Heap corruption when running OpenCL with Intel integrated GPUs

[MichaelZivid]

We are seeing a crash while running OpenCL code on various Intel integrated GPUs.

In particular, the crash occurs after running `clEnqueueWriteBuffer` with `blocking_write=CL_FALSE`, then running a kernel, and then calling `clFinish`. The latter call seems to cause memory heap corruption. The next time the application allocates or deallocates memory, there is an access violation and the program terminates.

The heap corruption is seemingly random or timing-related. We typically only see it every few hours when running the reproducer. We have reproduced this on several different machines with different hardware.

Setting the mentioned blocking write to `blocking_write=CL_TRUE` instead, seems to resolve the issue.

We have seen this crash at least on the following integrated GPUS and GFX driver versions:

- Intel® HD Graphics 630 (30.0.101.1660)
- Intel® UHD Graphics 770 (31.0.101.5333)
- Intel® Iris® Plus Graphics (27.20.100.9664)

We have only been able to reproduce this issue on Windows, not on other platforms. It seems to be an issue isolated to Intel, in particular, we have been testing extensively on Nvidia and AMD as well without seeing a similar issue.

An example of the actual exception causing the crash is as follows:

Unhandled exception at 0x00007FF8AB2F6E25 (ntdll.dll) in opencl_crash_reproducer.exe.732.dmp: 0xC0000005: Access violation reading location 0x000001B4C2A24000.

Please see the attachment for a full reproducer. In addition, detailed system information, build and run instructions, and a core crash dump is included.

Please let us know if we can provide anymore details.

[NormanS_Intel]

Hello MichaelZivid,

Thank you for posting in the community!

To ensure you receive the most specialized assistance, we have a dedicated forum that addresses these specific concerns. Therefore, I will be moving this discussion to our GPU Compute Software Forum. This will allow our knowledgeable community and experts to provide you with timely and accurate solutions.

Best regards,

Norman S.

Intel Customer Support Engineer