February 2021 - petsc-dev - mailman3.cels.anl.gov

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Jed Brown 14 Feb '21

14 Feb '21

Barry Smith <bsmith(a)petsc.dev> writes: >> On Feb 14, 2021, at 12:13 PM, Jed Brown <jed(a)jedbrown.org> wrote: >> >> Barry Smith <bsmith(a)petsc.dev> writes: >> >>>> This is a reasonable message to print on the screen, but I don’t think this is a reasonable flag to impose by default. >>>> You are basically asking all package managers to add a new flag (-march=generic) which was previously not needed. >>> >>> This is a tough constraint, package managers should not have to do anything to get portability but users get great performance without needing to be sophisticated. Seems to put the burden on the unsophisticated folks (all users) and not on the sophisticated folks (packages). >> >> The sophisticated folks are us, the upstream developers who know what needs to be optimized for specific hardware (src/mat/impls/) and what does not (most of the rest of PETSc). We've been typical self-centered scientific software developers who just assume people who care about performance will build a custom library for each machine they run on. > > I'm still bit confused. We need to build fat binaries otherwise the stuff in src/mat/impls/ will either be not portable or will be generic and potentially slower, right? Are you saying it is not worth building fat binaries for most of the source? So we mark either by directory or file where we want fat binaries built? Yes, binaries only need multiple versions for specific parts (kernels that benefit from vectorization). That keeps them smaller and the build fast. We can do a better job than "fat" compiler flags by working out dispatch in MatCreate_XXX() instead of for each function.

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Jed Brown 14 Feb '21

14 Feb '21

Barry Smith <bsmith(a)petsc.dev> writes: > My feeling is 90+% of users don't care about portability, they want to get fast performance on the machine they are compiling with (or a collection of machines they have around). This is an exclusionary view of users. PETSc could and would be used in more places if it was easy to install, e.g., if "pip install petsc" was a 3-second binary install instead of several minutes that requires compilers.

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Barry Smith 14 Feb '21

14 Feb '21

> On Feb 14, 2021, at 1:25 PM, Zhang, Hong <hongzhang(a)anl.gov> wrote: > > > >> On Feb 14, 2021, at 12:04 PM, Barry Smith <bsmith(a)petsc.dev> wrote: >> >> >> For our handcoded AVX functions this is fine, we can handle the dispatching ourselves. > > Cool. _may_i_use_cpu_feature() would be very useful to determine the optimal AVX code path at runtime. Theoretically we just need to query for the needed features once and cache the results. > >> >> But what about all the tons of regular code in PETSc, somehow we need to have the same function compiled twice and dispatched properly. Do we use what Hong suggested with fat binaries? So fat-binaries PLUS _may_i_use_cpu_feature together are the way to portable transportable libraries? >> >> >> And we do this always --with-debugging=0 so everyone, packages and users get portable but also the best performance possible. > > IMHO, only package managers should consider using -ax options. On our side, if we want to satisfy the needs of different parties (developers, users, package managers), better be conservative than aggressive. -march=native brings huge performance improvement But this means most our users are year after year throwing lots of performance on the floor and don't even know it. I think we pander for portability too much. > but it has never been the default for many compilers with a good reason. Even -O3 does not enable the advanced vector instructions. I just did a quick check on petsc-02: > > hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep SSE > #define __SSE__ 1 > #define __SSE_MATH__ 1 > #define __SSE2__ 1 > #define __SSE2_MATH__ 1 > hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep avx > hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ > > What Jed usually does (--with-debugging=0 COPTFLAGS='-O2 -march=native’) can be suggested to anyone who does not need to care about portability. If you do not want users to specify the magic options, perhaps we can provide a configure option like --with-portability. If it is set to false, we add aggressive flags automatically. My feeling is 90+% of users don't care about portability, they want to get fast performance on the machine they are compiling with (or a collection of machines they have around). Can we build aggressively for their system (except package managers and for people who provide the -march) and have PetscInitialize() produce a very useful error message if they then run the code on a system where it will not work? Any system calls to get that type of information? Barry > > Hong > >> >> Barry >> >> >>> On Feb 14, 2021, at 11:50 AM, Jed Brown <jed(a)jedbrown.org> wrote: >>> >>>> >>> >>> immintrin.h provides >>> >>> if (_may_i_use_cpu_feature(_FEATURE_FMA | _FEATURE_AVX2) { >>> fancy_version_that_needs_fma_and_avx2(); >>> } else { >>> fallback_version(); >>> } >>> >>> https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_may_i_u… >>> >>> I believe this function is slightly expensive because it probably calls the CPUID instruction each time. BLIS has code to cache the result and query features with simple bitwise math. >>> >>> https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.h >>> https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.c >>> >>> Of course this bit of dispatch should typically be done at object creation time, not every iteration. >> >

