Created
April 14, 2021 20:16
-
-
Save staticfloat/164a795e6eb9ae4ec4cfaecd66458aa4 to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
include("util.jl") | |
# First, do OpenBLAS64 vs. OpenBLAS32 testing (should be about the same) | |
BLAS.lbt_forward(OpenBLAS_jll.libopenblas_path; clear=true) | |
BLAS.lbt_forward(OpenBLAS32_jll.libopenblas_path) | |
config = BLAS.get_config() | |
@info("OpenBLAS 64/32") | |
@show config | |
@show gemm_test(500, 64) | |
@show gemm_test(500, 32) | |
# Next, do OpenBLAS64 vs. Accelerate testing | |
BLAS.lbt_forward(OpenBLAS_jll.libopenblas_path; clear=true) | |
BLAS.lbt_forward("/System/Library/Frameworks/Accelerate.framework/Versions/A/Accelerate") | |
@info("Accelerate 32") | |
@show config | |
@show gemm_test(500, 32) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[deps] | |
OpenBLAS32_jll = "656ef2d0-ae68-5445-9ca0-591084a874a2" |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
using OpenBLAS_jll, OpenBLAS32_jll, libblastrampoline_jll, LinearAlgebra, Libdl | |
get_dgemm_addr(::Val{Int32}) = dlsym(libblastrampoline_jll.libblastrampoline_handle, :dgemm_) | |
get_dgemm_addr(::Val{Int64}) = dlsym(libblastrampoline_jll.libblastrampoline_handle, :dgemm_64_) | |
function gemm!(transA::AbstractChar, transB::AbstractChar, | |
alpha::Union{(Float64), Bool}, | |
A::AbstractVecOrMat{Float64}, B::AbstractVecOrMat{Float64}, | |
beta::Union{(Float64), Bool}, | |
C::AbstractVecOrMat{Float64}, ::Val{BlasInt}) where {BlasInt <: Union{Int32,Int64}} | |
m = size(A, transA == 'N' ? 1 : 2) | |
ka = size(A, transA == 'N' ? 2 : 1) | |
kb = size(B, transB == 'N' ? 1 : 2) | |
n = size(B, transB == 'N' ? 2 : 1) | |
if ka != kb || m != size(C,1) || n != size(C,2) | |
throw(DimensionMismatch("A has size ($m,$ka), B has size ($kb,$n), C has size $(size(C))")) | |
end | |
fptr = get_dgemm_addr(Val(BlasInt)) | |
ccall(:jl_breakpoint, Cvoid, (Any,), fptr) | |
ccall(fptr, Cvoid, (Ref{UInt8}, Ref{UInt8}, Ref{BlasInt}, Ref{BlasInt}, | |
Ref{BlasInt}, Ref{Float64}, Ptr{Float64}, Ref{BlasInt}, | |
Ptr{Float64}, Ref{BlasInt}, Ref{Float64}, Ptr{Float64}, | |
Ref{BlasInt}, Clong, Clong), | |
transA, transB, m, n, | |
ka, alpha, A, max(1,stride(A,2)), | |
B, max(1,stride(B,2)), beta, C, | |
max(1,stride(C,2)), 1, 1) | |
C | |
end | |
function gemm(transA::AbstractChar, transB::AbstractChar, alpha::(Float64), A::AbstractMatrix{Float64}, B::AbstractMatrix{Float64}, ::Val{BlasInt}) where {BlasInt <: Union{Int32,Int64}} | |
gemm!(transA, transB, alpha, A, B, zero(Float64), similar(B, Float64, (size(A, transA == 'N' ? 1 : 2), size(B, transB == 'N' ? 2 : 1))), Val(BlasInt)) | |
end | |
function gemm_test(n::Int, interface::Int) | |
BlasInt = Int32 | |
if interface == 64 | |
BlasInt = Int64 | |
end | |
# First, warmup | |
X = fill(0.1, 200, 200) | |
gemm('N', 'N', 1.0, X, X, Val(BlasInt)) | |
# Next, actually call it; do so ten times, then return the minimum | |
X = fill(0.1, n, n) | |
ts = [@elapsed gemm('N', 'N', 1.0, X, X, Val(BlasInt)) for _ in 1:10] | |
return minimum(ts) | |
end |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment