jfalcou / eve

Expressive Vector Engine - SIMD in C++ Goes Brrrr

⚡ Short Intro ⚡

I am Joel Falcou, Destroyer of World, Terror of the Compilers.

In my spare time, I am an associated professor at the University Paris-Saclay and researcher at the Laboratoire de Recherche d’Informatique in Orsay, France. My research focuses on studying generative programming idioms and techniques to design tools for parallel software development.

I also have a rather personal take on humor as you may have noticed already.

❓ Research Activities ❓

The main parts of my research topic are:

• the exploration of Embedded Domain Specific Language design for parallel computing on various architectures;
• the definition of a formal framework for reasoning about meta-programs.

As I need something to pad my academic paper up to at least eight pages, I usually play around with various application fields like real-time image processing on embedded architectures or HPC on multi-core clusters.

👯C++ Community 👯

I am the co-host of the C++FRUG Meetup, president of the C++FRUG Association and I co-organize the CPPP Conference.

You can find me on Mastodon or on the #include Discord

eve's Issues

bitwise tests may lead to denormals or other unsupported values

bitwise_* must use values that are correct even in real and don't generate values that some architecture may not support when used.

Make a skeleton .rst/.md for the doc

• Basic pages
• Basic intermission pages
• Basic function pages

Issues when starting

Tried to do the setup:

Docs ==========================================
The big one at the moment is that docs are not readable as a generated web-page.
Read some top ones as markdown, looked through references

Potential typos

2 times scalar_value:

eve/docs/glossary.html

Line 29 in 412d84a

family if it satisfies the [scalar_value](reference/concepts.html#simd_value).

17 and forward?

eve/docs/index.html

Line 13 in 412d84a

**EVE** is a C++20 and forward implementation of a type based wrapper around for SIMD computations. It wraps SIMD instruction sets for most current architectures. It aims at showing how C++20 can be used to design and implement efficient, low level, high abstraction library suited for high performances.

Do you want to maybe add how to run tests?

Building library/verifying tests

cmake -G Ninja -DEVE_BUILD_TEST ..
Parse error in command line argument: -DEVE_BUILD_TEST
Should be: VAR:type=value
CMake Error: No cmake script provided.
CMake Error: Problem processing arguments. Aborting.

dyarosh@dyarosh-u:~/space/eve/build$ cmake -G Ninja  ..
-- The CXX compiler identification is Clang 11.0.0
-- Check for working CXX compiler: /usr/local/bin/clang++
-- Check for working CXX compiler: /usr/local/bin/clang++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- [eve] Building  mode with: 
-- Found OpenMP_CXX: -fopenmp=libomp (found version "4.5") 
-- Found OpenMP: TRUE (found version "4.5")  
-- Configuring done
-- Generating done
-- Build files have been written to: /home/dyarosh/space/eve/build/tts-download
-- Found OpenMP_CXX: -fopenmp=libomp (found version "4.5") 
-- [tts] TTS is used as a in-project dependency - No install targets available
-- [eve] Boost found in /usr/include - Boost dependent tests activated
-- Configuring done
-- Generating done
-- Build files have been written to: /home/dyarosh/space/eve/build
dyarosh@dyarosh-u:~/space/eve/build$ ninja -j 8 tests
[12/16123] Building CXX object tts-build/src/CMakeFiles/tts.dir/info.cpp.o
FAILED: tts-build/src/CMakeFiles/tts.dir/info.cpp.o 
/usr/local/bin/clang++  -DTTS_API_EXPORTS=1 -Dtts_EXPORTS -Itts-src/include -Itts-src/src -fPIC   -Wall -fvisibility=hidden -std=gnu++17 -MD -MT tts-build/src/CMakeFiles/tts.dir/info.cpp.o -MF tts-build/src/CMakeFiles/tts.dir/info.cpp.o.d -o tts-build/src/CMakeFiles/tts.dir/info.cpp.o -c tts-src/src/info.cpp
tts-src/src/info.cpp:16:10: fatal error: 'filesystem' file not found
#include <filesystem>
         ^~~~~~~~~~~~
1 error generated.
[19/16123] Building CXX object tts-build/src/CMakeFiles/tts.dir/args.cpp.o
ninja: build stopped: subcommand failed.

filesystem was not found. I have file system in libc++ .

Unfortunately, the best I could find in terms of building with libc++ was:
cmake -G Ninja -DCMAKE_CXX_FLAGS=-stdlib=libc++ ..
Which breaks spy

dyarosh@dyarosh-u:~/space/eve/build$ cmake -G Ninja -DCMAKE_CXX_FLAGS=-stdlib=libc++ ..
-- [eve] Building  mode with: -stdlib=libc++
-- Found OpenMP_CXX: -fopenmp=libomp  
-- Configuring done
-- Generating done
-- Build files have been written to: /home/dyarosh/space/eve/build/tts-download
-- Found OpenMP_CXX: -fopenmp=libomp  
-- [tts] TTS is used as a in-project dependency - No install targets available
-- [eve] Boost found in /usr/include - Boost dependent tests activated
-- Configuring done
-- Generating done
-- Build files have been written to: /home/dyarosh/space/eve/build
dyarosh@dyarosh-u:~/space/eve/build$ ninja -j 8 tests
[1/16123] Building CXX object test/doc/constant/CMakeFiles/doc.constant.valmin.unit.dir/valmin.cpp.o
FAILED: test/doc/constant/CMakeFiles/doc.constant.valmin.unit.dir/valmin.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/constant/CMakeFiles/doc.constant.valmin.unit.dir/valmin.cpp.o -MF test/doc/constant/CMakeFiles/doc.constant.valmin.unit.dir/valmin.cpp.o.d -o test/doc/constant/CMakeFiles/doc.constant.valmin.unit.dir/valmin.cpp.o -c ../test/doc/constant/valmin.cpp
In file included from ../test/doc/constant/valmin.cpp:1:
In file included from ../include/eve/constant/valmin.hpp:13:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
[2/16123] Building CXX object test/doc/core/CMakeFiles/doc.core.sqrt.unit.dir/sqrt.cpp.o
FAILED: test/doc/core/CMakeFiles/doc.core.sqrt.unit.dir/sqrt.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/core/CMakeFiles/doc.core.sqrt.unit.dir/sqrt.cpp.o -MF test/doc/core/CMakeFiles/doc.core.sqrt.unit.dir/sqrt.cpp.o.d -o test/doc/core/CMakeFiles/doc.core.sqrt.unit.dir/sqrt.cpp.o -c ../test/doc/core/sqrt.cpp
In file included from ../test/doc/core/sqrt.cpp:1:
In file included from ../include/eve/function/sqrt.hpp:14:
In file included from ../include/eve/function/definition/sqrt.hpp:14:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
[3/16123] Building CXX object test/doc/core/CMakeFiles/doc.core.tofloat.unit.dir/tofloat.cpp.o
FAILED: test/doc/core/CMakeFiles/doc.core.tofloat.unit.dir/tofloat.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/core/CMakeFiles/doc.core.tofloat.unit.dir/tofloat.cpp.o -MF test/doc/core/CMakeFiles/doc.core.tofloat.unit.dir/tofloat.cpp.o.d -o test/doc/core/CMakeFiles/doc.core.tofloat.unit.dir/tofloat.cpp.o -c ../test/doc/core/tofloat.cpp
In file included from ../test/doc/core/tofloat.cpp:1:
In file included from ../include/eve/function/tofloat.hpp:14:
In file included from ../include/eve/function/definition/tofloat.hpp:14:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
[4/16123] Building CXX object test/doc/core/CMakeFiles/doc.core.sub.unit.dir/sub.cpp.o
FAILED: test/doc/core/CMakeFiles/doc.core.sub.unit.dir/sub.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/core/CMakeFiles/doc.core.sub.unit.dir/sub.cpp.o -MF test/doc/core/CMakeFiles/doc.core.sub.unit.dir/sub.cpp.o.d -o test/doc/core/CMakeFiles/doc.core.sub.unit.dir/sub.cpp.o -c ../test/doc/core/sub.cpp
In file included from ../test/doc/core/sub.cpp:1:
In file included from ../include/eve/function/sub.hpp:14:
In file included from ../include/eve/function/definition/sub.hpp:13:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
[5/16123] Building CXX object test/doc/core/CMakeFiles/doc.core.toint.unit.dir/toint.cpp.o