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Zhang, Hong 14 Feb '21

14 Feb '21

Oops, a typo in the command line. Should be AVX. SSE3 or above and AVX are not used for -O3. hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep SSE #define __SSE__ 1 #define __SSE_MATH__ 1 #define __SSE2__ 1 #define __SSE2_MATH__ 1 hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep AVX hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ > On Feb 14, 2021, at 1:25 PM, Zhang, Hong via petsc-dev <petsc-dev(a)mcs.anl.gov> wrote: > > > >> On Feb 14, 2021, at 12:04 PM, Barry Smith <bsmith(a)petsc.dev> wrote: >> >> >> For our handcoded AVX functions this is fine, we can handle the dispatching ourselves. > > Cool. _may_i_use_cpu_feature() would be very useful to determine the optimal AVX code path at runtime. Theoretically we just need to query for the needed features once and cache the results. > >> >> But what about all the tons of regular code in PETSc, somehow we need to have the same function compiled twice and dispatched properly. Do we use what Hong suggested with fat binaries? So fat-binaries PLUS _may_i_use_cpu_feature together are the way to portable transportable libraries? >> >> >> And we do this always --with-debugging=0 so everyone, packages and users get portable but also the best performance possible. > > IMHO, only package managers should consider using -ax options. On our side, if we want to satisfy the needs of different parties (developers, users, package managers), better be conservative than aggressive. -march=native brings huge performance improvement but it has never been the default for many compilers with a good reason. Even -O3 does not enable the advanced vector instructions. I just did a quick check on petsc-02: > > hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep SSE > #define __SSE__ 1 > #define __SSE_MATH__ 1 > #define __SSE2__ 1 > #define __SSE2_MATH__ 1 > hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep avx > hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ > > What Jed usually does (--with-debugging=0 COPTFLAGS='-O2 -march=native’) can be suggested to anyone who does not need to care about portability. If you do not want users to specify the magic options, perhaps we can provide a configure option like --with-portability. If it is set to false, we add aggressive flags automatically. > > Hong > >> >> Barry >> >> >>> On Feb 14, 2021, at 11:50 AM, Jed Brown <jed(a)jedbrown.org> wrote: >>> >>>> >>> >>> immintrin.h provides >>> >>> if (_may_i_use_cpu_feature(_FEATURE_FMA | _FEATURE_AVX2) { >>> fancy_version_that_needs_fma_and_avx2(); >>> } else { >>> fallback_version(); >>> } >>> >>> https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_may_i_u… >>> >>> I believe this function is slightly expensive because it probably calls the CPUID instruction each time. BLIS has code to cache the result and query features with simple bitwise math. >>> >>> https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.h >>> https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.c >>> >>> Of course this bit of dispatch should typically be done at object creation time, not every iteration. >> >

