CompatHelper: bump compat for "Static" to "0.7"

Project.toml

@@ -1,7 +1,7 @@
 name = "Octavian"
 uuid = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
 authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
-version = "0.3.14"
+version = "0.3.15"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
@@ -23,7 +23,7 @@ LoopVectorization = "0.12.86"
 ManualMemory = "0.1.1"
 PolyesterWeave = "0.1.1"
 Requires = "1"
-Static = "0.2, 0.3, 0.4, 0.6"
+Static = "0.2, 0.3, 0.4, 0.6, 0.7"
 ThreadingUtilities = "0.5"
 VectorizationBase = "0.21.15"
 julia = "1.6"

src/Octavian.jl

@@ -17,6 +17,10 @@ using ManualMemory: MemoryBuffer, load, store!
 
 using ThreadingUtilities: _atomic_add!, _atomic_load, _atomic_store!, launch, wait, SPIN
 
+if !(StaticInt <: Base.Integer)
+const Integer = Union{Base.Integer, StaticInt}
+end
+
 export StaticInt
 export matmul!
 export matmul

src/funcptrs.jl

@@ -15,7 +15,7 @@ function (::LoopMulFunc{P,TC,TA,TB,Α,Β,Md,Kd,Nd})(p::Ptr{UInt}) where {P,TC,TA
 end
 @inline _call_loopmul!(C, A, B, α, β, M, K, N, ::Val{false}) = loopmul!(C, A, B, α, β, M, K, N)
 @inline function _call_loopmul!(C::StridedPointer{T}, A, B, α, β, M, K, N, ::Val{true}) where {T}
-  if M*K < first_cache_size(Val(T)) * R₂Default()
+  if M*K < ceil(Int,Float64(first_cache_size(Val(T)) * R₂Default()))
     packaloopmul!(C, A, B, α, β, M, K, N)
     return
   else

src/global_constants.jl

@@ -75,7 +75,5 @@ bcache_count() = VectorizationBase.num_cache(second_cache())
 const BCACHEPTR = Ref{Ptr{Cvoid}}(C_NULL)
 const BCACHE_LOCK = Threads.Atomic{UInt}(zero(UInt))
 
-@static if Sys.WORD_SIZE == 32
-  const ACACHEPTR = Ref{Ptr{Cvoid}}(C_NULL)
-end
+const ACACHEPTR = Ref{Ptr{Cvoid}}(C_NULL)
 

src/init.jl

@@ -24,15 +24,11 @@ function init_bcache()
   nothing
 end
 
-@static if Sys.WORD_SIZE == 32
-  function init_acache()
-    if ACACHEPTR[] == C_NULL
-      ACACHEPTR[] = VectorizationBase.valloc(first_cache_size() * init_num_tasks(), Cvoid, ccall(:jl_getpagesize, Int, ()))
-    end
-    nothing
+function init_acache()
+  if ACACHEPTR[] == C_NULL
+    ACACHEPTR[] = VectorizationBase.valloc(first_cache_size() * init_num_tasks(), Cvoid, ccall(:jl_getpagesize, Int, ()))
   end
-else
-  init_acache() = nothing
+  nothing
 end
 
 function init_num_tasks()

src/memory_buffer.jl

@@ -2,14 +2,8 @@
 @inline function first_cache_buffer(::Val{T}) where {T}
   first_cache_buffer(Val{T}(), first_cache_size(Val(T)))
 end
-@static if Sys.WORD_SIZE == 32
-  @inline function first_cache_buffer(::Val{T}, N) where {T}
-    reinterpret(Ptr{T}, ACACHEPTR[] + ((Threads.threadid()-1) * N) * static_sizeof(T))
-  end
-else
-  @inline function first_cache_buffer(::Val{T}, ::StaticInt{N}) where {T,N}
-    MemoryBuffer{N,T}(undef)
-  end
+@inline function first_cache_buffer(::Val{T}, N) where {T}
+  reinterpret(Ptr{T}, ACACHEPTR[] + ((Threads.threadid()-1) * N) * static_sizeof(T))
 end
 
 BCache(i::Integer) = BCache(BCACHEPTR[]+cld_fast(second_cache_size()*i, Threads.nthreads()), i % UInt)