github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-02-13 08:05:10 -05:00
Code Issues Packages Projects Releases Wiki Activity

update llvm api + add cache key (#3140)

Browse Source 
* update llvm api + add cache key

* use_xcode is a different function

* types
This commit is contained in:
George Hotz
2024-01-15 17:25:32 -08:00
committed by GitHub
parent cec0a7bc37
commit 120c8b1841
7 changed files with 56 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
tinygrad/runtime/ops_gpu.py
View File

@@ -15,14 +15,13 @@ def check(status):
if status != 0: raise RuntimeError(f"OpenCL Error {status}")
def checked(ret, status): return (check(status.value), ret)[1]
def compile_cl(prg:str) -> bytes:
assert CLDevice.compiler_context is not None, 'OpenCL requires a "compiler_context" to compile, init a device before you call this'
program = checked(cl.clCreateProgramWithSource(CLDevice.compiler_context.context, 1, to_char_p_p([prg_bytes := prg.encode()]),
def compile_cl(device:CLDevice, prg:str) -> bytes:
program = checked(cl.clCreateProgramWithSource(device.context, 1, to_char_p_p([prg_bytes := prg.encode()]),
ctypes.byref(ctypes.c_size_t(len(prg_bytes))), ctypes.byref(status := ctypes.c_int32())), status)
status = cl.clBuildProgram(program, 1, ctypes.byref(CLDevice.compiler_context.device_id), None, cl.clBuildProgram.argtypes[4](), None)
status = cl.clBuildProgram(program, 1, ctypes.byref(device.device_id), None, cl.clBuildProgram.argtypes[4](), None)
if status != 0:
cl.clGetProgramBuildInfo(program, CLDevice.compiler_context.device_id, cl.CL_PROGRAM_BUILD_LOG, 0, None, ctypes.byref(log_size := ctypes.c_size_t())) # noqa: E501
cl.clGetProgramBuildInfo(program, CLDevice.compiler_context.device_id, cl.CL_PROGRAM_BUILD_LOG, log_size.value, mstr := ctypes.create_string_buffer(log_size.value), None) # noqa: E501
cl.clGetProgramBuildInfo(program, device.device_id, cl.CL_PROGRAM_BUILD_LOG, 0, None, ctypes.byref(log_size := ctypes.c_size_t()))
cl.clGetProgramBuildInfo(program, device.device_id, cl.CL_PROGRAM_BUILD_LOG, log_size.value, mstr := ctypes.create_string_buffer(log_size.value), None) # noqa: E501
raise RuntimeError(f"OpenCL Compile Error\n\n{ctypes.string_at(mstr, size=log_size.value).decode()}")
binary_sizes = init_c_var((ctypes.c_size_t * 1)(), lambda x: check(cl.clGetProgramInfo(program, cl.CL_PROGRAM_BINARY_SIZES, ctypes.sizeof(x), ctypes.byref(x), None))) # noqa: E501
binary = init_c_var(ctypes.create_string_buffer(binary_sizes[0]), lambda x: check(cl.clGetProgramInfo(program, cl.CL_PROGRAM_BINARIES, ctypes.sizeof(ctypes.c_void_p), ctypes.byref((ctypes.c_void_p * 1)(ctypes.addressof(x))), None))) # noqa: E501
@@ -76,7 +75,6 @@ class CLAllocator(LRUAllocator):
class CLDevice(Compiled):
device_ids = None # this is global and only initted once
compiler_context = None # this is the first created context. we make an assumption they are all the same for the compiler
def __init__(self, device:str=""):
if CLDevice.device_ids is None:
num_platforms = init_c_var(ctypes.c_uint32(), lambda x: check(cl.clGetPlatformIDs(0, None, ctypes.byref(x))))
@@ -90,10 +88,11 @@ class CLDevice(Compiled):
self.device_id = CLDevice.device_ids[0 if ":" not in device else int(device.split(":")[1])]
self.context = checked(cl.clCreateContext(None, 1, ctypes.byref(self.device_id), cl.clCreateContext.argtypes[3](), None, ctypes.byref(status := ctypes.c_int32())), status) # noqa: E501
if CLDevice.compiler_context is None: CLDevice.compiler_context = self
self.queue = checked(cl.clCreateCommandQueue(self.context, self.device_id, cl.CL_QUEUE_PROFILING_ENABLE, ctypes.byref(status)), status)
self.pending_copyin: List[memoryview] = []
super().__init__(CLAllocator(self), LinearizerOptions("GPU"), OpenCLRenderer, compile_cl, functools.partial(CLProgram, self))
# TODO: vary the cache key based on device name
super().__init__(CLAllocator(self), LinearizerOptions("GPU"), OpenCLRenderer,
functools.partial(compile_cl, self), "compile_cl", functools.partial(CLProgram, self))
def synchronize(self):
check(cl.clFinish(self.queue))
self.pending_copyin.clear()