Remove loops in RSP code (#136)

* Remove loops in RSP cose

* more

* more

* more

* more

* more

.cargo/config.toml

@@ -1,2 +1,5 @@
 [target.'cfg(target_arch = "x86_64")']
 rustflags = ["-C", "target-cpu=x86-64-v3"]
+
+[target.'cfg(target_arch = "aarch64")']
+rustflags = ["-C", "target-cpu=cortex-a76"]

.github/workflows/build.yml

@@ -12,64 +12,113 @@ env:
 jobs:
   build-linux:
     runs-on: ubuntu-latest
-    env:
-      TARGET: x86_64-unknown-linux-gnu
+    strategy:
+      matrix:
+        target: [x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu]
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
-      - name: Install sdl2 headers
-        run: |
-          sudo apt -y install libsdl2-dev
       - 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: Set aarch64 linker
+        if: matrix.target == 'aarch64-unknown-linux-gnu'
+        run: echo "CARGO_TARGET_AARCH64_UNKNOWN_LINUX_GNU_LINKER=aarch64-linux-gnu-gcc" >> $GITHUB_ENV
       - name: Build
         run: |
-          cargo build --release -v --target=${{ env.TARGET }}
+          cargo build --release -v --target=${{ matrix.target }}
       - name: Upload file
         uses: actions/upload-artifact@v4
         with:
-          name: gopher64-linux
-          path: target/${{ env.TARGET }}/release/gopher64
+          name: gopher64-${{ matrix.target }}
+          path: target/${{ matrix.target }}/release/gopher64
 
   build-windows:
     runs-on: windows-latest
-    env:
-      TARGET: x86_64-pc-windows-msvc
+    strategy:
+      matrix:
+        target: [x86_64-pc-windows-msvc]
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
       - uses: actions-rust-lang/setup-rust-toolchain@v1
         with:
+          target: ${{ matrix.target }}
           rustflags: ""
       - name: Build
         run: |
-          cargo build --release -v --target=${{ env.TARGET }}
+          cargo build --release -v --target=${{ matrix.target }}
       - name: Upload file
         uses: actions/upload-artifact@v4
         with:
-          name: gopher64-windows
-          path: target/${{ env.TARGET }}/release/gopher64.exe
+          name: gopher64-${{ matrix.target }}
+          path: target/${{ matrix.target }}/release/gopher64.exe
 
   build-macos:
     runs-on: macos-latest
-    env:
-      TARGET: x86_64-apple-darwin
+    strategy:
+      matrix:
+        target: [x86_64-apple-darwin, aarch64-apple-darwin]
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
       - uses: actions-rust-lang/setup-rust-toolchain@v1
         with:
-          target: ${{ env.TARGET }}
+          target: ${{ matrix.target }}
           rustflags: ""
       - name: Build
         run: |
-          cargo build --release -v --target=${{ env.TARGET }}
+          cargo build --release -v --target=${{ matrix.target }}
       - name: Upload file
         uses: actions/upload-artifact@v4
         with:
-          name: gopher64-mac
-          path: target/${{ env.TARGET }}/release/gopher64
+          name: gopher64-${{ matrix.target }}
+          path: target/${{ matrix.target }}/release/gopher64

Cargo.toml

@@ -1,12 +1,12 @@
 [package]
 name = "gopher64"
-version = "0.1.14"
+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"
+dirs = "6.0"
 zip = "2.1"
 governor = "0.8"
 sevenz-rust = "0.6"
@@ -14,7 +14,6 @@ chrono = "0.4"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 eframe = "0.30"
-epaint = "0.30"
 rfd = "0.15"
 sha2 = "0.10"
 rusttype = "0.9"
@@ -31,7 +30,8 @@ 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.

build.rs

@@ -1,6 +1,8 @@
 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)
@@ -52,11 +54,15 @@ fn main() {
         .include("parallel-rdp/parallel-rdp-standalone/util")
         .includes(std::env::var("DEP_SDL2_INCLUDE"));
 
-    #[cfg(target_os = "windows")]
-    {
-        #[cfg(target_arch = "x86_64")]
-        {
+    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");
 
@@ -64,14 +70,90 @@ fn main() {
             .set_icon("data/icon.ico")
             .compile()
             .unwrap();
-    }
-    #[cfg(any(target_os = "linux", target_os = "macos"))]
-    {
-        #[cfg(target_arch = "x86_64")]
-        {
+    } 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")
     }
 
     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

@@ -132,6 +132,37 @@ static const unsigned cmd_len_lut[64] = {
 	1,
 };
 
