August 2017 - Anl-intel-lrz-knr - mailman3.cels.anl.gov

[Anl-intel-csa] Canceled: ANL/Intel CSA Workload Characterization
by Petrini, Fabrizio 31 Aug '17

31 Aug '17

Placeholder for weekly ANL/Intel CSA Exascale workload characterization meeting ......................................................................................................................................... --> Join Skype Meeting<https://meet.intel.com/fabrizio.petrini/7K7FM32F> Trouble Joining? Try Skype Web App<https://meet.intel.com/fabrizio.petrini/7K7FM32F?sl=1> Join by phone +1(916)356-2663 (or your local bridge access #) Choose bridge 5.<tel:+1(916)356-2663%20(or%20your%20local%20bridge%20access%20#)%20Choose%20bridge%205.> (Global) English (United States) Find a local number<https://dial.intel.com> Conference ID: 791574018 Forgot your dial-in PIN?<https://dial.intel.com> |Help<https://o15.officeredir.microsoft.com/r/rlidLync15?clid=1033&p1=5&p2=2009> [!OC([1033])!] .........................................................................................................................................

1 0

[Anl-intel-csa] Call tomorrow Sept 1st (Friday) @1:00pm
by Kumaran, Kalyan 31 Aug '17

31 Aug '17

Our usual call scheduled for tomorrow at 1:00pm is canceled. We are going to meet as a small group with a few of the Intel folks to go over the performance projection numbers to be presented at the Design Review. Anyone involved with the applications, to be presented at the design review, and projections are welcome to attend. I have reserved 4313 for the Argonne folks. -Kumar

1 0

Re: [Anl-intel-csa] support for atomics in memory model
by Tannenbaum, Barry M 30 Aug '17

30 Aug '17

Hi Chris. This is a loose end that I hope to tie up this week. Thanks for reminding us about it. - Barry -----Original Message----- From: anl-intel-csa [mailto:[email protected]] On Behalf Of Christopher Knight Sent: Tuesday, August 29, 2017 11:39 AM To: anl-intel-csa(a)jlse.anl.gov Subject: [Anl-intel-csa] support for atomics in memory model Hello, Is there an ETA for when support for atomics would be included in the memory model? I see the message below. chris ================= [knightc@jlselogin1 bond_harmonic_csa_kernel]$ env OMP_NUM_THREADS=1 CSASIM_MODEL_MEMORY=1 ../../lammps-31Mar2017/src/lmp_jlse_csa -i in.rhodo-charmm LAMMPS (31 Mar 2017) using 1 OpenMP thread(s) per MPI task allocating csa in BondHarmonicCSA... assembling bond_harmonic_csa_kernel.s... lookup… ... BondHarmonicCSA::evflag combination not yet supported: eflag= 1 vflag= 2 Using simple memory model lmp_jlse_csa: /nfs/hd/disks/csaobj/comp/fsim/2017-08-24-r2-release/csa.h:7974: virtual CSASIM::Cycle CSASIM::atmaddf64_op::issuableAt(CSASIM::Cycle): Assertion `0 && "useAsim() for atomics is not yet implemented."' failed. Aborted _______________________________________________ anl-intel-csa mailing list anl-intel-csa(a)lists.jlse.anl.gov https://lists.jlse.anl.gov/mailman/listinfo/anl-intel-csa

1 0

[Anl-intel-csa] support for atomics in memory model
by Christopher Knight 29 Aug '17

29 Aug '17

Hello, Is there an ETA for when support for atomics would be included in the memory model? I see the message below. chris ================= [knightc@jlselogin1 bond_harmonic_csa_kernel]$ env OMP_NUM_THREADS=1 CSASIM_MODEL_MEMORY=1 ../../lammps-31Mar2017/src/lmp_jlse_csa -i in.rhodo-charmm LAMMPS (31 Mar 2017) using 1 OpenMP thread(s) per MPI task allocating csa in BondHarmonicCSA... assembling bond_harmonic_csa_kernel.s... lookup… ... BondHarmonicCSA::evflag combination not yet supported: eflag= 1 vflag= 2 Using simple memory model lmp_jlse_csa: /nfs/hd/disks/csaobj/comp/fsim/2017-08-24-r2-release/csa.h:7974: virtual CSASIM::Cycle CSASIM::atmaddf64_op::issuableAt(CSASIM::Cycle): Assertion `0 && "useAsim() for atomics is not yet implemented."' failed. Aborted

