topic Allan, in OpenCL* for CPU

OpenCL Code Builder Deep Analysis error: "Cannot return a color for more than 72"

allanmac1 — Wed, 25 Feb 2015 03:23:11 GMT

Deep Analysis returns the following error and shows an entirely blank "Execution Duration" tab:

The remaining tabs are populated.

I'm running a workgroup of 2688/224 global/local items which is one SIMD8 per hardware thread on an HD6000.

This is on a NUC5i5RYH HD6000 IGP on Win7/x64.

Both 32-bit and 64-bit Code Builder throw the same error.

Allan,

Robert_I_Intel — Fri, 27 Feb 2015 22:49:23 GMT

Allan,

Sorry for getting back to you so late: I was teaching OpenCL classes for the last couple of days.

What kind of kernel are you analyzing? Could you share the code and the steps to reproduce it?

Unfortunately I can't share

allanmac1 — Fri, 27 Feb 2015 23:08:00 GMT

Unfortunately I can't share the code. I was hoping your team could just look at the error and zero in on where it's failing.

There are 60+ registers per work item and 224 x 12 (2688) items in the workgroup.

Haswell has 70 h/w threads per subslice which is less than 72.

But the HD 6000 would have 56 per subslice but 3 subslices and 2 slices. Perhaps that's hitting a hardcoded limit.

The Deep Analysis "EU Statistics" panel shows a single invocation per hardware thread across all possible 7 x 8 x 3 x 2 threads (but mentions "Core" -- see below) so the kernel is executing as I expected.

---

Side note: the labelling scheme used in the EU Statistics page is confusing: "Core W, Slice X, Half-Slice Y, EU Z".

If the Gen8 Compute Architecture Guide is relevant, I would expect to see "Slice X, Sub-Slice Y, EU Z".

But I'm sure your team has its reasons for labelling it that way. :)

Allan,

Robert_I_Intel — Fri, 27 Feb 2015 23:16:56 GMT

Allan,

I just communicated this issue to the development team. I am behind the technology curve: my boss is in the process of ordering HD6000, so I should be able to reproduce issues like this in the very near future.

Just to clarify: Deep Analysis runs fine on Haswell, but bombs on Broadwell?

Sigh, I have an entirely

allanmac1 — Fri, 27 Feb 2015 23:24:00 GMT

Sigh, I have an entirely different Deep Analysis error on Haswell (the dreaded -34) which I wasn't even going to mention. This is with the latest Code Builder. The regular "Analyze" functionality works fine though. Also, Deep Analysis ran fine on the previous release of Code Builder.

So, to answer your question, I've only seen Deep Analysis partially work on Broadwell with the latest Builder. All panels except the "Execution Duration" panel are populated.

Get an HD 6000 NUC. It rocks. :)