I had weird issues where gcc (that I am using for my tests right now) wasn't vectorising properly (even enabling all flags, from tree-vectorize, to mavx). According to my tests, I know the Intel compiler was a bit better at that.
I actually did not know PETSc was doing some unrolling himself. On my machine, PETSc aligns the memory to 16 bytes, that might also be a cause. However, I have no idea on the ability of the different compilers to detect and vectorize codes, even when those are manually unrolled by the user. I see PETSc unrolls the loop for MAXPY up 4 right hand side vectors. Just by coincidence, the numbers I reported are actually for this latter case. I pushed my code to a public git repository. However let me stress that the code is extremely ugly and buggy. I actually feel shame in showing you the code :) : https://bitbucket.org/FilippoL/petscxx/commits/branch/master If you want to look at the API usage, you can find it in the file "/src/benchmarks/bench_vector.cpp" after line 153. If you want the expressions generating the kernels you have to look at the file "/src/base/vectorexpression.hpp", specifically the "evalat" members. On Tue, 4 Apr 2017 at 21:39 Matthew Knepley <knep...@gmail.com> wrote: On Tue, Apr 4, 2017 at 1:19 PM, Filippo Leonardi <filippo.l...@gmail.com> wrote: You are in fact right, it is the same speedup of approximatively 2.5x (with 2 ranks), my brain rounded up to 3. (This was just a test done in 10 min on my Workstation, so no pretence to be definite, I just wanted to have an indication). Hmm, it seems like PetscKernelAXPY4() is just not vectorizing correctly then. I would be interested to see your code. As you say, I am using OpenBLAS, so I wouldn't be surprised of those results. If/when I use MKL (or something similar), I really do not expect such an improvement). Since you seem interested (if you are interested, I can give you all the details): the comparison I make, is with "petscxx" which is my template code (which uses a single loop) using AVX (I force PETSc to align the memory to 32 bit boundary and then I use packets of 4 doubles). Also notice that I use vectors with nice lengths, so there is no need to "peel" the end of the loop. The "PETSc" simulation is using PETSc's VecMAXPY. Thanks, Matt On Tue, 4 Apr 2017 at 19:12 Barry Smith <bsm...@mcs.anl.gov> wrote: MAXPY isn't really a BLAS 1 since it can reuse some data in certain vectors. > On Apr 4, 2017, at 10:25 AM, Filippo Leonardi <filippo.l...@gmail.com> wrote: > > I really appreciate the feedback. Thanks. > > That of deadlock, when the order of destruction is not preserved, is a point I hadn't thought of. Maybe it can be cleverly addressed. > > PS: If you are interested, I ran some benchmark on BLAS1 stuff and, for a single processor, I obtain: > > Example for MAXPY, with expression templates: > BM_Vector_petscxx_MAXPY/8 38 ns 38 ns 18369805 > BM_Vector_petscxx_MAXPY/64 622 ns 622 ns 1364335 > BM_Vector_petscxx_MAXPY/512 281 ns 281 ns 2477718 > BM_Vector_petscxx_MAXPY/4096 2046 ns 2046 ns 349954 > BM_Vector_petscxx_MAXPY/32768 18012 ns 18012 ns 38788 > BM_Vector_petscxx_MAXPY_BigO 0.55 N 0.55 N > BM_Vector_petscxx_MAXPY_RMS 7 % 7 % > Direct call to MAXPY: > BM_Vector_PETSc_MAXPY/8 33 ns 33 ns 20973674 > BM_Vector_PETSc_MAXPY/64 116 ns 116 ns 5992878 > BM_Vector_PETSc_MAXPY/512 731 ns 731 ns 963340 > BM_Vector_PETSc_MAXPY/4096 5739 ns 5739 ns 122414 > BM_Vector_PETSc_MAXPY/32768 46346 ns 46346 ns 15312 > BM_Vector_PETSc_MAXPY_BigO 1.41 N 1.41 N > BM_Vector_PETSc_MAXPY_RMS 0 % 0 % > > And 3x speedup on 2 MPI ranks (not much communication here, anyway). I am now convinced that this warrants some further investigation/testing. > > > On Tue, 4 Apr 2017 at 01:08 Jed Brown <j...@jedbrown.org> wrote: > Matthew Knepley <knep...@gmail.com> writes: > > >> BLAS. (Here a interesting point opens: I assume an efficient BLAS > >> > >> implementation, but I am not so sure about how the different BLAS do > >> things > >> > >> internally. I work from the assumption that we have a very well tuned BLAS > >> > >> implementation at our disposal). > >> > > > > The speed improvement comes from pulling vectors through memory fewer > > times by merging operations (kernel fusion). > > Typical examples are VecMAXPY and VecMDot, but note that these are not > xGEMV because the vectors are independent arrays rather than single > arrays with a constant leading dimension. > > >> call VecGetArray. However I will inevitably foget to return the array to > >> > >> PETSc. I could have my new VecArray returning an object that restores the > >> > >> array > >> > >> when it goes out of scope. I can also flag the function with [[nodiscard]] > >> to > >> > >> prevent the user to destroy the returned object from the start. > >> > > > > Jed claims that this pattern is no longer preferred, but I have forgotten > > his argument. > > Jed? > > Destruction order matters and needs to be collective. If an error > condition causes destruction to occur in a different order on different > processes, you can get deadlock. I would much rather have an error > leave some resources (for the OS to collect) than escalate into > deadlock. > > > We have had this discussion for years on this list. Having separate names > > for each type > > is really ugly and does not achieve what we want. We want smooth > > interoperability between > > objects with different backing types, but it is still not clear how to do > > this. > > Hide it internally and implicitly promote. Only the *GetArray functions > need to be parametrized on numeric type. But it's a lot of work on the > backend. -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener