BLIS now provides 64 bit binaries, add a LoopVectorization matmul (#52)

* BLIS now provides 64 bit binaries, add a default LoopVectorization matrix multiply

* Bump version

* Require LoopVectorization 0.12

* Require Gaius 0.6.4

* Update ci.yml

Co-authored-by: Dilum Aluthge <[email protected]>

Author: chriselrod

.github/workflows/ci.yml

@@ -60,6 +60,8 @@ jobs:
             Pkg.develop(PackageSpec(path=pwd()))
             Pkg.instantiate()'
         shell: bash
+        env:
+          JULIA_PKG_SERVER: ""
       - run: |
           julia --project=docs -e '
             using Documenter: doctest

Project.toml

@@ -1,7 +1,7 @@
 name = "BLASBenchmarksCPU"
 uuid = "5fdc822c-4560-4d20-af7e-e5ee461714d5"
 authors = ["Chris Elrod <[email protected]> and contributors"]
-version = "0.3.1"
+version = "0.3.2"
 
 [deps]
 BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
@@ -29,12 +29,12 @@ Colors = "0.12"
 DataFrames = "0.22"
 Fontconfig = "0.4"
 Gadfly = "1.3"
-Gaius = "0.5,0.6"
-LoopVectorization = "0.10, 0.11"
+Gaius = "0.6.4"
+LoopVectorization = "0.12"
 Octavian = "0.2"
 ProgressMeter = "1.4"
 Tullio = "0.2"
-VectorizationBase = "0.16, 0.17, 0.18, 0.19"
+VectorizationBase = "0.19"
 julia = "1.5"
 
 [extras]

src/benchconfig.jl

@@ -3,10 +3,27 @@
 function tmul_threads!(C, A, B)
     @tullio C[m,n] = A[m,k] * B[k,n]
 end
-
 function tmul_no_threads!(C, A, B)
     @tullio C[m,n] = A[m,k] * B[k,n] threads=false
 end
+function lvmul_threads!(C, A, B)
+    @avxt for n ∈ indices((C,B), 2), m ∈ indices((C,A), 1)
+        Cmn = zero(eltype(C))
+        for k ∈ indices((A,B), (2,1))
+            Cmn += A[m,k] * B[k,n]
+        end
+        C[m,n] = Cmn
+    end
+end
+function lvmul_no_threads!(C, A, B)
+    @avx for n ∈ indices((C,B), 2), m ∈ indices((C,A), 1)
+        Cmn = zero(eltype(C))
+        for k ∈ indices((A,B), (2,1))
+            Cmn += A[m,k] * B[k,n]
+        end
+        C[m,n] = Cmn
+    end
+end
 
 function generic_matmul!(C, A, B)
     istransposed(C) === 'N' || (generic_matmul!(untransposed(C), _transpose(B), _transpose(A)); return C)
@@ -17,6 +34,8 @@ function generic_matmul!(C, A, B)
     LinearAlgebra.generic_matmatmul!(C, transA, transB, pA, pB)
 end
 
+
+
 function getfuncs(libs::Vector{Symbol}, threaded::Bool)::Vector{Function}
     map(libs) do i
         if i === :MKL
@@ -31,6 +50,8 @@ function getfuncs(libs::Vector{Symbol}, threaded::Bool)::Vector{Function}
             threaded ? tmul_threads! : tmul_no_threads!
         elseif i === :Gaius
             threaded ? Gaius.mul! : Gaius.mul_serial!
+        elseif i === :LoopVectorization
+            threaded ? lvmul_threads! : lvmul_no_threads!
         elseif i === :generic || i === :Generic || i === :GENERIC
             generic_matmul!
         else

src/ccallblas.jl

@@ -15,7 +15,7 @@ _transpose(A::Transpose) = transpose(A)
 for (name,typ,suff) ∈ [
     ("mkl", :Int32, ""),
     ("openblas", :Int64, "_64_"),
-    ("blis", :Int32, "_")
+    ("blis", :Int64, "_64_")
 ]
     uname = uppercase(name)
     lib = Symbol("lib", uname)

src/plotting.jl

@@ -3,7 +3,7 @@
 ####################################### Colors #####################################################
 ####################################################################################################
 
