delete half_prekernel (#3388)

* generic rendering of half and bf16

hotfix

* fix uops + regression test

* fix the test for metal's half4

* uop.uop fixup

* mypy with --strict-equality, fix ops_gpu
This commit is contained in:
qazal
2024-02-14 16:40:48 +02:00
committed by GitHub
parent 078a2603d5
commit 27f4de2ce4
4 changed files with 21 additions and 15 deletions

View File

@@ -122,6 +122,7 @@ def _test_ops(a_dtype:DType, b_dtype:DType, target_dtype=None):
if not is_dtype_supported(a_dtype) or not is_dtype_supported(b_dtype) or not is_dtype_supported(target_dtype): return
if a_dtype == dtypes.bool or b_dtype == dtypes.bool: return
_assert_eq(Tensor([1,2,3,4], dtype=a_dtype)+Tensor([1,2,3,4], dtype=b_dtype), target_dtype, [2,4,6,8])
_assert_eq((Tensor([1], dtype=a_dtype).cast(b_dtype)+Tensor([1], dtype=a_dtype).cast(b_dtype)).cast(a_dtype), a_dtype, [2])
_assert_eq(Tensor([1,2,3,4], dtype=a_dtype)*Tensor([1,2,3,4], dtype=b_dtype), target_dtype, [1,4,9,16])
_assert_eq(Tensor([[1,2],[3,4]], dtype=a_dtype)@Tensor.eye(2, dtype=b_dtype), target_dtype, [[1,2],[3,4]])
_assert_eq(Tensor([1,1,1,1], dtype=a_dtype)+Tensor.ones((4,4), dtype=b_dtype), target_dtype, 2*Tensor.ones(4,4).numpy())