tbgpu: shim binary + remote apl pci dev (#14124)

* shim binary + remote pci dev * v2 * rip out apl * cmds * rename * clean * remove * rm gitignore * ui * install * linter * um * cleaner * assets * normal install in ui * cleaner app * install script * support fd mmap * cleaner * kill server when disconn * rename + pcidevs * sign * install and reinstall * no sip install * will trigger update * nv * ugh * this * fix * nv * use nosip sign * auto install * remove * mypy * upd * ditto * print * simpler * ditto * um * simpler * upd * upd * cleaner * autogen * cleaner * move * annotations * server cleaner
2026-04-29 03:00:14 -04:00 · 2026-01-20 16:15:18 +03:00
parent 4548fcc1b8
commit dc82856084
26 changed files with 1485 additions and 406 deletions
--- a/.github/workflows/autogen.yml
+++ b/.github/workflows/autogen.yml
@@ -43,7 +43,7 @@ jobs:
      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev
    - name: Regenerate autogen files
      run: |
-        find tinygrad/runtime/autogen -type f -name "*.py" -not -name "__init__.py" -not -name "comgr_3.py" -not -name "metal.py" -not -name "libclang.py" -delete
+        find tinygrad/runtime/autogen -type f -name "*.py" -not -name "__init__.py" -not -name "comgr_3.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
        python3 -c "from tinygrad.runtime.autogen import opencl"
        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv"
        python3 -c "from tinygrad.runtime.autogen import comgr, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd"
@@ -83,8 +83,8 @@ jobs:
        llvm: 'true'
    - name: Regenerate autogen files
      run: |
-        rm tinygrad/runtime/autogen/metal.py
-        LIBCLANG_PATH=/opt/homebrew/opt/llvm@20/lib/libclang.dylib python3 -c "from tinygrad.runtime.autogen import metal"
+        rm tinygrad/runtime/autogen/metal.py tinygrad/runtime/autogen/iokit.py tinygrad/runtime/autogen/corefoundation.py
+        LIBCLANG_PATH=/opt/homebrew/opt/llvm@20/lib/libclang.dylib python3 -c "from tinygrad.runtime.autogen import metal, iokit, corefoundation"
    - name: Check for differences
      run: |
        if ! git diff --quiet; then
--- a/extra/usbgpu/tbgpu/install_tinygpu.sh
+++ b/extra/usbgpu/tbgpu/install_tinygpu.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+set -e
+
+APP_PATH="/Applications/TinyGPU.app"
+DEXT_ID="org.tinygrad.tinygpu.edriver"
+
+# Install app if not present. TODO: url
+if [[ ! -d "$APP_PATH" ]]; then
+  echo "TinyGPU.app not found in /Applications"
+  exit 1
+fi
+
+# Ask user to install
+read -n1 -p "Install TinyGPU driver extension now? [y/N] " answer
+echo
+
+if [[ "$answer" =~ ^[Yy]$ ]]; then
+  "$APP_PATH/Contents/MacOS/TinyGPU" install
+else
+  echo "Skipped."
+  exit 0
+fi
--- a/extra/usbgpu/tbgpu/installer/Shared/TinyGPU-Bridging-Header.h
+++ b/extra/usbgpu/tbgpu/installer/Shared/TinyGPU-Bridging-Header.h
@@ -0,0 +1 @@
+int run_server(const char *sock_path);
--- a/extra/usbgpu/tbgpu/installer/Shared/TinyGPUApp.swift
+++ b/extra/usbgpu/tbgpu/installer/Shared/TinyGPUApp.swift
@@ -1,19 +1,48 @@
-import AppKit
 import SwiftUI

-final class AppDelegate: NSObject, NSApplicationDelegate {
-	func applicationShouldTerminateAfterLastWindowClosed(_ sender: NSApplication) -> Bool {
-		true
-	}
-}
+private let dextID = "org.tinygrad.tinygpu.edriver"

@main
 struct TinyGPUApp: App {
-	@NSApplicationDelegateAdaptor(AppDelegate.self) var appDelegate
+  private static var runner: TinyGPUCLIRunner?  // prevent dealloc before callback
+  @State private var text = ""
+  @NSApplicationDelegateAdaptor(AppDelegate.self) var appDelegate

-    var body: some Scene {
-        WindowGroup {
-			TinyGPUView()
-        }
+  init() {
+    guard CommandLine.arguments.count > 1 else { return }
+    Self.runner = TinyGPUCLIRunner(dextID)
+    Self.runner?.run(args: CommandLine.arguments) { exit($0.rawValue) }
+    dispatchMain()
+  }
+
+  var body: some Scene {
+    WindowGroup("TinyGPU") {
+      ScrollView {
+        Text(text).font(.custom("Menlo", size: 11)).frame(maxWidth: .infinity, alignment: .leading).padding(8)
+      }
+      .frame(width: 500, height: 300).padding()
+      .onAppear { setup() }
    }
+    .commands { CommandGroup(replacing: .newItem) {} }
+  }
+
+  func setup() {
+    let bundlePath = Bundle.main.bundlePath
+    guard bundlePath.hasPrefix("/Applications/") else {
+      var error: NSDictionary?
+      NSAppleScript(source: "do shell script \"mv '\(bundlePath)' '/Applications/'\" with administrator privileges")?.executeAndReturnError(&error)
+      text = error == nil ? "Moved! Please reopen from /Applications/\n" : "Move TinyGPU to /Applications first.\n"
+      return
+    }
+    let state = TinyGPUCLIRunner.queryDextState(dextID)
+    if state == .unloaded || state == .activating {
+      Self.runner = TinyGPUCLIRunner(dextID)
+      Self.runner?.run(args: ["", "install"]) { _ in }
+    }
+    text = "TinyGPU - Remote PCI Device Server\n\n" + TinyGPUCLIRunner.statusText(state)
+  }
+}
+
+class AppDelegate: NSObject, NSApplicationDelegate {
+  func applicationShouldTerminateAfterLastWindowClosed(_ sender: NSApplication) -> Bool { true }
 }
--- a/extra/usbgpu/tbgpu/installer/Shared/TinyGPUCLIRunner.swift
+++ b/extra/usbgpu/tbgpu/installer/Shared/TinyGPUCLIRunner.swift
@@ -0,0 +1,122 @@
+import Foundation
+import SystemExtensions
+
+enum TinyGPUCLIExit: Int32 { case ok = 0, usage = 2, failed = 3, needsApproval = 4 }
+enum DextState { case unloaded, activating, needsApproval, activated }
+
+final class TinyGPUCLIRunner: NSObject, OSSystemExtensionRequestDelegate {
+  private let dextID: String
+  private var done: ((TinyGPUCLIExit) -> Void)?
+  private var isInstall = true
+
+  init(_ dextID: String) { self.dextID = dextID }
+
+  static func queryDextState(_ bundleID: String) -> DextState {
+    let p = Process()
+    p.executableURL = URL(fileURLWithPath: "/usr/bin/systemextensionsctl")
+    p.arguments = ["list"]
+    let pipe = Pipe()
+    p.standardOutput = pipe
+    p.standardError = Pipe()
+    guard (try? p.run()) != nil else { return .unloaded }
+    p.waitUntilExit()
+
+    guard let output = String(data: pipe.fileHandleForReading.readDataToEndOfFile(), encoding: .utf8),
+          let line = output.split(separator: "\n").first(where: { $0.contains(bundleID) }) else { return .unloaded }
+    if line.contains("[activated enabled]") { return .activated }
+    if line.contains("[activated waiting for user]") { return .needsApproval }
+    return line.contains("terminated waiting to uninstall") ? .unloaded : .activating
+  }
+
+  private static let approvalHelp = """
+    Please go to System Settings > Privacy & Security and allow the extension.
+
+    If previously disabled: System Settings > General > Login Items & Extensions > Driver Extensions > Toggle TinyGPU ON
+
+    """
+
+  static func statusText(_ state: DextState) -> String {
+    switch state {
+    case .unloaded: return "Driver extension not installed.\n\n"
+    case .activating: return "Extension is activating...\n\n"
+    case .needsApproval: return "Extension awaiting approval.\n\n" + approvalHelp
+    case .activated: return "Extension is ready! Run tinygrad to use your eGPU.\n\n"
+    }
+  }
+
+  func run(args: [String], done: @escaping (TinyGPUCLIExit) -> Void) {
+    self.done = done
+    guard args.count > 1 else { return usage() }
+
+    switch args[1] {
+    case "status":
+      print(Self.statusText(Self.queryDextState(dextID)))
+      done(.ok)
+    case "install":
+      if Self.queryDextState(dextID) == .needsApproval { print(Self.statusText(.needsApproval)); return done(.needsApproval) }
+      print("Installing TinyGPU driver extension...\nYou may need to approve in System Settings.\n")
+      submitRequest(activate: true)
+    case "uninstall":
+      guard Self.queryDextState(dextID) != .unloaded else { print("Not installed.\n"); return done(.ok) }
+      print("Uninstalling TinyGPU driver extension...\n")
+      isInstall = false
+      submitRequest(activate: false)
+    case "server":
+      guard args.count > 2 else { print("Error: server requires socket path\n"); return usage() }
+      done(run_server(args[2]) == 0 ? .ok : .failed)
+    case "help", "-h", "--help":
+      usage(); done(.ok)
+    default:
+      print("Unknown command: \(args[1])\n"); usage()
+    }
+  }
+
+  private func usage() {
+    print("""
+      Usage: TinyGPU <command>
+        status     Show extension status
+        install    Install the driver extension
+        uninstall  Remove the driver extension
+        server <path>  Start server on Unix socket
+      """)
+    done?(.usage)
+  }
+
+  private func submitRequest(activate: Bool) {
+    let req = activate
+      ? OSSystemExtensionRequest.activationRequest(forExtensionWithIdentifier: dextID, queue: .main)
+      : OSSystemExtensionRequest.deactivationRequest(forExtensionWithIdentifier: dextID, queue: .main)
+    req.delegate = self
+    OSSystemExtensionManager.shared.submitRequest(req)
+  }
+
+  // MARK: - OSSystemExtensionRequestDelegate
+  func requestNeedsUserApproval(_ request: OSSystemExtensionRequest) {
+    print("\nUser approval required!\n\n\(Self.approvalHelp)After approval, connect the gpu and use it with tinygrad.\n")
+    done?(.needsApproval)
+  }
+
+  func request(_ request: OSSystemExtensionRequest, didFinishWithResult result: OSSystemExtensionRequest.Result) {
+    switch result {
+    case .completed: print("Driver extension \(isInstall ? "installed" : "uninstalled") successfully!\n")
+    case .willCompleteAfterReboot: print("Will complete after reboot.\n")
+    @unknown default: print("Completed: \(result)\n")
+    }
+    done?(.ok)
+  }
+
+  func request(_ request: OSSystemExtensionRequest, didFailWithError error: Error) {
+    print("\nError: \(error.localizedDescription)\n")
+    let code = (error as NSError).code
+    if code == 4 { print("Missing entitlements. Rebuild with proper signing.\n") }
+    else if code == 8 { print("Extension not found in app bundle.\n") }
+    else if code == 9 { print("Extension disabled by user.\n\n\(Self.approvalHelp)") }
+    done?(.failed)
+  }
+
+  func request(_ request: OSSystemExtensionRequest, actionForReplacingExtension existing: OSSystemExtensionProperties,
+               withExtension ext: OSSystemExtensionProperties) -> OSSystemExtensionRequest.ReplacementAction {
+    print("Updating v\(existing.bundleShortVersion) -> v\(ext.bundleShortVersion)...\n")
+    return .replace
+  }
+}
--- a/extra/usbgpu/tbgpu/installer/Shared/TinyGPUView.swift
+++ b/extra/usbgpu/tbgpu/installer/Shared/TinyGPUView.swift
@@ -1,33 +0,0 @@
-import SwiftUI
-
-struct TinyGPUView: View {
-	@ObservedObject var viewModel = TinyGPUViewModel()
-
-	var body: some View {
-#if os(macOS)
-		VStack(alignment: .center) {
-			Text("TinyGPU Intsaller")
-				.padding()
-				.font(.title)
-			Text(self.viewModel.dextLoadingState)
-				.multilineTextAlignment(.center)
-			HStack {
-				Button(
-					action: {
-						self.viewModel.activateMyDext()
-					}, label: {
-						Text("Install extension")
-					}
-				)
-			}
-		}
-		.frame(width: 500, height: 200, alignment: .center)
-#endif
-	}
-}
-
-struct TinyGPUView_Previews: PreviewProvider {
-    static var previews: some View {
-		TinyGPUView()
-    }
-}
--- a/extra/usbgpu/tbgpu/installer/Shared/TinyGPUViewModel.swift
+++ b/extra/usbgpu/tbgpu/installer/Shared/TinyGPUViewModel.swift
@@ -1,142 +0,0 @@
-import Foundation
-import os.log
-import SystemExtensions
-
-class TinyGPUDriverLoadingStateMachine {
-  enum State { case unloaded, activating, needsApproval, activated, activationError }
-}
-
-class TinyGPUViewModel: NSObject {
-
-	@Published private var state: TinyGPUDriverLoadingStateMachine.State = .unloaded
-	
-	override init() {
-		super.init()
-		refreshInitialDextState()
-	}
-	
-	private func refreshInitialDextState() {
-#if os(macOS)
-		Task.detached { [dextIdentifier] in
-		  let newState = Self.queryDextState(bundleID: dextIdentifier)
-		  await MainActor.run { self.state = newState }
-		}
-#endif
-	}
-
-#if os(macOS)
-	private static func queryDextState(bundleID: String) -> TinyGPUDriverLoadingStateMachine.State {
-		let tool = "/usr/bin/systemextensionsctl"
-		let p = Process()
-		p.executableURL = URL(fileURLWithPath: tool)
-		p.arguments = ["list"]
-
-		let pipe = Pipe()
-		p.standardOutput = pipe
-		p.standardError = Pipe()
-
-		do {
-			try p.run()
-			p.waitUntilExit()
-			let data = pipe.fileHandleForReading.readDataToEndOfFile()
-			guard let output = String(data: data, encoding: .utf8) else { return .unloaded }
-
-			// Look for our bundle id line
-			if let line = output.split(separator: "\n").first(where: { $0.contains(bundleID) }) {
-				if line.contains("[activated enabled]") { return .activated }
-				if line.contains("[activated waiting for user]") { return .needsApproval }
-				if line.contains("terminated waiting to uninstall") { return .unloaded }
-				return .activating
-			} else {
-				return .unloaded
-			}
-		} catch {
-			return .unloaded
-		}
-	}
-#endif
-
-	private let dextIdentifier: String = "org.tinygrad.tinygpu.edriver"
-
-	public var dextLoadingState: String {
-		switch state {
-		case .unloaded:
-			return "TinyGPUDriver isn't loaded."
-		case .activating:
-			return "Activating TinyGPUDriver, please wait."
-		case .needsApproval:
-			return "Please follow the prompt to approve TinyGPUDriver."
-		case .activated:
-			return "TinyGPUDriver has been activated and is ready to use. You can close the installer."
-		case .activationError:
-			return "TinyGPUDriver has experienced an error during activation.\nPlease check the logs to find the error."
-		}
-	}
-}
-
-extension TinyGPUViewModel: ObservableObject {
-
-#if os(macOS)
-	func activateMyDext() {
-		activateExtension(dextIdentifier)
-	}
-	
-	func deactivateMyDext() {
-		deactivateExtension(dextIdentifier)
-	}
-
-	func activateExtension(_ dextIdentifier: String) {
-
-		let request = OSSystemExtensionRequest
-			.activationRequest(forExtensionWithIdentifier: dextIdentifier,
-							   queue: .main)
-		request.delegate = self
-		OSSystemExtensionManager.shared.submitRequest(request)
-
-		self.state = .activating
-	}
-	
-	func deactivateExtension(_ dextIdentifier: String) {
-
-		let request = OSSystemExtensionRequest.deactivationRequest(forExtensionWithIdentifier: dextIdentifier, queue: .main)
-		request.delegate = self
-		OSSystemExtensionManager.shared.submitRequest(request)
-
-		self.state = .unloaded
-	}
-#endif
-}
-
-#if os(macOS)
-extension TinyGPUViewModel: OSSystemExtensionRequestDelegate {
-
-	func request(
-		_ request: OSSystemExtensionRequest,
-		actionForReplacingExtension existing: OSSystemExtensionProperties,
-		withExtension ext: OSSystemExtensionProperties) -> OSSystemExtensionRequest.ReplacementAction {
-
-		var replacementAction: OSSystemExtensionRequest.ReplacementAction
-
-		os_log("sysex actionForReplacingExtension: %@ %@", existing, ext)
-
-		replacementAction = .replace
-		self.state = .activating
-		return replacementAction
-	}
-
-	func requestNeedsUserApproval(_ request: OSSystemExtensionRequest) {
-		os_log("sysex requestNeedsUserApproval")
-		self.state = .needsApproval
-	}
-
-	func request(_ request: OSSystemExtensionRequest, didFinishWithResult result: OSSystemExtensionRequest.Result) {
-		os_log("sysex didFinishWithResult: %d", result.rawValue)
-		self.state = .activated
-	}
-
-	func request(_ request: OSSystemExtensionRequest, didFailWithError error: Error) {
-		os_log("sysex didFailWithError: %@", error.localizedDescription)
-		self.state = .activationError
-	}
-}
-#endif
--- a/extra/usbgpu/tbgpu/installer/Shared/server.c
+++ b/extra/usbgpu/tbgpu/installer/Shared/server.c
@@ -0,0 +1,277 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include <unistd.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+#include <sys/mman.h>
+#include <fcntl.h>
+#include <errno.h>
+#include <dispatch/dispatch.h>
+#include <CoreFoundation/CoreFoundation.h>
+#include <IOKit/IOKitLib.h>
+#include <IOKit/IOMessage.h>
+#include <mach/mach.h>
+
+// Protocol
+
+enum {
+  CMD_MAP_BAR = 1,        // map PCI BAR, returns size
+  CMD_MAP_SYSMEM_FD = 2,  // alloc DMA memory, returns fd via SCM_RIGHTS
+  CMD_CFG_READ = 3,       // read PCI config space
+  CMD_CFG_WRITE = 4,      // write PCI config space
+  CMD_RESET = 5,          // reset device
+  CMD_MMIO_READ = 6,      // bulk read from BAR
+  CMD_MMIO_WRITE = 7,     // bulk write to BAR
+  RESP_OK = 0, RESP_ERR = 1,
+};
+
+typedef struct { uint8_t cmd, bar; uint64_t offset, size, value; } __attribute__((packed)) request_t;
+typedef struct { uint8_t status; uint64_t value, addr; } __attribute__((packed)) response_t;
+
+// Constants and state
+
+#define BULK_BUF_SIZE (64 << 20)
+#define MAX_BARS 6
+#define MAX_SYSMEM 128
+
+static uint8_t g_bulk_buf[BULK_BUF_SIZE];
+static io_connect_t g_conn = IO_OBJECT_NULL;
+static int g_client_active = 0;
+
+static struct { mach_vm_address_t addr; mach_vm_size_t size; } g_bars[MAX_BARS];
+static struct { mach_vm_address_t addr; mach_vm_size_t size; int shm_fd; char shm_name[32]; } g_sysmem[MAX_SYSMEM];
+static int g_sysmem_count = 0;
+
+// Utilities
+
+static void recvall(int fd, void *buf, size_t len) {
+  for (size_t off = 0; off < len; ) {
+    ssize_t r = recv(fd, (uint8_t*)buf + off, len - off, 0);
+    if (r <= 0) break;
+    off += r;
+  }
+}
+
+// MMIO requires 32-bit aligned volatile accesses
+static void mmio_copy(void *dst, void *src, size_t len) {
+  volatile uint32_t *d = dst, *s = src;
+  for (size_t i = 0; i < len / 4; i++) d[i] = s[i];
+  for (size_t i = len & ~3; i < len; i++) ((volatile uint8_t*)dst)[i] = ((volatile uint8_t*)src)[i];
+}
+
+static int send_response(int fd, response_t *resp, int send_fd) {
+  char cmsgbuf[CMSG_SPACE(sizeof(int))];
+  struct iovec iov = {resp, sizeof(*resp)};
+  struct msghdr msg = {.msg_iov = &iov, .msg_iovlen = 1};
+
+  if (send_fd >= 0) {
+    msg.msg_control = cmsgbuf;
+    msg.msg_controllen = sizeof(cmsgbuf);
+    struct cmsghdr *cmsg = CMSG_FIRSTHDR(&msg);
+    *cmsg = (struct cmsghdr){.cmsg_level = SOL_SOCKET, .cmsg_type = SCM_RIGHTS, .cmsg_len = CMSG_LEN(sizeof(int))};
+    memcpy(CMSG_DATA(cmsg), &send_fd, sizeof(int));
+  }
+  return sendmsg(fd, &msg, 0) > 0 ? 0 : -1;
+}
+
+static void send_error(int fd, const char *msg) {
+  response_t resp = {.status = RESP_ERR, .value = strlen(msg)};
+  send_response(fd, &resp, -1);
+  send(fd, msg, strlen(msg), 0);
+}
+
+// Driver interface
+
+static void on_disconnect(void *refcon, io_service_t svc, uint32_t msg, void *arg) {
+  if (msg == kIOMessageServiceIsTerminated) _exit(0);
+}
+
+static io_connect_t open_tinygpu(void) {
+  static io_object_t notif;
+  io_service_t svc = IOServiceGetMatchingService(kIOMainPortDefault, IOServiceNameMatching("tinygpu"));
+  if (!svc) return IO_OBJECT_NULL;
+
+  if (!notif) {
+    IONotificationPortRef port = IONotificationPortCreate(kIOMainPortDefault);
+    IONotificationPortSetDispatchQueue(port, dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_HIGH, 0));
+    IOServiceAddInterestNotification(port, svc, kIOGeneralInterest, on_disconnect, NULL, &notif);
+  }
+
+  io_connect_t conn;
+  kern_return_t kr = IOServiceOpen(svc, mach_task_self(), 0, &conn);
+  IOObjectRelease(svc);
+  return kr == KERN_SUCCESS ? conn : IO_OBJECT_NULL;
+}
+
+static int dext_rpc(uint32_t sel, uint64_t *in, uint32_t in_cnt, uint64_t *out_val) {
+  uint64_t out[2];
+  uint32_t out_cnt = 2;
+  if (IOConnectCallMethod(g_conn, sel, in, in_cnt, NULL, 0, out, &out_cnt, NULL, NULL) != KERN_SUCCESS) return -1;
+  if (out_val) *out_val = out[0];
+  return 0;
+}
+
+static int map_bar(uint32_t bar, response_t *resp) {
+  if (bar >= MAX_BARS) return -1;
+  if (!g_bars[bar].addr && IOConnectMapMemory64(g_conn, bar, mach_task_self(), &g_bars[bar].addr, &g_bars[bar].size, kIOMapAnywhere)) return -1;
+  resp->addr = g_bars[bar].addr;
+  resp->value = g_bars[bar].size;
+  return 0;
+}
+
+static int map_sysmem_fd(uint64_t size, response_t *resp, int *out_fd) {
+  if (g_sysmem_count >= MAX_SYSMEM) return -1;
+  int idx = g_sysmem_count;
+  int fd = -1;
+  void *ptr = MAP_FAILED;
+  char shm_name[32];
+
+  // page-align, min 16KB for IOMemoryDescriptor
+  size_t alloc_sz = (size + 0xfff) & ~0xfff;
+  if (alloc_sz < 0x4000) alloc_sz = 0x4000;
+
+  snprintf(shm_name, sizeof(shm_name), "/tinygpu_%d", idx);
+  shm_unlink(shm_name);
+  if ((fd = shm_open(shm_name, O_CREAT | O_RDWR, 0600)) < 0) goto fail;
+  if (ftruncate(fd, alloc_sz) < 0) goto fail;
+  if ((ptr = mmap(NULL, alloc_sz, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0)) == MAP_FAILED) goto fail;
+
+  // PrepareDMA writes physical addresses to output buffer, copy to shared mem
+  uint8_t paddr_buf[8192] = {0};
+  size_t out_sz = sizeof(paddr_buf);
+  if (IOConnectCallStructMethod(g_conn, 3, ptr, alloc_sz, paddr_buf, &out_sz) != KERN_SUCCESS) goto fail;
+  memcpy(ptr, paddr_buf, out_sz);
+
+  g_sysmem[idx] = (typeof(g_sysmem[idx])){.addr = (mach_vm_address_t)ptr, .size = alloc_sz, .shm_fd = fd};
+  strncpy(g_sysmem[idx].shm_name, shm_name, sizeof(g_sysmem[idx].shm_name));
+  g_sysmem_count++;
+
+  *resp = (response_t){.addr = idx, .value = alloc_sz};
+  *out_fd = fd;
+  return 0;
+
+fail:
+  if (ptr != MAP_FAILED) munmap(ptr, alloc_sz);
+  if (fd >= 0) close(fd);
+  shm_unlink(shm_name);
+  return -1;
+}
+
+static int validate_bar(uint8_t bar, uint64_t off, uint64_t sz) {
+  return (bar < MAX_BARS && g_bars[bar].addr && off + sz <= g_bars[bar].size && sz <= BULK_BUF_SIZE) ? 0 : -1;
+}
+
+static void cleanup(void) {
+  for (int i = 0; i < MAX_BARS; i++)
+    if (g_bars[i].addr) { IOConnectUnmapMemory64(g_conn, i, mach_task_self(), g_bars[i].addr); g_bars[i].addr = 0; }
+
+  for (int i = 0; i < g_sysmem_count; i++) {
+    munmap((void*)g_sysmem[i].addr, g_sysmem[i].size);
+    close(g_sysmem[i].shm_fd);
+    shm_unlink(g_sysmem[i].shm_name);
+  }
+
+  g_sysmem_count = 0;
+  if (g_conn != IO_OBJECT_NULL) { IOServiceClose(g_conn); g_conn = IO_OBJECT_NULL; }
+}
+
+static void handle_client(int fd) {
+  int bufsize = BULK_BUF_SIZE;
+  setsockopt(fd, SOL_SOCKET, SO_SNDBUF, &bufsize, sizeof(bufsize));
+  setsockopt(fd, SOL_SOCKET, SO_RCVBUF, &bufsize, sizeof(bufsize));
+  printf("client connected\n");
+
+  g_conn = open_tinygpu();
+  if (g_conn == IO_OBJECT_NULL) {
+    fprintf(stderr, "failed to connect to tinygpu driver\n");
+    request_t req; recv(fd, &req, sizeof(req), 0);
+    send_error(fd, "Driver not available. Check: System Report > PCI for GPU, System Settings > Privacy & Security.");
+    return;
+  }
+
+  request_t req;
+  response_t resp;
+  while (recv(fd, &req, sizeof(req), 0) == sizeof(req)) {
+    resp = (response_t){0};
+
+    switch (req.cmd) {
+    case CMD_MAP_BAR:
+      resp.status = map_bar(req.bar, &resp) ? 1 : 0;
+      break;
+
+    case CMD_MAP_SYSMEM_FD: {
+      int shm_fd = -1;
+      resp.status = map_sysmem_fd(req.size, &resp, &shm_fd) ? 1 : 0;
+      send_response(fd, &resp, shm_fd);
+      continue;
+    }
+
+    case CMD_CFG_READ: {
+      uint64_t in[2] = {req.offset, req.size};
+      resp.status = dext_rpc(0, in, 2, &resp.value) ? 1 : 0;
+      break;
+    }
+
+    case CMD_CFG_WRITE: {
+      uint64_t in[3] = {req.offset, req.size, req.value};
+      resp.status = dext_rpc(1, in, 3, NULL) ? 1 : 0;
+      break;
+    }
+
+    case CMD_RESET:
+      resp.status = dext_rpc(2, NULL, 0, NULL) ? 1 : 0;
+      break;
+
+    case CMD_MMIO_READ:
+      if (validate_bar(req.bar, req.offset, req.size)) { resp.status = 1; break; }
+      mmio_copy(g_bulk_buf, (void*)(g_bars[req.bar].addr + req.offset), req.size);
+      resp.value = req.size;
+      send_response(fd, &resp, -1);
+      send(fd, g_bulk_buf, req.size, 0);
+      continue;
+
+    case CMD_MMIO_WRITE:
+      recvall(fd, g_bulk_buf, req.size);
+      if (!validate_bar(req.bar, req.offset, req.size))
+        mmio_copy((void*)(g_bars[req.bar].addr + req.offset), g_bulk_buf, req.size);
+      continue;
+
+    default:
+      resp.status = 1;
+    }
+    send_response(fd, &resp, -1);
+  }
+
+  printf("client disconnected\n");
+  cleanup();
+}
+
+int run_server(const char *sock_path) {
+  int server_fd = socket(AF_UNIX, SOCK_STREAM, 0);
+  if (server_fd < 0) { perror("socket"); return 1; }
+
+  struct sockaddr_un addr = {.sun_family = AF_UNIX};
+  strncpy(addr.sun_path, sock_path, sizeof(addr.sun_path) - 1);
+  unlink(sock_path);
+
+  if (bind(server_fd, (struct sockaddr*)&addr, sizeof(addr)) < 0) { perror("bind"); close(server_fd); return 1; }
+  if (listen(server_fd, 1) < 0) { perror("listen"); close(server_fd); return 1; }
+  printf("listening on %s\n", sock_path);
+
+  while (1) {
+    int client_fd = accept(server_fd, NULL, NULL);
+    if (client_fd < 0) { if (errno == EINTR) continue; perror("accept"); break; }
+    if (g_client_active) { printf("rejected: client already connected\n"); close(client_fd); continue; }
+    g_client_active = 1;
+    handle_client(client_fd);
+    g_client_active = 0;
+    close(client_fd);
+  }
+
+  close(server_fd);
+  unlink(sock_path);
+  cleanup();
+  return 0;
+}
--- a/extra/usbgpu/tbgpu/installer/Shared/tiny_icon.icon/Assets/tiny.svg
+++ b/extra/usbgpu/tbgpu/installer/Shared/tiny_icon.icon/Assets/tiny.svg
@@ -0,0 +1,18 @@
+<svg viewBox="0 0 130 50" xmlns="http://www.w3.org/2000/svg">
+        <!-- t -->
+        <rect x="10" y="0" width="10" height="40" fill="#000000"></rect>
+        <rect x="0" y="10" width="30" height="10" fill="#000000"></rect>
+        <rect x="10" y="30" width="20" height="10" fill="#000000"></rect>
+        <!-- i -->
+        <rect x="40" y="0" width="10" height="10" fill="#000000"></rect>
+        <rect x="40" y="20" width="10" height="20" fill="#000000"></rect>
+        <!-- n -->
+        <rect x="60" y="10" width="10" height="30" fill="#000000"></rect>
+        <rect x="60" y="10" width="20" height="10" fill="#000000"></rect>
+        <rect x="80" y="20" width="10" height="20" fill="#000000"></rect>
+        <!-- y -->
+        <rect x="100" y="10" width="10" height="20" fill="#000000"></rect>
+        <rect x="100" y="20" width="30" height="10" fill="#000000"></rect>
+        <rect x="120" y="10" width="10" height="30" fill="#000000"></rect>
+        <rect x="100" y="40" width="20" height="10" fill="#000000"></rect>
+      </svg>
--- a/extra/usbgpu/tbgpu/installer/Shared/tiny_icon.icon/icon.json
+++ b/extra/usbgpu/tbgpu/installer/Shared/tiny_icon.icon/icon.json
@@ -0,0 +1,52 @@
+{
+  "fill" : "automatic",
+  "groups" : [
+    {
+      "layers" : [
+        {
+          "fill-specializations" : [
+            {
+              "value" : "automatic"
+            },
+            {
+              "appearance" : "dark",
+              "value" : {
+                "solid" : "display-p3:0.94011,0.96611,0.94301,1.00000"
+              }
+            },
+            {
+              "appearance" : "tinted",
+              "value" : {
+                "solid" : "display-p3:0.79528,0.79528,0.79528,1.00000"
+              }
+            }
+          ],
+          "hidden" : false,
+          "image-name" : "tiny.svg",
+          "name" : "tiny_svg",
+          "position" : {
+            "scale" : 5,
+            "translation-in-points" : [
+              0,
+              19
+            ]
+          }
+        }
+      ],
+      "shadow" : {
+        "kind" : "neutral",
+        "opacity" : 0.5
+      },
+      "translucency" : {
+        "enabled" : true,
+        "value" : 0.5
+      }
+    }
+  ],
+  "supported-platforms" : {
+    "circles" : [
+      "watchOS"
+    ],
+    "squares" : "shared"
+  }
+}
--- a/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension.xcodeproj/project.pbxproj
+++ b/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension.xcodeproj/project.pbxproj
@@ -7,13 +7,14 @@
 	objects = {

