 	static std::recursive_mutex s_exception_handler_mutex;
 	static bool s_in_exception_handler = false;
 	static bool s_installed = false;
 } // namespace PageFaultHandler
 
+#ifdef USE_MACH_EXCEPTION_PORTS
+
+#if defined(_M_X86)
+#define THREAD_STATE64_COUNT x86_THREAD_STATE64_COUNT
+#define THREAD_STATE64 x86_THREAD_STATE64
+#define thread_state64_t x86_thread_state64_t
+#elif defined(_M_ARM64)
+#define THREAD_STATE64_COUNT ARM_THREAD_STATE64_COUNT
+#define THREAD_STATE64 ARM_THREAD_STATE64
+#define thread_state64_t arm_thread_state64_t
+#else
+#error Unknown Darwin Platform
+#endif
+
+void PageFaultHandler::SignalHandler(mach_port_t port)
+{
+	Threading::SetNameOfCurrentThread("Mach Exception Thread");
+
+#pragma pack(4)
+	struct
+	{
+		mach_msg_header_t Head;
+		NDR_record_t NDR;
+		exception_type_t exception;
+		mach_msg_type_number_t codeCnt;
+		int64_t code[2];
+		int flavor;
+		mach_msg_type_number_t old_stateCnt;
+		natural_t old_state[THREAD_STATE64_COUNT];
+		mach_msg_trailer_t trailer;
+	} msg_in;
+
+	struct
+	{
+		mach_msg_header_t Head;
+		NDR_record_t NDR;
+		kern_return_t RetCode;
+		int flavor;
+		mach_msg_type_number_t new_stateCnt;
+		natural_t new_state[THREAD_STATE64_COUNT];
+	} msg_out;
+#pragma pack()
+	memset(&msg_in, 0xee, sizeof(msg_in));
+	memset(&msg_out, 0xee, sizeof(msg_out));
+	mach_msg_size_t send_size = 0;
+	mach_msg_option_t option = MACH_RCV_MSG;
+	while (true)
+	{
+		kern_return_t r;
+		if ((r = mach_msg_overwrite(&msg_out.Head, option, send_size, sizeof(msg_in), port,
+				 MACH_MSG_TIMEOUT_NONE, MACH_PORT_NULL, &msg_in.Head, 0)))
+		{
+			pxFail(fmt::format("CRITICAL: mach_msg_overwrite: {:x}", r).c_str());
+		}
+
+		if (msg_in.Head.msgh_id == MACH_NOTIFY_NO_SENDERS)
+		{
+			// the other thread exited
+			mach_port_deallocate(mach_task_self(), port);
+			return;
+		}
+
+		if (msg_in.Head.msgh_id != 2406)
+		{
+			pxFailRel("unknown message received");
+			return;
+		}
+
+		if (msg_in.flavor != THREAD_STATE64)
+		{
+			pxFailRel(fmt::format("unknown flavour {}, expected {}", msg_in.flavor, THREAD_STATE64).c_str());
+			return;
+		}
+
+		s_exception_handler_mutex.lock();
+		thread_state64_t* state = (thread_state64_t*)msg_in.old_state;
+
+		HandlerResult result = HandlerResult::ExecuteNextHandler;
+		if (!s_in_exception_handler)
+		{
+			s_in_exception_handler = true;
+
+#ifdef _M_ARM64
+			result = HandlePageFault(reinterpret_cast<void*>(state->__pc), reinterpret_cast<void*>(msg_in.code[1]), (msg_in.code[0] & 2) != 0);
+#else
+			result = HandlePageFault(reinterpret_cast<void*>(state->__rip), reinterpret_cast<void*>(msg_in.code[1]), (msg_in.code[0] & 2) != 0);
+#endif
+			s_in_exception_handler = false;
+		}
+
+		// Set up the reply.
+		msg_out.Head.msgh_bits = MACH_MSGH_BITS(MACH_MSGH_BITS_REMOTE(msg_in.Head.msgh_bits), 0);
+		msg_out.Head.msgh_remote_port = msg_in.Head.msgh_remote_port;
+		msg_out.Head.msgh_local_port = MACH_PORT_NULL;
+		msg_out.Head.msgh_id = msg_in.Head.msgh_id + 100;
+		msg_out.NDR = msg_in.NDR;
+
+		if (result != HandlerResult::ContinueExecution) // cooked
