Remove loops in RSP code (#136)

* Remove loops in RSP cose

* more

* more

* more

* more

* more

.cargo/config.toml

@@ -1,8 +1,5 @@
-[target.x86_64-unknown-linux-gnu]
+[target.'cfg(target_arch = "x86_64")']
 rustflags = ["-C", "target-cpu=x86-64-v3"]
 
-[target.x86_64-pc-windows-msvc]
-rustflags = ["-C", "target-cpu=x86-64-v3"]
-
-[target.x86_64-apple-darwin]
-rustflags = ["-C", "target-cpu=x86-64-v2"]
+[target.'cfg(target_arch = "aarch64")']
+rustflags = ["-C", "target-cpu=cortex-a76"]

.github/FUNDING.yml

@@ -0,0 +1,2 @@
+github: loganmc10
+patreon: loganmc10

.github/workflows/build.yml

@@ -12,60 +12,113 @@ env:
 jobs:
   build-linux:
     runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        target: [x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu]
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
-      - name: Install sdl2 headers
+      - uses: actions-rust-lang/setup-rust-toolchain@v1
+        with:
+          target: ${{ matrix.target }}
+          rustflags: ""
+      - name: Install SDL2
+        if: matrix.target == 'aarch64-unknown-linux-gnu'
         run: |
+          cat <<EOF > deb822sources
+          Types: deb
+          URIs: http://archive.ubuntu.com/ubuntu/
+          Suites: noble
+          Components: main restricted universe
+          Architectures: amd64
+
+          Types: deb
+          URIs: http://security.ubuntu.com/ubuntu/
+          Suites: noble-security
+          Components: main restricted universe
+          Architectures: amd64
+
+          Types: deb
+          URIs: http://archive.ubuntu.com/ubuntu/
+          Suites: noble-updates
+          Components: main restricted universe
+          Architectures: amd64
+
+          Types: deb
+          URIs: http://azure.ports.ubuntu.com/ubuntu-ports/
+          Suites: noble
+          Components: main restricted multiverse universe
+          Architectures: arm64
+
+          Types: deb
+          URIs: http://azure.ports.ubuntu.com/ubuntu-ports/
+          Suites: noble-updates
+          Components: main restricted multiverse universe
+          Architectures: arm64
+          EOF
+
+          sudo mv deb822sources /etc/apt/sources.list.d/ubuntu.sources
+          sudo dpkg --add-architecture arm64
+          sudo apt update
+          sudo apt -y install libsdl2-dev:arm64 g++-aarch64-linux-gnu
+      - name: Install SDL2
+        if: matrix.target == 'x86_64-unknown-linux-gnu'
+        run: |
+          sudo apt update
           sudo apt -y install libsdl2-dev
-      - name: Update rust
-        run: |
-          rustup update
+      - name: Set aarch64 linker
+        if: matrix.target == 'aarch64-unknown-linux-gnu'
+        run: echo "CARGO_TARGET_AARCH64_UNKNOWN_LINUX_GNU_LINKER=aarch64-linux-gnu-gcc" >> $GITHUB_ENV
       - name: Build
         run: |
-          cargo build --release -v
+          cargo build --release -v --target=${{ matrix.target }}
       - name: Upload file
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          name: gopher64-linux
-          path: target/release/gopher64
+          name: gopher64-${{ matrix.target }}
+          path: target/${{ matrix.target }}/release/gopher64
 
   build-windows:
     runs-on: windows-latest
+    strategy:
+      matrix:
+        target: [x86_64-pc-windows-msvc]
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
-      - name: Update rust
-        run: |
-          rustup update
+      - uses: actions-rust-lang/setup-rust-toolchain@v1
+        with:
+          target: ${{ matrix.target }}
+          rustflags: ""
       - name: Build
         run: |
-          cargo build --release -v
+          cargo build --release -v --target=${{ matrix.target }}
       - name: Upload file
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          name: gopher64-windows
-          path: target/release/gopher64.exe
+          name: gopher64-${{ matrix.target }}
+          path: target/${{ matrix.target }}/release/gopher64.exe
 
   build-macos:
     runs-on: macos-latest
+    strategy:
+      matrix:
+        target: [x86_64-apple-darwin, aarch64-apple-darwin]
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
-      - name: Install sdl2 headers
-        run: |
-          brew install sdl2
-      - name: Update rust
-        run: |
-          rustup update
+      - uses: actions-rust-lang/setup-rust-toolchain@v1
+        with:
+          target: ${{ matrix.target }}
+          rustflags: ""
       - name: Build
         run: |
-          cargo build --release -v
+          cargo build --release -v --target=${{ matrix.target }}
       - name: Upload file
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          name: gopher64-mac
-          path: target/release/gopher64
+          name: gopher64-${{ matrix.target }}
+          path: target/${{ matrix.target }}/release/gopher64

.github/workflows/lint.yml

@@ -16,13 +16,11 @@ jobs:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
-      - name: Install sdl2 headers
-        run: |
-          sudo apt -y install libsdl2-dev
-      - name: Update rust
-        run: |
-          rustup update
-          rustup component add clippy
+      - uses: actions-rust-lang/setup-rust-toolchain@v1
+        with:
+          components: clippy, rustfmt
+      - name: Rustfmt Check
+        uses: actions-rust-lang/rustfmt@v1
       - name: Lint
         run: |
           cargo clippy -- -Dwarnings

.vscode/settings.json

@@ -1,5 +0,0 @@
-{
-    "rust-analyzer.linkedProjects": [
-        "./Cargo.toml"
-    ]
-}

Cargo.toml

@@ -1,34 +1,37 @@
 [package]
 name = "gopher64"
-version = "0.1.0"
+version = "0.1.17"
 edition = "2021"
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
-dirs = "5.0"
-zip = "0.6"
-governor = "0.6"
-sevenz-rust = "0.5"
+dirs = "6.0"
+zip = "2.1"
+governor = "0.8"
+sevenz-rust = "0.6"
 chrono = "0.4"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
-eframe = "0.23"
-rfd = {version = "0.12",  default-features = false, features = ["xdg-portal"] }
+eframe = "0.30"
+rfd = "0.15"
 sha2 = "0.10"
 rusttype = "0.9"
 futures = "0.3"
-sdl2-sys = { git = "https://github.com/Rust-SDL2/rust-sdl2.git" }
+sdl2-sys = "0.37"
+spin_sleep = "1.3"
+encoding_rs = "0.8"
 clap = { version = "4.4", features = ["derive"] }
 
 [dependencies.sdl2]
-git = "https://github.com/Rust-SDL2/rust-sdl2.git"
+version = "0.37"
 default-features = false
 features = ["static-link","bundled"]
 
 [build-dependencies]
 winres = "0.1"
-cc = { version = "1.0", features = ["parallel"] }
+bindgen = "0.71"
+cc = { version = "1.2", features = ["parallel"] }
 
 [profile.release]
 lto = true

README.md

@@ -46,6 +46,10 @@ You can create you own mappings by running `./gopher64 --configure-input-profile
 
 In order to use a controller (for example, an Xbox controller), run `./gopher64 --list-controllers` to get a list of attached controllers, and then assign it by doing `./gopher64 --assign-controller <controller_number> --port 1`
 
+## portable mode
+
+If you would like to keep all the game data in the same folder as executable, you just need to create a file called "portable.txt" in the same directory as the executable.
+
 ## contributing
 
 I am very open to contributions! Please reach out to me via a GitHub issue, or via discord (loganmc10) before doing substantial work on a PR.
@@ -56,4 +60,4 @@ Gopher64 is licensed under the GPLv3 license. Many portions of gopher64 have bee
 
 ## where to download
 
-Builds can be found as artifacts on [GitHub Actions](https://github.com/gopher64/gopher64/actions?query=branch%3Amain)
+https://github.com/gopher64/gopher64/releases

build.rs

@@ -1,7 +1,12 @@
 fn main() {
+    println!("cargo::rerun-if-changed=parallel-rdp");
+    println!("cargo::rerun-if-changed=src/compat");
+
+    let mut simd_build = cc::Build::new();
     let mut build = cc::Build::new();
     build
         .cpp(true)
+        .warnings(false)
         .std("c++17")
         .file("parallel-rdp/parallel-rdp-standalone/parallel-rdp/command_ring.cpp")
         .file("parallel-rdp/parallel-rdp-standalone/parallel-rdp/rdp_device.cpp")
@@ -37,6 +42,7 @@ fn main() {
         .file("parallel-rdp/parallel-rdp-standalone/util/aligned_alloc.cpp")
         .file("parallel-rdp/parallel-rdp-standalone/util/timer.cpp")
         .file("parallel-rdp/parallel-rdp-standalone/util/timeline_trace_file.cpp")
+        .file("parallel-rdp/parallel-rdp-standalone/util/environment.cpp")
         .file("parallel-rdp/parallel-rdp-standalone/util/thread_name.cpp")
         .file("parallel-rdp/parallel-rdp-standalone/volk/volk.c")
         .file("parallel-rdp/interface.cpp")
@@ -48,20 +54,106 @@ fn main() {
         .include("parallel-rdp/parallel-rdp-standalone/util")
         .includes(std::env::var("DEP_SDL2_INCLUDE"));
 
-    #[cfg(target_os = "windows")]
-    {
+    let os = std::env::var("CARGO_CFG_TARGET_OS").unwrap();
+    let arch = std::env::var("CARGO_CFG_TARGET_ARCH").unwrap();
+    if os == "windows" {
+        if arch == "x86_64" {
+            build.flag("/arch:AVX2");
+        } else if arch == "aarch64" {
+            panic!("unsupported platform")
+        } else {
+            panic!("unknown arch")
+        }
         build.flag("-DVK_USE_PLATFORM_WIN32_KHR");
 
         winres::WindowsResource::new()
             .set_icon("data/icon.ico")
             .compile()
             .unwrap();
+    } else if os == "linux" || os == "macos" {
+        if arch == "x86_64" {
+            build.flag("-march=x86-64-v3");
+        } else if arch == "aarch64" {
+            build.flag("-march=armv8.2-a");
+            simd_build.flag("-march=armv8.2-a");
+            simd_build.flag("-DSSE2NEON_SUPPRESS_WARNINGS");
+            simd_build.file("src/compat/aarch64.c");
+        } else {
+            panic!("unknown arch")
+        }
+    } else {
+        panic!("unknown OS")
     }
-    #[cfg(any(target_os = "linux", target_os = "macos"))]
-    {
-        build
-            .flag("-Wno-missing-field-initializers")
-            .flag("-Wno-unused-parameter");
-    }
+
     build.compile("parallel-rdp");
+
+    let out_path = std::path::PathBuf::from(std::env::var("OUT_DIR").unwrap());
+
+    let parallel_bindings = bindgen::Builder::default()
+        .header("parallel-rdp/interface.hpp")
+        .allowlist_function("rdp_init")
+        .allowlist_function("rdp_close")
+        .allowlist_function("rdp_set_vi_register")
+        .allowlist_function("rdp_update_screen")
+        .allowlist_function("rdp_process_commands")
+        .allowlist_function("rdp_full_sync")
+        .parse_callbacks(Box::new(bindgen::CargoCallbacks::new()))
+        .generate()
+        .expect("Unable to generate bindings");
+
+    parallel_bindings
+        .write_to_file(out_path.join("parallel_bindings.rs"))
+        .expect("Couldn't write bindings!");
+
+    if arch == "aarch64" {
+        let simd_bindings = bindgen::Builder::default()
+            .header("src/compat/sse2neon.h")
+            .allowlist_function("_mm_setzero_si128")
+            .allowlist_function("_mm_set_epi8")
+            .allowlist_function("_mm_movemask_epi8")
+            .allowlist_function("_mm_shuffle_epi8")
+            .allowlist_function("_mm_packs_epi16")
+            .allowlist_function("_mm_set_epi16")
+            .allowlist_function("_mm_cmpeq_epi8")
+            .allowlist_function("_mm_and_si128")
+            .allowlist_function("_mm_set1_epi8")
+            .allowlist_function("_mm_mullo_epi16")
+            .allowlist_function("_mm_cmpeq_epi16")
+            .allowlist_function("_mm_add_epi16")
+            .allowlist_function("_mm_slli_epi16")
+            .allowlist_function("_mm_mulhi_epi16")
+            .allowlist_function("_mm_srai_epi16")
+            .allowlist_function("_mm_andnot_si128")
+            .allowlist_function("_mm_or_si128")
+            .allowlist_function("_mm_mulhi_epu16")
+            .allowlist_function("_mm_sub_epi16")
+            .allowlist_function("_mm_unpacklo_epi16")
+            .allowlist_function("_mm_unpackhi_epi16")
+            .allowlist_function("_mm_packs_epi32")
+            .allowlist_function("_mm_adds_epu16")
+            .allowlist_function("_mm_cmpgt_epi16")
+            .allowlist_function("_mm_blendv_epi8")
+            .allowlist_function("_mm_min_epi16")
+            .allowlist_function("_mm_max_epi16")
+            .allowlist_function("_mm_subs_epi16")
+            .allowlist_function("_mm_adds_epi16")
+            .allowlist_function("_mm_xor_si128")
+            .allowlist_function("_mm_cmplt_epi16")
+            .allowlist_function("_mm_subs_epu16")
+            .allowlist_function("_mm_set1_epi32")
+            .blocklist_type("__m128i")
+            .blocklist_type("int64x2_t")
+            .wrap_static_fns(true)
+            .parse_callbacks(Box::new(bindgen::CargoCallbacks::new()))
+            .generate()
+            .expect("Unable to generate bindings");
+
+        simd_bindings
+            .write_to_file(out_path.join("simd_bindings.rs"))
+            .expect("Couldn't write bindings!");
+
+        simd_build.file(std::env::temp_dir().join("bindgen").join("extern.c"));
+        simd_build.include(".");
+        simd_build.compile("simd");
+    }
 }

parallel-rdp/interface.cpp

@@ -51,16 +51,16 @@ enum vi_registers
 };
 
 static bool fullscreen;
-static void *rdram;
 static SDL_Window *window;
 static RDP::CommandProcessor *processor;
+static SDL_WSIPlatform *wsi_platform;
 static WSI *wsi;
 static uint32_t cmd_data[0x00040000 >> 2];
 static int cmd_cur;
 static int cmd_ptr;
-static uint8_t emu_running;
+static bool emu_running;
 static uint64_t rdp_sync_signal;
-static uint32_t vi_registers[14];
+static GFX_INFO gfx_info;
 
 static uint64_t last_frame_counter;
 static uint64_t frame_counter;
@@ -132,54 +132,6 @@ static const unsigned cmd_len_lut[64] = {
 	1,
 };
 
-enum
-{
-	MB_RDRAM_DRAM_ALIGNMENT_REQUIREMENT = 64 * 1024
-};
-
-void vk_init(void *mem_base, uint32_t rdram_size, uint8_t _fullscreen)
-{
-	fullscreen = _fullscreen != 0;
-	rdram = mem_base;
-	bool window_vsync = 0;
-	wsi = new WSI;
-	SDL_WSIPlatform *wsi_platform = new SDL_WSIPlatform;
-	wsi_platform->set_window(window);
-	wsi->set_platform(wsi_platform);
-	wsi->set_present_mode(window_vsync ? PresentMode::SyncToVBlank : PresentMode::UnlockedMaybeTear);
-	wsi->set_backbuffer_srgb(false);
-	Context::SystemHandles handles = {};
-	if (!::Vulkan::Context::init_loader(nullptr))
-	{
-		vk_close();
-	}
-	if (!wsi->init_simple(1, handles))
-	{
-		vk_close();
-	}
-	RDP::CommandProcessorFlags flags = 0;
-	processor = new RDP::CommandProcessor(wsi->get_device(), rdram, 0, rdram_size, rdram_size / 2, flags);
-
-	if (!processor->device_is_supported())
-	{
-		delete processor;
-		delete wsi;
-		processor = nullptr;
-		vk_close();
-	}
-	wsi->begin_frame();
-
-	emu_running = 1;
-	last_frame_counter = 0;
-	frame_counter = 0;
-}
-
-void vk_close()
-{
-	delete processor;
-	delete wsi;
-}
-
 int sdl_event_filter(void *userdata, SDL_Event *event)
 {
 	if (event->type == SDL_WINDOWEVENT)
@@ -187,7 +139,10 @@ int sdl_event_filter(void *userdata, SDL_Event *event)
 		switch (event->window.event)
 		{
 		case SDL_WINDOWEVENT_CLOSE:
-			emu_running = 0;
+			emu_running = false;
+			break;
+		case SDL_WINDOWEVENT_RESIZED:
+			wsi_platform->do_resize();
 			break;
 		default:
 			break;
@@ -198,7 +153,7 @@ int sdl_event_filter(void *userdata, SDL_Event *event)
 		switch (event->key.keysym.scancode)
 		{
 		case SDL_SCANCODE_ESCAPE:
-			emu_running = 0;
+			emu_running = false;
 			break;
 		default:
 			break;
@@ -208,10 +163,70 @@ int sdl_event_filter(void *userdata, SDL_Event *event)
 	return 0;
 }
 
-void set_sdl_window(void *_window)
+void rdp_init(void *_window, GFX_INFO _gfx_info, bool _fullscreen, bool _upscale)
 {
 	window = (SDL_Window *)_window;
 	SDL_SetEventFilter(sdl_event_filter, nullptr);
+
+	gfx_info = _gfx_info;
+	fullscreen = _fullscreen;
+	bool window_vsync = 0;
+	wsi = new WSI;
+	wsi_platform = new SDL_WSIPlatform;
+	wsi_platform->set_window(window);
+	wsi->set_platform(wsi_platform);
+	wsi->set_present_mode(window_vsync ? PresentMode::SyncToVBlank : PresentMode::UnlockedMaybeTear);
+	wsi->set_backbuffer_srgb(false);
+	Context::SystemHandles handles = {};
+	if (!::Vulkan::Context::init_loader(nullptr))
+	{
+		rdp_close();
+	}
+	if (!wsi->init_simple(1, handles))
+	{
+		rdp_close();
+	}
+	RDP::CommandProcessorFlags flags = 0;
+	if (_upscale)
+	{
+		flags |= RDP::COMMAND_PROCESSOR_FLAG_UPSCALING_2X_BIT;
+		flags |= RDP::COMMAND_PROCESSOR_FLAG_SUPER_SAMPLED_DITHER_BIT;
+	}
+	processor = new RDP::CommandProcessor(wsi->get_device(), gfx_info.RDRAM, 0, gfx_info.RDRAM_SIZE, gfx_info.RDRAM_SIZE / 2, flags);
+
+	if (!processor->device_is_supported())
+	{
+		delete processor;
+		delete wsi;
+		processor = nullptr;
+		rdp_close();
+	}
+	wsi->begin_frame();
+
+	emu_running = true;
+	last_frame_counter = 0;
+	frame_counter = 0;
+}
+
+void rdp_close()
+{
+	wsi->end_frame();
+
+	if (processor)
+	{
+		delete processor;
+		processor = nullptr;
+	}
+	if (wsi)
+	{
+		delete wsi;
+		wsi = nullptr;
+	}
+	if (wsi_platform)
+	{
+		delete wsi_platform;
+		wsi_platform = nullptr;
+	}
 }
 
 static void calculate_viewport(float *x, float *y, float *width, float *height)
@@ -249,6 +264,9 @@ static void calculate_viewport(float *x, float *y, float *width, float *height)
 static void render_frame(Vulkan::Device &device)
 {
 	RDP::ScanoutOptions options = {};
+	options.persist_frame_on_invalid_input = true;
+	options.blend_previous_frame = true;
+	options.upscale_deinterlacing = false;
 	Vulkan::ImageHandle image = processor->scanout(options);
 
 	// Normally reflection is automated.
@@ -297,10 +315,9 @@ static void render_frame(Vulkan::Device &device)
 void rdp_set_vi_register(uint32_t reg, uint32_t value)
 {
 	processor->set_vi_register(RDP::VIRegister(reg), value);
-	vi_registers[reg] = value;
 }
 
-uint8_t rdp_update_screen()
+bool rdp_update_screen()
 {
 	auto &device = wsi->get_device();
 	render_frame(device);
@@ -349,11 +366,11 @@ static uint32_t viCalculateVerticalHeight(uint32_t vstart, uint32_t yscale)
 	return (delta * scale) / 0x800;
 }
 
-uint64_t rdp_process_commands(uint32_t *dpc_regs, uint8_t *SP_DMEM)
+uint64_t rdp_process_commands()
 {
 	uint64_t interrupt_timer = 0;
-	const uint32_t DP_CURRENT = dpc_regs[DPC_CURRENT_REG] & 0x00FFFFF8;
-	const uint32_t DP_END = dpc_regs[DPC_END_REG] & 0x00FFFFF8;
+	const uint32_t DP_CURRENT = *gfx_info.DPC_CURRENT_REG & 0x00FFFFF8;
+	const uint32_t DP_END = *gfx_info.DPC_END_REG & 0x00FFFFF8;
 
 	int length = DP_END - DP_CURRENT;
 	if (length <= 0)
@@ -363,16 +380,16 @@ uint64_t rdp_process_commands(uint32_t *dpc_regs, uint8_t *SP_DMEM)
 	if ((cmd_ptr + length) & ~(0x0003FFFF >> 3))
 		return interrupt_timer;
 
-	dpc_regs[DPC_STATUS_REG] |= DP_STATUS_PIPE_BUSY | DP_STATUS_START_GCLK;
+	*gfx_info.DPC_STATUS_REG |= DP_STATUS_PIPE_BUSY | DP_STATUS_START_GCLK;
 
 	uint32_t offset = DP_CURRENT;
-	if (dpc_regs[DPC_STATUS_REG] & DP_STATUS_XBUS_DMA)
+	if (*gfx_info.DPC_STATUS_REG & DP_STATUS_XBUS_DMA)
 	{
 		do
 		{
 			offset &= 0xFF8;
-			cmd_data[2 * cmd_ptr + 0] = SDL_SwapBE32(*reinterpret_cast<const uint32_t *>(SP_DMEM + offset));
-			cmd_data[2 * cmd_ptr + 1] = SDL_SwapBE32(*reinterpret_cast<const uint32_t *>(SP_DMEM + offset + 4));
+			cmd_data[2 * cmd_ptr + 0] = SDL_SwapBE32(*reinterpret_cast<const uint32_t *>(gfx_info.DMEM + offset));
+			cmd_data[2 * cmd_ptr + 1] = SDL_SwapBE32(*reinterpret_cast<const uint32_t *>(gfx_info.DMEM + offset + 4));
 			offset += sizeof(uint64_t);
 			cmd_ptr++;
 		} while (--length > 0);
@@ -388,8 +405,8 @@ uint64_t rdp_process_commands(uint32_t *dpc_regs, uint8_t *SP_DMEM)
 			do
 			{
 				offset &= 0xFFFFF8;
-				cmd_data[2 * cmd_ptr + 0] = *reinterpret_cast<const uint32_t *>((uint8_t *)rdram + offset);
-				cmd_data[2 * cmd_ptr + 1] = *reinterpret_cast<const uint32_t *>((uint8_t *)rdram + offset + 4);
+				cmd_data[2 * cmd_ptr + 0] = *reinterpret_cast<const uint32_t *>(gfx_info.RDRAM + offset);
+				cmd_data[2 * cmd_ptr + 1] = *reinterpret_cast<const uint32_t *>(gfx_info.RDRAM + offset + 4);
 				offset += sizeof(uint64_t);
 				cmd_ptr++;
 			} while (--length > 0);
@@ -404,7 +421,7 @@ uint64_t rdp_process_commands(uint32_t *dpc_regs, uint8_t *SP_DMEM)
 
 		if (cmd_ptr - cmd_cur - cmd_length < 0)
 		{
-			dpc_regs[DPC_START_REG] = dpc_regs[DPC_CURRENT_REG] = dpc_regs[DPC_END_REG];
+			*gfx_info.DPC_START_REG = *gfx_info.DPC_CURRENT_REG = *gfx_info.DPC_END_REG;
 			return interrupt_timer;
 		}
 