+int sdl_event_filter(void *userdata, SDL_Event *event)
+{
+	if (event->type == SDL_WINDOWEVENT)
+	{
+		switch (event->window.event)
+		{
+		case SDL_WINDOWEVENT_CLOSE:
+			emu_running = false;
+			break;
+		case SDL_WINDOWEVENT_RESIZED:
+			wsi_platform->do_resize();
+			break;
+		default:
+			break;
+		}
+	}
+	else if (fullscreen && event->type == SDL_KEYDOWN)
+	{
+		switch (event->key.keysym.scancode)
+		{
+		case SDL_SCANCODE_ESCAPE:
+			emu_running = false;
+			break;
+		default:
+			break;
+		}
+	}
+
+	return 0;
+}
+
 void rdp_init(void *_window, GFX_INFO _gfx_info, bool _fullscreen, bool _upscale)
 {
 	window = (SDL_Window *)_window;
@@ -198,37 +229,6 @@ void rdp_close()
 	}
 }
 
-int sdl_event_filter(void *userdata, SDL_Event *event)
-{
-	if (event->type == SDL_WINDOWEVENT)
-	{
-		switch (event->window.event)
-		{
-		case SDL_WINDOWEVENT_CLOSE:
-			emu_running = false;
-			break;
-		case SDL_WINDOWEVENT_RESIZED:
-			wsi_platform->do_resize();
-			break;
-		default:
-			break;
-		}
-	}
-	else if (fullscreen && event->type == SDL_KEYDOWN)
-	{
-		switch (event->key.keysym.scancode)
-		{
-		case SDL_SCANCODE_ESCAPE:
-			emu_running = false;
-			break;
-		default:
-			break;
-		}
-	}
-
-	return 0;
-}
-
 static void calculate_viewport(float *x, float *y, float *width, float *height)
 {
 	bool window_widescreen = false;
@@ -264,6 +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.

parallel-rdp/interface.hpp

@@ -29,7 +29,6 @@ extern "C"
 	bool rdp_update_screen();
 	uint64_t rdp_process_commands();
 	void rdp_full_sync();
-	int sdl_event_filter(void *userdata, SDL_Event *event);
 
 #ifdef __cplusplus
 }

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

@@ -1,3 +1,15 @@
+#![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;
@@ -36,7 +48,12 @@ 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
@@ -59,7 +76,7 @@ pub fn run_game(file_path: &std::path::Path, device: &mut Device, fullscreen: bo
     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);
 
@@ -166,7 +183,7 @@ pub struct Device {
 }
 
 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
@@ -175,7 +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 {
@@ -223,7 +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],
@@ -323,21 +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],

src/device/cop1.rs

@@ -17,7 +17,8 @@ use crate::device;
 //const FCR31_CAUSE_INVALID_BIT: u32 = 1 << 16;
 const FCR31_CAUSE_UNIMP_BIT: u32 = 1 << 17;
 pub const FCR31_CMP_BIT: u32 = 1 << 23;
-const FCR31_FS_BIT: u32 = 1 << 24;
+//#[cfg(target_arch = "x86_64")]
+//const FCR31_FS_BIT: u32 = 1 << 24;
 
 const FCR31_CAUSE_MASK: u32 = 0b00000000000000111111000000000000;
 const FCR31_ENABLE_MASK: u32 = 0b00000000000000000000111110000000;
@@ -26,7 +27,8 @@ 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],
@@ -366,8 +368,9 @@ pub fn set_control_registers_fpu(device: &mut device::Device, index: u32, data:
             {
                 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/pi.rs

@@ -13,7 +13,8 @@ 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;
-pub const PI_REGS_COUNT: u32 = 13;
+//const UNKNOWN_REG: u32 = 13; //LibDragon
+pub const PI_REGS_COUNT: u32 = 14;
 
 /* PI_STATUS - read */
 const PI_STATUS_DMA_BUSY: u32 = 1 << 0;

src/device/rdram.rs

@@ -57,10 +57,11 @@ 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) {

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_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) {
@@ -481,8 +596,8 @@ 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,
     );
@@ -508,7 +623,7 @@ pub fn cfc2(device: &mut device::Device, opcode: u32) {
             lo = &mut device.rsp.cpu.vce;
         }
         _ => {
-            panic!("unknown ctc2")
+            panic!("unknown cfc2")
         }
     }
 
@@ -520,16 +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,
         );
     }
 }
@@ -574,7 +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,
@@ -589,10 +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],
     );
 }
 
@@ -601,15 +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;
+        }
     }
 }
 
@@ -618,15 +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;
+        }
     }
 }
 
@@ -635,15 +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;
+        }
     }
 }
 
@@ -652,15 +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;
+        }
     }
 }
 
@@ -670,14 +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;
+        }
     }
 }
 
@@ -687,16 +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,
+            );
     }
 }
 
@@ -706,16 +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,
+            );
     }
 }
 
@@ -725,16 +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,
+            );
     }
 }
 
@@ -746,30 +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;
     }
 }
@@ -784,28 +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;
     }
 }
 
@@ -814,7 +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) {
@@ -822,12 +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;
+        }
     }
 }
 
@@ -836,12 +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;
+        }
     }
 }
 
@@ -850,12 +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;
+        }
     }
 }
 
@@ -864,12 +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;
+        }
     }
 }
 
@@ -881,11 +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;
+        }
     }
 }
 
@@ -898,10 +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;
@@ -917,10 +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;
@@ -934,13 +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;
     }
 }
 
@@ -949,8 +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],
@@ -967,12 +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;
     }
 }
 
@@ -986,7 +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;
     }
@@ -1003,11 +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/vru.rs

@@ -117,21 +117,12 @@ pub fn process(device: &mut device::Device, channel: usize) {
                         for i in 0..length {
                             data.extend(device.vru.word_buffer[offset + i as usize].to_be_bytes());
                         }
-                        if device.cart.rom[0x3E] == /* Japan */ 0x4A {
-                            let (res, _enc, errors) = encoding_rs::SHIFT_JIS.decode(&data);
-                            if errors {
-                                panic!("Failed to decode Japanese word {:X?}", data);
-                            } else {
-                                device.vru.words.push(res.to_string());
-                            }
+
+                        let (res, _enc, errors) = encoding_rs::SHIFT_JIS.decode(&data);
+                        if errors {
+                            panic!("Failed to decode Japanese word {:X?}", data);
                         } else {
-                            let decoded: Result<String, std::string::FromUtf8Error> =
-                                String::from_utf8(data);
-                            if let Ok(result) = decoded {
-                                device.vru.words.push(result)
-                            } else {
-                                panic!("Could not decode VRU word")
-                            }
+                            device.vru.words.push(res.to_string());
                         }
                     } else {
                         offset += 1;

src/main.rs

@@ -54,18 +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);
@@ -116,7 +145,7 @@ 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 {
@@ -126,7 +155,11 @@ fn main() {
         eframe::run_native(
             "gopher64",
             options,
-            Box::new(|cc| Ok(Box::new(ui::gui::GopherEguiApp::new(cc)))),
+            Box::new(|cc| {
+                Ok(Box::new(ui::gui::GopherEguiApp::new(
+                    cc, config_dir, cache_dir, data_dir,
+                )))
+            }),
         )
         .unwrap();
     }

src/ui.rs

@@ -37,17 +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/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],
 }
 

src/ui/gui.rs

@@ -3,6 +3,9 @@ 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>,
@@ -38,10 +41,14 @@ fn get_controllers(game_ui: &ui::Ui) -> Vec<String> {
 }
 
 impl GopherEguiApp {
-    pub fn new(cc: &eframe::CreationContext<'_>) -> GopherEguiApp {
+    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();
+        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![];
@@ -66,6 +73,9 @@ 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(),
@@ -114,7 +124,7 @@ fn save_config(
 
 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,
@@ -140,8 +150,9 @@ 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;
@@ -171,6 +182,9 @@ impl eframe::App for GopherEguiApp {
                 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>>,
@@ -192,15 +206,15 @@ impl eframe::App for GopherEguiApp {
                     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();
+                        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,
@@ -214,18 +228,22 @@ impl eframe::App for GopherEguiApp {
                             device.vru.word_receiver = Some(vru_word_receiver);
                             device.vru.gui_ctx = Some(gui_ctx);
                         }
-                        device::run_game(std::path::Path::new(file.path()), &mut device, false);
-                        let _ = std::fs::remove_file(running_file.clone());
+                        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;
@@ -345,15 +363,15 @@ fn execute<F: std::future::Future<Output = ()> + Send + 'static>(f: F) {
 }
 
 fn add_japanese_font(ctx: &egui::Context) {
-    ctx.add_font(epaint::text::FontInsert::new(
+    ctx.add_font(eframe::epaint::text::FontInsert::new(
         "japanese_font",
         egui::FontData::from_static(include_bytes!("../../data/NotoSansJP-Regular.ttf")),
         vec![
-            epaint::text::InsertFontFamily {
+            eframe::epaint::text::InsertFontFamily {
                 family: egui::FontFamily::Proportional,
                 priority: egui::epaint::text::FontPriority::Lowest,
             },
-            epaint::text::InsertFontFamily {
+            eframe::epaint::text::InsertFontFamily {
                 family: egui::FontFamily::Monospace,
                 priority: egui::epaint::text::FontPriority::Lowest,
             },

