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

[Anl-intel-csa] Linker failures when building with OpenMP
by Tannenbaum, Barry M 31 May '17

31 May '17

I managed to reproduce the problem using the current llvm/clang/openmp trunk and a simple program without any CSA code at all. The Intel OpenMP group has acknowledged the bug and will work on it. In the meantime they've provided a workaround: compile all of your modules with the CSA clang compiler and the -fopenmp -fopenmp-targets options. For example: # This example breaks clang say-hello.cpp -c clang -fopenmp -fopenmp-targets=csa main.cpp say-hello.o -o test # This example works clang -fopenmp -fopenmp-targets=csa say-hello.cpp -c clang -fopenmp -fopenmp-targets=csa main.cpp say-hello.o -o test - Barry

1 0

Re: [Anl-intel-csa] complex data type?
by Kazutomo Yoshii 31 May '17

31 May '17

Something like std::complex would be nice to have! I have ported kernels that have lots of complex ops to OpenCL, which required some tedious work and can be error-prone. We may be able to leverage the nature of composted operations to optimize computation later. I'm curious to know some details (or plan) of the hw/sw interface. I'll start a new email thread about the hw/sw interface later. - kaz On 05/30/2017 03:06 PM, Tannenbaum, Barry M wrote: > Yes, I'm looking at the problems with using std::complex. In OpenMP offloaded code. It appears to be inlined properly when compiled ahead of time. > > -----Original Message----- > From: anl-intel-csa [mailto:[email protected]] On Behalf Of Glossop, Kent > Sent: Tuesday, May 30, 2017 2:26 PM > To: Kazutomo Yoshii <kazutomo(a)mcs.anl.gov>; anl-intel-csa(a)lists.jlse.anl.gov > Subject: Re: [Anl-intel-csa] complex data type? > > We would not expect any hardware implementation of complex, just like most traditional machines. Complex operations are simply compositions of the underlying operations. Having said that, I believe one of the team members is looking into issues with trying to use the complex in C/C++ that people ran into. > > At the level of the hw/sw interface, we expect it to be a series of values, just as is normally the case for register machines where the calling convention is a set of registers and stack locations. Normally, the software partitions the values as necessary. (e.g. if you pass a complex value or a struct of a couple of floats by value, the pieces are normally assigned to successive memory locations.) > > Particularly for the August deadline, I am not expecting a significant amount of functional change. What we need in general is specific items that are blocking you, or likely to be blocking you, that can't easily be worked around. e.g. if you have an application that is using complex, it would be useful to know which features/functions are needed. In terms of calling into CSA code, I would expect that if you are passing a complex number by value, you will likely need to pass the real and imaginary parts as separate pieces to CSA call. It may or may not be necessary to have the CSA routine receive them as separate pieces as well. > > Kent > > -----Original Message----- > From: anl-intel-csa [mailto:[email protected]] On Behalf Of Kazutomo Yoshii > Sent: Tuesday, May 30, 2017 2:17 PM > To: anl-intel-csa(a)lists.jlse.anl.gov > Subject: [Anl-intel-csa] complex data type? > > > Any plan to implement complex data type eventually? > > Also a (kind of) related question. Do you plan to change the CSA argument passing mechanism? "struct csa_arg" expects a fixed size. I guess a compound data structure like complex would be an issue. > > - kaz > > _______________________________________________ > anl-intel-csa mailing list > anl-intel-csa(a)lists.jlse.anl.gov > https://lists.jlse.anl.gov/mailman/listinfo/anl-intel-csa > _______________________________________________ > 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

Re: [Anl-intel-csa] complex data type?
by Tannenbaum, Barry M 30 May '17

30 May '17

Yes, I'm looking at the problems with using std::complex. In OpenMP offloaded code. It appears to be inlined properly when compiled ahead of time. -----Original Message----- From: anl-intel-csa [mailto:[email protected]] On Behalf Of Glossop, Kent Sent: Tuesday, May 30, 2017 2:26 PM To: Kazutomo Yoshii <kazutomo(a)mcs.anl.gov>; anl-intel-csa(a)lists.jlse.anl.gov Subject: Re: [Anl-intel-csa] complex data type? We would not expect any hardware implementation of complex, just like most traditional machines. Complex operations are simply compositions of the underlying operations. Having said that, I believe one of the team members is looking into issues with trying to use the complex in C/C++ that people ran into. At the level of the hw/sw interface, we expect it to be a series of values, just as is normally the case for register machines where the calling convention is a set of registers and stack locations. Normally, the software partitions the values as necessary. (e.g. if you pass a complex value or a struct of a couple of floats by value, the pieces are normally assigned to successive memory locations.) Particularly for the August deadline, I am not expecting a significant amount of functional change. What we need in general is specific items that are blocking you, or likely to be blocking you, that can't easily be worked around. e.g. if you have an application that is using complex, it would be useful to know which features/functions are needed. In terms of calling into CSA code, I would expect that if you are passing a complex number by value, you will likely need to pass the real and imaginary parts as separate pieces to CSA call. It may or may not be necessary to have the CSA routine receive them as separate pieces as well. Kent -----Original Message----- From: anl-intel-csa [mailto:[email protected]] On Behalf Of Kazutomo Yoshii Sent: Tuesday, May 30, 2017 2:17 PM To: anl-intel-csa(a)lists.jlse.anl.gov Subject: [Anl-intel-csa] complex data type? Any plan to implement complex data type eventually? Also a (kind of) related question. Do you plan to change the CSA argument passing mechanism? "struct csa_arg" expects a fixed size. I guess a compound data structure like complex would be an issue. - kaz _______________________________________________ anl-intel-csa mailing list anl-intel-csa(a)lists.jlse.anl.gov https://lists.jlse.anl.gov/mailman/listinfo/anl-intel-csa _______________________________________________ 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

