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
We are developing with RK3588 and tried yolov7 model with RKNN Api version 1.5.2 and the Kernel version is 5.10.
With RKNPU 0.9.2 the kernel crash in the log file occurs (BUG: spinlock recursion on CPU#0) if the NPU core selection is auto mod.
In our 4-threads scenario we have 4 different rknn contexts, the device is crashing in the default auto core selection mode. Our tests:
Calling rknn_set_core_mask(ctx, RKNN_NPU_CORE_0_1_2) in all threads led to kernel crash
Calling rknn_set_core_mask(ctx, RKNN_NPU_CORE_0) in 1st and 2nd thread, calling rknn_set_core_mask(ctx, RKNN_NPU_CORE_1) in 3rd thread and calling rknn_set_core_mask(ctx, RKNN_NPU_CORE_2) in 4th thread; led to kernel crash
Calling rknn_set_core_mask(ctx, RKNN_NPU_CORE_0) in all threads worked but the inference time notably increased
When RKNPU driver is downgraded to 0.9.1, the problem is solved. Please, fix the current version, 0.9.2.
Hello,
We are developing with RK3588 and tried yolov7 model with RKNN Api version 1.5.2 and the Kernel version is 5.10.
With RKNPU 0.9.2 the kernel crash in the log file occurs (BUG: spinlock recursion on CPU#0) if the NPU core selection is auto mod.
RK3588-SpinlockRecursion.log
The usage scenario is below.
In our 4-threads scenario we have 4 different rknn contexts, the device is crashing in the default auto core selection mode. Our tests:
When RKNPU driver is downgraded to 0.9.1, the problem is solved. Please, fix the current version, 0.9.2.
Regards,
Muhammed
muhammed@dtsis.com
The text was updated successfully, but these errors were encountered: