November 2019 - Ecp-aurora-sdk-discuss

ComputeCpp v1.1.6 available
by Knight, Christopher J. 19 Nov '19

19 Nov '19

Just an FYI that the sycl_codeplay modules have been updated to point to v1.1.6 as of yesterday morning. For example, this is the version that ‘module load sycl_codeplay’ will make accessible. The README file has been updated with reported improvements and bug fixes. In particular, the change in how work-items are mapped is now publicly available. We would be interested to hear about any considerable performance changes you observed compared to v1.1.5. If issues are encountered with this newer version, then we can roll back to v1.1.5. chris ============= Downloaded v1.1.6 from https://developer.codeplay.com/computecppce/latest/download on November 18, 2019. Community Edition 1.1.6 Improvement Enqueue index is now double flipped (see blog post for details) SYCL alias to cl::sycl is enabled by default on modern compilers Bug Fixes Transaction order is now defended against accidental reordering preventing unexpected scheduler pending list changes. Improvement Changes to Work-item Mapping Optimization: https://codeplay.com/portal/11-18-19-computecpp-v1-1-6-changes-to-work-item…

1 0

Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today
by Rosales-fernandez, Carlos 19 Nov '19

19 Nov '19

Just an update to the list. After additional debug work by Thomas we have been able to reproduce the issue in a simplified code. It seems to be a problem in the NEO driver itself. We have filed a bug with the developer team and will be tracking this together with other issues related to the project. At the end the consistent runtime error received was the following: Abort was called at 92 line in file: /builddir/build/BUILD/compute-runtime-19.43.14583/runtime/os_interface/linux/drm_command_stream.inl Aborted So if you see this in your tests feel free to let us know, but we are already tracking this particular case. Thank you Thomas for the follow up work isolating the issue! Carlos From: Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss-bounces(a)lists.jlse.anl.gov> on behalf of "Rosales-fernandez, Carlos via Ecp-aurora-sdk-discuss" <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Reply-To: "Rosales-fernandez, Carlos" <carlos.rosales-fernandez(a)intel.com> Date: Friday, November 15, 2019 at 9:16 AM To: "Peterka, Thomas" <tpeterka(a)mcs.anl.gov> Cc: Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Subject: Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today Strange – I can also build and run these two in the iris nodes with a minor change (remove the “sycl::” in the usm allocator in example 1, which makes sense since I used the cl::sycl namespace). There must be an issue with your environment that is causing this – can we follow up offline and drop the list? We can update everyone later if we find a general issue. Carlos From: "Peterka, Thomas" <tpeterka(a)mcs.anl.gov> Date: Friday, November 15, 2019 at 8:34 AM To: "Rosales-fernandez, Carlos" <carlos.rosales-fernandez(a)intel.com> Cc: Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Subject: Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today Hi Carlos, I think there may be something wrong with this latest drop, because SYCL code that used to work now crashes, and reverting to the old version version of dpcpp/2019.9.30.013b continues to work. Specifically, I see two problems: 1. My test of USM generates the following run time error: symbol lookup error: ./fixed: undefined symbol: _ZN2cl4sycl6detail3usm12alignedAllocEmmRKNS0_7contextENS0_3usm5allocE The code is this: auto dev = Queue.get_device(); auto ctxt = Queue.get_context(); sycl::usm_allocator<int, sycl::usm::alloc::host> alloc(ctxt, dev); std::vector<int, decltype(alloc)> n(alloc); n.resize(1); The last line, resize, is where the error occurs. The same problem happens if I give the size in the constructor and don’t call the resize. 2. The following trivial example generates a segmentation fault: sycl::queue Queue; sycl::buffer<sycl::cl_int, 1> Buffer(4); sycl::range<1> NumOfWorkItems{Buffer.get_count()}; Queue.submit([&](sycl::handler &cgh) { auto Accessor = Buffer.get_access<sycl::access::mode::write>(cgh); cgh.parallel_for<class FillBuffer>(NumOfWorkItems, [=](sycl::id<1> WIid) { Accessor[WIid] = (cl::sycl::cl_int)WIid.get(0); }); }); The seg fault happens in the parallel_for loop. Tom ----------------- Tom Peterka Mathematics and Computer Science Division Argonne National Laboratory 9700 S. Cass Ave. Lemont, IL 60439 Bldg 240, Rm 2148 e-mail: tpeterka(a)mcs.anl.gov<mailto:[email protected]> web: www.mcs.anl.gov/~tpeterka phone: 630-252-7198 On Nov 11, 2019, at 3:19 PM, Rosales-fernandez, Carlos via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]>> wrote: Hi all, Just a heads up that we will be pushing a new SDK around 9pm Central Time to JLSE. At that point if you have scripts using module names only ,without an explicit version, your default loads will shift to the new SDK version. Older versions will remain in the system as usual. A notice will go out at that time with a link to the release notes. Regards, Carlos -- Ecp-aurora-sdk-discuss mailing list Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]> https://lists.jlse.anl.gov/mailman/listinfo/ecp-aurora-sdk-discuss