FAILED: test/doc/core/CMakeFiles/doc.core.toint.unit.dir/toint.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/core/CMakeFiles/doc.core.toint.unit.dir/toint.cpp.o -MF test/doc/core/CMakeFiles/doc.core.toint.unit.dir/toint.cpp.o.d -o test/doc/core/CMakeFiles/doc.core.toint.unit.dir/toint.cpp.o -c ../test/doc/core/toint.cpp
In file included from ../test/doc/core/toint.cpp:1:
In file included from ../include/eve/function/toint.hpp:14:
In file included from ../include/eve/function/definition/toint.hpp:14:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
[6/16123] Building CXX object test/doc/core/CMakeFiles/doc.core.ulpdist.unit.dir/ulpdist.cpp.o
FAILED: test/doc/core/CMakeFiles/doc.core.ulpdist.unit.dir/ulpdist.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/core/CMakeFiles/doc.core.ulpdist.unit.dir/ulpdist.cpp.o -MF test/doc/core/CMakeFiles/doc.core.ulpdist.unit.dir/ulpdist.cpp.o.d -o test/doc/core/CMakeFiles/doc.core.ulpdist.unit.dir/ulpdist.cpp.o -c ../test/doc/core/ulpdist.cpp
In file included from ../test/doc/core/ulpdist.cpp:1:
In file included from ../include/eve/function/ulpdist.hpp:14:
In file included from ../include/eve/function/definition/ulpdist.hpp:14:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
[7/16123] Building CXX object test/doc/core/CMakeFiles/doc.core.trunc.unit.dir/trunc.cpp.o
FAILED: test/doc/core/CMakeFiles/doc.core.trunc.unit.dir/trunc.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/core/CMakeFiles/doc.core.trunc.unit.dir/trunc.cpp.o -MF test/doc/core/CMakeFiles/doc.core.trunc.unit.dir/trunc.cpp.o.d -o test/doc/core/CMakeFiles/doc.core.trunc.unit.dir/trunc.cpp.o -c ../test/doc/core/trunc.cpp
In file included from ../test/doc/core/trunc.cpp:1:
In file included from ../include/eve/function/trunc.hpp:14:
In file included from ../include/eve/function/definition/trunc.hpp:14:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
[8/16123] Building CXX object test/doc/core/CMakeFiles/doc.core.tanh.unit.dir/tanh.cpp.o
FAILED: test/doc/core/CMakeFiles/doc.core.tanh.unit.dir/tanh.cpp.o 
/usr/local/bin/clang++   -Itts-src/include -I../test -I../include -stdlib=libc++   -std=c++17 -Wall -Wno-missing-braces -MD -MT test/doc/core/CMakeFiles/doc.core.tanh.unit.dir/tanh.cpp.o -MF test/doc/core/CMakeFiles/doc.core.tanh.unit.dir/tanh.cpp.o.d -o test/doc/core/CMakeFiles/doc.core.tanh.unit.dir/tanh.cpp.o -c ../test/doc/core/tanh.cpp
In file included from ../test/doc/core/tanh.cpp:1:
In file included from ../include/eve/function/tanh.hpp:14:
In file included from ../include/eve/function/definition/tanh.hpp:14:
In file included from ../include/eve/detail/overload.hpp:14:
In file included from ../include/eve/arch/spec.hpp:15:
In file included from ../include/eve/arch/x86/spec.hpp:14:
In file included from ../include/eve/arch/x86/predef.hpp:45:
../include/eve/detail/spy.hpp:465:31: error: no template named 'stdcpp_t' in namespace 'spy::detail'
  using stdlib_type = detail::stdcpp_t<(_LIBCPP_VERSION/1000)%10,0,_LIBCPP_VERSION%1000,0>;
                      ~~~~~~~~^
../include/eve/detail/spy.hpp:475:20: error: unknown type name 'stdlib_type'
  constexpr inline stdlib_type stdlib;
                   ^
../include/eve/detail/spy.hpp:480:45: error: no return statement in constexpr function
  inline constexpr stdlib_info<SLIB,M,N,P>::operator bool() const noexcept
                                            ^
3 errors generated.
ninja: build stopped: subcommand failed.