@@ -423,19 +440,19 @@ uint64_t rdp_process_commands(uint32_t *dpc_regs, uint8_t *SP_DMEM)
 				rdp_sync_signal = 0;
 			}
 
-			uint32_t width = viCalculateHorizonalWidth(vi_registers[VI_H_START_REG], vi_registers[VI_X_SCALE_REG], vi_registers[VI_WIDTH_REG]);
+			uint32_t width = viCalculateHorizonalWidth(*gfx_info.VI_H_START_REG, *gfx_info.VI_X_SCALE_REG, *gfx_info.VI_WIDTH_REG);
 			if (width == 0)
 			{
 				width = 320;
 			}
-			uint32_t height = viCalculateVerticalHeight(vi_registers[VI_V_START_REG], vi_registers[VI_Y_SCALE_REG]);
+			uint32_t height = viCalculateVerticalHeight(*gfx_info.VI_V_START_REG, *gfx_info.VI_Y_SCALE_REG);
 			if (height == 0)
 			{
 				height = 240;
 			}
 			interrupt_timer = width * height * 4;
 
-			dpc_regs[DPC_STATUS_REG] &= ~(DP_STATUS_PIPE_BUSY | DP_STATUS_START_GCLK);
+			*gfx_info.DPC_STATUS_REG &= ~(DP_STATUS_PIPE_BUSY | DP_STATUS_START_GCLK);
 		}
 
 		cmd_cur += cmd_length;
@@ -443,13 +460,13 @@ uint64_t rdp_process_commands(uint32_t *dpc_regs, uint8_t *SP_DMEM)
 
 	cmd_ptr = 0;
 	cmd_cur = 0;
-	dpc_regs[DPC_CURRENT_REG] = dpc_regs[DPC_END_REG];
-	dpc_regs[DPC_STATUS_REG] |= DP_STATUS_CBUF_READY;
+	*gfx_info.DPC_CURRENT_REG = *gfx_info.DPC_END_REG;
+	*gfx_info.DPC_STATUS_REG |= DP_STATUS_CBUF_READY;
 
 	return interrupt_timer;
 }
 
-void full_sync()
+void rdp_full_sync()
 {
 	if (rdp_sync_signal)
 	{

parallel-rdp/interface.hpp

@@ -1,17 +1,34 @@
 #pragma once
 
 #ifdef __cplusplus
+#include <cstdint>
+
 extern "C"
 {
 #endif
 
-    void vk_init(void *mem_base, uint32_t rdram_size, uint8_t fullscreen);
-    void vk_close();
-    void set_sdl_window(void *_window);
-    void rdp_set_vi_register(uint32_t reg, uint32_t value);
-    uint8_t rdp_update_screen();
-    uint64_t rdp_process_commands(uint32_t *dpc_regs, uint8_t *SP_DMEM);
-    void full_sync();
+	typedef struct
+	{
+		uint8_t *RDRAM;
+		uint8_t *DMEM;
+		uint32_t RDRAM_SIZE;
+		uint32_t *DPC_CURRENT_REG;
+		uint32_t *DPC_START_REG;
+		uint32_t *DPC_END_REG;
+		uint32_t *DPC_STATUS_REG;
+		uint32_t *VI_H_START_REG;
+		uint32_t *VI_V_START_REG;
+		uint32_t *VI_X_SCALE_REG;
+		uint32_t *VI_Y_SCALE_REG;
+		uint32_t *VI_WIDTH_REG;
+	} GFX_INFO;
+
+	void rdp_init(void *_window, GFX_INFO _gfx_info, bool fullscreen, bool _upscale);
+	void rdp_close();
+	void rdp_set_vi_register(uint32_t reg, uint32_t value);
+	bool rdp_update_screen();
+	uint64_t rdp_process_commands();
+	void rdp_full_sync();
 
 #ifdef __cplusplus
 }

parallel-rdp/wsi_platform.cpp

@@ -54,7 +54,17 @@ void SDL_WSIPlatform::poll_input()
 	SDL_PumpEvents();
 }
 
+void SDL_WSIPlatform::poll_input_async(Granite::InputTrackerHandler *handler)
+{
+	SDL_PumpEvents();
+}
+
 void SDL_WSIPlatform::set_window(SDL_Window *_window)
 {
 	window = _window;
 }
+
+void SDL_WSIPlatform::do_resize()
+{
+	resize = true;
+}

parallel-rdp/wsi_platform.hpp

@@ -13,9 +13,10 @@ public:
 	uint32_t get_surface_height() override;
 	bool alive(Vulkan::WSI &wsi) override;
 	void poll_input() override;
+	void poll_input_async(Granite::InputTrackerHandler *handler) override;
 	void set_window(SDL_Window *_window);
+	void do_resize();
 
 private:
-	VkSurfaceKHR surface;
 	SDL_Window *window;
 };

rust-toolchain.toml

@@ -1,2 +0,0 @@
-[toolchain]
-channel = "nightly"

src/compat/aarch64.c

@@ -0,0 +1,197 @@
+#include "src/compat/sse2neon.h"
+
+// Static wrappers
+
+__m128i _mm_srli_epi16__extern(__m128i __a, int __imm8) { return _mm_srli_epi16(__a, __imm8); }
+
+__m128i _mm_insert_epi8__extern(__m128i __a, int __i, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_insert_epi8(__a, __i, 0);
+    case 1:
+        return _mm_insert_epi8(__a, __i, 1);
+    case 2:
+        return _mm_insert_epi8(__a, __i, 2);
+    case 3:
+        return _mm_insert_epi8(__a, __i, 3);
+    case 4:
+        return _mm_insert_epi8(__a, __i, 4);
+    case 5:
+        return _mm_insert_epi8(__a, __i, 5);
+    case 6:
+        return _mm_insert_epi8(__a, __i, 6);
+    case 7:
+        return _mm_insert_epi8(__a, __i, 7);
+    case 8:
+        return _mm_insert_epi8(__a, __i, 8);
+    case 9:
+        return _mm_insert_epi8(__a, __i, 9);
+    case 10:
+        return _mm_insert_epi8(__a, __i, 10);
+    case 11:
+        return _mm_insert_epi8(__a, __i, 11);
+    case 12:
+        return _mm_insert_epi8(__a, __i, 12);
+    case 13:
+        return _mm_insert_epi8(__a, __i, 13);
+    case 14:
+        return _mm_insert_epi8(__a, __i, 14);
+    case 15:
+        return _mm_insert_epi8(__a, __i, 15);
+    default:
+        abort();
+    }
+}
+
+int _mm_extract_epi8__extern(__m128i __a, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_extract_epi8(__a, 0);
+    case 1:
+        return _mm_extract_epi8(__a, 1);
+    case 2:
+        return _mm_extract_epi8(__a, 2);
+    case 3:
+        return _mm_extract_epi8(__a, 3);
+    case 4:
+        return _mm_extract_epi8(__a, 4);
+    case 5:
+        return _mm_extract_epi8(__a, 5);
+    case 6:
+        return _mm_extract_epi8(__a, 6);
+    case 7:
+        return _mm_extract_epi8(__a, 7);
+    case 8:
+        return _mm_extract_epi8(__a, 8);
+    case 9:
+        return _mm_extract_epi8(__a, 9);
+    case 10:
+        return _mm_extract_epi8(__a, 10);
+    case 11:
+        return _mm_extract_epi8(__a, 11);
+    case 12:
+        return _mm_extract_epi8(__a, 12);
+    case 13:
+        return _mm_extract_epi8(__a, 13);
+    case 14:
+        return _mm_extract_epi8(__a, 14);
+    case 15:
+        return _mm_extract_epi8(__a, 15);
+    default:
+        abort();
+    }
+}
+
+__m128i _mm_insert_epi16__extern(__m128i __a, int __i, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_insert_epi16(__a, __i, 0);
+    case 1:
+        return _mm_insert_epi16(__a, __i, 1);
+    case 2:
+        return _mm_insert_epi16(__a, __i, 2);
+    case 3:
+        return _mm_insert_epi16(__a, __i, 3);
+    case 4:
+        return _mm_insert_epi16(__a, __i, 4);
+    case 5:
+        return _mm_insert_epi16(__a, __i, 5);
+    case 6:
+        return _mm_insert_epi16(__a, __i, 6);
+    case 7:
+        return _mm_insert_epi16(__a, __i, 7);
+    default:
+        abort();
+    }
+}
+
+int _mm_extract_epi16__extern(__m128i __a, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_extract_epi16(__a, 0);
+    case 1:
+        return _mm_extract_epi16(__a, 1);
+    case 2:
+        return _mm_extract_epi16(__a, 2);
+    case 3:
+        return _mm_extract_epi16(__a, 3);
+    case 4:
+        return _mm_extract_epi16(__a, 4);
+    case 5:
+        return _mm_extract_epi16(__a, 5);
+    case 6:
+        return _mm_extract_epi16(__a, 6);
+    case 7:
+        return _mm_extract_epi16(__a, 7);
+    default:
+        abort();
+    }
+}
+
+__m128i _mm_insert_epi32__extern(__m128i __a, int __i, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_insert_epi32(__a, __i, 0);
+    case 1:
+        return _mm_insert_epi32(__a, __i, 1);
+    case 2:
+        return _mm_insert_epi32(__a, __i, 2);
+    case 3:
+        return _mm_insert_epi32(__a, __i, 3);
+    default:
+        abort();
+    }
+}
+
+int _mm_extract_epi32__extern(__m128i __a, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_extract_epi32(__a, 0);
+    case 1:
+        return _mm_extract_epi32(__a, 1);
+    case 2:
+        return _mm_extract_epi32(__a, 2);
+    case 3:
+        return _mm_extract_epi32(__a, 3);
+    default:
+        abort();
+    }
+}
+
+__m128i _mm_insert_epi64__extern(__m128i __a, int64_t __i, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_insert_epi64(__a, __i, 0);
+    case 1:
+        return _mm_insert_epi64(__a, __i, 1);
+    default:
+        abort();
+    }
+}
+
+int64_t _mm_extract_epi64__extern(__m128i __a, const int __imm8)
+{
+    switch (__imm8)
+    {
+    case 0:
+        return _mm_extract_epi64(__a, 0);
+    case 1:
+        return _mm_extract_epi64(__a, 1);
+    default:
+        abort();
+    }
+}

src/compat/aarch64.rs

@@ -0,0 +1,47 @@
+unsafe extern "C" {
+    #[link_name = "_mm_srli_epi16__extern"]
+    pub fn _mm_srli_epi16(__a: __m128i, __imm8: ::std::os::raw::c_int) -> __m128i;
+
+    #[link_name = "_mm_insert_epi8__extern"]
+    pub fn _mm_insert_epi8(
+        __a: __m128i,
+        __i: ::std::os::raw::c_int,
+        __imm8: ::std::os::raw::c_int,
+    ) -> __m128i;
+
+    #[link_name = "_mm_extract_epi8__extern"]
+    pub fn _mm_extract_epi8(__a: __m128i, __imm8: ::std::os::raw::c_int) -> ::std::os::raw::c_int;
+
+    #[link_name = "_mm_insert_epi16__extern"]
+    pub fn _mm_insert_epi16(
+        __a: __m128i,
+        __i: ::std::os::raw::c_int,
+        __imm8: ::std::os::raw::c_int,
+    ) -> __m128i;
+
+    #[link_name = "_mm_extract_epi16__extern"]
+    pub fn _mm_extract_epi16(__a: __m128i, __imm8: ::std::os::raw::c_int) -> ::std::os::raw::c_int;
+
+    #[link_name = "_mm_insert_epi32__extern"]
+    pub fn _mm_insert_epi32(
+        __a: __m128i,
+        __i: ::std::os::raw::c_int,
+        __imm8: ::std::os::raw::c_int,
+    ) -> __m128i;
+
+    #[link_name = "_mm_extract_epi32__extern"]
+    pub fn _mm_extract_epi32(__a: __m128i, __imm8: ::std::os::raw::c_int) -> ::std::os::raw::c_int;
+
+    #[link_name = "_mm_insert_epi64__extern"]
+    pub fn _mm_insert_epi64(
+        __a: __m128i,
+        __i: ::std::os::raw::c_longlong,
+        __imm8: ::std::os::raw::c_int,
+    ) -> __m128i;
+
+    #[link_name = "_mm_extract_epi64__extern"]
+    pub fn _mm_extract_epi64(
+        __a: __m128i,
+        __imm8: ::std::os::raw::c_int,
+    ) -> ::std::os::raw::c_longlong;
+}

src/device.rs

@@ -1,4 +1,17 @@
+#![allow(non_camel_case_types)]
+#![allow(improper_ctypes)]
+
+#[cfg(target_arch = "aarch64")]
+#[derive(Copy, Clone)]
+#[allow(dead_code)]
+pub struct __m128i(std::arch::aarch64::int64x2_t);
+#[cfg(target_arch = "aarch64")]
+include!(concat!(env!("OUT_DIR"), "/simd_bindings.rs"));
+#[cfg(target_arch = "x86_64")]
+use std::arch::x86_64::*;
+
 use crate::ui;
+use std::collections::HashMap;
 use std::fs;
 use std::io::Read;
 
@@ -33,32 +46,44 @@ pub mod sram;
 pub mod tlb;
 pub mod unmapped;
 pub mod vi;
+pub mod vru;
 
-pub fn run_game(file_path: &std::path::Path, device: &mut Device, fullscreen: bool) {
+pub fn run_game(
+    file_path: &std::path::Path,
+    data_dir: std::path::PathBuf,
+    device: &mut Device,
+    fullscreen: bool,
+) {
     let rom_contents = get_rom_contents(file_path);
 
     cart_rom::init(device, rom_contents); // cart needs to come before rdram
 
     // rdram pointer is shared with parallel-rdp
-    let (rdram_ptr, rdram_size) = rdram::init(device);
+    rdram::init(device);
 
     ui::audio::init(&mut device.ui, 33600);
-    ui::video::init(&mut device.ui, rdram_ptr, rdram_size, fullscreen);
+    ui::video::init(device, fullscreen);
     ui::input::init(&mut device.ui);
 
     mi::init(device);
     pif::init(device);
+    if device.ui.config.input.emulate_vru {
+        vru::init(device);
+    }
     memory::init(device);
     rsp_interface::init(device);
     rdp::init(device);
     vi::init(device);
     cpu::init(device);
 
-    ui::storage::init(&mut device.ui);
+    ui::storage::init(&mut device.ui, data_dir);
     ui::storage::load_saves(&mut device.ui);
     cart_rom::load_rom_save(device);
 
     cpu::run(device);
+
+    ui::video::close();
+    ui::storage::write_saves(&device.ui);
 }
 
 fn swap_rom(contents: Vec<u8>) -> Vec<u8> {
@@ -89,7 +114,7 @@ fn swap_rom(contents: Vec<u8>) -> Vec<u8> {
 
 fn get_rom_contents(file_path: &std::path::Path) -> Vec<u8> {
     let mut contents = vec![];
-    if file_path.extension().unwrap().to_ascii_lowercase() == "zip" {
+    if file_path.extension().unwrap().eq_ignore_ascii_case("zip") {
         let zip_file = fs::File::open(file_path).unwrap();
         let mut archive = zip::ZipArchive::new(zip_file).unwrap();
         for i in 0..archive.len() {
@@ -106,7 +131,7 @@ fn get_rom_contents(file_path: &std::path::Path) -> Vec<u8> {
                 break;
             }
         }
-    } else if file_path.extension().unwrap().to_ascii_lowercase() == "7z" {
+    } else if file_path.extension().unwrap().eq_ignore_ascii_case("7z") {
         let mut archive =
             sevenz_rust::SevenZReader::open(file_path, sevenz_rust::Password::empty()).unwrap();
 
@@ -144,20 +169,21 @@ pub struct Device {
     pif: pif::Pif,
     cart: cart_rom::Cart,
     memory: memory::Memory,
-    rsp: rsp_interface::Rsp,
-    rdp: rdp::Rdp,
+    pub rsp: rsp_interface::Rsp,
+    pub rdp: rdp::Rdp,
     pub rdram: rdram::Rdram,
     mi: mi::Mi,
     pi: pi::Pi,
-    vi: vi::Vi,
+    pub vi: vi::Vi,
     ai: ai::Ai,
     si: si::Si,
     ri: ri::Ri,
     flashram: sram::Flashram,
+    pub vru: vru::Vru,
 }
 
 impl Device {
-    pub fn new() -> Device {
+    pub fn new(config_dir: std::path::PathBuf) -> Device {
         let mut byte_swap: usize = 0;
         let test: [u8; 4] = [1, 2, 3, 4];
         // if the host computer is little endian, that means the RDRAM will be stored as little endian
@@ -166,7 +192,7 @@ impl Device {
             byte_swap = 3;
         }
         Device {
-            ui: ui::Ui::new(),
+            ui: ui::Ui::new(config_dir),
             byte_swap,
             cpu: cpu::Cpu {
                 cop0: cop0::Cop0 {
@@ -214,7 +240,8 @@ impl Device {
                 cop1: cop1::Cop1 {
                     fcr0: 0,
                     fcr31: 0,
-                    flush_mode: 0,
+                    //#[cfg(target_arch = "x86_64")]
+                    //flush_mode: 0,
                     fgr32: [[0; 4]; 32],
                     fgr64: [[0; 8]; 32],
                     instrs: [cop1::reserved; 32],
@@ -239,7 +266,7 @@ impl Device {
                 llbit: false,
                 lo: 0,
                 hi: 0,
-                running: 0,
+                running: false,
                 instrs: [cop0::reserved; 64],
                 special_instrs: [cop0::reserved; 64],
                 regimm_instrs: [cop0::reserved; 32],
@@ -247,7 +274,7 @@ impl Device {
                     enabled: false,
                     count: u64::MAX,
                     handler: events::dummy_event,
-                }; events::EventType::EventTypeCount as usize],
+                }; events::EventType::Count as usize],
                 next_event_count: u64::MAX,
                 next_event: 0,
             },
@@ -292,7 +319,10 @@ impl Device {
                     words: [0; 4],
                 }; 512],
             },
-            rdram: rdram::Rdram { mem: vec![] },
+            rdram: rdram::Rdram {
+                mem: vec![],
+                size: 0x800000,
+            },
             rsp: rsp_interface::Rsp {
                 cpu: rsp_cpu::Cpu {
                     instructions: [rsp_cpu::Instructions {
@@ -311,21 +341,21 @@ impl Device {
                     sync_point: false,
                     cycle_counter: 0,
                     gpr: [0; 32],
-                    vpr: [0; 32],
+                    vpr: unsafe { [_mm_setzero_si128(); 32] },
                     reciprocals: [0; 512],
                     inverse_square_roots: [0; 512],
                     divdp: false,
                     divin: 0,
                     divout: 0,
-                    shuffle: unsafe { [std::arch::x86_64::_mm_setzero_si128(); 16] },
-                    vcol: unsafe { std::arch::x86_64::_mm_setzero_si128() },
-                    vcoh: unsafe { std::arch::x86_64::_mm_setzero_si128() },
-                    vccl: unsafe { std::arch::x86_64::_mm_setzero_si128() },
-                    vcch: unsafe { std::arch::x86_64::_mm_setzero_si128() },
-                    vce: unsafe { std::arch::x86_64::_mm_setzero_si128() },
-                    accl: unsafe { std::arch::x86_64::_mm_setzero_si128() },
-                    accm: unsafe { std::arch::x86_64::_mm_setzero_si128() },
-                    acch: unsafe { std::arch::x86_64::_mm_setzero_si128() },
+                    shuffle: unsafe { [_mm_setzero_si128(); 16] },
+                    vcol: unsafe { _mm_setzero_si128() },
+                    vcoh: unsafe { _mm_setzero_si128() },
+                    vccl: unsafe { _mm_setzero_si128() },
+                    vcch: unsafe { _mm_setzero_si128() },
+                    vce: unsafe { _mm_setzero_si128() },
+                    accl: unsafe { _mm_setzero_si128() },
+                    accm: unsafe { _mm_setzero_si128() },
+                    acch: unsafe { _mm_setzero_si128() },
                     special_instrs: [rsp_su_instructions::reserved; 64],
                     regimm_instrs: [rsp_su_instructions::reserved; 32],
                     cop0_instrs: [rsp_su_instructions::reserved; 32],
@@ -380,6 +410,7 @@ impl Device {
                 delay: 0,
                 count_per_scanline: 0,
                 limiter: None,
+                vi_counter: 0,
             },
             flashram: sram::Flashram {
                 status: 0,
@@ -388,6 +419,19 @@ impl Device {
                 silicon_id: [sram::FLASHRAM_TYPE_ID, sram::MX29L1100_ID],
                 mode: sram::FlashramMode::ReadArray,
             },
+            vru: vru::Vru {
+                status: 0,
+                voice_state: 0,
+                load_offset: 0,
+                voice_init: 0,
+                word_buffer: [0; 40],
+                words: Vec::new(),
+                talking: false,
+                word_mappings: HashMap::new(),
+                window_notifier: None,
+                word_receiver: None,
+                gui_ctx: None,
+            },
         }
     }
 }

src/device/ai.rs

@@ -1,16 +1,16 @@
 use crate::device;
 use crate::ui;
 
-pub const AI_DRAM_ADDR_REG: u32 = 0;
-pub const AI_LEN_REG: u32 = 1;
-//pub const AI_CONTROL_REG: u32 = 2;
-pub const AI_STATUS_REG: u32 = 3;
-pub const AI_DACRATE_REG: u32 = 4;
-//pub const AI_BITRATE_REG: u32 = 5;
+const AI_DRAM_ADDR_REG: u32 = 0;
+const AI_LEN_REG: u32 = 1;
+// const AI_CONTROL_REG: u32 = 2;
+const AI_STATUS_REG: u32 = 3;
+const AI_DACRATE_REG: u32 = 4;
+// const AI_BITRATE_REG: u32 = 5;
 pub const AI_REGS_COUNT: u32 = 6;
 
