cargo update

* update to rust 1.90

* more

.cargo/config.toml

@@ -1,4 +1,4 @@
-[target.'cfg(target_os = "linux")']
+[target.'cfg(all(target_arch = "aarch64", target_os = "linux"))']
 rustflags = ["-C", "link-arg=-fuse-ld=lld"]
 
 [target.'cfg(target_os = "windows")']

Cargo.toml

@@ -1,21 +1,21 @@
 [package]
 name = "gopher64"
-version = "1.1.5"
+version = "1.1.6"
 edition = "2024"
-rust-version = "1.89"
+rust-version = "1.90"
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
 dirs = "6.0"
-zip = "4.3"
+zip = "5.1"
 governor = "0.10"
-sevenz-rust2 = { version = "0.17", default-features = false }
+sevenz-rust2 = { version = "0.18", default-features = false }
 chrono = "0.4"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 serde-big-array = "0.5"
-slint = { version = "1.12", default-features = false, features = ["compat-1-2", "std", "backend-winit", "renderer-femtovg"] }
+slint = { version = "1.12", default-features = false, features = ["compat-1-2", "std", "backend-winit", "renderer-femtovg-wgpu"] }
 open = "5.3"
 sha2 = "0.10"
 sdl3-sys = { version = "0.5", features = ["build-from-source-static"] }
@@ -45,3 +45,7 @@ cc = { version = "1.2", features = ["parallel"] }
 [profile.release]
 panic = "abort"
 lto = "thin"
+
+[profile.release-with-debug]
+inherits = "release"
+debug = true

build.rs

@@ -54,6 +54,7 @@ fn main() {
         .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/util/slab_allocator.cpp")
         .file("parallel-rdp/interface.cpp")
         .file("parallel-rdp/wsi_platform.cpp")
         .include("parallel-rdp/parallel-rdp-standalone/parallel-rdp")
@@ -116,6 +117,7 @@ fn main() {
         .allowlist_function("rdp_onscreen_message")
         .allowlist_function("rdp_check_callback")
         .allowlist_function("rdp_new_processor")
+        .allowlist_function("rdp_check_framebuffers")
         .allowlist_function("rdp_state_size")
         .allowlist_function("rdp_save_state")
         .allowlist_function("rdp_load_state")
@@ -163,6 +165,7 @@ fn main() {
             .allowlist_function("_mm_cmplt_epi16")
             .allowlist_function("_mm_subs_epu16")
             .allowlist_function("_mm_set1_epi32")
+            .allowlist_function("_mm_set1_epi16")
             .blocklist_type("__m128i")
             .blocklist_type("int64x2_t")
             .wrap_static_fns(true)

parallel-rdp/interface.cpp

@@ -53,14 +53,26 @@ enum vi_registers
 	VI_REGS_COUNT
 };
 
+typedef struct
+{
+	uint32_t depthbuffer_address;
+	uint32_t framebuffer_address;
+	uint32_t framebuffer_y_offset;
+	uint32_t texture_address;
+	uint32_t framebuffer_pixel_size;
+	uint32_t framebuffer_width;
+	uint32_t texture_pixel_size;
+	uint32_t texture_width;
+	uint32_t framebuffer_height;
+} FrameBufferInfo;
+
 typedef struct
 {
 	uint32_t cmd_data[0x00040000 >> 2];
 	int cmd_cur;
 	int cmd_ptr;
 	uint32_t region;
-	uint64_t vi_counter;
-	uint64_t last_sync;
+	FrameBufferInfo frame_buffer_info;
 } RDP_DEVICE;
 
 static SDL_Window *window;
@@ -75,6 +87,9 @@ static GFX_INFO gfx_info;
 static const uint32_t *fragment_spirv;
 static size_t fragment_size;
 
+std::vector<bool> rdram_dirty;
+uint64_t sync_signal;
+
 static TTF_Font *message_font;
 static std::queue<std::string> messages;
 static uint64_t message_timer;
@@ -240,6 +255,9 @@ void rdp_new_processor(GFX_INFO _gfx_info)
 {
 	gfx_info = _gfx_info;
 
+	sync_signal = 0;
+	rdram_dirty.assign(gfx_info.RDRAM_SIZE >> 3, false);
+
 	if (processor)
 	{
 		delete processor;
@@ -527,7 +545,6 @@ void rdp_render_frame()
 {
 	auto &device = wsi->get_device();
 	render_frame(device);
-	rdp_device.vi_counter += 1;
 }
 
 void rdp_update_screen()
@@ -547,6 +564,27 @@ CALL_BACK rdp_check_callback()
 	return return_value;
 }
 