clarify and revisit the behavior relations between values, logical values and bit sets

make for benches the same layout than for other tests

the directories tree and the hierarchy of make targets ought to be the same (or similar) for benches and for random tests.
Also the the benches cpp ought to be generated autmagically from hpp sources.

upgrade & downgrade traits

Take a type and make it one notch smaller or bigger (int->short or short->int).

• upgrade
• downgrade

Support AVX512

New plan of actions:

• Add support for detecting AVX512* ABI and related register type
• Adapt logical to AVX512 types
• Adapt internal operators
• Optimize other functions

saturate

bitwise_cast and wide_cast have the syntax xxx_cast(x,as())
saturate has saturate(as(),x)
perhaps we must change saturate parameter order to be consistent

Implement element_type

element_type<T> is value_type_t<T> except for types like logical and complex where it's identity

rewrite bitwise_select and if_else

the scalar version of bitwise_select is in fact generic and could be displaced.
Once done, if_else common version could be written in terms of bitwise_select

Add ordering for API/ABI object

if constexpr(current_api >= sse4_1_ should be a thing

what remains to do to achieve a first step ?

TO REMOVE DEFINITIVELY:
bitfloating.hpp
compare_equal.hpp
compare_greater_equal.hpp
compare_greater.hpp
compare_less_equal.hpp
compare_less.hpp
compare_not_equal.hpp
divides.hpp
enumerate.hpp
dot.hpp
fix.hpp
genmaskc.hpp
genmask.hpp
group.hpp
hmsb.hpp
ifix.hpp
if_allbits_else_zero.hpp
if_dec.hpp
if_inc.hpp
if_minus.hpp
if_neg.hpp
ifnot_dec.hpp
ifnot_inc.hpp
ifnot_minus.hpp
ifnot_neg.hpp
ifnot_plus.hpp
if_plus.hpp
if_zero_else_allbits.hpp
ilogb.hpp
insert.hpp
majority.hpp
mask2logical.hpp
minmod.hpp
nearbyint.hpp
neg.hpp
nextpow2.hpp
predecessor.hpp
rem_2pi.hpp
rem_pio2_straight.hpp
remquo.hpp
repeat_lower_half.hpp
repeat_upper_half.hpp
reverse.hpp
sbits.hpp
shift_left.hpp
shift_right.hpp
sincospi.hpp
slide.hpp
split_high.hpp
split.hpp
split_low.hpp
split_multiplies.hpp
successor.hpp
tenpower.hpp <- a fusionner dans exp10
deinterleave_first.hpp
deinterleave_second.hpp
interleave_even.hpp
interleave_first.hpp
interleave_odd.hpp
interleave_second.hpp

