rsx: initial capture/replay functionality (#4510)

* rsx: initial capture/replay functionality

.gitmodules

@@ -42,3 +42,6 @@
 [submodule "3rdparty/pugixml"]
 	path = 3rdparty/pugixml
 	url = https://github.com/zeux/pugixml
+[submodule "3rdparty/xxHash"]
+	path = 3rdparty/xxHash
+	url = https://github.com/Cyan4973/xxHash

.travis.yml

@@ -47,7 +47,7 @@ install:
     fi;
 
 before_script:
-  - git submodule update --init asmjit 3rdparty/ffmpeg 3rdparty/pugixml 3rdparty/GSL 3rdparty/libpng Utilities/yaml-cpp 3rdparty/cereal 3rdparty/hidapi 3rdparty/Optional Vulkan/glslang Vulkan/Vulkan-LoaderAndValidationLayers
+  - git submodule update --init asmjit 3rdparty/ffmpeg 3rdparty/pugixml 3rdparty/GSL 3rdparty/libpng Utilities/yaml-cpp 3rdparty/cereal 3rdparty/hidapi 3rdparty/Optional 3rdparty/xxHash Vulkan/glslang Vulkan/Vulkan-LoaderAndValidationLayers
   - mkdir build ; cd build
   - export CMAKE_PREFIX_PATH=~/Qt/${QTVER}/gcc_64/lib/cmake
   - if [ "$TRAVIS_PULL_REQUEST" = false ]; then

3rdparty/xxHash

@@ -0,0 +1 @@
+Subproject commit 3064d42e7d74b0921bdd1818395d9cb37bb8976a

3rdparty/xxhash.vcxproj

@@ -0,0 +1,59 @@
+<?xml version="1.0" encoding="utf-8"?>
+<Project DefaultTargets="Build" ToolsVersion="15.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+  <ItemGroup Label="ProjectConfigurations">
+    <ProjectConfiguration Include="Debug|x64">
+      <Configuration>Debug</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|x64">
+      <Configuration>Release</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+  </ItemGroup>
+  <PropertyGroup Label="Globals">
+    <ProjectGuid>{939FE206-1182-ABC3-1234-FEAB88E98404}</ProjectGuid>
+    <WindowsTargetPlatformVersion>8.1</WindowsTargetPlatformVersion>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
+  <PropertyGroup Label="Configuration">
+    <ConfigurationType>StaticLibrary</ConfigurationType>
+    <PlatformToolset>v140</PlatformToolset>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Debug'" Label="Configuration">
+    <UseDebugLibraries>true</UseDebugLibraries>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Release'" Label="Configuration">
+    <UseDebugLibraries>false</UseDebugLibraries>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
+  <ImportGroup Label="ExtensionSettings">
+  </ImportGroup>
+    <ImportGroup Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+    <Import Project="..\rpcs3_default.props" />
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets" Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <Import Project="..\rpcs3_debug.props" />
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets" Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <Import Project="..\rpcs3_release.props" />
+  </ImportGroup>
+  <PropertyGroup Label="UserMacros" />
+  <PropertyGroup />
+  <ItemDefinitionGroup>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <ClCompile Include="xxHash/xxhash.c" />
+  </ItemGroup>
+  <ItemGroup>
+    <Text Include="LICENSE" />
+    <Text Include="CMakeLists.txt" />
+  </ItemGroup>
+  <ItemGroup>
+    <ClInclude Include="xxHash/xxhash.h" />
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+  </ImportGroup>
+</Project>

CMakeLists.txt

@@ -44,6 +44,7 @@ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_RELWITHDEBINFO "${PROJECT_BINARY_DIR}/bin")
 
 add_subdirectory( Vulkan )
 add_subdirectory( rpcs3 )
+add_subdirectory(3rdparty/xxHash/cmake_unofficial)
 
 include_directories(3rdparty/hidapi/hidapi)
 if(APPLE)

appveyor.yml

@@ -46,6 +46,7 @@ install:
       3rdparty/libpng `
       3rdparty/Optional `
       3rdparty/pugixml `
+      3rdparty/xxHash `
       3rdparty/zlib `
       asmjit `
       Utilities/yaml-cpp `

rpcs3.sln

@@ -74,6 +74,10 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "hidapi", "hidapi", "{FA1E6C
 EndProject
 Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "hidapi", "3rdparty\hidapi\windows\hidapi.vcxproj", "{A107C21C-418A-4697-BB10-20C3AA60E2E4}"
 EndProject
+Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "xxhash", "xxhash", "{D16E245C-CC5A-4B9A-8BAB-1176F02C1631}"
+EndProject
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "xxhash", "3rdparty\xxhash.vcxproj", "{939FE206-1182-ABC3-1234-FEAB88E98404}"
+EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
 		Debug - LLVM|x64 = Debug - LLVM|x64
@@ -223,6 +227,16 @@ Global
 		{A107C21C-418A-4697-BB10-20C3AA60E2E4}.Release - LLVM|x64.Build.0 = Release|x64
 		{A107C21C-418A-4697-BB10-20C3AA60E2E4}.Release|x64.ActiveCfg = Release|x64
 		{A107C21C-418A-4697-BB10-20C3AA60E2E4}.Release|x64.Build.0 = Release|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Debug - LLVM|x64.ActiveCfg = Debug|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Debug - LLVM|x64.Build.0 = Debug|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Debug - MemLeak|x64.ActiveCfg = Debug|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Debug - MemLeak|x64.Build.0 = Debug|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Debug|x64.ActiveCfg = Debug|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Debug|x64.Build.0 = Debug|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Release - LLVM|x64.ActiveCfg = Release|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Release - LLVM|x64.Build.0 = Release|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Release|x64.ActiveCfg = Release|x64
+		{939FE206-1182-ABC3-1234-FEAB88E98404}.Release|x64.Build.0 = Release|x64
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE
@@ -242,5 +256,6 @@ Global
 		{3EE5F075-B546-42C4-B6A8-E3CCEF38B78D} = {10FBF193-D532-4CCF-B875-4C7091A7F6C2}
 		{FDC361C5-7734-493B-8CFB-037308B35122} = {DDF904CA-2771-441A-8629-5DF2EB922A79}
 		{A107C21C-418A-4697-BB10-20C3AA60E2E4} = {FA1E6C16-CA63-45F8-8D52-E21DF396BE36}
+		{939FE206-1182-ABC3-1234-FEAB88E98404} = {D16E245C-CC5A-4B9A-8BAB-1176F02C1631}
 	EndGlobalSection
 EndGlobal

rpcs3/CMakeLists.txt

@@ -300,6 +300,7 @@ ${LLVM_INCLUDE_DIRS}
 "${RPCS3_SRC_DIR}/../3rdparty/cereal/include"
 "${RPCS3_SRC_DIR}/../3rdparty/Optional"
 "${RPCS3_SRC_DIR}/../3rdparty/discord-rpc/include"
+"${RPCS3_SRC_DIR}/../3rdparty/xxHash"
 )
 
 if(WIN32)
@@ -408,6 +409,8 @@ if(UNIX)
 	target_link_libraries(rpcs3 ${X11_LIBRARIES})
 endif()
 
+target_link_libraries(rpcs3 xxhash)
+
 if(WIN32)
 	target_link_libraries(rpcs3 ws2_32.lib Winmm.lib Psapi.lib gdi32.lib VKstatic.1 glslang OSDependent OGLCompiler SPIRV HLSL setupapi.lib hidapi-hid Shlwapi.lib)
 	if(NOT MSVC)

