Merge bceb269 into fe21ae1

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "Octavian"
 uuid = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
 authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
-version = "0.2.8"
+version = "0.2.9"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
@@ -11,9 +11,9 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
 ArrayInterface = "3"
-LoopVectorization = "0.11"
+LoopVectorization = "0.11.2"
 ThreadingUtilities = "0.2"
-VectorizationBase = "0.17"
+VectorizationBase = "0.18.1"
 julia = "1.5"
 
 [extras]

benchmark/staticarraybench.jl

@@ -63,11 +63,11 @@ rename!(df, matmulmethodnames);
 df.Size = sizerange
 
 function pick_suffix(desc = "")
-    suffix = if Octavian.VectorizationBase.has_feature("x86_64_avx512f")
+    suffix = if Bool(Octavian.VectorizationBase.has_feature(Val(:x86_64_avx512f)))
         "AVX512"
-    elseif Octavian.VectorizationBase.has_feature("x86_64_avx2")
+    elseif Bool(Octavian.VectorizationBase.has_feature(Val(:x86_64_avx2)))
         "AVX2"
-    elseif Octavian.VectorizationBase.has_feature("x86_64_avx")
+    elseif Bool(Octavian.VectorizationBase.has_feature(Val(:x86_64_avx)))
         "AVX"
     else
         "REGSIZE$(Octavian.VectorizationBase.register_size())"

src/Octavian.jl

@@ -3,11 +3,12 @@ module Octavian
 using VectorizationBase, ArrayInterface, LoopVectorization
 
 using VectorizationBase: align, AbstractStridedPointer, zstridedpointer,
-    static_sizeof, lazymul, StridedPointer, gesp, pause, pick_vector_width_val,
-    snum_cache_levels, scache_size, snum_cores, num_cores, cache_inclusivity, scacheline_size
+    static_sizeof, lazymul, StridedPointer, gesp, pause, pick_vector_width, has_feature,
+    num_cache_levels, cache_size, num_cores, num_cores, cache_inclusive, cache_linesize, ifelse
 using LoopVectorization: maybestaticsize, matmul_params, preserve_buffer, CloseOpen
 using ArrayInterface: StaticInt, Zero, One, OptionallyStaticUnitRange, size, strides, offsets, indices,
-    static_length, static_first, static_last, axes, dense_dims, stride_rank
+    static_length, static_first, static_last, axes, dense_dims, stride_rank,
+    StaticBool, True, False, gt, eq
 
 using ThreadingUtilities:
     _atomic_add!, _atomic_umax!, _atomic_umin!,

src/block_sizes.jl

@@ -1,7 +1,7 @@
 
 
 function block_sizes(::Type{T}, _α, _β, R₁, R₂) where {T}
-    W = pick_vector_width_val(T)
+    W = pick_vector_width(T)
     α = _α * W
     β = _β * W
     L₁ₑ = first_cache_size(T) * R₁
@@ -158,7 +158,7 @@ Note that for synchronization on `B`, all threads must have the same values for
 independently of `M`, this algorithm guarantees all threads are on the same page.
 """
 @inline function solve_block_sizes(::Type{T}, M, K, N, _α, _β, R₂, R₃, Wfactor) where {T}
-    W = pick_vector_width_val(T)
+    W = pick_vector_width(T)
     α = _α * W
     β = _β * W
     L₁ₑ =  first_cache_size(T) * R₂
@@ -177,7 +177,7 @@ independently of `M`, this algorithm guarantees all threads are on the same page
 end
 # Takes Nc, calcs Mc and Kc
 @inline function solve_McKc(::Type{T}, M, K, Nc, _α, _β, R₂, R₃, Wfactor) where {T}
-    W = pick_vector_width_val(T)
+    W = pick_vector_width(T)
     α = _α * W
     β = _β * W
     L₁ₑ =  first_cache_size(T) * R₂

src/global_constants.jl

@@ -2,7 +2,7 @@
 const OCTAVIAN_NUM_TASKS = Ref(1)
 _nthreads() = OCTAVIAN_NUM_TASKS[]
 
-@generated function calc_factors(::Union{Val{nc},StaticInt{nc}} = snum_cores()) where {nc}
+@generated function calc_factors(::Union{Val{nc},StaticInt{nc}} = num_cores()) where {nc}
     t = Expr(:tuple)
     for i ∈ nc:-1:1
         d, r = divrem(nc, i)
@@ -12,87 +12,58 @@ _nthreads() = OCTAVIAN_NUM_TASKS[]
 end
 # const CORE_FACTORS = calc_factors()
 
-@generated function MᵣW_mul_factor()
-    f = VectorizationBase.has_feature("x86_64_avx512f") ? 4 : 9
-    Expr(:call, Expr(:curly, :StaticInt, f))
-end
+MᵣW_mul_factor(::True) = StaticInt{4}()
+MᵣW_mul_factor(::False) = StaticInt{9}()
+MᵣW_mul_factor() = MᵣW_mul_factor(has_feature(Val(:x86_64_avx512f)))
 
