C++ wrappers for SIMD intrinsics and parallelized, optimized mathematical functions (SSE, AVX, AVX512, NEON, SVE))
Home Page: https://xsimd.readthedocs.io/
License: BSD 3-Clause "New" or "Revised" License
I have benchmarked posix_memalign and _mm_malloc, and they are quite a bit slower for 32 bit alignments than the xaligned_malloc/free.
Or we could enable/disable them by some compile time flags?
pow of complex<float> has some accuracy issues with AVX512.
Currently the build fails for me with the latest Intel Compiler on MSVC (and Visual Studio 2015). The build aborts in cmake configuration with error
-- Performing Test HAS_CPP11_FLAG
-- Performing Test HAS_CPP11_FLAG - Failed
CMake Error at test/CMakeLists.txt:53 (message):
Unsupported compiler -- xsimd requires C++11 support!
-- Configuring incomplete, errors occurred!
I assume the detection is not perfect, because the Intel Compiler has support for C++11. Is there anything I can do to help?
We need the modulo operator in xsimd for operations in xtensor.
An aligned pool allocator that keeps memory around until explicit flush/cleanup would be a nice additional feature.
This is a requirement for logarithm functions.
In xsimd_avx_float.hpp, the function batch<float, 8>::store_unaligned() calls the aligned intrinsic _mm256_store_ps() instead of the unaligned _mm256_storeu_ps(), leading to a segfault when the address is actually unaligned. I didn't check if the same problem occurs for other types.
I am not sure if this is a bug or feature, but there is no way to do a load/store of a vector which is smaller than the native vector width (e.g. a 128-bit batch<float, 4> on AVX) using the free function interface from xsimd/xsimd.hpp.
The way the current interface works, one has to use the constructor/method syntax for this kind of loads and stores.
Basi functions which are missing:
Classification functions:
Following changes are required:
Logarithm functions:
Hello!
I tried compiling the simple test program you include in your documentation
#include <iostream>
#include "xsimd/xsimd.hpp"
namespace xs = xsimd;
int main(int argc, char* argv[])
{
xs::batch<double, 4> a(1.5, 2.5, 3.5, 4.5);
xs::batch<double, 4> b(2.5, 3.5, 4.5, 5.5);
auto mean = (a + b) / 2;
std::cout << mean << std::endl;
return 0;
}
I compile using
g++ -std=c++14 -o out -I xsimd/include/ main.cpp
where gcc is version 5.4.1 .
I get the following error
In file included from xsimd/types/xsimd_types_include.hpp:15:0,
from xsimd/types/xsimd_traits.hpp:12,
from xsimd/xsimd.hpp:14,
from main.cpp:2:
xsimd/types/xsimd_sse_int32.hpp: In function ‘xsimd::batch<int, 4ul> xsimd::select(const xsimd::batch_bool<int, 4ul>&, const xsimd::batch<int, 4ul>&, const xsimd::batch<int, 4ul>&)’:
xsimd/types/xsimd_sse_int32.hpp:441:70: error: ‘s’ was not declared in this scope
return _mm_or_si128(_mm_and_si128(cond, a), _mm_andnot_si128(s, b));
^
In file included from xsimd/types/xsimd_types_include.hpp:16:0,
from xsimd/types/xsimd_traits.hpp:12,
from xsimd/xsimd.hpp:14,
from main.cpp:2:
xsimd/types/xsimd_sse_int64.hpp: In function ‘xsimd::batch<long int, 2ul> xsimd::select(const xsimd::batch_bool<long int, 2ul>&, const xsimd::batch<long int, 2ul>&, const xsimd::batch<long int, 2ul>&)’:
xsimd/types/xsimd_sse_int64.hpp:460:70: error: ‘s’ was not declared in this scope
return _mm_or_si128(_mm_and_si128(cond, a), _mm_andnot_si128(s, b));
^
I'm sure it's something simple I'm missing.
Thanks!
The issue is fully described here
The current implementations of this operators rely on operator[] and separately operates on elements of a batch.
This should be refactored with specifics intrinsics for better performances.