rpcs3/Emu/RSX/Capture/rsx_capture.cpp

@@ -0,0 +1,627 @@
+#include "stdafx.h"
+#include "rsx_capture.h"
+#include "Emu/RSX/Common/BufferUtils.h"
+#include "Emu/RSX/Common/TextureUtils.h"
+#include "Emu/RSX/Common/surface_store.h"
+#include "Emu/RSX/GCM.h"
+#include "Emu/RSX/RSXThread.h"
+#include "Emu/Memory/Memory.h"
+
+#include "xxhash.h"
+
+namespace rsx
+{
+	namespace capture
+	{
+		u32 get_io_offset(u32 offset, u32 location)
+		{
+			switch (location)
+			{
+			case CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER:
+			case CELL_GCM_LOCATION_MAIN:
+			{
+				if (u32 result = RSXIOMem.RealAddr(offset))
+				{
+					return offset;
+				}
+			}
+			case CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN:
+			{
+				if (u32 result = RSXIOMem.RealAddr(0x0e000000 + offset))
+				{
+					return 0x0e000000 + offset;
+				}
+			}
+			default: return 0xFFFFFFFFu;
+			}
+		}
+
+		void insert_mem_block_in_map(std::unordered_set<u64>& mem_changes, frame_capture_data::memory_block&& block, frame_capture_data::memory_block_data&& data)
+		{
+			u64 data_hash = 0;
+			if (data.data.size() > 0)
+			{
+				data_hash = XXH64(data.data.data(), data.data.size(), 0);
+				// using 0 to signify no block in use, so this one is 'reserved'
+				if (data_hash == 0)
+					fmt::throw_exception("Memory block data hash equal to 0");
+
+				block.size       = data.data.size();
+				block.data_state = data_hash;
+
+				auto it = frame_capture.memory_data_map.find(data_hash);
+				if (it != frame_capture.memory_data_map.end())
+				{
+					if (it->second.data != data.data)
+						// screw this
+						fmt::throw_exception("Memory map hash collision detected...cant capture");
+				}
+				else
+					frame_capture.memory_data_map.insert(std::make_pair(data_hash, std::move(data)));
+			}
+
+			u64 block_hash = XXH64(&block, sizeof(frame_capture_data::memory_block), 0);
+			mem_changes.insert(block_hash);
+			if (frame_capture.memory_map.find(block_hash) == frame_capture.memory_map.end())
+				frame_capture.memory_map.insert(std::make_pair(block_hash, std::move(block)));
+		}
+
+		void capture_draw_memory(thread* rsx)
+		{
+			// the idea here is to copy any memory that is needed to make the calls work
+			// todo:
+			//  - tile / zcull state changing during other commands
+			//  - track memory that is rendered into and ignore saving it later, this one will be tough
+
+			if (frame_capture.replay_commands.empty())
+				fmt::throw_exception("no replay commands to attach memory state to");
+
+			// shove the mem_changes onto the last issued command
+			std::unordered_set<u64>& mem_changes = frame_capture.replay_commands.back().memory_state;
+
+			// capture fragment shader mem
+			const u32 shader_program = method_registers.shader_program_address();
+			if (shader_program != 0)
+			{
+				const u32 program_location = (shader_program & 0x3) - 1;
+				const u32 program_offset   = (shader_program & ~0x3);
+
+				const u32 addr          = get_address(program_offset, program_location);
+				const u32 program_start = program_hash_util::fragment_program_utils::get_fragment_program_start(vm::base(addr));
+				const u32 ucode_size    = program_hash_util::fragment_program_utils::get_fragment_program_ucode_size(vm::base(addr + program_start));
+
+				frame_capture_data::memory_block block;
+				block.addr     = addr;
+				block.ioOffset = get_io_offset(program_offset, program_location);
+				frame_capture_data::memory_block_data block_data;
+				block_data.data.resize(ucode_size + program_start);
+				std::memcpy(block_data.data.data(), vm::base(addr), ucode_size + program_start);
+				insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
+			}
+
+			// vertex shader is passed in registers, so it can be ignored
+
+			// save fragment tex mem
+			for (const auto& tex : method_registers.fragment_textures)
+			{
+				if (!tex.enabled())
+					continue;
+
+				const u32 texaddr = get_address(tex.offset(), tex.location());
+				auto layout       = get_subresources_layout(tex);
+
+				// todo: dont use this function and just get size somehow
+				size_t texSize = 0;
+				for (const auto& l : layout)
+					texSize += l.data.size();
+
+				if (!texaddr || !texSize)
+					continue;
+
+				frame_capture_data::memory_block block;
+				block.addr     = texaddr;
+				block.ioOffset = get_io_offset(tex.offset(), tex.location());
+
+				frame_capture_data::memory_block_data block_data;
+				block_data.data.resize(texSize);
+				std::memcpy(block_data.data.data(), vm::base(texaddr), texSize);
+				insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
+			}
+
+			// save vertex texture mem
+			for (const auto& tex : method_registers.vertex_textures)
+			{
+				if (!tex.enabled())
+					continue;
+
+				const u32 texaddr = get_address(tex.offset(), tex.location());
+				auto layout       = get_subresources_layout(tex);
+
+				// todo: dont use this function and just get size somehow
+				size_t texSize = 0;
+				for (const auto& l : layout)
+					texSize += l.data.size();
+
+				if (!texaddr || !texSize)
+					continue;
+
+				frame_capture_data::memory_block block;
+				block.addr     = texaddr;
+				block.ioOffset = get_io_offset(tex.offset(), tex.location());
+				frame_capture_data::memory_block_data block_data;
+				block_data.data.resize(texSize);
+				std::memcpy(block_data.data.data(), vm::base(texaddr), texSize);
+				insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
+			}
+
+			// save vertex buffer memory
+			if (method_registers.current_draw_clause.command == draw_command::array)
+			{
+				const u32 input_mask = method_registers.vertex_attrib_input_mask();
+				for (u8 index = 0; index < limits::vertex_count; ++index)
+				{
+					const bool enabled = !!(input_mask & (1 << index));
+					if (!enabled)
+						continue;
+
+					const auto& info = method_registers.vertex_arrays_info[index];
+					if (!info.size())
+						continue;
+
+					// vert buffer
+					const u32 base_address    = get_vertex_offset_from_base(method_registers.vertex_data_base_offset(), info.offset() & 0x7fffffff);
+					const u32 memory_location = info.offset() >> 31;
+
+					const u32 addr       = get_address(base_address, memory_location);
+					const u32 vertSize   = get_vertex_type_size_on_host(info.type(), info.size());
+					const u32 vertStride = info.stride();
+
+					for (const auto& count : method_registers.current_draw_clause.first_count_commands)
+					{
+						const u32 vertCount     = count.second;
+						const size_t bufferSize = vertCount * vertStride + vertSize;
+
+						frame_capture_data::memory_block block;
+						block.addr     = addr;
+						block.ioOffset = get_io_offset(base_address, memory_location);
+						block.offset   = (count.first * vertStride);
+
+						frame_capture_data::memory_block_data block_data;
+						block_data.data.resize(bufferSize);
+						std::memcpy(block_data.data.data(), vm::base(addr + block.offset), bufferSize);
+						insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
+					}
+				}
+			}
+			// save index buffer if used
+			else if (method_registers.current_draw_clause.command == draw_command::indexed)
+			{
+				const u32 input_mask = method_registers.vertex_attrib_input_mask();
+
+				const u32 base_address    = method_registers.index_array_address();
+				const u32 memory_location = method_registers.index_array_location();
+
+				const u32 base_addr   = get_address(base_address, memory_location);
+				const u32 type_size   = get_index_type_size(method_registers.index_type());
+				const auto index_type = method_registers.index_type();
+
+				// manually parse index buffer and copy vertex buffer
+				u32 min_index = 0xFFFF, max_index = 0;
+				if (index_type == index_array_type::u32)
+					min_index = 0xFFFFFFFF;
+
+				const bool is_primitive_restart_enabled = method_registers.restart_index_enabled();
+				const u32 primitive_restart_index       = method_registers.restart_index();
+
+				for (const auto& count : method_registers.current_draw_clause.first_count_commands)
+				{
+					const u32 idxFirst = count.first;
+					const u32 idxCount = count.second;
+					const u32 idxAddr  = base_addr + (idxFirst * type_size);
+
+					const size_t bufferSize = idxCount * type_size;
+
+					frame_capture_data::memory_block block;
+					block.addr     = base_addr;
+					block.ioOffset = get_io_offset(base_address, memory_location);
+					block.offset   = (idxFirst * type_size);
+
+					frame_capture_data::memory_block_data block_data;
+					block_data.data.resize(bufferSize);
+					std::memcpy(block_data.data.data(), vm::base(idxAddr), bufferSize);
+					insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
+
+					switch (index_type)
+					{
+					case index_array_type::u16:
+					{
+						auto fifo = vm::ptr<u16>::make(idxAddr);
+						for (u32 i = 0; i < idxCount; ++i)
+						{
+							u16 index = fifo[i];
+							if (is_primitive_restart_enabled && index == (u16)primitive_restart_index)
+								continue;
+							index     = (u16)get_index_from_base(index, method_registers.vertex_data_base_index());
+							min_index = (u16)std::min(index, (u16)min_index);
+							max_index = (u16)std::max(index, (u16)max_index);
+						}
+						break;
+					}
+					case index_array_type::u32:
+					{
+						auto fifo = vm::ptr<u32>::make(idxAddr);
+						for (u32 i = 0; i < idxCount; ++i)
+						{
+							u32 index = fifo[i];
+							if (is_primitive_restart_enabled && index == primitive_restart_index)
+								continue;
+							index     = get_index_from_base(index, method_registers.vertex_data_base_index());
+							min_index = std::min(index, min_index);
+							max_index = std::max(index, max_index);
+						}
+						break;
+					}
+					}
+				}
+
+				if (min_index > max_index)
+				{
+					// ignore?
+				}
+
+				for (u8 index = 0; index < limits::vertex_count; ++index)
+				{
+					const bool enabled = !!(input_mask & (1 << index));
+					if (!enabled)
+						continue;
+
+					const auto& info = method_registers.vertex_arrays_info[index];
+					if (!info.size())
+						continue;
+
+					// vert buffer
+					const u32 vertStride      = info.stride();
+					const u32 base_address    = get_vertex_offset_from_base(method_registers.vertex_data_base_offset(), info.offset() & 0x7fffffff);
+					const u32 memory_location = info.offset() >> 31;
+
+					const u32 addr     = get_address(base_address, memory_location);
+					const u32 vertSize = get_vertex_type_size_on_host(info.type(), info.size());
+
+					const u32 bufferSize = vertStride * (max_index - min_index + 1) + vertSize;
+
+					frame_capture_data::memory_block block;
+					block.addr     = addr;
+					block.ioOffset = get_io_offset(base_address, memory_location);
+					block.offset   = (min_index * vertStride);
+
+					frame_capture_data::memory_block_data block_data;
+					block_data.data.resize(bufferSize);
+					std::memcpy(block_data.data.data(), vm::base(addr + block.offset), bufferSize);
+					insert_mem_block_in_map(mem_changes, std::move(block), std::move(block_data));
+				}
+			}
+
+			capture_display_tile_state(rsx, frame_capture.replay_commands.back());
+			capture_surface_state(rsx, frame_capture.replay_commands.back());
+		}
+
+		// i realize these are a slight copy pasta of the rsx_method implementations but its kinda unavoidable currently
+		void capture_image_in(thread* rsx, frame_capture_data::replay_command& replay_command)
+		{
+			const rsx::blit_engine::transfer_operation operation = method_registers.blit_engine_operation();
+
+			const u16 out_x = method_registers.blit_engine_output_x();
+			const u16 out_y = method_registers.blit_engine_output_y();
+			const u16 out_w = method_registers.blit_engine_output_width();
+			const u16 out_h = method_registers.blit_engine_output_height();
+
+			const u16 in_w = method_registers.blit_engine_input_width();
+			const u16 in_h = method_registers.blit_engine_input_height();
+
