A high-performance inference engine for AI models
use raw_msg_send
eugene0575151f9d79
diff --git a/crates/uzu/src/backends/metal/metal_extensions/device_extensions.rs b/crates/uzu/src/backends/metal/metal_extensions/device_extensions.rsindex 3186c3eb9..8eb91bbb2 100644--- a/crates/uzu/src/backends/metal/metal_extensions/device_extensions.rs+++ b/crates/uzu/src/backends/metal/metal_extensions/device_extensions.rs@@ -1,13 +1,33 @@+use core::mem::transmute;+use bytesize::ByteSize;use metal::MTLDevice;use objc2::{- Message, msg_send,+ Message,+ ffi::objc_msgSend,+ msg_send,rc::Retained,- runtime::{NSObjectProtocol, ProtocolObject},+ runtime::{NSObjectProtocol, ProtocolObject, Sel},sel,};use objc2_foundation::NSString;+// Used to bypass objc2's debug-mode class validation for proxy objects like+// `CaptureMTLDevice` which forward messages at the ObjC runtime level (via+// `forwardingTargetForSelector:`) but don't have the methods in their own class+// table — causing `msg_send!`'s debug check to panic even though the message+// would succeed.+//+// The caller must verify with `respondsToSelector:` first.+unsafe fn raw_msg_send<T: Message + ?Sized, R>(+ obj: &T,+ sel: Sel,+) -> R {+ let object_pointer: *const T = obj;+ let send: unsafe extern "C" fn(*const T, Sel) -> R = unsafe { transmute(objc_msgSend as *const ()) };+ unsafe { send(object_pointer, sel) }+}+#[derive(Debug, Clone, Copy, PartialEq, Eq)]pub enum DeviceGeneration {Gen13,@@ -49,8 +69,8 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Human-readable chip name, e.g. "M2 Max", "M4 Pro", "A17 Pro".fn family_name(&self) -> String {if self.respondsToSelector(sel!(familyName)) {- let ns: Retained<NSString> = unsafe { msg_send![self, familyName] };- ns.to_string()+ let family_name: Retained<NSString> = unsafe { msg_send![self, familyName] };+ family_name.to_string()} else {"Unknown".into()}@@ -59,7 +79,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Number of GPU shader cores.fn gpu_core_count(&self) -> u32 {if self.respondsToSelector(sel!(gpuCoreCount)) {- unsafe { msg_send![self, gpuCoreCount] }+ unsafe { raw_msg_send(self, sel!(gpuCoreCount)) }} else {8}@@ -68,7 +88,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Total unified (shared) memory.fn shared_memory_size(&self) -> ByteSize {if self.respondsToSelector(sel!(sharedMemorySize)) {- ByteSize(unsafe { msg_send![self, sharedMemorySize] })+ ByteSize(unsafe { raw_msg_send(self, sel!(sharedMemorySize)) })} else {ByteSize::gib(8)}@@ -77,7 +97,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Whether the GPU supports SIMD group operations.fn supports_simd_group(&self) -> bool {if self.respondsToSelector(sel!(supportsSIMDGroup)) {- unsafe { msg_send![self, supportsSIMDGroup] }+ unsafe { raw_msg_send(self, sel!(supportsSIMDGroup)) }} else {false}@@ -86,7 +106,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Whether the GPU supports `simdgroup_matrix`.fn supports_simd_group_matrix(&self) -> bool {if self.respondsToSelector(sel!(supportsSIMDGroupMatrix)) {- unsafe { msg_send![self, supportsSIMDGroupMatrix] }+ unsafe { raw_msg_send(self, sel!(supportsSIMDGroupMatrix)) }} else {false}@@ -95,7 +115,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Whether the GPU supports SIMD reduction operations.fn supports_simd_reduction(&self) -> bool {if self.respondsToSelector(sel!(supportsSIMDReduction)) {- unsafe { msg_send![self, supportsSIMDReduction] }+ unsafe { raw_msg_send(self, sel!(supportsSIMDReduction)) }} else {false}@@ -104,7 +124,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Whether the GPU supports SIMD shuffle-and-fill operations.fn supports_simd_shuffle_and_fill(&self) -> bool {if self.respondsToSelector(sel!(supportsSIMDShuffleAndFill)) {- unsafe { msg_send![self, supportsSIMDShuffleAndFill] }+ unsafe { raw_msg_send(self, sel!(supportsSIMDShuffleAndFill)) }} else {false}@@ -113,7 +133,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Whether the GPU supports SIMD shuffles and broadcast.fn supports_simd_shuffles_and_broadcast(&self) -> bool {if self.respondsToSelector(sel!(supportsSIMDShufflesAndBroadcast)) {- unsafe { msg_send![self, supportsSIMDShufflesAndBroadcast] }+ unsafe { raw_msg_send(self, sel!(supportsSIMDShufflesAndBroadcast)) }} else {false}@@ -123,7 +143,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// True on M5+ (Gen18+), false on M1-M4.fn supports_mxu(&self) -> bool {if self.respondsToSelector(sel!(supportsMXU)) {- unsafe { msg_send![self, supportsMXU] }+ unsafe { raw_msg_send(self, sel!(supportsMXU)) }} else {false}@@ -132,7 +152,7 @@ pub trait DeviceExt: MTLDevice + Message + NSObjectProtocol + Sized {/// Whether the GPU supports Thread-Local Storage.fn supports_tls(&self) -> bool {if self.respondsToSelector(sel!(supportsTLS)) {- unsafe { msg_send![self, supportsTLS] }+ unsafe { raw_msg_send(self, sel!(supportsTLS)) }} else {false}