julialinearalgebra / blis.jl Goto Github PK

View Code? Open in Web Editor NEW

26.0 4.0 4.0 1.56 MB

This repo plans to provide a low-level Julia wrapper for BLIS typed interface.

License: BSD 3-Clause "New" or "Revised" License

Julia 100.00%

blis wrapper julia matrix-multiplications

blis.jl's People

Contributors

Stargazers

Watchers

Forkers

ikirill ranocha jd-foster amontoison

blis.jl's Issues

Installing BLIS.jl yields an older version

When I install BLIS.jl with julia 1.7.3, it gives BLIS v0.2.0 and blis_jll v0.8.1, despite there being more recent updates to the package. In particular, I'm also getting the crash fixed in #12

I think this commit: f805070 might have bumped the version down by accident so Pkg.jl automatically installs an older version. I had to do dev BLIS to work around this.

▶ testjulia1 % julia --project=@.                                                                                                                                                    
               _                                                                                                                                                                     
   _       _ _(_)_     |  Documentation: https://docs.julialang.org                                                                                                                  
  (_)     | (_) (_)    |                                                                                                                                                             
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.                                                                                                                      
  | | | | | | |/ _` |  |                                                                                                                                                             
  | | |_| | | | (_| |  |  Version 1.7.3 (2022-05-06)                                                                                                                                 
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release                                                                                                                    
|__/                   |                                                                                                                                                             
                                                                                                                                                                                     
(@v1.7) pkg> activate .                                                                                                                                                              
  Activating new project at `~/Sandboxes/testjulia1`                                                                                                                                 
                                                                                                                                                                                     
(testjulia1) pkg> add BLIS                                                                                                                                                           
    Updating registry at `~/.julia/registries/General.toml`                                                                                                                          
   Resolving package versions...                                                                                                                                                     
    Updating `~/Sandboxes/testjulia1/Project.toml`                                                                                                                                   
  [32f2db32] + BLIS v0.2.0                                                                                                                                                           
    Updating `~/Sandboxes/testjulia1/Manifest.toml`                                                                                                                                  
  [32f2db32] + BLIS v0.2.0                                                                                                                                                           
  [692b3bcd] + JLLWrappers v1.4.1                                                                                                                                                    
  [21216c6a] + Preferences v1.3.0                                                                                                                                                    
  [1d63c593] + LLVMOpenMP_jll v14.0.4+0                                                                                                                                              
  [6136c539] + blis_jll v0.8.1+3                                                                                                                                                     
etc.

(the version numbers for BLIS and blis_jll are wrong)

Crash on multiplying large dense matrices

This runs in the base environment in a new Julia session:

  | | |_| | | | (_| |  |  Version 1.7.1 (2021-12-22)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

julia> n = 1000; a = fill(1., n, n); a*a;

However:

julia> using BLIS
[ Info: blis_jll yields BLIS installation: ~/.julia/artifacts/567e1b2234ebc730bc16e33871144bf9b561f1fa/lib/libblis.4.0.0.dylib.

julia> n = 1000; a = fill(1., n, n); a*a;

signal (10): Bus error: 10
in expression starting at REPL[4]:1
jl_system_image_data at /Applications/Julia-1.7.app/Contents/Resources/julia/lib/julia/sys.dylib (unknown line)
Allocations: 12825733 (Pool: 12821375; Big: 4358); GC: 13
[1]    4388 bus error  julia

Move repo to JuliaLinearAlgebra org?

To improve visibility and maintainability, would you be comfortable moving this repo to https://github.com/JuliaLinearAlgebra ownership?

TagBot trigger issue

This issue is used to trigger TagBot; feel free to unsubscribe.

If you haven't already, you should update your TagBot.yml to include issue comment triggers.
Please see this post on Discourse for instructions and more details.

Update to use libblastrampoline

It would be great to use the same mechanism that MKL.jl uses now, and leverage libblastrampoline.

https://github.com/JuliaLinearAlgebra/MKL.jl/blob/master/src/MKL.jl

Evict BLIS Backends to Another Package

cf. discussions at the end of #20

Crash when benchmarking `BLIS.BLASInterface.axpy!`

Simple reproducer:

julia> versioninfo()
Julia Version 1.7.2
Commit bf53498635 (2022-02-06 15:21 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-4870HQ CPU @ 2.50GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, haswell)

(tmp) pkg> st -m BLIS blis_jll
      Status `/tmp/Manifest.toml`
  [32f2db32] BLIS v0.2.0
  [6136c539] blis_jll v0.8.1+0

julia> using BenchmarkTools, BLIS
[ Info: blis_jll yields BLIS installation: /home/mose/.julia/artifacts/3bd06aa4f9939a35f5cb53773bd177f94ec99078/lib/libblis.so.

julia> @benchmark BLIS.BLASInterface.axpy!(a, x, y) setup=(T=Float32; N=100; a=randn(T); x=randn(T, N); y=randn(T, N)) evals=1
GC error (probable corruption) :
Allocations: 21278104 (Pool: 21271094; Big: 7010); GC: 26