+			const blit_engine::transfer_origin in_origin                    = method_registers.blit_engine_input_origin();
+			const blit_engine::transfer_interpolator in_inter               = method_registers.blit_engine_input_inter();
+			const rsx::blit_engine::transfer_source_format src_color_format = method_registers.blit_engine_src_color_format();
+
+			const f32 in_x = std::ceil(method_registers.blit_engine_in_x());
+			const f32 in_y = std::ceil(method_registers.blit_engine_in_y());
+
+			u16 in_pitch = method_registers.blit_engine_input_pitch();
+
+			if (in_w == 0 || in_h == 0 || out_w == 0 || out_h == 0)
+			{
+				return;
+			}
+
+			const u32 src_offset = method_registers.blit_engine_input_offset();
+			const u32 src_dma    = method_registers.blit_engine_input_location();
+
+			const u32 in_bpp              = (src_color_format == rsx::blit_engine::transfer_source_format::r5g6b5) ? 2 : 4; // bytes per pixel
+			const u32 in_offset           = u32(in_x * in_bpp + in_pitch * in_y);
+			const tiled_region src_region = rsx->get_tiled_address(src_offset + in_offset, src_dma & 0xf);
+
+			frame_capture_data::memory_block block;
+			block.addr     = src_region.address;
+			block.ioOffset = get_io_offset(src_region.tile ? src_region.base : src_offset + in_offset, src_dma & 0xf);
+
+			u8* pixels_src = src_region.tile ? src_region.ptr + src_region.base : src_region.ptr;
+
+			if (in_pitch == 0)
+			{
+				in_pitch = in_bpp * in_w;
+			}
+
+			rsx->read_barrier(src_region.address, in_pitch * in_h);
+
+			frame_capture_data::memory_block_data block_data;
+			block_data.data.resize(in_pitch * in_h);
+			std::memcpy(block_data.data.data(), pixels_src, in_pitch * in_h);
+			insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
+
+			// 'capture' destination to ensure memory is alloc'd and usable in replay
+			u32 dst_offset = 0;
+			u32 dst_dma    = 0;
+			rsx::blit_engine::transfer_destination_format dst_color_format;
+			u32 out_pitch    = 0;
+			u32 out_aligment = 64;
+
+			switch (method_registers.blit_engine_context_surface())
+			{
+			case blit_engine::context_surface::surface2d:
+				dst_dma          = method_registers.blit_engine_output_location_nv3062();
+				dst_offset       = method_registers.blit_engine_output_offset_nv3062();
+				dst_color_format = method_registers.blit_engine_nv3062_color_format();
+				out_pitch        = method_registers.blit_engine_output_pitch_nv3062();
+				out_aligment     = method_registers.blit_engine_output_alignment_nv3062();
+				break;
+
+			case blit_engine::context_surface::swizzle2d:
+				dst_dma          = method_registers.blit_engine_nv309E_location();
+				dst_offset       = method_registers.blit_engine_nv309E_offset();
+				dst_color_format = method_registers.blit_engine_output_format_nv309E();
+				break;
+
+			default: break;
+			}
+
+			const u32 out_bpp             = (dst_color_format == rsx::blit_engine::transfer_destination_format::r5g6b5) ? 2 : 4;
+			const s32 out_offset          = out_x * out_bpp + out_pitch * out_y;
+			const tiled_region dst_region = rsx->get_tiled_address(dst_offset + out_offset, dst_dma & 0xf);
+
+			frame_capture_data::memory_block blockDst;
+			blockDst.addr     = dst_region.address;
+			blockDst.ioOffset = get_io_offset(dst_region.tile ? dst_region.base : dst_offset + out_offset, dst_dma & 0xf);
+			if (blockDst.ioOffset != -1)
+			{
+				u32 blockSize = method_registers.blit_engine_context_surface() != blit_engine::context_surface::swizzle2d ? out_pitch * out_h : out_bpp * next_pow2(out_w) * next_pow2(out_h);
+
+				blockDst.size = blockSize;
+				frame_capture_data::memory_block_data block_data;
+				insert_mem_block_in_map(replay_command.memory_state, std::move(blockDst), std::move(block_data));
+			}
+
+			capture_display_tile_state(rsx, replay_command);
+		}
+
+		void capture_buffer_notify(thread* rsx, frame_capture_data::replay_command& replay_command)
+		{
+			s32 in_pitch          = method_registers.nv0039_input_pitch();
+			const u32 line_length = method_registers.nv0039_line_length();
+			const u32 line_count  = method_registers.nv0039_line_count();
+			const u8 in_format    = method_registers.nv0039_input_format();
+
+			if (!in_pitch)
+			{
+				in_pitch = line_length;
+			}
+
+			u32 src_offset = method_registers.nv0039_input_offset();
+			u32 src_dma    = method_registers.nv0039_input_location();
+			u32 src_addr   = get_address(src_offset, src_dma);
+
+			rsx->read_barrier(src_addr, in_pitch * line_count);
+
+			const u8* src = (u8*)vm::base(src_addr);
+
+			frame_capture_data::memory_block block;
+			block.addr     = src_addr;
+			block.ioOffset = get_io_offset(src_offset, src_dma);
+
+			frame_capture_data::memory_block_data block_data;
+			block_data.data.resize(in_pitch * line_count);
+
+			for (u32 i = 0; i < line_count; ++i)
+			{
+				std::memcpy(block_data.data.data() + (line_length * i), src, line_length);
+				src += in_pitch;
+			}
+
+			insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
+
+			// we 'capture' destination mostly to ensure that the location is allocated when replaying
+			u32 dst_offset = method_registers.nv0039_output_offset();
+			u32 dst_dma    = method_registers.nv0039_output_location();
+			u32 dst_addr   = get_address(dst_offset, dst_dma);
+
+			s32 out_pitch = method_registers.nv0039_output_pitch();
+			if (!out_pitch)
+			{
+				out_pitch = line_length;
+			}
+
+			frame_capture_data::memory_block blockDst;
+			blockDst.addr     = dst_addr;
+			blockDst.ioOffset = get_io_offset(dst_offset, dst_dma);
+
+			// only check for iooffset'd data
+			if (blockDst.ioOffset != -1)
+			{
+				frame_capture_data::memory_block_data block_data;
+				blockDst.size = out_pitch * line_count;
+				insert_mem_block_in_map(replay_command.memory_state, std::move(blockDst), std::move(block_data));
+			}
+
+			capture_display_tile_state(rsx, replay_command);
+		}
+
+		void capture_display_tile_state(thread* rsx, frame_capture_data::replay_command& replay_command)
+		{
+			frame_capture_data::display_buffers_state dbstate;
+			dbstate.count = rsx->display_buffers_count;
+			// should this only happen on flip?
+			for (u32 i = 0; i < rsx->display_buffers_count; ++i)
+			{
+				const auto& db            = rsx->display_buffers[i];
+				dbstate.buffers[i].height = db.height;
+				dbstate.buffers[i].width  = db.width;
+				dbstate.buffers[i].offset = db.offset;
+				dbstate.buffers[i].pitch  = db.pitch;
+			}
+
+			const u64 dbnum = XXH64(&dbstate, sizeof(frame_capture_data::display_buffers_state), 0);
+			if (frame_capture.display_buffers_map.find(dbnum) == frame_capture.display_buffers_map.end())
+				frame_capture.display_buffers_map.insert(std::make_pair(dbnum, std::move(dbstate)));
+
+			// todo: hook tile call sys_rsx call or something
+			frame_capture_data::tile_state tilestate;
+			for (u32 i = 0; i < limits::tiles_count; ++i)
+			{
+				const auto& tile = rsx->tiles[i];
+				auto& tstate     = tilestate.tiles[i];
+				tstate.bank      = tile.bank;
+				tstate.base      = tile.base;
+				tstate.binded    = tile.binded;
+				tstate.comp      = tile.comp;
+				tstate.location  = tile.location;
+				tstate.offset    = tile.offset;
+				tstate.pitch     = tile.pitch;
+				tstate.size      = tile.size;
+			}
+
+			for (u32 i = 0; i < limits::zculls_count; ++i)
+			{
+				const auto& zc      = rsx->zculls[i];
+				auto& zcstate       = tilestate.zculls[i];
+				zcstate.aaFormat    = zc.aaFormat;
+				zcstate.binded      = zc.binded;
+				zcstate.cullStart   = zc.cullStart;
+				zcstate.height      = zc.height;
+				zcstate.offset      = zc.offset;
+				zcstate.sFunc       = zc.sFunc;
+				zcstate.sMask       = zc.sMask;
+				zcstate.sRef        = zc.sRef;
+				zcstate.width       = zc.width;
+				zcstate.zcullDir    = zc.zcullDir;
+				zcstate.zcullFormat = zc.zcullFormat;
+				zcstate.zFormat     = zc.zFormat;
+			}
+
+			const u64 tsnum = XXH64(&tilestate, sizeof(frame_capture_data::tile_state), 0);
+
+			if (frame_capture.tile_map.find(tsnum) == frame_capture.tile_map.end())
+				frame_capture.tile_map.insert(std::make_pair(tsnum, std::move(tilestate)));
+
+			replay_command.display_buffer_state = dbnum;
+			replay_command.tile_state           = tsnum;
+		}
+
+		// for the most part capturing this is just to make sure the iomemory is recorded/allocated and accounted for in replay
+		void capture_get_report(thread* rsx, frame_capture_data::replay_command& replay_command, u32 arg)
+		{
+			u32 location = 0;
+			u32 offset   = arg & 0xffffff;
+
+			blit_engine::context_dma report_dma = method_registers.context_dma_report();
+
+			switch (report_dma)
+			{
+			// ignore regular report location
+			case blit_engine::context_dma::to_memory_get_report: location = CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_LOCAL; return;
+			case blit_engine::context_dma::report_location_main: location = CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN; break;
+			case blit_engine::context_dma::memory_host_buffer: location = CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER; break;
+			default: return;
+			}
+
+			u32 addr = get_address(offset, location);
+
+			frame_capture_data::memory_block block;
+			block.addr     = addr;
+			block.ioOffset = get_io_offset(offset, location);
+			block.size     = 16;
+
+			frame_capture_data::memory_block_data block_data;
+			insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
+		}
+
+		// This just checks the color/depth addresses and makes sure they are accounted for in io allocations
+		// Hopefully this works without fully having to calculate actual size
+		void capture_surface_state(thread* rsx, frame_capture_data::replay_command& replay_command)
+		{
+			const auto target = rsx::method_registers.surface_color_target();
+
+			u32 offset_color[] =
+			{
+			    rsx::method_registers.surface_a_offset(),
+			    rsx::method_registers.surface_b_offset(),
+			    rsx::method_registers.surface_c_offset(),
+			    rsx::method_registers.surface_d_offset(),
+			};
+
+			u32 context_dma_color[] =
+			{
+			    rsx::method_registers.surface_a_dma(),
+			    rsx::method_registers.surface_b_dma(),
+			    rsx::method_registers.surface_c_dma(),
+			    rsx::method_registers.surface_d_dma(),
+			};
+
+			auto check_add = [&replay_command](u32 offset, u32 dma) -> void
+			{
+				u32 ioOffset = get_io_offset(offset, dma);
+				if (ioOffset == -1)
+					return;
+
+				u32 addr = get_address(offset, dma);
+
+				frame_capture_data::memory_block block;
+				block.addr     = addr;
+				block.ioOffset = ioOffset;
+				block.size     = 64;
+
+				frame_capture_data::memory_block_data block_data;
+				insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
+			};
+
+			for (const auto& index : utility::get_rtt_indexes(target))
+			{
+				check_add(offset_color[index], context_dma_color[index]);
+			}
+
+			check_add(rsx::method_registers.surface_z_offset(), rsx::method_registers.surface_z_dma());
+		}
+
+		void capture_inline_transfer(thread* rsx, frame_capture_data::replay_command& replay_command, u32 idx, u32 arg)
+		{
+			const u16 x = method_registers.nv308a_x();
+			const u16 y = method_registers.nv308a_y();
+
+			const u32 pixel_offset = (method_registers.blit_engine_output_pitch_nv3062() * y) + (x << 2);
+			const u32 addr_offset  = method_registers.blit_engine_output_offset_nv3062() + pixel_offset + idx * 4;
+
+			// just need to capture dst for allocation later if in iomem
+
+			u32 ioOffset = get_io_offset(addr_offset, method_registers.blit_engine_output_location_nv3062());
+			if (ioOffset == -1)
+				return;
+
+			u32 addr = get_address(addr_offset, method_registers.blit_engine_output_location_nv3062());
+
+			frame_capture_data::memory_block block;
+			block.addr     = addr;
+			block.ioOffset = ioOffset;
+			block.size     = 4;
+
+			frame_capture_data::memory_block_data block_data;
+			insert_mem_block_in_map(replay_command.memory_state, std::move(block), std::move(block_data));
+		}
+	}
+}