-const LIBRARIES = [:Octavian, :MKL, :OpenBLAS, :blis, :Tullio, :Gaius, :Generic];
+const LIBRARIES = [:Octavian, :MKL, :OpenBLAS, :blis, :Tullio, :Gaius, :LoopVectorization, :Generic];
 """
 Defines the mapping between libraries and colors
 """# #0071c5 == Intel Blue
@@ -27,14 +27,14 @@ isjulialib(x) = x ∈ JULIA_LIBS
 
 
 function pick_suffix(desc = "")
-    suffix = if VectorizationBase.has_feature("x86_64_avx512f")
+    suffix = if Bool(VectorizationBase.has_feature(Val(:x86_64_avx512f)))
         "AVX512"
-    elseif VectorizationBase.has_feature("x86_64_avx2")
+    elseif Bool(VectorizationBase.has_feature(Val(:x86_64_avx2)))
         "AVX2"
-    elseif VectorizationBase.has_feature("x86_64_avx")
+    elseif Bool(VectorizationBase.has_feature(Val(:x86_64_avx)))
         "AVX"
     else
-        "REGSIZE$(VectorizationBase.register_size())"
+        "REGSIZE$(Int(VectorizationBase.register_size()))"
     end
     if desc != ""
         suffix *= '_' * desc
@@ -90,7 +90,6 @@ end
 """
 function Gadfly.plot(br::BenchmarkResult{T}; kwargs...) where {T}
     _plot(br; kwargs...)
-    return nothing
 end
 roundint(x) = round(Int,x)
 # `_plot` is just like `plot`, except _plot returns the filenames

src/runbenchmark.jl

@@ -4,12 +4,12 @@ struct BenchmarkResult{T,I<:Union{Int,NTuple{3,Int}}}
     gflops::Array{Float64,3}
     times::Array{Float64,3}
     threaded::Bool
-    function BenchmarkResult{T}(libraries, sizes, gflops, times, threaded) where {T}
-        gflopsperm = permutedims(gflops, (2,3,1))
-        timesperm = permutedims(times, (2,3,1))
-        I = eltype(sizes)
-        new{T,I}(libraries, convert(Vector{I},sizes), gflopsperm, timesperm, threaded)
-    end
+end
+function BenchmarkResult{T}(libraries, sizes, gflops, times, threaded) where {T}
+    gflopsperm = permutedims(gflops, (2,3,1))
+    timesperm = permutedims(times, (2,3,1))
+    I = eltype(sizes)
+    BenchmarkResult{T,I}(libraries, convert(Vector{I},sizes), gflopsperm, timesperm, threaded)
 end
 
 """
@@ -74,7 +74,6 @@ end
 
 function benchmark_fun!(
     f!::F,
-    summarystat,
     C,
     A,
     B,
@@ -203,6 +202,7 @@ function all_libs()
         :Octavian,
         :OpenBLAS,
         :Tullio,
+        :LoopVectorization
     ]
     return libs
 end
@@ -258,8 +258,7 @@ function runbench(
     threaded::Bool = Threads.nthreads() > 1,
     A_transform = identity,
     B_transform = identity,
-    sleep_time = 0.0,
-    summarystat = minimum
+    sleep_time = 0.0
 ) where {T}
     if threaded
         mkl_set_num_threads(num_cores())
@@ -303,7 +302,6 @@ function runbench(
             comment = "lib=$(lib), M=$(M), K=$(K), N=$(N)"
             t = benchmark_fun!(
                 funcs[i],
-                summarystat,
                 C,
                 A,
                 B,