!!! ERROR in jl_ -- ABORTING !!!
0x7f70fc77f010: Root object: 0x7f7152ae4010 :: 0x7f7149ad2070 (bits: 3)
        of type Task
0x7f70fc77f028: Root object: 0x7f71498c1760 :: 0x7f7149859b30 (bits: 3)
        of type Module
0x7f70fc77f040: Root object: 0x7f714985a620 :: 0x7f714985a2b0 (bits: 3)
        of type Array{Any, 1}
0x7f70fc77f058: Root object: 0x7f714e0ed220 :: 0x7f714985a380 (bits: 3)
        of type UnionAll
0x7f70fc77f070: Root object: 0x7f714d451b60 :: 0x7f71498d7f50 (bits: 3)
        of type Core.TypeMapEntry
0x7f70fc77f088: Root object: 0x7f714d5bbfa0 :: 0x7f71498d7f50 (bits: 3)
        of type Core.TypeMapEntry
0x7f70fc77f0a0: Root object: 0x7f714b87b220 :: 0x7f71498d7f50 (bits: 3)
        of type Core.TypeMapEntry

# Continues a _huge_ wall of errors

Perhaps some variables shared with the library aren't rooted correctly?

Unexpected "No Transpose" behaviour

Thanks for putting together this package.

I wanted to check that you were aware that the 'C' ("Conjugate Transpose) option when applied to real matrices has the same behaviour as the 'N' ("No Transpose") option, and when applied to complex matrices, it is the same as just conjugating the entries but without transpose.
My expectation follows from Reference BLAS notes: http://www.netlib.org/blas/blasqr.pdf
and here: http://www.math.utah.edu/software/lapack/lapack-blas/zgemm.html

I guess the reason for the behaviour is clear from the definition in types.jl:

char_to_trans = Dict( 'N' => BLIS_NO_TRANSPOSE,
                      'n' => BLIS_NO_TRANSPOSE,
                      'T' => BLIS_TRANSPOSE,
                      't' => BLIS_TRANSPOSE,
                      'C' => BLIS_CONJ_NO_TRANSPOSE,
                      'c' => BLIS_CONJ_NO_TRANSPOSE )

Here is a MWE.

A = Float64[1. 2.; 3. 4.]
B = Float64[5. 6.; 7. 8.]
C = zeros(Float64, 2, 2);

# # Note! : 'C' is equivalent to 'N', not 'T' on real matrices:
BLIS.BLASInterface.gemm!('N', 'C', 1.0, A, B, 1.0, copy(C)) == BLIS.BLASInterface.gemm!('N', 'N', 1.0, A, B, 1.0, copy(C))
# true

I was expecting

BLIS.BLASInterface.gemm!('N', 'C', 1.0, A, B, 1.0, copy(C)) == BLIS.BLASInterface.gemm!('N', 'T', 1.0, A, B, 1.0, copy(C))
# false : not expected

But now I see that LinearAlgebra.BLAS seems to follow the same convention too!

A = Matrix(UniformScaling(1.0+0.0im), 2, 2)
B = rand(ComplexF64, 2, 2);
@assert LinearAlgebra.BLAS.gemm('N','N',A,B) == B
@assert LinearAlgebra.BLAS.gemm('N','T',A,B) == transpose(B)
@assert LinearAlgebra.BLAS.gemm('N','C',A,B) == conj(B)
@assert LinearAlgebra.BLAS.gemm('N','C',A, Matrix(B')) == transpose(B)

Clarification welcome!

Strided GEMM Failing

Platform: macOS and some Linux
Problem: multiplying A & B with at least one side generically-strided (e.g. A = view(Abase, 1:Mst:M, ...) | Mst>1) yields all-0 results.
Possible cause: upstream (reproduced in C).
Status: investigating.
Workaround: as is stated in README.jl, failed methods can be blacklisted to block from the LinearAlgebra interface:

echo "gemm" > ~/.blis_jlbla_blacklist
# Trigger JIT reload.
touch $(find ~/.julia/packages/ -name "BLIS.jl" | tail -n 1)

Crashing on Windows with julia 1.8.4 or above

Here is a MWE crash reproducer (fails at time of writing onJulia Version 1.10.0-DEV.1059) extracted from init_test_mmul.jl:

using LinearAlgebra

T = Float32
αr = 1.1
βr = 1.1
αu = T(αr)
βu = T(βr)

χlarge = 500
Alarge_base = rand(ComplexF64, χlarge, χlarge)
Blarge_base = rand(ComplexF64, χlarge, χlarge)
Clarge_base = rand(ComplexF64, χlarge, χlarge)

Alarge = zeros(T, χlarge, χlarge)
Blarge = zeros(T, χlarge, χlarge)

elconv = x -> real(x)

Alarge .= elconv.(Alarge_base)
Blarge .= elconv.(Blarge_base)

Clarge_gemm_cur  = T.(elconv.(Clarge_base))

BLAS.gemm!('N', 'N', αu, Alarge, Blarge, βu, Clarge_gemm_cur)

X_jl = deepcopy(Clarge_gemm_cur)

using BLIS

Clarge_gemm_cur  = T.(elconv.(Clarge_base))
BLAS.gemm!('N', 'N', αu, Alarge, Blarge, βu, Clarge_gemm_cur)

X_bl = deepcopy(Clarge_gemm_cur)

@assert all(X_jl .≈ X_bl)

v0.3.0 To-do List

Waiting for more Arm64 works in the upstream (by myself, though).
Introduce more tests (sizes, transpose, etc.)
Unsure Option to not override the LinearAlgebra module.
Mixed precision

Update to Version 0.8.1

https://github.com/flame/blis/releases/tag/0.8.1

Likely would first require updating the build at https://github.com/JuliaPackaging/Yggdrasil/tree/master/B/blis