rpcs3/Emu/RSX/Capture/rsx_capture.h

@@ -0,0 +1,17 @@
+#pragma once
+#include "rsx_replay.h"
+
+namespace rsx
+{
+	class thread;
+	namespace capture
+	{
+		void capture_draw_memory(thread* rsx);
+		void capture_image_in(thread* rsx, frame_capture_data::replay_command& replay_command);
+		void capture_buffer_notify(thread* rsx, frame_capture_data::replay_command& replay_command);
+		void capture_display_tile_state(thread* rsx, frame_capture_data::replay_command& replay_command);
+		void capture_surface_state(thread* rsx, frame_capture_data::replay_command& replay_command);
+		void capture_get_report(thread* rsx, frame_capture_data::replay_command& replay_command, u32 arg);
+		void capture_inline_transfer(thread* rsx, frame_capture_data::replay_command& replay_command, u32 idx, u32 arg);
+	}
+}

rpcs3/Emu/RSX/Capture/rsx_replay.cpp

@@ -0,0 +1,340 @@
+#include "stdafx.h"
+#include "rsx_replay.h"
+
+#include "Emu/System.h"
+#include "Emu/Cell/lv2/sys_rsx.h"
+#include "Emu/Memory/Memory.h"
+#include "Emu/RSX/GSRender.h"
+
+#include <map>
+
+namespace rsx
+{
+	be_t<u32> rsx_replay_thread::allocate_context()
+	{
+		const u32 contextAddr = vm::alloc(sizeof(rsx_context), vm::main);
+		if (contextAddr == 0)
+			fmt::throw_exception("Capture Replay: context alloc failed");
+		const auto& contextInfo = vm::_ref<rsx_context>(contextAddr);
+
+		if (sys_rsx_memory_allocate(vm::get_addr(&contextInfo.mem_handle), vm::get_addr(&contextInfo.mem_addr), 0x0F900000, 0, 0, 0, 0) != CELL_OK)
+			fmt::throw_exception("Capture Replay: sys_rsx_memory_allocate failed!");
+
+		if (sys_rsx_context_allocate(vm::get_addr(&contextInfo.context_id), vm::get_addr(&contextInfo.dma_addr), vm::get_addr(&contextInfo.driver_info), vm::get_addr(&contextInfo.reports_addr), contextInfo.mem_handle, 0) != CELL_OK)
+			fmt::throw_exception("Capture Replay: sys_rsx_context_allocate failed!");
+
+		// 'fake' initialize usermemory
+		// todo: seriously, need to probly watch the replay memory map and just make sure its mapped before we copy rather than do this
+		vm::falloc(0x20000000, 0x10000000, vm::user_space);
+
+		return contextInfo.context_id;
+	}
+
+	std::tuple<u32, u32> rsx_replay_thread::get_usable_fifo_range()
+	{
+		u32 fifo_size = 4;
+
+		// run through replay commands to figure out how big command buffer needs to be 
+		// technically we could do this in batches if it gets too big, but we should be fine
+		// as we aren't allocating anything on main memory, although it may make issues with iooffset later
+		for (const auto& rc : frame->replay_commands)
+		{
+			const u32 count = (rc.rsx_command.first >> 18) & 0x7ff;
+			// allocate for register plus w/e number of arguments it has
+			fifo_size += (count + 1) * 4;
+		}
+
+		// safety check for now
+		// since we are allocating iobuffer, we need to make sure that any memory we use is not being used by the replay
+		std::map<u32, u32> ioOffsets;
+		u32 lowest_iooffset = 0xFFFFFFFF;
+		for (const auto& mm : frame->memory_map)
+		{
+			u32 offset = mm.second.ioOffset;
+			lowest_iooffset = std::min(lowest_iooffset, offset);
+			if (offset != 0xFFFFFFFF)
+			{
+				u32 iosize = mm.second.size + mm.second.offset;
+				auto it = ioOffsets.find(offset);
+				if (it == ioOffsets.end())
+					ioOffsets[offset] = iosize;
+				else
+					ioOffsets[offset] = std::max(ioOffsets[offset], iosize);
+			}
+		}
+
+		// if we cant use fifo starting at 0, try to find a block between them
+		u32 fifo_start_addr = 0;
+		if (fifo_size >= lowest_iooffset)
+		{
+			u32 largest_free_block = 0;
+			u32 largest_cur_end = 0; // this keeps track of largest end, in case allocations 'overlap'
+			for (auto io = ioOffsets.begin(); io != ioOffsets.end(); ++io)
+			{
+				auto next = std::next(io);
+
+				// last 'offset' is just end of memory
+				u32 nextOffset = 0x0F900000;
+				if (next != ioOffsets.end())
+					nextOffset = next->first - 4;
+
+				largest_cur_end = std::max(largest_cur_end, io->first + io->second + 4);
+				if (largest_cur_end < nextOffset)
+				{
+					u32 freeSize = nextOffset - largest_cur_end;
+					if (freeSize > largest_free_block)
+					{
+						fifo_start_addr = largest_cur_end;
+						largest_free_block = freeSize;
+					}
+
+					if (largest_free_block > fifo_size)
+						break;
+				}
+
+				if (next == ioOffsets.end())
+					break;
+			}
+
+			// todo: figure out *another* way to inject fifo if both ideas above fail
+			if (largest_free_block < fifo_size)
+				fmt::throw_exception("Capture Replay: no space in io for fifo commands! size: 0x%x, lowest in capture: 0x%x, largest_free_block: 0x%x", fifo_size, lowest_iooffset, largest_free_block);
+		}
+
+		return std::make_tuple(fifo_start_addr, fifo_size);
+	}
+
+	std::vector<u32> rsx_replay_thread::alloc_write_fifo(be_t<u32> context_id, u32 fifo_start_addr, u32 fifo_size)
+	{
+		const u32 fifo_mem = vm::alloc(fifo_size, vm::main);
+		if (fifo_mem == 0)
+			fmt::throw_exception("Capture Replay: fifo alloc failed! size: 0x%x", fifo_size);
+
+		// copy commands into fifo buffer
+		// todo: could change rsx_command to just be values to avoid this loop,
+		auto fifo_addr = vm::ptr<u32>::make(fifo_mem);
+		u32 count = 0;
+		std::vector<u32> fifo_stops;
+		u32 currentOffset = fifo_start_addr;
+		for (const auto& rc : frame->replay_commands)
+		{
+			bool hasState = (rc.memory_state.size() > 0) || (rc.display_buffer_state != 0) || (rc.tile_state != 0);
+			if (hasState)
+			{
+				if (count != 0)
+				{
+					// todo: support memory state in the middle of incrememented command
+					// This shouldn't ever happen as long as captures stay in 'strict' aka non-multidraw mode
+					fmt::throw_exception("capture replay: state change not supported between increment commands");
+				}
+
+				fifo_stops.emplace_back(currentOffset);
+			}
+
+			// spit out command
+			if (count == 0)
+			{
+				count = (rc.rsx_command.first >> 18) & 0x7ff;
+				*fifo_addr = rc.rsx_command.first;
+				fifo_addr++;
+				currentOffset += 4;
+			}
+
+			if (count != 0)
+			{
+				*fifo_addr = rc.rsx_command.second;
+				fifo_addr++;
+				count--;
+				currentOffset += 4;
+			}
+		}
+
+		fifo_stops.emplace_back(currentOffset);
+
+		if (sys_rsx_context_iomap(context_id, fifo_start_addr, fifo_mem, fifo_size, 0) != CELL_OK)
+			fmt::throw_exception("Capture Replay: fifo mapping failed");
+
+		return fifo_stops;
+	}
+
+	void rsx_replay_thread::apply_frame_state(be_t<u32> context_id, const frame_capture_data::replay_command& replay_cmd)
+	{
+		// apply memory needed for command
+		for (const auto& state : replay_cmd.memory_state)
+		{
+			auto it = frame->memory_map.find(state);
+			if (it == frame->memory_map.end())
+				fmt::throw_exception("requested memory state for command not found in memory_map");
+
+			if (it->second.data_state != 0)
+			{
+				const auto& memblock = it->second;
+				auto it_data = frame->memory_data_map.find(it->second.data_state);
+				if (it_data == frame->memory_data_map.end())
+					fmt::throw_exception("requested memory data state for command not found in memory_data_map");
+
+				const auto& data_block = it_data->second;
+				std::memcpy(vm::base(memblock.addr + memblock.offset), data_block.data.data(), data_block.data.size());
+			}
+		}
+
+		if (replay_cmd.display_buffer_state != 0 && replay_cmd.display_buffer_state != cs.display_buffer_hash)
+		{
+			auto it = frame->display_buffers_map.find(replay_cmd.display_buffer_state);
+			if (it == frame->display_buffers_map.end())
+				fmt::throw_exception("requested display buffer for command not found");
+
+			const auto& dbstate = it->second;
+			for (u32 i = 0; i < dbstate.count; ++i)
+			{
+				const auto& buf = dbstate.buffers[i];
+				if (cs.display_buffer_hash != 0 && memcmp(&cs.buffer_state.buffers[i], &buf, sizeof(rsx::frame_capture_data::buffer_state)) == 0)
+					continue;
+
+				cs.buffer_state.buffers[i] = buf;
+				sys_rsx_context_attribute(context_id, 0x104, i,
+					(u64)dbstate.buffers[i].width << 32 | dbstate.buffers[i].height, (u64)dbstate.buffers[i].pitch << 32 | dbstate.buffers[i].offset, 0);
+			}
+			cs.display_buffer_hash = replay_cmd.display_buffer_state;
+		}
+
+		if (replay_cmd.tile_state != 0 && replay_cmd.tile_state != cs.tile_hash)
+		{
+			auto it = frame->tile_map.find(replay_cmd.tile_state);
+			if (it == frame->tile_map.end())
+				fmt::throw_exception("requested tile state command not found");
+
+			const auto& tstate = it->second;
+			for (u32 i = 0; i < limits::tiles_count; ++i)
+			{
+				const auto& tstile = tstate.tiles[i];
+				if (cs.tile_hash != 0 && memcmp(&cs.tile_state.tiles[i], &tstile, sizeof(rsx::frame_capture_data::tile_info)) == 0)
+					continue;
+
+				cs.tile_state.tiles[i] = tstile;
+
+				GcmTileInfo t;
+				t.bank = tstile.bank;
+				t.base = tstile.base;
+				t.binded = tstile.binded;
+				t.comp = tstile.comp;
+				t.location = tstile.location;
+				t.offset = tstile.offset;
+				t.pitch = tstile.pitch;
+				t.size = tstile.size;
+
+				const auto& ti = t.pack();
+				sys_rsx_context_attribute(context_id, 0x300, i, (u64)ti.tile << 32 | ti.limit, (u64)ti.pitch << 32 | ti.format, 0);
+			}
+
+			for (u32 i = 0; i < limits::zculls_count; ++i)
+			{
+				const auto& zctile = tstate.zculls[i];
+				if (cs.tile_hash != 0 && memcmp(&cs.tile_state.zculls[i], &zctile, sizeof(rsx::frame_capture_data::zcull_info)) == 0)
+					continue;
+
+				cs.tile_state.zculls[i] = zctile;
+
+				GcmZcullInfo zc;
+				zc.aaFormat = zctile.aaFormat;
+				zc.binded = zctile.binded;
+				zc.cullStart = zctile.cullStart;
+				zc.height = zctile.height;
+				zc.offset = zctile.offset;
+				zc.sFunc = zctile.sFunc;
+				zc.sMask = zctile.sMask;
+				zc.sRef = zctile.sRef;
+				zc.width = zctile.width;
+				zc.zcullDir = zctile.zcullDir;
+				zc.zcullFormat = zctile.zcullFormat;
+				zc.zFormat = zctile.zFormat;
+
+				const auto& zci = zc.pack();
+				sys_rsx_context_attribute(context_id, 0x301, i, (u64)zci.region << 32 | zci.size, (u64)zci.start << 32 | zci.offset, (u64)zci.status0 << 32 | zci.status1);
+			}
+
+			cs.tile_hash = replay_cmd.tile_state;
+		}
+	}
+
+	void rsx_replay_thread::cpu_task()
+	{
+		be_t<u32> context_id = allocate_context();
+
+		auto fifo_info = get_usable_fifo_range();
+
+		const u32 fifo_start_addr = std::get<0>(fifo_info);
+		const u32 fifo_size = std::get<1>(fifo_info);
+
+		auto fifo_stops = alloc_write_fifo(context_id, fifo_start_addr, fifo_size);
+
+		// map game io
+		for (const auto it : frame->memory_map)
+		{
+			const auto& memblock = it.second;
+			if (memblock.ioOffset == 0xFFFFFFFF)
+				continue;
+
+			// sanity check
+			if (memblock.ioOffset <= fifo_start_addr + fifo_size && fifo_start_addr <= memblock.size + memblock.offset)
+				fmt::throw_exception("Capture Replay: overlap detected between game io allocs and fifo alloc, algorithms botched.");
+
+			if (sys_rsx_context_iomap(context_id, memblock.ioOffset, memblock.addr, memblock.size + memblock.offset, 0) != CELL_OK)
+				fmt::throw_exception("rsx io map failed for block");
+		}
+
+		while (!Emu.IsStopped())
+		{
+			// start up fifo buffer by dumping the put ptr to first stop
+			sys_rsx_context_attribute(context_id, 0x001, fifo_start_addr, fifo_stops[0], 0, 0);
+
+			auto renderer = fxm::get<GSRender>();
+			size_t stopIdx = 0;
+			for (const auto& replay_cmd : frame->replay_commands)
+			{
+				while (Emu.IsPaused())
+					std::this_thread::sleep_for(10ms);
+
+				if (Emu.IsStopped())
+					break;
+
+				// Loop and hunt down our next state change that needs to be done
+				if (!((replay_cmd.memory_state.size() > 0) || (replay_cmd.display_buffer_state != 0) || (replay_cmd.tile_state != 0)))
+					continue;
+
+				// wait until rsx idle and at our first 'stop' to apply state
+				while (!Emu.IsStopped() && (renderer->ctrl->get != renderer->ctrl->put) && (renderer->ctrl->get != fifo_stops[stopIdx]))
+				{
+					while (Emu.IsPaused())
+						std::this_thread::sleep_for(10ms);
+					std::this_thread::yield();
+				}
+
+				stopIdx++;
+
+				apply_frame_state(context_id, replay_cmd);
+
+				// move put ptr to next stop
+				if (stopIdx >= fifo_stops.size())
+					fmt::throw_exception("Capture Replay: StopIdx greater than size of fifo_stops");
+
+				renderer->ctrl->put = fifo_stops[stopIdx];
+			}
+
+			// dump put to end of stops, which should have actual end
+			u32 end = fifo_stops.back();
+			renderer->ctrl->put = end;
+
+			while (renderer->ctrl->get != end && !Emu.IsStopped())
+			{
+				while (Emu.IsPaused())
+					std::this_thread::sleep_for(10ms);
+			}
+
+			// random pause to not destroy gpu
+			std::this_thread::sleep_for(10ms);
+		}
+
+		state += cpu_flag::exit;
+	}
+}