+void rdp_check_framebuffers(uint32_t address, uint32_t length)
+{
+	if (sync_signal)
+	{
+		address >>= 3;
+		length = (length + 7) >> 3;
+		for (uint32_t i = 0; i < length; i++)
+		{
+			if (address + i >= rdram_dirty.size())
+				return;
+			if (rdram_dirty[address + i])
+			{
+				processor->wait_for_timeline(sync_signal);
+				rdram_dirty.assign(gfx_info.RDRAM_SIZE >> 3, false);
+				sync_signal = 0;
+				return;
+			}
+		}
+	}
+}
+
 size_t rdp_state_size()
 {
 	return sizeof(RDP_DEVICE);
@@ -554,6 +592,7 @@ size_t rdp_state_size()
 
 void rdp_save_state(uint8_t *state)
 {
+	processor->wait_for_timeline(processor->signal_timeline());
 	memcpy(state, &rdp_device, sizeof(RDP_DEVICE));
 }
 
@@ -569,6 +608,24 @@ void rdp_onscreen_message(const char *_message)
 	messages.push(_message);
 }
 
+uint32_t pixel_size(uint32_t pixel_type, uint32_t area)
+{
+	switch (pixel_type)
+	{
+	case 0:
+		return area / 2;
+	case 1:
+		return area;
+	case 2:
+		return area * 2;
+	case 3:
+		return area * 4;
+	default:
+		printf("Invalid pixel size: %u\n", pixel_type);
+		return 0;
+	}
+}
+
 uint64_t rdp_process_commands()
 {
 	uint64_t interrupt_timer = 0;
@@ -632,6 +689,69 @@ uint64_t rdp_process_commands()
 
 		switch (RDP::Op(command))
 		{
+		case RDP::Op::FillZBufferTriangle:
+		case RDP::Op::TextureZBufferTriangle:
+		case RDP::Op::ShadeZBufferTriangle:
+		case RDP::Op::ShadeTextureZBufferTriangle:
+		{
+			uint32_t offset_address = (rdp_device.frame_buffer_info.depthbuffer_address + pixel_size(2, rdp_device.frame_buffer_info.framebuffer_y_offset * rdp_device.frame_buffer_info.framebuffer_width)) >> 3;
+			if (offset_address < rdram_dirty.size() && !rdram_dirty[offset_address])
+			{
+				std::fill_n(rdram_dirty.begin() + offset_address, (pixel_size(2, rdp_device.frame_buffer_info.framebuffer_width * rdp_device.frame_buffer_info.framebuffer_height) + 7) >> 3, true);
+			}
+		}
+		case RDP::Op::FillTriangle:
+		case RDP::Op::TextureTriangle:
+		case RDP::Op::ShadeTriangle:
+		case RDP::Op::ShadeTextureTriangle:
+		case RDP::Op::TextureRectangle:
+		case RDP::Op::TextureRectangleFlip:
+		case RDP::Op::FillRectangle:
+		{
+			uint32_t offset_address = (rdp_device.frame_buffer_info.framebuffer_address + pixel_size(rdp_device.frame_buffer_info.framebuffer_pixel_size, rdp_device.frame_buffer_info.framebuffer_y_offset * rdp_device.frame_buffer_info.framebuffer_width)) >> 3;
+			if (offset_address < rdram_dirty.size() && !rdram_dirty[offset_address])
+			{
+				std::fill_n(rdram_dirty.begin() + offset_address, (pixel_size(rdp_device.frame_buffer_info.framebuffer_pixel_size, rdp_device.frame_buffer_info.framebuffer_width * rdp_device.frame_buffer_info.framebuffer_height) + 7) >> 3, true);
+			}
+		}
+		break;
+		case RDP::Op::LoadTLut:
+		case RDP::Op::LoadTile:
+		{
+			uint32_t upper_left_t = (w1 & 0xFFF) >> 2;
+			uint32_t offset_address = (rdp_device.frame_buffer_info.texture_address + pixel_size(rdp_device.frame_buffer_info.texture_pixel_size, upper_left_t * rdp_device.frame_buffer_info.texture_width)) >> 3;
+			if (offset_address < rdram_dirty.size() && !rdram_dirty[offset_address])
+			{
+				uint32_t lower_right_t = (w2 & 0xFFF) >> 2;
+				std::fill_n(rdram_dirty.begin() + offset_address, (pixel_size(rdp_device.frame_buffer_info.texture_pixel_size, (lower_right_t - upper_left_t) * rdp_device.frame_buffer_info.texture_width) + 7) >> 3, true);
+			}
+		}
+		break;
+		case RDP::Op::LoadBlock:
+		{
+			uint32_t upper_left_s = ((w1 >> 12) & 0xFFF) >> 2;
+			uint32_t upper_left_t = (w1 & 0xFFF) >> 2;
+			uint32_t offset_address = (rdp_device.frame_buffer_info.texture_address + pixel_size(rdp_device.frame_buffer_info.texture_pixel_size, upper_left_s + upper_left_t * rdp_device.frame_buffer_info.texture_width)) >> 3;
+			if (offset_address < rdram_dirty.size() && !rdram_dirty[offset_address])
+			{
+				uint32_t lower_right_s = ((w2 >> 12) & 0xFFF) >> 2;
+				std::fill_n(rdram_dirty.begin() + offset_address, (pixel_size(rdp_device.frame_buffer_info.texture_pixel_size, lower_right_s - upper_left_s) + 7) >> 3, true);
+			}
+		}
+		break;
+		case RDP::Op::SetColorImage:
+			rdp_device.frame_buffer_info.framebuffer_address = (w2 & 0x00FFFFFF);
+			rdp_device.frame_buffer_info.framebuffer_pixel_size = (w1 >> 19) & 0x3;
+			rdp_device.frame_buffer_info.framebuffer_width = (w1 & 0x3FF) + 1;
+			break;
+		case RDP::Op::SetMaskImage:
+			rdp_device.frame_buffer_info.depthbuffer_address = (w2 & 0x00FFFFFF);
+			break;
+		case RDP::Op::SetTextureImage:
+			rdp_device.frame_buffer_info.texture_address = (w2 & 0x00FFFFFF);
+			rdp_device.frame_buffer_info.texture_pixel_size = (w1 >> 19) & 0x3;
+			rdp_device.frame_buffer_info.texture_width = (w1 & 0x3FF) + 1;
+			break;
 		case RDP::Op::SetScissor:
 		{
 			uint32_t upper_left_x = ((w1 >> 12) & 0xFFF) >> 2;
@@ -646,14 +766,14 @@ uint64_t rdp_process_commands()
 			{
 				rdp_device.region = 0;
 			}
-			break;
+
+			rdp_device.frame_buffer_info.framebuffer_y_offset = upper_left_y;
+			rdp_device.frame_buffer_info.framebuffer_height = lower_right_y - upper_left_y;
 		}
+		break;
 		case RDP::Op::SyncFull:
-			if (rdp_device.last_sync != rdp_device.vi_counter)
-			{
-				processor->wait_for_timeline(processor->signal_timeline());
-				rdp_device.last_sync = rdp_device.vi_counter;
-			}
+			sync_signal = processor->signal_timeline();
+
 			interrupt_timer = rdp_device.region;
 			if (interrupt_timer == 0)
 				interrupt_timer = 5000;

