tinygrad 0.10.0__py3-none-any.whl → 0.10.2__py3-none-any.whl

tinygrad/runtime/ops_metal.py

@@ -1,8 +1,7 @@
-from __future__ import annotations
-import os, subprocess, pathlib, ctypes, tempfile, functools
-from typing import List, Any, Tuple, Optional, cast
-from tinygrad.helpers import prod, getenv, T
-from tinygrad.device import Compiled, Compiler, CompileError, LRUAllocator
+import os, pathlib, struct, ctypes, tempfile, functools, contextlib, decimal, platform
+from typing import Any, Union, cast
+from tinygrad.helpers import prod, to_mv, getenv, round_up, cache_dir, T, init_c_struct_t, PROFILE
+from tinygrad.device import Compiled, Compiler, CompileError, LRUAllocator, cpu_profile, ProfileDeviceEvent, ProfileRangeEvent
 from tinygrad.renderer.cstyle import MetalRenderer
 
 class objc_id(ctypes.c_void_p): # This prevents ctypes from converting response to plain int, and dict.fromkeys() can use it to dedup
@@ -10,10 +9,11 @@ class objc_id(ctypes.c_void_p): # This prevents ctypes from converting response
   def __eq__(self, other): return self.value == other.value
 
 class objc_instance(objc_id): # method with name "new", "alloc" should be freed after use
-  def __del__(self): msg(self, "release")
-
-@functools.lru_cache(None)
-def sel(name: str): return libobjc.sel_registerName(name.encode())
+  def __del__(self):
+    # CPython doesn't make any guarantees about order in which globals (like `msg` or `libobjc`) are destroyed when the interpreter shuts down
+    # https://github.com/tinygrad/tinygrad/pull/8949 triggered the unlucky ordering which lead to a bunch of errors at exit
+    # TODO: Why isn't `sys.is_finalizing` working?
+    if msg is not None and libobjc is not None: msg("release")(self)
 
 class MTLResourceOptions:
   MTLResourceCPUCacheModeDefaultCache = 0
@@ -22,6 +22,9 @@ class MTLResourceOptions:
 class MTLPipelineOption:
   MTLPipelineOptionNone = 0
 
+# 13 is requestType that metal uses to compile source code into MTLB, there aren't any docs or symbols.
+REQUEST_TYPE_COMPILE = 13
+
 libobjc = ctypes.CDLL("/usr/lib/libobjc.dylib")
 libmetal = ctypes.CDLL("/System/Library/Frameworks/Metal.framework/Metal")
 # Must be loaded for default Metal Device: https://developer.apple.com/documentation/metal/1433401-mtlcreatesystemdefaultdevice?language=objc
@@ -32,53 +35,107 @@ libobjc.sel_registerName.restype = objc_id
 libmetal.MTLCreateSystemDefaultDevice.restype = objc_instance
 libdispatch.dispatch_data_create.restype = objc_instance
 
-# Ignore mypy error reporting incompatible default, because typevar default only works on python 3.12
-def msg(ptr: objc_id, selector: str, /, *args: Any, restype: type[T] = objc_id) -> T: # type: ignore [assignment]
+@functools.lru_cache(None)
+def msg(selector: str, restype: type[T] = objc_id):  # type: ignore [assignment]
+  resname = libobjc.sel_registerName(selector.encode())
   sender = libobjc["objc_msgSend"] # Using attribute access returns a new reference so setting restype is safe
   sender.restype = restype
-  return sender(ptr, sel(selector), *args)
+  def _msg(ptr: objc_id, *args: Any) -> T: return sender(ptr, resname, *args)
+  return _msg
 
-def to_ns_str(s: str): return msg(libobjc.objc_getClass(b"NSString"), "stringWithUTF8String:", s.encode(), restype=objc_instance)
+@functools.lru_cache(None)
+def to_ns_str(s: str): return msg("stringWithUTF8String:", objc_instance)(libobjc.objc_getClass(b"NSString"), s.encode())
+def from_ns_str(s): return bytes(msg("UTF8String", ctypes.c_char_p)(s)).decode()
 
-def to_struct(*t: int, _type: type = ctypes.c_ulong):
-  class Struct(ctypes.Structure): pass
-  Struct._fields_ = [(f"field{i}", _type) for i in range(len(t))]
-  return Struct(*t)
+def to_struct(*t: int, _type: type = ctypes.c_ulong): return init_c_struct_t(tuple([(f"field{i}", _type) for i in range(len(t))]))(*t)
 
 def wait_check(cbuf: Any):
-  msg(cbuf, "waitUntilCompleted")
-  error_check(msg(cbuf, "error", restype=objc_instance))
+  msg("waitUntilCompleted")(cbuf)
+  error_check(msg("error", objc_instance)(cbuf))
 