-pub const AI_STATUS_BUSY: u32 = 0x40000000;
-pub const AI_STATUS_FULL: u32 = 0x80000000;
+const AI_STATUS_BUSY: u32 = 0x40000000;
+const AI_STATUS_FULL: u32 = 0x80000000;
 
 pub struct Ai {
     pub regs: [u32; AI_REGS_COUNT as usize],

src/device/cart.rs

@@ -4,19 +4,19 @@ use chrono::Timelike;
 use crate::device;
 use crate::ui;
 
-pub const JCMD_STATUS: u8 = 0x00;
-pub const JCMD_EEPROM_READ: u8 = 0x04;
-pub const JCMD_EEPROM_WRITE: u8 = 0x05;
-pub const JCMD_AF_RTC_STATUS: u8 = 0x06;
-pub const JCMD_AF_RTC_READ: u8 = 0x07;
-pub const JCMD_AF_RTC_WRITE: u8 = 0x08;
-pub const JCMD_RESET: u8 = 0xff;
+const JCMD_STATUS: u8 = 0x00;
+const JCMD_EEPROM_READ: u8 = 0x04;
+const JCMD_EEPROM_WRITE: u8 = 0x05;
+const JCMD_AF_RTC_STATUS: u8 = 0x06;
+const JCMD_AF_RTC_READ: u8 = 0x07;
+const JCMD_AF_RTC_WRITE: u8 = 0x08;
+const JCMD_RESET: u8 = 0xff;
 
-pub const JDT_AF_RTC: u16 = 0x1000; /* RTC */
-pub const JDT_EEPROM_4K: u16 = 0x8000; /* 4k EEPROM */
-pub const JDT_EEPROM_16K: u16 = 0xc000; /* 16k EEPROM */
-pub const EEPROM_BLOCK_SIZE: usize = 8;
-pub const EEPROM_MAX_SIZE: usize = 0x800;
+const JDT_AF_RTC: u16 = 0x1000; /* RTC */
+const JDT_EEPROM_4K: u16 = 0x8000; /* 4k EEPROM */
+const JDT_EEPROM_16K: u16 = 0xc000; /* 16k EEPROM */
+const EEPROM_BLOCK_SIZE: usize = 8;
+const EEPROM_MAX_SIZE: usize = 0x800;
 
 pub struct AfRtc {
     pub control: u16,
@@ -159,7 +159,7 @@ pub fn af_rtc_write_block(device: &mut device::Device, block: usize, offset: usi
     match device.pif.ram[block] {
         0 => {
             device.cart.rtc.control =
-                (device.pif.ram[offset + 1] as u16) << 8 | device.pif.ram[offset] as u16;
+                ((device.pif.ram[offset + 1] as u16) << 8) | device.pif.ram[offset] as u16;
             device.pif.ram[status] = 0x00;
         }
         1 => {

src/device/cart_rom.rs

@@ -10,7 +10,7 @@ pub enum CicType {
     CicNus5167,
 }
 
-pub const CART_MASK: usize = 0xFFFFFFF;
+const CART_MASK: usize = 0xFFFFFFF;
 pub struct Cart {
     pub rom: Vec<u8>,
     pub is_viewer_buffer: [u8; 0xFFFF],
@@ -80,14 +80,16 @@ pub fn dma_read(
     cart_addr &= CART_MASK as u32;
 
     for i in 0..length {
-        device.cart.rom[(cart_addr + i) as usize] =
-            device.rdram.mem[(dram_addr + i) as usize ^ device.byte_swap];
+        if cart_addr + i < device.cart.rom.len() as u32 {
+            device.cart.rom[(cart_addr + i) as usize] =
+                device.rdram.mem[(dram_addr + i) as usize ^ device.byte_swap];
 
-        device.ui.saves.romsave.0.insert(
-            (cart_addr + i).to_string(),
-            serde_json::to_value(device.rdram.mem[(dram_addr + i) as usize ^ device.byte_swap])
-                .unwrap(),
-        );
+            device.ui.saves.romsave.0.insert(
+                (cart_addr + i).to_string(),
+                serde_json::to_value(device.rdram.mem[(dram_addr + i) as usize ^ device.byte_swap])
+                    .unwrap(),
+            );
+        }
     }
 
     device.ui.saves.romsave.1 = true;
@@ -127,6 +129,9 @@ pub fn init(device: &mut device::Device, rom_file: Vec<u8>) {
     device.ui.game_hash = calculate_hash(&device.cart.rom);
 
     device.ui.game_id = String::from_utf8(device.cart.rom[0x3B..0x3E].to_vec()).unwrap();
+    if device.ui.game_id.contains('\0') {
+        device.ui.game_id = String::from("UNK");
+    }
 }
 
 pub fn load_rom_save(device: &mut device::Device) {

src/device/controller.rs

@@ -3,24 +3,18 @@ use crate::ui;
 
 pub const JCMD_STATUS: u8 = 0x00;
 pub const JCMD_CONTROLLER_READ: u8 = 0x01;
-pub const JCMD_PAK_READ: u8 = 0x02;
-pub const JCMD_PAK_WRITE: u8 = 0x03;
-//pub const JCMD_VRU_READ: u8 = 0x09;
-//pub const JCMD_VRU_WRITE: u8 = 0x0A;
-//pub const JCMD_VRU_READ_STATUS: u8 = 0x0B;
-//pub const JCMD_VRU_WRITE_CONFIG: u8 = 0x0C;
-//pub const JCMD_VRU_WRITE_INIT: u8 = 0x0D;
+const JCMD_PAK_READ: u8 = 0x02;
+const JCMD_PAK_WRITE: u8 = 0x03;
 pub const JCMD_RESET: u8 = 0xff;
 
-//pub const JDT_NONE: u16 = 0x0000;
-pub const JDT_JOY_ABS_COUNTERS: u16 = 0x0001; /* joystick with absolute coordinates */
-//pub const JDT_JOY_REL_COUNTERS: u16 = 0x0002; /* joystick with relative coordinates (= mouse) */
-pub const JDT_JOY_PORT: u16 = 0x0004; /* has port for external paks */
-//pub const JDT_VRU: u16 = 0x0100; /* VRU */
-pub const PAK_CHUNK_SIZE: usize = 0x20;
-pub const CONT_STATUS_PAK_NOT_PRESENT: u8 = 0;
-pub const CONT_STATUS_PAK_PRESENT: u8 = 1;
-pub const CONT_FLAVOR: u16 = JDT_JOY_ABS_COUNTERS | JDT_JOY_PORT;
+//const JDT_NONE: u16 = 0x0000;
+const JDT_JOY_ABS_COUNTERS: u16 = 0x0001; /* joystick with absolute coordinates */
+//const JDT_JOY_REL_COUNTERS: u16 = 0x0002; /* joystick with relative coordinates (= mouse) */
+const JDT_JOY_PORT: u16 = 0x0004; /* has port for external paks */
+const PAK_CHUNK_SIZE: usize = 0x20;
+const CONT_STATUS_PAK_NOT_PRESENT: u8 = 0;
+const CONT_STATUS_PAK_PRESENT: u8 = 1;
+const CONT_FLAVOR: u16 = JDT_JOY_ABS_COUNTERS | JDT_JOY_PORT;
 
 #[derive(Copy, Clone)]
 pub struct PakHandler {
@@ -75,14 +69,14 @@ pub fn pak_read_block(
     channel: usize,
 ) {
     let address =
-        (device.pif.ram[addr_acrc] as u16) << 8 | (device.pif.ram[addr_acrc + 1] & 0xe0) as u16;
+        ((device.pif.ram[addr_acrc] as u16) << 8) | (device.pif.ram[addr_acrc + 1] & 0xe0) as u16;
     let handler = device.pif.channels[channel].pak_handler;
 
     if handler.is_some() {
         (handler.unwrap().read)(device, channel, address, data, PAK_CHUNK_SIZE);
-        device.pif.ram[dcrc] = pak_data_crc(device, data, PAK_CHUNK_SIZE)
+        device.pif.ram[dcrc] = data_crc(device, data, PAK_CHUNK_SIZE)
     } else {
-        device.pif.ram[dcrc] = !pak_data_crc(device, data, PAK_CHUNK_SIZE)
+        device.pif.ram[dcrc] = !data_crc(device, data, PAK_CHUNK_SIZE)
     }
 }
 
@@ -94,18 +88,18 @@ pub fn pak_write_block(
     channel: usize,
 ) {
     let address =
-        (device.pif.ram[addr_acrc] as u16) << 8 | (device.pif.ram[addr_acrc + 1] & 0xe0) as u16;
+        ((device.pif.ram[addr_acrc] as u16) << 8) | (device.pif.ram[addr_acrc + 1] & 0xe0) as u16;
     let handler = device.pif.channels[channel].pak_handler;
 
     if handler.is_some() {
         (handler.unwrap().write)(device, channel, address, data, PAK_CHUNK_SIZE);
-        device.pif.ram[dcrc] = pak_data_crc(device, data, PAK_CHUNK_SIZE)
+        device.pif.ram[dcrc] = data_crc(device, data, PAK_CHUNK_SIZE)
     } else {
-        device.pif.ram[dcrc] = !pak_data_crc(device, data, PAK_CHUNK_SIZE)
+        device.pif.ram[dcrc] = !data_crc(device, data, PAK_CHUNK_SIZE)
     }
 }
 
-pub fn pak_data_crc(device: &device::Device, data_offset: usize, size: usize) -> u8 {
+pub fn data_crc(device: &device::Device, data_offset: usize, size: usize) -> u8 {
     let mut i = 0;
     let mut crc = 0;
 

src/device/cop0.rs

@@ -1,37 +1,37 @@
 use crate::device;
 
 pub const COP0_INDEX_REG: u32 = 0;
-pub const COP0_RANDOM_REG: u32 = 1;
+const COP0_RANDOM_REG: u32 = 1;
 pub const COP0_ENTRYLO0_REG: u32 = 2;
 pub const COP0_ENTRYLO1_REG: u32 = 3;
 pub const COP0_CONTEXT_REG: u32 = 4;
 pub const COP0_PAGEMASK_REG: u32 = 5;
-pub const COP0_WIRED_REG: u32 = 6;
-//pub const COP0_UNUSED_7: u32 = 7;
+const COP0_WIRED_REG: u32 = 6;
+//const COP0_UNUSED_7: u32 = 7;
 pub const COP0_BADVADDR_REG: u32 = 8;
 pub const COP0_COUNT_REG: u32 = 9;
 pub const COP0_ENTRYHI_REG: u32 = 10;
-pub const COP0_COMPARE_REG: u32 = 11;
+const COP0_COMPARE_REG: u32 = 11;
 pub const COP0_STATUS_REG: u32 = 12;
 pub const COP0_CAUSE_REG: u32 = 13;
 pub const COP0_EPC_REG: u32 = 14;
-pub const COP0_PREVID_REG: u32 = 15;
-pub const COP0_CONFIG_REG: u32 = 16;
+const COP0_PREVID_REG: u32 = 15;
+const COP0_CONFIG_REG: u32 = 16;
 pub const COP0_LLADDR_REG: u32 = 17;
-//pub const COP0_WATCHLO_REG: u32 = 18;
-//pub const COP0_WATCHHI_REG: u32 = 19;
+//const COP0_WATCHLO_REG: u32 = 18;
+//const COP0_WATCHHI_REG: u32 = 19;
 pub const COP0_XCONTEXT_REG: u32 = 20;
-//pub const COP0_UNUSED_21: u32 = 21;
-//pub const COP0_UNUSED_22: u32 = 22;
-//pub const COP0_UNUSED_23: u32 = 23;
-//pub const COP0_UNUSED_24: u32 = 24;
-//pub const COP0_UNUSED_25: u32 = 25;
-//pub const COP0_PARITYERR_REG: u32 = 26;
-//pub const COP0_CACHEERR_REG: u32 = 27;
+//const COP0_UNUSED_21: u32 = 21;
+//const COP0_UNUSED_22: u32 = 22;
+//const COP0_UNUSED_23: u32 = 23;
+//const COP0_UNUSED_24: u32 = 24;
+//const COP0_UNUSED_25: u32 = 25;
+//const COP0_PARITYERR_REG: u32 = 26;
+//const COP0_CACHEERR_REG: u32 = 27;
 pub const COP0_TAGLO_REG: u32 = 28;
-//pub const COP0_TAGHI_REG: u32 = 29;
-pub const COP0_ERROREPC_REG: u32 = 30;
-//pub const COP0_UNUSED_31: u32 = 31;
+//const COP0_TAGHI_REG: u32 = 29;
+const COP0_ERROREPC_REG: u32 = 30;
+//const COP0_UNUSED_31: u32 = 31;
 pub const COP0_REGS_COUNT: u32 = 32;
 
 pub const COP0_STATUS_IE: u64 = 1 << 0;
@@ -42,12 +42,12 @@ pub const COP0_STATUS_FR: u64 = 1 << 26;
 pub const COP0_STATUS_CU1: u64 = 1 << 29;
 pub const COP0_STATUS_CU2: u64 = 1 << 30;
 
-//pub const COP0_CAUSE_EXCCODE_INTR: u64 = 0 << 2;
+//const COP0_CAUSE_EXCCODE_INTR: u64 = 0 << 2;
 pub const COP0_CAUSE_EXCCODE_MOD: u64 = 1 << 2;
 pub const COP0_CAUSE_EXCCODE_TLBL: u64 = 2 << 2;
 pub const COP0_CAUSE_EXCCODE_TLBS: u64 = 3 << 2;
-//pub const COP0_CAUSE_EXCCODE_ADEL: u64 = 4 << 2;
-//pub const COP0_CAUSE_EXCCODE_ADES: u64 = 5 << 2;
+//const COP0_CAUSE_EXCCODE_ADEL: u64 = 4 << 2;
+//const COP0_CAUSE_EXCCODE_ADES: u64 = 5 << 2;
 pub const COP0_CAUSE_EXCCODE_SYS: u64 = 8 << 2;
 pub const COP0_CAUSE_EXCCODE_BP: u64 = 9 << 2;
 pub const COP0_CAUSE_EXCCODE_RI: u64 = 10 << 2;
@@ -55,7 +55,7 @@ pub const COP0_CAUSE_EXCCODE_CPU: u64 = 11 << 2;
 pub const COP0_CAUSE_EXCCODE_TR: u64 = 13 << 2;
 pub const COP0_CAUSE_EXCCODE_FPE: u64 = 15 << 2;
 pub const COP0_CAUSE_IP2: u64 = 1 << 10;
-pub const COP0_CAUSE_IP7: u64 = 1 << 15;
+const COP0_CAUSE_IP7: u64 = 1 << 15;
 pub const COP0_CAUSE_BD: u64 = 1 << 31;
 
 pub const COP0_CAUSE_CE1: u64 = 1 << 28;

src/device/cop1.rs

@@ -1,32 +1,34 @@
 use crate::device;
 
-//pub const FCR31_FLAG_INEXACT_BIT: u32 = 1 << 2;
-//pub const FCR31_FLAG_UNDERFLOW_BIT: u32 = 1 << 3;
-//pub const FCR31_FLAG_OVERFLOW_BIT: u32 = 1 << 4;
-//pub const FCR31_FLAG_DIVBYZERO_BIT: u32 = 1 << 5;
-//pub const FCR31_FLAG_INVALID_BIT: u32 = 1 << 6;
-//pub const FCR31_ENABLE_INEXACT_BIT: u32 = 1 << 7;
-//pub const FCR31_ENABLE_UNDERFLOW_BIT: u32 = 1 << 8;
-//pub const FCR31_ENABLE_OVERFLOW_BIT: u32 = 1 << 9;
-//pub const FCR31_ENABLE_DIVBYZERO_BIT: u32 = 1 << 10;
-//pub const FCR31_ENABLE_INVALID_BIT: u32 = 1 << 11;
-//pub const FCR31_CAUSE_INEXACT_BIT: u32 = 1 << 12;
-//pub const FCR31_CAUSE_UNDERFLOW_BIT: u32 = 1 << 13;
-//pub const FCR31_CAUSE_OVERFLOW_BIT: u32 = 1 << 14;
-//pub const FCR31_CAUSE_DIVBYZERO_BIT: u32 = 1 << 15;
-//pub const FCR31_CAUSE_INVALID_BIT: u32 = 1 << 16;
-pub const FCR31_CAUSE_UNIMP_BIT: u32 = 1 << 17;
+//const FCR31_FLAG_INEXACT_BIT: u32 = 1 << 2;
+//const FCR31_FLAG_UNDERFLOW_BIT: u32 = 1 << 3;
+//const FCR31_FLAG_OVERFLOW_BIT: u32 = 1 << 4;
+//const FCR31_FLAG_DIVBYZERO_BIT: u32 = 1 << 5;
+//const FCR31_FLAG_INVALID_BIT: u32 = 1 << 6;
+//const FCR31_ENABLE_INEXACT_BIT: u32 = 1 << 7;
+//const FCR31_ENABLE_UNDERFLOW_BIT: u32 = 1 << 8;
+//const FCR31_ENABLE_OVERFLOW_BIT: u32 = 1 << 9;
+//const FCR31_ENABLE_DIVBYZERO_BIT: u32 = 1 << 10;
+//const FCR31_ENABLE_INVALID_BIT: u32 = 1 << 11;
+//const FCR31_CAUSE_INEXACT_BIT: u32 = 1 << 12;
+//const FCR31_CAUSE_UNDERFLOW_BIT: u32 = 1 << 13;
+//const FCR31_CAUSE_OVERFLOW_BIT: u32 = 1 << 14;
+//const FCR31_CAUSE_DIVBYZERO_BIT: u32 = 1 << 15;
+//const FCR31_CAUSE_INVALID_BIT: u32 = 1 << 16;
+const FCR31_CAUSE_UNIMP_BIT: u32 = 1 << 17;
 pub const FCR31_CMP_BIT: u32 = 1 << 23;
-pub const FCR31_FS_BIT: u32 = 1 << 24;
+//#[cfg(target_arch = "x86_64")]
+//const FCR31_FS_BIT: u32 = 1 << 24;
 
-pub const FCR31_CAUSE_MASK: u32 = 0b00000000000000111111000000000000;
-pub const FCR31_ENABLE_MASK: u32 = 0b00000000000000000000111110000000;
-pub const FCR31_WRITE_MASK: u32 = 0b00000001100000111111111111111111;
+const FCR31_CAUSE_MASK: u32 = 0b00000000000000111111000000000000;
+const FCR31_ENABLE_MASK: u32 = 0b00000000000000000000111110000000;
+const FCR31_WRITE_MASK: u32 = 0b00000001100000111111111111111111;
 
 pub struct Cop1 {
     pub fcr0: u32,
     pub fcr31: u32,
-    pub flush_mode: u32,
+    //#[cfg(target_arch = "x86_64")]
+    //pub flush_mode: u32,
     pub fgr32: [[u8; 4]; 32],
     pub fgr64: [[u8; 8]; 32],
     pub instrs: [fn(&mut device::Device, u32); 32],
@@ -359,15 +361,16 @@ pub fn set_control_registers_fpu(device: &mut device::Device, index: u32, data:
             device.cpu.cop1.fcr31 = data & FCR31_WRITE_MASK;
             // the Cause bits are ANDed with the Enable bits to check for exceptions
             // "Unimplemented Operation" has no Enable bit and always causes an exception
-            if (device.cpu.cop1.fcr31 & FCR31_CAUSE_MASK) >> 5
+            if ((device.cpu.cop1.fcr31 & FCR31_CAUSE_MASK) >> 5)
                 & (device.cpu.cop1.fcr31 & FCR31_ENABLE_MASK)
                 != 0
                 || device.cpu.cop1.fcr31 & FCR31_CAUSE_UNIMP_BIT != 0
             {
                 device::exceptions::floating_point_exception(device)
             }
-
-            unsafe {
+            /*
+            #[cfg(target_arch = "x86_64")]
+            {
                 let flush_mode;
                 if (device.cpu.cop1.fcr31 & 2) != 0 {
                     if (device.cpu.cop1.fcr31 & FCR31_FS_BIT) != 0 {
@@ -380,10 +383,13 @@ pub fn set_control_registers_fpu(device: &mut device::Device, index: u32, data:
                 }
                 if flush_mode != device.cpu.cop1.flush_mode {
                     #[allow(deprecated)]
-                    std::arch::x86_64::_MM_SET_FLUSH_ZERO_MODE(flush_mode);
+                    unsafe {
+                        std::arch::x86_64::_MM_SET_FLUSH_ZERO_MODE(flush_mode)
+                    };
                     device.cpu.cop1.flush_mode = flush_mode;
                 }
             }
+            */
         }
         _ => {
             panic!("unknown FCR register")

src/device/cpu.rs

@@ -25,13 +25,13 @@ pub struct Cpu {
     pub pc_phys: u64,
     pub lo: u64,
     pub hi: u64,
-    pub running: u8,
+    pub running: bool,
     pub llbit: bool,
     pub clock_rate: u64,
     pub instrs: [fn(&mut device::Device, u32); 64],
     pub special_instrs: [fn(&mut device::Device, u32); 64],
     pub regimm_instrs: [fn(&mut device::Device, u32); 32],
-    pub events: [device::events::Event; device::events::EventType::EventTypeCount as usize],
+    pub events: [device::events::Event; device::events::EventType::Count as usize],
     pub next_event_count: u64,
     pub next_event: usize,
 }
@@ -231,8 +231,8 @@ pub fn in_delay_slot_taken(device: &device::Device) -> bool {
 }
 
 pub fn run(device: &mut device::Device) {
-    device.cpu.running = 1;
-    while device.cpu.running == 1 {
+    device.cpu.running = true;
+    while device.cpu.running {
         device.cpu.gpr[0] = 0; // gpr 0 is read only
         let (cached, err);
         (device.cpu.pc_phys, cached, err) = device::memory::translate_address(

src/device/cpu_instructions.rs

@@ -346,12 +346,12 @@ pub fn lwl(device: &mut device::Device, opcode: u32) {
 
     device.cpu.gpr[rt(opcode) as usize] = se32(
         ((device.cpu.gpr[rt(opcode) as usize] as u32) & mask
-            | device::memory::data_read(
+            | (device::memory::data_read(
                 device,
                 phys_address,
                 device::memory::AccessSize::Word,
                 cached,
-            ) << shift) as i32,
+            ) << shift)) as i32,
     )
 }
 