1 0

Re: [Anl-intel-csa] CSA builtin parallel functions
by Laosooksathit, Supada 28 Aug '17

28 Aug '17

Hi Kent, Got it! Thank you. It makes more sense now. Best, Supada From: Glossop, Kent Sent: Monday, August 28, 2017 11:56 AM To: Laosooksathit, Supada <supada.laosooksathit(a)intel.com>; Woodworth, Daniel <daniel.woodworth(a)intel.com>; anl-intel-csa(a)lists.jlse.anl.gov Subject: RE: CSA builtin parallel functions > Do I have to put a __builtin_csa_parallel_section_entry after every __builtin_csa_parallel_region_entry? A parallel region describes a range in which independent sections (tasks) may exist. A region by itself isn’t useful. Even a region with a single dynamically execute section isn’t particularly useful. What is useful is a region in which multiple sections can be executed in parallel, allowing the compiler to understand there are no dependencies between the independent sections, prior to exiting the region. The simplest case of that is typically a loop, where you have: region_entry() loop start section_entry() // every dynamic instance of the loop body is described as independent … section_exit() loop end region_exit() > Do you know if I can use multiple regions in the nested loop? Each region has 0 or more sections in it, but there can be multiple regions, and they can be nested. The following would be valid: r0 = region_entry() // a parallel region containing 2 independent loops r0s0 = section_entry(r0) // this region and r0s1 can execute in parallel loop1 start r1 = region_entry() // a nested region containing a loop with independent iterations loop1n start r1s0 = section_entry(r1) // iterations can overlap, be reordered, etc. … section_exit(r1s0) loop1n end region_exit(r1) loop1 end section_exit(r0s0) r0s1 = section_entry(r0) loop2 start … loop2 end section_exit(r0s1) region_exit(r0) From: anl-intel-csa [mailto:[email protected]] On Behalf Of Laosooksathit, Supada Sent: Monday, August 28, 2017 11:52 AM To: Woodworth, Daniel <daniel.woodworth(a)intel.com<mailto:[email protected]>>; anl-intel-csa(a)lists.jlse.anl.gov<mailto:[email protected]> Subject: Re: [Anl-intel-csa] CSA builtin parallel functions Daniel, Thank you for your response. Do you know if I can use multiple regions in the nested loop? Do I have to put a __builtin_csa_parallel_section_entry after every __builtin_csa_parallel_region_entry? Thank you, Supada From: Woodworth, Daniel Sent: Monday, August 28, 2017 10:13 AM To: Laosooksathit, Supada <supada.laosooksathit(a)intel.com<mailto:[email protected]>>; anl-intel-csa(a)lists.jlse.anl.gov<mailto:[email protected]> Subject: RE: CSA builtin parallel functions There’s a quick example at the end of the release notes where these were added, but it could probably use a little more documentation. Adding these intrinsics to a loop is pretty straightforward, though: 1. Take the loop that needs to be converted (if you aren’t sure which one it is, compile with -g and the warning message will give you line numbers) and delete the __builtin_csa_parallel_loop intrinsic. 2. Put a __builtin_csa_parallel_region_entry before the start of the loop. This takes one int argument, which is used to identify the region and can be any number as long as it is unique (the compiler starts assigning these at 1000 when it expands __builtin_csa_parallel_loops, so numbers below 1000 are advised). The value returned from this intrinsic is another int which is passed into the __builtin_csa_parallel_region_exit and __builtin_csa_parallel_section_entry intrinsics to identify which region entry they correspond to, so you’ll want to save that in a variable. 3. Put a __builtin_csa_parallel_region_exit after the end of the loop. This takes one argument which should be the return value from the __builtin_csa_parallel_region_entry earlier. 4. Put a __builtin_csa_parallel_section_entry inside of the loop before all of the memory operations in the loop that you want to be parallelized. This takes one argument which should be the __builtin_csa_parallel_region_entry return value. It returns another token value that’s used to identify the section, and that is consumed by __builtin_csa_parallel_section_exit. 5. Put a __builtin_ csa_parallel_section_exit inside of the loop after all of the memory operations that you want parallelized. This takes one argument which should be the __builtin_csa_parallel_section_entry return value. So, if you started with a loop that looks like this: __builtin_csa_parallel_loop(); for (int i = 0; i < n; ++i) { dest[i] = src[i]; } you should end up with a loop like this after converting it: int region_tok = __builtin_csa_parallel_region_entry(0); for (int i = 0; i < n; ++i) { int section_tok = __builtin_csa_parallel_section_entry(region_tok); dest[i] = src[i]; __builtin_csa_parallel_section_exit(section_tok); } __builtin_csa_parallel_region_exit(region_tok); From: anl-intel-csa [mailto:[email protected]] On Behalf Of Laosooksathit, Supada Sent: Monday, August 28, 2017 10:32 AM To: anl-intel-csa(a)lists.jlse.anl.gov<mailto:[email protected]> Subject: [Anl-intel-csa] CSA builtin parallel functions Hello, I am not sure if anyone has seen this. I am trying to run sgemm with __builtin_csa_parallel_loop(), but received the following warning: !! WARNING: COULD NOT PARALLELIZE LOOP !! We were unable to automatically identify a unique section for a loop marked with __builtin_csa_parallel_loop() (use -g for location information). This was likely caused by either having multiple loop exits or by having memory operations in the loop control. Please mark the regions and sections for this loop explicitly with __builtin_csa_parallel_(region|section)_(entry|exit) instead. I am also trying to use __builtin_csa_parallel_(region|section)_(entry|exit) as suggested, but very confused with how to use it. Thank you, Supada