There are some small failures with loading 64bit integers from char for -arch=nocona
[ RUN ] xsimd.api_load
1: lhs = 67305985 - rhs = 1
2: lhs = 134678021 - rhs = 2
load uchar -> int64 : BAD
Nb diff : 2 (100%)
1: lhs = 67305985 - rhs = 1
2: lhs = 134678021 - rhs = 2
loadu uchar -> int64 : BAD
Nb diff : 2 (100%)
/home/wolfv/Programs/xsimd/test/xsimd_api_test.cpp:70: Failure
Value of: res
Actual: false
Expected: true
[ FAILED ] xsimd.api_load (0 ms)
and
[ RUN ] xsimd.sse_load
1: lhs = 67305985 - rhs = 1
2: lhs = 134678021 - rhs = 2
load uchar -> int64 : BAD
Nb diff : 2 (100%)
cvttps uses truncation and
3/5 == 0 becomes true which seems to be the default behavior in C++
Hello xsimd developers,
I was wondering if you have any timeline on AVX512 support?
I recently gained access to a Skylake-AVX512 Xeon server, and wanted to test out how the new instructions would fare on my xsimd-enhanced libraries. I went ahead and hacked together some very preliminary support for AVX512 in xsimd which is nevertheless sufficient for some initial experiments (single-precision floats, basic maths). Posting the link here, in case it is of any use:
(note that I am pretty much a newbie when it comes to SIMD programming, so probably there are inefficiencies and mistakes)
Thanks for the great library!
char loading to other batches currently makes use of lax conversion enabled by default on clang (apparently).
We should remove the need to enable this compiler flag on GCC so that it compiles without hassle (by using static casts where appropriate).
We should implement a transpose operation to transpose NxN matrix blocks (where N is batch width).
The interface should probably look like the one for haddp, e.g. taking a pointer of rows.
template <class T, N>
void transpose(batch<T, N>* rows)
{
... inplace transpose ...
}I think the tile says it all :)
When running "make/ninja xbenchmark" on my Haswell-based machine, a pair of "neon" rows is present in the table of timings, even though this CPU certainly has no neon instructions available.
Currently haddp is untested (and unused).
There is support for a batch of int32_t, but nothing for uint32_t. Same applies to uint64_t.
Trigonometric functions:
@iamthebot finally run into our missing support for char loading.
So we need to add support to this ASAP.
Maybe that's because the search box is currently broken. But I did not find documentation for hadd and other batch reducers on ReadTheDocs, even though I expected it to be generated from this Doxygen.
power, trigonometric and hyperbolic tests are failing for NEON instructions because of potential truncation.
We probably need to detect ARM or 32bit better.
It could be interesting to support intels scatter and gather instructions (however, scatter does seem to be available only since AVX512).
In the STL it's std::isnan and it's also xsimd::isinf.
So is there a reason for xsimd::is_nan or would it make sense to rename it to xsimd::isnan to adhere to STL and unify the xsimd interface?
SIMDified arange can give a ~4x improvement about std::iota and for loops.
So, I tried to integrate #100 in my project to see if it solves the problem it's intended to solve, and I discovered a bunch of issues related to constructors of batch and batch_bool. This means that we don't yet have unit tests for them.
We're using a __m256d floating point type to store the batch bool for float/double in SSE and AVX and do equality comparison using variants of mm_cmp_pd/ps(...)
The problem is that theses functions check for NaN as NaNs are incomparable -- and you can select a mode to get to a desired result (e.g. ordered comparsion results in NaN and number to be false or whatever).
The other fact is that a true value is represented by setting all bits to 1 – including the NaN indicating bit.
So currently, if you compare a batch_bool<double, 4> a(true); a == a with itself, the result will be filled with false as you compare two NaN numbers with each other.
Solution a) switch to integer type inside of batch_bool. This reduces also the amount of implementations (as we can share for int/float int64/double).
Just need to add a cast to or constructor from __m256d (as that's still the result type of the comparison of float/double batches).
Solution b) cast to int before comparison and cast back for storage.
Some of the mathematical functions need both the sine and the cosine of their argument. Since these functions share some steps in their computation (among them, the reduction), a sincos method computing both sine and cosine in a single pass may improve performances.