-def elapsed_time(cbuf: objc_id):
-  return cast(float, msg(cbuf, "GPUEndTime", restype=ctypes.c_double)) - cast(float, msg(cbuf, "GPUStartTime", restype=ctypes.c_double))
+def cmdbuf_label(cbuf: objc_id) -> str|None: return from_ns_str(label) if (label:=msg("label", objc_id)(cbuf)).value is not None else None
+def cmdbuf_st_time(cbuf: objc_id) -> float: return cast(float, msg("GPUStartTime", ctypes.c_double)(cbuf))
+def cmdbuf_en_time(cbuf: objc_id) -> float: return cast(float, msg("GPUEndTime", ctypes.c_double)(cbuf))
 
 def error_check(error: objc_instance, error_constructor: type[Exception] = RuntimeError):
   if error.value is None: return None
-  raise error_constructor(bytes(msg(msg(error, "localizedDescription", restype=objc_instance), "UTF8String", restype=ctypes.c_char_p)).decode())
+  raise error_constructor(from_ns_str(msg("localizedDescription", objc_instance)(error)))
+
+class MetalDevice(Compiled):
+  def __init__(self, device:str):
+    self.sysdevice = libmetal.MTLCreateSystemDefaultDevice()
+    self.mtl_queue = msg("newCommandQueueWithMaxCommandBufferCount:", objc_instance)(self.sysdevice, 1024)
+    if self.mtl_queue is None: raise RuntimeError("Cannot allocate a new command queue")
+    self.mtl_buffers_in_flight: list[Any] = []
+    self.timeline_signal = msg("newSharedEvent", objc_instance)(self.sysdevice)
+    self.timeline_value = 0
+
+    Compiled.profile_events += [ProfileDeviceEvent(device)]
+
+    from tinygrad.runtime.graph.metal import MetalGraph
+    super().__init__(device, MetalAllocator(self), MetalRenderer(), MetalCompiler() if getenv("METAL_DIRECT", 1) else Compiler(),
+                     functools.partial(MetalProgram, self), MetalGraph)
+
+  def synchronize(self):
+    for cbuf in self.mtl_buffers_in_flight:
+      wait_check(cbuf)
+      st, en = decimal.Decimal(cmdbuf_st_time(cbuf)) * 1000000, decimal.Decimal(cmdbuf_en_time(cbuf)) * 1000000
+      if PROFILE and (lb:=cmdbuf_label(cbuf)) is not None:
+        Compiled.profile_events += [ProfileRangeEvent(self.device, lb, st, en, is_copy=lb.startswith("COPY"))]
+    self.mtl_buffers_in_flight.clear()
 
 def metal_src_to_library(device:MetalDevice, src:str) -> objc_instance:
-  options = msg(libobjc.objc_getClass(b"MTLCompileOptions"), "new", restype=objc_instance)
-  msg(options, "setFastMathEnabled:", getenv("METAL_FAST_MATH"))
-  library = msg(device.device, "newLibraryWithSource:options:error:", to_ns_str(src), options,
-                ctypes.byref(compileError:=objc_instance()), restype=objc_instance)
+  options = msg("new", objc_instance)(libobjc.objc_getClass(b"MTLCompileOptions"))
+  msg("setFastMathEnabled:")(options, getenv("METAL_FAST_MATH"))
+  library = msg("newLibraryWithSource:options:error:", objc_instance)(device.sysdevice, to_ns_str(src),
+                                                                      options, ctypes.byref(compileError:=objc_instance()))
   error_check(compileError, CompileError)
   return library
 