@@ -431,12 +431,12 @@ pub fn lwr(device: &mut device::Device, opcode: u32) {
 
     device.cpu.gpr[rt(opcode) as usize] = se32(
         ((device.cpu.gpr[rt(opcode) as usize] as u32) & (mask as u32)
-            | device::memory::data_read(
+            | (device::memory::data_read(
                 device,
                 phys_address,
                 device::memory::AccessSize::Word,
                 cached,
-            ) >> shift) as i32,
+            ) >> shift)) as i32,
     )
 }
 

src/device/events.rs

@@ -11,7 +11,8 @@ pub enum EventType {
     InterruptCheck,
     SPDma,
     Compare,
-    EventTypeCount,
+    Vru,
+    Count,
 }
 
 #[derive(PartialEq, Copy, Clone)]

src/device/is_viewer.rs

@@ -1,6 +1,6 @@
 use crate::device;
 
-pub const IS_VIEWER_MASK: usize = 0xFFFF;
+const IS_VIEWER_MASK: usize = 0xFFFF;
 
 pub fn read_mem(
     device: &mut device::Device,

src/device/memory.rs

@@ -1,24 +1,24 @@
 use crate::device;
 
-pub const MM_RDRAM_DRAM: usize = 0x00000000;
-pub const MM_RDRAM_REGS: usize = 0x03f00000;
-pub const MM_RSP_MEM: usize = 0x04000000;
-pub const MM_RSP_REGS: usize = 0x04040000;
-pub const MM_RSP_REGS_PC: usize = 0x04080000;
-pub const MM_DPC_REGS: usize = 0x04100000;
-pub const MM_DPS_REGS: usize = 0x04200000;
-pub const MM_MI_REGS: usize = 0x04300000;
-pub const MM_VI_REGS: usize = 0x04400000;
-pub const MM_AI_REGS: usize = 0x04500000;
-pub const MM_PI_REGS: usize = 0x04600000;
-pub const MM_RI_REGS: usize = 0x04700000;
-pub const MM_SI_REGS: usize = 0x04800000;
+const MM_RDRAM_DRAM: usize = 0x00000000;
+const MM_RDRAM_REGS: usize = 0x03f00000;
+const MM_RSP_MEM: usize = 0x04000000;
+const MM_RSP_REGS: usize = 0x04040000;
+const MM_RSP_REGS_PC: usize = 0x04080000;
+const MM_DPC_REGS: usize = 0x04100000;
+const MM_DPS_REGS: usize = 0x04200000;
+const MM_MI_REGS: usize = 0x04300000;
+const MM_VI_REGS: usize = 0x04400000;
+const MM_AI_REGS: usize = 0x04500000;
+const MM_PI_REGS: usize = 0x04600000;
+const MM_RI_REGS: usize = 0x04700000;
+const MM_SI_REGS: usize = 0x04800000;
 //const MM_DOM2_ADDR1: usize = 0x05000000;
 pub const MM_DOM2_ADDR2: usize = 0x08000000;
 pub const MM_CART_ROM: usize = 0x10000000;
-pub const MM_PIF_MEM: usize = 0x1fc00000;
-//pub const MM_DOM1_ADDR3: usize = 0x1fd00000;
-pub const MM_IS_VIEWER: usize = 0x13ff0000;
+const MM_PIF_MEM: usize = 0x1fc00000;
+//const MM_DOM1_ADDR3: usize = 0x1fd00000;
+const MM_IS_VIEWER: usize = 0x13ff0000;
 
 #[derive(PartialEq)]
 pub enum AccessType {

src/device/mempak.rs

@@ -1,7 +1,7 @@
 use crate::device;
 
-pub const MEMPAK_SIZE: usize = 0x8000;
-pub const MPK_PAGE_SIZE: usize = 256;
+const MEMPAK_SIZE: usize = 0x8000;
+const MPK_PAGE_SIZE: usize = 256;
 
 pub fn format_mempak(device: &mut device::Device) {
     if device.ui.saves.mempak.0.len() < MEMPAK_SIZE * 4 {

src/device/mi.rs

@@ -1,9 +1,9 @@
 use crate::device;
 
-pub const MI_INIT_MODE_REG: u32 = 0;
-pub const MI_VERSION_REG: u32 = 1;
-pub const MI_INTR_REG: u32 = 2;
-pub const MI_INTR_MASK_REG: u32 = 3;
+const MI_INIT_MODE_REG: u32 = 0;
+const MI_VERSION_REG: u32 = 1;
+const MI_INTR_REG: u32 = 2;
+const MI_INTR_MASK_REG: u32 = 3;
 pub const MI_REGS_COUNT: u32 = 4;
 
 /* read */
@@ -15,34 +15,34 @@ pub const MI_INTR_PI: u32 = 1 << 4;
 pub const MI_INTR_DP: u32 = 1 << 5;
 
 /* write */
-pub const MI_CLR_SP: u32 = 1 << 0;
-pub const MI_SET_SP: u32 = 1 << 1;
-pub const MI_CLR_SI: u32 = 1 << 2;
-pub const MI_SET_SI: u32 = 1 << 3;
-pub const MI_CLR_AI: u32 = 1 << 4;
-pub const MI_SET_AI: u32 = 1 << 5;
-pub const MI_CLR_VI: u32 = 1 << 6;
-pub const MI_SET_VI: u32 = 1 << 7;
-pub const MI_CLR_PI: u32 = 1 << 8;
-pub const MI_SET_PI: u32 = 1 << 9;
-pub const MI_CLR_DP: u32 = 1 << 10;
-pub const MI_SET_DP: u32 = 1 << 11;
+const MI_CLR_SP: u32 = 1 << 0;
+const MI_SET_SP: u32 = 1 << 1;
+const MI_CLR_SI: u32 = 1 << 2;
+const MI_SET_SI: u32 = 1 << 3;
+const MI_CLR_AI: u32 = 1 << 4;
+const MI_SET_AI: u32 = 1 << 5;
+const MI_CLR_VI: u32 = 1 << 6;
+const MI_SET_VI: u32 = 1 << 7;
+const MI_CLR_PI: u32 = 1 << 8;
+const MI_SET_PI: u32 = 1 << 9;
+const MI_CLR_DP: u32 = 1 << 10;
+const MI_SET_DP: u32 = 1 << 11;
 
 /* mode read */
-pub const MI_INIT_MODE: u32 = 1 << 7;
-pub const MI_EBUS_MODE: u32 = 1 << 8;
-pub const MI_RDRAM_MODE: u32 = 1 << 9;
+const MI_INIT_MODE: u32 = 1 << 7;
+const MI_EBUS_MODE: u32 = 1 << 8;
+const MI_RDRAM_MODE: u32 = 1 << 9;
 
 /* mode write */
-pub const MI_CLR_INIT: u32 = 1 << 7;
-pub const MI_SET_INIT: u32 = 1 << 8;
-pub const MI_CLR_EBUS: u32 = 1 << 9;
-pub const MI_SET_EBUS: u32 = 1 << 10;
-pub const MI_CLR_DP_INTR: u32 = 1 << 11;
-pub const MI_CLR_RDRAM: u32 = 1 << 12;
-pub const MI_SET_RDRAM: u32 = 1 << 13;
+const MI_CLR_INIT: u32 = 1 << 7;
+const MI_SET_INIT: u32 = 1 << 8;
+const MI_CLR_EBUS: u32 = 1 << 9;
+const MI_SET_EBUS: u32 = 1 << 10;
+const MI_CLR_DP_INTR: u32 = 1 << 11;
+const MI_CLR_RDRAM: u32 = 1 << 12;
+const MI_SET_RDRAM: u32 = 1 << 13;
 
-pub const MI_INIT_LENGTH_MASK: u32 = 0b1111111;
+const MI_INIT_LENGTH_MASK: u32 = 0b1111111;
 
 pub struct Mi {
     pub regs: [u32; MI_REGS_COUNT as usize],

src/device/pi.rs

@@ -1,29 +1,30 @@
 use crate::device;
 
-pub const PI_DRAM_ADDR_REG: u32 = 0;
-pub const PI_CART_ADDR_REG: u32 = 1;
-pub const PI_RD_LEN_REG: u32 = 2;
-pub const PI_WR_LEN_REG: u32 = 3;
+const PI_DRAM_ADDR_REG: u32 = 0;
+const PI_CART_ADDR_REG: u32 = 1;
+const PI_RD_LEN_REG: u32 = 2;
+const PI_WR_LEN_REG: u32 = 3;
 pub const PI_STATUS_REG: u32 = 4;
-pub const PI_BSD_DOM1_LAT_REG: u32 = 5;
-pub const PI_BSD_DOM1_PWD_REG: u32 = 6;
-pub const PI_BSD_DOM1_PGS_REG: u32 = 7;
-pub const PI_BSD_DOM1_RLS_REG: u32 = 8;
-pub const PI_BSD_DOM2_LAT_REG: u32 = 9;
-pub const PI_BSD_DOM2_PWD_REG: u32 = 10;
-pub const PI_BSD_DOM2_PGS_REG: u32 = 11;
-pub const PI_BSD_DOM2_RLS_REG: u32 = 12;
-pub const PI_REGS_COUNT: u32 = 13;
+const PI_BSD_DOM1_LAT_REG: u32 = 5;
+const PI_BSD_DOM1_PWD_REG: u32 = 6;
+const PI_BSD_DOM1_PGS_REG: u32 = 7;
+const PI_BSD_DOM1_RLS_REG: u32 = 8;
+const PI_BSD_DOM2_LAT_REG: u32 = 9;
+const PI_BSD_DOM2_PWD_REG: u32 = 10;
+const PI_BSD_DOM2_PGS_REG: u32 = 11;
+const PI_BSD_DOM2_RLS_REG: u32 = 12;
+//const UNKNOWN_REG: u32 = 13; //LibDragon
+pub const PI_REGS_COUNT: u32 = 14;
 
 /* PI_STATUS - read */
-pub const PI_STATUS_DMA_BUSY: u32 = 1 << 0;
+const PI_STATUS_DMA_BUSY: u32 = 1 << 0;
 pub const PI_STATUS_IO_BUSY: u32 = 1 << 1;
-//pub const PI_STATUS_ERROR: u32 = 1 << 2;
-pub const PI_STATUS_INTERRUPT: u32 = 1 << 3;
+//const PI_STATUS_ERROR: u32 = 1 << 2;
+const PI_STATUS_INTERRUPT: u32 = 1 << 3;
 
 /* PI_STATUS - write */
-pub const PI_STATUS_RESET: u32 = 1 << 0;
-pub const PI_STATUS_CLR_INTR: u32 = 1 << 1;
+const PI_STATUS_RESET: u32 = 1 << 0;
+const PI_STATUS_CLR_INTR: u32 = 1 << 1;
 
 pub struct Pi {
     pub regs: [u32; PI_REGS_COUNT as usize],

src/device/pif.rs

@@ -99,7 +99,7 @@ pub fn update_pif_ram(device: &mut device::Device) -> u64 {
     for k in 0..PIF_CHANNELS_COUNT {
         active_channels += process_channel(device, k)
     }
-    (24000 + (active_channels * 28000)) as u64
+    (24000 + (active_channels * 30000)) as u64
 }
 
 pub fn disable_pif_channel(channel: &mut PifChannel) {
@@ -229,8 +229,16 @@ pub fn init(device: &mut device::Device) {
         write: device::mempak::write,
     };
 
-    device.pif.channels[0].pak_handler = Some(mempak_handler);
-    device.pif.channels[0].process = Some(device::controller::process);
+    for i in 0..4 {
+        if device.ui.config.input.controller_enabled[i] {
+            device.pif.channels[i].pak_handler = Some(mempak_handler);
+            device.pif.channels[i].process = Some(device::controller::process);
+        }
+    }
+    if device.ui.config.input.emulate_vru {
+        device.pif.channels[3].pak_handler = None;
+        device.pif.channels[3].process = Some(device::vru::process);
+    }
     device.pif.channels[4].process = Some(device::cart::process)
 }
 

src/device/rdp.rs

@@ -5,41 +5,41 @@ pub const DPC_START_REG: u32 = 0;
 pub const DPC_END_REG: u32 = 1;
 pub const DPC_CURRENT_REG: u32 = 2;
 pub const DPC_STATUS_REG: u32 = 3;
-pub const DPC_CLOCK_REG: u32 = 4;
-pub const DPC_BUFBUSY_REG: u32 = 5;
-pub const DPC_PIPEBUSY_REG: u32 = 6;
-pub const DPC_TMEM_REG: u32 = 7;
+const DPC_CLOCK_REG: u32 = 4;
+const DPC_BUFBUSY_REG: u32 = 5;
+const DPC_PIPEBUSY_REG: u32 = 6;
+const DPC_TMEM_REG: u32 = 7;
 pub const DPC_REGS_COUNT: u32 = 8;
 
-//pub const DPS_TBIST_REG: u32 = 0;
-//pub const DPS_TEST_MODE_REG: u32 = 1;
-//pub const DPS_BUFTEST_ADDR_REG: u32 = 2;
-//pub const DPS_BUFTEST_DATA_REG: u32 = 3;
+//const DPS_TBIST_REG: u32 = 0;
+//const DPS_TEST_MODE_REG: u32 = 1;
+//const DPS_BUFTEST_ADDR_REG: u32 = 2;
+//const DPS_BUFTEST_DATA_REG: u32 = 3;
 pub const DPS_REGS_COUNT: u32 = 4;
 
 /* DPC status - read */
-pub const DPC_STATUS_XBUS_DMEM_DMA: u32 = 1 << 0;
-pub const DPC_STATUS_FREEZE: u32 = 1 << 1;
-pub const DPC_STATUS_FLUSH: u32 = 1 << 2;
-pub const DPC_STATUS_START_GCLK: u32 = 1 << 3;
-pub const DPC_STATUS_TMEM_BUSY: u32 = 1 << 4;
-pub const DPC_STATUS_PIPE_BUSY: u32 = 1 << 5;
-pub const DPC_STATUS_CMD_BUSY: u32 = 1 << 6;
-pub const DPC_STATUS_CBUF_READY: u32 = 1 << 7;
-//pub const DPC_STATUS_DMA_BUSY: u32 = 1 << 8;
-//pub const DPC_STATUS_END_VALID: u32 = 1 << 9;
-pub const DPC_STATUS_START_VALID: u32 = 1 << 10;
+const DPC_STATUS_XBUS_DMEM_DMA: u32 = 1 << 0;
+const DPC_STATUS_FREEZE: u32 = 1 << 1;
+const DPC_STATUS_FLUSH: u32 = 1 << 2;
+const DPC_STATUS_START_GCLK: u32 = 1 << 3;
+const DPC_STATUS_TMEM_BUSY: u32 = 1 << 4;
+const DPC_STATUS_PIPE_BUSY: u32 = 1 << 5;
+const DPC_STATUS_CMD_BUSY: u32 = 1 << 6;
+const DPC_STATUS_CBUF_READY: u32 = 1 << 7;
+//const DPC_STATUS_DMA_BUSY: u32 = 1 << 8;
+//const DPC_STATUS_END_VALID: u32 = 1 << 9;
+const DPC_STATUS_START_VALID: u32 = 1 << 10;
 /* DPC status - write */
-pub const DPC_CLR_XBUS_DMEM_DMA: u32 = 1 << 0;
-pub const DPC_SET_XBUS_DMEM_DMA: u32 = 1 << 1;
-pub const DPC_CLR_FREEZE: u32 = 1 << 2;
-pub const DPC_SET_FREEZE: u32 = 1 << 3;
-pub const DPC_CLR_FLUSH: u32 = 1 << 4;
-pub const DPC_SET_FLUSH: u32 = 1 << 5;
-pub const DPC_CLR_TMEM_CTR: u32 = 1 << 6;
-pub const DPC_CLR_PIPE_CTR: u32 = 1 << 7;
-pub const DPC_CLR_CMD_CTR: u32 = 1 << 8;
-pub const DPC_CLR_CLOCK_CTR: u32 = 1 << 9;
+const DPC_CLR_XBUS_DMEM_DMA: u32 = 1 << 0;
+const DPC_SET_XBUS_DMEM_DMA: u32 = 1 << 1;
+const DPC_CLR_FREEZE: u32 = 1 << 2;
+const DPC_SET_FREEZE: u32 = 1 << 3;
+const DPC_CLR_FLUSH: u32 = 1 << 4;
+const DPC_SET_FLUSH: u32 = 1 << 5;
+const DPC_CLR_TMEM_CTR: u32 = 1 << 6;
+const DPC_CLR_PIPE_CTR: u32 = 1 << 7;
+const DPC_CLR_CMD_CTR: u32 = 1 << 8;
+const DPC_CLR_CLOCK_CTR: u32 = 1 << 9;
 
 pub struct Rdp {
     pub regs_dpc: [u32; DPC_REGS_COUNT as usize],
@@ -92,7 +92,8 @@ pub fn write_regs_dpc(device: &mut device::Device, address: u64, value: u32, mas
 }
 
 pub fn run_rdp(device: &mut device::Device) {
-    let timer = ui::video::process_rdp_list(&mut device.rdp.regs_dpc, &mut device.rsp.mem);
+    let timer = ui::video::process_rdp_list();
+
     if timer != 0 {
         device::events::create_event(
             device,
@@ -173,6 +174,7 @@ pub fn init(device: &mut device::Device) {
 }
 
 pub fn rdp_interrupt_event(device: &mut device::Device) {
-    ui::video::rdp_full_sync();
+    ui::video::full_sync();
+
     device::mi::set_rcp_interrupt(device, device::mi::MI_INTR_DP)
 }

src/device/rdram.rs

@@ -3,10 +3,9 @@ use std::alloc::{alloc_zeroed, Layout};
 
 pub const RDRAM_MASK: usize = 0xFFFFFF;
 
-pub const RDRAM_SIZE: usize = 0x800000; // 8MB
-
 pub struct Rdram {
     pub mem: Vec<u8>,
+    pub size: u32,
 }
 
 pub fn read_mem_fast(
@@ -32,7 +31,7 @@ pub fn read_mem(
         rdram_calculate_cycles(access_size as u64) / (access_size as u64 / 4),
     );
     let masked_address = address as usize & RDRAM_MASK;
-    if address < RDRAM_SIZE as u64 {
+    if address < device.rdram.size as u64 {
         u32::from_ne_bytes(
             device.rdram.mem[masked_address..masked_address + 4]
                 .try_into()
@@ -44,7 +43,7 @@ pub fn read_mem(
 }
 
 pub fn write_mem(device: &mut device::Device, address: u64, value: u32, mask: u32) {
-    if address < RDRAM_SIZE as u64 {
+    if address < device.rdram.size as u64 {
         let mut data = u32::from_ne_bytes(
             device.rdram.mem[address as usize..(address + 4) as usize]
                 .try_into()
@@ -58,29 +57,30 @@ pub fn write_mem(device: &mut device::Device, address: u64, value: u32, mask: u3
 
 pub fn read_regs(
     _device: &mut device::Device,
-    _address: u64,
+    address: u64,
     _access_size: device::memory::AccessSize,
 ) -> u32 {
-    panic!("rdram read reg");
+    println!("Warning: read from RDRAM reg {:#x}", address);
+    0
 }
 
 pub fn write_regs(_device: &mut device::Device, _address: u64, _value: u32, _mask: u32) {
     panic!("rdram write reg");
 }
 
-pub fn init(device: &mut device::Device) -> (*mut u8, usize) {
+pub fn init(device: &mut device::Device) {
     let alignment = 64 * 1024;
-    let layout = Layout::from_size_align(RDRAM_SIZE, alignment).expect("Invalid layout");
+    let layout =
+        Layout::from_size_align(device.rdram.size as usize, alignment).expect("Invalid layout");
     let ptr = unsafe { alloc_zeroed(layout) };
-    device.rdram.mem = unsafe { Vec::from_raw_parts(ptr, RDRAM_SIZE, RDRAM_SIZE) };
+    device.rdram.mem =
+        unsafe { Vec::from_raw_parts(ptr, device.rdram.size as usize, device.rdram.size as usize) };
 
     // hack, skip RDRAM initialization
-    let data: u32 = RDRAM_SIZE as u32;
     device.rdram.mem[device.cart.rdram_size_offset..device.cart.rdram_size_offset + 4]
-        .copy_from_slice(&data.to_ne_bytes());
-    (ptr, RDRAM_SIZE)
+        .copy_from_slice(&device.rdram.size.to_ne_bytes());
 }
 
 pub fn rdram_calculate_cycles(length: u64) -> u64 {
-    31 + (length / 3)// https://hcs64.com/dma.html, https://github.com/rasky/n64-systembench
+    31 + (length / 3) // https://hcs64.com/dma.html, https://github.com/rasky/n64-systembench
 }

src/device/ri.rs

@@ -1,13 +1,13 @@
 use crate::device;
 
-//pub const RI_MODE_REG: u32 = 0;
-//pub const RI_CONFIG_REG: u32 = 1;
-//pub const RI_CURRENT_LOAD_REG: u32 = 2;
-pub const RI_SELECT_REG: u32 = 3;
-//pub const RI_REFRESH_REG: u32 = 4;
-//pub const RI_LATENCY_REG: u32 = 5;
-//pub const RI_ERROR_REG: u32 = 6;
-//pub const RI_WERROR_REG: u32 = 7;
+//const RI_MODE_REG: u32 = 0;
+//const RI_CONFIG_REG: u32 = 1;
+//const RI_CURRENT_LOAD_REG: u32 = 2;
+const RI_SELECT_REG: u32 = 3;
+//const RI_REFRESH_REG: u32 = 4;
+//const RI_LATENCY_REG: u32 = 5;
+//const RI_ERROR_REG: u32 = 6;
+//const RI_WERROR_REG: u32 = 7;
 pub const RI_REGS_COUNT: u32 = 8;
 
 pub struct Ri {

src/device/rsp_cpu.rs

@@ -1,4 +1,9 @@
+#[cfg(target_arch = "aarch64")]
+use device::__m128i;
+#[cfg(target_arch = "aarch64")]
+include!(concat!(env!("OUT_DIR"), "/simd_bindings.rs"));
 use crate::device;
+#[cfg(target_arch = "x86_64")]
 use std::arch::x86_64::*;
 
 pub struct BranchState {
@@ -30,17 +35,17 @@ pub struct Cpu {
     pub cycle_counter: u64,
     pub shuffle: [__m128i; 16],
     pub gpr: [u32; 32],
-    pub vpr: [u128; 32],
+    pub vpr: [__m128i; 32],
     pub reciprocals: [u16; 512],
     pub inverse_square_roots: [u16; 512],
-    pub vcol: std::arch::x86_64::__m128i,
-    pub vcoh: std::arch::x86_64::__m128i,
-    pub vccl: std::arch::x86_64::__m128i,
-    pub vcch: std::arch::x86_64::__m128i,
-    pub vce: std::arch::x86_64::__m128i,
-    pub accl: std::arch::x86_64::__m128i,
-    pub accm: std::arch::x86_64::__m128i,
-    pub acch: std::arch::x86_64::__m128i,
+    pub vcol: __m128i,
+    pub vcoh: __m128i,
+    pub vccl: __m128i,
+    pub vcch: __m128i,
+    pub vce: __m128i,
+    pub accl: __m128i,
+    pub accm: __m128i,
+    pub acch: __m128i,
     pub divdp: bool,
     pub divin: i16,
     pub divout: i16,

src/device/rsp_interface.rs

@@ -1,64 +1,64 @@
 use crate::device;
 
