try to fix an atomic intrinsic problem and add ptxas in builder to validate

Author: brandonros

crates/cuda_builder/src/lib.rs

@@ -366,6 +366,18 @@ impl CudaBuilder {
 
         let path = invoke_rustc(&self)?;
         println!("cargo:warning=Build completed successfully, PTX file generated at: {}", path.display());
+
+        // run ptxas on it to validate
+        // example: ptxas -v -arch=sm_120 /home/brandon/vanity-miner-rs/target/release/build/gpu_runner-34ef728ec383c50f/out/kernels.ptx -o /tmp/temp.cubin
+        let ptxas_output = Command::new("ptxas")
+            .arg("-v")
+            .arg(format!("-arch={}", self.arch.to_string()))
+            .arg(&path)
+            .arg("-o")
+            .arg("/tmp/temp.cubin")
+            .output()
+            .expect("Failed to run ptxas");
+        println!("cargo:warning=PTXAS output: {:?}", ptxas_output);
 
         if let Some(copy_path) = &self.ptx_file_copy_path {
             println!("cargo:warning=Copying PTX file from {} to {}", path.display(), copy_path.display());

crates/cuda_std/src/atomic/intrinsics.rs

@@ -70,7 +70,7 @@ macro_rules! load {
                 pub unsafe fn [<atomic_load_ $ordering _ $width _ $scope>](ptr: *const [<u $width>]) -> [<u $width>] {
                     let mut out;
                     asm!(
-                        concat!("ld.", stringify!($ordering), load_scope!($ordering, $scope), ".", stringify!([<u $width>]), " {}, [{}];"),
+                        concat!("ld.", stringify!($ordering), load_scope!($ordering, $scope_asm), ".", stringify!([<u $width>]), " {}, [{}];"),
                         out([<reg $width>]) out,
                         in(reg64) ptr
                     );
@@ -116,7 +116,7 @@ macro_rules! store {
                 #[doc = concat!("Performs a ", stringify!($ordering), " atomic store at the ", stringify!($scope), " level with a width of ", stringify!($width), " bits")]
                 pub unsafe fn [<atomic_store_ $ordering _ $width _ $scope>](ptr: *mut [<u $width>], val: [<u $width>]) {
                     asm!(
-                        concat!("st.", stringify!($ordering), load_scope!($ordering, $scope), ".", stringify!([<u $width>]), " [{}], {};"),
+                        concat!("st.", stringify!($ordering), load_scope!($ordering, $scope_asm), ".", stringify!([<u $width>]), " [{}], {};"),
                         in(reg64) ptr,
                         in([<reg $width>]) val,
                     );