+		{
+			msg_out.RetCode = KERN_FAILURE;
+			msg_out.flavor = 0;
+			msg_out.new_stateCnt = 0;
+
+			// The crash handler on macOS or Linux doesn't use context passed to it
+			// Stubbing it here is fine
+			CrashHandler::CrashSignalHandler(-1, nullptr, nullptr);
+
+			pxFailRel("CrashSignalHandler returned when it should have terminated us!");
+		}
+		else
+		{
+			// Resumes execution right where we left off (re-executes instruction that caused the SIGSEGV)
+			msg_out.RetCode = KERN_SUCCESS;
+			msg_out.flavor = THREAD_STATE64;
+			msg_out.new_stateCnt = THREAD_STATE64_COUNT;
+			memcpy(msg_out.new_state, msg_in.old_state, THREAD_STATE64_COUNT * sizeof(natural_t));
+		}
+
+		msg_out.Head.msgh_size =
+			offsetof(__typeof__(msg_out), new_state) + msg_out.new_stateCnt * sizeof(natural_t);
+		send_size = msg_out.Head.msgh_size;
+		option |= MACH_SEND_MSG;
+
+		s_exception_handler_mutex.unlock();
+	}
+}
+
+bool PageFaultHandler::Install(Error* error)
+{
+	exception_mask_t masks[EXC_TYPES_COUNT];
+	mach_port_t ports[EXC_TYPES_COUNT];
+	exception_behavior_t behaviors[EXC_TYPES_COUNT];
+	thread_state_flavor_t flavors[EXC_TYPES_COUNT];
+	mach_msg_type_number_t count = EXC_TYPES_COUNT;
+
+	kern_return_t r = task_get_exception_ports(mach_task_self(), EXC_MASK_ALL,
+		masks, &count, ports, behaviors, flavors);
+
+	mach_port_t port;
+	if ((r = mach_port_allocate(mach_task_self(), MACH_PORT_RIGHT_RECEIVE, &port)))
+	{
+		pxFailRel(fmt::format("mach_port_allocate: {:x}", r).c_str());
+		return false;
+	}
+
+	std::thread sig_thread(PageFaultHandler::SignalHandler, port);
+	sig_thread.detach();
+
+	if ((r = mach_port_insert_right(mach_task_self(), port, port, MACH_MSG_TYPE_MAKE_SEND)))
+	{
+		mach_port_deallocate(mach_task_self(), port);
+		pxFailRel(fmt::format("mach_port_insert_right: {:x}", r).c_str());
+		return false;
+	}
+
+	task_set_exception_ports(mach_task_self(), EXC_MASK_BAD_ACCESS, MACH_PORT_NULL, EXCEPTION_DEFAULT, THREAD_STATE_NONE);
+
+	if ((r = thread_set_exception_ports(mach_thread_self(), EXC_MASK_BAD_ACCESS, port, EXCEPTION_STATE | MACH_EXCEPTION_CODES, THREAD_STATE64)))
+	{
+		mach_port_deallocate(mach_task_self(), port);
+		pxFailRel(fmt::format("thread_set_exception_ports: {:x}", r).c_str());
+		return false;
+	}
+
+	if ((r = mach_port_mod_refs(mach_task_self(), port, MACH_PORT_RIGHT_SEND, -1)))
+	{
+		mach_port_deallocate(mach_task_self(), port);
+		pxFailRel(fmt::format("mach_port_mod_refs: {:x}", r).c_str());
+		return false;
+	}
+
+	mach_port_t previous;
+	if ((r = mach_port_request_notification(mach_task_self(), port, MACH_NOTIFY_NO_SENDERS, 0, port, MACH_MSG_TYPE_MAKE_SEND_ONCE, &previous)))
+	{
+		mach_port_deallocate(mach_task_self(), port);
+		pxFailRel(fmt::format("mach_port_mod_refs: {:x}", r).c_str());
+		return false;
+	}
+
+	s_installed = true;
+	return true;
+}
+
+#else
+
 void PageFaultHandler::SignalHandler(int sig, siginfo_t* info, void* ctx)
 {
 #if defined(_M_X86)
@@ -644,3 +839,4 @@ bool PageFaultHandler::Install(Error* error)
 	s_installed = true;
 	return true;
 }
+#endif

