OneDNN with Level Zero GPU runtime leaks memory

[jkasprza]

Summary

When running OneDNN with Level Zero GPU runtime and enabling Level Zero validation layers memory leaks are detected. Originally this issue was observed during OpenVINO inference, but I was able to reproduce it with benchdnn.

Version

OneDNN githash: 470e87e

Steps to reproduce

1. Build OneDNN with L0 GPU runtime -DDNNL_GPU_RUNTIME=ZE -DONEDNN_BUILD_GRAPH=OFF
2. Setup environment variables to enable L0 validation layers with leak detector

ZE_ENABLE_VALIDATION_LAYER=1
ZEL_ENABLE_BASIC_LEAK_CHECKER=1

3. Run any GPU primitive with benchdnn. For example:

benchdnn --matmul --engine=gpu --stag=abc --dtag=abc --wtag=cab --dt=f32 --bia-dt=f32 --bia_mask=4 --attr-scratchpad=user 1x30x512:1x512x37

Observed behavior

Benchdnn executes primitive and finishes successfully but leaks are detected.

ZE_ENABLE_VALIDATION_LAYER=1 ZEL_ENABLE_BASIC_LEAK_CHECKER=1 ./benchdnn --matmul --engine=gpu --stag=abc --dtag=abc --wtag=cab --dt=f32 --bia-dt=f32 --bia_mask=4 --attr-scratchpad=user 1x30x512:1x512x37
0:PASSED (134 ms) __REPRO: --matmul --engine=gpu --stag=abc --wtag=cab --dtag=abc --bia-dt=f32 --bia_mask=4 --attr-scratchpad=user 1x30x512:1x512x37
============================================================
= Implementation statistics (--summary=no-impl to disable) =
============================================================
| jit:gemm:any : 1 (100%)                                  |
============================================================
tests:1 passed:1 skipped:0 mistrusted:0 unimplemented:0 invalid_arguments:0 failed:0 listed:0
total: 0.14s; create_pd: 0.02s (13%); create_prim: 0.00s (0%); fill: 0.00s (2%); execute: 0.00s (1%); compute_ref: 0.00s (0%); compare: 0.00s (0%);
Check balance of create/destroy calls
----------------------------------------------------------
               zeContextCreate = 1     \--->              zeContextDestroy = 0     ---> LEAK = 1
          zeCommandQueueCreate = 0     \--->         zeCommandQueueDestroy = 0
                zeModuleCreate = 2     \--->               zeModuleDestroy = 1     ---> LEAK = 1
                zeKernelCreate = 2     \--->               zeKernelDestroy = 1     ---> LEAK = 1
             zeEventPoolCreate = 0     \--->            zeEventPoolDestroy = 0
  zeCommandListCreateImmediate = 2     |
           zeCommandListCreate = 0     \--->          zeCommandListDestroy = 2
                 zeEventCreate = 0     |
   zexCounterBasedEventCreate2 = 0     \--->                zeEventDestroy = 0
                 zeFenceCreate = 0     \--->                zeFenceDestroy = 0
                 zeImageCreate = 0     |
          zeImageViewCreateExt = 0     \--->                zeImageDestroy = 0
               zeSamplerCreate = 0     \--->              zeSamplerDestroy = 0
              zeMemAllocDevice = 4     |
                zeMemAllocHost = 8     |
              zeMemAllocShared = 0     \--->                     zeMemFree = 12
                                       \--->                  zeMemFreeExt = 0

Expected behavior

OneDNN should correctly handle and free all owned resources.