github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-23 05:48:08 -05:00
Code Issues Packages Projects Releases Wiki Activity

Actually flip local_max in CUDA (#1462)

Browse Source 
* Actually do the flip

* Fixed typo

---------

Co-authored-by: terafo <terafo@protonmail.com>
This commit is contained in:
terafo
2023-08-06 20:35:25 +03:00
committed by GitHub
parent d7d1011f1e
commit 24933ab551
2 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tinygrad/codegen/linearizer.py
View File

@@ -139,6 +139,7 @@ class LinearizerOptions(NamedTuple):
supports_float4: bool = True
supports_float4_alu: bool = True
has_local: bool = True
# NOTE: these two should be in z,y,x(reversed) order for cstyle backends, they are flipped when kernel is rendered
global_max: Optional[List[int]] = None
local_max: Optional[List[int]] = None

2
tinygrad/runtime/ops_cuda.py
View File

@@ -92,4 +92,4 @@ renderer = functools.partial(uops_to_cstyle, CStyleLanguage(
__device__ __forceinline__ explicit operator float4() const {return make_float4(__half2float(x.x), __half2float(x.y), __half2float(y.x), __half2float(y.y)); }
};
"""))
CUDABuffer = Compiled(RawCUDABuffer, LinearizerOptions(supports_float4_alu=False, global_max = [65535, 65535, 2147483647], local_max = [1024, 1024, 64]), renderer, CUDAProgram, cuda.Context.synchronize)
CUDABuffer = Compiled(RawCUDABuffer, LinearizerOptions(supports_float4_alu=False, global_max = [65535, 65535, 2147483647], local_max = [64, 1024, 1024]), renderer, CUDAProgram, cuda.Context.synchronize)