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -2181,13 +2181,13 @@ void GSDevice12::IASetIndexBuffer(const void* index, size_t count)
 	m_index_stream_buffer.CommitMemory(size);
 }
 
-void GSDevice12::OMSetRenderTargets(GSTexture* rt, GSTexture* ds, const GSVector4i& scissor)
+void GSDevice12::OMSetRenderTargets(GSTexture* rt, GSTexture* ds, const GSVector4i& scissor, bool depth_read)
 {
 	GSTexture12* vkRt = static_cast<GSTexture12*>(rt);
 	GSTexture12* vkDs = static_cast<GSTexture12*>(ds);
 	pxAssert(vkRt || vkDs);
 
-	if (m_current_render_target != vkRt || m_current_depth_target != vkDs)
+	if (m_current_render_target != vkRt || m_current_depth_target != vkDs || m_current_depth_read_only != depth_read)
 	{
 		// framebuffer change
 		EndRenderPass();
@@ -2214,13 +2214,14 @@ void GSDevice12::OMSetRenderTargets(GSTexture* rt, GSTexture* ds, const GSVector
 
 	m_current_render_target = vkRt;
 	m_current_depth_target = vkDs;
+	m_current_depth_read_only = depth_read;
 
 	if (!InRenderPass())
 	{
 		if (vkRt)
 			vkRt->TransitionToState(D3D12_RESOURCE_STATE_RENDER_TARGET);
 		if (vkDs)
-			vkDs->TransitionToState(D3D12_RESOURCE_STATE_DEPTH_WRITE);
+			vkDs->TransitionToState(depth_read ? (D3D12_RESOURCE_STATE_DEPTH_READ | D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE) : D3D12_RESOURCE_STATE_DEPTH_WRITE);
 	}
 
 	// This is used to set/initialize the framebuffer for tfx rendering.
@@ -3330,6 +3331,7 @@ void GSDevice12::UnbindTexture(GSTexture12* tex)
 	{
 		EndRenderPass();
 		m_current_depth_target = nullptr;
+		m_current_depth_read_only = false;
 	}
 }
 
@@ -3450,7 +3452,7 @@ void GSDevice12::BeginRenderPass(D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE color_b
 	D3D12_RENDER_PASS_DEPTH_STENCIL_DESC ds = {};
 	if (m_current_depth_target)
 	{
-		ds.cpuDescriptor = m_current_depth_target->GetWriteDescriptor();
+		ds.cpuDescriptor = m_current_depth_read_only ? m_current_depth_target->GetReadDepthViewDescriptor() : m_current_depth_target->GetWriteDescriptor();
 		ds.DepthEndingAccess.Type = depth_end;
 		ds.DepthBeginningAccess.Type = depth_begin;
 		if (depth_begin == D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE_CLEAR)
@@ -3470,7 +3472,8 @@ void GSDevice12::BeginRenderPass(D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE color_b
 	}
 
