CL_OUT_OF_RESOURCES in Command Queue during kernel enqueue

Designing error handling for a small OpenCL application that uses a single in-order Command Queue for kernel execution. The goal is to decide when CL_OUT_OF_RESOURCES should trigger a retry with reduced work-group size versus a hard failure.

Constraints include vendor-dependent ICD behavior, the distinction between global memory allocation failure, register pressure, and local memory over-allocation, and the requirement for explicit synchronization to guarantee memory consistency. Buffer migration between host and device requires explicit enqueue operations, and memory visibility is only guaranteed after a synchronization barrier.

Uncertainty remains about the precise source reported by the runtime and the effect of pending asynchronous operations on resource accounting. Does CL_OUT_OF_RESOURCES returned by clEnqueueNDRangeKernel indicate device memory exhaustion or command queue resource exhaustion? Is the error consistently reported for local memory over-allocation across vendor implementations? What synchronization state is required to reliably distinguish transient queue pressure from a persistent device limit?