rsx: Synchronization improvements

rpcs3/Emu/RSX/Host/MM.cpp

@@ -2,6 +2,7 @@
 #include "MM.h"
 #include <Emu/RSX/Common/simple_array.hpp>
 #include <Emu/RSX/RSXOffload.h>
+#include <Emu/RSX/rsx_utils.h>
 
 #include <Emu/Memory/vm.h>
 #include <Emu/IdManager.h>
@@ -120,7 +121,7 @@ namespace rsx
 			return;
 		}
 
-		g_deferred_mprotect_queue.push_back({ range, prot });
+		g_deferred_mprotect_queue.push_back({ range, prot, rsx::get_shared_tag() });
 	}
 
 	void mm_protect(void* ptr, u64 length, utils::protection prot)
@@ -221,4 +222,24 @@ namespace rsx
 		auto& rsxdma = g_fxo->get<rsx::dma_manager>();
 		rsxdma.backend_ctrl(mm_backend_ctrl::cmd_mm_flush, nullptr);
 	}
+
+	void mm_flush_partial(u64 last_tag)
+	{
+		std::lock_guard lock(g_mprotect_queue_lock);
+
+		u32 count = 0;
+		for (const auto& block : g_deferred_mprotect_queue)
+		{
+			if (block.sync_tag > last_tag)
+			{
+				break;
+			}
+			count++;
+		}
+
+		if (count)
+		{
+			mm_flush_mprotect_queue_internal(count);
+		}
+	}
 }

rpcs3/Emu/RSX/Host/MM.h

@@ -12,6 +12,7 @@ namespace rsx
 	{
 		utils::address_range64 range;
 		utils::protection prot;
+		u64 sync_tag;
 
 		inline void merge(const utils::address_range64& other)
 		{
@@ -47,5 +48,6 @@ namespace rsx
 	void mm_flush_lazy();
 	void mm_flush(u32 vm_address);
 	void mm_flush(const rsx::simple_array<utils::address_range64>& ranges);
+	void mm_flush_partial(u64 tag);
 	void mm_flush();
 }

rpcs3/Emu/RSX/NV47/HW/nv0039.cpp

@@ -107,7 +107,7 @@ namespace rsx
 				result == rsx::result_zcull_intr)
 			{
 				// This transfer overlaps will zcull data pool
-				if (RSX(ctx)->copy_zcull_stats(read_address, read_length, write_address) == write_length)
+				if (RSX(ctx)->copy_zcull_stats(read_address, read_length, write_address) >= write_length)
 				{
 					// All writes deferred
 					return;
@@ -126,7 +126,7 @@ namespace rsx
 			u8* dst = vm::_ptr<u8>(write_address);
 			const u8* src = vm::_ptr<u8>(read_address);
 
-			rsx::simple_array<utils::address_range64> flush_mm_ranges =
+			const rsx::simple_array<utils::address_range64> flush_mm_ranges =
 			{
 				utils::address_range64::start_length(reinterpret_cast<u64>(dst), write_length),
 				utils::address_range64::start_length(reinterpret_cast<u64>(src), read_length)

rpcs3/Emu/RSX/NV47/HW/nv3089.cpp

@@ -217,7 +217,7 @@ namespace rsx
 				if (const auto result = RSX(ctx)->read_barrier(src_address, data_length, false);
 					result == rsx::result_zcull_intr)
 				{
-					if (RSX(ctx)->copy_zcull_stats(src_address, data_length, dst_address) == data_length)
+					if (RSX(ctx)->copy_zcull_stats(src_address, data_length, dst_address) >= data_length)
 					{
 						// All writes deferred
 						return { false, src_info, dst_info };

rpcs3/Emu/RSX/NV47/HW/nv4097.cpp

@@ -4,6 +4,7 @@
 
 #include "Emu/RSX/RSXThread.h"
 #include "Emu/RSX/Common/BufferUtils.h"
+#include "Emu/RSX/Host/MM.h"
 
 #define RSX(ctx) ctx->rsxthr
 #define REGS(ctx) (&rsx::method_registers)
@@ -719,9 +720,11 @@ namespace rsx
 			}
 
 			const u32 addr = RSX(ctx)->iomap_table.get_addr(0xf100000 + (index * 0x40));
-
 			ensure(addr != umax);
 