 	GetCommandList()->BeginRenderPass(m_current_render_target ? 1 : 0,
-		m_current_render_target ? &rt : nullptr, m_current_depth_target ? &ds : nullptr, D3D12_RENDER_PASS_FLAG_NONE);
+		m_current_render_target ? &rt : nullptr, m_current_depth_target ? &ds : nullptr,
+		(m_current_depth_target && m_current_depth_read_only) ? (D3D12_RENDER_PASS_FLAG_BIND_READ_ONLY_DEPTH) : D3D12_RENDER_PASS_FLAG_NONE);
 }
 
 void GSDevice12::EndRenderPass()
@@ -3552,11 +3555,13 @@ __ri void GSDevice12::ApplyBaseState(u32 flags, ID3D12GraphicsCommandList* cmdli
 		if (m_current_render_target)
 		{
 			cmdlist->OMSetRenderTargets(1, &m_current_render_target->GetWriteDescriptor().cpu_handle, FALSE,
-				m_current_depth_target ? &m_current_depth_target->GetWriteDescriptor().cpu_handle : nullptr);
+				m_current_depth_target ?
+					(m_current_depth_read_only ? &m_current_depth_target->GetReadDepthViewDescriptor().cpu_handle : &m_current_depth_target->GetWriteDescriptor().cpu_handle) :
+					nullptr);
 		}
 		else if (m_current_depth_target)
 		{
-			cmdlist->OMSetRenderTargets(0, nullptr, FALSE, &m_current_depth_target->GetWriteDescriptor().cpu_handle);
+			cmdlist->OMSetRenderTargets(0, nullptr, FALSE, m_current_depth_read_only ? &m_current_depth_target->GetReadDepthViewDescriptor().cpu_handle : &m_current_depth_target->GetWriteDescriptor().cpu_handle);
 		}
 	}
 }
@@ -3924,15 +3929,6 @@ void GSDevice12::RenderHW(GSHWDrawConfig& config)
 	if (config.blend.constant_enable)
 		SetBlendConstants(config.blend.constant);
 
-	// Depth testing and sampling, bind resource as dsv read only and srv at the same time without the need of a copy.
-	if (config.tex && config.tex == config.ds)
-	{
-		EndRenderPass();
-
-		// Transition dsv as read only.
-		draw_ds->TransitionToState(D3D12_RESOURCE_STATE_DEPTH_READ);
-	}
-
 	// Primitive ID tracking DATE setup.
 	GSTexture12* date_image = nullptr;
 	if (config.destination_alpha == GSHWDrawConfig::DestinationAlphaMode::PrimIDTracking)
@@ -4034,7 +4030,8 @@ void GSDevice12::RenderHW(GSHWDrawConfig& config)
 			Console.Warning("D3D12: Failed to allocate temp texture for RT copy.");
 	}
 
-	OMSetRenderTargets(draw_rt, draw_ds, config.scissor);
+	// For depth testing and sampling, use a read only dsv, otherwise use a write dsv
+	OMSetRenderTargets(draw_rt, draw_ds, config.scissor, config.tex && config.tex == config.ds);
 
 	// Begin render pass if new target or out of the area.
 	if (!m_in_render_pass)

pcsx2/GS/Renderers/DX12/GSDevice12.h

@@ -459,7 +459,7 @@ public:
 	void PSSetShaderResource(int i, GSTexture* sr, bool check_state, bool feedback = false);
 	void PSSetSampler(GSHWDrawConfig::SamplerSelector sel);
 
-	void OMSetRenderTargets(GSTexture* rt, GSTexture* ds, const GSVector4i& scissor);
+	void OMSetRenderTargets(GSTexture* rt, GSTexture* ds, const GSVector4i& scissor, bool depth_read = false);
 
 	void SetVSConstantBuffer(const GSHWDrawConfig::VSConstantBuffer& cb);
 	void SetPSConstantBuffer(const GSHWDrawConfig::PSConstantBuffer& cb);
@@ -580,6 +580,7 @@ private:
 
 	GSTexture12* m_current_render_target = nullptr;
 	GSTexture12* m_current_depth_target = nullptr;
+	bool m_current_depth_read_only = false;
 