-pub const SP_MEM_ADDR_REG: u32 = 0;
-pub const SP_DRAM_ADDR_REG: u32 = 1;
-pub const SP_RD_LEN_REG: u32 = 2;
-pub const SP_WR_LEN_REG: u32 = 3;
+const SP_MEM_ADDR_REG: u32 = 0;
+const SP_DRAM_ADDR_REG: u32 = 1;
+const SP_RD_LEN_REG: u32 = 2;
+const SP_WR_LEN_REG: u32 = 3;
 pub const SP_STATUS_REG: u32 = 4;
-pub const SP_DMA_FULL_REG: u32 = 5;
-pub const SP_DMA_BUSY_REG: u32 = 6;
-pub const SP_SEMAPHORE_REG: u32 = 7;
+const SP_DMA_FULL_REG: u32 = 5;
+const SP_DMA_BUSY_REG: u32 = 6;
+const SP_SEMAPHORE_REG: u32 = 7;
 pub const SP_REGS_COUNT: u32 = 8;
 
 pub const SP_PC_REG: u32 = 0;
-//pub const SP_IBIST_REG: u32 = 1;
+//const SP_IBIST_REG: u32 = 1;
 pub const SP_REGS2_COUNT: u32 = 2;
 
 /* SP_STATUS - read */
 pub const SP_STATUS_HALT: u32 = 1 << 0;
-pub const SP_STATUS_BROKE: u32 = 1 << 1;
-pub const SP_STATUS_DMA_BUSY: u32 = 1 << 2;
-pub const SP_STATUS_DMA_FULL: u32 = 1 << 3;
-//pub const SP_STATUS_IO_FULL: u32 = 1 << 4;
-pub const SP_STATUS_SSTEP: u32 = 1 << 5;
-pub const SP_STATUS_INTR_BREAK: u32 = 1 << 6;
-pub const SP_STATUS_SIG0: u32 = 1 << 7;
-pub const SP_STATUS_SIG1: u32 = 1 << 8;
-pub const SP_STATUS_SIG2: u32 = 1 << 9;
-pub const SP_STATUS_SIG3: u32 = 1 << 10;
-pub const SP_STATUS_SIG4: u32 = 1 << 11;
-pub const SP_STATUS_SIG5: u32 = 1 << 12;
-pub const SP_STATUS_SIG6: u32 = 1 << 13;
-pub const SP_STATUS_SIG7: u32 = 1 << 14;
+const SP_STATUS_BROKE: u32 = 1 << 1;
+const SP_STATUS_DMA_BUSY: u32 = 1 << 2;
+const SP_STATUS_DMA_FULL: u32 = 1 << 3;
+//const SP_STATUS_IO_FULL: u32 = 1 << 4;
+const SP_STATUS_SSTEP: u32 = 1 << 5;
+const SP_STATUS_INTR_BREAK: u32 = 1 << 6;
+const SP_STATUS_SIG0: u32 = 1 << 7;
+const SP_STATUS_SIG1: u32 = 1 << 8;
+const SP_STATUS_SIG2: u32 = 1 << 9;
+const SP_STATUS_SIG3: u32 = 1 << 10;
+const SP_STATUS_SIG4: u32 = 1 << 11;
+const SP_STATUS_SIG5: u32 = 1 << 12;
+const SP_STATUS_SIG6: u32 = 1 << 13;
+const SP_STATUS_SIG7: u32 = 1 << 14;
 
 /* SP_STATUS - write */
-pub const SP_CLR_HALT: u32 = 1 << 0;
+const SP_CLR_HALT: u32 = 1 << 0;
 pub const SP_SET_HALT: u32 = 1 << 1;
-pub const SP_CLR_BROKE: u32 = 1 << 2;
-pub const SP_CLR_INTR: u32 = 1 << 3;
-pub const SP_SET_INTR: u32 = 1 << 4;
-pub const SP_CLR_SSTEP: u32 = 1 << 5;
-pub const SP_SET_SSTEP: u32 = 1 << 6;
-pub const SP_CLR_INTR_BREAK: u32 = 1 << 7;
-pub const SP_SET_INTR_BREAK: u32 = 1 << 8;
-pub const SP_CLR_SIG0: u32 = 1 << 9;
-pub const SP_SET_SIG0: u32 = 1 << 10;
-pub const SP_CLR_SIG1: u32 = 1 << 11;
-pub const SP_SET_SIG1: u32 = 1 << 12;
-pub const SP_CLR_SIG2: u32 = 1 << 13;
-pub const SP_SET_SIG2: u32 = 1 << 14;
-pub const SP_CLR_SIG3: u32 = 1 << 15;
-pub const SP_SET_SIG3: u32 = 1 << 16;
-pub const SP_CLR_SIG4: u32 = 1 << 17;
-pub const SP_SET_SIG4: u32 = 1 << 18;
-pub const SP_CLR_SIG5: u32 = 1 << 19;
-pub const SP_SET_SIG5: u32 = 1 << 20;
-pub const SP_CLR_SIG6: u32 = 1 << 21;
-pub const SP_SET_SIG6: u32 = 1 << 22;
-pub const SP_CLR_SIG7: u32 = 1 << 23;
-pub const SP_SET_SIG7: u32 = 1 << 24;
+const SP_CLR_BROKE: u32 = 1 << 2;
+const SP_CLR_INTR: u32 = 1 << 3;
+const SP_SET_INTR: u32 = 1 << 4;
+const SP_CLR_SSTEP: u32 = 1 << 5;
+const SP_SET_SSTEP: u32 = 1 << 6;
+const SP_CLR_INTR_BREAK: u32 = 1 << 7;
+const SP_SET_INTR_BREAK: u32 = 1 << 8;
+const SP_CLR_SIG0: u32 = 1 << 9;
+const SP_SET_SIG0: u32 = 1 << 10;
+const SP_CLR_SIG1: u32 = 1 << 11;
+const SP_SET_SIG1: u32 = 1 << 12;
+const SP_CLR_SIG2: u32 = 1 << 13;
+const SP_SET_SIG2: u32 = 1 << 14;
+const SP_CLR_SIG3: u32 = 1 << 15;
+const SP_SET_SIG3: u32 = 1 << 16;
+const SP_CLR_SIG4: u32 = 1 << 17;
+const SP_SET_SIG4: u32 = 1 << 18;
+const SP_CLR_SIG5: u32 = 1 << 19;
+const SP_SET_SIG5: u32 = 1 << 20;
+const SP_CLR_SIG6: u32 = 1 << 21;
+const SP_SET_SIG6: u32 = 1 << 22;
+const SP_CLR_SIG7: u32 = 1 << 23;
+const SP_SET_SIG7: u32 = 1 << 24;
 
-pub const RSP_MEM_MASK: usize = 0x1FFF;
+const RSP_MEM_MASK: usize = 0x1FFF;
 
 #[derive(PartialEq, Copy, Clone)]
 pub enum DmaDir {
@@ -165,11 +165,14 @@ pub fn do_dma(device: &mut device::Device, dma: RspDma) {
         while j < count {
             let mut i = 0;
             while i < length {
-                let data = u32::from_ne_bytes(
-                    device.rdram.mem[dram_addr as usize..dram_addr as usize + 4]
-                        .try_into()
-                        .unwrap(),
-                );
+                let mut data = 0;
+                if dram_addr < device.rdram.size {
+                    data = u32::from_ne_bytes(
+                        device.rdram.mem[dram_addr as usize..dram_addr as usize + 4]
+                            .try_into()
+                            .unwrap(),
+                    );
+                }
                 if offset != 0 {
                     // imem being updated
                     device.rsp.cpu.instructions[((mem_addr & 0xFFF) / 4) as usize].func =

src/device/rsp_su_instructions.rs

@@ -1,4 +1,11 @@
+#[cfg(target_arch = "aarch64")]
+use device::__m128i;
+#[cfg(target_arch = "aarch64")]
+include!(concat!(env!("OUT_DIR"), "/simd_bindings.rs"));
+#[cfg(target_arch = "aarch64")]
+include!("../compat/aarch64.rs");
 use crate::device;
+#[cfg(target_arch = "x86_64")]
 use std::arch::x86_64::*;
 
 pub fn rd(opcode: u32) -> u32 {
@@ -39,28 +46,136 @@ pub fn sign_extend_7bit_offset(offset: u8, shift_amount: u32) -> u32 {
     (((soffset) as i32) as u32) << shift_amount
 }
 
-pub fn modify_vpr_byte(vpr: &mut u128, value: u8, element: u8) {
-    let pos = 15 - (element & 15);
-    let mask = 0xFF << (pos * 8);
-    *vpr &= !mask;
-    *vpr |= (value as u128) << (pos * 8);
+pub fn modify_vpr8(vpr: &mut __m128i, element: u8, value: u8) {
+    unsafe {
+        *vpr = match element & 15 {
+            0 => _mm_insert_epi8(*vpr, value as i32, 15),
+            1 => _mm_insert_epi8(*vpr, value as i32, 14),
+            2 => _mm_insert_epi8(*vpr, value as i32, 13),
+            3 => _mm_insert_epi8(*vpr, value as i32, 12),
+            4 => _mm_insert_epi8(*vpr, value as i32, 11),
+            5 => _mm_insert_epi8(*vpr, value as i32, 10),
+            6 => _mm_insert_epi8(*vpr, value as i32, 9),
+            7 => _mm_insert_epi8(*vpr, value as i32, 8),
+            8 => _mm_insert_epi8(*vpr, value as i32, 7),
+            9 => _mm_insert_epi8(*vpr, value as i32, 6),
+            10 => _mm_insert_epi8(*vpr, value as i32, 5),
+            11 => _mm_insert_epi8(*vpr, value as i32, 4),
+            12 => _mm_insert_epi8(*vpr, value as i32, 3),
+            13 => _mm_insert_epi8(*vpr, value as i32, 2),
+            14 => _mm_insert_epi8(*vpr, value as i32, 1),
+            15 => _mm_insert_epi8(*vpr, value as i32, 0),
+            _ => unreachable!(),
+        };
+    }
 }
 
-pub fn get_vpr_byte(vpr: u128, element: u8) -> u8 {
-    let pos = 15 - (element & 15);
-    (vpr >> (pos * 8)) as u8
+pub fn get_vpr8(vpr: __m128i, element: u8) -> u8 {
+    unsafe {
+        match element & 15 {
+            0 => _mm_extract_epi8(vpr, 15) as u8,
+            1 => _mm_extract_epi8(vpr, 14) as u8,
+            2 => _mm_extract_epi8(vpr, 13) as u8,
+            3 => _mm_extract_epi8(vpr, 12) as u8,
+            4 => _mm_extract_epi8(vpr, 11) as u8,
+            5 => _mm_extract_epi8(vpr, 10) as u8,
+            6 => _mm_extract_epi8(vpr, 9) as u8,
+            7 => _mm_extract_epi8(vpr, 8) as u8,
+            8 => _mm_extract_epi8(vpr, 7) as u8,
+            9 => _mm_extract_epi8(vpr, 6) as u8,
+            10 => _mm_extract_epi8(vpr, 5) as u8,
+            11 => _mm_extract_epi8(vpr, 4) as u8,
+            12 => _mm_extract_epi8(vpr, 3) as u8,
+            13 => _mm_extract_epi8(vpr, 2) as u8,
+            14 => _mm_extract_epi8(vpr, 1) as u8,
+            15 => _mm_extract_epi8(vpr, 0) as u8,
+            _ => unreachable!(),
+        }
+    }
 }
 
-pub fn modify_vpr_element(vpr: &mut u128, value: u16, element: u8) {
-    let pos = 7 - (element & 7);
-    let mask = 0xFFFF << (pos * 16);
-    *vpr &= !mask;
-    *vpr |= (value as u128) << (pos * 16);
+pub fn modify_vpr16(vpr: &mut __m128i, element: u8, value: u16) {
+    unsafe {
+        *vpr = match element & 7 {
+            0 => _mm_insert_epi16(*vpr, value as i32, 7),
+            1 => _mm_insert_epi16(*vpr, value as i32, 6),
+            2 => _mm_insert_epi16(*vpr, value as i32, 5),
+            3 => _mm_insert_epi16(*vpr, value as i32, 4),
+            4 => _mm_insert_epi16(*vpr, value as i32, 3),
+            5 => _mm_insert_epi16(*vpr, value as i32, 2),
+            6 => _mm_insert_epi16(*vpr, value as i32, 1),
+            7 => _mm_insert_epi16(*vpr, value as i32, 0),
+            _ => unreachable!(),
+        };
+    }
 }
 
-pub fn get_vpr_element(vpr: u128, element: u8) -> u16 {
-    let pos = 7 - (element & 7);
-    (vpr >> (pos * 16)) as u16
+pub fn get_vpr16(vpr: __m128i, element: u8) -> u16 {
+    unsafe {
+        match element & 7 {
+            0 => _mm_extract_epi16(vpr, 7) as u16,
+            1 => _mm_extract_epi16(vpr, 6) as u16,
+            2 => _mm_extract_epi16(vpr, 5) as u16,
+            3 => _mm_extract_epi16(vpr, 4) as u16,
+            4 => _mm_extract_epi16(vpr, 3) as u16,
+            5 => _mm_extract_epi16(vpr, 2) as u16,
+            6 => _mm_extract_epi16(vpr, 1) as u16,
+            7 => _mm_extract_epi16(vpr, 0) as u16,
+            _ => unreachable!(),
+        }
+    }
+}
+
+pub fn modify_vpr32(vpr: &mut __m128i, element: u8, value: u32) {
+    unsafe {
+        *vpr = match element & 3 {
+            0 => _mm_insert_epi32(*vpr, value as i32, 3),
+            1 => _mm_insert_epi32(*vpr, value as i32, 2),
+            2 => _mm_insert_epi32(*vpr, value as i32, 1),
+            3 => _mm_insert_epi32(*vpr, value as i32, 0),
+            _ => unreachable!(),
+        };
+    }
+}
+
+pub fn get_vpr32(vpr: __m128i, element: u8) -> u32 {
+    unsafe {
+        match element & 3 {
+            0 => _mm_extract_epi32(vpr, 3) as u32,
+            1 => _mm_extract_epi32(vpr, 2) as u32,
+            2 => _mm_extract_epi32(vpr, 1) as u32,
+            3 => _mm_extract_epi32(vpr, 0) as u32,
+            _ => unreachable!(),
+        }
+    }
+}
+
+pub fn modify_vpr64(vpr: &mut __m128i, element: u8, value: u64) {
+    unsafe {
+        *vpr = match element & 1 {
+            0 => _mm_insert_epi64(*vpr, value as i64, 1),
+            1 => _mm_insert_epi64(*vpr, value as i64, 0),
+            _ => unreachable!(),
+        };
+    }
+}
+
+pub fn get_vpr64(vpr: __m128i, element: u8) -> u64 {
+    unsafe {
+        match element & 1 {
+            0 => _mm_extract_epi64(vpr, 1) as u64,
+            1 => _mm_extract_epi64(vpr, 0) as u64,
+            _ => unreachable!(),
+        }
+    }
+}
+
+pub fn modify_vpr128(vpr: &mut __m128i, value: u128) {
+    unsafe { *vpr = std::mem::transmute::<u128, __m128i>(value) }
+}
+
+pub fn get_vpr128(vpr: __m128i) -> u128 {
+    unsafe { std::mem::transmute::<__m128i, u128>(vpr) }
 }
 
 pub fn j(device: &mut device::Device, opcode: u32) {
@@ -208,7 +323,7 @@ pub fn lw(device: &mut device::Device, opcode: u32) {
     w[3] = device.rsp.mem[(address as usize + 3) & 0xFFF];
 
     device.rsp.cpu.gpr[rt(opcode) as usize] =
-        (w[0] as u32) << 24 | (w[1] as u32) << 16 | (w[2] as u32) << 8 | (w[3] as u32)
+        ((w[0] as u32) << 24) | ((w[1] as u32) << 16) | ((w[2] as u32) << 8) | (w[3] as u32)
 }
 
 pub fn lbu(device: &mut device::Device, opcode: u32) {
@@ -240,7 +355,7 @@ pub fn lwu(device: &mut device::Device, opcode: u32) {
     w[3] = device.rsp.mem[(address as usize + 3) & 0xFFF];
 
     device.rsp.cpu.gpr[rt(opcode) as usize] =
-        (w[0] as u32) << 24 | (w[1] as u32) << 16 | (w[2] as u32) << 8 | (w[3] as u32)
+        ((w[0] as u32) << 24) | ((w[1] as u32) << 16) | ((w[2] as u32) << 8) | (w[3] as u32)
 }
 
 pub fn sb(device: &mut device::Device, opcode: u32) {
@@ -481,12 +596,13 @@ pub fn mtc0(device: &mut device::Device, opcode: u32) {
 }
 
 pub fn mfc2(device: &mut device::Device, opcode: u32) {
-    let hi = get_vpr_byte(device.rsp.cpu.vpr[rd(opcode) as usize], velement(opcode));
-    let lo = get_vpr_byte(
+    let hi = get_vpr8(device.rsp.cpu.vpr[rd(opcode) as usize], velement(opcode));
+    let lo = get_vpr8(
         device.rsp.cpu.vpr[rd(opcode) as usize],
         velement(opcode) + 1,
     );
-    device.rsp.cpu.gpr[rt(opcode) as usize] = ((hi as u16) << 8 | (lo as u16)) as i16 as i32 as u32
+    device.rsp.cpu.gpr[rt(opcode) as usize] =
+        (((hi as u16) << 8) | (lo as u16)) as i16 as i32 as u32
 }
 
 pub fn cfc2(device: &mut device::Device, opcode: u32) {
@@ -507,7 +623,7 @@ pub fn cfc2(device: &mut device::Device, opcode: u32) {
             lo = &mut device.rsp.cpu.vce;
         }
         _ => {
-            panic!("unknown ctc2")
+            panic!("unknown cfc2")
         }
     }
 
@@ -519,16 +635,16 @@ pub fn cfc2(device: &mut device::Device, opcode: u32) {
 }
 
 pub fn mtc2(device: &mut device::Device, opcode: u32) {
-    modify_vpr_byte(
+    modify_vpr8(
         &mut device.rsp.cpu.vpr[rd(opcode) as usize],
-        (device.rsp.cpu.gpr[rt(opcode) as usize] >> 8) as u8,
         velement(opcode),
+        (device.rsp.cpu.gpr[rt(opcode) as usize] >> 8) as u8,
     );
     if velement(opcode) != 15 {
-        modify_vpr_byte(
+        modify_vpr8(
             &mut device.rsp.cpu.vpr[rd(opcode) as usize],
-            device.rsp.cpu.gpr[rt(opcode) as usize] as u8,
             velement(opcode) + 1,
+            device.rsp.cpu.gpr[rt(opcode) as usize] as u8,
         );
     }
 }
@@ -573,7 +689,7 @@ pub fn ctc2(device: &mut device::Device, opcode: u32) {
             ),
             _mm_setzero_si128(),
         );
-        *hi = std::arch::x86_64::_mm_cmpeq_epi8(
+        *hi = _mm_cmpeq_epi8(
             _mm_and_si128(
                 _mm_set1_epi8(!(device.rsp.cpu.gpr[rt(opcode) as usize] >> 8) as i8),
                 mask,
@@ -588,10 +704,10 @@ pub fn lbv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 0));
 
     let element = velement(opcode);
-    modify_vpr_byte(
+    modify_vpr8(
         &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-        device.rsp.mem[(address & 0xFFF) as usize],
         element,
+        device.rsp.mem[(address & 0xFFF) as usize],
     );
 }
 