+			// Notify ticks are strongly ordered
+			RSX(ctx)->sync();
+
 			vm::_ptr<atomic_t<RsxNotify>>(addr)->store(
 			{
 				RSX(ctx)->timestamp(),

rpcs3/Emu/RSX/NV47/HW/nv47_sync.hpp

@@ -24,17 +24,18 @@ namespace rsx
 				if (vm::_ref<RsxSemaphore>(address) == data)
 				{
 					// It's a no-op to write the same value (although there is a delay in real-hw so it's more accurate to allow GPU label in this case)
+					// There is no possible way for the guest to know that the label has been processed so we can skip MM sync here.
 					return;
 				}
 
 				if constexpr (FlushDMA || FlushPipe)
 				{
-					// Release op must be acoompanied by MM flush.
-					// FlushPipe implicitly does a MM flush but FlushDMA does not. Trigger the flush here
-					rsx::mm_flush();
-
 					if constexpr (FlushDMA)
 					{
+						// Release op must be acoompanied by MM flush.
+						// FlushPipe implicitly does a MM flush but FlushDMA does not. Trigger the flush here
+						rsx::mm_flush();
+
 						// If the backend handled the request, this call will basically be a NOP
 						g_fxo->get<rsx::dma_manager>().sync();
 					}

rpcs3/Emu/RSX/RSXZCULL.cpp

@@ -1,6 +1,7 @@
 #include "stdafx.h"
 #include "Core/RSXEngLock.hpp"
 #include "Core/RSXReservationLock.hpp"
+#include "Host/MM.h"
 #include "RSXThread.h"
 
 namespace rsx
@@ -155,6 +156,9 @@ namespace rsx
 
 				if (m_pending_writes.empty())
 				{
+					// Immediate write, flush MM queue
+					rsx::mm_flush();
+
 					// No need to queue this if there is no pending request in the pipeline anyway
 					write(sink, ptimer->timestamp(), type, m_statistics_map[m_statistics_tag_id].result);
 					return;
@@ -173,6 +177,7 @@ namespace rsx
 					It->counter_tag = m_statistics_tag_id;
 					It->sink = sink;
 					It->type = type;
+					It->sync_tag = rsx::get_shared_tag();
 
 					if (forwarder != &(*It))
 					{
@@ -369,6 +374,9 @@ namespace rsx
 
 		void ZCULL_control::write(queued_report_write* writer, u64 timestamp, u32 value)
 		{
+			// Reports are strongly ordered.
+			rsx::mm_flush_partial(writer->sync_tag);
+
 			write(writer->sink, timestamp, writer->type, value);
 			on_report_completed(writer->sink);
 
@@ -799,6 +807,8 @@ namespace rsx
 		u32 ZCULL_control::copy_reports_to(u32 start, u32 range, u32 dest)
 		{
 			u32 bytes_to_write = 0;
+			std::unordered_set<u32> unique_addresses;
+
 			const auto memory_range = utils::address_range32::start_length(start, range);
 			for (auto& writer : m_pending_writes)
 			{
@@ -807,8 +817,14 @@ namespace rsx
 
 				if (!writer.forwarder && memory_range.overlaps(writer.sink))
 				{
-					u32 address = (writer.sink - start) + dest;
+					const u32 address = (writer.sink - start) + dest;
 					writer.sink_alias.push_back(vm::cast(address));
+
+					if (!unique_addresses.contains(address))
+					{
+						bytes_to_write += sizeof(RsxReport);
+						unique_addresses.insert(address);
+					}
 				}
 			}
 

rpcs3/Emu/RSX/RSXZCULL.h

@@ -46,6 +46,7 @@ namespace rsx
 		{
 			u32 type = CELL_GCM_ZPASS_PIXEL_CNT;
 			u32 counter_tag;
+			u64 sync_tag;
 			occlusion_query_info* query;
 			queued_report_write* forwarder;
 

rpcs3/tests/rpcs3_test.vcxproj

@@ -35,7 +35,7 @@
   <PropertyGroup>
     <PreferredToolArchitecture>x64</PreferredToolArchitecture>
     <IntDir>$(SolutionDir)build\tmp\$(ProjectName)-$(Configuration)-$(Platform)\</IntDir>
-    <IncludePath>.\;$(SolutionDir);$(SolutionDir)rpcs3;$(VC_IncludePath);$(WindowsSDK_IncludePath);$(UniversalCRT_IncludePath);</IncludePath>
+    <IncludePath>.\;$(SolutionDir);$(SolutionDir)rpcs3;$(SolutionDir)3rdparty/ffmpeg/include;$(VC_IncludePath);$(WindowsSDK_IncludePath);$(UniversalCRT_IncludePath);</IncludePath>
     <LibraryPath>$(SolutionDir)build\lib\$(Configuration)-$(Platform)\;$(UniversalCRT_LibraryPath_x64);$(LibraryPath)</LibraryPath>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">