Re: [Anl-intel-csa] complex data type?
by Glossop, Kent 30 May '17

30 May '17

We would not expect any hardware implementation of complex, just like most traditional machines. Complex operations are simply compositions of the underlying operations. Having said that, I believe one of the team members is looking into issues with trying to use the complex in C/C++ that people ran into. At the level of the hw/sw interface, we expect it to be a series of values, just as is normally the case for register machines where the calling convention is a set of registers and stack locations. Normally, the software partitions the values as necessary. (e.g. if you pass a complex value or a struct of a couple of floats by value, the pieces are normally assigned to successive memory locations.) Particularly for the August deadline, I am not expecting a significant amount of functional change. What we need in general is specific items that are blocking you, or likely to be blocking you, that can't easily be worked around. e.g. if you have an application that is using complex, it would be useful to know which features/functions are needed. In terms of calling into CSA code, I would expect that if you are passing a complex number by value, you will likely need to pass the real and imaginary parts as separate pieces to CSA call. It may or may not be necessary to have the CSA routine receive them as separate pieces as well. Kent -----Original Message----- From: anl-intel-csa [mailto:[email protected]] On Behalf Of Kazutomo Yoshii Sent: Tuesday, May 30, 2017 2:17 PM To: anl-intel-csa(a)lists.jlse.anl.gov Subject: [Anl-intel-csa] complex data type? Any plan to implement complex data type eventually? Also a (kind of) related question. Do you plan to change the CSA argument passing mechanism? "struct csa_arg" expects a fixed size. I guess a compound data structure like complex would be an issue. - kaz _______________________________________________ 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] complex data type?
by Kazutomo Yoshii 30 May '17

30 May '17

Any plan to implement complex data type eventually? Also a (kind of) related question. Do you plan to change the CSA argument passing mechanism? "struct csa_arg" expects a fixed size. I guess a compound data structure like complex would be an issue. - kaz

1 0

[Anl-intel-csa] CSA math library function estimates for size / latency
by Glossop, Kent 30 May '17

30 May '17

There were questions on CSA math function size / latency. The following estimate came from some of our math library team members. Our expectation at this point is we will likely start out sharing instances of math functions as a single instance, only replicating ("inlining") if the overall graph is small enough and that actually helps performance. (e.g. if you had a loop body with 2 calls to sin(), there would likely be one body by default, with picks/switches as appropriate for the call/return.) Operation Operation est Latency est sqrt64 14 42 sin64 30 15*3 cos64 30 15*3 log64 24 12*3 exp64 14 10*3 sqrt64 11 11*3 invsqrt64 11 11*3 Floating div 64 15 15*3 Floating inverse 13 13*3 floor64 3 (1 with ROUND instruction) 3*3 mod64 5 (2 with ROUND instruction) 5*3 ceil64 3 (1 with ROUND instruction) 3*3 round64 5 5*3 cvt single int 1 2 cvt double int 1 2 cvt single<=>double 1 2 Kent

1 0

Re: [Anl-intel-csa] Metagenomics apps of interest
by Kazutomo Yoshii 26 May '17

26 May '17

I guess other two are also easy to build. The first step is to find right data set. I'll find local genomics experts who are familiar with diamond and mash next week. - kaz On 05/26/2017 03:01 PM, Kumaran, Kalyan wrote: > Elliott, here’s the complete list: > >> >> 1. Diamond (https://github.com/bbuchfink/diamond) >> 2. BWA (https://github.com/lh3/bwa) >> 3. Mash (https://github.com/marbl/mash) >> > _______________________________________________ > 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] Metagenomics apps of interest
by Kumaran, Kalyan 26 May '17

26 May '17

Elliott, here’s the complete list: > > > 1. Diamond (https://github.com/bbuchfink/diamond) > 2. BWA (https://github.com/lh3/bwa) > 3. Mash (https://github.com/marbl/mash) >

1 0

[Anl-intel-csa] Fwd: QMCPACK
by Kumaran, Kalyan 26 May '17

26 May '17

Begin forwarded message: [cid:1DE547160A4FD34A8F0D65873758C7B9@localhost] Sent from my iPhone

1 0

[Anl-intel-csa] bwa
by Kazutomo Yoshii 26 May '17

26 May '17

$ git clone /soft/restricted/intel_csa/apps/bwa/bwa.git/ $ cd bwa $ git checkout -t origin/csa $ source /soft/restricted/intel_csa/csasetup.bash $ make $ ./bench_extend2 testdata/extend2-qlen26-tlen212-tsc28860-knl04.bindat # test bindat was collected from a BWA run. To see my modifications $ git diff 5961611c3 !(testdata) - kaz

1 0