#
✕ nick:linkmauve
#asahi · 2026-09-15 06:51
linkmauveWhat are blobs?
#asahi · 2026-09-15 06:50
linkmauveI haven’t tried using a VM in years, but venus is Vulkan being serialized through the VM and then executed by virglrenderer, whereas DRM native contexts is about running the native driver and just replacing the ioctl()s, IIRC.
#asahi · 2026-09-15 06:48
linkmauvenort_, you probably want to use DRM native contexts instead of venus, if you want usable performances.
#asahi · 2026-09-14 21:54
linkmauveRight, that’s what I was thinking about.
#asahi · 2026-09-14 21:52
linkmauvechaos_princess, oh?
#asahi · 2026-09-14 21:52
linkmauveTurns out having 512-bit vectors is actually useful, so now they’re back to trying to put it in actual cores.
#asahi · 2026-09-14 21:52
linkmauveThey cancelled it now.
#asahi · 2026-09-14 21:48
linkmauveI got no computer with any of the AVX512 wilderness of extensions.
#asahi · 2026-09-14 21:47
linkmauveOops, just did.
#asahi · 2026-09-14 21:42
linkmauvechadmed, thanks! Just signed my commits.
#asahi · 2026-09-14 18:52
linkmauveAh, I missed the 1.98.0 release blog post due to being at a festival, that’s why I had never heard about it.
#asahi · 2026-09-14 18:50
linkmauveTIL, this is -ffast-math but localized to a single operation, nice!
#asahi · 2026-09-14 18:48
linkmauveThis code will exclusively run on an ecore right?
#asahi · 2026-09-14 18:48
linkmauveNeat, which new Rust feature did you need?
#asahi · 2026-09-14 18:01
linkmauveThat’s great!
#asahi · 2026-09-14 18:01
linkmauveWow, so about twice as fast as before!
#asahi · 2026-09-14 17:48
linkmauveI just pushed a new version, the only change is to use vfmaq_n_f32() instead of vfmaq_f32() for the last step, which takes a f32 instead of a float32x4 which I had to vdupq_n_f32() at first.
#asahi · 2026-09-14 17:47
linkmauveFor the example no, it shouldn’t.
#asahi · 2026-09-14 17:43
linkmauveWouldn’t it be more interesting to do it on an actual Apple core though?
#asahi · 2026-09-14 17:41
linkmauveOh…
#asahi · 2026-09-14 17:41
linkmauvechaos_princess, you mean line numbers from the Rust source code? In the hot loop, it’s basically 1:1 Rust-asm.
#asahi · 2026-09-14 17:26
linkmauveI wanted to bump its syn dependency, to reduce compilation times (Cortex-A53 is slow, turns out), but there might be better use of my time.
#asahi · 2026-09-14 17:25
linkmauveOw, the lv2 crate hasn’t seen a single commit in the past five years…
#asahi · 2026-09-14 17:22
linkmauveperf report might give something sensible still.
#asahi · 2026-09-14 17:22
linkmauvehttps://linkmauve.fr/files/performance_test.O2.perf.data
#asahi · 2026-09-14 17:21
linkmauveWould you be able to make use of a perf.data?
#asahi · 2026-09-14 17:20
linkmauveAt O2 or O3?
#asahi · 2026-09-14 17:20
linkmauveI do.
#asahi · 2026-09-14 17:19
linkmauveopt-level = 2 is much slower, at 715ms mean out of ten runs.
#asahi · 2026-09-14 17:17
linkmauveOn my PinePhone.
#asahi · 2026-09-14 17:17
linkmauvechaos_princess, your suggestions lowered the total time for the “benchmark” from 775ms to 701ms mean out of ten runs!
#asahi · 2026-09-14 17:03
linkmauveThe instructions seem less reordered to take into account microarchitectural details though.
#asahi · 2026-09-14 17:03
linkmauveIt still is at opt-level = 2, not at "s".
#asahi · 2026-09-14 16:59
linkmauveAh I just figured out why all iterations get unrolled: you define opt-level = 3 in Cargo.toml, at 2 or "s" it emits a proper loop.
#asahi · 2026-09-14 16:46
linkmauveRight.
#asahi · 2026-09-14 16:45
linkmauveBut it would be better to do it in the code I guess, to avoid the next person to wonder about that.
#asahi · 2026-09-14 16:45
linkmauveAh yeah, I had thought about that but figured out the compiler already did it.
#asahi · 2026-09-14 16:43
linkmauveI don’t understand why everything gets unrolled, all nine iterations.
#asahi · 2026-09-14 16:42
linkmauvechaos_princess, https://linkmauve.fr/files/hilbert_iir.S
#asahi · 2026-09-14 16:30
linkmauveThere, pushed again.
#asahi · 2026-09-14 16:29
linkmauveIt does, just checked. :)
#asahi · 2026-09-14 16:28
linkmauveAh yeah, makes sense.
#asahi · 2026-09-14 16:27
linkmauveThe scalar fma instruction, which could be used to reduce input[k][i] * self.direct + vaddvq_f32(ra) by one instruction.
#asahi · 2026-09-14 16:26
linkmauveWut, there is no intrinsic whatsoever for the FMADD instruction?!
#asahi · 2026-09-14 16:16
linkmauvechaos_princess, how do you do that without miscomputing the state?
#asahi · 2026-09-14 15:54
linkmauveI didn’t do it in the x86 version, since that’s an extension not part of SSE, so not guaranteed to be present on all amd64 CPUs.
#asahi · 2026-09-14 15:54
linkmauveI just pushed your version.
#asahi · 2026-09-14 15:54
linkmauvechaos_princess, oh indeed, that removes four instructions for every iteration, thanks!
#asahi · 2026-09-14 14:45
linkmauvechadmed, JamesCalligeros[m], I just opened https://github.com/chadmed/triforce/pull/8 if you’re interested in testing that on Apple hardware.
#asahi · 2026-09-12 22:52
linkmauveBoth auto-vectorized and manually written.

plus ›