mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-24 07:04:00 +08:00
Lei Wang
6722690697
* Fix L2 persistent lowering for fp16 host stubs Avoid typed access_ptr casts in stream access policy window setup and add issue #1810 regression test. * test and lint fix * Add regression test for issue #1810 Add a regression test for issue #1810 related to CUDA stream access policy.
33 lines
1.2 KiB
Python
33 lines
1.2 KiB
Python
import tilelang
|
|
import tilelang.testing
|
|
from tilelang import language as T
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_issue_1810_l2_persistent_float16_host_stub_no_half(monkeypatch):
|
|
"""Regression for issue #1810.
|
|
|
|
`annotate_l2_hit_ratio` lowers to CUDA stream access policy window calls.
|
|
The runtime API only needs an opaque base pointer, but older lowering used a
|
|
typed access pointer. For float16 buffers, TVM's C host codegen prints the
|
|
type as `half`, which is not defined by the stable C ABI headers, causing
|
|
host-side compilation failures when exporting the executable.
|
|
"""
|
|
|
|
@T.prim_func
|
|
def minimal_kernel(A: T.Buffer((1,), "float16")):
|
|
with T.Kernel():
|
|
T.annotate_l2_hit_ratio({A: 0.9})
|
|
T.evaluate(0)
|
|
|
|
kernel = tilelang.compile(minimal_kernel, execution_backend="tvm_ffi", target="cuda")
|
|
source = kernel.get_host_source()
|
|
assert "__tvm_cuda_stream_set_access_policy_window_packed" in source
|
|
assert "__tvm_cuda_stream_reset_access_policy_window_packed" in source
|
|
assert "half*" not in source
|
|
assert "((half*)" not in source
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|