@@ -600,15 +716,25 @@ pub fn lsv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 1));
 
     let mut element = velement(opcode);
-    let end = std::cmp::min(element + 2, 16);
-    while element < end {
-        modify_vpr_byte(
+
+    if element % 2 == 0 {
+        modify_vpr16(
             &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            device.rsp.mem[(address & 0xFFF) as usize],
-            element,
+            element / 2,
+            (device.rsp.mem[(address & 0xFFF) as usize] as u16) << 8
+                | (device.rsp.mem[((address + 1) & 0xFFF) as usize] as u16),
         );
-        address += 1;
-        element += 1;
+    } else {
+        let end = std::cmp::min(element + 2, 16);
+        while element < end {
+            modify_vpr8(
+                &mut device.rsp.cpu.vpr[rt(opcode) as usize],
+                element,
+                device.rsp.mem[(address & 0xFFF) as usize],
+            );
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -617,15 +743,27 @@ pub fn llv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 2));
 
     let mut element = velement(opcode);
-    let end = std::cmp::min(element + 4, 16);
-    while element < end {
-        modify_vpr_byte(
+
+    if element % 4 == 0 {
+        modify_vpr32(
             &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            device.rsp.mem[(address & 0xFFF) as usize],
-            element,
+            element / 4,
+            (device.rsp.mem[(address & 0xFFF) as usize] as u32) << (8 * 3)
+                | (device.rsp.mem[((address + 1) & 0xFFF) as usize] as u32) << (8 * 2)
+                | (device.rsp.mem[((address + 2) & 0xFFF) as usize] as u32) << 8
+                | (device.rsp.mem[((address + 3) & 0xFFF) as usize] as u32),
         );
-        address += 1;
-        element += 1;
+    } else {
+        let end = std::cmp::min(element + 4, 16);
+        while element < end {
+            modify_vpr8(
+                &mut device.rsp.cpu.vpr[rt(opcode) as usize],
+                element,
+                device.rsp.mem[(address & 0xFFF) as usize],
+            );
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -634,15 +772,31 @@ pub fn ldv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 3));
 
     let mut element = velement(opcode);
-    let end = std::cmp::min(element + 8, 16);
-    while element < end {
-        modify_vpr_byte(
+
+    if element % 8 == 0 {
+        modify_vpr64(
             &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            device.rsp.mem[(address & 0xFFF) as usize],
-            element,
+            element / 8,
+            (device.rsp.mem[(address & 0xFFF) as usize] as u64) << (8 * 7)
+                | (device.rsp.mem[((address + 1) & 0xFFF) as usize] as u64) << (8 * 6)
+                | (device.rsp.mem[((address + 2) & 0xFFF) as usize] as u64) << (8 * 5)
+                | (device.rsp.mem[((address + 3) & 0xFFF) as usize] as u64) << (8 * 4)
+                | (device.rsp.mem[((address + 4) & 0xFFF) as usize] as u64) << (8 * 3)
+                | (device.rsp.mem[((address + 5) & 0xFFF) as usize] as u64) << (8 * 2)
+                | (device.rsp.mem[((address + 6) & 0xFFF) as usize] as u64) << 8
+                | (device.rsp.mem[((address + 7) & 0xFFF) as usize] as u64),
         );
-        address += 1;
-        element += 1;
+    } else {
+        let end = std::cmp::min(element + 8, 16);
+        while element < end {
+            modify_vpr8(
+                &mut device.rsp.cpu.vpr[rt(opcode) as usize],
+                element,
+                device.rsp.mem[(address & 0xFFF) as usize],
+            );
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -651,15 +805,38 @@ pub fn lqv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 4));
 
     let mut element = velement(opcode);
-    let end = std::cmp::min(16 + element - ((address & 15) as u8), 16);
-    while element < end {
-        modify_vpr_byte(
+
+    if element == 0 && address % 16 == 0 {
+        modify_vpr128(
             &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            device.rsp.mem[(address & 0xFFF) as usize],
-            element,
+            (device.rsp.mem[(address & 0xFFF) as usize] as u128) << (8 * 15)
+                | (device.rsp.mem[((address + 1) & 0xFFF) as usize] as u128) << (8 * 14)
+                | (device.rsp.mem[((address + 2) & 0xFFF) as usize] as u128) << (8 * 13)
+                | (device.rsp.mem[((address + 3) & 0xFFF) as usize] as u128) << (8 * 12)
+                | (device.rsp.mem[((address + 4) & 0xFFF) as usize] as u128) << (8 * 11)
+                | (device.rsp.mem[((address + 5) & 0xFFF) as usize] as u128) << (8 * 10)
+                | (device.rsp.mem[((address + 6) & 0xFFF) as usize] as u128) << (8 * 9)
+                | (device.rsp.mem[((address + 7) & 0xFFF) as usize] as u128) << (8 * 8)
+                | (device.rsp.mem[((address + 8) & 0xFFF) as usize] as u128) << (8 * 7)
+                | (device.rsp.mem[((address + 9) & 0xFFF) as usize] as u128) << (8 * 6)
+                | (device.rsp.mem[((address + 10) & 0xFFF) as usize] as u128) << (8 * 5)
+                | (device.rsp.mem[((address + 11) & 0xFFF) as usize] as u128) << (8 * 4)
+                | (device.rsp.mem[((address + 12) & 0xFFF) as usize] as u128) << (8 * 3)
+                | (device.rsp.mem[((address + 13) & 0xFFF) as usize] as u128) << (8 * 2)
+                | (device.rsp.mem[((address + 14) & 0xFFF) as usize] as u128) << 8
+                | (device.rsp.mem[((address + 15) & 0xFFF) as usize] as u128),
         );
-        address += 1;
-        element += 1;
+    } else {
+        let end = std::cmp::min(16 + element - ((address & 15) as u8), 16);
+        while element < end {
+            modify_vpr8(
+                &mut device.rsp.cpu.vpr[rt(opcode) as usize],
+                element,
+                device.rsp.mem[(address & 0xFFF) as usize],
+            );
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -669,14 +846,36 @@ pub fn lrv(device: &mut device::Device, opcode: u32) {
 
     let mut element = 16u8.wrapping_sub(((address & 15) as u8).wrapping_sub(velement(opcode)));
     address &= !15;
-    while element < 16 {
-        modify_vpr_byte(
+    if element == 0 {
+        modify_vpr128(
             &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            device.rsp.mem[(address & 0xFFF) as usize],
-            element,
+            (device.rsp.mem[(address & 0xFFF) as usize] as u128) << (8 * 15)
+                | (device.rsp.mem[((address + 1) & 0xFFF) as usize] as u128) << (8 * 14)
+                | (device.rsp.mem[((address + 2) & 0xFFF) as usize] as u128) << (8 * 13)
+                | (device.rsp.mem[((address + 3) & 0xFFF) as usize] as u128) << (8 * 12)
+                | (device.rsp.mem[((address + 4) & 0xFFF) as usize] as u128) << (8 * 11)
+                | (device.rsp.mem[((address + 5) & 0xFFF) as usize] as u128) << (8 * 10)
+                | (device.rsp.mem[((address + 6) & 0xFFF) as usize] as u128) << (8 * 9)
+                | (device.rsp.mem[((address + 7) & 0xFFF) as usize] as u128) << (8 * 8)
+                | (device.rsp.mem[((address + 8) & 0xFFF) as usize] as u128) << (8 * 7)
+                | (device.rsp.mem[((address + 9) & 0xFFF) as usize] as u128) << (8 * 6)
+                | (device.rsp.mem[((address + 10) & 0xFFF) as usize] as u128) << (8 * 5)
+                | (device.rsp.mem[((address + 11) & 0xFFF) as usize] as u128) << (8 * 4)
+                | (device.rsp.mem[((address + 12) & 0xFFF) as usize] as u128) << (8 * 3)
+                | (device.rsp.mem[((address + 13) & 0xFFF) as usize] as u128) << (8 * 2)
+                | (device.rsp.mem[((address + 14) & 0xFFF) as usize] as u128) << 8
+                | (device.rsp.mem[((address + 15) & 0xFFF) as usize] as u128),
         );
-        address += 1;
-        element += 1;
+    } else {
+        while element < 16 {
+            modify_vpr8(
+                &mut device.rsp.cpu.vpr[rt(opcode) as usize],
+                element,
+                device.rsp.mem[(address & 0xFFF) as usize],
+            );
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -686,16 +885,34 @@ pub fn lpv(device: &mut device::Device, opcode: u32) {
 
     let index = ((address & 7) as u8).wrapping_sub(velement(opcode));
     address &= !7;
-    let mut offset: u8 = 0;
-    while offset < 8 {
-        modify_vpr_element(
-            &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(offset)) & 15) as u32))
-                & 0xFFF) as usize] as u16)
-                << 8,
-            offset,
-        );
-        offset += 1;
+    unsafe {
+        device.rsp.cpu.vpr[rt(opcode) as usize] =
+            _mm_set_epi16(
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(0)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(1)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(2)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(3)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(4)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(5)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(6)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(7)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 8,
+            );
     }
 }
 
@@ -705,16 +922,34 @@ pub fn luv(device: &mut device::Device, opcode: u32) {
 
     let index = ((address & 7) as u8).wrapping_sub(velement(opcode));
     address &= !7;
-    let mut offset: u8 = 0;
-    while offset < 8 {
-        modify_vpr_element(
-            &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(offset)) & 15) as u32))
-                & 0xFFF) as usize] as u16)
-                << 7,
-            offset,
-        );
-        offset += 1;
+    unsafe {
+        device.rsp.cpu.vpr[rt(opcode) as usize] =
+            _mm_set_epi16(
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(0)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(1)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(2)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(3)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(4)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(5)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(6)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(7)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+            );
     }
 }
 
@@ -724,16 +959,34 @@ pub fn lhv(device: &mut device::Device, opcode: u32) {
 
     let index = ((address & 7) as u8).wrapping_sub(velement(opcode));
     address &= !7;
-    let mut offset: u8 = 0;
-    while offset < 8 {
-        modify_vpr_element(
-            &mut device.rsp.cpu.vpr[rt(opcode) as usize],
-            (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(offset * 2)) & 15) as u32))
-                & 0xFFF) as usize] as u16)
-                << 7,
-            offset,
-        );
-        offset += 1;
+    unsafe {
+        device.rsp.cpu.vpr[rt(opcode) as usize] =
+            _mm_set_epi16(
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(0)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(2)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(4)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(6)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(8)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(10)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(12)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+                (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(14)) & 15) as u32))
+                    & 0xFFF) as usize] as i16)
+                    << 7,
+            );
     }
 }
 
@@ -745,30 +998,30 @@ pub fn lfv(device: &mut device::Device, opcode: u32) {
     address &= !7;
     let start = velement(opcode);
     let end = std::cmp::min(start + 8, 16);
-    let mut tmp: u128 = 0;
+    let mut tmp: __m128i = unsafe { _mm_setzero_si128() };
     let mut offset: u8 = 0;
     while offset < 4 {
-        modify_vpr_element(
+        modify_vpr16(
             &mut tmp,
+            offset,
             (device.rsp.mem[((address.wrapping_add(((index.wrapping_add(offset * 4)) & 15) as u32))
                 & 0xFFF) as usize] as u16)
                 << 7,
-            offset,
         );
-        modify_vpr_element(
+        modify_vpr16(
             &mut tmp,
+            offset + 4,
             (device.rsp.mem[((address
                 .wrapping_add(((index.wrapping_add(offset * 4).wrapping_add(8)) & 15) as u32))
                 & 0xFFF) as usize] as u16)
                 << 7,
-            offset + 4,
         );
         offset += 1;
     }
     offset = start;
     while offset < end {
-        let value = get_vpr_byte(tmp, offset);
-        modify_vpr_byte(&mut device.rsp.cpu.vpr[rt(opcode) as usize], value, offset);
+        let value = get_vpr8(tmp, offset);
+        modify_vpr8(&mut device.rsp.cpu.vpr[rt(opcode) as usize], offset, value);
         offset += 1;
     }
 }
@@ -783,28 +1036,26 @@ pub fn ltv(device: &mut device::Device, opcode: u32) {
     address = begin + (((velement(opcode)) as u32 + (address & 8)) & 15);
     let vtbase = rt(opcode) & !7;
     let mut vtoff = (velement(opcode)) as u32 >> 1;
-    let mut i = 0;
-    while i < 8 {
-        modify_vpr_byte(
+    for i in 0..8 {
+        modify_vpr8(
             &mut device.rsp.cpu.vpr[(vtbase + vtoff) as usize],
-            device.rsp.mem[(address & 0xFFF) as usize],
             i * 2,
+            device.rsp.mem[(address & 0xFFF) as usize],
         );
         address += 1;
         if address == begin + 16 {
             address = begin
         }
-        modify_vpr_byte(
+        modify_vpr8(
             &mut device.rsp.cpu.vpr[(vtbase + vtoff) as usize],
-            device.rsp.mem[(address & 0xFFF) as usize],
             i * 2 + 1,
+            device.rsp.mem[(address & 0xFFF) as usize],
         );
         address += 1;
         if address == begin + 16 {
             address = begin
         }
         vtoff = (vtoff + 1) & 7;
-        i += 1;
     }
 }
 
@@ -813,7 +1064,7 @@ pub fn sbv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 0));
 
     device.rsp.mem[(address & 0xFFF) as usize] =
-        get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], velement(opcode))
+        get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], velement(opcode))
 }
 
 pub fn ssv(device: &mut device::Device, opcode: u32) {
@@ -821,12 +1072,20 @@ pub fn ssv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 1));
 
     let mut element = velement(opcode);
-    let end = element + 2;
-    while element < end {
-        device.rsp.mem[(address & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], element);
-        address += 1;
-        element += 1;
+
+    if element % 2 == 0 {
+        device.rsp.mem[(address & 0xFFF) as usize..((address + 2) & 0xFFF) as usize]
+            .copy_from_slice(
+                &get_vpr16(device.rsp.cpu.vpr[rt(opcode) as usize], element / 2).to_be_bytes(),
+            );
+    } else {
+        let end = element + 2;
+        while element < end {
+            device.rsp.mem[(address & 0xFFF) as usize] =
+                get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], element);
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -835,12 +1094,20 @@ pub fn slv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 2));
 
     let mut element = velement(opcode);
-    let end = element + 4;
-    while element < end {
-        device.rsp.mem[(address & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], element);
-        address += 1;
-        element += 1;
+
+    if element % 4 == 0 {
+        device.rsp.mem[(address & 0xFFF) as usize..((address + 4) & 0xFFF) as usize]
+            .copy_from_slice(
+                &get_vpr32(device.rsp.cpu.vpr[rt(opcode) as usize], element / 4).to_be_bytes(),
+            );
+    } else {
+        let end = element + 4;
+        while element < end {
+            device.rsp.mem[(address & 0xFFF) as usize] =
+                get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], element);
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -849,12 +1116,20 @@ pub fn sdv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 3));
 
     let mut element = velement(opcode);
-    let end = element + 8;
-    while element < end {
-        device.rsp.mem[(address & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], element);
-        address += 1;
-        element += 1;
+
+    if element % 8 == 0 {
+        device.rsp.mem[(address & 0xFFF) as usize..((address + 8) & 0xFFF) as usize]
+            .copy_from_slice(
+                &get_vpr64(device.rsp.cpu.vpr[rt(opcode) as usize], element / 8).to_be_bytes(),
+            );
+    } else {
+        let end = element + 8;
+        while element < end {
+            device.rsp.mem[(address & 0xFFF) as usize] =
+                get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], element);
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -863,12 +1138,18 @@ pub fn sqv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 4));
 
     let mut element = velement(opcode);
-    let end = element + (16 - (address & 15)) as u8;
-    while element < end {
-        device.rsp.mem[(address & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], element);
-        address += 1;
-        element += 1;
+
+    if element == 0 && address % 16 == 0 {
+        device.rsp.mem[(address & 0xFFF) as usize..((address + 16) & 0xFFF) as usize]
+            .copy_from_slice(&get_vpr128(device.rsp.cpu.vpr[rt(opcode) as usize]).to_be_bytes());
+    } else {
+        let end = element + (16 - (address & 15)) as u8;
+        while element < end {
+            device.rsp.mem[(address & 0xFFF) as usize] =
+                get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], element);
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -880,11 +1161,17 @@ pub fn srv(device: &mut device::Device, opcode: u32) {
     let end = element + (address & 15) as u8;
     let base = (16 - (address & 15)) as u8;
     address &= !15;
-    while element < end {
-        device.rsp.mem[(address & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], element + base);
-        address += 1;
-        element += 1;
+
+    if element == 0 && base == 0 {
+        device.rsp.mem[(address & 0xFFF) as usize..((address + 16) & 0xFFF) as usize]
+            .copy_from_slice(&get_vpr128(device.rsp.cpu.vpr[rt(opcode) as usize]).to_be_bytes());
+    } else {
+        while element < end {
+            device.rsp.mem[(address & 0xFFF) as usize] =
+                get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], element + base);
+            address += 1;
+            element += 1;
+        }
     }
 }
 
@@ -897,10 +1184,10 @@ pub fn spv(device: &mut device::Device, opcode: u32) {
     while element < end {
         if (element & 15) < 8 {
             device.rsp.mem[(address & 0xFFF) as usize] =
-                get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], (element & 7) << 1);
+                get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], (element & 7) << 1);
         } else {
             device.rsp.mem[(address & 0xFFF) as usize] =
-                (get_vpr_element(device.rsp.cpu.vpr[rt(opcode) as usize], element) >> 7) as u8;
+                (get_vpr16(device.rsp.cpu.vpr[rt(opcode) as usize], element) >> 7) as u8;
         }
         address += 1;
         element += 1;
@@ -916,10 +1203,10 @@ pub fn suv(device: &mut device::Device, opcode: u32) {
     while element < end {
         if (element & 15) < 8 {
             device.rsp.mem[(address & 0xFFF) as usize] =
-                (get_vpr_element(device.rsp.cpu.vpr[rt(opcode) as usize], element) >> 7) as u8;
+                (get_vpr16(device.rsp.cpu.vpr[rt(opcode) as usize], element) >> 7) as u8;
         } else {
             device.rsp.mem[(address & 0xFFF) as usize] =
-                get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], (element & 7) << 1);
+                get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], (element & 7) << 1);
         }
         address += 1;
         element += 1;
@@ -933,13 +1220,11 @@ pub fn shv(device: &mut device::Device, opcode: u32) {
     let element = velement(opcode);
     let index = (address & 7) as u8;
     address &= !7;
-    let mut offset = 0;
-    while offset < 8 {
+    for offset in 0..8 {
         let byte_val = element + offset * 2;
-        let value = get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], byte_val) << 1
-            | get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], byte_val + 1) >> 7;
+        let value = (get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], byte_val) << 1)
+            | (get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], byte_val + 1) >> 7);
         device.rsp.mem[((address + ((index + offset * 2) & 15) as u32) & 0xFFF) as usize] = value;
-        offset += 1;
     }
 }
 
@@ -948,8 +1233,7 @@ pub fn sfv(device: &mut device::Device, opcode: u32) {
         .wrapping_add(sign_extend_7bit_offset(voffset(opcode), 4));
     let base = address & 7;
     address &= !7;
-    let element = velement(opcode);
-    let elements = match element {
+    let elements = match velement(opcode) {
         0 | 15 => [0, 1, 2, 3],
         1 => [6, 7, 4, 5],
         4 => [1, 2, 3, 0],
@@ -966,12 +1250,10 @@ pub fn sfv(device: &mut device::Device, opcode: u32) {
         }
     };
     let mut offset = 0;
-    let mut i = 0;
-    while i < 4 {
+    for element in elements {
         device.rsp.mem[((address + ((base + offset) & 15)) & 0xFFF) as usize] =
-            (get_vpr_element(device.rsp.cpu.vpr[rt(opcode) as usize], elements[i]) >> 7) as u8;
+            (get_vpr16(device.rsp.cpu.vpr[rt(opcode) as usize], element) >> 7) as u8;
         offset += 4;
-        i += 1;
     }
 }
 
@@ -985,7 +1267,7 @@ pub fn swv(device: &mut device::Device, opcode: u32) {
     address &= !7;
     while element < end {
         device.rsp.mem[((address + (base & 15)) & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[rt(opcode) as usize], element);
+            get_vpr8(device.rsp.cpu.vpr[rt(opcode) as usize], element);
         base += 1;
         element += 1;
     }
@@ -1002,11 +1284,11 @@ pub fn stv(device: &mut device::Device, opcode: u32) {
     let mut offset = start;
     while offset < end {
         device.rsp.mem[((address + (base & 15)) & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[offset as usize], element);
+            get_vpr8(device.rsp.cpu.vpr[offset as usize], element);
         base = base.wrapping_add(1);
         element += 1;
         device.rsp.mem[((address + (base & 15)) & 0xFFF) as usize] =
-            get_vpr_byte(device.rsp.cpu.vpr[offset as usize], element);
+            get_vpr8(device.rsp.cpu.vpr[offset as usize], element);
         base = base.wrapping_add(1);
         element += 1;
         offset += 1;

src/device/si.rs

@@ -1,18 +1,18 @@
 use crate::device;
 
-pub const SI_DRAM_ADDR_REG: u32 = 0;
-pub const SI_PIF_ADDR_RD64B_REG: u32 = 1;
-//pub const SI_R2_REG: u32 = 2;
-//pub const SI_R3_REG: u32 = 3;
-pub const SI_PIF_ADDR_WR64B_REG: u32 = 4;
-//pub const SI_R5_REG: u32 = 5;
+const SI_DRAM_ADDR_REG: u32 = 0;
+const SI_PIF_ADDR_RD64B_REG: u32 = 1;
+//const SI_R2_REG: u32 = 2;
+//const SI_R3_REG: u32 = 3;
+const SI_PIF_ADDR_WR64B_REG: u32 = 4;
+//const SI_R5_REG: u32 = 5;
 pub const SI_STATUS_REG: u32 = 6;
 pub const SI_REGS_COUNT: u32 = 7;
 
 pub const SI_STATUS_DMA_BUSY: u32 = 1 << 0;
 pub const SI_STATUS_IO_BUSY: u32 = 1 << 1;
-//pub const SI_STATUS_DMA_ERROR: u32 = 1 << 3;
-pub const SI_STATUS_INTERRUPT: u32 = 1 << 12;
+//const SI_STATUS_DMA_ERROR: u32 = 1 << 3;
+const SI_STATUS_INTERRUPT: u32 = 1 << 12;
 
 #[derive(PartialEq)]
 pub enum DmaDir {
@@ -59,7 +59,7 @@ pub fn dma_write(device: &mut device::Device) {
     device::events::create_event(
         device,
         device::events::EventType::SI,
-        device.cpu.cop0.regs[device::cop0::COP0_COUNT_REG as usize] + 5200, //based on https://github.com/rasky/n64-systembench
+        device.cpu.cop0.regs[device::cop0::COP0_COUNT_REG as usize] + 6000, //based on https://github.com/rasky/n64-systembench
         dma_event,
     )
 }

src/device/sram.rs

@@ -1,13 +1,13 @@
 use crate::device;
 use crate::ui;
 
-pub const SRAM_MASK: usize = 0xFFFF;
-pub const SRAM_SIZE: usize = 0x8000;
-pub const FLASHRAM_SIZE: usize = 0x20000;
+const SRAM_MASK: usize = 0xFFFF;
+const SRAM_SIZE: usize = 0x8000;
+const FLASHRAM_SIZE: usize = 0x20000;
 pub const FLASHRAM_TYPE_ID: u32 = 0x11118001;
 pub const MX29L1100_ID: u32 = 0x00c2001e;
-pub const MX29L0000_ID: u32 = 0x00c20000;
-pub const MX29L0001_ID: u32 = 0x00c20001;
+const MX29L0000_ID: u32 = 0x00c20000;
+const MX29L0001_ID: u32 = 0x00c20001;
 
 #[derive(PartialEq)]
 pub enum FlashramMode {
@@ -141,7 +141,7 @@ pub fn dma_read_sram(
 
     format_sram(device);
 
-    while i < dram_addr + length {
+    while i < dram_addr + length && i < device.rdram.size {
         device.ui.saves.sram.0[j as usize] = device.rdram.mem[i as usize ^ device.byte_swap];
         i += 1;
         j += 1;
@@ -191,7 +191,7 @@ pub fn dma_write_sram(
 
     format_sram(device);
 
-    while i < dram_addr + length {
+    while i < dram_addr + length && i < device.rdram.size {
         device.rdram.mem[i as usize ^ device.byte_swap] = device.ui.saves.sram.0[j as usize];
         i += 1;
         j += 1;

src/device/tlb.rs

@@ -38,8 +38,8 @@ pub fn read(device: &mut device::Device, index: u64) {
         device.cpu.cop0.tlb_entries[index as usize].mask << 13;
 
     device.cpu.cop0.regs[device::cop0::COP0_ENTRYHI_REG as usize] =
-        (device.cpu.cop0.tlb_entries[index as usize].region as u64) << 62
-            | device.cpu.cop0.tlb_entries[index as usize].vpn2 << 13
+        ((device.cpu.cop0.tlb_entries[index as usize].region as u64) << 62)
+            | (device.cpu.cop0.tlb_entries[index as usize].vpn2 << 13)
             | (device.cpu.cop0.tlb_entries[index as usize].asid) as u64;
 
     device.cpu.cop0.regs[device::cop0::COP0_ENTRYLO0_REG as usize] =
@@ -136,7 +136,8 @@ pub fn probe(device: &mut device::Device) {
         if e.region != (device.cpu.cop0.regs[device::cop0::COP0_ENTRYHI_REG as usize] >> 62) as u8 {
             continue;
         }
-        if e.g == 0 && e.asid != device.cpu.cop0.regs[device::cop0::COP0_ENTRYHI_REG as usize] as u8 {
+        if e.g == 0 && e.asid != device.cpu.cop0.regs[device::cop0::COP0_ENTRYHI_REG as usize] as u8
+        {
             continue;
         }
         device.cpu.cop0.regs[device::cop0::COP0_INDEX_REG as usize] = pos as u64;
@@ -185,8 +186,11 @@ pub fn tlb_unmap(device: &mut device::Device, index: u64) {
 pub fn tlb_map(device: &mut device::Device, index: u64) {
     let e = &mut device.cpu.cop0.tlb_entries[index as usize];
 