 class MetalCompiler(Compiler):
-  def __init__(self, device:Optional[MetalDevice]=None):
-    self.device = device
-    super().__init__("compile_metal_xcode" if self.device is None else "compile_metal")
+  # Opening METAL after LLVM doesn't fail because ctypes.CDLL opens with RTLD_LOCAL but MTLCompiler opens it's own llvm with RTLD_GLOBAL
+  # This means that MTLCompiler's llvm will create it's own instances of global state because RTLD_LOCAL doesn't export symbols, but if RTLD_GLOBAL
+  # library is loaded first then RTLD_LOCAL library will just use it's symbols. On linux there is RTLD_DEEPBIND to prevent that, but on macos there
+  # doesn't seem to be anything we can do.
+  with contextlib.suppress(FileNotFoundError):
+    import tinygrad.runtime.autogen.llvm # noqa: F401
+  support = ctypes.CDLL("/System/Library/PrivateFrameworks/MTLCompiler.framework/MTLCompiler")
+  support.MTLCodeGenServiceCreate.restype = ctypes.c_void_p
+
+  def __init__(self):
+    self.cgs = ctypes.c_void_p(MetalCompiler.support.MTLCodeGenServiceCreate(b"tinygrad"))
+    super().__init__("compile_metal_direct")
+  def __reduce__(self): return (MetalCompiler,()) # force pickle to create new instance for each multiprocessing fork
   def compile(self, src:str) -> bytes:
-    if self.device is None:
-      # NOTE: if you run llvm-dis on "air" you can see the llvm bytecode
-      air = subprocess.check_output(['xcrun', '-sdk', 'macosx', 'metal', '-x', 'metal', '-c', '-', '-o', '-'], input=src.encode('utf-8'))
-      lib = subprocess.check_output(['xcrun', '-sdk', 'macosx', 'metallib', '-', '-o', '-'], input=air)
-    else:
-      library = metal_src_to_library(self.device, src)
-      library_contents = msg(library, "libraryDataContents", restype=objc_instance)
-      lib = ctypes.string_at(msg(library_contents, "bytes"), cast(int, msg(library_contents, "length", restype=ctypes.c_ulong)))
-    assert lib[:4] == b"MTLB", "Invalid Metal library. Using conda? Corrupt XCode?"
-    return lib
+    ret: Union[Exception, bytes] = CompileError("MTLCodeGenServiceBuildRequest returned without calling the callback")
+    @ctypes.CFUNCTYPE(None, ctypes.c_void_p, ctypes.c_int32, ctypes.c_void_p, ctypes.c_size_t, ctypes.c_char_p)
+    def callback(blockptr, error, dataPtr, dataLen, errorMessage):
+      nonlocal ret
+      if error == 0:
+        reply = bytes(to_mv(dataPtr, dataLen))
+        # offset from beginning to data = header size + warning size
+        ret = reply[sum(struct.unpack('<LL', reply[8:16])):]
+      else:
+        ret = CompileError(errorMessage.decode())
+
+    # no changes for compute in 2.0 - 2.4 specs, use 2.0 as default for old versions.
+    macos_major = int(platform.mac_ver()[0].split('.')[0])
+    metal_version = "metal3.1" if macos_major >= 14 else "metal3.0" if macos_major >= 13 else "macos-metal2.0"
+
+    # llvm will create modules.timestamp in cache path and cache compilation of metal stdlib (250ms => 8ms compilation time)
+    # note that llvm won't necessarily create anything else here as apple has prebuilt versions of many standard libraries
+    params = f'-fno-fast-math -std={metal_version} --driver-mode=metal -x metal -fmodules-cache-path="{cache_dir}" -fno-caret-diagnostics'
+    # source blob has to be padded to multiple of 4 but at least one 'b\x00' should be added, params blob just has to be null terminated
+    src_padded, params_padded = src.encode() + b'\x00'*(round_up(len(src) + 1, 4) - len(src)), params.encode() + b'\x00'
+    request = struct.pack('<QQ', len(src_padded), len(params_padded)) + src_padded + params_padded
+    # The callback is actually not a callback but a block which is apple's non-standard extension to add closures to C.
+    # See https://clang.llvm.org/docs/Block-ABI-Apple.html#high-level for struct layout.
+    # Fields other than invoke are unused in this case so we can just use ctypes.byref with negative offset to invoke field, add blockptr as a first
+    # argument and pretend it's a normal callback
+    MetalCompiler.support.MTLCodeGenServiceBuildRequest(self.cgs, None, REQUEST_TYPE_COMPILE, request, len(request), ctypes.byref(callback, -0x10))
+    if isinstance(ret, Exception): raise ret
+    assert ret[:4] == b"MTLB" and ret[-4:] == b"ENDT", f"Invalid Metal library. {ret!r}"
+    return ret
   def disassemble(self, lib:bytes):
     with tempfile.NamedTemporaryFile(delete=True) as shader:
       shader.write(lib)
@@ -87,102 +144,81 @@ class MetalCompiler(Compiler):
       if ret: print("Disassembler Error: Make sure you have https://github.com/dougallj/applegpu cloned to tinygrad/extra/disassemblers/applegpu")
 
