Precomp for >= 1.8.1beta (#132)

* Adjust blocking behavior to cap cache use at L2e and L3e

* Precompile

* Bump version

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 = "0.3.11"
+version = "0.3.12"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/Octavian.jl

@@ -23,6 +23,8 @@ export matmul
 export matmul_serial!
 export matmul_serial
 
+debug() = false
+
 include("global_constants.jl")
 include("types.jl")
 include("integerdivision.jl")
@@ -36,4 +38,21 @@ include("complex_matmul.jl")
 
 include("init.jl") # `Octavian.__init__()` is defined in this file
 
+@static if VERSION >= v"1.8.0-beta1"
+  let
+    __init__()
+    A64 = rand(100,100)
+    matmul(A64,A64)
+    matmul(A64',A64)
+    matmul(A64,A64')
+    matmul(A64',A64')
+    A32 = rand(Float32,100,100)
+    matmul(A32,A32)
+    matmul(A32',A32)
+    matmul(A32,A32')
+    matmul(A32',A32')
+  end
+end
+
+
 end # module Octavian

src/global_constants.jl

@@ -16,10 +16,10 @@ MᵣW_mul_factor(::True) = StaticInt{4}()
 MᵣW_mul_factor(::False) = StaticInt{9}()
 MᵣW_mul_factor() = MᵣW_mul_factor(has_feature(Val(:x86_64_avx512f)))
 
-W₁Default(::True) = StaticFloat64{0.0007423708195588264}()
-W₂Default(::True) = StaticFloat64{0.7398765624419478}()
-R₁Default(::True) = StaticFloat64{0.4697043382682602}()
-R₂Default(::True) = StaticFloat64{0.6342912896800855}()
+W₁Default(::True) = StaticFloat64{0.0009898277594117685}()
+W₂Default(::True) = StaticFloat64{0.9865020832559304}()
+R₁Default(::True) = StaticFloat64{0.5820044063603483}()
+R₂Default(::True) = StaticFloat64{0.7580885846640107}()
 
 W₁Default_arch(::Val{:znver1}) = StaticFloat64{0.053918949422353986}()
 W₂Default_arch(::Val{:znver1}) = StaticFloat64{0.3013238122374886}()
@@ -73,7 +73,7 @@ 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
+@static if Sys.WORD_SIZE == 32
   const ACACHEPTR = Ref{Ptr{Cvoid}}(C_NULL)
 end
 

src/init.jl

@@ -12,31 +12,35 @@ function __init__()
     )
     @warn msg
   end
-  reseet_bcache_lock!()
+  reset_bcache_lock!()
 end
 
 function init_bcache()
   if bcache_count() ≢ Zero()
-    BCACHEPTR[] = VectorizationBase.valloc(second_cache_size() * bcache_count(), Cvoid, ccall(:jl_getpagesize, Int, ()))
+    if BCACHEPTR[] == C_NULL
+      BCACHEPTR[] = VectorizationBase.valloc(second_cache_size() * bcache_count(), Cvoid, ccall(:jl_getpagesize, Int, ()))
+    end
   end
   nothing
 end
 
-@static if Sys.WORD_SIZE ≤ 32
+@static if Sys.WORD_SIZE == 32
   function init_acache()
-    ACACHEPTR[] = VectorizationBase.valloc(first_cache_size() * init_num_tasks(), Cvoid, ccall(:jl_getpagesize, Int, ()))
+    if ACACHEPTR[] == C_NULL
+      ACACHEPTR[] = VectorizationBase.valloc(first_cache_size() * init_num_tasks(), Cvoid, ccall(:jl_getpagesize, Int, ()))
+    end
     nothing
   end
 else
   init_acache() = nothing
 end
 
 function init_num_tasks()
-  num_tasks = _read_environment_num_tasks()::Int
+  num_tasks = _read_environment_num_tasks()
   OCTAVIAN_NUM_TASKS[] = num_tasks
 end
 
-function _read_environment_num_tasks()
+function _read_environment_num_tasks()::Int
   environment_variable = get(ENV, "OCTAVIAN_NUM_TASKS", "")::String
   nt = min(Threads.nthreads(), VectorizationBase.num_cores())::Int
   if isempty(environment_variable)

src/macrokernels.jl

@@ -132,8 +132,11 @@ function packaloopmul!(
     α, β, M, K, N
 ) where {T}
   Ãₚ, buffer = alloc_a_pack(K, Val(T))
+  Mᵣ, Nᵣ = matmul_params(Val(T))
+  if (debug() && (first_cache_size(Val(T)) < Mᵣ*pick_vector_width(T)*K*cld(M,Mᵣ*pick_vector_width(T))))
+    throw("First cache size too small: M: $M, K: $K, N: $N, first cache size: $(first_cache_size(Val(T)))")
+  end
   GC.@preserve buffer begin
-    Mᵣ, Nᵣ = matmul_params(Val(T))
     packamul!(C, Ãₚ, A, B, α, β, M, K, Nᵣ)
     ploopmul!(gesp(C, (Zero(), Nᵣ)), Ãₚ, gesp(B, (Zero(), Nᵣ)), α, β, M, K, N - Nᵣ)
   end

src/memory_buffer.jl

@@ -24,7 +24,7 @@ function _use_bcache()
   end
   return BCache(nothing)
 end
-@inline _free_bcache!(b::BCache{Nothing}) = reseet_bcache_lock!()
+@inline _free_bcache!(::BCache{Nothing}) = reset_bcache_lock!()
 
 _use_bcache(::Nothing) = _use_bcache()
 function _use_bcache(i)
@@ -42,5 +42,5 @@ _free_bcache!(b::BCache{UInt}) = (Threads.atomic_xor!(BCACHE_LOCK, one(UInt) <<
 Currently not using try/finally in matmul routine, despite locking.
 So if it errors for some reason, you may need to manually call `reset_bcache_lock!()`.
 """
-@inline reseet_bcache_lock!() = (BCACHE_LOCK[] = zero(UInt); nothing)
+@inline reset_bcache_lock!() = (BCACHE_LOCK[] = zero(UInt); nothing)
 

test/aqua.jl

@@ -1,3 +1,4 @@
 @testset "Aqua.jl" begin
-    Aqua.test_all(Octavian)
+  Aqua.test_all(Octavian, ambiguities=false)
+  @test isempty(Test.detect_ambiguities(Octavian))
 end

test/runtests.jl

@@ -18,6 +18,7 @@ include("test_suite_preamble.jl")
 @info("Octavian.OCTAVIAN_NUM_TASKS[] is $(Octavian.OCTAVIAN_NUM_TASKS[]) tasks")
 
 Random.seed!(123)
+Octavian.debug() = true
 
 include("block_sizes.jl")
 include("init.jl")