-    if e.v_even != 0 && e.start_even < e.end_even
-            && !(e.start_even >= 0x80000000 && e.end_even < 0xC0000000) && e.phys_even < 0x20000000 {
+    if e.v_even != 0
+        && e.start_even < e.end_even
+        && !(e.start_even >= 0x80000000 && e.end_even < 0xC0000000)
+        && e.phys_even < 0x20000000
+    {
         let mut i = e.start_even;
         while i < e.end_even {
             device.cpu.cop0.tlb_lut_r[(i >> 12) as usize].address =
@@ -205,8 +209,11 @@ pub fn tlb_map(device: &mut device::Device, index: u64) {
         }
     }
 
-    if e.v_odd != 0 && e.start_odd < e.end_odd
-            && !(e.start_odd >= 0x80000000 && e.end_odd < 0xC0000000) && e.phys_odd < 0x20000000 {
+    if e.v_odd != 0
+        && e.start_odd < e.end_odd
+        && !(e.start_odd >= 0x80000000 && e.end_odd < 0xC0000000)
+        && e.phys_odd < 0x20000000
+    {
         let mut i = e.start_odd;
         while i < e.end_odd {
             device.cpu.cop0.tlb_lut_r[(i >> 12) as usize].address =

src/device/vi.rs

@@ -2,20 +2,20 @@ use crate::device;
 use crate::ui;
 use governor::clock::Clock;
 
-pub const VI_STATUS_REG: u32 = 0;
-//pub const VI_ORIGIN_REG: u32 = 1;
-//pub const VI_WIDTH_REG: u32 = 2;
-//pub const VI_V_INTR_REG: u32 = 3;
-pub const VI_CURRENT_REG: u32 = 4;
-//pub const VI_BURST_REG: u32 = 5;
-pub const VI_V_SYNC_REG: u32 = 6;
-pub const VI_H_SYNC_REG: u32 = 7;
-//pub const VI_LEAP_REG: u32 = 8;
-//pub const VI_H_START_REG: u32 = 9;
-//pub const VI_V_START_REG: u32 = 10;
-//pub const VI_V_BURST_REG: u32 = 11;
-//pub const VI_X_SCALE_REG: u32 = 12;
-//pub const VI_Y_SCALE_REG: u32 = 13;
+const VI_STATUS_REG: u32 = 0;
+//const VI_ORIGIN_REG: u32 = 1;
+pub const VI_WIDTH_REG: u32 = 2;
+//const VI_V_INTR_REG: u32 = 3;
+const VI_CURRENT_REG: u32 = 4;
+//const VI_BURST_REG: u32 = 5;
+const VI_V_SYNC_REG: u32 = 6;
+const VI_H_SYNC_REG: u32 = 7;
+//const VI_LEAP_REG: u32 = 8;
+pub const VI_H_START_REG: u32 = 9;
+pub const VI_V_START_REG: u32 = 10;
+//const VI_V_BURST_REG: u32 = 11;
+pub const VI_X_SCALE_REG: u32 = 12;
+pub const VI_Y_SCALE_REG: u32 = 13;
 pub const VI_REGS_COUNT: u32 = 14;
 
 pub struct Vi {
@@ -25,10 +25,13 @@ pub struct Vi {
     pub field: u32,
     pub limiter: Option<governor::DefaultDirectRateLimiter>,
     pub count_per_scanline: u64,
+    pub vi_counter: u64,
 }
 
 //static mut FRAME_COUNTER: u64 = 0;
 
+const LIMIT_BUFFER: u64 = 3;
+
 pub fn set_expected_refresh_rate(device: &mut device::Device) {
     let expected_refresh_rate = device.vi.clock as f64
         / (device.vi.regs[VI_V_SYNC_REG as usize] + 1) as f64
@@ -39,7 +42,7 @@ pub fn set_expected_refresh_rate(device: &mut device::Device) {
         device.vi.delay / (device.vi.regs[VI_V_SYNC_REG as usize] + 1) as u64;
 
     let quota = governor::Quota::with_period(std::time::Duration::from_secs_f64(
-        1.0 / expected_refresh_rate,
+        (1.0 / expected_refresh_rate) * LIMIT_BUFFER as f64,
     ))
     .unwrap();
     device.vi.limiter = Some(governor::RateLimiter::direct(quota))
@@ -97,21 +100,19 @@ pub fn write_regs(device: &mut device::Device, address: u64, value: u32, mask: u
                 device::memory::masked_write_32(&mut device.vi.regs[reg as usize], value, mask);
                 set_vertical_interrupt(device);
                 set_expected_refresh_rate(device);
-                ui::video::set_register(reg as u32, device.vi.regs[reg as usize])
             }
         }
         VI_H_SYNC_REG => {
             if device.vi.regs[reg as usize] != value & mask {
                 device::memory::masked_write_32(&mut device.vi.regs[reg as usize], value, mask);
                 set_expected_refresh_rate(device);
-                ui::video::set_register(reg as u32, device.vi.regs[reg as usize])
             }
         }
         _ => {
             device::memory::masked_write_32(&mut device.vi.regs[reg as usize], value, mask);
-            ui::video::set_register(reg as u32, device.vi.regs[reg as usize])
         }
     }
+    ui::video::set_register(reg as u32, device.vi.regs[reg as usize])
 }
 
 pub fn vertical_interrupt_event(device: &mut device::Device) {
@@ -123,7 +124,10 @@ pub fn vertical_interrupt_event(device: &mut device::Device) {
         }
     */
 
-    speed_limiter(device);
+    device.vi.vi_counter += 1;
+    if device.vi.vi_counter % LIMIT_BUFFER == 0 {
+        speed_limiter(device);
+    }
 
     /* toggle vi field if in interlaced mode */
     device.vi.field ^= (device.vi.regs[VI_STATUS_REG as usize] >> 6) & 0x1;
@@ -163,8 +167,8 @@ pub fn speed_limiter(device: &device::Device) {
     if result.is_err() {
         let outcome = result.unwrap_err();
         let dur = outcome.wait_time_from(governor::clock::DefaultClock::default().now());
-        std::thread::sleep(dur);
+        spin_sleep::sleep(dur);
 
-        device.vi.limiter.as_ref().unwrap().check().unwrap();
+        let _ = device.vi.limiter.as_ref().unwrap().check();
     }
 }

src/main.rs

@@ -1,5 +1,3 @@
-#![feature(round_ties_even)]
-#![feature(string_remove_matches)]
 #![cfg_attr(not(debug_assertions), windows_subsystem = "windows")] // hide console window on Windows in release
 
 mod device;
@@ -56,18 +54,47 @@ struct Args {
 }
 
 fn main() {
-    let config_dir = dirs::config_dir().unwrap().join("gopher64");
-    let cache_dir = dirs::cache_dir().unwrap().join("gopher64");
+    let exe_path = std::env::current_exe().unwrap();
+    let portable_dir = exe_path.parent();
+    let portable = portable_dir.unwrap().join("portable.txt").exists();
+    let config_dir;
+    let cache_dir;
+    let data_dir;
+    if portable {
+        config_dir = portable_dir.unwrap().join("portable_data").join("config");
+        cache_dir = portable_dir.unwrap().join("portable_data").join("cache");
+        data_dir = portable_dir.unwrap().join("portable_data").join("data");
+    } else {
+        config_dir = dirs::config_dir().unwrap().join("gopher64");
+        cache_dir = dirs::cache_dir().unwrap().join("gopher64");
+        data_dir = dirs::data_dir().unwrap().join("gopher64");
+    };
 
-    let _ = std::fs::create_dir_all(config_dir.clone());
-    let _ = std::fs::create_dir_all(cache_dir.clone());
-    let _ = std::fs::remove_file(cache_dir.clone().join("game_running"));
+    let mut result = std::fs::create_dir_all(config_dir.clone());
+    if result.is_err() {
+        panic!("could not create config dir: {}", result.err().unwrap())
+    }
+    result = std::fs::create_dir_all(cache_dir.clone());
+    if result.is_err() {
+        panic!("could not create cache dir: {}", result.err().unwrap())
+    }
+    result = std::fs::create_dir_all(data_dir.clone().join("saves"));
+    if result.is_err() {
+        panic!("could not create data dir: {}", result.err().unwrap())
+    }
+    let running_file = cache_dir.join("game_running");
+    if running_file.exists() {
+        result = std::fs::remove_file(running_file);
+        if result.is_err() {
+            panic!("could not remove running file: {}", result.err().unwrap())
+        }
+    }
 
     let args = Args::parse();
     let args_as_strings: Vec<String> = std::env::args().collect();
     let args_count = args_as_strings.len();
     if args_count > 1 {
-        let mut device = device::Device::new();
+        let mut device = device::Device::new(config_dir);
 
         if args.clear_input_bindings {
             ui::input::clear_bindings(&mut device.ui);
@@ -118,17 +145,21 @@ fn main() {
 
         if args.game.is_some() {
             let file_path = std::path::Path::new(args.game.as_ref().unwrap());
-            device::run_game(file_path, &mut device, args.fullscreen);
+            device::run_game(file_path, data_dir, &mut device, args.fullscreen);
         }
     } else {
         let options = eframe::NativeOptions {
-            initial_window_size: Some(eframe::egui::vec2(640.0, 480.0)),
+            viewport: eframe::egui::ViewportBuilder::default().with_inner_size([640.0, 480.0]),
             ..Default::default()
         };
         eframe::run_native(
             "gopher64",
             options,
-            Box::new(|_cc| Box::new(ui::gui::GopherEguiApp::new())),
+            Box::new(|cc| {
+                Ok(Box::new(ui::gui::GopherEguiApp::new(
+                    cc, config_dir, cache_dir, data_dir,
+                )))
+            }),
         )
         .unwrap();
     }

src/ui.rs

@@ -4,6 +4,7 @@ pub mod gui;
 pub mod input;
 pub mod storage;
 pub mod video;
+pub mod vru;
 
 pub struct Ui {
     pub controllers: [input::Controllers; 4],
@@ -18,6 +19,7 @@ pub struct Ui {
     pub video_subsystem: Option<sdl2::VideoSubsystem>,
     pub audio_subsystem: Option<sdl2::AudioSubsystem>,
     pub joystick_subsystem: Option<sdl2::JoystickSubsystem>,
+    #[allow(dead_code)]
     pub controller_subsystem: Option<sdl2::GameControllerSubsystem>,
     pub window: Option<sdl2::video::Window>,
     pub audio_device: Option<sdl2::audio::AudioQueue<i16>>,
@@ -25,7 +27,6 @@ pub struct Ui {
 
 impl Drop for Ui {
     fn drop(&mut self) {
-        storage::write_saves(self);
         write_config(self);
     }
 }
@@ -36,17 +37,14 @@ fn write_config(ui: &Ui) {
 }
 
 impl Ui {
-    pub fn new() -> Ui {
+    pub fn new(config_dir: std::path::PathBuf) -> Ui {
         let sdl_context = sdl2::init().unwrap();
         let video_subsystem = sdl_context.video().unwrap();
         let audio_subsystem = sdl_context.audio().unwrap();
         let joystick_subsystem = sdl_context.joystick().unwrap();
         let controller_subsystem = sdl_context.game_controller().unwrap();
 
-        let config_file_path = dirs::config_dir()
-            .unwrap()
-            .join("gopher64")
-            .join("config.json");
+        let config_file_path = config_dir.join("config.json");
         let config_file = std::fs::read(config_file_path.clone());
         let mut config_map = config::Config::new();
         if config_file.is_ok() {

src/ui/audio.rs

@@ -24,11 +24,11 @@ pub fn play_audio(device: &mut device::Device, dram_addr: usize, length: u64) {
     while i < length as usize / 2 {
         // Left channel
         primary_buffer[i] = device.rdram.mem[dram_addr + (i * 2) + 2] as i16
-            | (device.rdram.mem[dram_addr + (i * 2) + 3] as i16) << 8;
+            | ((device.rdram.mem[dram_addr + (i * 2) + 3] as i16) << 8);
 
         // Right channel
         primary_buffer[i + 1] = device.rdram.mem[dram_addr + (i * 2)] as i16
-            | (device.rdram.mem[dram_addr + (i * 2) + 1] as i16) << 8;
+            | ((device.rdram.mem[dram_addr + (i * 2) + 1] as i16) << 8);
         i += 2;
     }
 

src/ui/config.rs

@@ -6,7 +6,7 @@ pub struct InputProfile {
     pub controller_buttons: [(bool, i32); 14],
     pub controller_axis: [(bool, i32, i16); 18],
     pub joystick_buttons: [(bool, u32); 14],
-    pub joystick_hat: [(bool, u32, i8); 14],
+    pub joystick_hat: [(bool, u32, u8); 14],
     pub joystick_axis: [(bool, u32, i16); 18],
 }
 
@@ -15,10 +15,17 @@ pub struct Input {
     pub input_profiles: std::collections::HashMap<String, InputProfile>,
     pub input_profile_binding: [String; 4],
     pub controller_assignment: [Option<String>; 4],
+    pub controller_enabled: [bool; 4],
+    pub emulate_vru: bool,
+}
+#[derive(serde::Serialize, serde::Deserialize)]
+pub struct Video {
+    pub upscale: bool,
 }
 #[derive(serde::Serialize, serde::Deserialize)]
 pub struct Config {
     pub input: Input,
+    pub video: Video,
 }
 
 impl Config {
@@ -35,7 +42,10 @@ impl Config {
                 ],
                 controller_assignment: [None, None, None, None],
                 input_profiles,
+                controller_enabled: [true, false, false, false],
+                emulate_vru: false,
             },
+            video: Video { upscale: false },
         }
     }
 }

src/ui/gui.rs

@@ -3,12 +3,22 @@ use crate::ui;
 use eframe::egui;
 
 pub struct GopherEguiApp {
+    config_dir: std::path::PathBuf,
+    cache_dir: std::path::PathBuf,
+    data_dir: std::path::PathBuf,
     configure_profile: bool,
     profile_name: String,
     controllers: Vec<String>,
     selected_controller: [i32; 4],
     selected_profile: [String; 4],
     input_profiles: Vec<String>,
+    controller_enabled: [bool; 4],
+    upscale: bool,
+    emulate_vru: bool,
+    show_vru_dialog: bool,
+    vru_window_receiver: Option<std::sync::mpsc::Receiver<Vec<String>>>,
+    vru_word_notifier: Option<std::sync::mpsc::Sender<String>>,
+    vru_word_list: Vec<String>,
 }
 
 fn get_input_profiles(game_ui: &ui::Ui) -> Vec<String> {
@@ -31,8 +41,14 @@ fn get_controllers(game_ui: &ui::Ui) -> Vec<String> {
 }
 
 impl GopherEguiApp {
-    pub fn new() -> GopherEguiApp {
-        let game_ui = ui::Ui::new();
+    pub fn new(
+        cc: &eframe::CreationContext<'_>,
+        config_dir: std::path::PathBuf,
+        cache_dir: std::path::PathBuf,
+        data_dir: std::path::PathBuf,
+    ) -> GopherEguiApp {
+        add_japanese_font(&cc.egui_ctx);
+        let game_ui = ui::Ui::new(config_dir.clone());
         let joystick_subsystem = game_ui.joystick_subsystem.as_ref().unwrap();
         let num_joysticks = joystick_subsystem.num_joysticks().unwrap();
         let mut guids: Vec<String> = vec![];
@@ -57,17 +73,34 @@ impl GopherEguiApp {
             }
         }
         GopherEguiApp {
+            cache_dir: cache_dir.clone(),
+            config_dir: config_dir.clone(),
+            data_dir: data_dir.clone(),
             configure_profile: false,
             profile_name: "".to_string(),
             selected_profile: game_ui.config.input.input_profile_binding.clone(),
             selected_controller,
             controllers: get_controllers(&game_ui),
             input_profiles: get_input_profiles(&game_ui),
+            controller_enabled: game_ui.config.input.controller_enabled,
+            upscale: game_ui.config.video.upscale,
+            emulate_vru: game_ui.config.input.emulate_vru,
+            show_vru_dialog: false,
+            vru_window_receiver: None,
+            vru_word_notifier: None,
+            vru_word_list: Vec::new(),
         }
     }
 }
 
-fn save_config(game_ui: &mut ui::Ui, selected_controller: [i32; 4], selected_profile: [String; 4]) {
+fn save_config(
+    game_ui: &mut ui::Ui,
+    selected_controller: [i32; 4],
+    selected_profile: [String; 4],
+    controller_enabled: [bool; 4],
+    upscale: bool,
+    emulate_vru: bool,
+) {
     let joystick_subsystem = game_ui.joystick_subsystem.as_ref().unwrap();
     for (pos, item) in selected_controller.iter().enumerate() {
         if *item != -1 {
@@ -83,15 +116,22 @@ fn save_config(game_ui: &mut ui::Ui, selected_controller: [i32; 4], selected_pro
     }
 
     game_ui.config.input.input_profile_binding = selected_profile;
+    game_ui.config.input.controller_enabled = controller_enabled;
+
+    game_ui.config.video.upscale = upscale;
+    game_ui.config.input.emulate_vru = emulate_vru;
 }
 
 impl Drop for GopherEguiApp {
     fn drop(&mut self) {
-        let mut game_ui = ui::Ui::new();
+        let mut game_ui = ui::Ui::new(self.config_dir.clone());
         save_config(
             &mut game_ui,
             self.selected_controller,
             self.selected_profile.clone(),
+            self.controller_enabled,
+            self.upscale,
+            self.emulate_vru,
         );
     }
 }
@@ -110,12 +150,15 @@ impl eframe::App for GopherEguiApp {
                     ui.horizontal(|ui| {
                         if ui.button("Configure Profile").clicked() {
                             let profile_name = self.profile_name.clone();
+                            let config_dir = self.config_dir.clone();
                             execute(async {
-                                let mut game_ui = ui::Ui::new();
+                                let mut game_ui = ui::Ui::new(config_dir);
                                 ui::input::configure_input_profile(&mut game_ui, profile_name);
                             });
                             self.configure_profile = false;
-                            if !self.profile_name.is_empty() && self.profile_name != "default" {
+                            if !self.profile_name.is_empty()
+                                && !self.input_profiles.contains(&self.profile_name)
+                            {
                                 self.input_profiles.push(self.profile_name.clone())
                             }
                         };
@@ -127,39 +170,80 @@ impl eframe::App for GopherEguiApp {
         }
 
         egui::CentralPanel::default().show(ctx, |ui| {
-            ui.set_enabled(!self.configure_profile);
+            if self.configure_profile {
+                ui.disable()
+            }
 
             if ui.button("Open ROM").clicked() {
                 // Spawn dialog on main thread
                 let task = rfd::AsyncFileDialog::new().pick_file();
                 let selected_controller = self.selected_controller;
                 let selected_profile = self.selected_profile.clone();
+                let controller_enabled = self.controller_enabled;
+                let upscale = self.upscale;
+                let emulate_vru = self.emulate_vru;
+                let config_dir = self.config_dir.clone();
+                let cache_dir = self.cache_dir.clone();
+                let data_dir = self.data_dir.clone();
+
+                let (vru_window_notifier, vru_window_receiver): (
+                    std::sync::mpsc::Sender<Vec<String>>,
+                    std::sync::mpsc::Receiver<Vec<String>>,
+                ) = std::sync::mpsc::channel();
+
+                let (vru_word_notifier, vru_word_receiver): (
+                    std::sync::mpsc::Sender<String>,
+                    std::sync::mpsc::Receiver<String>,
+                ) = std::sync::mpsc::channel();
+
+                if emulate_vru {
+                    self.vru_window_receiver = Some(vru_window_receiver);
+                    self.vru_word_notifier = Some(vru_word_notifier);
+                }
+
+                let gui_ctx = ctx.clone();
                 execute(async move {
                     let file = task.await;
 
                     if let Some(file) = file {
-                        let running_file = dirs::cache_dir()
-                            .unwrap()
-                            .join("gopher64")
-                            .join("game_running");
+                        let running_file = cache_dir.join("game_running");
                         if running_file.exists() {
                             return;
                         }
-                        let _ = std::fs::File::create(running_file.clone());
-                        let mut device = device::Device::new();
-                        save_config(&mut device.ui, selected_controller, selected_profile);
-                        device::run_game(std::path::Path::new(file.path()), &mut device, false);
-                        let _ = std::fs::remove_file(running_file.clone());
+                        let result = std::fs::File::create(running_file.clone());
+                        if result.is_err() {
+                            panic!("could not create running file: {}", result.err().unwrap())
+                        }
+                        let mut device = device::Device::new(config_dir);
+                        save_config(
+                            &mut device.ui,
+                            selected_controller,
+                            selected_profile,
+                            controller_enabled,
+                            upscale,
+                            emulate_vru,
+                        );
+                        if emulate_vru {
+                            device.vru.window_notifier = Some(vru_window_notifier);
+                            device.vru.word_receiver = Some(vru_word_receiver);
+                            device.vru.gui_ctx = Some(gui_ctx);
+                        }
+                        device::run_game(
+                            std::path::Path::new(file.path()),
+                            data_dir,
+                            &mut device,
+                            false,
+                        );
+                        let result = std::fs::remove_file(running_file.clone());
+                        if result.is_err() {
+                            panic!("could not remove running file: {}", result.err().unwrap())
+                        }
                     }
                 });
             }
 
             if ui.button("Configure Input Profile").clicked() {
-                let running_file = dirs::cache_dir()
-                    .unwrap()
-                    .join("gopher64")
-                    .join("game_running");
-                if running_file.exists() {
+                if self.cache_dir.join("game_running").exists() {
                     return;
                 }
                 self.configure_profile = true;
@@ -167,15 +251,17 @@ impl eframe::App for GopherEguiApp {
 
             ui.add_space(32.0);
             ui.label("Controller Config:");
-            egui::Grid::new("some_unique_id").show(ui, |ui| {
+            egui::Grid::new("controller_config").show(ui, |ui| {
                 ui.label("Port");
+                ui.label("Enabled");
                 ui.label("Profile");
                 ui.label("Controller");
                 ui.end_row();
                 for i in 0..4 {
                     ui.label(format!("{}", i + 1));
+                    ui.checkbox(&mut self.controller_enabled[i], "");
 
