DPAS instruction on PVC #14371
-
|
I'm trying to port some code that I've developed for an A770 to PVC that uses DPAS instructions. From what I gather, the difference in the interface for the DPAS instructions is that the N dimension of the matrix multiplication (referred to as systolic depth) increases from 8 to 16. When I change this parameter to the esimd::xmx::dpas<> template, I get a static assertion error from sycl/ext/intel/esimd/xmx/dpas.hpp stating that the systolic depth must be 8. Is this a bug or is there another way I should be using DPAS instructions on PVC? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 4 replies
-
|
@intel/dpcpp-esimd-reviewers, FYI. |
Beta Was this translation helpful? Give feedback.
-
|
@fineg74 Do you mind taking a look at this? Thanks |
Beta Was this translation helpful? Give feedback.
-
|
I might be mixing up systolic depth and execution width. I'll see if I can change the parameters around a bit and see if I can get things to work. |
Beta Was this translation helpful? Give feedback.
-
|
Apologies, it looks like I was mixing up the dimensions. Systolic depth refers to part of the reduction K dimension, repeat count refers to M, and execution width refers to N. Execution width is implicit whereas systolic depth and repeat count are arguments to the dpas template. Reseting systolic depth back to 8 fixed this, after adjusting the esimd vector sizes appropriately for the new execution width. Thanks for the quick triage on this issue! |
Beta Was this translation helpful? Give feedback.
-
|
Glad it's working! |
Beta Was this translation helpful? Give feedback.
Apologies, it looks like I was mixing up the dimensions. Systolic depth refers to part of the reduction K dimension, repeat count refers to M, and execution width refers to N. Execution width is implicit whereas systolic depth and repeat count are arguments to the dpas template. Reseting systolic depth back to 8 fixed this, after adjusting the esimd vector sizes appropriately for the new execution width. Thanks for the quick triage on this issue!