 	D3D12_VIEWPORT m_viewport = {0.0f, 0.0f, 1.0f, 1.0f, 0.0f, 1.0f};
 	GSVector4i m_scissor = GSVector4i::zero();

pcsx2/GS/Renderers/DX12/GSTexture12.cpp

@@ -17,13 +17,15 @@
 GSTexture12::GSTexture12(Type type, Format format, int width, int height, int levels, DXGI_FORMAT dxgi_format,
 	wil::com_ptr_nothrow<ID3D12Resource> resource, wil::com_ptr_nothrow<ID3D12Resource> resource_fbl,
 	wil::com_ptr_nothrow<D3D12MA::Allocation> allocation, const D3D12DescriptorHandle& srv_descriptor,
-	const D3D12DescriptorHandle& write_descriptor, const D3D12DescriptorHandle& uav_descriptor,
-	const D3D12DescriptorHandle& fbl_descriptor, WriteDescriptorType wdtype, D3D12_RESOURCE_STATES resource_state)
+	const D3D12DescriptorHandle& write_descriptor, const D3D12DescriptorHandle& ro_dsv_descriptor,
+	const D3D12DescriptorHandle& uav_descriptor, const D3D12DescriptorHandle& fbl_descriptor,
+	WriteDescriptorType wdtype, D3D12_RESOURCE_STATES resource_state)
 	: m_resource(std::move(resource))
 	, m_resource_fbl(std::move(resource_fbl))
 	, m_allocation(std::move(allocation))
 	, m_srv_descriptor(srv_descriptor)
 	, m_write_descriptor(write_descriptor)
+	, m_read_dsv_descriptor(ro_dsv_descriptor)
 	, m_uav_descriptor(uav_descriptor)
 	, m_fbl_descriptor(fbl_descriptor)
 	, m_write_descriptor_type(wdtype)
@@ -58,6 +60,7 @@ void GSTexture12::Destroy(bool defer)
 				break;
 			case WriteDescriptorType::DSV:
 				dev->DeferDescriptorDestruction(dev->GetDSVHeapManager(), &m_write_descriptor);
+				dev->DeferDescriptorDestruction(dev->GetDSVHeapManager(), &m_read_dsv_descriptor);
 				break;
 			case WriteDescriptorType::None:
 			default:
@@ -87,6 +90,7 @@ void GSTexture12::Destroy(bool defer)
 				break;
 			case WriteDescriptorType::DSV:
 				dev->GetDSVHeapManager().Free(&m_write_descriptor);
+				dev->GetDSVHeapManager().Free(&m_read_dsv_descriptor);
 				break;
 			case WriteDescriptorType::None:
 			default:
@@ -237,7 +241,7 @@ std::unique_ptr<GSTexture12> GSTexture12::Create(Type type, Format format, int w
 		}
 	}
 
-	D3D12DescriptorHandle srv_descriptor, write_descriptor, uav_descriptor, fbl_descriptor;
+	D3D12DescriptorHandle srv_descriptor, write_descriptor, ro_dsv_descriptor, uav_descriptor, fbl_descriptor;
 	WriteDescriptorType write_descriptor_type = WriteDescriptorType::None;
 	if (srv_format != DXGI_FORMAT_UNKNOWN)
 	{
@@ -261,11 +265,17 @@ std::unique_ptr<GSTexture12> GSTexture12::Create(Type type, Format format, int w
 		case Type::DepthStencil:
 		{
 			write_descriptor_type = WriteDescriptorType::DSV;
-			if (!CreateDSVDescriptor(resource.get(), dsv_format, &write_descriptor))
+			if (!CreateDSVDescriptor(resource.get(), dsv_format, &write_descriptor, false))
 			{
 				dev->GetDescriptorHeapManager().Free(&srv_descriptor);
 				return {};
 			}
+			if (!CreateDSVDescriptor(resource.get(), dsv_format, &ro_dsv_descriptor, true))
+			{
+				dev->GetDSVHeapManager().Free(&write_descriptor);
+				dev->GetDescriptorHeapManager().Free(&srv_descriptor);
+				return {};
+			}
 		}
 		break;
 