-                    egui::ComboBox::from_id_source(format!("profile-combo-{}", i))
+                    egui::ComboBox::from_id_salt(format!("profile-combo-{}", i))
                         .selected_text(self.selected_profile[i].clone())
                         .show_ui(ui, |ui| {
                             for j in 0..self.input_profiles.len() {
@@ -192,7 +278,7 @@ impl eframe::App for GopherEguiApp {
                     } else {
                         self.controllers[self.selected_controller[i] as usize].clone()
                     };
-                    egui::ComboBox::from_id_source(format!("controller-combo-{}", i))
+                    egui::ComboBox::from_id_salt(format!("controller-combo-{}", i))
                         .selected_text(controller_text)
                         .show_ui(ui, |ui| {
                             ui.selectable_value(
@@ -211,10 +297,84 @@ impl eframe::App for GopherEguiApp {
                     ui.end_row();
                 }
             });
+            ui.add_space(32.0);
+            ui.checkbox(&mut self.upscale, "High-Res Graphics");
+            ui.checkbox(
+                &mut self.emulate_vru,
+                "Emulate VRU (connects VRU to controller port 4)",
+            );
+            ui.add_space(32.0);
+            ui.label(format!("Version: {}", env!("CARGO_PKG_VERSION")));
         });
+
+        if self.emulate_vru && self.vru_window_receiver.is_some() {
+            let result = self.vru_window_receiver.as_ref().unwrap().try_recv();
+            if result.is_ok() {
+                self.show_vru_dialog = true;
+                self.vru_word_list = result.unwrap();
+            }
+        }
+
+        if self.show_vru_dialog {
+            ctx.show_viewport_immediate(
+                egui::ViewportId::from_hash_of("vru_dialog"),
+                egui::ViewportBuilder::default()
+                    .with_title("What would you like to say?")
+                    .with_always_on_top(),
+                |ctx, class| {
+                    assert!(
+                        class == egui::ViewportClass::Immediate,
+                        "This egui backend doesn't support multiple viewports"
+                    );
+                    egui::CentralPanel::default().show(ctx, |ui| {
+                        egui::Grid::new("vru_words").show(ui, |ui| {
+                            for (i, v) in self.vru_word_list.iter().enumerate() {
+                                if i % 5 == 0 {
+                                    ui.end_row();
+                                }
+                                if ui.button((*v).to_string()).clicked() {
+                                    self.vru_word_notifier
+                                        .as_ref()
+                                        .unwrap()
+                                        .send(v.clone())
+                                        .unwrap();
+                                    self.show_vru_dialog = false;
+                                }
+                            }
+                        });
+                    });
+
+                    if ctx.input(|i| i.viewport().close_requested()) {
+                        self.vru_word_notifier
+                            .as_ref()
+                            .unwrap()
+                            .send(String::from(""))
+                            .unwrap();
+                        self.show_vru_dialog = false;
+                    }
+                },
+            );
+        }
     }
 }
 
 fn execute<F: std::future::Future<Output = ()> + Send + 'static>(f: F) {
     std::thread::spawn(move || futures::executor::block_on(f));
 }
+
+fn add_japanese_font(ctx: &egui::Context) {
+    ctx.add_font(eframe::epaint::text::FontInsert::new(
+        "japanese_font",
+        egui::FontData::from_static(include_bytes!("../../data/NotoSansJP-Regular.ttf")),
+        vec![
+            eframe::epaint::text::InsertFontFamily {
+                family: egui::FontFamily::Proportional,
+                priority: egui::epaint::text::FontPriority::Lowest,
+            },
+            eframe::epaint::text::InsertFontFamily {
+                family: egui::FontFamily::Monospace,
+                priority: egui::epaint::text::FontPriority::Lowest,
+            },
+        ],
+    ));
+}

src/ui/input.rs

@@ -2,29 +2,29 @@ use std::ops::Neg;
 
 use crate::ui;
 
-pub const R_DPAD: usize = 0;
-pub const L_DPAD: usize = 1;
-pub const D_DPAD: usize = 2;
-pub const U_DPAD: usize = 3;
-pub const START_BUTTON: usize = 4;
-pub const Z_TRIG: usize = 5;
-pub const B_BUTTON: usize = 6;
-pub const A_BUTTON: usize = 7;
-pub const R_CBUTTON: usize = 8;
-pub const L_CBUTTON: usize = 9;
-pub const D_CBUTTON: usize = 10;
-pub const U_CBUTTON: usize = 11;
-pub const R_TRIG: usize = 12;
-pub const L_TRIG: usize = 13;
-pub const X_AXIS: usize = 16;
-pub const Y_AXIS: usize = 24;
+const R_DPAD: usize = 0;
+const L_DPAD: usize = 1;
+const D_DPAD: usize = 2;
+const U_DPAD: usize = 3;
+const START_BUTTON: usize = 4;
+const Z_TRIG: usize = 5;
+const B_BUTTON: usize = 6;
+const A_BUTTON: usize = 7;
+const R_CBUTTON: usize = 8;
+const L_CBUTTON: usize = 9;
+const D_CBUTTON: usize = 10;
+const U_CBUTTON: usize = 11;
+const R_TRIG: usize = 12;
+const L_TRIG: usize = 13;
+const X_AXIS: usize = 16;
+const Y_AXIS: usize = 24;
 
-pub const AXIS_LEFT: usize = 14;
-pub const AXIS_RIGHT: usize = 15;
-pub const AXIS_UP: usize = 16;
-pub const AXIS_DOWN: usize = 17;
+const AXIS_LEFT: usize = 14;
+const AXIS_RIGHT: usize = 15;
+const AXIS_UP: usize = 16;
+const AXIS_DOWN: usize = 17;
 
-pub const MAX_AXIS_VALUE: f64 = 85.0;
+const MAX_AXIS_VALUE: f64 = 85.0;
 
 pub struct Controllers {
     pub game_controller: Option<sdl2::controller::GameController>,
@@ -78,74 +78,110 @@ pub fn set_axis_from_joystick(
     (x, y)
 }
 
+fn get_axis_from_i32(axis: i32) -> sdl2::controller::Axis {
+    match axis {
+        0 => sdl2::controller::Axis::LeftX,
+        1 => sdl2::controller::Axis::LeftY,
+        2 => sdl2::controller::Axis::RightX,
+        3 => sdl2::controller::Axis::RightY,
+        4 => sdl2::controller::Axis::TriggerLeft,
+        5 => sdl2::controller::Axis::TriggerRight,
+        _ => panic!("Invalid axis"),
+    }
+}
+
+fn get_button_from_i32(button: i32) -> sdl2::controller::Button {
+    match button {
+        0 => sdl2::controller::Button::A,
+        1 => sdl2::controller::Button::B,
+        2 => sdl2::controller::Button::X,
+        3 => sdl2::controller::Button::Y,
+        4 => sdl2::controller::Button::Back,
+        5 => sdl2::controller::Button::Guide,
+        6 => sdl2::controller::Button::Start,
+        7 => sdl2::controller::Button::LeftStick,
+        8 => sdl2::controller::Button::RightStick,
+        9 => sdl2::controller::Button::LeftShoulder,
+        10 => sdl2::controller::Button::RightShoulder,
+        11 => sdl2::controller::Button::DPadUp,
+        12 => sdl2::controller::Button::DPadDown,
+        13 => sdl2::controller::Button::DPadLeft,
+        14 => sdl2::controller::Button::DPadRight,
+        _ => panic!("Invalid button"),
+    }
+}
+
 pub fn set_axis_from_controller(
     profile: &ui::config::InputProfile,
     controller: &sdl2::controller::GameController,
 ) -> (f64, f64) {
-    unsafe {
-        let mut x = 0.0;
-        let mut y = 0.0;
-        if profile.controller_axis[AXIS_LEFT].0 {
-            let axis_position =
-                controller.axis(std::mem::transmute(profile.controller_axis[AXIS_LEFT].1));
-            if axis_position as isize * profile.controller_axis[AXIS_LEFT].2 as isize > 0 {
-                x = axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64;
-            }
+    let mut x = 0.0;
+    let mut y = 0.0;
+    if profile.controller_axis[AXIS_LEFT].0 {
+        let axis_position =
+            controller.axis(get_axis_from_i32(profile.controller_axis[AXIS_LEFT].1));
+        if axis_position as isize * profile.controller_axis[AXIS_LEFT].2 as isize > 0 {
+            x = axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64;
         }
-        if profile.controller_axis[AXIS_RIGHT].0 {
-            let axis_position =
-                controller.axis(std::mem::transmute(profile.controller_axis[AXIS_RIGHT].1));
-            if axis_position as isize * profile.controller_axis[AXIS_RIGHT].2 as isize > 0 {
-                x = axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64;
-            }
-        }
-        if profile.controller_axis[AXIS_DOWN].0 {
-            let axis_position =
-                controller.axis(std::mem::transmute(profile.controller_axis[AXIS_DOWN].1));
-            if axis_position as isize * profile.controller_axis[AXIS_DOWN].2 as isize > 0 {
-                y = (axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64).neg();
-            }
-        }
-        if profile.controller_axis[AXIS_UP].0 {
-            let axis_position =
-                controller.axis(std::mem::transmute(profile.controller_axis[AXIS_UP].1));
-            if axis_position as isize * profile.controller_axis[AXIS_UP].2 as isize > 0 {
-                y = (axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64).neg();
-            }
-        }
-        (x, y)
     }
+    if profile.controller_axis[AXIS_RIGHT].0 {
+        let axis_position =
+            controller.axis(get_axis_from_i32(profile.controller_axis[AXIS_RIGHT].1));
+        if axis_position as isize * profile.controller_axis[AXIS_RIGHT].2 as isize > 0 {
+            x = axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64;
+        }
+    }
+    if profile.controller_axis[AXIS_DOWN].0 {
+        let axis_position =
+            controller.axis(get_axis_from_i32(profile.controller_axis[AXIS_DOWN].1));
+        if axis_position as isize * profile.controller_axis[AXIS_DOWN].2 as isize > 0 {
+            y = (axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64).neg();
+        }
+    }
+    if profile.controller_axis[AXIS_UP].0 {
+        let axis_position = controller.axis(get_axis_from_i32(profile.controller_axis[AXIS_UP].1));
+        if axis_position as isize * profile.controller_axis[AXIS_UP].2 as isize > 0 {
+            y = (axis_position as f64 * MAX_AXIS_VALUE / i16::MAX as f64).neg();
+        }
+    }
+    (x, y)
 }
 
 pub fn set_axis_from_keys(
     profile: &ui::config::InputProfile,
     keyboard_state: sdl2::keyboard::KeyboardState,
 ) -> (f64, f64) {
-    unsafe {
-        let mut x = 0.0;
-        let mut y = 0.0;
-        if profile.keys[AXIS_LEFT].0
-            && keyboard_state.is_scancode_pressed(std::mem::transmute(profile.keys[AXIS_LEFT].1))
-        {
-            x = -MAX_AXIS_VALUE
-        }
-        if profile.keys[AXIS_RIGHT].0
-            && keyboard_state.is_scancode_pressed(std::mem::transmute(profile.keys[AXIS_RIGHT].1))
-        {
-            x = MAX_AXIS_VALUE
-        }
-        if profile.keys[AXIS_DOWN].0
-            && keyboard_state.is_scancode_pressed(std::mem::transmute(profile.keys[AXIS_DOWN].1))
-        {
-            y = -MAX_AXIS_VALUE
-        }
-        if profile.keys[AXIS_UP].0
-            && keyboard_state.is_scancode_pressed(std::mem::transmute(profile.keys[AXIS_UP].1))
-        {
-            y = MAX_AXIS_VALUE
-        }
-        (x, y)
+    let mut x = 0.0;
+    let mut y = 0.0;
+    if profile.keys[AXIS_LEFT].0
+        && keyboard_state.is_scancode_pressed(
+            sdl2::keyboard::Scancode::from_i32(profile.keys[AXIS_LEFT].1).unwrap(),
+        )
+    {
+        x = -MAX_AXIS_VALUE
     }
+    if profile.keys[AXIS_RIGHT].0
+        && keyboard_state.is_scancode_pressed(
+            sdl2::keyboard::Scancode::from_i32(profile.keys[AXIS_RIGHT].1).unwrap(),
+        )
+    {
+        x = MAX_AXIS_VALUE
+    }
+    if profile.keys[AXIS_DOWN].0
+        && keyboard_state.is_scancode_pressed(
+            sdl2::keyboard::Scancode::from_i32(profile.keys[AXIS_DOWN].1).unwrap(),
+        )
+    {
+        y = -MAX_AXIS_VALUE
+    }
+    if profile.keys[AXIS_UP].0
+        && keyboard_state.is_scancode_pressed(
+            sdl2::keyboard::Scancode::from_i32(profile.keys[AXIS_UP].1).unwrap(),
+        )
+    {
+        y = MAX_AXIS_VALUE
+    }
+    (x, y)
 }
 
 pub fn set_buttons_from_joystick(
@@ -162,7 +198,7 @@ pub fn set_buttons_from_joystick(
     let profile_joystick_hat = profile.joystick_hat[i];
     if profile_joystick_hat.0
         && joystick.hat(profile_joystick_hat.1).unwrap()
-            == unsafe { std::mem::transmute(profile_joystick_hat.2) }
+            == sdl2::joystick::HatState::from_raw(profile_joystick_hat.2)
     {
         *keys |= 1 << i;
     }
@@ -186,16 +222,12 @@ pub fn set_buttons_from_controller(
 ) {
     let profile_controller_button = profile.controller_buttons[i];
     if profile_controller_button.0 {
-        unsafe {
-            *keys |=
-                (controller.button(std::mem::transmute(profile_controller_button.1)) as u32) << i;
-        }
+        *keys |= (controller.button(get_button_from_i32(profile_controller_button.1)) as u32) << i;
     }
 
     let profile_controller_axis = profile.controller_axis[i];
     if profile_controller_axis.0 {
-        let axis_position =
-            unsafe { controller.axis(std::mem::transmute(profile_controller_axis.1)) };
+        let axis_position = controller.axis(get_axis_from_i32(profile_controller_axis.1));
         if axis_position as isize * profile_controller_axis.2 as isize > 0
             && axis_position.saturating_abs() > i16::MAX / 2
         {
@@ -218,11 +250,10 @@ pub fn get(ui: &mut ui::Ui, channel: usize) -> u32 {
         if profile_name != "default" || channel == 0 {
             let profile_key = profile.keys[i];
             if profile_key.0 {
-                unsafe {
-                    keys |= (keyboard_state.is_scancode_pressed(std::mem::transmute(profile_key.1))
-                        as u32)
-                        << i;
-                }
+                keys |= (keyboard_state
+                    .is_scancode_pressed(sdl2::keyboard::Scancode::from_i32(profile_key.1).unwrap())
+                    as u32)
+                    << i;
             }
         }
 
@@ -377,8 +408,11 @@ pub fn configure_input_profile(ui: &mut ui::Ui, profile: String) {
                         key_set = true
                     }
                     sdl2::event::Event::JoyHatMotion { hat_idx, state, .. } => {
-                        new_joystick_hat[*value] =
-                            unsafe { (true, hat_idx as u32, std::mem::transmute(state)) };
+                        new_joystick_hat[*value] = (
+                            true,
+                            hat_idx as u32,
+                            sdl2::joystick::HatState::to_raw(state),
+                        );
                         key_set = true
                     }
                     sdl2::event::Event::JoyAxisMotion {
@@ -507,7 +541,10 @@ pub fn init(ui: &mut ui::Ui) {
                         .unwrap()
                         .open(joystick_index);
                     if joystick_result.is_err() {
-                        println!("could not connect joystick")
+                        println!(
+                            "could not connect joystick: {}",
+                            joystick_result.err().unwrap()
+                        )
                     } else {
                         ui.controllers[i].joystick = Some(joystick_result.unwrap());
                     }

src/ui/storage.rs

@@ -86,38 +86,33 @@ pub fn get_save_type(game_id: &str) -> Vec<SaveTypes> {
     }
 }
 
-pub fn init(ui: &mut ui::Ui) {
+pub fn init(ui: &mut ui::Ui, data_dir: std::path::PathBuf) {
     let id = ui.game_id.as_str();
     ui.save_type = get_save_type(id);
 
-    let base_path = dirs::data_dir().unwrap().join("gopher64").join("saves");
+    let base_path = data_dir.join("saves");
 
-    let result = std::fs::create_dir_all(base_path.clone());
-    if result.is_err() {
-        panic!("could not create save dir")
-    }
-
-    ui.paths.eep_file_path = base_path.clone();
+    ui.paths.eep_file_path.clone_from(&base_path);
     ui.paths
         .eep_file_path
         .push(ui.game_id.to_owned() + "-" + &ui.game_hash + ".eep");
 
-    ui.paths.sra_file_path = base_path.clone();
+    ui.paths.sra_file_path.clone_from(&base_path);
     ui.paths
         .sra_file_path
         .push(ui.game_id.to_owned() + "-" + &ui.game_hash + ".sra");
 
-    ui.paths.fla_file_path = base_path.clone();
+    ui.paths.fla_file_path.clone_from(&base_path);
     ui.paths
         .fla_file_path
         .push(ui.game_id.to_owned() + "-" + &ui.game_hash + ".fla");
 
-    ui.paths.pak_file_path = base_path.clone();
+    ui.paths.pak_file_path.clone_from(&base_path);
     ui.paths
         .pak_file_path
         .push(ui.game_id.to_owned() + "-" + &ui.game_hash + ".mpk");
 
-    ui.paths.romsave_file_path = base_path.clone();
+    ui.paths.romsave_file_path.clone_from(&base_path);
     ui.paths
         .romsave_file_path
         .push(ui.game_id.to_owned() + "-" + &ui.game_hash + ".romsave");
@@ -148,7 +143,7 @@ pub fn load_saves(ui: &mut ui::Ui) {
 
 pub fn write_rom_save(ui: &ui::Ui) {
     let f = std::fs::File::create(ui.paths.romsave_file_path.clone()).unwrap();
-    serde_json::to_writer(f, &ui.saves.romsave).unwrap();
+    serde_json::to_writer(f, &ui.saves.romsave.0).unwrap();
 }
 
 pub fn write_saves(ui: &ui::Ui) {

src/ui/video.rs

@@ -1,36 +1,57 @@
-use crate::ui;
+#![allow(non_snake_case)]
+include!(concat!(env!("OUT_DIR"), "/parallel_bindings.rs"));
+use crate::device;
 
-extern "C" {
-    pub fn vk_init(rdram_ptr: usize, rdram_size: u32, fullscreen: u8);
-    pub fn set_sdl_window(window: usize);
-    pub fn rdp_update_screen() -> u8;
-    pub fn rdp_set_vi_register(reg: u32, value: u32);
-    pub fn rdp_process_commands(dpc_regs: &mut [u32; 8], SP_DMEM: &mut [u8; 8192]) -> u64;
-    pub fn full_sync();
-}
-
-pub fn init(ui: &mut ui::Ui, rdram_ptr: *mut u8, rdram_size: usize, fullscreen: bool) {
-    let mut builder = ui
+pub fn init(device: &mut device::Device, fullscreen: bool) {
+    let mut builder = device
+        .ui
         .video_subsystem
         .as_ref()
         .unwrap()
         .window("gopher64", 640, 480);
+
     builder.position_centered().vulkan();
+
     if fullscreen {
         builder.fullscreen_desktop();
     } else {
         builder.resizable();
     }
-    ui.window = Some(builder.build().unwrap());
+    device.ui.window = Some(builder.build().unwrap());
+
+    let gfx_info = GFX_INFO {
+        RDRAM: device.rdram.mem.as_mut_ptr(),
+        DMEM: device.rsp.mem.as_mut_ptr(),
+        RDRAM_SIZE: device.rdram.size,
+        DPC_CURRENT_REG: &mut device.rdp.regs_dpc[device::rdp::DPC_CURRENT_REG as usize],
+        DPC_START_REG: &mut device.rdp.regs_dpc[device::rdp::DPC_START_REG as usize],
+        DPC_END_REG: &mut device.rdp.regs_dpc[device::rdp::DPC_END_REG as usize],
+        DPC_STATUS_REG: &mut device.rdp.regs_dpc[device::rdp::DPC_STATUS_REG as usize],
+        VI_H_START_REG: &mut device.vi.regs[device::vi::VI_H_START_REG as usize],
+        VI_V_START_REG: &mut device.vi.regs[device::vi::VI_V_START_REG as usize],
+        VI_X_SCALE_REG: &mut device.vi.regs[device::vi::VI_X_SCALE_REG as usize],
+        VI_Y_SCALE_REG: &mut device.vi.regs[device::vi::VI_Y_SCALE_REG as usize],
+        VI_WIDTH_REG: &mut device.vi.regs[device::vi::VI_WIDTH_REG as usize],
+    };
+
     unsafe {
-        set_sdl_window(ui.window.as_mut().unwrap().raw() as usize);
-        vk_init(rdram_ptr as usize, rdram_size as u32, fullscreen as u8)
+        rdp_init(
+            device.ui.window.as_mut().unwrap().raw() as *mut std::ffi::c_void,
+            gfx_info,
+            fullscreen,
+            device.ui.config.video.upscale,
+        )
     }
 }
 
-pub fn update_screen() -> u8 {
-    // when the window is closed, running is set to 0
-    
+pub fn close() {
+    unsafe {
+        rdp_close();
+    }
+}
+
+pub fn update_screen() -> bool {
+    // when the window is closed, running is set to false
     unsafe { rdp_update_screen() }
 }
 
@@ -40,14 +61,13 @@ pub fn set_register(reg: u32, value: u32) {
     }
 }
 
-pub fn process_rdp_list(dpc_regs: &mut [u32; 8], sp_dmem: &mut [u8; 8192]) -> u64 {
-    
-    unsafe { rdp_process_commands(dpc_regs, sp_dmem) }
+pub fn process_rdp_list() -> u64 {
+    unsafe { rdp_process_commands() }
 }
 
-pub fn rdp_full_sync() {
+pub fn full_sync() {
     unsafe {
-        full_sync();
+        rdp_full_sync();
     }
 }
 

src/ui/vru.rs

@@ -0,0 +1,28 @@
+use eframe::egui;
+
+use crate::ui;
+
+pub fn prompt_for_match(
+    words: &[String],
+    window_notifier: &std::sync::mpsc::Sender<Vec<String>>,
+    word_index_notifier: &std::sync::mpsc::Receiver<String>,
+    gui_ctx: &egui::Context,
+) -> u16 {
+    let mut dedup_words = words.to_owned();
+    dedup_words.sort();
+    dedup_words.dedup();
+    window_notifier.send(dedup_words).unwrap();
+    gui_ctx.request_repaint(); // this is so the window pops up right away
+    let mut result = word_index_notifier.try_recv();
+    while result.is_err() {
+        result = word_index_notifier.try_recv();
+        std::thread::sleep(std::time::Duration::from_secs_f64(1.0 / 60.0));
+        ui::video::update_screen(); // so the OS doesn't complain about the game window being frozen
+    }
+    for (i, v) in words.iter().enumerate() {
+        if *v == *result.as_ref().unwrap() {
+            return i as u16;
+        }
+    }
+    0x7FFF
+}