 /* Begin PBXBuildFile section */
+		0A52E8692F18FC6900A816CD /* tiny_icon.icon in Resources */ = {isa = PBXBuildFile; fileRef = 0A52E8682F18FC6900A816CD /* tiny_icon.icon */; };
+		0A5C11DD2F18E466006DBBCA /* server.c in Sources */ = {isa = PBXBuildFile; fileRef = 0A5C11DC2F18E461006DBBCA /* server.c */; };
 		0ACB55392E9CB880007029EF /* PCIDriverKit.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 0ACB55382E9CB880007029EF /* PCIDriverKit.framework */; };
+		0AD7C2E52F18DEBC00562D1A /* TinyGPUCLIRunner.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AD7C2E42F18DEB800562D1A /* TinyGPUCLIRunner.swift */; };
 		54798269286A3512009785F6 /* CoreAudio.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 54798268286A3512009785F6 /* CoreAudio.framework */; };
-		549EB121286A1A37009D38AB /* TinyGPUViewModel.swift in Sources */ = {isa = PBXBuildFile; fileRef = 549EB11F286A1A37009D38AB /* TinyGPUViewModel.swift */; };
 		549EB123286A1D48009D38AB /* org.tinygrad.tinygpu.edriver.dext in Embed System Extensions */ = {isa = PBXBuildFile; fileRef = C5B7D9BC26128AC50089B4C3 /* org.tinygrad.tinygpu.edriver.dext */; settings = {ATTRIBUTES = (RemoveHeadersOnCopy, ); }; };
 		549EB131286A2B98009D38AB /* IOKit.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549EB130286A2B98009D38AB /* IOKit.framework */; };
 		54E42BC8286A1697000E1E9A /* TinyGPUApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 54E42BB8286A1696000E1E9A /* TinyGPUApp.swift */; };
-		54E42BCA286A1697000E1E9A /* TinyGPUView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 54E42BB9286A1696000E1E9A /* TinyGPUView.swift */; };
 		54E42BCC286A1697000E1E9A /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 54E42BBA286A1697000E1E9A /* Assets.xcassets */; };
 		C5B7D9C326128AC50089B4C3 /* TinyGPUDriver.cpp in Sources */ = {isa = PBXBuildFile; fileRef = C5B7D9C226128AC50089B4C3 /* TinyGPUDriver.cpp */; };
 		C5B7D9C526128AC50089B4C3 /* TinyGPUDriver.iig in Sources */ = {isa = PBXBuildFile; fileRef = C5B7D9C426128AC50089B4C3 /* TinyGPUDriver.iig */; };
@@ -34,6 +35,15 @@
 /* End PBXContainerItemProxy section */

 /* Begin PBXCopyFilesBuildPhase section */
+		0AD7C2DB2F18D7D500562D1A /* CopyFiles */ = {
+			isa = PBXCopyFilesBuildPhase;
+			buildActionMask = 2147483647;
+			dstPath = "";
+			dstSubfolderSpec = 6;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
 		549EB122286A1D3A009D38AB /* Embed System Extensions */ = {
 			isa = PBXCopyFilesBuildPhase;
 			buildActionMask = 2147483647;
@@ -48,13 +58,18 @@
 /* End PBXCopyFilesBuildPhase section */

 /* Begin PBXFileReference section */
+		0A52E8682F18FC6900A816CD /* tiny_icon.icon */ = {isa = PBXFileReference; lastKnownFileType = folder.iconcomposer.icon; path = tiny_icon.icon; sourceTree = "<group>"; };
+		0A5C11DC2F18E461006DBBCA /* server.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; path = server.c; sourceTree = "<group>"; };
+		0A5C11DE2F18E468006DBBCA /* TinyGPU-Bridging-Header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "TinyGPU-Bridging-Header.h"; sourceTree = "<group>"; };
 		0ACB55382E9CB880007029EF /* PCIDriverKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = PCIDriverKit.framework; path = System/DriverKit/System/Library/Frameworks/PCIDriverKit.framework; sourceTree = SDKROOT; };
+		0AD7C2D62F18D3DB00562D1A /* IOKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = IOKit.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX26.2.sdk/System/Library/Frameworks/IOKit.framework; sourceTree = DEVELOPER_DIR; };
+		0AD7C2D82F18D3E300562D1A /* CoreFoundation.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = CoreFoundation.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX26.2.sdk/System/Library/Frameworks/CoreFoundation.framework; sourceTree = DEVELOPER_DIR; };
+		0AD7C2E42F18DEB800562D1A /* TinyGPUCLIRunner.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TinyGPUCLIRunner.swift; sourceTree = "<group>"; };
+		0AFA851D2F1CE486005FDAC2 /* TinyGPUDriver.Release.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = TinyGPUDriver.Release.entitlements; sourceTree = "<group>"; };
 		54798268286A3512009785F6 /* CoreAudio.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = CoreAudio.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX13.0.sdk/System/Library/Frameworks/CoreAudio.framework; sourceTree = DEVELOPER_DIR; };
-		549EB11F286A1A37009D38AB /* TinyGPUViewModel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TinyGPUViewModel.swift; sourceTree = "<group>"; usesTabs = 1; };
 		549EB130286A2B98009D38AB /* IOKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = IOKit.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX13.0.sdk/System/Library/Frameworks/IOKit.framework; sourceTree = DEVELOPER_DIR; };
 		549EB132286A2B9D009D38AB /* IOKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = IOKit.framework; path = Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS16.0.sdk/System/Library/Frameworks/IOKit.framework; sourceTree = DEVELOPER_DIR; };
 		54E42BB8286A1696000E1E9A /* TinyGPUApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TinyGPUApp.swift; sourceTree = "<group>"; };
-		54E42BB9286A1696000E1E9A /* TinyGPUView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TinyGPUView.swift; sourceTree = "<group>"; };
 		54E42BBA286A1697000E1E9A /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
 		54E42BC4286A1697000E1E9A /* TinyGPU.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = TinyGPU.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		54E42BC6286A1697000E1E9A /* macOS.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = macOS.entitlements; sourceTree = "<group>"; };
@@ -102,10 +117,12 @@
 		54E42BB7286A1696000E1E9A /* Shared */ = {
 			isa = PBXGroup;
 			children = (
+				0A5C11DC2F18E461006DBBCA /* server.c */,
+				0AD7C2E42F18DEB800562D1A /* TinyGPUCLIRunner.swift */,
 				54E42BB8286A1696000E1E9A /* TinyGPUApp.swift */,
-				54E42BB9286A1696000E1E9A /* TinyGPUView.swift */,
-				549EB11F286A1A37009D38AB /* TinyGPUViewModel.swift */,
 				54E42BBA286A1697000E1E9A /* Assets.xcassets */,
+				0A52E8682F18FC6900A816CD /* tiny_icon.icon */,
+				0A5C11DE2F18E468006DBBCA /* TinyGPU-Bridging-Header.h */,
 			);
 			path = Shared;
 			sourceTree = "<group>";
@@ -142,6 +159,8 @@
 		C5B7D9BE26128AC50089B4C3 /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
+				0AD7C2D82F18D3E300562D1A /* CoreFoundation.framework */,
+				0AD7C2D62F18D3DB00562D1A /* IOKit.framework */,
 				0ACB55382E9CB880007029EF /* PCIDriverKit.framework */,
 				54798268286A3512009785F6 /* CoreAudio.framework */,
 				549EB130286A2B98009D38AB /* IOKit.framework */,
@@ -168,6 +187,7 @@
 				C5D787AB261667FC006047E5 /* TinyGPUDriverUserClient.iig */,
 				C5B7D9C626128AC50089B4C3 /* Info.plist */,
 				C5B7D9CE26128B150089B4C3 /* TinyGPUDriver.entitlements */,
+				0AFA851D2F1CE486005FDAC2 /* TinyGPUDriver.Release.entitlements */,
 			);
 			path = TinyGPUDriverExtension;
 			sourceTree = "<group>";
@@ -185,21 +205,22 @@
 /* End PBXHeadersBuildPhase section */

 /* Begin PBXNativeTarget section */
-		54E42BC3286A1697000E1E9A /* TinyGPU Installer (macOS) */ = {
+		54E42BC3286A1697000E1E9A /* TinyGPU */ = {
 			isa = PBXNativeTarget;
-			buildConfigurationList = 54E42BD2286A1697000E1E9A /* Build configuration list for PBXNativeTarget "TinyGPU Installer (macOS)" */;
+			buildConfigurationList = 54E42BD2286A1697000E1E9A /* Build configuration list for PBXNativeTarget "TinyGPU" */;
 			buildPhases = (
 				54E42BC0286A1697000E1E9A /* Sources */,
 				54E42BC1286A1697000E1E9A /* Frameworks */,
 				54E42BC2286A1697000E1E9A /* Resources */,
 				549EB122286A1D3A009D38AB /* Embed System Extensions */,
+				0AD7C2DB2F18D7D500562D1A /* CopyFiles */,
 			);
 			buildRules = (
 			);
 			dependencies = (
 				549EB127286A1D66009D38AB /* PBXTargetDependency */,
 			);
-			name = "TinyGPU Installer (macOS)";
+			name = TinyGPU;
 			productName = "SimpleAudioDriverExtension2 (macOS)";
 			productReference = 54E42BC4286A1697000E1E9A /* TinyGPU.app */;
 			productType = "com.apple.product-type.application";
@@ -236,7 +257,7 @@
 				TargetAttributes = {
 					54E42BC3286A1697000E1E9A = {
 						CreatedOnToolsVersion = 14.0;
-						LastSwiftMigration = 1400;
+						LastSwiftMigration = 2620;
 					};
 					C5B7D9BB26128AC50089B4C3 = {
 						CreatedOnToolsVersion = 13.0;
@@ -257,7 +278,7 @@
 			projectRoot = "";
 			targets = (
 				C5B7D9BB26128AC50089B4C3 /* TinyGPUDriver */,
-				54E42BC3286A1697000E1E9A /* TinyGPU Installer (macOS) */,
+				54E42BC3286A1697000E1E9A /* TinyGPU */,
 			);
 		};
 /* End PBXProject section */
@@ -268,6 +289,7 @@
 			buildActionMask = 2147483647;
 			files = (
 				54E42BCC286A1697000E1E9A /* Assets.xcassets in Resources */,
+				0A52E8692F18FC6900A816CD /* tiny_icon.icon in Resources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -285,9 +307,9 @@
 			isa = PBXSourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				549EB121286A1A37009D38AB /* TinyGPUViewModel.swift in Sources */,
-				54E42BCA286A1697000E1E9A /* TinyGPUView.swift in Sources */,
 				54E42BC8286A1697000E1E9A /* TinyGPUApp.swift in Sources */,
+				0A5C11DD2F18E466006DBBCA /* server.c in Sources */,
+				0AD7C2E52F18DEBC00562D1A /* TinyGPUCLIRunner.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -316,12 +338,12 @@
 		54E42BCF286A1697000E1E9A /* Debug */ = {
 			isa = XCBuildConfiguration;
 			buildSettings = {
-				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_APPICON_NAME = tiny_icon;
 				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_ENTITLEMENTS = macOS/macOS.entitlements;
-				CODE_SIGN_IDENTITY = "-";
-				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "-";
+				CODE_SIGN_IDENTITY = "Apple Development";
+				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
 				CURRENT_PROJECT_VERSION = 1;
@@ -331,19 +353,21 @@
 				ENABLE_PREVIEWS = YES;
 				ENABLE_USER_SCRIPT_SANDBOXING = YES;
 				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSHumanReadableCopyright = "";
 				LD_RUNPATH_SEARCH_PATHS = (
 					"$(inherited)",
 					"@executable_path/../Frameworks",
 				);
 				MACOSX_DEPLOYMENT_TARGET = 12.1;
-				MARKETING_VERSION = 1.0;
+				MARKETING_VERSION = 1.0.0;
 				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.installer;
 				PRODUCT_NAME = TinyGPU;
 				PROVISIONING_PROFILE_SPECIFIER = "";
 				SDKROOT = macosx;
 				SWIFT_ACTIVE_COMPILATION_CONDITIONS = DEBUG;
 				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OBJC_BRIDGING_HEADER = "Shared/TinyGPU-Bridging-Header.h";
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 5.0;
 			};
@@ -352,12 +376,12 @@
 		54E42BD0286A1697000E1E9A /* Release */ = {
 			isa = XCBuildConfiguration;
 			buildSettings = {
-				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_APPICON_NAME = tiny_icon;
 				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_ENTITLEMENTS = macOS/macOS.entitlements;
-				CODE_SIGN_IDENTITY = "-";
-				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "-";
+				CODE_SIGN_IDENTITY = "Apple Development";
+				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
 				CURRENT_PROJECT_VERSION = 1;
@@ -367,19 +391,21 @@
 				ENABLE_PREVIEWS = YES;
 				ENABLE_USER_SCRIPT_SANDBOXING = YES;
 				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_KEY_LSApplicationCategoryType = "public.app-category.utilities";
 				INFOPLIST_KEY_NSHumanReadableCopyright = "";
 				LD_RUNPATH_SEARCH_PATHS = (
 					"$(inherited)",
 					"@executable_path/../Frameworks",
 				);
 				MACOSX_DEPLOYMENT_TARGET = 12.1;
-				MARKETING_VERSION = 1.0;
+				MARKETING_VERSION = 1.0.0;
 				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.installer;
 				PRODUCT_NAME = TinyGPU;
 				PROVISIONING_PROFILE_SPECIFIER = "";
 				SDKROOT = macosx;
 				SWIFT_COMPILATION_MODE = wholemodule;
 				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OBJC_BRIDGING_HEADER = "Shared/TinyGPU-Bridging-Header.h";
 				SWIFT_OPTIMIZATION_LEVEL = "-O";
 				SWIFT_VERSION = 5.0;
 			};
@@ -477,6 +503,7 @@
 				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
 				CLANG_WARN_UNREACHABLE_CODE = YES;
 				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				CODE_SIGN_IDENTITY = "Apple Development";
 				COPY_PHASE_STRIP = NO;
 				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
 				DRIVERKIT_DEPLOYMENT_TARGET = 21.0;
@@ -502,26 +529,29 @@
 			buildSettings = {
 				AD_HOC_CODE_SIGNING_ALLOWED = YES;
 				CODE_SIGN_ENTITLEMENTS = TinyGPUDriverExtension/TinyGPUDriver.entitlements;
-				CODE_SIGN_IDENTITY = "-";
+				CODE_SIGN_IDENTITY = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
-				CURRENT_PROJECT_VERSION = 1;
+				CURRENT_PROJECT_VERSION = 2;
 				DEVELOPMENT_TEAM = 9YG3G8543N;
 				DRIVERKIT_DEPLOYMENT_TARGET = 21.0;
 				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				EXCLUDED_ARCHS = "";
 				FRAMEWORK_SEARCH_PATHS = (
 					"$(inherited)",
 					"$(SDKROOT)/System/DriverKit/System/Library/Frameworks",
 				);
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = TinyGPUDriverExtension/Info.plist;
-				INFOPLIST_KEY_OSBundleUsageDescription = "Sample Code Audio Driver Kit Extension";
-				MARKETING_VERSION = 1.0;
+				INFOPLIST_KEY_OSBundleUsageDescription = "TinyGPU Driver";
+				IPHONEOS_DEPLOYMENT_TARGET = 17.6;
+				MARKETING_VERSION = 1.0.0;
 				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.edriver;
 				PRODUCT_NAME = "$(inherited)";
 				PROVISIONING_PROFILE_SPECIFIER = "";
 				RUN_CLANG_STATIC_ANALYZER = YES;
 				SDKROOT = driverkit;
 				SKIP_INSTALL = YES;
+				SUPPORTED_PLATFORMS = driverkit;
 			};
 			name = Debug;
 		};
@@ -529,34 +559,40 @@
 			isa = XCBuildConfiguration;
 			buildSettings = {
 				AD_HOC_CODE_SIGNING_ALLOWED = YES;
-				CODE_SIGN_ENTITLEMENTS = TinyGPUDriverExtension/TinyGPUDriver.entitlements;
-				CODE_SIGN_IDENTITY = "-";
-				CODE_SIGN_STYLE = Automatic;
-				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_TEAM = 9YG3G8543N;
+				CODE_SIGN_ENTITLEMENTS = TinyGPUDriverExtension/TinyGPUDriver.Release.entitlements;
+				CODE_SIGN_IDENTITY = "Apple Development";
+				"CODE_SIGN_IDENTITY[sdk=driverkit*]" = "Apple Development";
+				CODE_SIGN_STYLE = Manual;
+				CURRENT_PROJECT_VERSION = 2;
+				DEVELOPMENT_TEAM = "";
+				"DEVELOPMENT_TEAM[sdk=driverkit*]" = 9YG3G8543N;
 				DRIVERKIT_DEPLOYMENT_TARGET = 21.0;
 				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				EXCLUDED_ARCHS = "";
 				FRAMEWORK_SEARCH_PATHS = (
 					"$(inherited)",
 					"$(SDKROOT)/System/DriverKit/System/Library/Frameworks",
 				);
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_FILE = TinyGPUDriverExtension/Info.plist;
-				INFOPLIST_KEY_OSBundleUsageDescription = "Sample Code Audio Driver Kit Extension";
-				MARKETING_VERSION = 1.0;
+				INFOPLIST_KEY_OSBundleUsageDescription = "TinyGPU Driver";
+				IPHONEOS_DEPLOYMENT_TARGET = 17.6;
+				MARKETING_VERSION = 1.0.0;
+				ONLY_ACTIVE_ARCH = YES;
 				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.edriver;
 				PRODUCT_NAME = "$(inherited)";
 				PROVISIONING_PROFILE_SPECIFIER = "";
 				RUN_CLANG_STATIC_ANALYZER = YES;
 				SDKROOT = driverkit;
 				SKIP_INSTALL = YES;
+				SUPPORTED_PLATFORMS = driverkit;
 			};
 			name = Release;
 		};
 /* End XCBuildConfiguration section */