@@ -281,6 +291,7 @@ std::unique_ptr<GSTexture12> GSTexture12::Create(Type type, Format format, int w
 				dev->GetRTVHeapManager().Free(&write_descriptor);
 				break;
 			case WriteDescriptorType::DSV:
+				dev->GetDSVHeapManager().Free(&ro_dsv_descriptor);
 				dev->GetDSVHeapManager().Free(&write_descriptor);
 				break;
 			default:
@@ -301,6 +312,7 @@ std::unique_ptr<GSTexture12> GSTexture12::Create(Type type, Format format, int w
 					dev->GetRTVHeapManager().Free(&write_descriptor);
 					break;
 				case WriteDescriptorType::DSV:
+					dev->GetDSVHeapManager().Free(&ro_dsv_descriptor);
 					dev->GetDSVHeapManager().Free(&write_descriptor);
 					break;
 				default:
@@ -313,7 +325,7 @@ std::unique_ptr<GSTexture12> GSTexture12::Create(Type type, Format format, int w
 
 	return std::unique_ptr<GSTexture12>(
 		new GSTexture12(type, format, width, height, levels, dxgi_format, std::move(resource), std::move(resource_fbl), std::move(allocation),
-			srv_descriptor, write_descriptor, uav_descriptor, fbl_descriptor, write_descriptor_type, state));
+			srv_descriptor, write_descriptor, ro_dsv_descriptor, uav_descriptor, fbl_descriptor, write_descriptor_type, state));
 }
 
 std::unique_ptr<GSTexture12> GSTexture12::Adopt(wil::com_ptr_nothrow<ID3D12Resource> resource, Type type, Format format,
@@ -322,7 +334,7 @@ std::unique_ptr<GSTexture12> GSTexture12::Adopt(wil::com_ptr_nothrow<ID3D12Resou
 {
 	const D3D12_RESOURCE_DESC desc = resource->GetDesc();
 
-	D3D12DescriptorHandle srv_descriptor, write_descriptor, uav_descriptor;
+	D3D12DescriptorHandle srv_descriptor, write_descriptor, ro_dsv_descriptor, uav_descriptor;
 	WriteDescriptorType write_descriptor_type = WriteDescriptorType::None;
 	if (srv_format != DXGI_FORMAT_UNKNOWN)
 	{
@@ -342,11 +354,17 @@ std::unique_ptr<GSTexture12> GSTexture12::Adopt(wil::com_ptr_nothrow<ID3D12Resou
 	else if (type == Type::DepthStencil)
 	{
 		write_descriptor_type = WriteDescriptorType::DSV;
-		if (!CreateDSVDescriptor(resource.get(), dsv_format, &write_descriptor))
+		if (!CreateDSVDescriptor(resource.get(), dsv_format, &write_descriptor, false))
 		{
 			GSDevice12::GetInstance()->GetDescriptorHeapManager().Free(&srv_descriptor);
 			return {};
 		}
+		if (!CreateDSVDescriptor(resource.get(), dsv_format, &ro_dsv_descriptor, true))
+		{
+			GSDevice12::GetInstance()->GetDSVHeapManager().Free(&write_descriptor);
+			GSDevice12::GetInstance()->GetDescriptorHeapManager().Free(&srv_descriptor);
+			return {};
+		}
 	}
 
 	if (uav_format != DXGI_FORMAT_UNKNOWN)
@@ -359,6 +377,7 @@ std::unique_ptr<GSTexture12> GSTexture12::Adopt(wil::com_ptr_nothrow<ID3D12Resou
 					GSDevice12::GetInstance()->GetRTVHeapManager().Free(&write_descriptor);
 					break;
 				case WriteDescriptorType::DSV:
+					GSDevice12::GetInstance()->GetDSVHeapManager().Free(&ro_dsv_descriptor);
 					GSDevice12::GetInstance()->GetDSVHeapManager().Free(&write_descriptor);
 					break;
 				default:
@@ -371,7 +390,7 @@ std::unique_ptr<GSTexture12> GSTexture12::Adopt(wil::com_ptr_nothrow<ID3D12Resou
 	}
 
 	return std::unique_ptr<GSTexture12>(new GSTexture12(type, format, static_cast<u32>(desc.Width), desc.Height,
-		desc.MipLevels, desc.Format, std::move(resource), {}, {}, srv_descriptor, write_descriptor, uav_descriptor,
+		desc.MipLevels, desc.Format, std::move(resource), {}, {}, srv_descriptor, write_descriptor, {}, uav_descriptor,
 		{}, write_descriptor_type, resource_state));
 }
 