At least on x86, the fastest intrinsics for shuffling the contents of a vector or blending data from two vectors take an immediate operand, which must be a compile-time constant. So there would be a use case for a compile-time version of xsimd::select(), as it could use these faster instructions.
An example of prior art for this is the shuffle() instruction family in bSIMD:
Hyperbolic functions:
We have a prototype for a cumulative sum using xsimd here: https://gist.github.com/jpivarski/b2a04778124e7dc790d87fcdfd399e1e/
We should work to integrate it into the library.
Error and gamma functions:
Power functions:
sqrt has its specific intrinsics.
Mathematical functions of xsimd should accept integral types when their standard counterparts do.
Exponential functions:
Using version 3.0.1 + fix for select, here is the output from the tests:
[==========] Running 22 tests from 1 test case.
[----------] Global test environment set-up.
[----------] 22 tests from xsimd
[ RUN ] xsimd.sse_float_trigonometric
[ OK ] xsimd.sse_float_trigonometric (9 ms)
[ RUN ] xsimd.sse_double_trigonometric
[ OK ] xsimd.sse_double_trigonometric (9 ms)
[ RUN ] xsimd.sse_float_rounding
/<<PKGBUILDDIR>>/test/xsimd_rounding_test.cpp:34: Failure
Value of: res
Actual: false
Expected: true
[ FAILED ] xsimd.sse_float_rounding (0 ms)
[ RUN ] xsimd.sse_double_rounding
/<<PKGBUILDDIR>>/test/xsimd_rounding_test.cpp:41: Failure
Value of: res
Actual: false
Expected: true
[ FAILED ] xsimd.sse_double_rounding (0 ms)
[ RUN ] xsimd.sse_float_power
[ OK ] xsimd.sse_float_power (6 ms)
[ RUN ] xsimd.sse_double_power
[ OK ] xsimd.sse_double_power (4 ms)
[ RUN ] xsimd.sse_float_hyperbolic
[ OK ] xsimd.sse_float_hyperbolic (8 ms)
[ RUN ] xsimd.sse_double_hyperbolic
[ OK ] xsimd.sse_double_hyperbolic (5 ms)
[ RUN ] xsimd.sse_float_fp_manipulation
[ OK ] xsimd.sse_float_fp_manipulation (0 ms)
[ RUN ] xsimd.sse_double_fp_manipulation
[ OK ] xsimd.sse_double_fp_manipulation (0 ms)
[ RUN ] xsimd.sse_float_exponential
/<<PKGBUILDDIR>>/test/xsimd_exponential_test.cpp:35: Failure
Value of: res
Actual: false
Expected: true
[ FAILED ] xsimd.sse_float_exponential (10 ms)
[ RUN ] xsimd.sse_double_exponential
/<<PKGBUILDDIR>>/test/xsimd_exponential_test.cpp:42: Failure
Value of: res
Actual: false
Expected: true
[ FAILED ] xsimd.sse_double_exponential (7 ms)
[ RUN ] xsimd.sse_float_error_gamma
/<<PKGBUILDDIR>>/test/xsimd_error_gamma_test.cpp:35: Failure
Value of: res
Actual: false
Expected: true
[ FAILED ] xsimd.sse_float_error_gamma (23 ms)
[ RUN ] xsimd.sse_double_error_gamma
/<<PKGBUILDDIR>>/test/xsimd_error_gamma_test.cpp:42: Failure
Value of: res
Actual: false
Expected: true
[ FAILED ] xsimd.sse_double_error_gamma (13 ms)
[ RUN ] xsimd.sse_float_basic_math
[ OK ] xsimd.sse_float_basic_math (0 ms)
[ RUN ] xsimd.sse_double_basic_math
[ OK ] xsimd.sse_double_basic_math (0 ms)
[ RUN ] xsimd.sse_float_basic
[ OK ] xsimd.sse_float_basic (0 ms)
[ RUN ] xsimd.sse_double_basic
[ OK ] xsimd.sse_double_basic (0 ms)
[ RUN ] xsimd.sse_int32_basic
[ OK ] xsimd.sse_int32_basic (0 ms)
[ RUN ] xsimd.sse_int64_basic
[ OK ] xsimd.sse_int64_basic (0 ms)
[ RUN ] xsimd.sse_conversion
[ OK ] xsimd.sse_conversion (0 ms)
[ RUN ] xsimd.sse_cast
[ OK ] xsimd.sse_cast (0 ms)
[----------] 22 tests from xsimd (94 ms total)
[----------] Global test environment tear-down
[==========] 22 tests from 1 test case ran. (94 ms total)
[ PASSED ] 16 tests.
