You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
test_runtime / test_runtime_external on Linux: stale kernel-name expectations, and an all-zero dec1 out from the host emulator #278
On Linux (Ubuntu 24.04, clang 18), after the saturation expectation in the tests is corrected (RGBA16F storage saturates to +-65504, so the
expected value is clipped before the FP16 cast; this also covers the +inf mismatches in test_direct_fp16_output and test_fused_post_whole_frame), 6 test_runtime tests still fail:
swin_fused is dec2,dec1 where the tests expect dec1 (test_default_kernels_whole_frame in two classes, test_swin_mid_whole_frame, test_fused_dec1_dec0_whole_frame expects ('fused','none') but gets ('fused','dec2')). Is the new default intended, so the tests
should change, or should the default stay dec1?
test_config_option and test_environment_variable_and_stable (via check_frame / run_sequence): the host emulator's dec1 out
is all zero while the reference is not (16379 of 16384 elements differ). Is this a real emulator/runtime problem, or an expectation
that changed with the fused dec2,dec1 path? Not analysed further.
In test_runtime_external, test_half_output_default_kernels and test_half_output_pack12_default expect filter_last = post_filter_c1_s4_o16 but get post_filter_c1_s4_h_band. Which is intended?
Question about the infinity outputs: with an unclipped reference, the fused-post test (test_runtime.py:317) reports an infinity mismatch
(the kernel gives 65504 where astype(np.float16) gives inf). That is the saturation expectation above, not a defect we can see; but if
the saturation is not intended for the fused _o16 kernels, that expectation should be revisited. With the clipped expectation the test
passes (Ubuntu 24.04 and Fedora, below).
Activity