 /* Begin XCConfigurationList section */
-		54E42BD2286A1697000E1E9A /* Build configuration list for PBXNativeTarget "TinyGPU Installer (macOS)" */ = {
+		54E42BD2286A1697000E1E9A /* Build configuration list for PBXNativeTarget "TinyGPU" */ = {
 			isa = XCConfigurationList;
 			buildConfigurations = (
 				54E42BCF286A1697000E1E9A /* Debug */,
--- a/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.NoSIP.entitlements
+++ b/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.NoSIP.entitlements
@@ -0,0 +1,17 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>com.apple.developer.driverkit</key>
+	<true/>
+	<key>com.apple.developer.driverkit.transport.pci</key>
+	<array>
+		<dict>
+			<key>IOPCIPrimaryMatch</key>
+			<string>0xFFFFFFFF&amp;0x00000000</string>
+		</dict>
+	</array>
+	<key>com.apple.developer.driverkit.allow-any-userclient-access</key>
+	<true/>
+</dict>
+</plist>
--- a/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.Release.entitlements
+++ b/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.Release.entitlements
@@ -0,0 +1,17 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>com.apple.application-identifier</key>
+	<string>9YG3G8543N.org.tinygrad.tinygpu.edriver</string>
+	<key>com.apple.developer.driverkit</key>
+	<true/>
+	<key>com.apple.developer.driverkit.transport.pci</key>
+	<array>
+		<dict>
+			<key>IOPCIPrimaryMatch</key>
+			<string>4098</string>
+		</dict>
+	</array>
+</dict>
+</plist>
--- a/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.entitlements
+++ b/extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.entitlements
@@ -11,7 +11,5 @@
 			<string>0xFFFFFFFF&amp;0x00000000</string>
 		</dict>
 	</array>
-	<key>com.apple.developer.driverkit.allow-any-userclient-access</key>
-	<true/>
 </dict>
 </plist>
--- a/extra/usbgpu/tbgpu/installer/build_and_sign.sh
+++ b/extra/usbgpu/tbgpu/installer/build_and_sign.sh
@@ -0,0 +1,33 @@
+#!/bin/bash
+set -e
+
+xcodebuild clean build CODE_SIGN_IDENTITY="" CODE_SIGNING_REQUIRED=NO -alltargets -configuration Release build
+
+cp "../profiles/devid_provisioning.provisionprofile" "./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext/embedded.provisionprofile"
+cp "../profiles/installer_provisioning.provisionprofile" "./build/Release/TinyGPU.app/Contents/embedded.provisionprofile"
+
+codesign \
+    --sign "Developer ID Application: tinygrad, Corp. (9YG3G8543N)" \
+    --entitlements ./TinyGPUDriverExtension/TinyGPUDriver.Release.entitlements \
+    --verbose \
+    --options runtime \
+    --timestamp \
+    --force \
+    ./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext
+
+codesign \
+    --sign "Developer ID Application: tinygrad, Corp. (9YG3G8543N)" \
+    --entitlements ./macOS/macOS.entitlements \
+    --options runtime \
+    --verbose \
+    --timestamp \
+    --force \
+    ./build/Release/TinyGPU.app
+
+codesign --verify --deep --strict --verbose=4 ./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext
+
+codesign --verify --deep --strict --verbose=4 ./build/Release/TinyGPU.app
+
+spctl -a -vv ./build/Release/TinyGPU.app
+
+spctl -a -vv ./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext
--- a/extra/usbgpu/tbgpu/installer/install_nosip.sh
+++ b/extra/usbgpu/tbgpu/installer/install_nosip.sh
@@ -0,0 +1,52 @@
+#!/bin/bash
+set -e
+
+# Check SIP status if not building only
+if [[ "$1" != "--build" ]]; then
+  SIP_STATUS=$(csrutil status 2>&1)
+  if [[ "$SIP_STATUS" == *"enabled"* ]]; then
+    echo "ERROR: System Integrity Protection (SIP) is enabled."
+    echo "This dev build requires SIP to be disabled to load unsigned dexts."
+    echo ""
+    echo "To disable SIP:"
+    echo "  1. Restart and hold Power button (M1+) or Cmd+R (Intel)"
+    echo "  2. Open Terminal from Recovery menu"
+    echo "  3. Run: csrutil disable"
+    echo "  4. Restart"
+    exit 1
+  fi
+fi
+
+echo "SIP is disabled, proceeding with dev build..."
+
+cd "$(dirname "$0")"
+
+# Build without code signing
+xcodebuild clean build CODE_SIGN_IDENTITY="" CODE_SIGNING_REQUIRED=NO -alltargets -configuration Debug build
+
+APP_PATH="./build/Debug/TinyGPU.app"
+DEXT_PATH="$APP_PATH/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext"
+
+# Ad-hoc sign with dev entitlements (matches any GPU)
+codesign --sign - --entitlements ./TinyGPUDriverExtension/TinyGPUDriver.NoSIP.entitlements --force --timestamp --verbose "$DEXT_PATH"
+codesign --sign - --entitlements ./macOS/macOS.entitlements --force --timestamp --verbose "$APP_PATH"
+
+echo "Build complete: $APP_PATH"
+
+if [[ "$1" == "--build" ]]; then
+  exit 0
+fi
+
+# Install
+echo "Installing to /Applications..."
+
+if [ -d "/Applications/TinyGPU.app" ]; then
+  echo "Removing existing /Applications/TinyGPU.app..."
+  rm -rf "/Applications/TinyGPU.app"
+fi
+
+cp -r "$APP_PATH" /Applications/
+echo "Installed to /Applications/TinyGPU.app"
+
+echo "Activating driver extension..."
+/Applications/TinyGPU.app/Contents/MacOS/TinyGPU install
--- a/extra/usbgpu/tbgpu/installer/macOS/macOS.entitlements
+++ b/extra/usbgpu/tbgpu/installer/macOS/macOS.entitlements
@@ -8,5 +8,9 @@
 	<true/>
 	<key>com.apple.developer.system-extension.install</key>
 	<true/>
+	<key>com.apple.developer.driverkit.userclient-access</key>
+	<array>
+		<string>org.tinygrad.tinygpu.edriver</string>
+	</array>
 </dict>
 </plist>
--- a/extra/usbgpu/tbgpu/installer/notary_tool.sh
+++ b/extra/usbgpu/tbgpu/installer/notary_tool.sh
@@ -0,0 +1,5 @@
+#!/bin/bash
+set -e
+
+ditto -c -k --keepParent ./build/Release/TinyGPU.app ./build/Release/TinyGPU.zip
+xcrun notarytool submit ./build/Release/TinyGPU.zip --keychain-profile "hgwJFhdheiIEy82nDN" --wait
--- a/extra/usbgpu/tbgpu/kill_tinygpu.sh
+++ b/extra/usbgpu/tbgpu/kill_tinygpu.sh
@@ -0,0 +1,2 @@
+#!/bin/bash
+pkill -f "tinygpu.sock"
--- a/extra/usbgpu/tbgpu/main.cpp
+++ b/extra/usbgpu/tbgpu/main.cpp
@@ -1,39 +0,0 @@
-#include <CoreFoundation/CoreFoundation.h>
-#include <IOKit/IOKitLib.h>
-#include <mach/mach.h>
-#include <stdio.h>
-#include <inttypes.h>
-
-static io_connect_t open_uc_by_name(const char *svc_name) {
-  io_connect_t conn = IO_OBJECT_NULL;
-  io_service_t service = IOServiceGetMatchingService(kIOMasterPortDefault, IOServiceNameMatching(svc_name));
-  if (!service) { fprintf(stderr, "service not found: %s\n", svc_name); return IO_OBJECT_NULL; }
-  kern_return_t kr = IOServiceOpen(service, mach_task_self(), /*type*/0, &conn);
-  IOObjectRelease(service);
-  if (kr) { fprintf(stderr, "IOServiceOpen 0x%x\n", kr); return IO_OBJECT_NULL; }
-  return conn;
-}
-
-int main(int argc, char **argv) {
-  uint32_t bar = (argc > 1) ? (uint32_t)strtoul(argv[1], NULL, 0) : 0;  // pick BAR index
-  io_connect_t conn = open_uc_by_name("tinygpu");
-  if (!conn) return 2;
-
-  mach_vm_address_t addr = 0;
-  mach_vm_size_t    size = 0;
-  kern_return_t kr = IOConnectMapMemory64(conn, bar, mach_task_self(), &addr, &size, kIOMapAnywhere);
-  if (kr) { fprintf(stderr, "Map BAR%u failed 0x%x\n", bar, kr); IOServiceClose(conn); return 3; }
-
-  printf("BAR%u mapped at 0x%llx, size 0x%llx\n", bar, (unsigned long long)addr, (unsigned long long)size);
-
-  // example: read a 32-bit register at offset 0x0 (make sure it’s safe!)
-  volatile uint32_t *mmio = (volatile uint32_t*)(uintptr_t)addr;
-  uint32_t v = mmio[0];
-  printf("mmio[0]=0x%08x\n", v);
-
-  kr = IOConnectUnmapMemory64(conn, bar, mach_task_self(), addr);
-  if (kr) fprintf(stderr, "Unmap failed 0x%x\n", kr);
-
-  IOServiceClose(conn);
-  return 0;
-}
--- a/extra/usbgpu/tbgpu/main.py
+++ b/extra/usbgpu/tbgpu/main.py
@@ -1,82 +0,0 @@
-import ctypes, ctypes.util, sys
-
-cf = ctypes.CDLL(ctypes.util.find_library("CoreFoundation"))
-iokit = ctypes.CDLL(ctypes.util.find_library("IOKit"))
-libsys = ctypes.CDLL(ctypes.util.find_library("System"))
-
-kern_return_t = ctypes.c_int
-mach_port_t = ctypes.c_uint
-io_object_t = mach_port_t
-io_service_t = io_object_t
-io_connect_t = mach_port_t
-CFMutableDictionaryRef = ctypes.c_void_p
-CFStringRef = ctypes.c_void_p
-
-kIOMasterPortDefault = mach_port_t(0)
-
-libsys.mach_task_self_.restype = mach_port_t
-
-iokit.IOServiceNameMatching.argtypes = [ctypes.c_char_p]
-iokit.IOServiceNameMatching.restype  = CFMutableDictionaryRef
-
-iokit.IOServiceGetMatchingService.argtypes = [mach_port_t, CFMutableDictionaryRef]
-iokit.IOServiceGetMatchingService.restype  = io_service_t
-
-iokit.IOObjectRelease.argtypes = [io_object_t]
-iokit.IOObjectRelease.restype  = kern_return_t
-
-iokit.IOServiceOpen.argtypes = [io_service_t, mach_port_t, ctypes.c_uint32, ctypes.POINTER(io_connect_t)]
-iokit.IOServiceOpen.restype  = kern_return_t
-
-iokit.IOConnectCallMethod.argtypes = [io_connect_t, ctypes.c_uint32, ctypes.POINTER(ctypes.c_uint64), ctypes.c_uint32, ctypes.c_void_p,
-  ctypes.c_size_t, ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(ctypes.c_uint32), ctypes.c_void_p, ctypes.POINTER(ctypes.c_size_t)]
-iokit.IOConnectCallMethod.restype = kern_return_t
-
-def open_userclient_by_name(name: str, uc_type: int = 0) -> io_connect_t:
-  mdict = iokit.IOServiceNameMatching(name.encode("utf-8"))
-  if not mdict: raise RuntimeError("IOServiceNameMatching returned NULL")
-
-  # Grab the first matching service
-  service = iokit.IOServiceGetMatchingService(kIOMasterPortDefault, mdict)
-  if not service: raise RuntimeError(f'service "{name}" not found')
-
-  # print("lol", service)
-  # print(libsys.mach_task_self_)
-  # cast libsys.mach_task_self_ to uint and print
-  # print("lol", ctypes.cast(libsys.mach_task_self_, ctypes.POINTER(ctypes.c_uint)).contents.value)
-
-  try:
-    # Open user client (type -> passed to NewUserClient_Impl)
-    conn = io_connect_t(0)
-    # print("lol", libsys.mach_task_self_)
-    kr = iokit.IOServiceOpen(service, ctypes.cast(libsys.mach_task_self_, ctypes.POINTER(ctypes.c_uint)).contents.value,
-                             ctypes.c_uint32(uc_type), ctypes.byref(conn))
-    if kr != 0: raise OSError(kr, f"IOServiceOpen failed (0x{kr:08x})")
-    return conn
-  finally: iokit.IOObjectRelease(service)
-
-def external_method(conn: io_connect_t, selector: int = 0) -> int:
-  # no scalars in/out, no struct in/out — just ping selector 0
-  in_scalars = ctypes.POINTER(ctypes.c_uint64)()  # NULL
-  out_scalars = (ctypes.c_uint64 * 1)()       # space if driver returns something
-  out_scalars_cnt = ctypes.c_uint32(0)      # driver can set this
-
-  return iokit.IOConnectCallMethod(conn, ctypes.c_uint32(selector), in_scalars, ctypes.c_uint32(0), None, ctypes.c_size_t(0),
-    out_scalars, ctypes.byref(out_scalars_cnt), None, ctypes.byref(ctypes.c_size_t(0)))
-
-def close_userclient(conn: io_connect_t) -> None:
-  # IOServiceClose is a macro; exported symbol is IOServiceClose in IOKit
-  iokit.IOServiceClose.argtypes = [io_connect_t]
-  iokit.IOServiceClose.restype  = kern_return_t
-  iokit.IOServiceClose(conn)
-
-if __name__ == "__main__":
-  try:
-    conn = open_userclient_by_name("tinygpu", uc_type=0)
-    kr = external_method(conn, selector=0)
-    print(f"ExternalMethod(0) -> 0x{kr:08x}")
-  except Exception as e:
-    print(e)
-    sys.exit(1)
-  finally:
-    if 'conn' in locals() and conn.value: close_userclient(conn)
--- a/tinygrad/runtime/autogen/init.py
+++ b/tinygrad/runtime/autogen/init.py
@@ -141,4 +141,9 @@ def __getattr__(nm):
      return load("metal", "'Metal'", [f"{macossdk}/System/Library/Frameworks/Metal.framework/Headers/MTL{s}.h" for s in
                  ["ComputeCommandEncoder", "ComputePipeline", "CommandQueue", "Device", "IndirectCommandBuffer", "Resource", "CommandEncoder"]],
                  args=["-xobjective-c","-isysroot",macossdk], types={"dispatch_data_t":"objc.id_"})
+    case "iokit": return load("iokit", "'IOKit'", [f"{macossdk}/System/Library/Frameworks/IOKit.framework/Headers/IOKitLib.h"],
+                              args=["-isysroot", macossdk])
+    case "corefoundation": return load("corefoundation", "'CoreFoundation'",
+                                       [f"{macossdk}/System/Library/Frameworks/CoreFoundation.framework/Headers/CF{s}.h" for s in ["String", "Data"]],
+                                       args=["-isysroot", macossdk])
    case _: raise AttributeError(f"no such autogen: {nm}")
--- a/tinygrad/runtime/autogen/corefoundation.py
+++ b/tinygrad/runtime/autogen/corefoundation.py
@@ -0,0 +1,345 @@
+# mypy: disable-error-code="empty-body"
+from __future__ import annotations
+import ctypes
+from typing import Annotated, Literal, TypeAlias
+from tinygrad.runtime.support.c import _IO, _IOW, _IOR, _IOWR
+from tinygrad.runtime.support import c
+dll = c.DLL('corefoundation', 'CoreFoundation')
+CFStringEncoding: TypeAlias = Annotated[int, ctypes.c_uint32]
+CFStringBuiltInEncodings: TypeAlias = Annotated[int, ctypes.c_uint32]
+class _anonenum0(Annotated[int, ctypes.c_uint32], c.Enum): pass
+kCFStringEncodingMacRoman = _anonenum0.define('kCFStringEncodingMacRoman', 0)
+kCFStringEncodingWindowsLatin1 = _anonenum0.define('kCFStringEncodingWindowsLatin1', 1280)
+kCFStringEncodingISOLatin1 = _anonenum0.define('kCFStringEncodingISOLatin1', 513)
+kCFStringEncodingNextStepLatin = _anonenum0.define('kCFStringEncodingNextStepLatin', 2817)
+kCFStringEncodingASCII = _anonenum0.define('kCFStringEncodingASCII', 1536)
+kCFStringEncodingUnicode = _anonenum0.define('kCFStringEncodingUnicode', 256)
+kCFStringEncodingUTF8 = _anonenum0.define('kCFStringEncodingUTF8', 134217984)
+kCFStringEncodingNonLossyASCII = _anonenum0.define('kCFStringEncodingNonLossyASCII', 3071)
+kCFStringEncodingUTF16 = _anonenum0.define('kCFStringEncodingUTF16', 256)
+kCFStringEncodingUTF16BE = _anonenum0.define('kCFStringEncodingUTF16BE', 268435712)
+kCFStringEncodingUTF16LE = _anonenum0.define('kCFStringEncodingUTF16LE', 335544576)
+kCFStringEncodingUTF32 = _anonenum0.define('kCFStringEncodingUTF32', 201326848)
+kCFStringEncodingUTF32BE = _anonenum0.define('kCFStringEncodingUTF32BE', 402653440)
+kCFStringEncodingUTF32LE = _anonenum0.define('kCFStringEncodingUTF32LE', 469762304)
+
+CFTypeID: TypeAlias = Annotated[int, ctypes.c_uint64]
+@dll.bind
+def CFStringGetTypeID() -> CFTypeID: ...
+class struct___CFAllocator(ctypes.Structure): pass
+CFAllocatorRef: TypeAlias = c.POINTER[struct___CFAllocator]
+ConstStr255Param: TypeAlias = c.POINTER[Annotated[int, ctypes.c_ubyte]]
+class struct___CFString(ctypes.Structure): pass
+CFStringRef: TypeAlias = c.POINTER[struct___CFString]
+@dll.bind
+def CFStringCreateWithPascalString(alloc:CFAllocatorRef, pStr:ConstStr255Param, encoding:CFStringEncoding) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateWithCString(alloc:CFAllocatorRef, cStr:c.POINTER[Annotated[bytes, ctypes.c_char]], encoding:CFStringEncoding) -> CFStringRef: ...
+UInt8: TypeAlias = Annotated[int, ctypes.c_ubyte]
+CFIndex: TypeAlias = Annotated[int, ctypes.c_int64]
+Boolean: TypeAlias = Annotated[int, ctypes.c_ubyte]
+@dll.bind
+def CFStringCreateWithBytes(alloc:CFAllocatorRef, bytes:c.POINTER[UInt8], numBytes:CFIndex, encoding:CFStringEncoding, isExternalRepresentation:Boolean) -> CFStringRef: ...
+UniChar: TypeAlias = Annotated[int, ctypes.c_uint16]
+@dll.bind
+def CFStringCreateWithCharacters(alloc:CFAllocatorRef, chars:c.POINTER[UniChar], numChars:CFIndex) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateWithPascalStringNoCopy(alloc:CFAllocatorRef, pStr:ConstStr255Param, encoding:CFStringEncoding, contentsDeallocator:CFAllocatorRef) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateWithCStringNoCopy(alloc:CFAllocatorRef, cStr:c.POINTER[Annotated[bytes, ctypes.c_char]], encoding:CFStringEncoding, contentsDeallocator:CFAllocatorRef) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateWithBytesNoCopy(alloc:CFAllocatorRef, bytes:c.POINTER[UInt8], numBytes:CFIndex, encoding:CFStringEncoding, isExternalRepresentation:Boolean, contentsDeallocator:CFAllocatorRef) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateWithCharactersNoCopy(alloc:CFAllocatorRef, chars:c.POINTER[UniChar], numChars:CFIndex, contentsDeallocator:CFAllocatorRef) -> CFStringRef: ...
+@c.record
+class CFRange(c.Struct):
+  SIZE = 16
+  location: Annotated[CFIndex, 0]
+  length: Annotated[CFIndex, 8]
+@dll.bind
+def CFStringCreateWithSubstring(alloc:CFAllocatorRef, str:CFStringRef, range:CFRange) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateCopy(alloc:CFAllocatorRef, theString:CFStringRef) -> CFStringRef: ...
+class struct___CFDictionary(ctypes.Structure): pass
+CFDictionaryRef: TypeAlias = c.POINTER[struct___CFDictionary]
+@dll.bind
+def CFStringCreateWithFormat(alloc:CFAllocatorRef, formatOptions:CFDictionaryRef, format:CFStringRef) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateWithFormatAndArguments(alloc:CFAllocatorRef, formatOptions:CFDictionaryRef, format:CFStringRef, arguments:Annotated[int, ctypes.c_int32]) -> CFStringRef: ...
+class struct___CFError(ctypes.Structure): pass
+CFErrorRef: TypeAlias = c.POINTER[struct___CFError]
+@dll.bind
+def CFStringCreateStringWithValidatedFormat(alloc:CFAllocatorRef, formatOptions:CFDictionaryRef, validFormatSpecifiers:CFStringRef, format:CFStringRef, errorPtr:c.POINTER[CFErrorRef]) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateStringWithValidatedFormatAndArguments(alloc:CFAllocatorRef, formatOptions:CFDictionaryRef, validFormatSpecifiers:CFStringRef, format:CFStringRef, arguments:Annotated[int, ctypes.c_int32], errorPtr:c.POINTER[CFErrorRef]) -> CFStringRef: ...
+CFMutableStringRef: TypeAlias = c.POINTER[struct___CFString]
+@dll.bind
+def CFStringCreateMutable(alloc:CFAllocatorRef, maxLength:CFIndex) -> CFMutableStringRef: ...