[ FAILED ] 6 tests, listed below:
[ FAILED ] xsimd.sse_float_rounding
[ FAILED ] xsimd.sse_double_rounding
[ FAILED ] xsimd.sse_float_exponential
[ FAILED ] xsimd.sse_double_exponential
[ FAILED ] xsimd.sse_float_error_gamma
[ FAILED ] xsimd.sse_double_error_gamma
6 FAILED TESTS
Batch constructor taking an initializer list should be available. Helpers allowing to build batches from initializer lists without knowing their real type should be implemented too.
I included xsimd.hpp from cloning the repo, then going into includes, and making a main.cpp at the same path:
#include "xsimd.hpp"
#include <iostream>
int main(){
std::cout << "Hello world.";
}
and it fails with the following error when I try to compile:
root@dc0238c74c63:/.../native/third_party/xsimd/include/xsimd# g++ main.cpp -o main
In file included from types/xsimd_types_include.hpp:22,
from types/xsimd_traits.hpp:14,
from xsimd.hpp:14,
from main.cpp:1:
types/xsimd_sse_int8.hpp: In static member function 'static xsimd::detail::sse_int8_batch_kernel<signed char>::batch_type xsimd::detail::batch_kernel<signed char, 16>::abs(const batch_type&)':
types/xsimd_sse_int8.hpp:590:32: error: '_mm_srai_epi8' was not declared in this scope
__m128i sign = _mm_srai_epi8(rhs, 31);
^~~~~~~~~~~~~
types/xsimd_sse_int8.hpp:590:32: note: suggested alternative: '_mm_srai_epi32'
__m128i sign = _mm_srai_epi8(rhs, 31);
^~~~~~~~~~~~~
_mm_srai_epi32
Here is my system information:
root@dc0238c74c63:/# cat /proc/cpuinfo
processor : 0
vendor_id : GenuineIntel
cpu family : 6
model : 63
model name : Intel(R) Xeon(R) CPU @ 2.30GHz
stepping : 0
microcode : 0x1
cpu MHz : 2300.000
cache size : 46080 KB
physical id : 0
siblings : 2
core id : 0
cpu cores : 1
apicid : 0
initial apicid : 0
fpu : yes
fpu_exception : yes
cpuid level : 13
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid xsaveopt
bugs :
bogomips : 4600.00
clflush size : 64
cache_alignment : 64
address sizes : 46 bits physical, 48 bits virtual
power management:
processor : 1
vendor_id : GenuineIntel
cpu family : 6
model : 63
model name : Intel(R) Xeon(R) CPU @ 2.30GHz
stepping : 0
microcode : 0x1
cpu MHz : 2300.000
cache size : 46080 KB
physical id : 0
siblings : 2
core id : 0
cpu cores : 1
apicid : 1
initial apicid : 1
fpu : yes
fpu_exception : yes
cpuid level : 13
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid xsaveopt
bugs :
bogomips : 4600.00
clflush size : 64
cache_alignment : 64
address sizes : 46 bits physical, 48 bits virtual
power management:
root@dc0238c74c63:/# gcc --version
gcc (Ubuntu 8.1.0-1ubuntu1) 8.1.0
Copyright (C) 2018 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
root@dc0238c74c63:/# lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description: Ubuntu 18.04.1 LTS
Release: 18.04
Codename: bionic
For API completness, it would be great to have support for nextafter
For some reason, the ReadTheDocs documentation does not have a working search index yet. That is sad, as it is a very important feature for finding functions in the docs when one doesn't know how the docs are organized.
P.S. This issue likely also affects other QuantStack projects using the same doc generation toolchain.
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.