1 0

[Anl-intel-csa] CSA SDK Release Notes for 2017/08/24 version
by Glossop, Kent 28 Aug '17

28 Aug '17

CSA SDK Release Notes for 2017/08/24 version The default release on JLSE has been changed to 2017-08-24-r2-release Compiler / Runtime * Has various fixes to the support for _builtin_csa_parallel* functions for reported issues Simulator * A simple model of the memory subsystem has been added to the functional simulator. Since there's a performance penalty for using the simple memory model, it is disabled by default. To enable the simple memory model, define the environment variable CSASIM_MODEL_MEMORY before running your application. For example: export CSASIM_MODEL_MEMORY=1 When the simple memory model initializes, it will print out "Using simple memory model" to confirm that it's activated. The simple memory model writes additional statistics to the .stat file, including CSA cache misses, CSA cache hits, CSA cache bandwidth, and memory bandwidth. Note: We recommend any performance estimates use this model rather than relying only on the perfect cache model. The values will be lower, but more realistic. Note that they may still be slightly optimistic, but should be fairly close to the currently projected design.

1 0

Re: [Anl-intel-csa] CSA builtin parallel functions
by Glossop, Kent 28 Aug '17

28 Aug '17

> Do I have to put a __builtin_csa_parallel_section_entry after every __builtin_csa_parallel_region_entry? A parallel region describes a range in which independent sections (tasks) may exist. A region by itself isn’t useful. Even a region with a single dynamically execute section isn’t particularly useful. What is useful is a region in which multiple sections can be executed in parallel, allowing the compiler to understand there are no dependencies between the independent sections, prior to exiting the region. The simplest case of that is typically a loop, where you have: region_entry() loop start section_entry() // every dynamic instance of the loop body is described as independent … section_exit() loop end region_exit() > Do you know if I can use multiple regions in the nested loop? Each region has 0 or more sections in it, but there can be multiple regions, and they can be nested. The following would be valid: r0 = region_entry() // a parallel region containing 2 independent loops r0s0 = section_entry(r0) // this region and r0s1 can execute in parallel loop1 start r1 = region_entry() // a nested region containing a loop with independent iterations loop1n start r1s0 = section_entry(r1) // iterations can overlap, be reordered, etc. … section_exit(r1s0) loop1n end region_exit(r1) loop1 end section_exit(r0s0) r0s1 = section_entry(r0) loop2 start … loop2 end section_exit(r0s1) region_exit(r0) From: anl-intel-csa [mailto:[email protected]] On Behalf Of Laosooksathit, Supada Sent: Monday, August 28, 2017 11:52 AM To: Woodworth, Daniel <daniel.woodworth(a)intel.com>; anl-intel-csa(a)lists.jlse.anl.gov Subject: Re: [Anl-intel-csa] CSA builtin parallel functions Daniel, Thank you for your response. Do you know if I can use multiple regions in the nested loop? Do I have to put a __builtin_csa_parallel_section_entry after every __builtin_csa_parallel_region_entry? Thank you, Supada From: Woodworth, Daniel Sent: Monday, August 28, 2017 10:13 AM To: Laosooksathit, Supada <supada.laosooksathit(a)intel.com<mailto:[email protected]>>; anl-intel-csa(a)lists.jlse.anl.gov<mailto:[email protected]> Subject: RE: CSA builtin parallel functions There’s a quick example at the end of the release notes where these were added, but it could probably use a little more documentation. Adding these intrinsics to a loop is pretty straightforward, though: 1. Take the loop that needs to be converted (if you aren’t sure which one it is, compile with -g and the warning message will give you line numbers) and delete the __builtin_csa_parallel_loop intrinsic. 2. Put a __builtin_csa_parallel_region_entry before the start of the loop. This takes one int argument, which is used to identify the region and can be any number as long as it is unique (the compiler starts assigning these at 1000 when it expands __builtin_csa_parallel_loops, so numbers below 1000 are advised). The value returned from this intrinsic is another int which is passed into the __builtin_csa_parallel_region_exit and __builtin_csa_parallel_section_entry intrinsics to identify which region entry they correspond to, so you’ll want to save that in a variable. 3. Put a __builtin_csa_parallel_region_exit after the end of the loop. This takes one argument which should be the return value from the __builtin_csa_parallel_region_entry earlier. 4. Put a __builtin_csa_parallel_section_entry inside of the loop before all of the memory operations in the loop that you want to be parallelized. This takes one argument which should be the __builtin_csa_parallel_region_entry return value. It returns another token value that’s used to identify the section, and that is consumed by __builtin_csa_parallel_section_exit. 5. Put a __builtin_ csa_parallel_section_exit inside of the loop after all of the memory operations that you want parallelized. This takes one argument which should be the __builtin_csa_parallel_section_entry return value. So, if you started with a loop that looks like this: __builtin_csa_parallel_loop(); for (int i = 0; i < n; ++i) { dest[i] = src[i]; } you should end up with a loop like this after converting it: int region_tok = __builtin_csa_parallel_region_entry(0); for (int i = 0; i < n; ++i) { int section_tok = __builtin_csa_parallel_section_entry(region_tok); dest[i] = src[i]; __builtin_csa_parallel_section_exit(section_tok); } __builtin_csa_parallel_region_exit(region_tok); From: anl-intel-csa [mailto:[email protected]] On Behalf Of Laosooksathit, Supada Sent: Monday, August 28, 2017 10:32 AM To: anl-intel-csa(a)lists.jlse.anl.gov<mailto:[email protected]> Subject: [Anl-intel-csa] CSA builtin parallel functions Hello, I am not sure if anyone has seen this. I am trying to run sgemm with __builtin_csa_parallel_loop(), but received the following warning: !! WARNING: COULD NOT PARALLELIZE LOOP !! We were unable to automatically identify a unique section for a loop marked with __builtin_csa_parallel_loop() (use -g for location information). This was likely caused by either having multiple loop exits or by having memory operations in the loop control. Please mark the regions and sections for this loop explicitly with __builtin_csa_parallel_(region|section)_(entry|exit) instead. I am also trying to use __builtin_csa_parallel_(region|section)_(entry|exit) as suggested, but very confused with how to use it. Thank you, Supada

