Feature/trim compatible

Project.toml

@@ -1,7 +1,7 @@
 name = "CPUSummary"
 uuid = "2a0fbf3d-bb9c-48f3-b0a9-814d99fd7ab9"
 authors = ["chriselrod <[email protected]> and contributors"]
-version = "0.2.6"
+version = "0.2.7"
 
 [deps]
 CpuId = "adafc99b-e345-5852-983c-f28acb93d879"

src/CPUSummary.jl

@@ -1,7 +1,6 @@
 module CPUSummary
-if isdefined(Base, :Experimental) &&
-   isdefined(Base.Experimental, Symbol("@max_methods"))
-    @eval Base.Experimental.@max_methods 1
+if isdefined(Base, :Experimental) && isdefined(Base.Experimental, Symbol("@max_methods"))
+  @eval Base.Experimental.@max_methods 1
 end
 
 using Static
@@ -52,22 +51,22 @@ if (Sys.ARCH === :x86_64)
 else
   include("generic_topology.jl")
 end
-function __init__()
-  ccall(:jl_generating_output, Cint, ()) == 1 && return
-  nc = _get_num_cores()
-  syst = Sys.CPU_THREADS::Int
-  if nc != num_l1cache()
-    @eval num_l1cache() = static($nc)
-  end
-  if nc != num_cores()
-    @eval num_cores() = static($nc)
-  end
-  if syst != sys_threads()
-    @eval sys_threads() = static($syst)
-  end
-  _extra_init()
-  return nothing
-end
+# function __init__()
+#   ccall(:jl_generating_output, Cint, ()) == 1 && return
+#   nc = _get_num_cores()
+#   syst = Sys.CPU_THREADS::Int
+#   if nc != num_l1cache()
+#     @eval num_l1cache() = static($nc)
+#   end
+#   if nc != num_cores()
+#     @eval num_cores() = static($nc)
+#   end
+#   if syst != sys_threads()
+#     @eval sys_threads() = static($syst)
+#   end
+#   # _extra_init()
+#   return nothing
+# end
 
 
 # end
@@ -80,7 +79,7 @@ cache_linesize() = cache_linesize(Val(1))
 function num_cache_levels()
   numl4 = num_l4cache()
   numl4 === nothing && return nothing
-  ifelse(
+  return ifelse(
     eq(numl4, Zero()),
     ifelse(
       eq(num_l3cache(), Zero()),
@@ -96,8 +95,8 @@ function num_cache_levels()
 end
 
 # explicit precompilation only on Julia v1.9 and newer
-if VERSION >= v"1.9"
-  include("precompile.jl")
-end
+# if VERSION >= v"1.9"
+#     include("precompile.jl")
+# end
 
 end

src/generic_topology.jl

@@ -35,14 +35,16 @@ end
 cache_associativity(::Union{Val{2},StaticInt{2}}) = StaticInt{0}()
 cache_type(::Union{Val{2},StaticInt{2}}) = Val{:Unified}()
 cache_inclusive(_) = False()
-@static if Sys.ARCH === :aarch64 && (Sys.isapple() || occursin("apple", Sys.CPU_NAME::String))
+@static if Sys.ARCH === :aarch64 &&
+           (Sys.isapple() || occursin("apple", Sys.CPU_NAME::String))
   cache_linesize(_) = StaticInt{128}() # assume...
 else
   cache_linesize(_) = StaticInt{64}() # assume...
 end
 cache_size(_) = StaticInt{0}()
 
-@static if Sys.ARCH === :aarch64 && (Sys.isapple() || occursin("apple", Sys.CPU_NAME::String))
+@static if Sys.ARCH === :aarch64 &&
+           (Sys.isapple() || occursin("apple", Sys.CPU_NAME::String))
 else
   cache_type(::Union{Val{3},StaticInt{3}}) = Val{:Unified}()
   cache_size(::Union{Val{3},StaticInt{3}}) = StaticInt{1441792}()

src/x86.jl

@@ -20,19 +20,76 @@ const PrecompiledCacheSize = let cs = CpuId.cachesize()
   ntuple(i -> i == 3 ? cs[3] ÷ _get_num_cores() : cs[i], length(cs))
 end
 const PrecompiledCacheInclusive = CpuId.cacheinclusive()
-cache_inclusive(_) = False()
-@noinline function _eval_cache_size(cachesize)
-  for (i, csi) in enumerate(cachesize)
-    @eval cache_size(::Union{Val{$i},StaticInt{$i}}) = $(static(csi))
+# cache_inclusive(_) = False()
+# @noinline function _eval_cache_size(cachesize)
+#   for (i, csi) in enumerate(cachesize)
+#     @eval cache_size(::Union{Val{$i},StaticInt{$i}}) = $(static(csi))
+#   end
+# end
+# @noinline function _eval_cache_inclusive(cacheinclusive)
+#   for (i, cii) in enumerate(cacheinclusive)
+#     @eval cache_inclusive(::Union{Val{$i},StaticInt{$i}}) = $(static(cii != 0))
+#   end
+# end
+# _eval_cache_size(PrecompiledCacheSize)
+# _eval_cache_inclusive(PrecompiledCacheInclusive)
+
+
+cache_size(::Val{S}) where {S} = cache_size(S)
+cache_size(::StaticInt{S}) where {S} = cache_size(S)
+
+@inline @generated function cache_size(cachesize)
+  cs = let cs = CpuId.cachesize()
+    ntuple(i -> i == 3 ? cs[3] ÷ _get_num_cores() : cs[i], length(cs))
+  end
+
+  cache_sizes = map(enumerate(cs)) do (i, csi)
+
+    return :(
+      if cachesize == $i
+        return static($csi)
+      end
+    )
+  end
+
+  return quote
+    begin
+      $(cache_sizes...)
+    end
   end
+
 end
-@noinline function _eval_cache_inclusive(cacheinclusive)
-  for (i, cii) in enumerate(cacheinclusive)
-    @eval cache_inclusive(::Union{Val{$i},StaticInt{$i}}) = $(static(cii != 0))
+cache_inclusive(::Val{S}) where {S} = cache_inclusive(S)
+cache_inclusive(::StaticInt{S}) where {S} = cache_inclusive(S)
+
+@inline @generated function cache_inclusive(cacheinclusive)
+  ci = CpuId.cacheinclusive()
+
+  cache_inclusives = map(enumerate(ci)) do (i, cii)
+    val = cii != 0
+    return :(
+      if cacheinclusive == $i
+        return static($val)
+      end
+    )
+  end
+
+  if !isempty(cache_inclusives)
+    push!(cache_inclusives, :(return False()))
+  else
+    cache_inclusives = [:(return False())]
+  end
+
+  return quote
+    begin
+      $(cache_inclusives...)
+    end
   end
+
+
 end
-_eval_cache_size(PrecompiledCacheSize)
-_eval_cache_inclusive(PrecompiledCacheInclusive)
+
+
 # TODO: implement
 cache_associativity(_) = static(0)
 
@@ -43,7 +100,7 @@ cache_type(_) = Val{:Unified}()
 let lnsize = static(CpuId.cachelinesize())
   global cache_linesize(_) = lnsize
 end
-cache_size(_) = StaticInt{0}()
+# cache_size(_) = StaticInt{0}()
 
 # cache_size(::Union{Val{3},StaticInt{3}}) = num_cores() * StaticInt{1441792}()
 function _extra_init()
@@ -55,5 +112,3 @@ function _extra_init()
   ci !== PrecompiledCacheInclusive && _eval_cache_inclusive(ci)
   return nothing
 end
-
-