+@dll.bind
+def CFStringCreateMutableCopy(alloc:CFAllocatorRef, maxLength:CFIndex, theString:CFStringRef) -> CFMutableStringRef: ...
+@dll.bind
+def CFStringCreateMutableWithExternalCharactersNoCopy(alloc:CFAllocatorRef, chars:c.POINTER[UniChar], numChars:CFIndex, capacity:CFIndex, externalCharactersAllocator:CFAllocatorRef) -> CFMutableStringRef: ...
+@dll.bind
+def CFStringGetLength(theString:CFStringRef) -> CFIndex: ...
+@dll.bind
+def CFStringGetCharacterAtIndex(theString:CFStringRef, idx:CFIndex) -> UniChar: ...
+@dll.bind
+def CFStringGetCharacters(theString:CFStringRef, range:CFRange, buffer:c.POINTER[UniChar]) -> None: ...
+StringPtr: TypeAlias = c.POINTER[Annotated[int, ctypes.c_ubyte]]
+@dll.bind
+def CFStringGetPascalString(theString:CFStringRef, buffer:StringPtr, bufferSize:CFIndex, encoding:CFStringEncoding) -> Boolean: ...
+@dll.bind
+def CFStringGetCString(theString:CFStringRef, buffer:c.POINTER[Annotated[bytes, ctypes.c_char]], bufferSize:CFIndex, encoding:CFStringEncoding) -> Boolean: ...
+ConstStringPtr: TypeAlias = c.POINTER[Annotated[int, ctypes.c_ubyte]]
+@dll.bind
+def CFStringGetPascalStringPtr(theString:CFStringRef, encoding:CFStringEncoding) -> ConstStringPtr: ...
+@dll.bind
+def CFStringGetCStringPtr(theString:CFStringRef, encoding:CFStringEncoding) -> c.POINTER[Annotated[bytes, ctypes.c_char]]: ...
+@dll.bind
+def CFStringGetCharactersPtr(theString:CFStringRef) -> c.POINTER[UniChar]: ...
+@dll.bind
+def CFStringGetBytes(theString:CFStringRef, range:CFRange, encoding:CFStringEncoding, lossByte:UInt8, isExternalRepresentation:Boolean, buffer:c.POINTER[UInt8], maxBufLen:CFIndex, usedBufLen:c.POINTER[CFIndex]) -> CFIndex: ...
+class struct___CFData(ctypes.Structure): pass
+CFDataRef: TypeAlias = c.POINTER[struct___CFData]
+@dll.bind
+def CFStringCreateFromExternalRepresentation(alloc:CFAllocatorRef, data:CFDataRef, encoding:CFStringEncoding) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateExternalRepresentation(alloc:CFAllocatorRef, theString:CFStringRef, encoding:CFStringEncoding, lossByte:UInt8) -> CFDataRef: ...
+@dll.bind
+def CFStringGetSmallestEncoding(theString:CFStringRef) -> CFStringEncoding: ...
+@dll.bind
+def CFStringGetFastestEncoding(theString:CFStringRef) -> CFStringEncoding: ...
+@dll.bind
+def CFStringGetSystemEncoding() -> CFStringEncoding: ...
+@dll.bind
+def CFStringGetMaximumSizeForEncoding(length:CFIndex, encoding:CFStringEncoding) -> CFIndex: ...
+@dll.bind
+def CFStringGetFileSystemRepresentation(string:CFStringRef, buffer:c.POINTER[Annotated[bytes, ctypes.c_char]], maxBufLen:CFIndex) -> Boolean: ...
+@dll.bind
+def CFStringGetMaximumSizeOfFileSystemRepresentation(string:CFStringRef) -> CFIndex: ...
+@dll.bind
+def CFStringCreateWithFileSystemRepresentation(alloc:CFAllocatorRef, buffer:c.POINTER[Annotated[bytes, ctypes.c_char]]) -> CFStringRef: ...
+CFStringCompareFlags: TypeAlias = Annotated[int, ctypes.c_uint64]
+class _anonenum1(Annotated[int, ctypes.c_uint32], c.Enum): pass
+kCFCompareCaseInsensitive = _anonenum1.define('kCFCompareCaseInsensitive', 1)
+kCFCompareBackwards = _anonenum1.define('kCFCompareBackwards', 4)
+kCFCompareAnchored = _anonenum1.define('kCFCompareAnchored', 8)
+kCFCompareNonliteral = _anonenum1.define('kCFCompareNonliteral', 16)
+kCFCompareLocalized = _anonenum1.define('kCFCompareLocalized', 32)
+kCFCompareNumerically = _anonenum1.define('kCFCompareNumerically', 64)
+kCFCompareDiacriticInsensitive = _anonenum1.define('kCFCompareDiacriticInsensitive', 128)
+kCFCompareWidthInsensitive = _anonenum1.define('kCFCompareWidthInsensitive', 256)
+kCFCompareForcedOrdering = _anonenum1.define('kCFCompareForcedOrdering', 512)
+
+class struct___CFLocale(ctypes.Structure): pass
+CFLocaleRef: TypeAlias = c.POINTER[struct___CFLocale]
+CFComparisonResult: TypeAlias = Annotated[int, ctypes.c_int64]
+@dll.bind
+def CFStringCompareWithOptionsAndLocale(theString1:CFStringRef, theString2:CFStringRef, rangeToCompare:CFRange, compareOptions:CFStringCompareFlags, locale:CFLocaleRef) -> CFComparisonResult: ...
+@dll.bind
+def CFStringCompareWithOptions(theString1:CFStringRef, theString2:CFStringRef, rangeToCompare:CFRange, compareOptions:CFStringCompareFlags) -> CFComparisonResult: ...
+@dll.bind
+def CFStringCompare(theString1:CFStringRef, theString2:CFStringRef, compareOptions:CFStringCompareFlags) -> CFComparisonResult: ...
+@dll.bind
+def CFStringFindWithOptionsAndLocale(theString:CFStringRef, stringToFind:CFStringRef, rangeToSearch:CFRange, searchOptions:CFStringCompareFlags, locale:CFLocaleRef, result:c.POINTER[CFRange]) -> Boolean: ...
+@dll.bind
+def CFStringFindWithOptions(theString:CFStringRef, stringToFind:CFStringRef, rangeToSearch:CFRange, searchOptions:CFStringCompareFlags, result:c.POINTER[CFRange]) -> Boolean: ...
+class struct___CFArray(ctypes.Structure): pass
+CFArrayRef: TypeAlias = c.POINTER[struct___CFArray]
+@dll.bind
+def CFStringCreateArrayWithFindResults(alloc:CFAllocatorRef, theString:CFStringRef, stringToFind:CFStringRef, rangeToSearch:CFRange, compareOptions:CFStringCompareFlags) -> CFArrayRef: ...
+@dll.bind
+def CFStringFind(theString:CFStringRef, stringToFind:CFStringRef, compareOptions:CFStringCompareFlags) -> CFRange: ...
+@dll.bind
+def CFStringHasPrefix(theString:CFStringRef, prefix:CFStringRef) -> Boolean: ...
+@dll.bind
+def CFStringHasSuffix(theString:CFStringRef, suffix:CFStringRef) -> Boolean: ...
+@dll.bind
+def CFStringGetRangeOfComposedCharactersAtIndex(theString:CFStringRef, theIndex:CFIndex) -> CFRange: ...
+class struct___CFCharacterSet(ctypes.Structure): pass
+CFCharacterSetRef: TypeAlias = c.POINTER[struct___CFCharacterSet]
+@dll.bind
+def CFStringFindCharacterFromSet(theString:CFStringRef, theSet:CFCharacterSetRef, rangeToSearch:CFRange, searchOptions:CFStringCompareFlags, result:c.POINTER[CFRange]) -> Boolean: ...
+@dll.bind
+def CFStringGetLineBounds(theString:CFStringRef, range:CFRange, lineBeginIndex:c.POINTER[CFIndex], lineEndIndex:c.POINTER[CFIndex], contentsEndIndex:c.POINTER[CFIndex]) -> None: ...
+@dll.bind
+def CFStringGetParagraphBounds(string:CFStringRef, range:CFRange, parBeginIndex:c.POINTER[CFIndex], parEndIndex:c.POINTER[CFIndex], contentsEndIndex:c.POINTER[CFIndex]) -> None: ...
+CFOptionFlags: TypeAlias = Annotated[int, ctypes.c_uint64]
+UTF32Char: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def CFStringGetHyphenationLocationBeforeIndex(string:CFStringRef, location:CFIndex, limitRange:CFRange, options:CFOptionFlags, locale:CFLocaleRef, character:c.POINTER[UTF32Char]) -> CFIndex: ...
+@dll.bind
+def CFStringIsHyphenationAvailableForLocale(locale:CFLocaleRef) -> Boolean: ...
+@dll.bind
+def CFStringCreateByCombiningStrings(alloc:CFAllocatorRef, theArray:CFArrayRef, separatorString:CFStringRef) -> CFStringRef: ...
+@dll.bind
+def CFStringCreateArrayBySeparatingStrings(alloc:CFAllocatorRef, theString:CFStringRef, separatorString:CFStringRef) -> CFArrayRef: ...
+SInt32: TypeAlias = Annotated[int, ctypes.c_int32]
+@dll.bind
+def CFStringGetIntValue(str:CFStringRef) -> SInt32: ...
+@dll.bind
+def CFStringGetDoubleValue(str:CFStringRef) -> Annotated[float, ctypes.c_double]: ...
+@dll.bind
+def CFStringAppend(theString:CFMutableStringRef, appendedString:CFStringRef) -> None: ...
+@dll.bind
+def CFStringAppendCharacters(theString:CFMutableStringRef, chars:c.POINTER[UniChar], numChars:CFIndex) -> None: ...
+@dll.bind
+def CFStringAppendPascalString(theString:CFMutableStringRef, pStr:ConstStr255Param, encoding:CFStringEncoding) -> None: ...
+@dll.bind
+def CFStringAppendCString(theString:CFMutableStringRef, cStr:c.POINTER[Annotated[bytes, ctypes.c_char]], encoding:CFStringEncoding) -> None: ...
+@dll.bind
+def CFStringAppendFormat(theString:CFMutableStringRef, formatOptions:CFDictionaryRef, format:CFStringRef) -> None: ...
+@dll.bind
+def CFStringAppendFormatAndArguments(theString:CFMutableStringRef, formatOptions:CFDictionaryRef, format:CFStringRef, arguments:Annotated[int, ctypes.c_int32]) -> None: ...
+@dll.bind
+def CFStringInsert(str:CFMutableStringRef, idx:CFIndex, insertedStr:CFStringRef) -> None: ...
+@dll.bind
+def CFStringDelete(theString:CFMutableStringRef, range:CFRange) -> None: ...
+@dll.bind
+def CFStringReplace(theString:CFMutableStringRef, range:CFRange, replacement:CFStringRef) -> None: ...
+@dll.bind
+def CFStringReplaceAll(theString:CFMutableStringRef, replacement:CFStringRef) -> None: ...
+@dll.bind
+def CFStringFindAndReplace(theString:CFMutableStringRef, stringToFind:CFStringRef, replacementString:CFStringRef, rangeToSearch:CFRange, compareOptions:CFStringCompareFlags) -> CFIndex: ...
+@dll.bind
+def CFStringSetExternalCharactersNoCopy(theString:CFMutableStringRef, chars:c.POINTER[UniChar], length:CFIndex, capacity:CFIndex) -> None: ...
+@dll.bind
+def CFStringPad(theString:CFMutableStringRef, padString:CFStringRef, length:CFIndex, indexIntoPad:CFIndex) -> None: ...
+@dll.bind
+def CFStringTrim(theString:CFMutableStringRef, trimString:CFStringRef) -> None: ...
+@dll.bind
+def CFStringTrimWhitespace(theString:CFMutableStringRef) -> None: ...
+@dll.bind
+def CFStringLowercase(theString:CFMutableStringRef, locale:CFLocaleRef) -> None: ...
+@dll.bind
+def CFStringUppercase(theString:CFMutableStringRef, locale:CFLocaleRef) -> None: ...
+@dll.bind
+def CFStringCapitalize(theString:CFMutableStringRef, locale:CFLocaleRef) -> None: ...
+CFStringNormalizationForm: TypeAlias = Annotated[int, ctypes.c_int64]
+class _anonenum2(Annotated[int, ctypes.c_uint32], c.Enum): pass
+kCFStringNormalizationFormD = _anonenum2.define('kCFStringNormalizationFormD', 0)
+kCFStringNormalizationFormKD = _anonenum2.define('kCFStringNormalizationFormKD', 1)
+kCFStringNormalizationFormC = _anonenum2.define('kCFStringNormalizationFormC', 2)
+kCFStringNormalizationFormKC = _anonenum2.define('kCFStringNormalizationFormKC', 3)
+
+@dll.bind
+def CFStringNormalize(theString:CFMutableStringRef, theForm:CFStringNormalizationForm) -> None: ...
+@dll.bind
+def CFStringFold(theString:CFMutableStringRef, theFlags:CFStringCompareFlags, theLocale:CFLocaleRef) -> None: ...
+@dll.bind
+def CFStringTransform(string:CFMutableStringRef, range:c.POINTER[CFRange], transform:CFStringRef, reverse:Boolean) -> Boolean: ...
+try: kCFStringTransformStripCombiningMarks = CFStringRef.in_dll(dll, 'kCFStringTransformStripCombiningMarks') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformToLatin = CFStringRef.in_dll(dll, 'kCFStringTransformToLatin') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformFullwidthHalfwidth = CFStringRef.in_dll(dll, 'kCFStringTransformFullwidthHalfwidth') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinKatakana = CFStringRef.in_dll(dll, 'kCFStringTransformLatinKatakana') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinHiragana = CFStringRef.in_dll(dll, 'kCFStringTransformLatinHiragana') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformHiraganaKatakana = CFStringRef.in_dll(dll, 'kCFStringTransformHiraganaKatakana') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformMandarinLatin = CFStringRef.in_dll(dll, 'kCFStringTransformMandarinLatin') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinHangul = CFStringRef.in_dll(dll, 'kCFStringTransformLatinHangul') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinArabic = CFStringRef.in_dll(dll, 'kCFStringTransformLatinArabic') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinHebrew = CFStringRef.in_dll(dll, 'kCFStringTransformLatinHebrew') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinThai = CFStringRef.in_dll(dll, 'kCFStringTransformLatinThai') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinCyrillic = CFStringRef.in_dll(dll, 'kCFStringTransformLatinCyrillic') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformLatinGreek = CFStringRef.in_dll(dll, 'kCFStringTransformLatinGreek') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformToXMLHex = CFStringRef.in_dll(dll, 'kCFStringTransformToXMLHex') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformToUnicodeName = CFStringRef.in_dll(dll, 'kCFStringTransformToUnicodeName') # type: ignore
+except (ValueError,AttributeError): pass
+try: kCFStringTransformStripDiacritics = CFStringRef.in_dll(dll, 'kCFStringTransformStripDiacritics') # type: ignore
+except (ValueError,AttributeError): pass
+@dll.bind
+def CFStringIsEncodingAvailable(encoding:CFStringEncoding) -> Boolean: ...
+@dll.bind
+def CFStringGetListOfAvailableEncodings() -> c.POINTER[CFStringEncoding]: ...
+@dll.bind
+def CFStringGetNameOfEncoding(encoding:CFStringEncoding) -> CFStringRef: ...
+@dll.bind
+def CFStringConvertEncodingToNSStringEncoding(encoding:CFStringEncoding) -> Annotated[int, ctypes.c_uint64]: ...
+@dll.bind
+def CFStringConvertNSStringEncodingToEncoding(encoding:Annotated[int, ctypes.c_uint64]) -> CFStringEncoding: ...
+UInt32: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def CFStringConvertEncodingToWindowsCodepage(encoding:CFStringEncoding) -> UInt32: ...
+@dll.bind
+def CFStringConvertWindowsCodepageToEncoding(codepage:UInt32) -> CFStringEncoding: ...
+@dll.bind
+def CFStringConvertIANACharSetNameToEncoding(theString:CFStringRef) -> CFStringEncoding: ...
+@dll.bind
+def CFStringConvertEncodingToIANACharSetName(encoding:CFStringEncoding) -> CFStringRef: ...
+@dll.bind
+def CFStringGetMostCompatibleMacStringEncoding(encoding:CFStringEncoding) -> CFStringEncoding: ...
+@c.record
+class CFStringInlineBuffer(c.Struct):
+  SIZE = 184
+  buffer: Annotated[c.Array[UniChar, Literal[64]], 0]
+  theString: Annotated[CFStringRef, 128]
+  directUniCharBuffer: Annotated[c.POINTER[UniChar], 136]
+  directCStringBuffer: Annotated[c.POINTER[Annotated[bytes, ctypes.c_char]], 144]
+  rangeToBuffer: Annotated[CFRange, 152]
+  bufferedRangeStart: Annotated[CFIndex, 168]
+  bufferedRangeEnd: Annotated[CFIndex, 176]
+CFTypeRef: TypeAlias = ctypes.c_void_p
+@dll.bind
+def CFShow(obj:CFTypeRef) -> None: ...
+@dll.bind
+def CFShowStr(str:CFStringRef) -> None: ...
+@dll.bind
+def __CFStringMakeConstantString(cStr:c.POINTER[Annotated[bytes, ctypes.c_char]]) -> CFStringRef: ...
+CFMutableDataRef: TypeAlias = c.POINTER[struct___CFData]
+@dll.bind
+def CFDataGetTypeID() -> CFTypeID: ...
+@dll.bind
+def CFDataCreate(allocator:CFAllocatorRef, bytes:c.POINTER[UInt8], length:CFIndex) -> CFDataRef: ...
+@dll.bind
+def CFDataCreateWithBytesNoCopy(allocator:CFAllocatorRef, bytes:c.POINTER[UInt8], length:CFIndex, bytesDeallocator:CFAllocatorRef) -> CFDataRef: ...
+@dll.bind
+def CFDataCreateCopy(allocator:CFAllocatorRef, theData:CFDataRef) -> CFDataRef: ...
+@dll.bind
+def CFDataCreateMutable(allocator:CFAllocatorRef, capacity:CFIndex) -> CFMutableDataRef: ...
+@dll.bind
+def CFDataCreateMutableCopy(allocator:CFAllocatorRef, capacity:CFIndex, theData:CFDataRef) -> CFMutableDataRef: ...
+@dll.bind
+def CFDataGetLength(theData:CFDataRef) -> CFIndex: ...
+@dll.bind
+def CFDataGetBytePtr(theData:CFDataRef) -> c.POINTER[UInt8]: ...
+@dll.bind
+def CFDataGetMutableBytePtr(theData:CFMutableDataRef) -> c.POINTER[UInt8]: ...
+@dll.bind
+def CFDataGetBytes(theData:CFDataRef, range:CFRange, buffer:c.POINTER[UInt8]) -> None: ...
+@dll.bind
+def CFDataSetLength(theData:CFMutableDataRef, length:CFIndex) -> None: ...
+@dll.bind
+def CFDataIncreaseLength(theData:CFMutableDataRef, extraLength:CFIndex) -> None: ...
+@dll.bind
+def CFDataAppendBytes(theData:CFMutableDataRef, bytes:c.POINTER[UInt8], length:CFIndex) -> None: ...
+@dll.bind
+def CFDataReplaceBytes(theData:CFMutableDataRef, range:CFRange, newBytes:c.POINTER[UInt8], newLength:CFIndex) -> None: ...
+@dll.bind
+def CFDataDeleteBytes(theData:CFMutableDataRef, range:CFRange) -> None: ...
+CFDataSearchFlags: TypeAlias = Annotated[int, ctypes.c_uint64]
+class _anonenum3(Annotated[int, ctypes.c_uint32], c.Enum): pass
+kCFDataSearchBackwards = _anonenum3.define('kCFDataSearchBackwards', 1)
+kCFDataSearchAnchored = _anonenum3.define('kCFDataSearchAnchored', 2)
+
+@dll.bind
+def CFDataFind(theData:CFDataRef, dataToFind:CFDataRef, searchRange:CFRange, compareOptions:CFDataSearchFlags) -> CFRange: ...
+c.init_records()
+__COREFOUNDATION_CFSTRING__ = 1 # type: ignore
+kCFStringEncodingInvalidId = (0xffffffff) # type: ignore
+CF_FORMAT_FUNCTION = lambda F,A: __attribute__((format(CFString, F, A))) # type: ignore
+CF_FORMAT_ARGUMENT = lambda A: __attribute__((format_arg(A))) # type: ignore
+__kCFStringInlineBufferLength = 64 # type: ignore
+__COREFOUNDATION_CFDATA__ = 1 # type: ignore
--- a/tinygrad/runtime/autogen/iokit.py
+++ b/tinygrad/runtime/autogen/iokit.py
@@ -0,0 +1,279 @@
+# mypy: disable-error-code="empty-body"
+from __future__ import annotations
+import ctypes
+from typing import Annotated, Literal, TypeAlias
+from tinygrad.runtime.support.c import _IO, _IOW, _IOR, _IOWR
+from tinygrad.runtime.support import c
+dll = c.DLL('iokit', 'IOKit')
+class struct_IONotificationPort(ctypes.Structure): pass
+IONotificationPortRef: TypeAlias = c.POINTER[struct_IONotificationPort]
+IOServiceMatchingCallback: TypeAlias = c.CFUNCTYPE[None, [ctypes.c_void_p, Annotated[int, ctypes.c_uint32]]]
+IOServiceInterestCallback: TypeAlias = c.CFUNCTYPE[None, [ctypes.c_void_p, Annotated[int, ctypes.c_uint32], Annotated[int, ctypes.c_uint32], ctypes.c_void_p]]
+mach_port_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+try: kIOMainPortDefault = mach_port_t.in_dll(dll, 'kIOMainPortDefault') # type: ignore
+except (ValueError,AttributeError): pass
+kern_return_t: TypeAlias = Annotated[int, ctypes.c_int32]