1 0

Re: [Anl-intel-csa] CSA builtin parallel functions
by Laosooksathit, Supada 28 Aug '17

28 Aug '17

Daniel, Thank you for your response. Do you know if I can use multiple regions in the nested loop? Do I have to put a __builtin_csa_parallel_section_entry after every __builtin_csa_parallel_region_entry? Thank you, Supada From: Woodworth, Daniel Sent: Monday, August 28, 2017 10:13 AM To: Laosooksathit, Supada <supada.laosooksathit(a)intel.com>; anl-intel-csa(a)lists.jlse.anl.gov Subject: RE: CSA builtin parallel functions There’s a quick example at the end of the release notes where these were added, but it could probably use a little more documentation. Adding these intrinsics to a loop is pretty straightforward, though: 1. Take the loop that needs to be converted (if you aren’t sure which one it is, compile with -g and the warning message will give you line numbers) and delete the __builtin_csa_parallel_loop intrinsic. 2. Put a __builtin_csa_parallel_region_entry before the start of the loop. This takes one int argument, which is used to identify the region and can be any number as long as it is unique (the compiler starts assigning these at 1000 when it expands __builtin_csa_parallel_loops, so numbers below 1000 are advised). The value returned from this intrinsic is another int which is passed into the __builtin_csa_parallel_region_exit and __builtin_csa_parallel_section_entry intrinsics to identify which region entry they correspond to, so you’ll want to save that in a variable. 3. Put a __builtin_csa_parallel_region_exit after the end of the loop. This takes one argument which should be the return value from the __builtin_csa_parallel_region_entry earlier. 4. Put a __builtin_csa_parallel_section_entry inside of the loop before all of the memory operations in the loop that you want to be parallelized. This takes one argument which should be the __builtin_csa_parallel_region_entry return value. It returns another token value that’s used to identify the section, and that is consumed by __builtin_csa_parallel_section_exit. 5. Put a __builtin_ csa_parallel_section_exit inside of the loop after all of the memory operations that you want parallelized. This takes one argument which should be the __builtin_csa_parallel_section_entry return value. So, if you started with a loop that looks like this: __builtin_csa_parallel_loop(); for (int i = 0; i < n; ++i) { dest[i] = src[i]; } you should end up with a loop like this after converting it: int region_tok = __builtin_csa_parallel_region_entry(0); for (int i = 0; i < n; ++i) { int section_tok = __builtin_csa_parallel_section_entry(region_tok); dest[i] = src[i]; __builtin_csa_parallel_section_exit(section_tok); } __builtin_csa_parallel_region_exit(region_tok); From: anl-intel-csa [mailto:[email protected]] On Behalf Of Laosooksathit, Supada Sent: Monday, August 28, 2017 10:32 AM To: anl-intel-csa(a)lists.jlse.anl.gov<mailto:[email protected]> Subject: [Anl-intel-csa] CSA builtin parallel functions Hello, I am not sure if anyone has seen this. I am trying to run sgemm with __builtin_csa_parallel_loop(), but received the following warning: !! WARNING: COULD NOT PARALLELIZE LOOP !! We were unable to automatically identify a unique section for a loop marked with __builtin_csa_parallel_loop() (use -g for location information). This was likely caused by either having multiple loop exits or by having memory operations in the loop control. Please mark the regions and sections for this loop explicitly with __builtin_csa_parallel_(region|section)_(entry|exit) instead. I am also trying to use __builtin_csa_parallel_(region|section)_(entry|exit) as suggested, but very confused with how to use it. Thank you, Supada