parallel-rdp/interface.hpp

@@ -47,6 +47,7 @@ extern "C"
 	uint64_t rdp_process_commands();
 	void rdp_onscreen_message(const char *_message);
 	void rdp_new_processor(GFX_INFO _gfx_info);
+	void rdp_check_framebuffers(uint32_t address, uint32_t length);
 	size_t rdp_state_size();
 	void rdp_save_state(uint8_t *state);
 	void rdp_load_state(const uint8_t *state);

rust-toolchain.toml

@@ -1,2 +1,2 @@
 [toolchain]
-channel = "1.89.0"
+channel = "1.90.0"

src/device/mi.rs

@@ -81,7 +81,8 @@ fn update_init_mode(device: &mut device::Device, w: u32) {
         device.mi.regs[MI_INIT_MODE_REG as usize] &= !MI_INIT_MODE
     }
     if w & MI_SET_INIT != 0 {
-        device.mi.regs[MI_INIT_MODE_REG as usize] |= MI_INIT_MODE
+        panic!("MI_SET_INIT not implemented")
+        //device.mi.regs[MI_INIT_MODE_REG as usize] |= MI_INIT_MODE
     }
     if w & MI_CLR_EBUS != 0 {
         device.mi.regs[MI_INIT_MODE_REG as usize] &= !MI_EBUS_MODE

src/device/pi.rs

@@ -1,4 +1,5 @@
 use crate::device;
+use crate::ui;
 use rand_chacha::rand_core::RngCore;
 
 const PI_DRAM_ADDR_REG: u32 = 0;
@@ -92,6 +93,9 @@ fn dma_write(device: &mut device::Device) {
         length -= dram_addr & 0x7;
     }
 