@@ -405,7 +424,7 @@ bool GSTexture12::CreateRTVDescriptor(ID3D12Resource* resource, DXGI_FORMAT form
 	return true;
 }
 
-bool GSTexture12::CreateDSVDescriptor(ID3D12Resource* resource, DXGI_FORMAT format, D3D12DescriptorHandle* dh)
+bool GSTexture12::CreateDSVDescriptor(ID3D12Resource* resource, DXGI_FORMAT format, D3D12DescriptorHandle* dh, bool read_only)
 {
 	if (!GSDevice12::GetInstance()->GetDSVHeapManager().Allocate(dh))
 	{
@@ -413,7 +432,7 @@ bool GSTexture12::CreateDSVDescriptor(ID3D12Resource* resource, DXGI_FORMAT form
 		return false;
 	}
 
-	const D3D12_DEPTH_STENCIL_VIEW_DESC desc = {format, D3D12_DSV_DIMENSION_TEXTURE2D, D3D12_DSV_FLAG_NONE};
+	const D3D12_DEPTH_STENCIL_VIEW_DESC desc = {format, D3D12_DSV_DIMENSION_TEXTURE2D, read_only ? D3D12_DSV_FLAG_READ_ONLY_DEPTH : D3D12_DSV_FLAG_NONE };
 	GSDevice12::GetInstance()->GetDevice()->CreateDepthStencilView(resource, &desc, dh->cpu_handle);
 	return true;
 }
@@ -708,7 +727,36 @@ void GSTexture12::TransitionToState(ID3D12GraphicsCommandList* cmdlist, D3D12_RE
 	if (m_resource_state == state)
 		return;
 
-	TransitionSubresourceToState(cmdlist, D3D12_RESOURCE_BARRIER_ALL_SUBRESOURCES, m_resource_state, state);
+	// Read only depth requires special handling as we might want to write stencil.
+	// Also batch the transition barriers as per recommendation from docs.
+	if (state == (D3D12_RESOURCE_STATE_DEPTH_READ | D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE))
+	{
+		// Transition to read depth/write stencil
+		const D3D12_RESOURCE_BARRIER barriers[2] = {
+			{D3D12_RESOURCE_BARRIER_TYPE_TRANSITION, D3D12_RESOURCE_BARRIER_FLAG_NONE,
+				{{m_resource.get(), 0, m_resource_state, (D3D12_RESOURCE_STATE_DEPTH_READ | D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE)}}},
+			{D3D12_RESOURCE_BARRIER_TYPE_TRANSITION, D3D12_RESOURCE_BARRIER_FLAG_NONE,
+				{{m_resource.get(), 1, m_resource_state, D3D12_RESOURCE_STATE_DEPTH_WRITE}}},
+		};
+		GSDevice12::GetInstance()->GetCommandList()->ResourceBarrier(m_resource_state == D3D12_RESOURCE_STATE_DEPTH_WRITE ? 1 : 2, barriers);
+	}
+	else if (m_resource_state == (D3D12_RESOURCE_STATE_DEPTH_READ | D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE))
+	{
+		// Transition from read depth/write stencil
+		const D3D12_RESOURCE_BARRIER barriers[2] = {
+			{D3D12_RESOURCE_BARRIER_TYPE_TRANSITION, D3D12_RESOURCE_BARRIER_FLAG_NONE,
+				{{m_resource.get(), 0, (D3D12_RESOURCE_STATE_DEPTH_READ | D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE), state}}},
+			{D3D12_RESOURCE_BARRIER_TYPE_TRANSITION, D3D12_RESOURCE_BARRIER_FLAG_NONE,
+				{{m_resource.get(), 1, D3D12_RESOURCE_STATE_DEPTH_WRITE, state}}},
+		};
+		GSDevice12::GetInstance()->GetCommandList()->ResourceBarrier(state == D3D12_RESOURCE_STATE_DEPTH_WRITE ? 1 : 2, barriers);
+	}
+	else
+	{
+		// Normal transition
+		TransitionSubresourceToState(cmdlist, D3D12_RESOURCE_BARRIER_ALL_SUBRESOURCES, m_resource_state, state);
+	}
+
 	m_resource_state = state;
 }
 