rpcs3/Emu/RSX/Capture/rsx_replay.h

@@ -0,0 +1,241 @@
+#pragma once
+
+#include "Emu/System.h"
+#include "Emu/IdManager.h"
+#include "Emu/Cell/PPUModule.h"
+#include "Emu/Cell/lv2/sys_sync.h"
+#include "Emu/Cell/lv2/sys_ppu_thread.h"
+
+#include <cereal/types/vector.hpp>
+#include <cereal/types/array.hpp>
+#include <cereal/types/string.hpp>
+#include <cereal/types/utility.hpp>
+#include <cereal/types/unordered_set.hpp>
+
+namespace rsx
+{
+	constexpr u32 FRAME_CAPTURE_MAGIC = 0x52524300; // ascii 'RRC/0'
+	constexpr u32 FRAME_CAPTURE_VERSION = 0x1;
+	struct frame_capture_data
+	{
+
+		struct memory_block_data
+		{
+			std::vector<u8> data;
+			template<typename Archive>
+			void serialize(Archive& ar)
+			{
+				ar(data);
+			}
+		};
+
+		// simple block to hold ps3 address and data
+		struct memory_block
+		{
+			u32 addr{0};
+			u32 ioOffset{0xFFFFFFFF}; // rsx ioOffset, -1 signifies unused
+			u32 offset{0};			  // offset into addr/ioOffset to copy state into
+			u32 size{0};			  // size of block needed
+			u64 data_state{0};		  // this can be 0, in which case its just needed as an alloc
+			template<typename Archive>
+			void serialize(Archive & ar)
+			{
+				ar(addr);
+				ar(ioOffset);
+				ar(offset);
+				ar(size);
+				ar(data_state);
+			}
+		};
+
+		struct replay_command
+		{
+			std::pair<u32, u32> rsx_command;      // fifo command
+			std::unordered_set<u64> memory_state; // index into memory_map for the various memory blocks that need applying before this command can run
+			u64 tile_state{0};                    // tile state for this command
+			u64 display_buffer_state{0};
+
+			template<typename Archive>
+			void serialize(Archive & ar)
+			{
+				ar(rsx_command);
+				ar(memory_state);
+				ar(tile_state);
+				ar(display_buffer_state);
+			}
+		};
+
+		// same thing as gcmtileinfo
+		struct tile_info
+		{
+			u32 location{0};
+			u32 offset{0};
+			u32 size{0};
+			u32 pitch{0};
+			u32 comp{0};
+			u32 base{0};
+			u32 bank{0};
+			bool binded{false};
+
+			template<typename Archive>
+			void serialize(Archive & ar)
+			{
+				ar(location);
+				ar(offset);
+				ar(size);
+				ar(pitch);
+				ar(comp);
+				ar(base);
+				ar(bank);
+				ar(binded);
+			}
+		};
+
+		// same thing as gcmzcullinfo
+		struct zcull_info
+		{
+			u32 offset{0};
+			u32 width{0};
+			u32 height{0};
+			u32 cullStart{0};
+			u32 zFormat{0};
+			u32 aaFormat{0};
+			u32 zcullDir{0};
+			u32 zcullFormat{0};
+			u32 sFunc{0};
+			u32 sRef{0};
+			u32 sMask{0};
+			bool binded{false};
+
+			template<typename Archive>
+			void serialize(Archive & ar)
+			{
+				ar(offset);
+				ar(width);
+				ar(cullStart);
+				ar(zFormat);
+				ar(aaFormat);
+				ar(zcullDir);
+				ar(zcullFormat);
+				ar(sFunc);
+				ar(sRef);
+				ar(sMask);
+				ar(binded);
+			}
+		};
+
+		// bleh, may need to break these out, might be unneccessary to do both always
+		struct tile_state
+		{
+			tile_info tiles[15];
+			zcull_info zculls[8];
+
+			template<typename Archive>
+			void serialize(Archive & ar)
+			{
+				ar(tiles);
+				ar(zculls);
+			}
+		};
+
+		struct buffer_state
+		{
+			u32 width{0};
+			u32 height{0};
+			u32 pitch{0};
+			u32 offset{0};
+
+			template<typename Archive>
+			void serialize(Archive & ar)
+			{
+				ar(width);
+				ar(height);
+				ar(pitch);
+				ar(offset);
+			}
+		};
+
+		struct display_buffers_state
+		{
+			std::array<buffer_state, 8> buffers;
+			u32 count{0};
+
+			template<typename Archive>
+			void serialize(Archive & ar)
+			{
+				ar(buffers);
+				ar(count);
+			}
+		};
+
+		u32 magic;
+		u32 version;
+		// hashmap of holding various states for tile
+		std::unordered_map<u64, tile_state> tile_map;
+		// hashmap of various memory 'changes' that can be applied to ps3 memory
+		std::unordered_map<u64, memory_block> memory_map;
+		// hashmap of memory blocks that can be applied, this is split from above for size decrease
+		std::unordered_map<u64, memory_block_data> memory_data_map;
+		// display buffer state map
+		std::unordered_map<u64, display_buffers_state> display_buffers_map;
+		// actual command queue to hold everything above
+		std::vector<replay_command> replay_commands;
+
+		template<typename Archive>
+		void serialize(Archive & ar)
+		{
+			ar(magic);
+			ar(version);
+			ar(tile_map);
+			ar(memory_map);
+			ar(memory_data_map);
+			ar(display_buffers_map);
+			ar(replay_commands);
+		}
+
+		void reset()
+		{
+			magic = FRAME_CAPTURE_MAGIC;
+			version = FRAME_CAPTURE_VERSION;
+			tile_map.clear();
+			memory_map.clear();
+			replay_commands.clear();
+		}
+	};
+
+
+	class rsx_replay_thread : public ppu_thread
+	{
+		struct rsx_context
+		{
+			be_t<u32> mem_handle;
+			be_t<u32> context_id;
+			be_t<u64> mem_addr;
+			be_t<u64> dma_addr;
+			be_t<u64> reports_addr;
+			be_t<u64> driver_info;
+		};
+
+		struct current_state
+		{
+			u64 tile_hash{0};
+			u64 display_buffer_hash{0};
+			frame_capture_data::display_buffers_state buffer_state;
+			frame_capture_data::tile_state tile_state;
+		};
+
+		current_state cs;
+		std::unique_ptr<frame_capture_data> frame;
+
+	public:
+		rsx_replay_thread(std::unique_ptr<frame_capture_data>&& frame_data)
+			: ppu_thread("Rsx Capture Replay Thread"), frame(std::move(frame_data)) {};
+
+		virtual void cpu_task() override;
+	private:
+		be_t<u32> allocate_context();
+		std::tuple<u32, u32> get_usable_fifo_range();
+		std::vector<u32> alloc_write_fifo(be_t<u32> context_id, u32 fifo_start_addr, u32 fifo_size);
+		void apply_frame_state(be_t<u32> context_id, const frame_capture_data::replay_command& replay_cmd);
+	};
+}