 class MetalProgram:
-  def __init__(self, device:MetalDevice, name:str, lib:bytes):
-    self.device, self.name, self.lib = device, name, lib
+  def __init__(self, dev:MetalDevice, name:str, lib:bytes):
+    self.dev, self.name, self.lib = dev, name, lib
     if lib[:4] == b"MTLB":
       # binary metal library
       data = libdispatch.dispatch_data_create(lib, len(lib), None, None)
-      error_library_creation = objc_instance()
-      self.library = msg(self.device.device, "newLibraryWithData:error:", data, ctypes.byref(error_library_creation), restype=objc_instance)
-      error_check(error_library_creation)
+      self.library = msg("newLibraryWithData:error:", objc_instance)(self.dev.sysdevice, data, ctypes.byref(error_lib:=objc_instance()))
+      error_check(error_lib)
     else:
       # metal source. rely on OS caching
-      try: self.library = metal_src_to_library(self.device, lib.decode())
+      try: self.library = metal_src_to_library(self.dev, lib.decode())
       except CompileError as e: raise RuntimeError from e
-    self.fxn = msg(self.library, "newFunctionWithName:", to_ns_str(name), restype=objc_instance)
-    descriptor = msg(libobjc.objc_getClass(b"MTLComputePipelineDescriptor"), "new", restype=objc_instance)
-    msg(descriptor, "setComputeFunction:", self.fxn)
-    msg(descriptor, "setSupportIndirectCommandBuffers:", True)
-    self.pipeline_state = msg(self.device.device, "newComputePipelineStateWithDescriptor:options:reflection:error:",
-      descriptor, MTLPipelineOption.MTLPipelineOptionNone, None, ctypes.byref(error_pipeline_creation:=objc_instance()), restype=objc_instance)
+    self.fxn = msg("newFunctionWithName:", objc_instance)(self.library, to_ns_str(name))
+    descriptor = msg("new", objc_instance)(libobjc.objc_getClass(b"MTLComputePipelineDescriptor"))
+    msg("setComputeFunction:")(descriptor, self.fxn)
+    msg("setSupportIndirectCommandBuffers:")(descriptor, True)
+    self.pipeline_state = msg("newComputePipelineStateWithDescriptor:options:reflection:error:", objc_instance)(self.dev.sysdevice,
+      descriptor, MTLPipelineOption.MTLPipelineOptionNone, None, ctypes.byref(error_pipeline_creation:=objc_instance()))
     error_check(error_pipeline_creation)
-
-  def __call__(self, *bufs, global_size:Tuple[int,int,int]=(1,1,1), local_size:Tuple[int,int,int]=(1,1,1), vals:Tuple[int, ...]=(), wait=False):
-    max_total_threads = msg(self.pipeline_state, "maxTotalThreadsPerThreadgroup", restype=ctypes.c_ulong)
-    if prod(local_size) > cast(int, max_total_threads):
-      exec_width = msg(self.pipeline_state, "threadExecutionWidth", restype=ctypes.c_ulong)
-      memory_length = msg(self.pipeline_state, "staticThreadgroupMemoryLength", restype=ctypes.c_ulong)
-      raise RuntimeError(f"local size {local_size} bigger than {max_total_threads} with exec width {exec_width} memory length {memory_length}")
-    command_buffer = msg(self.device.mtl_queue, "commandBuffer", restype=objc_instance)
-    encoder = msg(command_buffer, "computeCommandEncoder", restype=objc_instance)
-    msg(encoder, "setComputePipelineState:", self.pipeline_state)
-    for i,a in enumerate(bufs): msg(encoder, "setBuffer:offset:atIndex:", a.buf, a.offset, i)
-    for i,a in enumerate(vals,start=len(bufs)): msg(encoder, "setBytes:length:atIndex:", bytes(ctypes.c_int(a)), 4, i)
-    msg(encoder, "dispatchThreadgroups:threadsPerThreadgroup:", to_struct(*global_size), to_struct(*local_size))
-    msg(encoder, "endEncoding")
-    msg(command_buffer, "commit")
+    # cache these msg calls
+    self.max_total_threads: int = cast(int, msg("maxTotalThreadsPerThreadgroup", ctypes.c_ulong)(self.pipeline_state))
+
+  def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int, ...]=(), wait=False):
+    if prod(local_size) > self.max_total_threads:
+      exec_width = msg("threadExecutionWidth", ctypes.c_ulong)(self.pipeline_state)
+      memory_length = msg("staticThreadgroupMemoryLength", ctypes.c_ulong)(self.pipeline_state)
+      raise RuntimeError(f"local size {local_size} bigger than {self.max_total_threads} with exec width {exec_width} memory length {memory_length}")
+    command_buffer = msg("commandBuffer", objc_instance)(self.dev.mtl_queue)
+    encoder = msg("computeCommandEncoder", objc_instance)(command_buffer)
+    msg("setComputePipelineState:")(encoder, self.pipeline_state)
+    for i,a in enumerate(bufs): msg("setBuffer:offset:atIndex:")(encoder, a.buf, a.offset, i)
+    for i,a in enumerate(vals, start=len(bufs)): msg("setBytes:length:atIndex:")(encoder, bytes(ctypes.c_int(a)), 4, i)
+    msg("dispatchThreadgroups:threadsPerThreadgroup:")(encoder, to_struct(*global_size), to_struct(*local_size))
+    msg("endEncoding")(encoder)
+    msg("setLabel:")(command_buffer, to_ns_str(self.name)) # TODO: is this always needed?
+    msg("commit")(command_buffer)
+    self.dev.mtl_buffers_in_flight.append(command_buffer)
     if wait:
       wait_check(command_buffer)
-      return elapsed_time(command_buffer)
-    self.device.mtl_buffers_in_flight.append(command_buffer)
+      return cmdbuf_en_time(command_buffer) - cmdbuf_st_time(command_buffer)
 