pcsx2/GS/Renderers/DX12/GSTexture12.h

@@ -30,6 +30,7 @@ public:
 
 	__fi const D3D12DescriptorHandle& GetSRVDescriptor() const { return m_srv_descriptor; }
 	__fi const D3D12DescriptorHandle& GetWriteDescriptor() const { return m_write_descriptor; }
+	__fi const D3D12DescriptorHandle& GetReadDepthViewDescriptor() const { return m_read_dsv_descriptor; }
 	__fi const D3D12DescriptorHandle& GetUAVDescriptor() const { return m_uav_descriptor; }
 	__fi const D3D12DescriptorHandle& GetFBLDescriptor() const { return m_fbl_descriptor; }
 	__fi D3D12_RESOURCE_STATES GetResourceState() const { return m_resource_state; }
@@ -72,13 +73,14 @@ private:
 	GSTexture12(Type type, Format format, int width, int height, int levels, DXGI_FORMAT dxgi_format,
 		wil::com_ptr_nothrow<ID3D12Resource> resource, wil::com_ptr_nothrow<ID3D12Resource> resource_fbl,
 		wil::com_ptr_nothrow<D3D12MA::Allocation> allocation, const D3D12DescriptorHandle& srv_descriptor,
-		const D3D12DescriptorHandle& write_descriptor, const D3D12DescriptorHandle& uav_descriptor,
-		const D3D12DescriptorHandle& fbl_descriptor, WriteDescriptorType wdtype, D3D12_RESOURCE_STATES resource_state);
+		const D3D12DescriptorHandle& write_descriptor, const D3D12DescriptorHandle& ro_dsv_descriptor,
+		const D3D12DescriptorHandle& uav_descriptor, const D3D12DescriptorHandle& fbl_descriptor,
+		WriteDescriptorType wdtype, D3D12_RESOURCE_STATES resource_state);
 
 	static bool CreateSRVDescriptor(
 		ID3D12Resource* resource, u32 levels, DXGI_FORMAT format, D3D12DescriptorHandle* dh);
 	static bool CreateRTVDescriptor(ID3D12Resource* resource, DXGI_FORMAT format, D3D12DescriptorHandle* dh);
-	static bool CreateDSVDescriptor(ID3D12Resource* resource, DXGI_FORMAT format, D3D12DescriptorHandle* dh);
+	static bool CreateDSVDescriptor(ID3D12Resource* resource, DXGI_FORMAT format, D3D12DescriptorHandle* dh, bool read_only);
 	static bool CreateUAVDescriptor(ID3D12Resource* resource, DXGI_FORMAT format, D3D12DescriptorHandle* dh);
 
 	ID3D12GraphicsCommandList* GetCommandBufferForUpdate();
@@ -91,6 +93,7 @@ private:
 
 	D3D12DescriptorHandle m_srv_descriptor = {};
 	D3D12DescriptorHandle m_write_descriptor = {};
+	D3D12DescriptorHandle m_read_dsv_descriptor = {};
 	D3D12DescriptorHandle m_uav_descriptor = {};
 	D3D12DescriptorHandle m_fbl_descriptor = {};
 	WriteDescriptorType m_write_descriptor_type = WriteDescriptorType::None;