1 0

Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today
by Rosales-fernandez, Carlos 15 Nov '19

15 Nov '19

Strange – I can also build and run these two in the iris nodes with a minor change (remove the “sycl::” in the usm allocator in example 1, which makes sense since I used the cl::sycl namespace). There must be an issue with your environment that is causing this – can we follow up offline and drop the list? We can update everyone later if we find a general issue. Carlos From: "Peterka, Thomas" <tpeterka(a)mcs.anl.gov> Date: Friday, November 15, 2019 at 8:34 AM To: "Rosales-fernandez, Carlos" <carlos.rosales-fernandez(a)intel.com> Cc: Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Subject: Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today Hi Carlos, I think there may be something wrong with this latest drop, because SYCL code that used to work now crashes, and reverting to the old version version of dpcpp/2019.9.30.013b continues to work. Specifically, I see two problems: 1. My test of USM generates the following run time error: symbol lookup error: ./fixed: undefined symbol: _ZN2cl4sycl6detail3usm12alignedAllocEmmRKNS0_7contextENS0_3usm5allocE The code is this: auto dev = Queue.get_device(); auto ctxt = Queue.get_context(); sycl::usm_allocator<int, sycl::usm::alloc::host> alloc(ctxt, dev); std::vector<int, decltype(alloc)> n(alloc); n.resize(1); The last line, resize, is where the error occurs. The same problem happens if I give the size in the constructor and don’t call the resize. 2. The following trivial example generates a segmentation fault: sycl::queue Queue; sycl::buffer<sycl::cl_int, 1> Buffer(4); sycl::range<1> NumOfWorkItems{Buffer.get_count()}; Queue.submit([&](sycl::handler &cgh) { auto Accessor = Buffer.get_access<sycl::access::mode::write>(cgh); cgh.parallel_for<class FillBuffer>(NumOfWorkItems, [=](sycl::id<1> WIid) { Accessor[WIid] = (cl::sycl::cl_int)WIid.get(0); }); }); The seg fault happens in the parallel_for loop. Tom ----------------- Tom Peterka Mathematics and Computer Science Division Argonne National Laboratory 9700 S. Cass Ave. Lemont, IL 60439 Bldg 240, Rm 2148 e-mail: tpeterka(a)mcs.anl.gov<mailto:[email protected]> web: www.mcs.anl.gov/~tpeterka phone: 630-252-7198 On Nov 11, 2019, at 3:19 PM, Rosales-fernandez, Carlos via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]>> wrote: Hi all, Just a heads up that we will be pushing a new SDK around 9pm Central Time to JLSE. At that point if you have scripts using module names only ,without an explicit version, your default loads will shift to the new SDK version. Older versions will remain in the system as usual. A notice will go out at that time with a link to the release notes. Regards, Carlos -- Ecp-aurora-sdk-discuss mailing list Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]> https://lists.jlse.anl.gov/mailman/listinfo/ecp-aurora-sdk-discuss

1 0

Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today
by Rosales-fernandez, Carlos 15 Nov '19