+@dll.bind
+def IOMainPort(bootstrapPort:mach_port_t, mainPort:c.POINTER[mach_port_t]) -> kern_return_t: ...
+try: kIOMasterPortDefault = mach_port_t.in_dll(dll, 'kIOMasterPortDefault') # type: ignore
+except (ValueError,AttributeError): pass
+@dll.bind
+def IOMasterPort(bootstrapPort:mach_port_t, mainPort:c.POINTER[mach_port_t]) -> kern_return_t: ...
+@dll.bind
+def IONotificationPortCreate(mainPort:mach_port_t) -> IONotificationPortRef: ...
+@dll.bind
+def IONotificationPortDestroy(notify:IONotificationPortRef) -> None: ...
+class struct___CFRunLoopSource(ctypes.Structure): pass
+CFRunLoopSourceRef: TypeAlias = c.POINTER[struct___CFRunLoopSource]
+@dll.bind
+def IONotificationPortGetRunLoopSource(notify:IONotificationPortRef) -> CFRunLoopSourceRef: ...
+@dll.bind
+def IONotificationPortGetMachPort(notify:IONotificationPortRef) -> mach_port_t: ...
+@dll.bind
+def IONotificationPortSetImportanceReceiver(notify:IONotificationPortRef) -> kern_return_t: ...
+class struct_dispatch_queue_s(ctypes.Structure): pass
+dispatch_queue_t: TypeAlias = c.POINTER[struct_dispatch_queue_s]
+@dll.bind
+def IONotificationPortSetDispatchQueue(notify:IONotificationPortRef, queue:dispatch_queue_t) -> None: ...
+@c.record
+class mach_msg_header_t(c.Struct):
+  SIZE = 24
+  msgh_bits: Annotated[mach_msg_bits_t, 0]
+  msgh_size: Annotated[mach_msg_size_t, 4]
+  msgh_remote_port: Annotated[mach_port_t, 8]
+  msgh_local_port: Annotated[mach_port_t, 12]
+  msgh_voucher_port: Annotated[mach_port_name_t, 16]
+  msgh_id: Annotated[mach_msg_id_t, 20]
+mach_msg_bits_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+mach_msg_size_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+mach_port_name_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+mach_msg_id_t: TypeAlias = Annotated[int, ctypes.c_int32]
+@dll.bind
+def IODispatchCalloutFromMessage(unused:ctypes.c_void_p, msg:c.POINTER[mach_msg_header_t], reference:ctypes.c_void_p) -> None: ...
+uint32_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def IOCreateReceivePort(msgType:uint32_t, recvPort:c.POINTER[mach_port_t]) -> kern_return_t: ...
+io_object_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def IOObjectRelease(object:io_object_t) -> kern_return_t: ...
+@dll.bind
+def IOObjectRetain(object:io_object_t) -> kern_return_t: ...
+io_name_t: TypeAlias = c.Array[Annotated[bytes, ctypes.c_char], Literal[128]]
+@dll.bind
+def IOObjectGetClass(object:io_object_t, className:io_name_t) -> kern_return_t: ...
+class struct___CFString(ctypes.Structure): pass
+CFStringRef: TypeAlias = c.POINTER[struct___CFString]
+@dll.bind
+def IOObjectCopyClass(object:io_object_t) -> CFStringRef: ...
+@dll.bind
+def IOObjectCopySuperclassForClass(classname:CFStringRef) -> CFStringRef: ...
+@dll.bind
+def IOObjectCopyBundleIdentifierForClass(classname:CFStringRef) -> CFStringRef: ...
+boolean_t: TypeAlias = Annotated[int, ctypes.c_int32]
+@dll.bind
+def IOObjectConformsTo(object:io_object_t, className:io_name_t) -> boolean_t: ...
+@dll.bind
+def IOObjectIsEqualTo(object:io_object_t, anObject:io_object_t) -> boolean_t: ...
+@dll.bind
+def IOObjectGetKernelRetainCount(object:io_object_t) -> uint32_t: ...
+@dll.bind
+def IOObjectGetUserRetainCount(object:io_object_t) -> uint32_t: ...
+@dll.bind
+def IOObjectGetRetainCount(object:io_object_t) -> uint32_t: ...
+io_iterator_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def IOIteratorNext(iterator:io_iterator_t) -> io_object_t: ...
+@dll.bind
+def IOIteratorReset(iterator:io_iterator_t) -> None: ...
+@dll.bind
+def IOIteratorIsValid(iterator:io_iterator_t) -> boolean_t: ...
+class struct___CFDictionary(ctypes.Structure): pass
+CFDictionaryRef: TypeAlias = c.POINTER[struct___CFDictionary]
+io_service_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def IOServiceGetMatchingService(mainPort:mach_port_t, matching:CFDictionaryRef) -> io_service_t: ...
+@dll.bind
+def IOServiceGetMatchingServices(mainPort:mach_port_t, matching:CFDictionaryRef, existing:c.POINTER[io_iterator_t]) -> kern_return_t: ...
+uintptr_t: TypeAlias = Annotated[int, ctypes.c_uint64]
+@dll.bind
+def IOServiceAddNotification(mainPort:mach_port_t, notificationType:io_name_t, matching:CFDictionaryRef, wakePort:mach_port_t, reference:uintptr_t, notification:c.POINTER[io_iterator_t]) -> kern_return_t: ...
+@dll.bind
+def IOServiceAddMatchingNotification(notifyPort:IONotificationPortRef, notificationType:io_name_t, matching:CFDictionaryRef, callback:IOServiceMatchingCallback, refCon:ctypes.c_void_p, notification:c.POINTER[io_iterator_t]) -> kern_return_t: ...
+@dll.bind
+def IOServiceAddInterestNotification(notifyPort:IONotificationPortRef, service:io_service_t, interestType:io_name_t, callback:IOServiceInterestCallback, refCon:ctypes.c_void_p, notification:c.POINTER[io_object_t]) -> kern_return_t: ...
+@dll.bind
+def IOServiceMatchPropertyTable(service:io_service_t, matching:CFDictionaryRef, matches:c.POINTER[boolean_t]) -> kern_return_t: ...
+@dll.bind
+def IOServiceGetBusyState(service:io_service_t, busyState:c.POINTER[uint32_t]) -> kern_return_t: ...
+@c.record
+class struct_mach_timespec(c.Struct):
+  SIZE = 8
+  tv_sec: Annotated[Annotated[int, ctypes.c_uint32], 0]
+  tv_nsec: Annotated[clock_res_t, 4]
+mach_timespec_t: TypeAlias = struct_mach_timespec
+clock_res_t: TypeAlias = Annotated[int, ctypes.c_int32]
+@dll.bind
+def IOServiceWaitQuiet(service:io_service_t, waitTime:c.POINTER[mach_timespec_t]) -> kern_return_t: ...
+@dll.bind
+def IOKitGetBusyState(mainPort:mach_port_t, busyState:c.POINTER[uint32_t]) -> kern_return_t: ...
+IOOptionBits: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def IOKitWaitQuietWithOptions(mainPort:mach_port_t, waitTime:c.POINTER[mach_timespec_t], options:IOOptionBits) -> kern_return_t: ...
+@dll.bind
+def IOKitWaitQuiet(mainPort:mach_port_t, waitTime:c.POINTER[mach_timespec_t]) -> kern_return_t: ...
+task_port_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+io_connect_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def IOServiceOpen(service:io_service_t, owningTask:task_port_t, type:uint32_t, connect:c.POINTER[io_connect_t]) -> kern_return_t: ...
+@dll.bind
+def IOServiceRequestProbe(service:io_service_t, options:uint32_t) -> kern_return_t: ...
+class _anonenum0(Annotated[int, ctypes.c_uint32], c.Enum): pass
+kIOServiceInteractionAllowed = _anonenum0.define('kIOServiceInteractionAllowed', 1)
+
+@dll.bind
+def IOServiceAuthorize(service:io_service_t, options:uint32_t) -> kern_return_t: ...
+@dll.bind
+def IOServiceOpenAsFileDescriptor(service:io_service_t, oflag:Annotated[int, ctypes.c_int32]) -> Annotated[int, ctypes.c_int32]: ...
+@dll.bind
+def IOServiceClose(connect:io_connect_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectAddRef(connect:io_connect_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectRelease(connect:io_connect_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectGetService(connect:io_connect_t, service:c.POINTER[io_service_t]) -> kern_return_t: ...
+@dll.bind
+def IOConnectSetNotificationPort(connect:io_connect_t, type:uint32_t, port:mach_port_t, reference:uintptr_t) -> kern_return_t: ...
+mach_vm_address_t: TypeAlias = Annotated[int, ctypes.c_uint64]
+mach_vm_size_t: TypeAlias = Annotated[int, ctypes.c_uint64]
+@dll.bind
+def IOConnectMapMemory(connect:io_connect_t, memoryType:uint32_t, intoTask:task_port_t, atAddress:c.POINTER[mach_vm_address_t], ofSize:c.POINTER[mach_vm_size_t], options:IOOptionBits) -> kern_return_t: ...
+@dll.bind
+def IOConnectMapMemory64(connect:io_connect_t, memoryType:uint32_t, intoTask:task_port_t, atAddress:c.POINTER[mach_vm_address_t], ofSize:c.POINTER[mach_vm_size_t], options:IOOptionBits) -> kern_return_t: ...
+@dll.bind
+def IOConnectUnmapMemory(connect:io_connect_t, memoryType:uint32_t, fromTask:task_port_t, atAddress:mach_vm_address_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectUnmapMemory64(connect:io_connect_t, memoryType:uint32_t, fromTask:task_port_t, atAddress:mach_vm_address_t) -> kern_return_t: ...
+CFTypeRef: TypeAlias = ctypes.c_void_p
+@dll.bind
+def IOConnectSetCFProperties(connect:io_connect_t, properties:CFTypeRef) -> kern_return_t: ...
+@dll.bind
+def IOConnectSetCFProperty(connect:io_connect_t, propertyName:CFStringRef, property:CFTypeRef) -> kern_return_t: ...
+uint64_t: TypeAlias = Annotated[int, ctypes.c_uint64]
+size_t: TypeAlias = Annotated[int, ctypes.c_uint64]
+@dll.bind
+def IOConnectCallMethod(connection:mach_port_t, selector:uint32_t, input:c.POINTER[uint64_t], inputCnt:uint32_t, inputStruct:ctypes.c_void_p, inputStructCnt:size_t, output:c.POINTER[uint64_t], outputCnt:c.POINTER[uint32_t], outputStruct:ctypes.c_void_p, outputStructCnt:c.POINTER[size_t]) -> kern_return_t: ...
+@dll.bind
+def IOConnectCallAsyncMethod(connection:mach_port_t, selector:uint32_t, wake_port:mach_port_t, reference:c.POINTER[uint64_t], referenceCnt:uint32_t, input:c.POINTER[uint64_t], inputCnt:uint32_t, inputStruct:ctypes.c_void_p, inputStructCnt:size_t, output:c.POINTER[uint64_t], outputCnt:c.POINTER[uint32_t], outputStruct:ctypes.c_void_p, outputStructCnt:c.POINTER[size_t]) -> kern_return_t: ...
+@dll.bind
+def IOConnectCallStructMethod(connection:mach_port_t, selector:uint32_t, inputStruct:ctypes.c_void_p, inputStructCnt:size_t, outputStruct:ctypes.c_void_p, outputStructCnt:c.POINTER[size_t]) -> kern_return_t: ...
+@dll.bind
+def IOConnectCallAsyncStructMethod(connection:mach_port_t, selector:uint32_t, wake_port:mach_port_t, reference:c.POINTER[uint64_t], referenceCnt:uint32_t, inputStruct:ctypes.c_void_p, inputStructCnt:size_t, outputStruct:ctypes.c_void_p, outputStructCnt:c.POINTER[size_t]) -> kern_return_t: ...
+@dll.bind
+def IOConnectCallScalarMethod(connection:mach_port_t, selector:uint32_t, input:c.POINTER[uint64_t], inputCnt:uint32_t, output:c.POINTER[uint64_t], outputCnt:c.POINTER[uint32_t]) -> kern_return_t: ...
+@dll.bind
+def IOConnectCallAsyncScalarMethod(connection:mach_port_t, selector:uint32_t, wake_port:mach_port_t, reference:c.POINTER[uint64_t], referenceCnt:uint32_t, input:c.POINTER[uint64_t], inputCnt:uint32_t, output:c.POINTER[uint64_t], outputCnt:c.POINTER[uint32_t]) -> kern_return_t: ...
+@dll.bind
+def IOConnectTrap0(connect:io_connect_t, index:uint32_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectTrap1(connect:io_connect_t, index:uint32_t, p1:uintptr_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectTrap2(connect:io_connect_t, index:uint32_t, p1:uintptr_t, p2:uintptr_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectTrap3(connect:io_connect_t, index:uint32_t, p1:uintptr_t, p2:uintptr_t, p3:uintptr_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectTrap4(connect:io_connect_t, index:uint32_t, p1:uintptr_t, p2:uintptr_t, p3:uintptr_t, p4:uintptr_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectTrap5(connect:io_connect_t, index:uint32_t, p1:uintptr_t, p2:uintptr_t, p3:uintptr_t, p4:uintptr_t, p5:uintptr_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectTrap6(connect:io_connect_t, index:uint32_t, p1:uintptr_t, p2:uintptr_t, p3:uintptr_t, p4:uintptr_t, p5:uintptr_t, p6:uintptr_t) -> kern_return_t: ...
+@dll.bind
+def IOConnectAddClient(connect:io_connect_t, client:io_connect_t) -> kern_return_t: ...
+io_registry_entry_t: TypeAlias = Annotated[int, ctypes.c_uint32]
+@dll.bind
+def IORegistryGetRootEntry(mainPort:mach_port_t) -> io_registry_entry_t: ...
+io_string_t: TypeAlias = c.Array[Annotated[bytes, ctypes.c_char], Literal[512]]
+@dll.bind
+def IORegistryEntryFromPath(mainPort:mach_port_t, path:io_string_t) -> io_registry_entry_t: ...
+@dll.bind
+def IORegistryEntryCopyFromPath(mainPort:mach_port_t, path:CFStringRef) -> io_registry_entry_t: ...
+class _anonenum1(Annotated[int, ctypes.c_uint32], c.Enum): pass
+kIORegistryIterateRecursively = _anonenum1.define('kIORegistryIterateRecursively', 1)
+kIORegistryIterateParents = _anonenum1.define('kIORegistryIterateParents', 2)
+
+@dll.bind
+def IORegistryCreateIterator(mainPort:mach_port_t, plane:io_name_t, options:IOOptionBits, iterator:c.POINTER[io_iterator_t]) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryCreateIterator(entry:io_registry_entry_t, plane:io_name_t, options:IOOptionBits, iterator:c.POINTER[io_iterator_t]) -> kern_return_t: ...
+@dll.bind
+def IORegistryIteratorEnterEntry(iterator:io_iterator_t) -> kern_return_t: ...
+@dll.bind
+def IORegistryIteratorExitEntry(iterator:io_iterator_t) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetName(entry:io_registry_entry_t, name:io_name_t) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetNameInPlane(entry:io_registry_entry_t, plane:io_name_t, name:io_name_t) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetLocationInPlane(entry:io_registry_entry_t, plane:io_name_t, location:io_name_t) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetPath(entry:io_registry_entry_t, plane:io_name_t, path:io_string_t) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryCopyPath(entry:io_registry_entry_t, plane:io_name_t) -> CFStringRef: ...
+@dll.bind
+def IORegistryEntryGetRegistryEntryID(entry:io_registry_entry_t, entryID:c.POINTER[uint64_t]) -> kern_return_t: ...
+CFMutableDictionaryRef: TypeAlias = c.POINTER[struct___CFDictionary]
+class struct___CFAllocator(ctypes.Structure): pass
+CFAllocatorRef: TypeAlias = c.POINTER[struct___CFAllocator]
+@dll.bind
+def IORegistryEntryCreateCFProperties(entry:io_registry_entry_t, properties:c.POINTER[CFMutableDictionaryRef], allocator:CFAllocatorRef, options:IOOptionBits) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryCreateCFProperty(entry:io_registry_entry_t, key:CFStringRef, allocator:CFAllocatorRef, options:IOOptionBits) -> CFTypeRef: ...
+@dll.bind
+def IORegistryEntrySearchCFProperty(entry:io_registry_entry_t, plane:io_name_t, key:CFStringRef, allocator:CFAllocatorRef, options:IOOptionBits) -> CFTypeRef: ...
+io_struct_inband_t: TypeAlias = c.Array[Annotated[bytes, ctypes.c_char], Literal[4096]]
+@dll.bind
+def IORegistryEntryGetProperty(entry:io_registry_entry_t, propertyName:io_name_t, buffer:io_struct_inband_t, size:c.POINTER[uint32_t]) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntrySetCFProperties(entry:io_registry_entry_t, properties:CFTypeRef) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntrySetCFProperty(entry:io_registry_entry_t, propertyName:CFStringRef, property:CFTypeRef) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetChildIterator(entry:io_registry_entry_t, plane:io_name_t, iterator:c.POINTER[io_iterator_t]) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetChildEntry(entry:io_registry_entry_t, plane:io_name_t, child:c.POINTER[io_registry_entry_t]) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetParentIterator(entry:io_registry_entry_t, plane:io_name_t, iterator:c.POINTER[io_iterator_t]) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryGetParentEntry(entry:io_registry_entry_t, plane:io_name_t, parent:c.POINTER[io_registry_entry_t]) -> kern_return_t: ...
+@dll.bind
+def IORegistryEntryInPlane(entry:io_registry_entry_t, plane:io_name_t) -> boolean_t: ...
+@dll.bind
+def IOServiceMatching(name:c.POINTER[Annotated[bytes, ctypes.c_char]]) -> CFMutableDictionaryRef: ...
+@dll.bind
+def IOServiceNameMatching(name:c.POINTER[Annotated[bytes, ctypes.c_char]]) -> CFMutableDictionaryRef: ...
+@dll.bind
+def IOBSDNameMatching(mainPort:mach_port_t, options:uint32_t, bsdName:c.POINTER[Annotated[bytes, ctypes.c_char]]) -> CFMutableDictionaryRef: ...
+@dll.bind
+def IOOpenFirmwarePathMatching(mainPort:mach_port_t, options:uint32_t, path:c.POINTER[Annotated[bytes, ctypes.c_char]]) -> CFMutableDictionaryRef: ...
+@dll.bind
+def IORegistryEntryIDMatching(entryID:uint64_t) -> CFMutableDictionaryRef: ...
+@dll.bind
+def IOServiceOFPathToBSDName(mainPort:mach_port_t, openFirmwarePath:io_name_t, bsdName:io_name_t) -> kern_return_t: ...
+IOAsyncCallback0: TypeAlias = c.CFUNCTYPE[None, [ctypes.c_void_p, Annotated[int, ctypes.c_int32]]]
+IOAsyncCallback1: TypeAlias = c.CFUNCTYPE[None, [ctypes.c_void_p, Annotated[int, ctypes.c_int32], ctypes.c_void_p]]
+IOAsyncCallback2: TypeAlias = c.CFUNCTYPE[None, [ctypes.c_void_p, Annotated[int, ctypes.c_int32], ctypes.c_void_p, ctypes.c_void_p]]
+IOAsyncCallback: TypeAlias = c.CFUNCTYPE[None, [ctypes.c_void_p, Annotated[int, ctypes.c_int32], c.POINTER[ctypes.c_void_p], Annotated[int, ctypes.c_uint32]]]
+vm_size_t: TypeAlias = Annotated[int, ctypes.c_uint64]
+@dll.bind
+def OSGetNotificationFromMessage(msg:c.POINTER[mach_msg_header_t], index:uint32_t, type:c.POINTER[uint32_t], reference:c.POINTER[uintptr_t], content:c.POINTER[ctypes.c_void_p], size:c.POINTER[vm_size_t]) -> kern_return_t: ...
+@dll.bind
+def IOCatalogueSendData(mainPort:mach_port_t, flag:uint32_t, buffer:c.POINTER[Annotated[bytes, ctypes.c_char]], size:uint32_t) -> kern_return_t: ...
+@dll.bind
+def IOCatalogueTerminate(mainPort:mach_port_t, flag:uint32_t, description:io_name_t) -> kern_return_t: ...
+@dll.bind
+def IOCatalogueGetData(mainPort:mach_port_t, flag:uint32_t, buffer:c.POINTER[c.POINTER[Annotated[bytes, ctypes.c_char]]], size:c.POINTER[uint32_t]) -> kern_return_t: ...
+@dll.bind
+def IOCatalogueModuleLoaded(mainPort:mach_port_t, name:io_name_t) -> kern_return_t: ...
+@dll.bind
+def IOCatalogueReset(mainPort:mach_port_t, flag:uint32_t) -> kern_return_t: ...
+c.init_records()
--- a/tinygrad/runtime/support/nv/ip.py
+++ b/tinygrad/runtime/support/nv/ip.py
@@ -539,7 +539,7 @@ class NV_GSP(NV_IP):
    self.stat_q.wait_resp(nv.NV_VGPU_MSG_FUNCTION_SET_PAGE_DIRECTORY)

