Archive/ZLUDA
0
0
Fork 0
mirror of https://github.com/vosen/ZLUDA.git synced 2025-04-28 21:47:57 +03:00
Code Issues Projects Releases Packages Wiki Activity Actions

Simplify compilation, make sure linking ockl works with floating point mode

Browse source
This commit is contained in:
Andrzej Janik 2025-03-16 00:23:01 +00:00
parent 2a8e167341
commit 9b7a685173
1 changed files with 24 additions and 24 deletions
Download patch file Download diff file Expand all files Collapse all files

48
comgr/src/lib.rs
View file

@ -133,21 +133,26 @@ pub fn compile_bitcode(
&linking_info, &linking_info,
amd_comgr_action_kind_t::AMD_COMGR_ACTION_LINK_BC_TO_BC, amd_comgr_action_kind_t::AMD_COMGR_ACTION_LINK_BC_TO_BC,
)?; )?;
let link_with_device_libs_info = ActionInfo::new()?; let compile_to_exec = ActionInfo::new()?;
link_with_device_libs_info.set_isa_name(gcn_arch)?; compile_to_exec.set_isa_name(gcn_arch)?;
link_with_device_libs_info.set_language(amd_comgr_language_t::AMD_COMGR_LANGUAGE_LLVM_IR)?; compile_to_exec.set_language(amd_comgr_language_t::AMD_COMGR_LANGUAGE_LLVM_IR)?;
// This makes no sense, but it makes ockl linking work let common_options = [
link_with_device_libs_info // This makes no sense, but it makes ockl linking work
.set_options([c"-Xclang", c"-mno-link-builtin-bitcode-postopt"].into_iter())?; c"-Xclang",
let with_device_libs = do_action( c"-mno-link-builtin-bitcode-postopt",
&linked_data_set, // Otherwise LLVM omits dynamic fp mode for ockl functions during linking
&link_with_device_libs_info, // and then fails to inline them
amd_comgr_action_kind_t::AMD_COMGR_ACTION_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC, c"-Xclang",
)?; c"-fdenormal-fp-math=dynamic",
let compile_action_info = ActionInfo::new()?; c"-O3",
compile_action_info.set_isa_name(gcn_arch)?; c"-mno-wavefrontsize64",
let common_options = [c"-O3", c"-mno-wavefrontsize64", c"-mcumode"].into_iter(); c"-mcumode",
// Useful for inlining reports, combined with AMD_COMGR_SAVE_TEMPS=1 AMD_COMGR_EMIT_VERBOSE_LOGS=1 AMD_COMGR_REDIRECT_LOGS=stderr
// c"-fsave-optimization-record=yaml",
]
.into_iter();
let opt_options = if cfg!(debug_assertions) { let opt_options = if cfg!(debug_assertions) {
//[c"-g", c"-mllvm", c"-print-before-all", c"", c""]
[c"-g", c"", c"", c"", c""] [c"-g", c"", c"", c"", c""]
} else { } else {
[ [
@ -159,19 +164,14 @@ pub fn compile_bitcode(
c"-inlinehint-threshold=3250", c"-inlinehint-threshold=3250",
] ]
}; };
compile_action_info.set_options(common_options.chain(opt_options))?; compile_to_exec.set_options(common_options.chain(opt_options))?;
let reloc_data_set = do_action(
&with_device_libs,
&compile_action_info,
amd_comgr_action_kind_t::AMD_COMGR_ACTION_CODEGEN_BC_TO_RELOCATABLE,
)?;
let exec_data_set = do_action( let exec_data_set = do_action(
&reloc_data_set, &linked_data_set,
&compile_action_info, &compile_to_exec,
amd_comgr_action_kind_t::AMD_COMGR_ACTION_LINK_RELOCATABLE_TO_EXECUTABLE, amd_comgr_action_kind_t::AMD_COMGR_ACTION_COMPILE_SOURCE_TO_EXECUTABLE,
)?; )?;
let executable = let executable =
exec_data_set.get_data(amd_comgr_data_kind_t::AMD_COMGR_DATA_KIND_EXECUTABLE, 0)?; exec_data_set.get_data(amd_comgr_data_kind_t::AMD_COMGR_DATA_KIND_EXECUTABLE, 0)?;
executable.copy_content() executable.copy_content()
} }