rpcs3/Emu/RSX/Capture/rsx_trace.h

@@ -0,0 +1,33 @@
+#pragma once
+#include <string>
+#include <array>
+#include <vector>
+#include "Utilities/types.h"
+#include "Emu/RSX/rsx_methods.h"
+
+namespace rsx
+{
+struct frame_trace_data
+{
+	struct draw_state
+	{
+		std::string name;
+		std::pair<std::string, std::string> programs;
+		rsx::rsx_state state;
+		std::array<std::vector<gsl::byte>, 4> color_buffer;
+		std::array<std::vector<gsl::byte>, 2> depth_stencil;
+		std::vector<gsl::byte> index;
+		u32 vertex_count;
+
+	};
+
+	std::vector<std::pair<u32, u32>> command_queue;
+	std::vector<draw_state> draw_calls;
+
+	void reset()
+	{
+		command_queue.clear();
+		draw_calls.clear();
+	}
+};
+}

rpcs3/Emu/RSX/RSXThread.cpp

@@ -8,6 +8,7 @@
 
 #include "Common/BufferUtils.h"
 #include "Common/texture_cache.h"
+#include "Capture/rsx_capture.h"
 #include "rsx_methods.h"
 #include "rsx_utils.h"
 
@@ -15,6 +16,7 @@
 #include "Utilities/StrUtil.h"
 
 #include <thread>
+#include <unordered_set>
 #include <fenv.h>
 
 class GSRender;
@@ -22,9 +24,8 @@ class GSRender;
 #define CMD_DEBUG 0
 
 bool user_asked_for_frame_capture = false;
-rsx::frame_capture_data frame_debug;
-
-
+rsx::frame_trace_data frame_debug;
+rsx::frame_capture_data frame_capture;
 
 namespace rsx
 {
@@ -255,22 +256,7 @@ namespace rsx
 
 	void thread::capture_frame(const std::string &name)
 	{
-		frame_capture_data::draw_state draw_state = {};
-
-		int clip_w = rsx::method_registers.surface_clip_width();
-		int clip_h = rsx::method_registers.surface_clip_height();
-		draw_state.state = rsx::method_registers;
-		draw_state.color_buffer = std::move(copy_render_targets_to_memory());
-		draw_state.depth_stencil = std::move(copy_depth_stencil_buffer_to_memory());
-
-		if (draw_state.state.current_draw_clause.command == rsx::draw_command::indexed)
-		{
-			draw_state.vertex_count = 0;
-			draw_state.vertex_count = draw_state.state.current_draw_clause.get_elements_count();
-			auto index_raw_data_ptr = get_raw_index_array(draw_state.state.current_draw_clause.first_count_commands);
-			draw_state.index.resize(index_raw_data_ptr.size_bytes());
-			std::copy(index_raw_data_ptr.begin(), index_raw_data_ptr.end(), draw_state.index.begin());
-		}
+		frame_trace_data::draw_state draw_state = {};
 
 		draw_state.programs = get_programs();
 		draw_state.name = name;
@@ -330,6 +316,9 @@ namespace rsx
 
 	void thread::end()
 	{
+		if (capture_current_frame)
+			capture::capture_draw_memory(this);
+
 		in_begin_end = false;
 
 		for (u8 index = 0; index < rsx::limits::vertex_count; ++index)
@@ -847,12 +836,44 @@ namespace rsx
 					}
 				}
 
-				method_registers.decode(reg, value);
-
 				if (capture_current_frame)
 				{
 					frame_debug.command_queue.push_back(std::make_pair(reg, value));
+
+					if (!(reg == NV406E_SET_REFERENCE || reg == NV406E_SEMAPHORE_RELEASE || reg == NV406E_SEMAPHORE_ACQUIRE))
+					{
+						// todo: handle nv406e methods better?, do we care about call/jumps?
+						rsx::frame_capture_data::replay_command replay_cmd;
+						replay_cmd.rsx_command = std::make_pair(i == 0 ? cmd : 0, value);
+
+						frame_capture.replay_commands.push_back(replay_cmd);
+
+						// to make this easier, use the replay command 'i' positions back
+						auto it = std::prev(frame_capture.replay_commands.end(), i + 1);
+
+						switch (reg)
+						{
+						case NV4097_GET_REPORT:
+							capture::capture_get_report(this, *it, value);
+							break;
+						case NV3089_IMAGE_IN:
+							capture::capture_image_in(this, *it);
+							break;
+						case NV0039_BUFFER_NOTIFY:
+							capture::capture_buffer_notify(this, *it);
+							break;
+						case NV4097_CLEAR_SURFACE:
+							capture::capture_surface_state(this, *it);
+							break;
+						default:
+							if (reg >= NV308A_COLOR && reg < NV3089_SET_OBJECT)
+								capture::capture_inline_transfer(this, *it, reg - NV308A_COLOR, value);
+							break;
 						}
+					}
+				}
+
+				method_registers.decode(reg, value);
 
 				if (execute_method_call)
 				{

rpcs3/Emu/RSX/RSXThread.h

@@ -17,7 +17,8 @@
 
 #include "Utilities/Thread.h"
 #include "Utilities/geometry.h"
-#include "rsx_trace.h"
+#include "Capture/rsx_trace.h"
+#include "Capture/rsx_replay.h"
 #include "restore_new.h"
 #include "Utilities/variant.hpp"
 #include "define_new_memleakdetect.h"
@@ -27,7 +28,8 @@
 extern u64 get_system_time();
 
 extern bool user_asked_for_frame_capture;
-extern rsx::frame_capture_data frame_debug;
+extern rsx::frame_trace_data frame_debug;
+extern rsx::frame_capture_data frame_capture;
 
 namespace rsx
 {

rpcs3/Emu/RSX/rsx_cache.h

@@ -301,7 +301,7 @@ namespace rsx
 		template <typename... Args>
 		void load(progress_dialog_helper* dlg, Args&& ...args)
 		{
-			if (g_cfg.video.disable_on_disk_shader_cache)
+			if (g_cfg.video.disable_on_disk_shader_cache || Emu.GetCachePath() == "")
 			{
 				return;
 			}
@@ -382,7 +382,7 @@ namespace rsx
 
 		void store(pipeline_storage_type &pipeline, RSXVertexProgram &vp, RSXFragmentProgram &fp)
 		{
-			if (g_cfg.video.disable_on_disk_shader_cache)
+			if (g_cfg.video.disable_on_disk_shader_cache || Emu.GetCachePath() == "")
 			{
 				return;
 			}

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -7,6 +7,7 @@
 #include "rsx_decode.h"
 #include "Emu/Cell/PPUCallback.h"
 #include "Emu/Cell/lv2/sys_rsx.h"
+#include "Capture/rsx_capture.h"
 