  def rpc_set_gsp_system_info(self):
-    def bdf_as_int(s): return 0x000 if s.startswith("usb") else (int(s[5:7],16)<<8) | (int(s[8:10],16)<<3) | int(s[-1],16)
+    def bdf_as_int(s): return 0x000 if s.startswith("usb") or s.startswith("remote") else (int(s[5:7],16)<<8) | (int(s[8:10],16)<<3) | int(s[-1],16)

    pcidev = self.nvdev.pci_dev
    data = nv.GspSystemInfo(gpuPhysAddr=pcidev.bar_info[0].addr, gpuPhysFbAddr=pcidev.bar_info[1].addr, gpuPhysInstAddr=pcidev.bar_info[3].addr,
--- a/tinygrad/runtime/support/system.py
+++ b/tinygrad/runtime/support/system.py
@@ -1,7 +1,8 @@
-import os, mmap, array, functools, ctypes, select, contextlib, dataclasses, sys, itertools
+from __future__ import annotations
+import os, mmap, array, functools, ctypes, select, contextlib, dataclasses, sys, itertools, struct, socket, subprocess, time, enum
 from typing import ClassVar
-from tinygrad.helpers import round_up, getenv, OSX, temp, ceildiv
-from tinygrad.runtime.autogen import libc, vfio, pci
+from tinygrad.helpers import round_up, getenv, OSX, temp, ceildiv, unwrap, fetch, system
+from tinygrad.runtime.autogen import libc, pci, vfio, iokit, corefoundation
 from tinygrad.runtime.support.hcq import FileIOInterface, MMIOInterface, HCQBuffer, hcq_filter_visible_devices
 from tinygrad.runtime.support.memory import MemoryManager, VirtMapping, AddrSpace
 from tinygrad.runtime.support.usb import ASM24Controller, USBMMIOInterface
@@ -20,15 +21,9 @@ class _System:
  @functools.cached_property
  def atomic_lib(self): return ctypes.CDLL(ctypes.util.find_library('atomic')) if sys.platform == "linux" else None

