wip

Author: simeonschaub

src/compiler/codegen.jl

@@ -65,7 +65,19 @@ GPUCompiler.@unlocked function GPUCompiler.mcgen(
     output = tempname(cleanup=false) * (filetype == "asm" ? ".s" : ".o")
     write(input, mod)
 
-    cmd = `$clang_path -x ir $input -mcpu=$(target.dev_isa) --target=$(GPUCompiler.llvm_triple(target)) -nogpulib $(filetype == "asm" ? "-S" : "-c") -o $output`
+    wavefrontsize64 = if hasproperty(job.config.params, :wavefrontsize64)
+        job.config.params.wavefrontsize64
+    else
+        true # Default fallback
+    end
+
+    devlib_paths = get_device_libs_paths(target; wavefrontsize64)
+    devlib_flags = String[]
+    for path in devlib_paths
+        push!(devlib_flags, "-Xclang", "-mlink-builtin-bitcode", "-Xclang", path)
+    end
+
+    cmd = `$clang_path -x ir $input -mcpu=$(target.dev_isa) --target=$(GPUCompiler.llvm_triple(target)) -nogpulib $devlib_flags $(filetype == "asm" ? "-S" : "-c") -o $output`
 
     try
         run(cmd)
@@ -180,7 +192,7 @@ function compiler_config(dev::HIP.HIPDevice;
 
     target = GCNCompilerTarget(; dev_isa, features)
     params = HIPCompilerParams(wavefrontsize64, unsafe_fp_atomics)
-    CompilerConfig(target, params; kernel, name, always_inline=true)
+    CompilerConfig(target, params; kernel, name, always_inline=true, validate=false)
 end
 
 const hipfunction_lock = ReentrantLock()

src/compiler/device_libs.jl

@@ -23,36 +23,28 @@ end
 
 const DEVICE_LIBS::Dict{String, DevLib} = Dict{String, DevLib}()
 
-function link_device_libs!(
-    target::GCNCompilerTarget, mod::LLVM.Module;
+function get_device_libs_paths(
+    target::GCNCompilerTarget;
     wavefrontsize64::Bool,
 )
-    isnothing(libdevice_libs) && return
-    if !isempty(AMDGPU.ROCmDiscovery.clang_path)
-        return
-    end
+    paths = String[]
+    isnothing(libdevice_libs) && return paths
 
     # 1. Load other libraries.
     lib_names = ("hc", "hip", "irif", "ockl", "opencl", "ocml")
     for lib_name in lib_names
-        devlib = get!(DEVICE_LIBS, lib_name) do
-            DevLib(lib_name, locate_lib(lib_name))
-        end
-        load_and_link!(devlib, mod)
+        path = locate_lib(lib_name)
+        !isnothing(path) && push!(paths, path)
     end
 
     # 2. Load OCLC library.
-    devlib = get!(DEVICE_LIBS, "oclc") do
-        isa_short = replace(target.dev_isa, "gfx"=>"")
-        DevLib("oclc", locate_lib("oclc_isa_version_$isa_short"))
-    end
-    load_and_link!(devlib, mod)
+    isa_short = replace(target.dev_isa, "gfx"=>"")
+    path = locate_lib("oclc_isa_version_$isa_short")
+    !isnothing(path) && push!(paths, path)
 
     # 3. Load OCLC ABI library.
-    devlib = get!(DEVICE_LIBS, "oclc_abi") do
-        DevLib("oclc_abi", locate_lib("oclc_abi_version_500"))
-    end
-    load_and_link!(devlib, mod)
+    path = locate_lib("oclc_abi_version_500")
+    !isnothing(path) && push!(paths, path)
 
     # 4. Load options libraries.
     options = (
@@ -65,8 +57,25 @@ function link_device_libs!(
     for (option, value) in options
         toggle = value ? "on" : "off"
         name = "oclc_$(option)_$(toggle)"
+        path = locate_lib(name)
+        !isnothing(path) && push!(paths, path)
+    end
+    return paths
+end
+
+function link_device_libs!(
+    target::GCNCompilerTarget, mod::LLVM.Module;
+    wavefrontsize64::Bool,
+)
+    isnothing(libdevice_libs) && return
+    if !isempty(AMDGPU.ROCmDiscovery.clang_path)
+        return
+    end
+
+    for path in get_device_libs_paths(target; wavefrontsize64)
+        name = basename(path)
         devlib = get!(DEVICE_LIBS, name) do
-            DevLib(name, locate_lib(name))
+            DevLib(name, path)
         end
         load_and_link!(devlib, mod)
     end

test/Project.toml

@@ -1,6 +1,7 @@
 [deps]
 AMDGPU = "21141c5a-9bdb-4563-92ae-f87d6854732e"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+AMDGPU = "0c46a032-3f7e-4b1d-8c5a-2e9f6b1d0f3e"
 BFloat16s = "ab4f0b2a-ad5b-11e8-123f-65d77653426b"
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 FFTW = "7a1cc6ca-52ef-59f5-83cd-3a7055c09341"