 class MetalBuffer:
   def __init__(self, buf:Any, size:int, offset=0): self.buf, self.size, self.offset = buf, size, offset
 
 class MetalAllocator(LRUAllocator):
-  def __init__(self, device:MetalDevice):
-    self.device:MetalDevice = device
+  def __init__(self, dev:MetalDevice):
+    self.dev:MetalDevice = dev
     super().__init__()
   def _alloc(self, size:int, options) -> MetalBuffer:
     # Buffer is explicitly released in _free() rather than garbage collected via reference count
-    ret = msg(self.device.device, "newBufferWithLength:options:", size, MTLResourceOptions.MTLResourceStorageModeShared, restype=objc_id)
+    ret = msg("newBufferWithLength:options:", objc_id)(self.dev.sysdevice, ctypes.c_ulong(size), MTLResourceOptions.MTLResourceStorageModeShared)
     if ret.value is None: raise MemoryError(f"Metal OOM while allocating {size=}")
     return MetalBuffer(ret, size)
-  def _free(self, opaque:MetalBuffer, options): msg(opaque.buf, "release")
-  def transfer(self, dest:MetalBuffer, src:MetalBuffer, sz:int, src_dev:MetalDevice, dest_dev:MetalDevice):
+  def _free(self, opaque:MetalBuffer, options): msg("release")(opaque.buf)
+  def _transfer(self, dest:MetalBuffer, src:MetalBuffer, sz:int, src_dev:MetalDevice, dest_dev:MetalDevice):
     dest_dev.synchronize()
-    src_command_buffer = msg(src_dev.mtl_queue, "commandBuffer", restype=objc_instance)
-    encoder = msg(src_command_buffer, "blitCommandEncoder", restype=objc_instance)
-    msg(encoder, "copyFromBuffer:sourceOffset:toBuffer:destinationOffset:size:", src.buf, ctypes.c_ulong(src.offset),
+    src_command_buffer = msg("commandBuffer", objc_instance)(src_dev.mtl_queue)
+    encoder = msg("blitCommandEncoder", objc_instance)(src_command_buffer)
+    msg("copyFromBuffer:sourceOffset:toBuffer:destinationOffset:size:")(encoder, src.buf, ctypes.c_ulong(src.offset),
         dest.buf, ctypes.c_ulong(dest.offset), ctypes.c_ulong(sz))
-    msg(encoder, "endEncoding")
+    msg("endEncoding")(encoder)
     if src_dev != dest_dev:
-      msg(src_command_buffer, "encodeSignalEvent:value:", src_dev.timeline_signal, src_dev.timeline_value)
-      dest_command_buffer = msg(dest_dev.mtl_queue, "commandBuffer", restype=objc_instance)
-      msg(dest_command_buffer, "encodeWaitForEvent:value:", src_dev.timeline_signal, src_dev.timeline_value)
-      msg(dest_command_buffer, "commit")
+      msg("encodeSignalEvent:value:")(src_command_buffer, src_dev.timeline_signal, src_dev.timeline_value)
+      dest_command_buffer = msg("commandBuffer", objc_instance)(dest_dev.mtl_queue)
+      msg("encodeWaitForEvent:value:")(dest_command_buffer, src_dev.timeline_signal, src_dev.timeline_value)
+      msg("commit")(dest_command_buffer)
       dest_dev.mtl_buffers_in_flight.append(dest_command_buffer)
       src_dev.timeline_value += 1
-    msg(src_command_buffer, "commit")
+    msg("setLabel:")(src_command_buffer, to_ns_str(f"COPY {src_dev.device} -> {dest_dev.device}"))
+    msg("commit")(src_command_buffer)
     src_dev.mtl_buffers_in_flight.append(src_command_buffer)