1 0

Re: [Anl-intel-csa] CSA builtin parallel functions
by Woodworth, Daniel 28 Aug '17

28 Aug '17

There’s a quick example at the end of the release notes where these were added, but it could probably use a little more documentation. Adding these intrinsics to a loop is pretty straightforward, though: 1. Take the loop that needs to be converted (if you aren’t sure which one it is, compile with -g and the warning message will give you line numbers) and delete the __builtin_csa_parallel_loop intrinsic. 2. Put a __builtin_csa_parallel_region_entry before the start of the loop. This takes one int argument, which is used to identify the region and can be any number as long as it is unique (the compiler starts assigning these at 1000 when it expands __builtin_csa_parallel_loops, so numbers below 1000 are advised). The value returned from this intrinsic is another int which is passed into the __builtin_csa_parallel_region_exit and __builtin_csa_parallel_section_entry intrinsics to identify which region entry they correspond to, so you’ll want to save that in a variable. 3. Put a __builtin_csa_parallel_region_exit after the end of the loop. This takes one argument which should be the return value from the __builtin_csa_parallel_region_entry earlier. 4. Put a __builtin_csa_parallel_section_entry inside of the loop before all of the memory operations in the loop that you want to be parallelized. This takes one argument which should be the __builtin_csa_parallel_region_entry return value. It returns another token value that’s used to identify the section, and that is consumed by __builtin_csa_parallel_section_exit. 5. Put a __builtin_ csa_parallel_section_exit inside of the loop after all of the memory operations that you want parallelized. This takes one argument which should be the __builtin_csa_parallel_section_entry return value. So, if you started with a loop that looks like this: __builtin_csa_parallel_loop(); for (int i = 0; i < n; ++i) { dest[i] = src[i]; } you should end up with a loop like this after converting it: int region_tok = __builtin_csa_parallel_region_entry(0); for (int i = 0; i < n; ++i) { int section_tok = __builtin_csa_parallel_section_entry(region_tok); dest[i] = src[i]; __builtin_csa_parallel_section_exit(section_tok); } __builtin_csa_parallel_region_exit(region_tok); From: anl-intel-csa [mailto:[email protected]] On Behalf Of Laosooksathit, Supada Sent: Monday, August 28, 2017 10:32 AM To: anl-intel-csa(a)lists.jlse.anl.gov Subject: [Anl-intel-csa] CSA builtin parallel functions Hello, I am not sure if anyone has seen this. I am trying to run sgemm with __builtin_csa_parallel_loop(), but received the following warning: !! WARNING: COULD NOT PARALLELIZE LOOP !! We were unable to automatically identify a unique section for a loop marked with __builtin_csa_parallel_loop() (use -g for location information). This was likely caused by either having multiple loop exits or by having memory operations in the loop control. Please mark the regions and sections for this loop explicitly with __builtin_csa_parallel_(region|section)_(entry|exit) instead. I am also trying to use __builtin_csa_parallel_(region|section)_(entry|exit) as suggested, but very confused with how to use it. Thank you, Supada

1 0

[Anl-intel-csa] CSA builtin parallel functions
by Laosooksathit, Supada 28 Aug '17

28 Aug '17

Hello, I am not sure if anyone has seen this. I am trying to run sgemm with __builtin_csa_parallel_loop(), but received the following warning: !! WARNING: COULD NOT PARALLELIZE LOOP !! We were unable to automatically identify a unique section for a loop marked with __builtin_csa_parallel_loop() (use -g for location information). This was likely caused by either having multiple loop exits or by having memory operations in the loop control. Please mark the regions and sections for this loop explicitly with __builtin_csa_parallel_(region|section)_(entry|exit) instead. I am also trying to use __builtin_csa_parallel_(region|section)_(entry|exit) as suggested, but very confused with how to use it. Thank you, Supada

1 0