Can't use nasm feature on arch linux (for faster rav1d video decoding)

[jleibs]

To reproduce:

cargo run --package rerun-cli --no-default-features --features native_viewer,nasm

Output:

  = note: /usr/bin/ld: /home/jleibs/rerun/target/debug/deps/librav1d-13aa763589daefd4.rlib(mc16_avx512.o): warning: relocation against `dav1d_mc_warp_filter' in read-only section `.text'
          /usr/bin/ld: /home/jleibs/rerun/target/debug/deps/librav1d-13aa763589daefd4.rlib(looprestoration_sse.o): relocation R_X86_64_PC32 against symbol `dav1d_sgr_x_by_x' can not be used when making a shared object; recompile with -fPIC
          /usr/bin/ld: final link failed: bad value
          collect2: error: ld returned 1 exit status

Affected platforms: Arch Linux

• Regression caused by: Always-on AV1 in --release; opt-in to nasm-optimizations #7661

I'm confirmed I have nasm installed:

➜  nasm --version
NASM version 2.16.03 compiled on May  4 2024

[jleibs]

Also fails in CI:

• https://github.com/rerun-io/rerun/actions/runs/11261410122/job/31314795062#step:6:386

[jprochazk]

I can repro this on Linux (Fedora). I tried setting -Clink-args=-fPIC, but that didn't help, it still claims it should be recompiled with -fPIC. Neither mold nor lld work.

Weirdly enough, it's just re_viewer. Compiling re_video by itself seems to work?

cargo run --release -p re_video --example frames -F nasm

But this does not:

cargo run --release --package rerun-cli --no-default-features --features native_viewer,nasm

[emilk]

So maybe some clash of dependencies then?

[teh-cmc]

TL;DR:

• Only x64 platforms go through the nasm-based path, everything else uses the system's default assembler (I'm not sure why, but naively trying to use the system's assembler for x64 fails miserably).
• On Linux x64, the nasm path generates relocation entries that fail to resolve at link-time for some unknown reason, even though the whole thing is fully statically link, only use relative relocs, etc. Even weirder, it only fails when linking the full viewer, simpler examples do link successfully (it's possible that the problematic symbols are unused and therefore optimized out before the reloc pass for these examples?).
• On Mac x64, the nasm path just works. Somehow.

[Wumpf]

On windows x64 it works like a charm. Given that it makes 8k video usable (completely unusable without), I think as a first step we should limit the changes of #7675 to Linux

[Wumpf]

:sigh: that is ofc running against rust-lang/cargo#1197, it's very hard to make this feature have only an effect on certain platforms

[emilk]

I think we can work-around it by adding this to a crate other than re_video

[target.'cfg( not(target_os = "linux") )'.dependencies] # TODO(#7671)
re_video = { workspace = true, features = [ "nasm" ] }

[teh-cmc]

Interestingly, the official dav1d (Dav1d, with a D!) libraries don't even have any of these problematic symbols:

readelf --wide -s /usr/lib/libdav1d.so | rg dav1d_
    39: 0000000000003284  1797 FUNC    GLOBAL DEFAULT   11 dav1d_open
    40: 000000000016a390   912 FUNC    GLOBAL DEFAULT   11 dav1d_flush
    41: 0000000000169910   610 FUNC    GLOBAL DEFAULT   11 dav1d_parse_sequence_header
    42: 0000000000003989    20 FUNC    GLOBAL DEFAULT   11 dav1d_close
    43: 0000000000169c60   113 FUNC    GLOBAL DEFAULT   11 dav1d_data_props_unref
    44: 0000000000169320   206 FUNC    GLOBAL DEFAULT   11 dav1d_data_wrap
    45: 0000000000002b81    12 FUNC    GLOBAL DEFAULT   11 dav1d_version
    46: 0000000000002d9f    96 FUNC    GLOBAL DEFAULT   11 dav1d_get_frame_delay
    47: 000000000016a060   413 FUNC    GLOBAL DEFAULT   11 dav1d_apply_grain
    48: 0000000000172f00   890 FUNC    GLOBAL DEFAULT   11 dav1d_get_picture
    49: 00000000001692a0   115 FUNC    GLOBAL DEFAULT   11 dav1d_data_create
    50: 0000000000169270    41 FUNC    GLOBAL DEFAULT   11 dav1d_get_event_flags
    51: 0000000000002020    11 FUNC    GLOBAL DEFAULT   11 dav1d_set_cpu_flags_mask
    52: 0000000000169ce0   253 FUNC    GLOBAL DEFAULT   11 dav1d_get_decode_error_data_props
    53: 0000000000002b8d    10 FUNC    GLOBAL DEFAULT   11 dav1d_version_api
    54: 0000000000169470     9 FUNC    GLOBAL DEFAULT   11 dav1d_data_unref
    55: 00000000001693f0   125 FUNC    GLOBAL DEFAULT   11 dav1d_data_wrap_user_data
    56: 0000000000172e50   164 FUNC    GLOBAL DEFAULT   11 dav1d_send_data
    57: 0000000000002b97    86 FUNC    GLOBAL DEFAULT   11 dav1d_default_settings
    58: 000000000016a720     9 FUNC    GLOBAL DEFAULT   11 dav1d_picture_unref

So I guess these routines are specific to the Rust implementation.

[emilk]

Should have been fixed by rerun-io/re_rav1d#3, but lets test it before closing (and maybe follow-up and remove TODO(#7671) from the Rerun codebase)

[emilk]

Apparently the -F nasm feature makes a 25x perf difference for @teh-cmc on Linux 64.

Based on that, I think we should disable av1 decoding without -F nasm on Linux and Windows (Mac is fast even without it).