Thanks for mmmetal and roserg! I did misunderstanding the SIMDgroup usage, you are right.
By the way, Anyone who intent to use this feature can refer in the implementation in TF-Lite:
https://github.com/alpa-projects/tensorflow-alpa/blob/ee8f6612b515ada4509fa53491c5ba5b3ef8524a/tensorflow/lite/delegates/gpu/common/tasks/conv_metal_simd.cc
PYNing
Users receive reputation points by contributing quality content to the forums.
Posts include post and replies published on the forums.
Post authors can mark replies as Accepted to indicate successful solutions.
Apple Developer Forums admins can mark replies as Apple Recommended to indicate an approved solution
Post
Replies
Boosts
Views
Activity
The bug is fixed.
Reason: Use untracked MTLHeaps, but not using MTLFence to protect the GPU could execute both filters in parallel, and thus read uninitialized dynamic texture data allocated from the heap.
Solution1: set MTLFence, see https://developer.apple.com/documentation/metal/synchronization/implementing_a_multistage_image_filter_using_heaps_and_events
Solution2: turn untracked MTLHeap to tracked MTLHeap, that it is, let heapDescriptor.hazarTrackingMode = MTLHazardTrackingModeTracked, see https://developer.apple.com/documentation/metal/mtlheapdescriptor/3131686-hazardtrackingmode