REMOVE NOW BUT TO PUT BACK LATER:
bits.hpp
bit_maskc <- a faire plus tard
clz.hpp <- a renommer
ctz.hpp <- a renommer
lexicographical_compare
cotd.hpp
cscd.hpp
cummax.hpp
cummin.hpp
cumprod.hpp
cumsum.hpp
deinterleave.hpp
erfc.hpp
erfcx.hpp
erf.hpp
exprecnegc.hpp
exprecneg.hpp
expx2.hpp
ffs.hpp <- a renommer
fpclassify.hpp
gamma.hpp
gammaln.hpp
hi.hpp
indeg.hpp -> radindeg.hpp
interleave.hpp
lo.hpp
ilog2.hpp <- a renommer
inrad.hpp -> deginrad.hpp
isincluded_c.hpp <- a renommer
isincluded.hpp <- a renommer
safe_max.hpp <- a mettre dans detail
safe_min.hpp <- a mettre dans detail
reversebits.hpp
rol.hpp
ror.hpp
rrol.hpp
rror.hpp
nthroot.hpp
secd.hpp
signbit.hpp
signgam.hpp
sindcosd.hpp <- a renommer
sind.hpp
tand.hpp
significants.hpp <- a renommer
sqrt1pm1.hpp
stirling.hpp
subadd.hpp

KEEP FOR RELEASE:
maximum.hpp
minimum.hpp
prod.hpp
shuffle.hpp
sinc.hpp
sincpi.hpp
sinhc.hpp
sum.hpp

TO INVESTIGATE:
reldist.hpp
swapbytes.hpp

conditional swap

a swap[cond](a,b) would be great.
It seems that could be a shuffle

Support ARM SVE

power

power works fine in avx2 and works all the way. But it has to be benched more thoroghly (in ssse4 for instance) to see if a convert from wide raw method to wide result + limits handling will not be faster than the direct implementation.

Provide object like decorator

so that complex_(sqrt) returns an object able to be decorated further

negate negatenz

these functions had been forgotten and but be implemented

Consider else option for conditional function

Something like :

auto f = (abs | v)[x<0]
auto f = (abs | v).not_[x<0]
auto f = (abs | zero)[x<0]

enforce_precision

g++-9 seems to have a flaw with combined -mavx2 and -mfma options that produce fast-math like
over-optimization as (x-c)+c evaluation to x in floating point computation.

this behaviour is like ub sensitive to code displacement or insered write orders and disappear with debug option

conformity with std C++20 naming

c++20 Numerics library définit un header bit
https://en.cppreference.com/w/cpp/header/bit
ne devrait-on pas adopter sa terminologie ?

• ceil2 -> bit_ceil
• floor2 -> bit_floor
• en particulier remplacer bitwise by bit dans nos noms de fonctions ?
• et mettre les bon noms pour clz etc. qd on les ajoutera
• et remplacer popcnt par popcount

we need indexed load and multi extraction syntax w=v[i] for wide v and wide i

indexed load will be useful in the new reduction scheme for trigo

the multi extraction must work this way
if v is of type wide<T,N> and ind of type wide<I,M> containing integers in the range 0..N-1
w=v[i] will return a wide<T,M> vector such that w[j] =v[i[j]] for each j in [0, N-1]

We need :

• a gather-like load
• a dynamic shuffle, maybe called lookup?

Write a GLossary page doc

This doc will define Elementwsie operation, and other non-Concepts design elements

popcount simd must be revisited

using perhaps:

https://github.com/WojciechMula
https://github.com/WojciechMula/sse-popcount/blob/master/popcnt-sse-bit-parallel.cpp
https://github.com/WojciechMula/sse-popcount/blob/master/popcnt-sse-bit-parallel-better.cpp

or

https://github.com/kimwalisch/libpopcnt/blob/master/libpopcnt.h

Add support fro constant generation from a bit pattern

constant<T,P>() should generate a constant of type T which bit patterns is P

Consolidate aggregator files

Files like arithmetic.hpp need to be cleared up and split if needed in accordance to the documentation.

Orphans files need to be cleaned up too.

XOP

si XOP est défini il faut donner une valeur à EVE_HW_X86

Use if_else instead of map in some abs implementation

Implement `is_equal`

Clarify inrad/indeg etc relevance

deginrad, inrad,indeg and inpi are in a boat and they're not consistent

logical_xxx and bit_xxx

naming ?

