Add fake ptxas binary

Cargo.lock

@@ -2649,6 +2649,17 @@ dependencies = [
  "syn 2.0.89",
 ]
 
+[[package]]
+name = "ptxas"
+version = "0.0.0"
+dependencies = [
+ "bpaf",
+ "comgr",
+ "hip_runtime-sys",
+ "ptx",
+ "ptx_parser",
+]
+
 [[package]]
 name = "quick-error"
 version = "1.2.3"

Cargo.toml

@@ -16,6 +16,7 @@ members = [
     "ptx_parser",
     "ptx_parser_macros",
     "ptx_parser_macros_impl",
+    "ptxas",
     "xtask",
     "zluda",
     "zluda_bindgen",

compiler/src/main.rs

@@ -10,6 +10,7 @@ use bpaf::Bpaf;
 
 mod error;
 use error::CompilerError;
+use hip_runtime_sys::{hipDeviceProp_tR0600, hipGetDevicePropertiesR0600, hipInit};
 
 const DEFAULT_ARCH: &'static str = "gfx1100";
 
@@ -58,9 +59,14 @@ fn main_core() -> Result<(), CompilerError> {
 
     let arch: String = match opts.arch {
         Some(s) => s,
-        None => get_gpu_arch()
+        None => {
+            unsafe { hipInit(0) }?;
+            let mut dev_props: hipDeviceProp_tR0600 = unsafe { mem::zeroed() };
+            unsafe { hipGetDevicePropertiesR0600(&mut dev_props, 0) }?;
+            get_gpu_arch(&mut dev_props)
                 .map(String::from)
-            .unwrap_or(DEFAULT_ARCH.to_owned()),
+                .unwrap_or(DEFAULT_ARCH.to_owned())
+        }
     };
 
     let ptx = fs::read(&ptx_path).map_err(CompilerError::from)?;
@@ -78,8 +84,8 @@ fn main_core() -> Result<(), CompilerError> {
         &comgr,
         &arch,
         &llvm.bitcode,
-        &llvm.attributes_bitcode,
         &llvm.linked_bitcode,
+        &llvm.attributes_bitcode,
         Some(&comgr_hook),
     )
     .map_err(CompilerError::from)?;
@@ -116,11 +122,8 @@ struct LLVMArtifacts {
     llvm_ir: Vec<u8>,
 }
 
-fn get_gpu_arch() -> Result<&'static str, CompilerError> {
+fn get_gpu_arch<'a>(dev_props: &'a mut hipDeviceProp_tR0600) -> Result<&'a str, CompilerError> {
-    use hip_runtime_sys::*;
+    unsafe { hipGetDevicePropertiesR0600(dev_props, 0) }?;
-    unsafe { hipInit(0) }?;
-    let mut dev_props: hipDeviceProp_tR0600 = unsafe { mem::zeroed() };
-    unsafe { hipGetDevicePropertiesR0600(&mut dev_props, 0) }?;
     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();

ptxas/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "ptxas"
+version = "0.0.0"
+authors = ["Andrzej Janik <vosen@vosen.pl>"]
+edition = "2021"
+
+[[bin]]
+name = "ptxas"
+path = "src/main.rs"
+
+[dependencies]
+comgr = { path = "../comgr" }
+ptx = { path = "../ptx" }
+ptx_parser = { path = "../ptx_parser" }
+hip_runtime-sys = { path = "../ext/hip_runtime-sys" }
+bpaf = { version = "0.9.19", features = ["derive"] }
+
+[package.metadata.zluda]

ptxas/src/main.rs

@@ -0,0 +1,65 @@
+use bpaf::{any, doc::Style, Bpaf, Parser};
+use hip_runtime_sys::{hipDeviceProp_tR0600, hipGetDevicePropertiesR0600};
+use std::{ffi::CStr, mem};
+
+#[derive(Debug, Clone, Bpaf)]
+#[allow(dead_code)]
+#[bpaf(options, version("V12.8.0"))]
+pub struct Options {
+    #[bpaf(short, long)]
+    output: String,
+    warn_on_spills: bool,
+    #[bpaf(short, long)]
+    verbose: bool,
+    #[bpaf(external)]
+    gpu_name: String,
+    #[bpaf(long, short('O'), fallback(3))]
+    opt_level: usize,
+    #[bpaf(positional)]
+    input: String,
+}
+
+// #[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())
+    })
+    .metavar(&[("-arch=", Style::Literal), ("ARG", 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)
+}