+    // check RDP before PI writes to RDRAM
+    ui::video::check_framebuffers(dram_addr, length);
+
     let cycles = (handler.write)(device, cart_addr, dram_addr, length);
 
     device::events::create_event(device, device::events::EVENT_TYPE_PI, cycles);

src/device/rdram.rs

@@ -1,4 +1,5 @@
 use crate::device;
+use crate::ui;
 use std::alloc::{Layout, alloc_zeroed};
 
 //const RDRAM_CONFIG_REG: u32 = 0;
@@ -50,6 +51,8 @@ pub fn read_mem(
     );
     let masked_address = address as usize & RDRAM_MASK;
 
+    ui::video::check_framebuffers(masked_address as u32, 4);
+
     u32::from_ne_bytes(
         device
             .rdram
@@ -62,6 +65,8 @@ pub fn read_mem(
 }
 
 pub fn write_mem(device: &mut device::Device, address: u64, value: u32, mask: u32) {
+    ui::video::check_framebuffers(address as u32, 4);
+
     let mut data = u32::from_ne_bytes(
         device
             .rdram

src/device/rsp_interface.rs

@@ -1,4 +1,5 @@
 use crate::device;
+use crate::ui;
 
 const SP_MEM_ADDR_REG: u32 = 0;
 const SP_DRAM_ADDR_REG: u32 = 1;
@@ -146,6 +147,7 @@ fn do_dma(device: &mut device::Device, dma: RspDma) {
     let mut dram_addr = dma.dramaddr & 0xfffff8;
     let offset = dma.memaddr & 0x1000;
 
+    ui::video::check_framebuffers(dram_addr, count * (length + skip) - skip);
     if dma.dir == DmaDir::Read {
         let mut j = 0;
         while j < count {
@@ -219,6 +221,8 @@ fn fifo_push(device: &mut device::Device, dir: DmaDir) {
         panic!("RSP DMA already full")
     }
 
+    device.rsp.cpu.sync_point = true;
+
     if device.rsp.regs[SP_DMA_BUSY_REG as usize] != 0 {
         device.rsp.fifo[1].dir = dir;
         if dir == DmaDir::Read {
@@ -276,21 +280,20 @@ pub fn read_regs(
     }
     let reg = (address & 0xFFFF) >> 2;
     match reg as u32 {
-        SP_DMA_BUSY_REG | SP_DMA_FULL_REG => {
-            if device.rsp.regs[reg as usize] != 0 {
-                device.rsp.cpu.sync_point = true;
-            }
-            device.rsp.regs[reg as usize]
-        }
         SP_STATUS_REG => {
             let value = device.rsp.regs[reg as usize]
-                & !(SP_STATUS_HALT | SP_STATUS_BROKE | SP_STATUS_INTR_BREAK);
+                & (SP_STATUS_SIG0
+                    | SP_STATUS_SIG1
+                    | SP_STATUS_SIG2
+                    | SP_STATUS_SIG3
+                    | SP_STATUS_SIG4
+                    | SP_STATUS_SIG5
+                    | SP_STATUS_SIG6
+                    | SP_STATUS_SIG7);
             if value == device.rsp.last_status_value && value != 0 {
                 device.rsp.cpu.sync_point = true;
             }
-            if value & (SP_STATUS_DMA_BUSY | SP_STATUS_DMA_FULL) != 0 {
-                device.rsp.cpu.sync_point = true;
-            }
+
             device.rsp.last_status_value = value;
             device.rsp.regs[reg as usize]
         }
@@ -464,6 +467,7 @@ fn update_sp_status(device: &mut device::Device, w: u32) {
 fn do_task(device: &mut device::Device) {
     device.rsp.cpu.sync_point = false;
     device.rsp.last_status_value = 0;
+    device.rdp.last_status_value = 0;
     if device.rsp.regs[SP_DMA_BUSY_REG as usize] == 1 {
         device.rsp.run_after_dma = true
     } else {

src/device/rsp_su_instructions.rs

@@ -304,25 +304,44 @@ pub fn lh(device: &mut device::Device, opcode: u32) {
     let address =
         device.rsp.cpu.gpr[rs(opcode) as usize].wrapping_add(imm(opcode) as i16 as i32 as u32);
 
-    let mut w = [0; 2];
-    w[0] = device.rsp.mem[address as usize & 0xFFF];
-    w[1] = device.rsp.mem[(address as usize + 1) & 0xFFF];
+    // Optimize for aligned access when possible
+    let addr_masked = address as usize & 0xFFF;
+    let value = if addr_masked <= 0xFFE {
+        u16::from_be_bytes(
+            device.rsp.mem[addr_masked..addr_masked + 2]
+                .try_into()
+                .unwrap(),
+        )
+    } else {
+        let w0 = device.rsp.mem[addr_masked];
+        let w1 = device.rsp.mem[(addr_masked + 1) & 0xFFF];
+        ((w0 as u16) << 8) | w1 as u16
+    };
 
