github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-02-12 23:54:58 -05:00
Code Issues Packages Projects Releases Wiki Activity

ruff checks the max line length is 150 (#2734)

Browse Source 
* ruff checks the max line length is 150

* fix tensor.py

* a lot more

* done
This commit is contained in:
George Hotz
2023-12-12 17:34:47 -08:00
committed by GitHub
parent 3635540ddb
commit 6d6eb9302d
59 changed files with 438 additions and 324 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
tinygrad/runtime/ops_gpu.py
View File

@@ -7,7 +7,8 @@ from tinygrad.codegen.kernel import LinearizerOptions
from tinygrad.renderer.cstyle import OpenCLRenderer
from tinygrad.device import Compiled, LRUAllocator
OSX_TIMING_RATIO = (125/3) if OSX else 1.0 # see test/external/external_osx_profiling.py to determine this ratio. it's in like GPU clocks or something
# see test/external/external_osx_profiling.py to determine this ratio. it's in like GPU clocks or something
OSX_TIMING_RATIO = (125/3) if OSX else 1.0
def check(status):
if status != 0: raise RuntimeError(f"OpenCL Error {status}")
@@ -16,22 +17,24 @@ def checked(ret, status): return (check(status.value), ret)[1]
@diskcache
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()]), ctypes.byref(ctypes.c_size_t(len(prg_bytes))), ctypes.byref(status := ctypes.c_int32())), status)
program = checked(cl.clCreateProgramWithSource(CLDevice.compiler_context.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)
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()))
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)
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
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)))
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)))
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
check(cl.clReleaseProgram(program))
return bytes(binary)
class CLProgram:
def __init__(self, device:CLDevice, name:str, lib:bytes):
self.device, self.name, self.lib = device, name, lib
self.program = checked(cl.clCreateProgramWithBinary(device.context, 1, ctypes.byref(device.device_id), (ctypes.c_size_t * 1)(len(lib)), to_char_p_p([lib], ctypes.c_ubyte),
ctypes.byref(binary_status := ctypes.c_int32()), ctypes.byref(errcode_ret := ctypes.c_int32())), errcode_ret)
self.program = checked(cl.clCreateProgramWithBinary(device.context, 1, ctypes.byref(device.device_id), (ctypes.c_size_t * 1)(len(lib)),
to_char_p_p([lib], ctypes.c_ubyte), ctypes.byref(binary_status := ctypes.c_int32()),
ctypes.byref(errcode_ret := ctypes.c_int32())), errcode_ret)
check(binary_status.value)
check(cl.clBuildProgram(self.program, 1, ctypes.byref(device.device_id), None, cl.clBuildProgram.argtypes[4](), None)) # NOTE: OSX requires this
self.kernel = checked(cl.clCreateKernel(self.program, name.encode(), ctypes.byref(status := ctypes.c_int32())), status)
@@ -40,16 +43,16 @@ class CLProgram:
check(cl.clReleaseKernel(self.kernel))
check(cl.clReleaseProgram(self.program))
def __call__(self, *bufs:cl.cl_mem, global_size:Tuple[int,...], local_size:Optional[Tuple[int,...]]=None, vals:Tuple[int, ...]=(), wait=False) -> Optional[float]:
def __call__(self, *bufs:cl.cl_mem, global_size:Tuple[int,...], local_size:Optional[Tuple[int,...]]=None, vals:Tuple[int, ...]=(), wait=False) -> Optional[float]: # noqa: E501
for i,b in enumerate(bufs): cl.clSetKernelArg(self.kernel, i, ctypes.sizeof(b), ctypes.byref(b))
for i,b in enumerate(vals,start=len(bufs)): cl.clSetKernelArg(self.kernel, i, 4, ctypes.byref(ctypes.c_int32(b)))
if local_size is not None: global_size = tuple(int(g*l) for g,l in zip(global_size, local_size))
event = cl.cl_event() if wait else None
check(cl.clEnqueueNDRangeKernel(self.device.queue, self.kernel, len(global_size), None, (ctypes.c_size_t * len(global_size))(*global_size), (ctypes.c_size_t * len(local_size))(*local_size) if local_size else None, 0, None, event))
check(cl.clEnqueueNDRangeKernel(self.device.queue, self.kernel, len(global_size), None, (ctypes.c_size_t * len(global_size))(*global_size), (ctypes.c_size_t * len(local_size))(*local_size) if local_size else None, 0, None, event)) # noqa: E501
if wait:
check(cl.clWaitForEvents(1, ctypes.byref(event)))
start = init_c_var(ctypes.c_ulong(), lambda x: check(cl.clGetEventProfilingInfo(event, cl.CL_PROFILING_COMMAND_START, ctypes.sizeof(x), ctypes.byref(x), None)))
end = init_c_var(ctypes.c_ulong(), lambda x: check(cl.clGetEventProfilingInfo(event, cl.CL_PROFILING_COMMAND_END, ctypes.sizeof(x), ctypes.byref(x), None)))
start = init_c_var(ctypes.c_ulong(), lambda x: check(cl.clGetEventProfilingInfo(event, cl.CL_PROFILING_COMMAND_START, ctypes.sizeof(x), ctypes.byref(x), None))) # noqa: E501
end = init_c_var(ctypes.c_ulong(), lambda x: check(cl.clGetEventProfilingInfo(event, cl.CL_PROFILING_COMMAND_END, ctypes.sizeof(x), ctypes.byref(x), None))) # noqa: E501
return float(end.value-start.value) * OSX_TIMING_RATIO * 1e-9
return None
@@ -83,10 +86,10 @@ class CLDevice(Compiled):
err = cl.clGetDeviceIDs(platform_ids[0], device_type, 0, None, ctypes.byref(num_devices))
if err == 0 and num_devices.value != 0: break
if DEBUG >= 1: print(f"CLDevice: got {num_platforms.value} platforms and {num_devices.value} devices")
CLDevice.device_ids = init_c_var((cl.cl_device_id * num_devices.value)(), lambda x: check(cl.clGetDeviceIDs(platform_ids[0], device_type, num_devices, x, None)))
CLDevice.device_ids = init_c_var((cl.cl_device_id * num_devices.value)(), lambda x: check(cl.clGetDeviceIDs(platform_ids[0], device_type, num_devices, x, None))) # noqa: E501
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)
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] = []