Remove ComputeCapability and CUDA_ARCH.

CUDA C++ has the `__CUDA_ARCH__` macro for conditional compilation.
rust-cuda has a `CUDA_ARCH` environment variable that is similar, and
the `from_cuda_arch_env` method parses the environment variable's value
to produce a value of type `ComputeCapability`, which can be queried for
conditional compilation.

But `ComputeCapability` has a big problem. It's missing all the
capabilities after 80, including the 'a' and 'f' suffix ones. We could
just add them, but it implements `PartialOrd`/`Ord` and uses ordering to
determine feature availability. This was valid before the 'a' and 'f'
suffixes were added but is no longer, because some pairs of values are
incomparable. E.g. `100a` and `101a` -- each one has some features the
other doesn't, so neither is clearly larger than the other, and they're
also not equal.

So, what to do? Well, `CUDA_ARCH` was added in 2022. More recently,
another mechanism for conditional compilation was added:
`target_feature`, in #239. This does work with the 'a' and 'f' suffix
targets, and it's more Rust-y.

So this commit just removes `CUDA_ARCH` and `ComputeCapability`
(removing two more places where the default compilation target is
specified) and changes the only uses (in `cuda_std/src/atomic/mid.rs`)
to use `target_feature` instead. We don't have any tests exercising
conditional compilation, alas, but I did some manual checking locally to
verify that it works the same.

Author: nnethercote

crates/cuda_builder/src/lib.rs

@@ -809,9 +809,6 @@ fn invoke_rustc(builder: &CudaBuilder) -> Result<PathBuf, CudaBuilderError> {
         }
     }
 
-    let arch = format!("{:?}0", builder.arch);
-    cargo.env("CUDA_ARCH", arch.strip_prefix("Compute").unwrap());
-
     let cargo_encoded_rustflags = join_checking_for_separators(rustflags, "\x1f");
 
     let build = cargo

crates/cuda_std/src/atomic/mid.rs

@@ -7,19 +7,14 @@
 #![allow(dead_code, unused_imports)]
 
 use super::intrinsics;
-use crate::cfg::ComputeCapability;
 use crate::gpu_only;
 use core::sync::atomic::Ordering::{self, *};
 use paste::paste;
 
-fn ge_sm70() -> bool {
-    ComputeCapability::from_cuda_arch_env() >= ComputeCapability::Compute70
-}
-
 #[gpu_only]
 pub fn device_thread_fence(ordering: Ordering) {
     unsafe {
-        if ge_sm70() {
+        if cfg!(target_feature = "compute_70") {
             if ordering == SeqCst {
                 return intrinsics::fence_sc_device();
             }
@@ -38,7 +33,7 @@ pub fn device_thread_fence(ordering: Ordering) {
 #[gpu_only]
 pub fn block_thread_fence(ordering: Ordering) {
     unsafe {
-        if ge_sm70() {
+        if cfg!(target_feature = "compute_70") {
             if ordering == SeqCst {
                 return intrinsics::fence_sc_block();
             }
@@ -57,7 +52,7 @@ pub fn block_thread_fence(ordering: Ordering) {
 #[gpu_only]
 pub fn system_thread_fence(ordering: Ordering) {
     unsafe {
-        if ge_sm70() {
+        if cfg!(target_feature = "compute_70") {
             if ordering == SeqCst {
                 return intrinsics::fence_sc_system();
             }
@@ -80,7 +75,7 @@ macro_rules! load {
                 #[$crate::gpu_only]
                 #[allow(clippy::missing_safety_doc)]
                 pub unsafe fn [<atomic_load_ $width _ $scope>](ptr: *mut $type, ordering: Ordering) -> $type {
-                    if ge_sm70() {
+                    if cfg!(target_feature = "compute_70") {
                         match ordering {
                             SeqCst => {
                                 intrinsics::[<fence_sc_ $scope>]();
@@ -136,7 +131,7 @@ macro_rules! store {
                 #[$crate::gpu_only]
                 #[allow(clippy::missing_safety_doc)]
                 pub unsafe fn [<atomic_store_ $width _ $scope>](ptr: *mut $type, ordering: Ordering, val: $type) {
-                    if ge_sm70() {
+                    if cfg!(target_feature = "compute_70") {
                         match ordering {
                             SeqCst => {
                                 intrinsics::[<fence_sc_ $scope>]();
@@ -185,7 +180,7 @@ macro_rules! inner_fetch_ops_1_param {
                 #[$crate::gpu_only]
                 #[allow(clippy::missing_safety_doc)]
                 pub unsafe fn [<atomic_fetch_ $op _ $type _ $scope>](ptr: *mut $type, ordering: Ordering, val: $type) -> $type {
-                    if ge_sm70() {
+                    if cfg!(target_feature = "compute_70") {
                         match ordering {
                             SeqCst => {
                                 intrinsics::[<fence_sc_ $scope>]();
@@ -259,7 +254,7 @@ macro_rules! inner_cas {
                 #[$crate::gpu_only]
                 #[allow(clippy::missing_safety_doc)]
                 pub unsafe fn [<atomic_compare_and_swap_ $type _ $scope>](ptr: *mut $type, current: $type, new: $type, ordering: Ordering) -> $type {
-                    if ge_sm70() {
+                    if cfg!(target_feature = "compute_70") {
                         match ordering {
                             SeqCst => {
                                 intrinsics::[<fence_sc_ $scope>]();

crates/cuda_std/src/cfg.rs

@@ -43,7 +43,6 @@ pub mod misc;
 // WIP
 // pub mod rt;
 pub mod atomic;
-pub mod cfg;
 pub mod ptr;
 pub mod shared;
 pub mod thread;

crates/cuda_std/src/lib.rs

@@ -17,12 +17,6 @@ static REQUIRED_MAJOR_LLVM_VERSION: u8 = 7;
 
 fn main() {
     rustc_llvm_build();
-
-    // this is set by cuda_builder, but in case somebody is using the codegen
-    // manually, default to 520 (which is what nvvm defaults to).
-    if option_env!("CUDA_ARCH").is_none() {
-        println!("cargo:rustc-env=CUDA_ARCH=520")
-    }
 }
 
 fn fail(s: &str) -> ! {

crates/rustc_codegen_nvvm/build.rs

@@ -241,7 +241,6 @@ fn build_deps(
         .arg("--target-dir")
         .arg(deps_target_dir)
         .env("CARGO_ENCODED_RUSTFLAGS", cargo_encoded_rustflags)
-        .env("CUDA_ARCH", "70")
         .stderr(std::process::Stdio::inherit())
         .stdout(std::process::Stdio::inherit())
         .status()