-    device.rsp.cpu.gpr[rt(opcode) as usize] =
-        (((w[0] as u16) << 8) | w[1] as u16) as i16 as i32 as u32
+    device.rsp.cpu.gpr[rt(opcode) as usize] = value as i16 as i32 as u32;
 }
 
 pub fn lw(device: &mut device::Device, opcode: u32) {
     let address =
         device.rsp.cpu.gpr[rs(opcode) as usize].wrapping_add(imm(opcode) as i16 as i32 as u32);
 
-    let mut w = [0; 4];
-    for (i, item) in w.iter_mut().enumerate() {
-        *item = device.rsp.mem[(address as usize + i) & 0xFFF];
-    }
+    // Optimize for aligned access when possible
+    let addr_masked = address as usize & 0xFFF;
+    let value = if addr_masked <= 0xFFC {
+        u32::from_be_bytes(
+            device.rsp.mem[addr_masked..addr_masked + 4]
+                .try_into()
+                .unwrap(),
+        )
+    } else {
+        let mut value = 0;
+        for i in 0..4 {
+            value |= (device.rsp.mem[(addr_masked + i) & 0xFFF] as u32) << (8 * (3 - i));
+        }
+        value
+    };
 
-    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)
+    device.rsp.cpu.gpr[rt(opcode) as usize] = value;
 }
 
 pub fn lbu(device: &mut device::Device, opcode: u32) {
@@ -336,24 +355,44 @@ pub fn lhu(device: &mut device::Device, opcode: u32) {
     let address =
         device.rsp.cpu.gpr[rs(opcode) as usize].wrapping_add(imm(opcode) as i16 as i32 as u32);
 
-    let mut w = [0; 2];
-    w[0] = device.rsp.mem[address as usize & 0xFFF];
-    w[1] = device.rsp.mem[(address as usize + 1) & 0xFFF];
+    // Optimize for aligned access when possible
+    let addr_masked = address as usize & 0xFFF;
+    let value = if addr_masked <= 0xFFE {
+        u16::from_be_bytes(
+            device.rsp.mem[addr_masked..addr_masked + 2]
+                .try_into()
+                .unwrap(),
+        )
+    } else {
+        let w0 = device.rsp.mem[addr_masked];
+        let w1 = device.rsp.mem[(addr_masked + 1) & 0xFFF];
+        ((w0 as u16) << 8) | w1 as u16
+    };
 
-    device.rsp.cpu.gpr[rt(opcode) as usize] = (((w[0] as u16) << 8) | w[1] as u16) as u32
+    device.rsp.cpu.gpr[rt(opcode) as usize] = value as u32;
 }
 
 pub fn lwu(device: &mut device::Device, opcode: u32) {
     let address =
         device.rsp.cpu.gpr[rs(opcode) as usize].wrapping_add(imm(opcode) as i16 as i32 as u32);
 
-    let mut w = [0; 4];
-    for (i, item) in w.iter_mut().enumerate() {
-        *item = device.rsp.mem[(address as usize + i) & 0xFFF];
-    }
+    // Optimize for aligned access when possible
+    let addr_masked = address as usize & 0xFFF;
+    let value = if addr_masked <= 0xFFC {
+        u32::from_be_bytes(
+            device.rsp.mem[addr_masked..addr_masked + 4]
+                .try_into()
+                .unwrap(),
+        )
+    } else {
+        let mut value = 0;
+        for i in 0..4 {
+            value |= (device.rsp.mem[(addr_masked + i) & 0xFFF] as u32) << (8 * (3 - i));
+        }
+        value
+    };
 
-    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)
+    device.rsp.cpu.gpr[rt(opcode) as usize] = value;
 }
 
 pub fn sb(device: &mut device::Device, opcode: u32) {
@@ -367,18 +406,30 @@ pub fn sh(device: &mut device::Device, opcode: u32) {
     let address =
         device.rsp.cpu.gpr[rs(opcode) as usize].wrapping_add(imm(opcode) as i16 as i32 as u32);
 