15 Nov '19

Thomas, These two examples work in our in-house systems with the current drop. Let me test in JLSE to see if I can find what is going on. Carlos From: "Peterka, Thomas" <tpeterka(a)mcs.anl.gov> Date: Friday, November 15, 2019 at 8:34 AM To: "Rosales-fernandez, Carlos" <carlos.rosales-fernandez(a)intel.com> Cc: Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Subject: Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today Hi Carlos, I think there may be something wrong with this latest drop, because SYCL code that used to work now crashes, and reverting to the old version version of dpcpp/2019.9.30.013b continues to work. Specifically, I see two problems: 1. My test of USM generates the following run time error: symbol lookup error: ./fixed: undefined symbol: _ZN2cl4sycl6detail3usm12alignedAllocEmmRKNS0_7contextENS0_3usm5allocE The code is this: auto dev = Queue.get_device(); auto ctxt = Queue.get_context(); sycl::usm_allocator<int, sycl::usm::alloc::host> alloc(ctxt, dev); std::vector<int, decltype(alloc)> n(alloc); n.resize(1); The last line, resize, is where the error occurs. The same problem happens if I give the size in the constructor and don’t call the resize. 2. The following trivial example generates a segmentation fault: sycl::queue Queue; sycl::buffer<sycl::cl_int, 1> Buffer(4); sycl::range<1> NumOfWorkItems{Buffer.get_count()}; Queue.submit([&](sycl::handler &cgh) { auto Accessor = Buffer.get_access<sycl::access::mode::write>(cgh); cgh.parallel_for<class FillBuffer>(NumOfWorkItems, [=](sycl::id<1> WIid) { Accessor[WIid] = (cl::sycl::cl_int)WIid.get(0); }); }); The seg fault happens in the parallel_for loop. Tom ----------------- Tom Peterka Mathematics and Computer Science Division Argonne National Laboratory 9700 S. Cass Ave. Lemont, IL 60439 Bldg 240, Rm 2148 e-mail: tpeterka(a)mcs.anl.gov<mailto:[email protected]> web: www.mcs.anl.gov/~tpeterka phone: 630-252-7198 On Nov 11, 2019, at 3:19 PM, Rosales-fernandez, Carlos via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]>> wrote: Hi all, Just a heads up that we will be pushing a new SDK around 9pm Central Time to JLSE. At that point if you have scripts using module names only ,without an explicit version, your default loads will shift to the new SDK version. Older versions will remain in the system as usual. A notice will go out at that time with a link to the release notes. Regards, Carlos -- Ecp-aurora-sdk-discuss mailing list Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]> https://lists.jlse.anl.gov/mailman/listinfo/ecp-aurora-sdk-discuss

1 0

Re: [Ecp-aurora-sdk-discuss] New A21 SDK coming to JLSE later today
by Peterka, Thomas 15 Nov '19

15 Nov '19

Hi Carlos, I think there may be something wrong with this latest drop, because SYCL code that used to work now crashes, and reverting to the old version version of dpcpp/2019.9.30.013b continues to work. Specifically, I see two problems: 1. My test of USM generates the following run time error: symbol lookup error: ./fixed: undefined symbol: _ZN2cl4sycl6detail3usm12alignedAllocEmmRKNS0_7contextENS0_3usm5allocE The code is this: auto dev = Queue.get_device(); auto ctxt = Queue.get_context(); sycl::usm_allocator<int, sycl::usm::alloc::host> alloc(ctxt, dev); std::vector<int, decltype(alloc)> n(alloc); n.resize(1); The last line, resize, is where the error occurs. The same problem happens if I give the size in the constructor and don’t call the resize. 2. The following trivial example generates a segmentation fault: sycl::queue Queue; sycl::buffer<sycl::cl_int, 1> Buffer(4); sycl::range<1> NumOfWorkItems{Buffer.get_count()}; Queue.submit([&](sycl::handler &cgh) { auto Accessor = Buffer.get_access<sycl::access::mode::write>(cgh); cgh.parallel_for<class FillBuffer>(NumOfWorkItems, [=](sycl::id<1> WIid) { Accessor[WIid] = (cl::sycl::cl_int)WIid.get(0); }); }); The seg fault happens in the parallel_for loop. Tom ----------------- Tom Peterka Mathematics and Computer Science Division Argonne National Laboratory 9700 S. Cass Ave. Lemont, IL 60439 Bldg 240, Rm 2148 e-mail: tpeterka(a)mcs.anl.gov<mailto:[email protected]> web: www.mcs.anl.gov/~tpeterka phone: 630-252-7198 On Nov 11, 2019, at 3:19 PM, Rosales-fernandez, Carlos via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]>> wrote: Hi all, Just a heads up that we will be pushing a new SDK around 9pm Central Time to JLSE. At that point if you have scripts using module names only ,without an explicit version, your default loads will shift to the new SDK version. Older versions will remain in the system as usual. A notice will go out at that time with a link to the release notes. Regards, Carlos -- Ecp-aurora-sdk-discuss mailing list Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]> https://lists.jlse.anl.gov/mailman/listinfo/ecp-aurora-sdk-discuss