-  def from_buffer(self, src:memoryview) -> Optional[Any]:
-    ptr = (ctypes.c_char * src.nbytes).from_buffer(src)
-    ret = msg(self.device.device, "newBufferWithBytesNoCopy:length:options:deallocator:", ptr, src.nbytes, 0, None, restype=objc_instance)
-    if ret: self.device.mv_in_metal.append(src)
-    return MetalBuffer(ret, src.nbytes)
-  def as_buffer(self, src:MetalBuffer) -> memoryview:
-    self.device.synchronize()
-    ptr = msg(src.buf, "contents", restype=objc_id) # Shared memory, do not release here
-    array = (ctypes.c_char * (src.offset + src.size)).from_address(ptr.value)
-    return memoryview(array).cast("B")[src.offset:]
-  def copyin(self, dest:MetalBuffer, src:memoryview): self.as_buffer(dest)[:] = src
-  def copyout(self, dest:memoryview, src:MetalBuffer): dest[:] = self.as_buffer(src)
-  def offset(self, buf:MetalBuffer, size:int, offset:int): return MetalBuffer(buf.buf, size, offset)
-
-class MetalDevice(Compiled):
-  def __init__(self, device:str):
-    self.device = libmetal.MTLCreateSystemDefaultDevice()
-    self.mtl_queue = msg(self.device, "newCommandQueueWithMaxCommandBufferCount:", 1024, restype=objc_instance)
-    if self.mtl_queue is None: raise RuntimeError("Cannot allocate a new command queue")
-    self.mtl_buffers_in_flight: List[Any] = []
-    self.mv_in_metal: List[memoryview] = []
-    self.timeline_signal = msg(self.device, "newSharedEvent", restype=objc_instance)
-    self.timeline_value = 0
-
-    from tinygrad.runtime.graph.metal import MetalGraph
-    super().__init__(device, MetalAllocator(self), MetalRenderer(), MetalCompiler() if getenv("METAL_XCODE") else Compiler(),
-                     functools.partial(MetalProgram, self), MetalGraph)
-  def synchronize(self):
-    for cbuf in self.mtl_buffers_in_flight: wait_check(cbuf)
-    self.mv_in_metal.clear()
-    self.mtl_buffers_in_flight.clear()
+  def _cp_mv(self, dst, src, prof_desc):
+    with cpu_profile(prof_desc, self.dev.device, is_copy=True): dst[:] = src
+  def _as_buffer(self, src:MetalBuffer) -> memoryview:
+    self.dev.synchronize()
+    return to_mv(cast(int, msg("contents", objc_id)(src.buf).value), src.size + src.offset)[src.offset:]
+  def _copyin(self, dest:MetalBuffer, src:memoryview): self._cp_mv(self._as_buffer(dest), src, "CPU -> METAL")
+  def _copyout(self, dest:memoryview, src:MetalBuffer): self._cp_mv(dest, self._as_buffer(src), "METAL -> CPU")
+  def _offset(self, buf:MetalBuffer, size:int, offset:int): return MetalBuffer(buf.buf, size, offset)

tinygrad/runtime/ops_npy.py

@@ -2,8 +2,8 @@ import numpy as np
 from tinygrad.helpers import flat_mv
 from tinygrad.device import Compiled, Allocator
 
-class NpyAllocator(Allocator):  # pylint: disable=abstract-method
-  def copyout(self, dest:memoryview, src:np.ndarray): dest[:] = flat_mv(np.require(src, requirements='C').data)
+class NpyAllocator(Allocator):
+  def _copyout(self, dest:memoryview, src:np.ndarray): dest[:] = flat_mv(np.require(src, requirements='C').data)
 
 class NpyDevice(Compiled):
   def __init__(self, device:str): super().__init__(device, NpyAllocator(), None, None, None)