-  @functools.cached_property
-  def iokit(self): return ctypes.CDLL(ctypes.util.find_library("IOKit"))
-
  @functools.cached_property
  def libsys(self): return ctypes.CDLL(ctypes.util.find_library("System"))

-  @functools.cached_property
-  def mach_task_self(self): return ctypes.cast(self.libsys.mach_task_self_, ctypes.POINTER(ctypes.c_uint)).contents.value
-
  @functools.cached_property
  def pagemap(self) -> FileIOInterface:
    self.write_sysfs("/proc/sys/vm/compact_unevictable_allowed", "0", "Failed to disable migration of locked pages")
@@ -46,28 +41,6 @@ class _System:
      return vfio_fd
    except OSError: return None

-  @functools.cached_property
-  def macos_tinygpu_conn(self):
-    self.iokit.IOServiceNameMatching.restype = ctypes.c_void_p # CFMutableDictionaryRef
-    if not (mdict:=self.iokit.IOServiceNameMatching("tinygpu".encode("utf-8"))): raise RuntimeError("IOServiceNameMatching returned NULL")
-    if not (service:=self.iokit.IOServiceGetMatchingService(ctypes.c_uint(0), ctypes.c_void_p(mdict))):
-      raise RuntimeError('Service "tinygpu" is not running')
-    if self.iokit.IOServiceOpen(service, self.mach_task_self, ctypes.c_uint32(0), ctypes.byref(conn:=ctypes.c_uint(0))):
-      raise RuntimeError("IOServiceOpen failed")
-    return conn
-
-  def iokit_pci_memmap(self, typ:int):
-    if self.iokit.IOConnectMapMemory64(self.macos_tinygpu_conn, ctypes.c_uint32(typ), System.mach_task_self,
-      ctypes.byref(addr:=ctypes.c_uint64(0)), ctypes.byref(size:=ctypes.c_uint64(0)), 0x1): raise RuntimeError(f"IOConnectMapMemory64({typ=}) failed")
-    return MMIOInterface(addr.value, size.value)
-
-  def iokit_pci_rpc(self, sel:int, *args:int):
-    in_scalars = (ctypes.c_uint64 * len(args))(*args) if args else ctypes.POINTER(ctypes.c_uint64)()
-    if (self.iokit.IOConnectCallMethod(self.macos_tinygpu_conn, sel, in_scalars, len(args), None, ctypes.c_size_t(0),
-        out_scalars:=(ctypes.c_uint64*16)(), ctypes.byref(outcnt:=ctypes.c_uint32(16)), None, ctypes.byref(ctypes.c_size_t(0)))):
-      raise RuntimeError(f"IOConnectCallMethod({sel=}, {args=}) failed")
-    return out_scalars[:outcnt.value]
-
  def reserve_hugepages(self, cnt): os.system(f"sudo sh -c 'echo {cnt} > /proc/sys/vm/nr_hugepages'")

  def memory_barrier(self): lib.atomic_thread_fence(__ATOMIC_SEQ_CST:=5) if (lib:=self.libsys if OSX else self.atomic_lib) is not None else None
@@ -79,14 +52,24 @@ class _System:
    self.pagemap.seek(vaddr // mmap.PAGESIZE * 8)
    return [(x & ((1<<55) - 1)) * mmap.PAGESIZE for x in array.array('Q', self.pagemap.read(size//mmap.PAGESIZE*8, binary=True))]

-  def pci_scan_bus(self, target_vendor:int, target_devices:list[tuple[int, list[int]]], base_class:int|None=None) -> list[str]:
-    result = []
-    for pcibus in FileIOInterface("/sys/bus/pci/devices").listdir():
-      vendor = int(FileIOInterface(f"/sys/bus/pci/devices/{pcibus}/vendor").read(), 16)
-      device = int(FileIOInterface(f"/sys/bus/pci/devices/{pcibus}/device").read(), 16)
-      if base_class is not None and int(FileIOInterface(f"/sys/bus/pci/devices/{pcibus}/class").read(), 16) >> 16 != base_class: continue
-      if vendor == target_vendor and any((device & mask) in devlist for mask, devlist in target_devices): result.append(pcibus)
-    return sorted(result)
+  def pci_scan_bus(self, vendor:int, devices:list[tuple[int, list[int]]], base_class:int|None=None) -> list[str]:
+    all_devs = []
+    if OSX:
+      def read_prop(svc, key) -> int:
+        cfkey = corefoundation.CFStringCreateWithCString(None, key.encode(), corefoundation.kCFStringEncodingUTF8)
+        cfdata = ctypes.cast(iokit.IORegistryEntryCreateCFProperty(svc, ctypes.cast(cfkey, iokit.CFStringRef), None, 0), corefoundation.CFDataRef)
+        corefoundation.CFDataGetBytes(cfdata, corefoundation.CFRange(0, corefoundation.CFDataGetLength(cfdata)), buf:=(ctypes.c_uint8*8)())
+        return int.from_bytes(bytes(buf), "little")
+
+      iokit.IOServiceGetMatchingServices(0, iokit.IOServiceMatching(b"IOPCIDevice"), ctypes.byref(iterator:=ctypes.c_uint()))
+      while svc:=iokit.IOIteratorNext(iterator): all_devs.append((v:=read_prop(svc, "vendor-id"), d:=read_prop(svc, "device-id"), f"{v:x}:{d:x}"))
+    else:
+      for pcibus in FileIOInterface("/sys/bus/pci/devices").listdir():
+        if base_class is not None and int(FileIOInterface(f"/sys/bus/pci/devices/{pcibus}/class").read(), 16) >> 16 != base_class: continue
+        all_devs.append((int(FileIOInterface(f"/sys/bus/pci/devices/{pcibus}/vendor").read(), 16),
+                         int(FileIOInterface(f"/sys/bus/pci/devices/{pcibus}/device").read(), 16), pcibus))
+
+    return sorted([val for vendor, device, val in all_devs if vendor == vendor and any((device & mask) in devlist for mask, devlist in devices)])

  def pci_setup_usb_bars(self, usb:ASM24Controller, gpu_bus:int, mem_base:int, pref_mem_base:int) -> dict[int, PCIBarInfo]:
    for bus in range(gpu_bus):
@@ -154,6 +137,8 @@ class _System:

 System = _System()

+# *** PCI Devices
+
 class PCIDevice:
  def __init__(self, devpref:str, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
    self.lock_fd = System.flock_acquire(f"{devpref.lower()}_{pcibus.lower()}.lock")
@@ -210,21 +195,6 @@ class PCIDevice:
    return MMIOInterface(loc, sz, fmt=fmt)
  def reset(self): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/devices/{self.pcibus}/reset'")

-class APLPCIDevice(PCIDevice):
-  def __init__(self, devpref:str, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
-    self.lock_fd = System.flock_acquire(f"{devpref.lower()}_{pcibus.lower()}.lock")
-    self.pcibus, self.bars = pcibus, {b: System.iokit_pci_memmap(b) for b in bars}
-    self.bar_info = {b:PCIBarInfo(0, self.bars[b].nbytes-1 if b in self.bars else 0) for b in range(6)} # NOTE: fake bar info for nv.
-  def alloc_sysmem(self, size:int, vaddr:int=0, contiguous:bool=False) -> tuple[MMIOInterface, list[int]]:
-    sysmem_view = System.iokit_pci_memmap(round_up(size, mmap.PAGESIZE))
-    paddrs = list(itertools.takewhile(lambda p: p[1] != 0, zip(sysmem_view.view(fmt='Q')[0::2], sysmem_view.view(fmt='Q')[1::2])))
-    assert not contiguous or len(paddrs) == 1, "not contiguous, but required"
-    return sysmem_view, [p + i for p, sz in paddrs for i in range(0, sz, 0x1000)][:ceildiv(size, 0x1000)]
-  def map_bar(self, bar:int, off:int=0, addr:int=0, size:int|None=None, fmt='B') -> MMIOInterface: return self.bars[bar].view(off, size, fmt)
-  def read_config(self, offset:int, size:int): return System.iokit_pci_rpc(__TinyGPURPCReadCfg:=0, offset, size)[0]
-  def write_config(self, offset:int, value:int, size:int): System.iokit_pci_rpc(__TinyGPURPCWriteCfg:=1, offset, size, value)
-  def reset(self): System.iokit_pci_rpc(__TinyGPURPCReset:=2)
-
 class USBPCIDevice(PCIDevice):
  def __init__(self, devpref:str, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
    self.lock_fd = System.flock_acquire(f"{devpref.lower()}_{pcibus.lower()}.lock")
@@ -286,10 +256,101 @@ class LNXPCIIfaceBase:

    self.dev_impl.mm.map_range(int(b.va_addr), round_up(b.size, 0x1000), paddrs, aspace=aspace, snooped=snooped, uncached=uncached)

-class APLPCIIfaceBase(LNXPCIIfaceBase):
-  def __init__(self, dev, dev_id, vendor, devices, bars, vram_bar, va_start, va_size, base_class:int|None=None):
-    self.pci_dev, self.dev, self.vram_bar = APLPCIDevice(dev.__class__.__name__[:2], pcibus=f'usb4:{dev_id}', bars=bars), dev, vram_bar
-    assert (read_vendor:=self.pci_dev.read_config(pci.PCI_VENDOR_ID, 2)) == vendor, f"Vendor ID mismatch: expected {vendor:#x}, got {read_vendor:#x}"
-  def map(self, b:HCQBuffer): raise RuntimeError(f"map failed: {b.owner} -> {self.dev}")
+# *** Remote PCI Devices

-PCIIfaceBase:type = APLPCIIfaceBase if OSX else LNXPCIIfaceBase
+class RemoteCmd(enum.IntEnum): MAP_BAR, MAP_SYSMEM_FD, CFG_READ, CFG_WRITE, RESET, MMIO_READ, MMIO_WRITE = 1, 2, 3, 4, 5, 6, 7
+
+class RemoteMMIOInterface(MMIOInterface):
+  def __init__(self, dev:RemotePCIDevice, residx:int, nbytes:int, fmt='B', off=0):
+    self.dev, self.residx, self.nbytes, self.fmt, self.off, self.el_sz = dev, residx, nbytes, fmt, off, struct.calcsize(fmt)
+
+  def __getitem__(self, index):
+    sl = index if isinstance(index, slice) else slice(index, index + 1)
+    start, stop = (sl.start or 0) * self.el_sz, (sl.stop or len(self)) * self.el_sz
+    data = self.dev._bulk_read(RemoteCmd.MMIO_READ, self.residx, self.off + start, stop - start)
+    result = data if self.fmt == 'B' else list(struct.unpack(f'<{(stop - start) // self.el_sz}{self.fmt}', data))
+    return result if isinstance(index, slice) else result[0]
+
+  def __setitem__(self, index, val):
+    start = (index.start or 0) * self.el_sz if isinstance(index, slice) else index * self.el_sz
+    data = (val if self.fmt == 'B' else struct.pack(f'<{len(val)}{self.fmt}', *val)) if isinstance(index, slice) else struct.pack(f'<{self.fmt}', val)
+    self.dev._bulk_write(RemoteCmd.MMIO_WRITE, self.residx, self.off + start, data)
+
+  def view(self, offset:int=0, size:int|None=None, fmt=None):
+    return RemoteMMIOInterface(self.dev, self.residx, size or (self.nbytes - offset), fmt or self.fmt, self.off + offset)
+
+class RemotePCIDevice(PCIDevice):
+  def __init__(self, devpref:str, pcibus:str, bars:list[int], sock:socket.socket):
+    self.lock_fd = System.flock_acquire(f"{devpref.lower()}_{pcibus.lower()}.lock")
+    self.pcibus, self.sock = pcibus, sock
+    for buft in [socket.SO_SNDBUF, socket.SO_RCVBUF]: self.sock.setsockopt(socket.SOL_SOCKET, buft, 64 << 20)
+    self.bar_info = {b: PCIBarInfo(0, self._rpc(RemoteCmd.MAP_BAR, b)[0]) for b in bars}
+
+  def _recvall(self, n:int) -> bytes:
+    data = b''
+    while len(data) < n and (chunk:=self.sock.recv(n - len(data))): data += chunk
+    if len(data) < n: raise RuntimeError("Connection closed")
+    return data
+
+  def _recv_with_fd(self) -> tuple[bytes, int|None]:
+    msg, anc, _, _ = self.sock.recvmsg(17, socket.CMSG_LEN(4))
+    return msg, struct.unpack('<i', anc[0][2][:4])[0]
+
+  def _rpc(self, cmd:int, *args:int, readout_size:int=0, has_fd=False) -> tuple[int, int, bytes|None, int|None]:
+    self.sock.sendall(struct.pack('<BBQQQ', cmd, *(*args, 0, 0, 0, 0)[:4]))
+    msg, fd = self._recv_with_fd() if has_fd else (self._recvall(17), None)
+    if (resp:=struct.unpack('<BQQ', msg))[0] != 0:
+      raise RuntimeError(f"RPC failed: {self._recvall(resp[1]).decode('utf-8') if resp[1] > 0 else 'unknown error'}")
+    return (resp[1], resp[2]) + ((self._recvall(readout_size) if readout_size > 0 else None),) + (fd,)
+
+  def _bulk_read(self, cmd:int, idx:int, offset:int, size:int) -> bytes: return unwrap(self._rpc(cmd, idx, offset, size, readout_size=size)[2])
+  def _bulk_write(self, cmd:int, idx:int, offset:int, data:bytes): self.sock.sendall(struct.pack('<BBQQQ', cmd, idx, offset, len(data), 0) + data)
+
+  def alloc_sysmem(self, size:int, vaddr:int=0, contiguous:bool=False) -> tuple[MMIOInterface, list[int]]:
+    mapped_size, _, _, fd = self._rpc(RemoteCmd.MAP_SYSMEM_FD, 0, 0, size, has_fd=True)
+    memview = MMIOInterface(FileIOInterface(fd=fd).mmap(0, mapped_size, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED, 0), mapped_size, fmt='B')
+
+    # paddrs are returned as (paddr, size) pairs until a (paddr=0, size=0) terminator in the beginning of the mapping.
+    paddrs_raw = list(itertools.takewhile(lambda p: p[1] != 0, zip(memview.view(fmt='Q')[0::2], memview.view(fmt='Q')[1::2])))
+    return memview, [p + i for p, sz in paddrs_raw for i in range(0, sz, 0x1000)][:ceildiv(size, 0x1000)]
+  def read_config(self, offset:int, size:int): return self._rpc(RemoteCmd.CFG_READ, 0, offset, size)[0]
+  def write_config(self, offset:int, value:int, size:int): self._rpc(RemoteCmd.CFG_WRITE, 0, offset, size, value)
+  def reset(self): self._rpc(RemoteCmd.RESET, 0, 0, 0)
+  def map_bar(self, bar:int, off:int=0, addr:int=0, size:int|None=None, fmt='B') -> MMIOInterface:
+    return RemoteMMIOInterface(self, bar, size or self.bar_info[bar].size, fmt).view(off, size, fmt)
+
+class APLRemotePCIDevice(RemotePCIDevice):
+  APP_PATH = "/Applications/TinyGPU.app/Contents/MacOS/TinyGPU"
+
+  @staticmethod
+  def install_tinygpu():
+    print("Downloading TinyGPU.app...")
+    system(f"ditto -xk {fetch('https://github.com/nimlgen/tinygpu_releases/raw/8120b5508b43149d27bf22f9a4e6d7c5a4b401e9/TinyGPU.zip')} /Applications")
+    print(system(f"{APLRemotePCIDevice.APP_PATH} install"))
+
+  def __init__(self, devpref:str, pcibus:str, bars:list[int], resize_bars:list[int]|None=None):
+    sock_path, sock = getenv("APL_REMOTE_SOCK", temp("tinygpu.sock")), socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    for i in range(100):
+      with contextlib.suppress(ConnectionRefusedError, FileNotFoundError):
+        sock.connect(sock_path)
+        break
+      if i == 0: subprocess.Popen([self.APP_PATH, "server", sock_path], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+      time.sleep(0.05)
+    else: raise RuntimeError(f"Failed to connect to TinyGPU server at {sock_path}.")
+    super().__init__(devpref, pcibus, bars, sock)
+
+class APLRemoteIfaceBase(LNXPCIIfaceBase):
+  def __init__(self, dev, dev_id, vendor, devices:list[tuple[int, list[int]]], bars, vram_bar, va_start, va_size, base_class:int|None=None):
+    if not (cls:=type(self)).gpus:
+      cls.gpus = System.pci_scan_bus(vendor, devices, base_class)
+      if not cls.gpus: raise RuntimeError("No supported GPUs found")
+      if not os.path.exists(APLRemotePCIDevice.APP_PATH): APLRemotePCIDevice.install_tinygpu()
+    self.pci_dev = APLRemotePCIDevice(dev.__class__.__name__[:2], f'remote:{dev_id}', bars)
+    self.dev, self.vram_bar = dev, vram_bar
+
+  def free(self, b:HCQBuffer):
+    for dev in b.mapped_devs[1:]: dev.iface.dev_impl.mm.unmap_range(b.va_addr, b.size)
+
+  def map(self, b:HCQBuffer): raise RuntimeError(f"P2P mapping not supported for remote devices: {b.owner} -> {self.dev}")
+
+PCIIfaceBase:type = APLRemoteIfaceBase if OSX else LNXPCIIfaceBase