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Zhang, Hong 14 Feb '21

14 Feb '21

> On Feb 14, 2021, at 12:04 PM, Barry Smith <bsmith(a)petsc.dev> wrote: > > > For our handcoded AVX functions this is fine, we can handle the dispatching ourselves. Cool. _may_i_use_cpu_feature() would be very useful to determine the optimal AVX code path at runtime. Theoretically we just need to query for the needed features once and cache the results. > > But what about all the tons of regular code in PETSc, somehow we need to have the same function compiled twice and dispatched properly. Do we use what Hong suggested with fat binaries? So fat-binaries PLUS _may_i_use_cpu_feature together are the way to portable transportable libraries? > > > And we do this always --with-debugging=0 so everyone, packages and users get portable but also the best performance possible. IMHO, only package managers should consider using -ax options. On our side, if we want to satisfy the needs of different parties (developers, users, package managers), better be conservative than aggressive. -march=native brings huge performance improvement but it has never been the default for many compilers with a good reason. Even -O3 does not enable the advanced vector instructions. I just did a quick check on petsc-02: hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep SSE #define __SSE__ 1 #define __SSE_MATH__ 1 #define __SSE2__ 1 #define __SSE2_MATH__ 1 hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ icc -O3 -E -dM - < /dev/null | grep avx hongzhang@petsc-02:/nfs/gce/projects/TSAdjoint$ What Jed usually does (--with-debugging=0 COPTFLAGS='-O2 -march=native’) can be suggested to anyone who does not need to care about portability. If you do not want users to specify the magic options, perhaps we can provide a configure option like --with-portability. If it is set to false, we add aggressive flags automatically. Hong > > Barry > > >> On Feb 14, 2021, at 11:50 AM, Jed Brown <jed(a)jedbrown.org> wrote: >> >>> >> >> immintrin.h provides >> >> if (_may_i_use_cpu_feature(_FEATURE_FMA | _FEATURE_AVX2) { >> fancy_version_that_needs_fma_and_avx2(); >> } else { >> fallback_version(); >> } >> >> https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_may_i_u… >> >> I believe this function is slightly expensive because it probably calls the CPUID instruction each time. BLIS has code to cache the result and query features with simple bitwise math. >> >> https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.h >> https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.c >> >> Of course this bit of dispatch should typically be done at object creation time, not every iteration. >

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Barry Smith 14 Feb '21

14 Feb '21

> On Feb 14, 2021, at 12:13 PM, Jed Brown <jed(a)jedbrown.org> wrote: > > Barry Smith <bsmith(a)petsc.dev> writes: > >>> This is a reasonable message to print on the screen, but I don’t think this is a reasonable flag to impose by default. >>> You are basically asking all package managers to add a new flag (-march=generic) which was previously not needed. >> >> This is a tough constraint, package managers should not have to do anything to get portability but users get great performance without needing to be sophisticated. Seems to put the burden on the unsophisticated folks (all users) and not on the sophisticated folks (packages). > > The sophisticated folks are us, the upstream developers who know what needs to be optimized for specific hardware (src/mat/impls/) and what does not (most of the rest of PETSc). We've been typical self-centered scientific software developers who just assume people who care about performance will build a custom library for each machine they run on. I'm still bit confused. We need to build fat binaries otherwise the stuff in src/mat/impls/ will either be not portable or will be generic and potentially slower, right? Are you saying it is not worth building fat binaries for most of the source? So we mark either by directory or file where we want fat binaries built?

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Jed Brown 14 Feb '21

14 Feb '21

Barry Smith <bsmith(a)petsc.dev> writes: >> This is a reasonable message to print on the screen, but I don’t think this is a reasonable flag to impose by default. >> You are basically asking all package managers to add a new flag (-march=generic) which was previously not needed. > > This is a tough constraint, package managers should not have to do anything to get portability but users get great performance without needing to be sophisticated. Seems to put the burden on the unsophisticated folks (all users) and not on the sophisticated folks (packages). The sophisticated folks are us, the upstream developers who know what needs to be optimized for specific hardware (src/mat/impls/) and what does not (most of the rest of PETSc). We've been typical self-centered scientific software developers who just assume people who care about performance will build a custom library for each machine they run on.

