SGEMM benchmark config for Vega FE? about tensile

[guacamoleo]

If you mean you were using sgemm_5760.yaml, that produces HIP kernels and 5 TFlops does sound like the best that our compiler can do at this time. If you use sgemm_asm.yaml, then Tensile will produce assembly kernels and you should see over 90% efficiency. David E. Tanner

…

_________________________________________________________________________________ MTS Software Engineer | Radeon Technologies Group – Open Compute From: eqy [mailto:[email protected]] Sent: Monday, October 16, 2017 2:26 PM To: ROCmSoftwarePlatform/Tensile <[email protected]> Cc: Subscribed <[email protected]> Subject: [ROCmSoftwarePlatform/Tensile] SGEMM benchmark config for Vega FE? (#122) Hi, I'm experimenting with SGEMM performance tuning on Vega FE and get around 5 GFLOP/s max with the 5760 benchmark config. I was wondering if there was a pointer to a current best config for Vega/Vega FE that I could use as a starting point that was closer to peak performance? Thanks, Eddie — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#122>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ACiWnhO_IKbkErK8Q9PNm7G_MkwngTmaks5ss63fgaJpZM4P7HuS>.

[eqy]

It is possible to run an apples-to-apples comparison with the sgemm_5760 config (in terms of input size) with the sgemm_asm.yaml config?
I'm getting terminate called after throwing an instance of 'std::bad_alloc' so I'm not sure if that's due to an input size that is too large.

[guacamoleo]

sgemm_5760.yaml is an example of non-batched gemm while sgemm_asm.yaml is an example of batched gemm. In the sgemm_asm.yaml file, find all instances of “Batched: True” and change to False. This should make the two much more similar. David E. Tanner

…

_________________________________________________________________________________ MTS Software Engineer | Radeon Technologies Group – Open Compute From: eqy [mailto:[email protected]] Sent: Monday, October 16, 2017 3:32 PM To: ROCmSoftwarePlatform/Tensile <[email protected]> Cc: Tanner, David <[email protected]>; Comment <[email protected]> Subject: Re: [ROCmSoftwarePlatform/Tensile] SGEMM benchmark config for Vega FE? (#122) It is possible to run an apples-to-apples comparison with the sgemm_5760 config (in terms of input size) with the sgemm_asm.yaml config? I'm getting terminate called after throwing an instance of 'std::bad_alloc' so I'm not sure if that's due to an input size that is too large. — You are receiving this because you commented. Reply to this email directly, view it on GitHub<#122 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ACiWnp5jVJc8nMf3rNm1ziQMuimk2wNIks5ss71dgaJpZM4P7HuS>.

[eqy]

Great, thanks!