May 2023 - Ecp-aurora-sdk-discuss - mailman3.cels.anl.gov

Re: [Ecp-aurora-sdk-discuss] feedback about MKL_DIRECT_CALL-like feature for GPU in MKL
by Bertoni, Colleen 15 May '23

15 May '23

Thanks for the responses! Good questions, I don't know the answer--we'll ask the MKL team at the next WG call. Colleen ________________________________ From: Jin, Xiaoyong <xjin(a)anl.gov> Sent: Friday, May 12, 2023 4:49 PM To: Bertoni, Colleen <bertoni(a)anl.gov> Cc: ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Subject: Re: [Ecp-aurora-sdk-discuss] feedback about MKL_DIRECT_CALL-like feature for GPU in MKL Hi Colleen, Do you know if there's way to measure the overhead of parameter checking currently with the default mode? And how can we compare this overhead to kernel launching overhead? Additionally, how much overhead is there currently for computational error checking? The second question is easy to answer. I would prefer an environment variable to avoid the need of recompiling everything. Best, Xiaoyong > On May 12, 2023, at 4:33 PM, Bertoni, Colleen via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> wrote: > > Hi everyone, > > We had a request from the MKL team to see if anyone was using or previously used MKL_DIRECT_CALL in your CPU MKL code (https://www.intel.com/content/www/us/en/docs/onemkl/developer-guide-linux/2…) They are considering implementing something similar for GPU (i.e. if you compile with a special flag or set an environment variable, it would avoid the argument checking MKL typically does to remove the overhead from those). > > In particular, they wanted feedback about: > > - What would be useful to disabled: input parameter checking (as with MKL_DIRECT_CALL on the CPU) or is disabling computational error more desirable (e.g. singular matrix check for LU)? > - Do you have any preference on how to enable/disable error checking? For example, an environment variable for all oneMKL or runtime setting or compiler setting? > > Thanks, > Colleen > -- > Ecp-aurora-sdk-discuss mailing list > Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov > https://lists.jlse.anl.gov/mailman/listinfo/ecp-aurora-sdk-discuss

1 0

Re: [Ecp-aurora-sdk-discuss] feedback about MKL_DIRECT_CALL-like feature for GPU in MKL
by Jin, Xiaoyong 12 May '23

12 May '23

Hi Colleen, Do you know if there's way to measure the overhead of parameter checking currently with the default mode? And how can we compare this overhead to kernel launching overhead? Additionally, how much overhead is there currently for computational error checking? The second question is easy to answer. I would prefer an environment variable to avoid the need of recompiling everything. Best, Xiaoyong > On May 12, 2023, at 4:33 PM, Bertoni, Colleen via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> wrote: > > Hi everyone, > > We had a request from the MKL team to see if anyone was using or previously used MKL_DIRECT_CALL in your CPU MKL code (https://www.intel.com/content/www/us/en/docs/onemkl/developer-guide-linux/2…) They are considering implementing something similar for GPU (i.e. if you compile with a special flag or set an environment variable, it would avoid the argument checking MKL typically does to remove the overhead from those). > > In particular, they wanted feedback about: > > - What would be useful to disabled: input parameter checking (as with MKL_DIRECT_CALL on the CPU) or is disabling computational error more desirable (e.g. singular matrix check for LU)? > - Do you have any preference on how to enable/disable error checking? For example, an environment variable for all oneMKL or runtime setting or compiler setting? > > Thanks, > Colleen > -- > Ecp-aurora-sdk-discuss mailing list > Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov > https://lists.jlse.anl.gov/mailman/listinfo/ecp-aurora-sdk-discuss

1 0

feedback about MKL_DIRECT_CALL-like feature for GPU in MKL
by Bertoni, Colleen 12 May '23

12 May '23

Hi everyone, We had a request from the MKL team to see if anyone was using or previously used MKL_DIRECT_CALL in your CPU MKL code (https://www.intel.com/content/www/us/en/docs/onemkl/developer-guide-linux/2…) They are considering implementing something similar for GPU (i.e. if you compile with a special flag or set an environment variable, it would avoid the argument checking MKL typically does to remove the overhead from those). In particular, they wanted feedback about: - What would be useful to disabled: input parameter checking (as with MKL_DIRECT_CALL on the CPU) or is disabling computational error more desirable (e.g. singular matrix check for LU)? - Do you have any preference on how to enable/disable error checking? For example, an environment variable for all oneMKL or runtime setting or compiler setting? Thanks, Colleen

1 0