From 5be8ccd0a9c231db760020d43dce22b4707e0841 Mon Sep 17 00:00:00 2001
From: Andrzej Janik <vosen@vosen.pl>
Date: Tue, 16 Sep 2025 21:28:50 +0000
Subject: [PATCH] Make fake ptxas just copy files, in case of an ELF binary
 just load empty module

---
 ptxas/src/main.rs        | 58 +++++++++++++++-------------------------
 zluda/src/impl/module.rs | 11 +++++++-
 2 files changed, 31 insertions(+), 38 deletions(-)

diff --git a/ptxas/src/main.rs b/ptxas/src/main.rs
index 0ffe841..f8a52aa 100644
--- a/ptxas/src/main.rs
+++ b/ptxas/src/main.rs
@@ -1,6 +1,4 @@
-use bpaf::{any, doc::Style, Bpaf, Parser};
-use hip_runtime_sys::{hipDeviceProp_tR0600, hipGetDevicePropertiesR0600};
-use std::{ffi::CStr, mem};
+use bpaf::{any, choice, doc::Style, literal, Bpaf, Parser};
 
 #[derive(Debug, Clone, Bpaf)]
 #[allow(dead_code)]
@@ -12,6 +10,8 @@ pub struct Options {
     #[bpaf(short, long)]
     verbose: bool,
     #[bpaf(external)]
+    lineinfo: bool,
+    #[bpaf(external)]
     gpu_name: String,
     #[bpaf(long, short('O'), fallback(3))]
     opt_level: usize,
@@ -19,48 +19,32 @@ pub struct Options {
     input: String,
 }
 
+fn lineinfo() -> impl Parser<bool> {
+    choice(["-lineinfo", "--lineinfo"].into_iter().map(|s| {
+        literal(s)
+            .anywhere()
+            .optional()
+            .map(|_| true)
+            .fallback(false)
+            .boxed()
+    }))
+}
+
 // #[bpaf(long, long("gpu_name"), fallback_with(default_arch))]
 fn gpu_name() -> impl Parser<String> {
     any("", move |s: String| {
-        Some(s.strip_prefix("-arch=")?.to_owned())
+        Some(
+            s.strip_prefix("-arch=")
+                .or_else(|| s.strip_prefix("--gpu-name="))?
+                .to_owned(),
+        )
     })
-    .metavar(&[("-arch=", Style::Literal), ("ARG", Style::Metavar)])
+    .metavar(&[("--gpu-name=", Style::Literal), ("SM", Style::Metavar)])
     .anywhere()
     .fallback_with(|| Ok::<String, &'static str>("sm_52".to_string()))
 }
 
 fn main() {
     let options = options().run();
-    let comgr = comgr::Comgr::new().unwrap();
-    unsafe { hip_runtime_sys::hipInit(0) }.unwrap();
-    let mut dev_props: hipDeviceProp_tR0600 = unsafe { mem::zeroed() };
-    let (gpu_arch, clock_rate) = get_gpu_arch_and_clock_rate(&mut dev_props);
-    let input = std::fs::read_to_string(options.input).unwrap();
-    let ast = ptx_parser::parse_module_checked(&input).unwrap();
-    let llvm = ptx::to_llvm_module(
-        ast,
-        ptx::Attributes {
-            clock_rate: clock_rate as u32,
-        },
-        |_| {},
-    )
-    .unwrap();
-    let elf_binary = comgr::compile_bitcode(
-        &comgr,
-        gpu_arch,
-        &*llvm.llvm_ir.write_bitcode_to_memory(),
-        &*llvm.linked_bitcode(),
-        &*llvm.attributes_ir.write_bitcode_to_memory(),
-        None,
-    )
-    .unwrap();
-    std::fs::write(options.output, elf_binary).unwrap();
-}
-
-fn get_gpu_arch_and_clock_rate<'a>(dev_props: &'a mut hipDeviceProp_tR0600) -> (&'a str, i32) {
-    unsafe { hipGetDevicePropertiesR0600(dev_props, 0) }.unwrap();
-    let gcn_arch_name = &dev_props.gcnArchName;
-    let gcn_arch_name = unsafe { CStr::from_ptr(gcn_arch_name.as_ptr()) };
-    let gcn_arch_name = gcn_arch_name.to_str();
-    (gcn_arch_name.unwrap(), dev_props.clockRate)
+    std::fs::copy(&options.input, &options.output).unwrap();
 }
diff --git a/zluda/src/impl/module.rs b/zluda/src/impl/module.rs
index 669fe1f..da7c145 100644
--- a/zluda/src/impl/module.rs
+++ b/zluda/src/impl/module.rs
@@ -20,6 +20,10 @@ impl ZludaObject for Module {
     }
 }
 
+static EMPTY_PTX: &str = ".version 6.5
+.target sm_30
+.address_size 64";
+
 // get_ptx takes an `image` that can be anything we support and returns a
 // String containing a ptx extracted from `image`.
 fn get_ptx<'a>(image: CodeLibraryRef<'a>) -> Result<Cow<'a, str>, CUerror> {
@@ -58,7 +62,12 @@ fn cow_bytes_to_str<'a>(data: Cow<'a, [u8]>) -> Option<Cow<'a, str>> {
 
 pub(crate) fn load_hip_module(library: CodeLibraryRef) -> Result<hipModule_t, CUerror> {
     let global_state = driver::global_state()?;
-    let text = get_ptx(library)?;
+    let maybe_ptx = get_ptx(library);
+    let text = if cfg!(debug_assertions) {
+        maybe_ptx?
+    } else {
+        maybe_ptx.unwrap_or_else(|_| Cow::Borrowed(EMPTY_PTX))
+    };
     let hip_properties = get_hip_properties()?;
     let gcn_arch = get_gcn_arch(&hip_properties)?;
     let attributes = ExtraCacheAttributes {