1 0

New A21 SDK available in JLSE
by Rosales-fernandez, Carlos 11 Nov '19

11 Nov '19

Hi all, We are pleased to announce that a new A21 SDK has been deployed in JLSE. This version 2019.11.15.015 includes multiple improvements. For details please see the release notes in the wiki: https://wiki.jlse.anl.gov/display/inteldga/DGA+SDK+Release+Notes?preview=/2… Modules will enable the newer SDK by default, but the older versions remain available for regression testing. As usual, please let us know of any issues you encounter and any improvements we can include. Two important notes: 1. This release does not contain a DAAL update since the team is working on stabilizing updates aimed at the December release. 2. You will not be able to collect profiles with the new VTune version until the SEP driver has been rebuilt and installed by the administrators. Regards, Carlos

1 0

New A21 SDK coming to JLSE later today
by Rosales-fernandez, Carlos 11 Nov '19

11 Nov '19

Hi all, Just a heads up that we will be pushing a new SDK around 9pm Central Time to JLSE. At that point if you have scripts using module names only ,without an explicit version, your default loads will shift to the new SDK version. Older versions will remain in the system as usual. A notice will go out at that time with a link to the release notes. Regards, Carlos

1 0

Intel oneAPI Workshop at SC19
by Ghadar, Yasaman 11 Nov '19

11 Nov '19

Dear Aurora programming models attendees, We just got word that Intel is holding two sessions of a oneAPI workshop at the SC19 conference next week. If any of you that are attending SC19 are interested, please consider yourself invited. oneAPI is Intel’s broader branding of the compilers, programming models, runtimes, and libraries that include support for their GPUs. Our SDK that we use on the JLSE systems has early versions of some of the oneAPI software. Intel oneAPI Project [Intel] Intel oneAPI Workshop Learn more ><https://software.seek.intel.com/OneAPI_Workshop_REG> We would like to invite you to an opportunity to hear about Intel’s oneAPI project. oneAPI is a unified programming model to optimize performance for a wide range of data-centric workloads across scalar, vector, matrix and spatial architectures deployed in CPU, GPU, AI, FPGA and other accelerators It includes the programming language, Data Parallel C++ and libraries that, together, deliver parallel programming productivity, and cross-architecture application performance while providing an alternative to single architecture programming approaches. Two dates to choose from: November, 19 SC19 – Located at One Intel Station 1-4 PM MDT – 3 hours November, 20 SC19 – Located at One Intel Station 1-4 PM MDT – 3 hours During this workshop you will: * See oneAPI tools in action * Learn DPC++ programming * Experience Intel® DevCloud using CPUs, GPUs and FPGAs * Run code samples via immersive hands on training Please bring your laptop to the workshop. Registration and more information is available by visiting our registration website<https://software.seek.intel.com/OneAPI_Workshop_REG>. Thanks for saving the date and we look forward to seeing you soon! Location: Ellie Caulkins Opera House at the Denver Performing Arts Center In the Studio Loft – Level 4 1385 Curtis Street Denver, Colorado Register now<https://software.seek.intel.com/OneAPI_Workshop_REG> [Facebook]<https://www.facebook.com/Intel> [LinkedIn] <https://www.linkedin.com/company/intel-corporation> [Twitter] <https://twitter.com/intel> [YouTube] <https://www.youtube.com/user/channelintel?sub_confirmation=1&elqTrackId=10a…> [Instagram] <http://www.instagram.com/intel> [AddThis] <https://www.intel.com/content/www/us/en/blogs-communities-social.html> Best regards, Yasaman Ghadar

1 0