 #include <sstream>
 #include <cereal/archives/binary.hpp>
@@ -990,22 +991,44 @@ namespace rsx
 
 	void flip_command(thread* rsx, u32, u32 arg)
 	{
-		if (user_asked_for_frame_capture)
+		if (user_asked_for_frame_capture && !g_cfg.video.strict_rendering_mode)
+		{
+			// not dealing with non-strict rendering capture for now
+			user_asked_for_frame_capture = false;
+			LOG_FATAL(RSX, "RSX Capture: Capture only supported when ran with strict rendering mode.");
+		}
+		else if (user_asked_for_frame_capture && !rsx->capture_current_frame)
 		{
 			rsx->capture_current_frame = true;
 			user_asked_for_frame_capture = false;
 			frame_debug.reset();
+			frame_capture.reset();
+
+			// random number just to jumpstart the size
+			frame_capture.replay_commands.reserve(8000);
+
+			// capture first tile state with nop cmd
+			rsx::frame_capture_data::replay_command replay_cmd;
+			replay_cmd.rsx_command = std::make_pair(NV4097_NO_OPERATION, 0);
+			frame_capture.replay_commands.push_back(replay_cmd);
+			capture::capture_display_tile_state(rsx, frame_capture.replay_commands.back());
 		}
 		else if (rsx->capture_current_frame)
 		{
 			rsx->capture_current_frame = false;
 			std::stringstream os;
 			cereal::BinaryOutputArchive archive(os);
-			archive(frame_debug);
+			const std::string& filePath = fs::get_config_dir() + "capture.rrc";
+			archive(frame_capture);
 			{
-				fs::file f(fs::get_config_dir() + "capture.txt", fs::rewrite);
+				// todo: 'dynamicly' create capture filename, also may want to compress this data?
+				fs::file f(filePath, fs::rewrite);
 				f.write(os.str());
 			}
+
+			LOG_SUCCESS(RSX, "capture successful: %s", filePath.c_str());
+			
+			frame_capture.reset();
 			Emu.Pause();
 		}
 

rpcs3/Emu/RSX/rsx_trace.h

@@ -1,55 +0,0 @@
-#pragma once
-#include <string>
-#include <array>
-#include <vector>
-#include "Utilities/types.h"
-#include "rsx_methods.h"
-
-#include <cereal/types/vector.hpp>
-#include <cereal/types/array.hpp>
-#include <cereal/types/string.hpp>
-#include <cereal/types/utility.hpp>
-
-namespace rsx
-{
-struct frame_capture_data
-{
-	struct draw_state
-	{
-		std::string name;
-		std::pair<std::string, std::string> programs;
-		rsx::rsx_state state;
-		std::array<std::vector<gsl::byte>, 4> color_buffer;
-		std::array<std::vector<gsl::byte>, 2> depth_stencil;
-		std::vector<gsl::byte> index;
-		u32 vertex_count;
-
-		template<typename Archive>
-		void serialize(Archive & ar)
-		{
-			ar(name);
-			ar(programs);
-			ar(state);
-			ar(color_buffer);
-			ar(depth_stencil);
-			ar(index);
-		}
-
-	};
-	std::vector<std::pair<u32, u32> > command_queue;
-	std::vector<draw_state> draw_calls;
-
-	template<typename Archive>
-	void serialize(Archive & ar)
-	{
-		ar(command_queue);
-		ar(draw_calls);
-	}
-
-	void reset()
-	{
-		command_queue.clear();
-		draw_calls.clear();
-	}
-};
-}

rpcs3/Emu/System.cpp

@@ -13,9 +13,11 @@
 #include "Emu/Cell/RawSPUThread.h"
 #include "Emu/Cell/lv2/sys_sync.h"
 #include "Emu/Cell/lv2/sys_prx.h"
+#include "Emu/Cell/lv2/sys_rsx.h"
 
 #include "Emu/IdManager.h"
 #include "Emu/RSX/GSRender.h"
+#include "Emu/RSX/Capture/rsx_replay.h"
 
 #include "Loader/PSF.h"
 #include "Loader/ELF.h"
@@ -26,9 +28,13 @@
 #include "../Crypto/unpkg.h"
 #include "yaml-cpp/yaml.h"
 
+#include "cereal/archives/binary.hpp"
+
 #include <thread>
 #include <typeinfo>
 #include <queue>
+#include <fstream>
+#include <memory>
 
 #include "Utilities/GDBDebugServer.h"
 
@@ -233,6 +239,50 @@ void Emulator::Init()
 	fxm::make_always<patch_engine>()->append(fs::get_config_dir() + "/patch.yml");
 }
 
+bool Emulator::BootRsxCapture(const std::string& path)
+{
+	if (!fs::is_file(path))
+		return false;
+
+	std::fstream f(path, std::ios::in | std::ios::binary);
+
+	cereal::BinaryInputArchive archive(f);
+	std::unique_ptr<rsx::frame_capture_data> frame = std::make_unique<rsx::frame_capture_data>();
+	archive(*frame);
+
+	if (frame->magic != rsx::FRAME_CAPTURE_MAGIC)
+	{
+		LOG_ERROR(LOADER, "Invalid rsx capture file!");
+		return false;
+	}
+
+	if (frame->version != rsx::FRAME_CAPTURE_VERSION)
+	{
+		LOG_ERROR(LOADER, "Rsx capture file version not supported! Expected %d, found %d", rsx::FRAME_CAPTURE_VERSION, frame->version);
+		return false;
+	}
+
+	Init();
+
+	vm::init();
+
+	// PS3 'executable'
+	m_state = system_state::ready;
+	GetCallbacks().on_ready();
+
+	auto gsrender = fxm::import<GSRender>(Emu.GetCallbacks().get_gs_render);
+	if (gsrender.get() == nullptr)
+		return false;
+
+	GetCallbacks().on_run();
+	m_state = system_state::running;
+
+	auto&& rsxcapture = idm::make_ptr<ppu_thread, rsx::rsx_replay_thread>(std::move(frame));
+	rsxcapture->run();
+
+	return true;
+}
+
 bool Emulator::BootGame(const std::string& path, bool direct, bool add_only)
 {
 	static const char* boot_list[] =

rpcs3/Emu/System.h

@@ -260,6 +260,7 @@ public:
 	}
 
 	bool BootGame(const std::string& path, bool direct = false, bool add_only = false);
+	bool BootRsxCapture(const std::string& path);
 	bool InstallPkg(const std::string& path);
 
 	static std::string GetEmuDir();

rpcs3/VKGSRender.vcxproj

@@ -97,6 +97,7 @@
   <ItemDefinitionGroup>
     <ClCompile>
       <AdditionalIncludeDirectories>..\Vulkan\Vulkan-LoaderAndValidationLayers\include;..\Vulkan\glslang\glslang\Public;..\Vulkan\glslang;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalOptions Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">/bigobj %(AdditionalOptions)</AdditionalOptions>
     </ClCompile>
   </ItemDefinitionGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />

rpcs3/emucore.vcxproj

@@ -277,6 +277,8 @@
     <ClCompile Include="Emu\Cell\SPUInterpreter.cpp" />
     <ClCompile Include="Emu\IdManager.cpp" />
     <ClCompile Include="Emu\Memory\wait_engine.cpp" />
+    <ClCompile Include="Emu\RSX\Capture\rsx_capture.cpp" />
+    <ClCompile Include="Emu\RSX\Capture\rsx_replay.cpp" />
     <ClCompile Include="Emu\RSX\CgBinaryFragmentProgram.cpp" />
     <ClCompile Include="Emu\RSX\CgBinaryVertexProgram.cpp" />
     <ClCompile Include="Emu\RSX\Common\BufferUtils.cpp" />
@@ -523,6 +525,9 @@
     <ClInclude Include="Emu\CPU\CPUDisAsm.h" />
     <ClInclude Include="Emu\CPU\CPUThread.h" />
     <ClInclude Include="Emu\Memory\wait_engine.h" />
+    <ClInclude Include="Emu\RSX\Capture\rsx_capture.h" />
+    <ClInclude Include="Emu\RSX\Capture\rsx_replay.h" />
+    <ClInclude Include="Emu\RSX\Capture\rsx_trace.h" />
     <ClInclude Include="Emu\RSX\Common\GLSLCommon.h" />
     <ClInclude Include="Emu\RSX\Common\TextGlyphs.h" />
     <ClInclude Include="Emu\RSX\Common\texture_cache.h" />
@@ -532,7 +537,6 @@
     <ClInclude Include="Emu\RSX\overlay_controls.h" />
     <ClInclude Include="Emu\RSX\rsx_cache.h" />
     <ClInclude Include="Emu\RSX\rsx_decode.h" />
-    <ClInclude Include="Emu\RSX\rsx_trace.h" />
     <ClInclude Include="Emu\RSX\rsx_vertex_data.h" />
     <ClInclude Include="Emu\VFS.h" />
     <ClInclude Include="Emu\GameInfo.h" />

rpcs3/emucore.vcxproj.filters

@@ -63,6 +63,9 @@
     <Filter Include="Emu\Cell\Modules">
       <UniqueIdentifier>{4317ac27-38e4-4f8d-9bac-496f9b00f615}</UniqueIdentifier>
     </Filter>
+    <Filter Include="Emu\GPU\RSX\Capture">
+      <UniqueIdentifier>{3f233c0b-4ee1-4c02-963f-0109d3d9c705}</UniqueIdentifier>
+    </Filter>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="Crypto\aes.cpp">
@@ -743,6 +746,12 @@
     <ClCompile Include="Emu\Cell\lv2\sys_gpio.cpp">
       <Filter>Emu\Cell\lv2</Filter>
     </ClCompile>
+    <ClCompile Include="Emu\RSX\Capture\rsx_replay.cpp">
+      <Filter>Emu\GPU\RSX\Capture</Filter>
+    </ClCompile>
+    <ClCompile Include="Emu\RSX\Capture\rsx_capture.cpp">
+      <Filter>Emu\GPU\RSX\Capture</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Crypto\aes.h">
@@ -1336,9 +1345,6 @@
     <ClInclude Include="Emu\RSX\rsx_decode.h">
       <Filter>Emu\GPU\RSX</Filter>
     </ClInclude>
-    <ClInclude Include="Emu\RSX\rsx_trace.h">
-      <Filter>Emu\GPU\RSX</Filter>
-    </ClInclude>
     <ClInclude Include="Emu\RSX\gcm_enums.h">
       <Filter>Emu\GPU\RSX</Filter>
     </ClInclude>
@@ -1435,5 +1441,14 @@
     <ClInclude Include="Emu\Cell\Modules\cellJpgEnc.h">
       <Filter>Emu\Cell\Modules</Filter>
     </ClInclude>
+    <ClInclude Include="Emu\RSX\Capture\rsx_replay.h">
+      <Filter>Emu\GPU\RSX\Capture</Filter>
+    </ClInclude>
+    <ClInclude Include="Emu\RSX\Capture\rsx_trace.h">
+      <Filter>Emu\GPU\RSX\Capture</Filter>
+    </ClInclude>
+    <ClInclude Include="Emu\RSX\Capture\rsx_capture.h">
+      <Filter>Emu\GPU\RSX\Capture</Filter>
+    </ClInclude>
   </ItemGroup>
 </Project>

