github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-02-16 09:37:11 -05:00
Code Issues Packages Projects Releases Wiki Activity

BufferSpec and ProgramSpec [pr] (#7814)

Browse Source 
* BufferSpec and ProgramSpec [pr]

* delete preallocate, it's unused

* Revert "delete preallocate, it's unused"

This reverts commit dcfcfaccde.
This commit is contained in:
George Hotz
2024-11-21 12:18:05 +08:00
committed by GitHub
parent 490a6130af
commit c5d458ce02
27 changed files with 139 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tinygrad/runtime/ops_gpu.py
View File

@@ -4,7 +4,7 @@ import ctypes, functools, hashlib
from tinygrad.runtime.autogen import opencl as cl
from tinygrad.helpers import init_c_var, to_char_p_p, from_mv, OSX, DEBUG, getenv, mv_address
from tinygrad.renderer.cstyle import OpenCLRenderer, IntelRenderer
from tinygrad.device import BufferOptions, LRUAllocator, Compiled, Compiler, CompileError
from tinygrad.device import BufferSpec, LRUAllocator, Compiled, Compiler, CompileError
# 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
@@ -44,7 +44,7 @@ class CLProgram:
if hasattr(self, 'kernel'): check(cl.clReleaseKernel(self.kernel))
if hasattr(self, 'program'): check(cl.clReleaseProgram(self.program))
def __call__(self, *bufs:Tuple[ctypes._CData, BufferOptions], global_size:Tuple[int,int,int]=(1,1,1), local_size:Optional[Tuple[int,int,int]]=None, vals:Tuple[int, ...]=(), wait=False) -> Optional[float]: # noqa: E501
def __call__(self, *bufs:Tuple[ctypes._CData, BufferSpec], global_size:Tuple[int,int,int]=(1,1,1), local_size:Optional[Tuple[int,int,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,v in enumerate(vals,start=len(bufs)): cl.clSetKernelArg(self.kernel, i, 4, ctypes.byref(ctypes.c_int32(v)))
if local_size is not None: global_size = cast(Tuple[int,int,int], tuple(int(g*l) for g,l in zip(global_size, local_size)))
@@ -62,14 +62,14 @@ class CLAllocator(LRUAllocator):
def __init__(self, dev:CLDevice):
self.dev = dev
super().__init__()
def _alloc(self, size:int, options:BufferOptions) -> Tuple[ctypes._CData, BufferOptions]:
def _alloc(self, size:int, options:BufferSpec) -> Tuple[ctypes._CData, BufferSpec]:
if options.image is not None:
return (checked(cl.clCreateImage2D(self.dev.context, cl.CL_MEM_READ_WRITE,
cl.cl_image_format(cl.CL_RGBA, {2: cl.CL_HALF_FLOAT, 4: cl.CL_FLOAT}[options.image.itemsize]),
options.image.shape[1], options.image.shape[0], 0, None, status := ctypes.c_int32()), status), options)
return (checked(cl.clCreateBuffer(self.dev.context, cl.CL_MEM_READ_WRITE, size, None, status := ctypes.c_int32()), status), options)
def _free(self, opaque:Tuple[ctypes._CData, BufferOptions], options:BufferOptions): check(cl.clReleaseMemObject(opaque[0]))
def _copyin(self, dest:Tuple[ctypes._CData, BufferOptions], src:memoryview):
def _free(self, opaque:Tuple[ctypes._CData, BufferSpec], options:BufferSpec): check(cl.clReleaseMemObject(opaque[0]))
def _copyin(self, dest:Tuple[ctypes._CData, BufferSpec], src:memoryview):
if dest[1].image is not None:
check(cl.clEnqueueWriteImage(self.dev.queue, dest[0], False, (ctypes.c_size_t * 3)(0,0,0),
(ctypes.c_size_t * 3)(dest[1].image.shape[1],dest[1].image.shape[0],1), 0, 0, from_mv(src), 0, None, None))
@@ -77,7 +77,7 @@ class CLAllocator(LRUAllocator):
if mv_address(src) % 16: src = memoryview(bytearray(src))
check(cl.clEnqueueWriteBuffer(self.dev.queue, dest[0], False, 0, len(src)*src.itemsize, from_mv(src), 0, None, None))
self.dev.pending_copyin.append(src) # NOTE: these can't be freed until the GPU actually executes this command
def _copyout(self, dest:memoryview, src:Tuple[ctypes._CData, BufferOptions]):
def _copyout(self, dest:memoryview, src:Tuple[ctypes._CData, BufferSpec]):
if src[1].image is not None:
check(cl.clEnqueueReadImage(self.dev.queue, src[0], False, (ctypes.c_size_t * 3)(0,0,0),
(ctypes.c_size_t * 3)(src[1].image.shape[1],src[1].image.shape[0],1), 0, 0, from_mv(dest), 0, None, None))