Backport arm asm improvements from dav1d 1.4.2 (#1300)

Improvements include implementation of various functions using `dotprod` and `i8mm` ISA extensions.
memorysafety · Jul 17, 2024 · b26781a · b26781a
2 parents 068be6b + 6b87a77
commit b26781a
Show file tree

Hide file tree

Showing 8 changed files with 1,960 additions and 197 deletions.
diff --git a/build.rs b/build.rs
@@ -103,6 +103,11 @@ mod asm {
         if let Arch::Arm(arch) = arch {
             define(Define::bool("ARCH_ARM", arch == ArchArm::Arm32));
             define(Define::bool("ARCH_AARCH64", arch == ArchArm::Arm64));
+
+            if arch == ArchArm::Arm64 {
+                define(Define::bool("HAVE_DOTPROD", true));
+                define(Define::bool("HAVE_I8MM", true));
+            }
         }
 
         if let Arch::X86(arch) = arch {
@@ -199,6 +204,7 @@ mod asm {
         ][..];
 
         let arm_generic = &["itx", "msac", "refmvs", "looprestoration_common"][..];
+        let arm_dotprod = &["mc_dotprod"][..];
         let arm_bpc8 = &[
             "cdef",
             "filmgrain",
@@ -243,11 +249,20 @@ mod asm {
             #[cfg(feature = "bitdepth_16")]
             arm_bpc16,
         ][..];
+        let arm64_all = &[
+            arm_generic,
+            arm_dotprod,
+            #[cfg(feature = "bitdepth_8")]
+            arm_bpc8,
+            #[cfg(feature = "bitdepth_16")]
+            arm_bpc16,
+        ][..];
 
         let asm_file_names = match arch {
             Arch::X86(ArchX86::X86_32) => x86_all,
             Arch::X86(ArchX86::X86_64) => x86_64_all,
-            Arch::Arm(..) => arm_all,
+            Arch::Arm(ArchArm::Arm32) => arm_all,
+            Arch::Arm(ArchArm::Arm64) => arm64_all,
         };
 
         let asm_file_dir = match arch {
@@ -293,8 +308,11 @@ mod asm {
             }
             cc.compile(rav1dasm);
         } else {
-            cc::Build::new()
-                .files(asm_file_paths)
+            let mut cc = cc::Build::new();
+            if arch == Arch::Arm(ArchArm::Arm64) {
+                cc.flag("-march=armv8.6-a");
+            }
+            cc.files(asm_file_paths)
                 .include(".")
                 .include(&out_dir)
                 .debug(cfg!(debug_assertions))

diff --git a/include/common/bitdepth.rs b/include/common/bitdepth.rs
@@ -457,7 +457,10 @@ macro_rules! bd_fn {
 ///
 /// Similar to [`bd_fn!`] except that it selects which [`BitDepth`] `fn`
 /// based on `$bpc:literal bpc` instead of `$BD:ty`.
-#[cfg(all(feature = "asm", any(target_arch = "x86", target_arch = "x86_64")))]
+#[cfg(all(
+    feature = "asm",
+    any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64")
+))]
 macro_rules! bpc_fn {
     ($bpc:literal bpc, $name:ident, $asm:ident) => {{
         use $crate::include::common::bitdepth::fn_identity;
@@ -487,7 +490,10 @@ macro_rules! fn_identity {
 ))]
 pub(crate) use bd_fn;
 
-#[cfg(all(feature = "asm", any(target_arch = "x86", target_arch = "x86_64")))]
+#[cfg(all(
+    feature = "asm",
+    any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64")
+))]
 pub(crate) use bpc_fn;
 
 #[allow(unused)]