cyme
Framework to facilitate SIMD programming, without any tedious SIMD intrinsics.
Installation (need boost, doxygen, gcc >= 4.4, the doc provides tutorials, etc ... ). There are a few variables to adjust Makefile generation with cmake (can be combined), see next examples
mkdir b
cd b
cmake .. # basic
cmake .. # your machine supports slurm, specify the environment variable SLURM_ACCOUNT (export SLURM_ACCOUNT="--account=proj16") before !
cmake .. -DSLURM_FOUND=false # your machine supports slurm but you do not care
cmake .. -DCOMMON_LIBRARY_TYPE=STATIC # you prefer static library (boost lib, e.g. BG/Q)
cmake .. -DCYME_POWER_VMX=true # force VMX backend for on BG/Q frontend ambiguity
make doxygen
Compilation:
-D__CYME_SIMD_VALUE__=sse -msse4.1 + optional -D__FMA__
-D__CYME_SIMD_VALUE__=avx -msvx+ optional -D__FMA__
-D__CYME_SIMD_VALUE__=qpx -D__FMA__
-D__CYME_SIMD_VALUE__=vmx -D__FMA__ -mcpu=power7 -maltivec -mvsx
Know issues:
- BG/Q: bug Newton-Raphson algorithm (division/sqrt).
DEBUG mode:
-
the core engine test: core_operator_bracket_torture - Newton-Raphson algo for the division fails:
it_AoS = block_a.begin(); for(; it_AoS != block_a.end(); ++it_AoS) (*it_AoS)[0] = (((((*it_AoS)[1] /(*it_AoS)[2])/(*it_AoS)[3])/(*it_AoS)[4])/(*it_AoS)[5]);
it_AoSoA_w = block_b.begin(); it_AoSoA_r = block_b.begin(); for(; it_AoSoA_r != block_b.end(); ++it_AoSoA_w,++it_AoSoA_r) (*it_AoSoA_w)[0] = (((((*it_AoSoA_r)[1]/(*it_AoSoA_r)[2])/(*it_AoSoA_r)[3])/(*it_AoSoA_r)[4])/(*it_AoSoA_r)[5]);
It fails for an unroll of 2 and double, only if the number of successive division is larger than 3. I check the corresponding wrapper __mm__rec() and _mm_nfma(), I do not find any issues. It works for float. Test works if I perform the normal division
- the validity test also fail for an unroll of 2
RELEASE:
The serial test unroll 4 fails, again for the division. It is very wierd as the operator /= call the system division and not Newton-Raphson
Release 1.7.0
- Optimize the backend AVX to AVX2 if the processor supports it
- Tune log(x) and exp(x) solver using AVX2
- Complete and Optimize all inequalities <, >, <=, >= for the DAG and SSE/AVX backend, with corresponding tests
- Add support for bitwise operator for 32 bits integer SSE/AVX only, with corresponding tests
- Optimize operator ==
- Clang format and typos
Release 1.6.0
- More support of inequalities for SSE/AVX backend
- Support masking for inequalities for SSE/aVX backend
- Better compatibility with Intel compiler
- Implementation of power(x,y) (node in the DAG + backend), all backends
- Minor fixes
- Swap boost by c++11
Release 1.5.0
- Merge rvec and wvec into vec
- fix bugs in the exponential all backends
- boundary limite for the exp with the new corresponding wrappers for integer manipulation
- Prepare work for the next release !
Release 1.4.0
- Google Summer code merge: work of Kai Langen
- 2^x, 10^x, log_2 and log_10 solver
- backend power7/power8 (with litle endian support)
- backend ARM
- trigonometric function backend SSE/AVX/NEON and VMX (no mic and not BG/Q)
- fabs
- printing functions
Release 1.3.0
- New license lgpl 3.0
- Open Source Standard
- Add new elementary function (1/sqrt)
- Tiny optimization in the tree
Release 1.2.0
- Add new elementary function (sqrt)
- Fix bug in log
- Better DAG construction for operations with scalar operations
- Better catch of the FMA
- Simplification of the template expression when I mix float/double, implement unity concept
- Prepare for open source release
- Ameliorate CMake during the installation, configuration file is created with the needed define
Release 1.1.0
- Add new elementary function (pow/log)
- Ameliorate the tree creation by distinction Read/Write operators - const/ non const
- better DAG construction for --
Release 1.0.0
- First release of cyme
Funding
The development of this software was supported by funding to the Blue Brain Project, a research center of the École polytechnique fédérale de Lausanne (EPFL), from the Swiss government’s ETH Board of the Swiss Federal Institutes of Technology.
Copyright (c) 2013-2021 Blue Brain Project/EPFL