-    device.rsp.mem[address as usize & 0xFFF] = (device.rsp.cpu.gpr[rt(opcode) as usize] >> 8) as u8;
-    device.rsp.mem[(address as usize + 1) & 0xFFF] =
-        (device.rsp.cpu.gpr[rt(opcode) as usize]) as u8;
+    let addr_masked = address as usize & 0xFFF;
+    let value = device.rsp.cpu.gpr[rt(opcode) as usize] as u16;
+    // Optimize for aligned access when possible
+    if addr_masked <= 0xFFE {
+        device.rsp.mem[addr_masked..addr_masked + 2].copy_from_slice(&value.to_be_bytes());
+    } else {
+        device.rsp.mem[addr_masked] = (value >> 8) as u8;
+        device.rsp.mem[(addr_masked + 1) & 0xFFF] = value as u8;
+    }
 }
 
 pub fn sw(device: &mut device::Device, opcode: u32) {
     let address =
         device.rsp.cpu.gpr[rs(opcode) as usize].wrapping_add(imm(opcode) as i16 as i32 as u32);
 
-    for i in 0..4 {
-        device.rsp.mem[(address as usize + i) & 0xFFF] =
-            (device.rsp.cpu.gpr[rt(opcode) as usize] >> ((3 - i) * 8)) as u8;
+    let addr_masked = address as usize & 0xFFF;
+    let value = device.rsp.cpu.gpr[rt(opcode) as usize];
+    // Optimize for aligned access when possible
+    if addr_masked <= 0xFFC {
+        device.rsp.mem[addr_masked..addr_masked + 4].copy_from_slice(&value.to_be_bytes());
+    } else {
+        for i in 0..4 {
+            device.rsp.mem[(addr_masked + i) & 0xFFF] = (value >> ((3 - i) * 8)) as u8;
+        }
     }
 }
 
@@ -737,10 +788,21 @@ pub fn llv(device: &mut device::Device, opcode: u32) {
     let mut element = velement(opcode);
 
     if element.is_multiple_of(4) {
-        let mut value = 0;
-        for i in 0..4 {
-            value |= (device.rsp.mem[((address + i) & 0xFFF) as usize] as u32) << (8 * (3 - i));
-        }
+        // Optimize for aligned access when possible
+        let addr_masked = address as usize & 0xFFF;
+        let value = if addr_masked <= 0xFFC {
+            u32::from_be_bytes(
+                device.rsp.mem[addr_masked..addr_masked + 4]
+                    .try_into()
+                    .unwrap(),
+            )
+        } else {
+            let mut value = 0;
+            for i in 0..4 {
+                value |= (device.rsp.mem[(addr_masked + i) & 0xFFF] as u32) << (8 * (3 - i));
+            }
+            value
+        };
         modify_vpr32(
             &mut device.rsp.cpu.vpr[rt(opcode) as usize],
             element / 4,
@@ -767,10 +829,21 @@ pub fn ldv(device: &mut device::Device, opcode: u32) {
     let mut element = velement(opcode);
 
     if element.is_multiple_of(8) {
-        let mut value = 0;
-        for i in 0..8 {
-            value |= (device.rsp.mem[((address + i) & 0xFFF) as usize] as u64) << (8 * (7 - i));
-        }
+        // Optimize for aligned access when possible
+        let addr_masked = address as usize & 0xFFF;
+        let value = if addr_masked <= 0xFF8 {
+            u64::from_be_bytes(
+                device.rsp.mem[addr_masked..addr_masked + 8]
+                    .try_into()
+                    .unwrap(),
+            )
+        } else {
+            let mut value = 0;
+            for i in 0..8 {
+                value |= (device.rsp.mem[(addr_masked + i) & 0xFFF] as u64) << (8 * (7 - i));
+            }
+            value
+        };
         modify_vpr64(
             &mut device.rsp.cpu.vpr[rt(opcode) as usize],
             element / 8,
@@ -797,10 +870,21 @@ pub fn lqv(device: &mut device::Device, opcode: u32) {
     let mut element = velement(opcode);
 
     if element == 0 && address.is_multiple_of(16) {
-        let mut value = 0;
-        for i in 0..16 {
-            value |= (device.rsp.mem[((address + i) & 0xFFF) as usize] as u128) << (8 * (15 - i));
-        }
+        // Optimize for aligned 16-byte access
+        let addr_masked = address as usize & 0xFFF;
+        let value = if addr_masked <= 0xFF0 {
+            u128::from_be_bytes(
+                device.rsp.mem[addr_masked..addr_masked + 16]
+                    .try_into()
+                    .unwrap(),
+            )
+        } else {
+            let mut value = 0;
+            for i in 0..16 {
+                value |= (device.rsp.mem[(addr_masked + i) & 0xFFF] as u128) << (8 * (15 - i));
+            }
+            value
+        };
         modify_vpr128(&mut device.rsp.cpu.vpr[rt(opcode) as usize], value);
     } else {
         let end = std::cmp::min(16 + element - ((address & 15) as u8), 16);
@@ -823,10 +907,21 @@ 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;
     if element == 0 {
-        let mut value = 0;
-        for i in 0..16 {
-            value |= (device.rsp.mem[((address + i) & 0xFFF) as usize] as u128) << (8 * (15 - i));
-        }
+        // Optimize for aligned 16-byte access
+        let addr_masked = address as usize & 0xFFF;
+        let value = if addr_masked <= 0xFF0 {
+            u128::from_be_bytes(
+                device.rsp.mem[addr_masked..addr_masked + 16]
+                    .try_into()
+                    .unwrap(),
+            )
+        } else {
+            let mut value = 0;
+            for i in 0..16 {
+                value |= (device.rsp.mem[(addr_masked + i) & 0xFFF] as u128) << (8 * (15 - i));
+            }
+            value
+        };
         modify_vpr128(&mut device.rsp.cpu.vpr[rt(opcode) as usize], value);
     } else {
         while element < 16 {

src/device/si.rs

@@ -1,6 +1,7 @@
 use rand_chacha::rand_core::RngCore;
 
 use crate::device;
+use crate::ui;
 
 const SI_DRAM_ADDR_REG: u32 = 0;
 const SI_PIF_ADDR_RD64B_REG: u32 = 1;
@@ -98,6 +99,8 @@ fn copy_pif_rdram(device: &mut device::Device) {
             i += 4;
         }
     } else if device.si.dma_dir == DmaDir::Read {
+        // check RDP before SI writes to RDRAM
+        ui::video::check_framebuffers(dram_addr as u32, device::pif::PIF_RAM_SIZE as u32);
         let mut i = 0;
         while i < device::pif::PIF_RAM_SIZE {
             let data = u32::from_be_bytes(device.pif.ram[i..i + 4].try_into().unwrap());

src/ui/gui/netplay_join.slint

@@ -9,6 +9,7 @@ export component NetplayJoin inherits Window {
     in-out property sessions <=> NetplayData.sessions;
     in-out property current_session <=> NetplayData.current_session;
     in-out property room_urls <=> NetplayData.room_urls;
+    in-out property room_ports <=> NetplayData.room_ports;
     in-out property game_name <=> NetplayData.game_name;
     in-out property game_hash <=> NetplayData.game_hash;
     in-out property game_cheats <=> NetplayData.game_cheats;
@@ -19,7 +20,7 @@ export component NetplayJoin inherits Window {
     in-out property rom_dir <=> NetplayData.rom_dir;
     callback refresh_session();
     callback select_rom(string);
-    callback join_session(string, string, string, string, string);
+    callback join_session(string, string, string, string, int);
     preferred-width: 700px;
 
     VerticalBox {
@@ -93,6 +94,7 @@ export component NetplayJoin inherits Window {
                     NetplayData.current_session = -1;
                     NetplayData.sessions = [];
                     NetplayData.room_urls = [];
+                    NetplayData.room_ports = [];
                     refresh_session();
                 }
             }
@@ -129,7 +131,7 @@ export component NetplayJoin inherits Window {
                     NetplayData.game_hash,
                     NetplayData.password,
                     NetplayData.room_urls[NetplayData.current_session],
-                    NetplayData.sessions[NetplayData.current_session][1].text);
+                    NetplayData.room_ports[NetplayData.current_session]);
                 }
             }
         }

src/ui/gui/netplay_page.slint

@@ -19,6 +19,7 @@ export global NetplayData {
     in-out property <string> session_name;
     in-out property <[[StandardListViewItem]]> sessions;
     in-out property <[string]> room_urls;
+    in-out property <[int]> room_ports;
     in-out property <int> current_session: -1;
     in-out property <string> rom_dir;
     in-out property <string> custom_server_url;

src/ui/netplay.rs

@@ -491,6 +491,7 @@ fn update_sessions(weak: slint::Weak<NetplayJoin>) {
             tokio::spawn(async move {
                 let mut sessions = vec![];
                 let mut room_urls = vec![];
+                let mut room_ports = vec![];
                 for (server_name, server_url) in servers.iter() {
                     if let Ok(Ok((socket, _response))) = tokio::time::timeout(
                         std::time::Duration::from_secs(2),
@@ -533,6 +534,7 @@ fn update_sessions(weak: slint::Weak<NetplayJoin>) {
                             for room in rooms {
                                 let mut session = vec![];
                                 room_urls.push(server_url.into());
+                                room_ports.push(room.port.unwrap());
 
                                 session.push(slint::StandardListViewItem::from(
                                     slint::SharedString::from(server_name),
@@ -585,6 +587,11 @@ fn update_sessions(weak: slint::Weak<NetplayJoin>) {
                             std::rc::Rc::new(room_urls_vec);
                         handle.set_room_urls(slint::ModelRc::from(room_urls_model));
 