Implement `logical_not`

Loading `aligned_ptr<const T>`

Test case:

TEST_CASE("eve_extra.load_const_aligned_ptr", "[eve_extra]") {
  using wide = eve::wide<char, eve::fixed<16>>;
  alignas(wide) std::array<char, 16> data;
  data.fill(1);

  eve::aligned_ptr<const char, 16> ptr{data.cbegin()};

  auto loaded = eve::load(ptr, eve::as_<wide>{});
  REQUIRE(loaded == wide{1});
}

Error message:

[1/2] Building CXX object src/test/CMakeFiles/tests.dir/eve_extra.t.cc.o
FAILED: src/test/CMakeFiles/tests.dir/eve_extra.t.cc.o 
/usr/local/bin/clang++   -I../src/../../eve/include -I../src/. -O3 -DNDEBUG   -Werror -Wall -Wextra -Wpedantic -Og -g -Wno-unused-parameter -DEVE_DISABLE_ASSERTS -fno-omit-frame-pointer --std=c++17 -stdlib=libc++ -fsanitize=address -march=native -std=gnu++17 -MD -MT src/test/CMakeFiles/tests.dir/eve_extra.t.cc.o -MF src/test/CMakeFiles/tests.dir/eve_extra.t.cc.o.d -o src/test/CMakeFiles/tests.dir/eve_extra.t.cc.o -c ../src/test/eve_extra.t.cc
../src/test/eve_extra.t.cc:115:17: error: no matching function for call to 'load'
  auto loaded = eve::load(ptr, eve::as_<wide>{});
                ^~~~~~~~~
../src/../../eve/include/eve/function/load.hpp:44:24: note: candidate template ignored: deduced conflicting types for parameter 'T' ('const char' vs. 'eve::wide<char, eve::fixed<16>, eve::sse_>')
  EVE_FORCEINLINE auto load(aligned_ptr<T, Align> ptr, as_<T> const &) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:76:24: note: candidate template ignored: deduced conflicting types for parameter 'T' ('const char' vs. 'char')
  EVE_FORCEINLINE auto load(aligned_ptr<T, Align> ptr, as_<wide<T, Size, ABI>> const &) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:38:24: note: candidate template ignored: could not match 'T *' against 'eve::aligned_ptr<const char, 16>'
  EVE_FORCEINLINE auto load(T *ptr, as_<T> const &) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:66:24: note: candidate template ignored: could not match 'T *' against 'eve::aligned_ptr<const char, 16>'
  EVE_FORCEINLINE auto load(T *ptr, as_<wide<T, Size, ABI>> const &) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:84:24: note: candidate template ignored: could not match 'logical<T> *' against 'eve::aligned_ptr<const char, 16>'
  EVE_FORCEINLINE auto load(logical<T> *ptr, as_<logical<wide<T, Size, ABI>>> const &) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:94:24: note: candidate template ignored: could not match 'logical<type-parameter-0-2>' against 'const char'
  EVE_FORCEINLINE auto load(aligned_ptr<logical<T>, Align> ptr,
                       ^
../src/../../eve/include/eve/function/load.hpp:26:24: note: candidate function template not viable: requires single argument 'ptr', but 2 arguments were provided
  EVE_FORCEINLINE auto load(T *ptr) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:32:24: note: candidate function template not viable: requires single argument 'ptr', but 2 arguments were provided
  EVE_FORCEINLINE auto load(aligned_ptr<T, Align> ptr) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:54:24: note: candidate function template not viable: requires single argument 'ptr', but 2 arguments were provided
  EVE_FORCEINLINE auto load(wide<T, Size, ABI> *ptr) noexcept
                       ^
../src/../../eve/include/eve/function/load.hpp:60:24: note: candidate function template not viable: requires single argument 'ptr', but 2 arguments were provided
  EVE_FORCEINLINE auto load(aligned_ptr<wide<T, Size, ABI>> ptr) noexcept
                       ^

Use SPY instead of hand-rolled platform thingy

also move SIMD detection to SPY

check() doesn't play nice with decorated functions

Apparently, the decoration mechanism has been implemented in a certain way while it was thought to be done otherwise. This leads to check() not interacting properly with decorated functions.

