Skip to content
Snippets Groups Projects

Draft: aarch64: add DAV1D_ARM_CPU_FLAG_DOTP

Closed Draft: aarch64: add DAV1D_ARM_CPU_FLAG_DOTP
janne/dav1d:arm64_cpuid into master
1 unresolved thread
Closed Janne Grunau requested to merge janne/dav1d:arm64_cpuid into master
1 unresolved thread

Runtime detection on Linux using HWCAP_CPUID for user space access to the CPU feature registers. See https://www.kernel.org/doc/html/latest/arm64/cpu-feature-registers.html

Runtime detection on MacOS using sysctlbyname("hw.cpufamily") and matching known ARMv8.4-A CPUs. MacOS unfortunately doesn't expose a flag for this feature.

marked as draft as it is pointless without users.

Merge request reports

Pipeline #65789 passed

Stage:
    Stage:
      Stage:

        Pipeline passed for a10bbd1f on janne:arm64_cpuid

        Test coverage 95.83% (0.02%) from 1 job
        Approval is optional

        Closed by Ronald S. BultjeRonald S. Bultje 1 year ago (Feb 24, 2024 6:05pm UTC)

        Merge details

        • The changes were not merged into master.

        Activity

        Filter activity
        • Approvals
        • Assignees & reviewers
        • Comments (from bots)
        • Comments (from users)
        • Commits & branches
        • Edits
        • Labels
        • Lock status
        • Mentions
        • Merge request status
        • Tracking
      • Janne Grunau resolved all threads

        resolved all threads

      • Janne Grunau added 1 commit

        added 1 commit

        • dcf4ce3c - aarch64: add DAV1D_ARM_CPU_FLAG_DOTP

        Compare with previous version

      • Martin Storsjö
      • Martin Storsjö
        Martin Storsjö @mstorsjo started a thread on commit dcf4ce3c
        • src/arm/cpu.c
        124 size_t size = sizeof(cpu_family);
        125 /* there is no explicit flag for dot product availability
        126 * enable it on know ARMv8.4-A CPUs
        127 */
        128 int ret = sysctlbyname("hw.cpufamily", &cpu_family, &size, NULL, 0);
        129 if (!ret) {
        130 switch (cpu_family)
        131 {
        132 case CPUFAMILY_ARM_LIGHTNING_THUNDER: // explicitit fall through
        133 case CPUFAMILY_ARM_FIRESTORM_ICESTORM:
        134 flags |= DAV1D_ARM_CPU_FLAG_DOTP;
        135 break;
        136 default:
        137 break;
        138 }
        139 }
        • Martin Storsjö
          Martin Storsjö @mstorsjo ·
          Developer

          Would it be good to somehow fold in a check for __ARM_FEATURE_DOTPROD too, which would imply that it's always available unconditionally? (Clang doesn't seem to set that one, but GCC does if building with -march=armv8.4-a.)

        • Janne Grunau
          Janne Grunau @janne ·
          Author Maintainer

          I'd say yes but I'm reluctant to add this to the current #idef hell. I'll try to think of something. Maybe it would make sense to split cpu.c into 32/64-bit files.

        • Martin Storsjö
          Martin Storsjö @mstorsjo ·
          Developer

          Yep, the code is a bit tricky. Maybe a separate #if defined(__aarch64__) && defined(__ARM_FEATURE_DOTPROD) (and we could have the same for __ARM_NEON and other hardcoded conditions) after the different cases of actually doing runtime probing?

        • Please register or sign in to reply
      • Janne Grunau added 1 commit

        added 1 commit

        • a10bbd1f - aarch64: add DAV1D_ARM_CPU_FLAG_DOTP

        Compare with previous version

      • Ronald S. Bultje mentioned in merge request !1609 (merged)

        mentioned in merge request !1609 (merged)

      • Ronald S. Bultje
        Ronald S. Bultje @rbultje ·
        Developer

        Superseeded by !1609 (merged).

      • Ronald S. Bultje closed

        closed

        • Please register or sign in to reply

        VideoLAN code repository instance