-@generated function W₁Default()
-    w = if VectorizationBase.has_feature("x86_64_avx512f")
-        0.006089395198610773
-    elseif (Sys.CPU_NAME === "znver2") || (Sys.CPU_NAME === "znver3") # these are znver2 values, I'm assuming they're better for znver3 than generic
-        0.1
-    elseif Sys.CPU_NAME === "znver1"
-        0.053918949422353986
-    else
-        0.1
-    end
-    Expr(:call, Expr(:curly, :StaticFloat, w))
-end
-@generated function W₂Default()
-    w = if VectorizationBase.has_feature("x86_64_avx512f")
-        0.7979822724696168
-    elseif (Sys.CPU_NAME === "znver2") || (Sys.CPU_NAME === "znver3") # these are znver2 values, I'm assuming they're better for znver3 than generic
-        0.993489411720157
-    elseif Sys.CPU_NAME === "znver1"
-        0.3013238122374886
-    else
-        0.15989396641218157
-    end
-    Expr(:call, Expr(:curly, :StaticFloat, w))
-end
-@generated function R₁Default()
-    w = if VectorizationBase.has_feature("x86_64_avx512f")
-        0.5900561503730485
-    elseif (Sys.CPU_NAME === "znver2") || (Sys.CPU_NAME === "znver3") # these are znver2 values, I'm assuming they're better for znver3 than generic
-        0.6052218809954467
-    elseif Sys.CPU_NAME === "znver1"
-        0.6077103834481342
-    else
-        0.4203583148344484
-    end
-    Expr(:call, Expr(:curly, :StaticFloat, w))
-end
-@generated function R₂Default()
-    w = if VectorizationBase.has_feature("x86_64_avx512f")
-        0.762152930709678
-    elseif (Sys.CPU_NAME === "znver2") || (Sys.CPU_NAME === "znver3") # these are znver2 values, I'm assuming they're better for znver3 than generic
-        0.7594052633561165
-    elseif Sys.CPU_NAME === "znver1"
-        0.8775382433240162
-    else
-        0.6344856142604789
-    end
-    Expr(:call, Expr(:curly, :StaticFloat, w))
-end
+W₁Default(::True) = StaticFloat{0.006089395198610773}()
+W₂Default(::True) = StaticFloat{0.7979822724696168}()
+R₁Default(::True) = StaticFloat{0.5900561503730485}()
+R₂Default(::True) = StaticFloat{0.762152930709678}()
+
+W₁Default_arch(::Val{:znver1}) = StaticFloat{0.053918949422353986}()
+W₂Default_arch(::Val{:znver1}) = StaticFloat{0.3013238122374886}()
+R₁Default_arch(::Val{:znver1}) = StaticFloat{0.6077103834481342}()
+R₂Default_arch(::Val{:znver1}) = StaticFloat{0.8775382433240162}()
+
+W₁Default_arch(::Union{Val{:znver2},Val{:znver3}}) = StaticFloat{0.1}()
+W₂Default_arch(::Union{Val{:znver2},Val{:znver3}}) = StaticFloat{0.993489411720157}()
+R₁Default_arch(::Union{Val{:znver2},Val{:znver3}}) = StaticFloat{0.6052218809954467}()
+R₂Default_arch(::Union{Val{:znver2},Val{:znver3}}) = StaticFloat{0.7594052633561165}()
+
+W₁Default_arch(_) = StaticFloat{0.1}()
+W₂Default_arch(_) = StaticFloat{0.15989396641218157}()
+R₁Default_arch(_) = StaticFloat{0.4203583148344484}()
+R₂Default_arch(_) = StaticFloat{0.8775382433240162}()
+
+W₁Default(::False) = W₁Default_arch(VectorizationBase.cpu_name())
+W₂Default(::False) = W₂Default_arch(VectorizationBase.cpu_name())
+R₁Default(::False) = R₁Default_arch(VectorizationBase.cpu_name())
+R₂Default(::False) = R₂Default_arch(VectorizationBase.cpu_name())
+
+W₁Default() = W₁Default(has_feature(Val(:x86_64_avx512f)))
+W₂Default() = W₂Default(has_feature(Val(:x86_64_avx512f)))
+R₁Default() = R₁Default(has_feature(Val(:x86_64_avx512f)))
+R₂Default() = R₂Default(has_feature(Val(:x86_64_avx512f)))
+
+
+
+
+first_cache() = ifelse(gt(num_cache_levels(), StaticInt{2}()), StaticInt{2}(), StaticInt{1}())
+second_cache() = first_cache() + One()
+
+_first_cache_size(fcs::StaticInt) = ifelse(eq(first_cache(), StaticInt(2)) & cache_inclusive(StaticInt(2)), fcs - cache_size(One()), fcs)
+_first_cache_size(::Nothing) = StaticInt(262144)
+first_cache_size() = _first_cache_size(cache_size(first_cache()))
+
+_second_cache_size(scs::StaticInt) = ifelse(cache_inclusive(second_cache()), scs - cache_size(first_cache()), scs)
+_second_cache_size(::Nothing) = StaticInt(3145728)
+second_cache_size() = _second_cache_size(cache_size(second_cache()))
 
-first_cache() = StaticInt{1}() + (snum_cache_levels() > StaticInt{2}() ? One() : Zero())
-second_cache() = StaticInt{2}() + (snum_cache_levels() > StaticInt{2}() ? One() : Zero())
-
-function first_cache_size()
-    fcs = scache_size(first_cache())
-    if fcs === Zero()
-        return StaticInt(262144)
-    elseif (first_cache() === StaticInt(2)) && cache_inclusivity()[2]
-        return fcs - scache_size(One())
-    else
-        return fcs
-    end
-end
-function second_cache_size()
-    scs = scache_size(second_cache())
-    if scs === Zero()
-        return StaticInt(3145728)
-    elseif cache_inclusivity()[second_cache()]
-        return scs - scache_size(first_cache())
-    else
-        return scs
-    end
-end
 first_cache_size(::Type{T}) where {T} = first_cache_size() ÷ static_sizeof(T)
 second_cache_size(::Type{T}) where {T} = second_cache_size() ÷ static_sizeof(T)
 
-bcache_count() = VectorizationBase.scache_count(second_cache())
+bcache_count() = VectorizationBase.num_cache(second_cache())
 
 const BCACHEPTR = Ref{Ptr{Cvoid}}(C_NULL)
 const BCACHE_LOCK = Threads.Atomic{UInt}(zero(UInt))