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Barry Smith 14 Feb '21

14 Feb '21

For our handcoded AVX functions this is fine, we can handle the dispatching ourselves. But what about all the tons of regular code in PETSc, somehow we need to have the same function compiled twice and dispatched properly. Do we use what Hong suggested with fat binaries? So fat-binaries PLUS _may_i_use_cpu_feature together are the way to portable transportable libraries? And we do this always --with-debugging=0 so everyone, packages and users get portable but also the best performance possible. Barry > On Feb 14, 2021, at 11:50 AM, Jed Brown <jed(a)jedbrown.org> wrote: > >> > > immintrin.h provides > > if (_may_i_use_cpu_feature(_FEATURE_FMA | _FEATURE_AVX2) { > fancy_version_that_needs_fma_and_avx2(); > } else { > fallback_version(); > } > > https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_may_i_u… > > I believe this function is slightly expensive because it probably calls the CPUID instruction each time. BLIS has code to cache the result and query features with simple bitwise math. > > https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.h > https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.c > > Of course this bit of dispatch should typically be done at object creation time, not every iteration.

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Barry Smith 14 Feb '21

14 Feb '21

> On Feb 14, 2021, at 10:09 AM, Pierre Jolivet <pierre(a)joliv.et> wrote: > > > >> On 14 Feb 2021, at 4:52 PM, Zhang, Hong via petsc-dev <petsc-dev(a)mcs.anl.gov <mailto:[email protected]>> wrote: >> >> >> > > This is a reasonable message to print on the screen, but I don’t think this is a reasonable flag to impose by default. > You are basically asking all package managers to add a new flag (-march=generic) which was previously not needed. This is a tough constraint, package managers should not have to do anything to get portability but users get great performance without needing to be sophisticated. Seems to put the burden on the unsophisticated folks (all users) and not on the sophisticated folks (packages). What about a configure option --with-package-system which packagers would add that insures portability and packager rules and it not set configure is free to make better optimization choices? Or is there some environmental variable etc that configure can check that we know is always around when package building? Barry > > I’m crossing my fingers Jed has a clever way of "making portable binaries that run-time detected when to use newer instructions where it matters”, because -march=native by default is just not practical when deploying software. > > Thanks, > Pierre > >> We should also inform users that tuning -march options may enable vectorization instructions such as SSE(3 and above) and AVX but generate nonportable binaries. >> >> If we add -march=native to the configure test, we will need to run executables to make sure the specified instruction sets are supported by the CPU where the code is running. For PETSc, the executables should cover all the intrinsics we use in the code ideally; otherwise, users will get run-time errors when there is a mismatch in vectorization instructions between compiler support and CPU support. >> >> Hong >> >>> >>> None of the examples in config/examples actually use -march=native, and this is a very common thing to do that, as you point out, isn't obvious until you know you have to do it, so it seems to be worth the screen space. >>> >>> >>> >>> >>> >>>> >>>> Barry >>>> >>>> >>>>> >>>>> Barry Smith <bsmith(a)petsc.dev <mailto:[email protected]>> writes: >>>>> >>>>>> Shouldn't configure be setting something appropriate for this automatically? This is nuts, it means when users do a ./configure make unless they pass weird arguments they sure as heck don't know about to the compiler they won't get any of the glory that they expect and that has been in almost all Intel systems forever. >>>>>> >>>>>> Barry >>>>>> >>>>>> I run ./configure --with-debugging=0 and I get none of the stuff added by Intel for 15+ years? >>>>>> >>>>>> >>>>>>> On Feb 13, 2021, at 11:26 PM, Jed Brown <jed(a)jedbrown.org <mailto:[email protected]>> wrote: >>>>>>> >>>>>>> Use -march=native or similar. The default target is basic x86_64, which has only SSE2. >>>>>>> >>>>>>> Barry Smith <bsmith(a)petsc.dev <mailto:[email protected]>> writes: >>>>>>> >>>>>>>> PETSc source has code like defined(__AVX2__) in the source but it does not seem to be able to find any of these macros (icc or gcc) on the petsc-02 system >>>>>>>> >>>>>>>> Are these macros supposed to be defined? How does on get them to be defined? Why are they not define? What am I doing wrong? >>>>>>>> >>>>>>>> Keep reading >>>>>>>> >>>>>>>> $ lscpu >>>>>>>> Architecture: x86_64 >>>>>>>> CPU op-mode(s): 32-bit, 64-bit >>>>>>>> Byte Order: Little Endian >>>>>>>> CPU(s): 64 >>>>>>>> On-line CPU(s) list: 0-63 >>>>>>>> Thread(s) per core: 2 >>>>>>>> Core(s) per socket: 16 >>>>>>>> Socket(s): 2 >>>>>>>> NUMA node(s): 2 >>>>>>>> Vendor ID: GenuineIntel >>>>>>>> CPU family: 6 >>>>>>>> Model: 85 >>>>>>>> Model name: Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz >>>>>>>> Stepping: 7 >>>>>>>> CPU MHz: 1000.603 >>>>>>>> CPU max MHz: 2301.0000 >>>>>>>> CPU min MHz: 1000.0000 >>>>>>>> BogoMIPS: 4600.00 >>>>>>>> Virtualization: VT-x >>>>>>>> L1d cache: 32K >>>>>>>> L1i cache: 32K >>>>>>>> L2 cache: 1024K >>>>>>>> L3 cache: 22528K >>>>>>>> NUMA node0 CPU(s): 0-15,32-47 >>>>>>>> NUMA node1 CPU(s): 16-31,48-63 >>>>>>>> Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts pku ospke avx512_vnni md_clear flush_l1d arch_capabilities >>>>>>>> >>>>>>>> Test program >>>>>>>> >>>>>>>> #if defined(__FMA__) >>>>>>>> #error FMA >>>>>>>> #endif >>>>>>>> >>>>>>>> #if defined(__AVX512F__) >>>>>>>> #error AVX512F >>>>>>>> #endif >>>>>>>> >>>>>>>> #if defined(__AVX2__) >>>>>>>> #error AVX2 >>>>>>>> #endif >>>>>>>> >>>>>>>> >>>>>>>> icc mytest.c >>>>>>>> /usr/lib/gcc/x86_64-linux-gnu/7/../../../x86_64-linux-gnu/Scrt1.o: In function `_start': >>>>>>>> (.text+0x20): undefined reference to `main' >