+                        let room_ports_vec = slint::VecModel::from(room_ports.to_vec());
+                        let room_ports_model: std::rc::Rc<slint::VecModel<i32>> =
+                            std::rc::Rc::new(room_ports_vec);
+                        handle.set_room_ports(slint::ModelRc::from(room_ports_model));
+
                         handle.set_current_session(-1);
                         handle.set_pending_refresh(false);
                     })
@@ -732,7 +739,7 @@ fn join_session(
     player_name: String,
     game_hash: String,
     password: String,
-    room_name: String,
+    room_port: i32,
     fullscreen: bool,
     weak_app: slint::Weak<AppWindow>,
     weak: slint::Weak<NetplayJoin>,
@@ -740,73 +747,26 @@ fn join_session(
     tokio::spawn(async move {
         let now_utc = chrono::Utc::now().timestamp_millis().to_string();
         let hasher = Sha256::new().chain_update(&now_utc).chain_update(EMU_NAME);
-        let request_rooms = NetplayMessage {
-            message_type: "request_get_rooms".to_string(),
-            player_name: None,
-            client_sha: None,
-            netplay_version: Some(NETPLAY_VERSION),
-            player_names: None,
-            emulator: Some(EMU_NAME.to_string()),
-            accept: None,
-            rooms: None,
-            message: None,
-            auth_time: Some(now_utc),
-            auth: Some(format!("{:x}", hasher.finalize())),
-            room: None,
-        };
-
-        netplay_write_sender.send(Some(request_rooms)).unwrap();
-
-        let mut port = 0;
-        match tokio::time::timeout(
-            std::time::Duration::from_secs(2),
-            netplay_read_receiver.recv(),
-        )
-        .await
-        {
-            Ok(Ok(message)) => {
-                if message.message_type == "reply_get_rooms"
-                    && message.accept.unwrap() == 0
-                    && let Some(rooms) = message.rooms
-                {
-                    for room in rooms {
-                        if room.room_name.unwrap() == room_name {
-                            port = room.port.unwrap_or(0);
-                        }
-                    }
-                }
-            }
-            Ok(Err(err)) => {
-                panic!("netplay_read_receiver error: {err}");
-            }
-            Err(_) => {
-                weak.upgrade_in_event_loop(move |handle| {
-                    handle.set_pending_session(false);
-                    show_netplay_error("Server did not respond".to_string());
-                })
-                .unwrap();
-            }
-        }
 
         let join_room = NetplayMessage {
             message_type: "request_join_room".to_string(),
             player_name: Some(player_name),
             client_sha: Some(env!("GIT_HASH").to_string()),
-            netplay_version: None,
-            emulator: None,
+            netplay_version: Some(NETPLAY_VERSION),
+            emulator: Some(EMU_NAME.to_string()),
             accept: None,
             message: None,
             rooms: None,
-            auth_time: None,
             player_names: None,
-            auth: None,
+            auth_time: Some(now_utc),
+            auth: Some(format!("{:x}", hasher.finalize())),
             room: Some(NetplayRoom {
                 room_name: None,
                 password: Some(password),
                 game_name: None,
                 md5: Some(game_hash),
                 protected: None,
-                port: Some(port),
+                port: Some(room_port),
                 features: None,
                 buffer_target: None,
             }),
@@ -1200,7 +1160,7 @@ pub fn setup_join_window(
     });
     let weak = join_window.as_weak();
     join_window.on_join_session(
-        move |player_name, game_hash, password, room_url, room_name| {
+        move |player_name, game_hash, password, room_url, room_port| {
             let _ = netplay_write_sender.send(None); // close current websocket if any
             manage_websocket(
                 room_url.to_string(),
@@ -1215,7 +1175,7 @@ pub fn setup_join_window(
                 player_name.to_string(),
                 game_hash.to_string(),
                 password.to_string(),
-                room_name.to_string(),
+                room_port,
                 fullscreen,
                 weak_app.clone(),
                 weak.clone(),

src/ui/video.rs

@@ -192,6 +192,10 @@ pub fn process_rdp_list() -> u64 {
     unsafe { rdp_process_commands() }
 }
 
+pub fn check_framebuffers(address: u32, length: u32) {
+    unsafe { rdp_check_framebuffers(address, length) }
+}
+
 pub fn onscreen_message(_ui: &ui::Ui, message: &str) {
     unsafe { rdp_onscreen_message(std::ffi::CString::new(message).unwrap().as_ptr()) };
 }