rpcs3/rpcs3.vcxproj

@@ -143,7 +143,7 @@
       <WarningLevel>Level3</WarningLevel>
     </ClCompile>
     <Link>
-      <AdditionalDependencies>..\hidapi.lib;winmm.lib;OpenAL.lib;XAudio.lib;D3D12GSRender.lib;GLGSRender.lib;shlwapi.lib;VKGSRender.lib;VKstatic.1.lib;glslang.lib;OSDependent.lib;OGLCompiler.lib;SPIRV.lib;HLSL.lib;Advapi32.lib;user32.lib;zlib.lib;..\libpng.lib;asmjit.lib;yaml-cpp.lib;discord-rpc.lib;emucore.lib;dxgi.lib;$(QTDIR)\lib\qtmain.lib;shell32.lib;opengl32.lib;$(QTDIR)\lib\Qt5OpenGL.lib;$(QTDIR)\lib\Qt5Widgets.lib;$(QTDIR)\lib\Qt5Gui.lib;$(QTDIR)\lib\Qt5Qml.lib;$(QTDIR)\lib\Qt5Network.lib;$(QTDIR)\lib\Qt5Core.lib;Qt5Core.lib;Qt5Gui.lib;Qt5Widgets.lib;Qt5WinExtras.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalDependencies>gdi32.lib;..\hidapi.lib;winmm.lib;OpenAL.lib;XAudio.lib;D3D12GSRender.lib;GLGSRender.lib;shlwapi.lib;VKGSRender.lib;VKstatic.1.lib;glslang.lib;OSDependent.lib;OGLCompiler.lib;SPIRV.lib;HLSL.lib;Advapi32.lib;user32.lib;zlib.lib;..\libpng.lib;asmjit.lib;yaml-cpp.lib;discord-rpc.lib;emucore.lib;dxgi.lib;$(QTDIR)\lib\qtmain.lib;shell32.lib;opengl32.lib;$(QTDIR)\lib\Qt5OpenGL.lib;$(QTDIR)\lib\Qt5Widgets.lib;$(QTDIR)\lib\Qt5Gui.lib;$(QTDIR)\lib\Qt5Qml.lib;$(QTDIR)\lib\Qt5Network.lib;$(QTDIR)\lib\Qt5Core.lib;Qt5Core.lib;Qt5Gui.lib;Qt5Widgets.lib;Qt5WinExtras.lib;%(AdditionalDependencies)</AdditionalDependencies>
       <AdditionalLibraryDirectories>..\3rdparty\OpenAL\libs\Win64;..\Vulkan\glslang-build\hlsl\Release;..\Vulkan\glslang-build\SPIRV\Release;..\Vulkan\glslang-build\OGLCompilersDLL\Release;..\Vulkan\glslang-build\glslang\OSDependent\Windows\Release;..\Vulkan\Vulkan-build\loader\Release;..\Vulkan\glslang-build\glslang\Release;..\lib\$(CONFIGURATION)-$(PLATFORM);..\3rdparty\minidx12\Lib;..\3rdparty\discord-rpc\lib;$(QTDIR)\lib;%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
       <AdditionalOptions>"/MANIFESTDEPENDENCY:type='win32' name='Microsoft.Windows.Common-Controls' version='6.0.0.0' publicKeyToken='6595b64144ccf1df' language='*' processorArchitecture='*'" %(AdditionalOptions)</AdditionalOptions>
       <DataExecutionPrevention>true</DataExecutionPrevention>

rpcs3/rpcs3qt/debugger_frame.cpp

@@ -45,7 +45,7 @@ debugger_frame::debugger_frame(std::shared_ptr<gui_settings> settings, QWidget *
 
 	m_go_to_addr = new QPushButton(tr("Go To Address"), this);
 	m_go_to_pc = new QPushButton(tr("Go To PC"), this);
-	m_btn_capture = new QPushButton(tr("Capture"), this);
+	m_btn_capture = new QPushButton(tr("RSX Capture"), this);
 	m_btn_step = new QPushButton(tr("Step"), this);
 	m_btn_step_over = new QPushButton(tr("Step Over"), this);
 	m_btn_run = new QPushButton(RunString, this);

rpcs3/rpcs3qt/main_window.cpp

@@ -331,6 +331,32 @@ void main_window::BootGame()
 	Boot(path);
 }
 
+void main_window::BootRsxCapture()
+{
+	bool stopped = false;
+	if (Emu.IsRunning())
+	{
+		Emu.Pause();
+		stopped = true;
+	}
+
+	QString filePath = QFileDialog::getOpenFileName(this, tr("Select RSX Capture"), "", tr("RRC files (*.rrc);;All files (*.*)"));
+	if (filePath.isEmpty())
+	{
+		if (stopped) Emu.Resume();
+		return;
+	}
+	Emu.SetForceBoot(true);
+	Emu.Stop();
+
+	const std::string path = sstr(filePath);
+
+	if (!Emu.BootRsxCapture(path))
+		LOG_ERROR(GENERAL, "Capture Boot Failed");
+	else
+		LOG_SUCCESS(LOADER, "Capture Boot Success");
+}   
+
 void main_window::InstallPkg(const QString& dropPath)
 {
 	QString filePath = dropPath;
@@ -1068,6 +1094,7 @@ void main_window::CreateConnects()
 {
 	connect(ui->bootElfAct, &QAction::triggered, this, &main_window::BootElf);
 	connect(ui->bootGameAct, &QAction::triggered, this, &main_window::BootGame);
+	connect(ui->actionopen_rsx_capture, &QAction::triggered, this, &main_window::BootRsxCapture);
 
 	connect(ui->bootRecentMenu, &QMenu::aboutToShow, [=]
 	{

rpcs3/rpcs3qt/main_window.h

@@ -90,6 +90,7 @@ private Q_SLOTS:
 	void Boot(const std::string& path, bool direct = false, bool add_only = false);
 	void BootElf();
 	void BootGame();
+	void BootRsxCapture();
 	void DecryptSPRXLibraries();
 
 	void SaveWindowState();

rpcs3/rpcs3qt/main_window.ui

@@ -221,6 +221,7 @@
     <addaction name="toolsStringSearchAct"/>
     <addaction name="separator"/>
     <addaction name="toolsDecryptSprxLibsAct"/>
+    <addaction name="actionopen_rsx_capture"/>
    </widget>
    <widget class="QMenu" name="menuView">
     <property name="title">
@@ -260,6 +261,7 @@
     <addaction name="showToolBarAct"/>
     <addaction name="separator"/>
     <addaction name="showGameListAct"/>
+    <addaction name="showGameToolBarAct"/>
     <addaction name="separator"/>
     <addaction name="showHiddenEntriesAct"/>
     <addaction name="separator"/>
@@ -951,6 +953,11 @@
     <string>Show Hidden Entries</string>
    </property>
   </action>
+  <action name="actionopen_rsx_capture">
+   <property name="text">
+    <string>Open RSX Capture</string>
+   </property>
+  </action>
  </widget>
  <layoutdefault spacing="6" margin="11"/>
  <resources>

rpcs3_default.props

@@ -3,7 +3,7 @@
   <ImportGroup Label="PropertySheets" />
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup>
-    <IncludePath>.\;..\;..\asmjit\src\asmjit;..\Utilities\yaml-cpp\include;..\3rdparty\ffmpeg\WindowsInclude;..\3rdparty\cereal\include;..\3rdparty\ffmpeg\Windows\x86_64\Include;$(VC_IncludePath);$(WindowsSDK_IncludePath);$(UniversalCRT_IncludePath);..\3rdparty\minidx12\Include;..\3rdparty\GSL\include;..\3rdparty\libpng;..\3rdparty\GL;..\3rdparty\stblib;..\3rdparty\OpenAL\include;..\3rdparty\pugixml\src;..\3rdparty\hidapi\hidapi;..\3rdparty\Optional</IncludePath>
+    <IncludePath>.\;..\;..\asmjit\src\asmjit;..\Utilities\yaml-cpp\include;..\3rdparty\ffmpeg\WindowsInclude;..\3rdparty\cereal\include;..\3rdparty\ffmpeg\Windows\x86_64\Include;$(VC_IncludePath);$(WindowsSDK_IncludePath);$(UniversalCRT_IncludePath);..\3rdparty\minidx12\Include;..\3rdparty\GSL\include;..\3rdparty\libpng;..\3rdparty\GL;..\3rdparty\stblib;..\3rdparty\OpenAL\include;..\3rdparty\pugixml\src;..\3rdparty\hidapi\hidapi;..\3rdparty\Optional;..\3rdparty\xxhash</IncludePath>
     <OutDir>$(SolutionDir)lib\$(Configuration)-$(Platform)\</OutDir>
     <LibraryPath>$(SolutionDir)lib\$(Configuration)-$(Platform)\;$(UniversalCRT_LibraryPath_x64);$(LibraryPath)</LibraryPath>
     <IntDir>$(SolutionDir)tmp\$(ProjectName)-$(Configuration)-$(Platform)\</IntDir>
@@ -21,7 +21,7 @@
       <AdditionalOptions>/Zc:throwingNew</AdditionalOptions>
     </ClCompile>
     <Link>
-      <AdditionalDependencies>ws2_32.lib;avcodec.lib;avformat.lib;avutil.lib;swresample.lib;swscale.lib</AdditionalDependencies>
+      <AdditionalDependencies>xxhash.lib;ws2_32.lib;avcodec.lib;avformat.lib;avutil.lib;swresample.lib;swscale.lib</AdditionalDependencies>
       <AdditionalLibraryDirectories>..\3rdparty\ffmpeg\Windows\x86_64\lib</AdditionalLibraryDirectories>
     </Link>
   </ItemDefinitionGroup>