This needs to be solved without changing the semantic of decoration w/r to conditional functions

g++10 pb with avx2 fma

big_(sincos) simd have failures that only appear and not on all systems with -mfma -mavx2 for double.
Is it a compiler pb ? The failure seems difficult to isolate and write orders in the code often makes
it disappear.

Implement `is_nez`

is_nez is is_not_equal(x,0) for arithmetic types and identity for logical

as_ status is unclear

multiple points:

• [x ] Use as<T>()in place of as_<T>{} calls
• we must provide shortcuts for specific type markup like single_, uchar_ etc...

lo and hi must be written with a shuffle

partitionning the wide in two parts containing the low and the hi parts and returning the good half

do ppc and arm version of all.hpp

Make a model function, constant, type page for the doc

Logical/Arithmetic converters

let's have a as_logical(v) and as_arithmetic(v) cnersions functions to ease the logical<->arithmetic conversion.

Implement `is_not_equal`

Eve doesn't disable asserts on clang -O3

I don't exaclty know how, but:

Given code:

using wide = eve::wide<char, eve::fixed<32>>;
eve::aligned_ptr<char, alignof(wide)> test(char* s) {
  return {s};
}

When I compile with:

clang++ -std=c++17 -O3  -march=native

I get

.L_Z4testPc$local:
	.cfi_startproc
# %bb.0:
	push	rax
	.cfi_def_cfa_offset 16
	mov	rsi, rdi
	mov	rdi, rsp
	call	_ZN3eve11aligned_ptrIcLm32EEC2EPc
	mov	rax, qword ptr [rsp]
	pop	rcx
	.cfi_def_cfa_offset 8
	ret
.Lfunc_end0:
... # lots of code with iostream

When I pass

clang++ -std=c++17 -DNDEBUG -O3  -march=native 

I get

.L_Z4testPc$local:
	.cfi_startproc
# %bb.0:
	mov	rax, rdi
	ret

which is what I'd expect.

Same for

clang++ -std=c++17 -DEVE_DISABLE_ASSERTS -O3  -march=native 

Could be a clang bug I don't know but still.

Make native, aggregated and emulated concept work for every types

With the caveat that we consider that abi_of<T> is cpu_ for scalar and therefore is considered native.

Implement `bitwise_mask`

bitwise_mask uses is_nez for arithmetics and bitwise_cast for logicals

scalaire pur

faut-il pouvoir tester eve sans aucun simd et si oui comment ?

logical and bool

It will be good to add two overload to each of the binary logical operators logical_xxx, between a logical/bool and bool/logical

why do we stil separate scalar and common ?

Is there still a reason to maintain separate files for scalar and common implementation of most functions ?
More if constexpr could perhaps do the job allowing the suppression of a bunch of include files ?

AVX2 equality comparison splits register in 2

After dissambing, found an issue:

Funciton:

auto test(eve::wide<char, eve::fixed<32>> x,
          eve::wide<char, eve::fixed<32>> y) {
  return x == y;
}

Assembly:

	vpcmpeqb	xmm2, xmm0, xmm1
	vextracti128	xmm0, ymm0, 1
	vextracti128	xmm1, ymm1, 1
	vpcmpeqb	xmm0, xmm0, xmm1
	vinserti128	ymm0, ymm2, xmm0, 1
	ret

Expected:

vpcmpeqb        ymm0, ymm0, ymm1

Compiler options:

clang++ -std=c++17 -O3 -DEVE_DISABLE_ASSERTS  -march=native

Prepare bench systems