1 0

Re: [petsc-dev] error with flags PETSc uses for determining AVX
by Jed Brown 14 Feb '21

14 Feb '21

Pierre Jolivet <pierre(a)joliv.et> writes: >>>> Expecting PETSc users to automatically add -march= is not realistic. I will try to rig something up in configure where if the user does not provide march something reasonable is selected. >>> A softer (yet trivial to implement) option might also be to just alert the user that these flags exist in the usual message about using default optimization flags. Something like this would encourage users to do what Jed is doing: >>> >>> ***** WARNING: Using default optimization C flags -g -O3 >>> You might consider manually setting optimal optimization flags for your system with >>> COPTFLAGS="optimization flags" see config/examples/arch-*-opt.py for examples. >>> In particular, you may want to supply specific flags (e.g. -march=native) >>> to take advantage of higher-performance instructions. >> >> I think this is a reasonable thing to do. > > This is a reasonable message to print on the screen, but I don’t think this is a reasonable flag to impose by default. > You are basically asking all package managers to add a new flag (-march=generic) which was previously not needed. > > I’m crossing my fingers Jed has a clever way of "making portable binaries that run-time detected when to use newer instructions where it matters”, because -march=native by default is just not practical when deploying software. immintrin.h provides if (_may_i_use_cpu_feature(_FEATURE_FMA | _FEATURE_AVX2) { fancy_version_that_needs_fma_and_avx2(); } else { fallback_version(); } https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_may_i_u… I believe this function is slightly expensive because it probably calls the CPUID instruction each time. BLIS has code to cache the result and query features with simple bitwise math. https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.h https://github.com/flame/blis/blob/master/frame/base/bli_cpuid.c Of course this bit of dispatch should typically be done at object creation time, not every iteration.

1 0