Posted by grigio 1 day ago
It needs no host-platform tools to compile code, so for example I have a signed binary on my iPhone, written and signed entirely on Linux.
The compiler (v 0.61, not the released v0.6) currently produces code that is between 5x and 0.5x the speed of clang when measured at -O3 on both, against ObjC code, which is pretty good for a nascent compiler. The geometric mean of 19 mini-benchmarks comes in at xc being roughly 6% faster than clang.
Since there is also a framework and gui-designer under way, I’d like to release it as more than a tar.bz2 file for Linux at some point. Flatpark would seem to be appropriate :)
Zynq is an FPGA with dual Arm A9 cores, on one of which I run the OS, the other is earmarked for the m68k/030 emulator. There’s a hardware 6502 in fabric which has ANTIC, POKEY, and HDMI output (with audio). The OS includes a compositing window manager which uses shared memory to give each GEM window a retained-mode drawing surface.
You can see it booted up, running the emulator in either of:
Https://0x0000ff.co.uk/mov/xt/xtos-aug-11.mov
https://0x0000ff.co.uk/mov/xt/ballblazer.mov
The compiler is self-hosting, so can run on the OS on the FPGA, but this is quite long-winded an explanation so I just shortened it to Zynq.
The “I” in quotes is (for those who didn’t notice) an indication that I used an LLM to help me do all this.
- The Zynq MPSoc is the Ultrascale Zynq (released 4 years later), not the plain Zynq. A “good” (7020) plain Zynq board with decent I/O ports costs about £100, the very cheapest Ultrascale boards are twice that with far less expansion ports - to get similar I/O you’ll be up in the 3-4x price bracket at least. They’re very different products, fabricated on different technologies (16nm, not 28nm), with fundamentally different internal architecture (eg: Ultrascale has built-in Ultraram, far better power management, quad A53 + dual R5F vs (max)dual A9, codecs on some, ddr4 support,…)
- This isn’t the first OS I (without the quotes) have written for the Zynq. I wrote the OS for the Apple Vision Pro, quite some time before LLM’s existed, when it was a prototype on my desk, consisting of a Zynq being the boot-up device and architecture controller for a bunch of the fastest and largest FPGAs on the planet interconnected with optical fibre. And this one did not rely on FreeRTOS (or any other OS) for anything.
- I would agree with you that it ought to be blatantly obvious that I’ve used an LLM to help out, the timescale itself should be a dead giveaway. However, the very first comment I got (“How is this supposed to be relevant to the article we're discussing?”) seemed to miss the quotes around the “I”. The signoff comment was more a “get a life” response to the subsequent down-voting haters than anything else. LLMs are just tools, you still need the skills and knowledge to direct them and correct them on anything as significant as a compiler.
- “modified” is carrying significant water… XTOS has shared libraries, dynamic linking, can launch external programs from disk not just compiled-in tasks, has shared memory, CoW pages, page-protection (nx, pid locking), swap, threading, system calls, a posix layer, networking, virtual filesystem support, boot-script support, interactive shells with ptys, and a graphical compositing window manager which outputs 1080p HDMI (ie: with the embedded audio islands, not just DVI over an HDMI interface). But yes, “I” still use FreeRTOS for context switching and interrupt control. It boots far far faster than Linux, which was the point. I wanted a switch-on-and-go appliance-like OS, not a visible boot sequence.