From f2b4a0a40999fab4bec3da55f94e83aecf211b1f Mon Sep 17 00:00:00 2001
From: Chris Elrod <elrodc@gmail.com>
Date: Mon, 3 Feb 2020 06:17:39 -0500
Subject: [PATCH] On Linux, it will now look for /usr/lib(64)/libmvec.so. If it
 finds it, it will use it in place of a few select functions.

---
 .gitignore          |  2 ++
 Project.toml        |  1 +
 deps/build.jl       | 68 +++++++++++++++++++++++++++++++++++++++++++++
 src/SLEEFPirates.jl |  1 +
 4 files changed, 72 insertions(+)
 create mode 100644 deps/build.jl

diff --git a/.gitignore b/.gitignore
index e92b2de..13a69f9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -8,3 +8,5 @@ src/#*#
 tests/#*#
 *~
 *.mem
+src/svmlwrap.jl
+deps/build.log
\ No newline at end of file
diff --git a/Project.toml b/Project.toml
index 812a378..dc6a10c 100644
--- a/Project.toml
+++ b/Project.toml
@@ -4,6 +4,7 @@ authors = ["chriselrod <elrodc@gmail.com>"]
 version = "0.3.2"
 
 [deps]
+Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 SIMDPirates = "21efa798-c60a-11e8-04d3-e1a92915a26a"
 VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
diff --git a/deps/build.jl b/deps/build.jl
new file mode 100644
index 0000000..430c144
--- /dev/null
+++ b/deps/build.jl
@@ -0,0 +1,68 @@
+using VectorizationBase: REGISTER_SIZE
+using Libdl
+
+function try_to_find_libmvec()
+    Sys.islinux() || return ""
+    libnames = [
+        "libmvec.so"
+    ]
+    paths = [
+        "/usr/lib64/", "/usr/lib"
+    ]
+    find_library(libnames, paths)
+end
+
+function create_svmlwrap_file(mveclib)
+    mveclib == "" && return mveclib
+    lib = dlopen(mveclib)
+    file = ["const MVECLIB = \"$mveclib\""]
+    sizes = [(:b,16),(:d,32),(:e,64)]
+    for f ∈ [:log, :exp, :pow, :sin, :cos]
+        for (l,s) ∈ sizes
+            s > REGISTER_SIZE && continue
+            for isdouble ∈ (true,false)
+                ts = s >> (2 + isdouble)
+                func = Symbol(:_ZGV, l, :N, ts, :v_, f, isdouble ? Symbol("") : :f)
+                sym = dlsym(lib, func, throw_error = false)
+                if sym != C_NULL
+                    typ = "Float$(isdouble ? 64 : 32)"
+                    vtyp = "NTuple{$ts,Core.VecElement{$typ}}"
+                    svtyp = "SVec{$ts,$typ}"
+                    def1 = "@inline $f(v::$vtyp) = ccall((:$func,MVECLIB), $vtyp, ($vtyp,), v)"
+                    def2 = "@inline $f(v::$svtyp) = SVec($f(extract_data(v)))"
+                    push!(file, def1)
+                    push!(file, def2)
+                end
+            end
+        end
+    end
+    let f = :pow
+        for (l,s) ∈ sizes
+            if s ≤ REGISTER_SIZE
+                for isdouble ∈ (true,false)
+                    ts = s >> (2 + isdouble)
+                    func = Symbol(:_ZGV, l, :N, ts, :vv_, f, isdouble ? Symbol("") : :f)
+                    sym = dlsym(lib, func, throw_error = false)
+                    if sym != C_NULL
+                        typ = "Float$(isdouble ? 64 : 32)"
+                        vtyp = "NTuple{$ts,Core.VecElement{$typ}}"
+                        svtyp = "SVec{$ts,$typ}"
+                        def1 = "@inline $f(v1::$vtyp, v2::$vtyp) = ccall((:$func,MVECLIB), $vtyp, ($vtyp,$vtyp), v1, v2)"
+                        def2 = "@inline $f(v1::$svtyp, v2::$svtyp) = SVec($f(extract_data(v1),extract_data(v2)))"
+                        push!(file, def1)
+                        push!(file, def2)
+                    end
+                end
+            end
+        end        
+    end
+    join(file, "\n")
+end
+
+open(joinpath(@__DIR__, "..", "src", "svmlwrap.jl"), "w") do f
+    write(f, create_svmlwrap_file(try_to_find_libmvec()))
+end
+
+
+
+
diff --git a/src/SLEEFPirates.jl b/src/SLEEFPirates.jl
index 3237fc9..48344d2 100644
--- a/src/SLEEFPirates.jl
+++ b/src/SLEEFPirates.jl
@@ -116,6 +116,7 @@ include("log.jl")    # logarithmic functions
 include("trig.jl")   # trigonometric and inverse trigonometric functions
 include("hyp.jl")    # hyperbolic and inverse hyperbolic functions
 include("misc.jl")   # miscallenous math functions including pow and cbrt
+include("svmlwrap.jl")
 
 # fallback definitions