src/ui/input.rs

@@ -78,94 +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::<i32, sdl2::controller::Axis>(
-                    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::<i32, sdl2::controller::Axis>(
-                    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::<i32, sdl2::controller::Axis>(
-                    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::<i32, sdl2::controller::Axis>(
-                    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::<i32, sdl2::keyboard::Scancode>(
-                    profile.keys[AXIS_LEFT].1,
-                ))
-        {
-            x = -MAX_AXIS_VALUE
-        }
-        if profile.keys[AXIS_RIGHT].0
-            && keyboard_state
-                .is_scancode_pressed(std::mem::transmute::<i32, sdl2::keyboard::Scancode>(
-                    profile.keys[AXIS_RIGHT].1,
-                ))
-        {
-            x = MAX_AXIS_VALUE
-        }
-        if profile.keys[AXIS_DOWN].0
-            && keyboard_state
-                .is_scancode_pressed(std::mem::transmute::<i32, sdl2::keyboard::Scancode>(
-                    profile.keys[AXIS_DOWN].1,
-                ))
-        {
-            y = -MAX_AXIS_VALUE
-        }
-        if profile.keys[AXIS_UP].0
-            && keyboard_state
-                .is_scancode_pressed(std::mem::transmute::<i32, sdl2::keyboard::Scancode>(
-                    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(
@@ -182,9 +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::<i8, sdl2::joystick::HatState>(profile_joystick_hat.2)
-            }
+            == sdl2::joystick::HatState::from_raw(profile_joystick_hat.2)
     {
         *keys |= 1 << i;
     }
@@ -208,21 +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::<i32, sdl2::controller::Button>(
-                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::<i32, sdl2::controller::Axis>(
-                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
         {
@@ -245,13 +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::<i32, sdl2::keyboard::Scancode>(
-                            profile_key.1,
-                        )) as u32)
-                        << i;
-                }
+                keys |= (keyboard_state
+                    .is_scancode_pressed(sdl2::keyboard::Scancode::from_i32(profile_key.1).unwrap())
+                    as u32)
+                    << i;
             }
         }
 
@@ -406,13 +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::<sdl2::joystick::HatState, i8>(state),
-                            )
-                        };
+                        new_joystick_hat[*value] = (
+                            true,
+                            hat_idx as u32,
+                            sdl2::joystick::HatState::to_raw(state),
+                        );
                         key_set = true
                     }
                     sdl2::event::Event::JoyAxisMotion {
@@ -541,7 +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

@@ -86,16 +86,11 @@ 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 result = std::fs::create_dir_all(base_path.clone());
-    if result.is_err() {
-        panic!("could not create save dir")
-    }
+    let base_path = data_dir.join("saves");
 
     ui.paths.eep_file_path.clone_from(&base_path);
     ui.paths

src/ui/video.rs

@@ -1,30 +1,7 @@
+#![allow(non_snake_case)]
+include!(concat!(env!("OUT_DIR"), "/parallel_bindings.rs"));
 use crate::device;
 
-#[repr(C)]
-pub struct GfxInfo {
-    pub rdram: *mut u8,
-    pub dmem: *mut u8,
-    pub rdram_size: u32,
-    pub dpc_current_reg: *mut u32,
-    pub dpc_start_reg: *mut u32,
-    pub dpc_end_reg: *mut u32,
-    pub dpc_status_reg: *mut u32,
-    pub vi_h_start_reg: *mut u32,
-    pub vi_v_start_reg: *mut u32,
-    pub vi_x_scale_reg: *mut u32,
-    pub vi_y_scale_reg: *mut u32,
-    pub vi_width_reg: *mut u32,
-}
-
-unsafe extern "C" {
-    pub fn rdp_init(window: usize, gfx_info: GfxInfo, fullscreen: bool, upscale: bool);
-    pub fn rdp_close();
-    pub fn rdp_update_screen() -> bool;
-    pub fn rdp_set_vi_register(reg: u32, value: u32);
-    pub fn rdp_process_commands() -> u64;
-    pub fn rdp_full_sync();
-}
-
 pub fn init(device: &mut device::Device, fullscreen: bool) {
     let mut builder = device
         .ui
@@ -42,24 +19,24 @@ pub fn init(device: &mut device::Device, fullscreen: bool) {
     }
     device.ui.window = Some(builder.build().unwrap());
 
-    let gfx_info = GfxInfo {
-        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],
+    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 {
         rdp_init(
-            device.ui.window.as_mut().unwrap().raw() as usize,
+            device.ui.window.as_mut().unwrap().raw() as *mut std::ffi::c_void,
             gfx_info,
             fullscreen,
             device.ui.config.video.upscale,