Optimize atomic float on NVPTX

README.md

@@ -70,7 +70,7 @@ See the [`atomic128` module's readme](https://github.com/taiki-e/portable-atomic
 - <a name="optional-features-float"></a>**`float`**<br>
   Provide `AtomicF{32,64}`.
 
-  Note that most of `fetch_*` operations of atomic floats are implemented using CAS loops, which can be slower than equivalent operations of atomic integers. ([GPU targets have atomic instructions for float, so we plan to use these instructions for GPU targets in the future.](https://github.com/taiki-e/portable-atomic/issues/34))
+  Note that most of `fetch_*` operations of atomic floats are implemented using CAS loops, which can be slower than equivalent operations of atomic integers. (GPU targets have atomic instructions for float, so we use these instructions for GPU targets on nightly.)
 
 - **`std`**<br>
   Use `std`.

build.rs

@@ -162,7 +162,8 @@ fn main() {
         // https://github.com/rust-lang/rust/pull/111331 merged in Rust 1.71 (nightly-2023-05-09).
         if !no_asm
             && (target_arch == "powerpc64" && version.probe(60, 2022, 2, 12)
-                || target_arch == "s390x" && version.probe(71, 2023, 5, 8))
+                || target_arch == "s390x" && version.probe(71, 2023, 5, 8)
+                || target_arch == "nvptx64")
             && is_allowed_feature("asm_experimental_arch")
         {
             println!("cargo:rustc-cfg=portable_atomic_unstable_asm_experimental_arch");
@@ -328,6 +329,36 @@ fn main() {
                 false,
             );
         }
+        "nvptx64" => {
+            let mut has_sm_70 = false;
+            if let Some(rustflags) = env::var_os("CARGO_ENCODED_RUSTFLAGS") {
+                for mut flag in rustflags.to_string_lossy().split('\x1f') {
+                    flag = strip_prefix(flag, "-C").unwrap_or(flag);
+                    if let Some(flag) = strip_prefix(flag, "target-feature=") {
+                        for s in flag.split(',') {
+                            // TODO: Handles cases where a specific target feature
+                            // implicitly enables another target feature.
+                            match (s.as_bytes().first(), s.get(1..)) {
+                                (Some(b'+'), Some(f)) => {
+                                    if let Some(sm) = strip_prefix(f, "sm_") {
+                                        if let Ok(sm) = sm.parse::<u32>() {
+                                            if sm >= 70 {
+                                                has_sm_70 = true;
+                                            }
+                                        }
+                                    }
+                                }
+                                (Some(b'-'), Some(_f)) => {
+                                    // TODO
+                                }
+                                _ => {}
+                            }
+                        }
+                    }
+                }
+            }
+            target_feature_if("sm_70", has_sm_70, &version, None, false);
+        }
         _ => {}
     }
 }

src/imp/float.rs

@@ -5,9 +5,8 @@
 // Note that most of `fetch_*` operations of atomic floats are implemented using
 // CAS loops, which can be slower than equivalent operations of atomic integers.
 //
-// GPU targets have atomic instructions for float, so GPU targets will use
-// architecture-specific implementations instead of this implementation in the
-// future: https://github.com/taiki-e/portable-atomic/issues/34
+// GPU targets have atomic instructions for float, so we use these instructions
+// for GPU targets on nightly (see nvptx.rs).
 
 #![cfg_attr(
     all(target_pointer_width = "16", not(feature = "fallback")),

src/imp/mod.rs

@@ -135,6 +135,14 @@ mod arm_linux;
 #[cfg(target_arch = "msp430")]
 pub(crate) mod msp430;
 
+#[cfg(feature = "float")]
+#[cfg(all(
+    target_arch = "nvptx64",
+    any(target_feature = "sm_70", portable_atomic_target_feature = "sm_70"),
+    portable_atomic_unstable_asm_experimental_arch,
+))]
+pub(crate) mod nvptx;
+
 // atomic load/store for RISC-V without A-extension
 #[cfg(any(test, not(feature = "critical-section")))]
 #[cfg_attr(portable_atomic_no_cfg_target_has_atomic, cfg(any(test, portable_atomic_no_atomic_cas)))]
@@ -219,8 +227,21 @@ mod interrupt;
 // Atomic float implementations
 
 #[cfg(feature = "float")]
+#[cfg(not(all(
+    target_arch = "nvptx64",
+    any(target_feature = "sm_70", portable_atomic_target_feature = "sm_70"),
+    portable_atomic_unstable_asm_experimental_arch,
+)))]
 pub(crate) mod float;
 
+#[cfg(feature = "float")]
+#[cfg(all(
+    target_arch = "nvptx64",
+    any(target_feature = "sm_70", portable_atomic_target_feature = "sm_70"),
+    portable_atomic_unstable_asm_experimental_arch,
+))]
+pub(crate) use nvptx as float;
+
 // -----------------------------------------------------------------------------
 
 #[cfg(not(any(