makegeneric

Project.toml

@@ -1,7 +1,7 @@
 name = "HostCPUFeatures"
 uuid = "3e5b6fbb-0976-4d2c-9146-d79de83f2fb0"
 authors = ["Chris Elrod <[email protected]> and contributors"]
-version = "0.1.10"
+version = "0.1.11"
 
 [deps]
 BitTwiddlingConvenienceFunctions = "62783981-4cbd-42fc-bca8-16325de8dc4b"

src/HostCPUFeatures.jl

@@ -40,6 +40,10 @@ unwrap(::StaticSymbol{S}) where {S} = S
 end
 const BASELINE_CPU_NAME = get_cpu_name()
 function __init__()
+  if Sys.ARCH === :x86_64
+    target = Base.unsafe_string(Base.JLOptions().cpu_target)
+    occursin("native",  target) || return make_generic(target)
+  end
   ccall(:jl_generating_output, Cint, ()) == 1 && return
   BASELINE_CPU_NAME == Sys.CPU_NAME::String || redefine()
   return nothing

src/cpu_info_x86.jl

@@ -31,3 +31,79 @@ fast_int64_to_double() = has_feature(Val(:x86_64_avx512dq))
 
 fast_half() = False()
 
+@noinline function setfeaturefalse(s)
+  if has_feature(Val(s)) === True()
+    @eval has_feature(::Val{:x86_64_avx512ifma}) = False()
+  end
+end
+@noinline function setfeaturetrue(s)
+  if has_feature(Val(s)) === False()
+    @eval has_feature(::Val{:x86_64_avx512ifma}) = True()
+  end
+end
+
+function make_generic(target)
+  if occursin("tigerlake", target) || occursin("znver4", target)
+    # most feature-complete architectures we use
+    setfeaturetrue(:x86_64_avx512ifma)
+    setfeaturetrue(:x86_64_avx512vl)
+    setfeaturetrue(:x86_64_avx512bw)
+    setfeaturetrue(:x86_64_avx512dq)
+    setfeaturetrue(:x86_64_avx512f)
+    setfeaturetrue(:x86_64_avx2)
+    setfeaturetrue(:x86_64_bmi2)
+    setfeaturetrue(:x86_64_fma)
+    setfeaturetrue(:x86_64_avx)
+  elseif occursin("icelake", target) || occursin("skylake-avx512", target) || occursin("rocketlake", target) || occursin("cascadelake", target)
+    # no ifma, but avx512f and avx512dq
+    setfeaturefalse(:x86_64_avx512ifma)
+    setfeaturetrue(:x86_64_avx512vl)
+    setfeaturetrue(:x86_64_avx512bw)
+    setfeaturetrue(:x86_64_avx512dq)
+    setfeaturetrue(:x86_64_avx512f)
+    setfeaturetrue(:x86_64_avx2)
+    setfeaturetrue(:x86_64_bmi2)
+    setfeaturetrue(:x86_64_fma)
+    setfeaturetrue(:x86_64_avx)
+    return
+  elseif occursin("znver", target) || occursin("lake", target) || occursin("well", target)
+    # no avx512, but avx2, fma, and bmi2
+    # znver tries to capture all zen < 4
+    # lake tries to capture lakes we didn't single out above as having avx512
+    # 
+    setfeaturefalse(:x86_64_avx512ifma)
+    setfeaturefalse(:x86_64_avx512vl)
+    setfeaturefalse(:x86_64_avx512bw)
+    setfeaturefalse(:x86_64_avx512dq)
+    setfeaturefalse(:x86_64_avx512f)
+    setfeaturetrue(:x86_64_avx2)
+    setfeaturetrue(:x86_64_bmi2)
+    setfeaturetrue(:x86_64_fma)
+    setfeaturetrue(:x86_64_avx)
+    return
+  elseif occursin("ivybridge", target) || occursin("sandybridge", target)
+    # has avx, and that is about it we care about
+    setfeaturefalse(:x86_64_avx512ifma)
+    setfeaturefalse(:x86_64_avx512vl)
+    setfeaturefalse(:x86_64_avx512bw)
+    setfeaturefalse(:x86_64_avx512dq)
+    setfeaturefalse(:x86_64_avx512f)
+    setfeaturefalse(:x86_64_avx2)
+    setfeaturefalse(:x86_64_bmi2)
+    setfeaturefalse(:x86_64_fma)
+    setfeaturetrue(:x86_64_avx)
+  else
+    # hopefully we didn't miss something
+    # TODO: sapphire rapids
+    setfeaturefalse(:x86_64_avx512ifma)
+    setfeaturefalse(:x86_64_avx512vl)
+    setfeaturefalse(:x86_64_avx512bw)
+    setfeaturefalse(:x86_64_avx512dq)
+    setfeaturefalse(:x86_64_avx512f)
+    setfeaturefalse(:x86_64_avx2)
+    setfeaturefalse(:x86_64_bmi2)
+    setfeaturefalse(:x86_64_fma)
+    setfeaturefalse(:x86_64_avx)
+  end
+end
+