| *** tpb has joined #tomu | 00:00 | |
| *** CarlFK has quit IRC | 00:54 | |
| *** squelch has joined #tomu | 01:05 | |
| *** squelch has quit IRC | 01:10 | |
| *** CarlFK has joined #tomu | 04:37 | |
| *** rohitksingh has joined #tomu | 04:48 | |
| *** krogozinski has quit IRC | 06:20 | |
| *** krogozinski has joined #tomu | 06:26 | |
| *** ln5 has joined #tomu | 07:55 | |
| *** rohitksingh has quit IRC | 09:00 | |
| *** gio has quit IRC | 09:59 | |
| *** gio has joined #tomu | 10:02 | |
| *** krogozinski has quit IRC | 11:32 | |
| *** im-tomu has left #tomu | 12:13 | |
| *** im-tomu has joined #tomu | 12:13 | |
| st-gourichon-fid | Hi ! | 13:17 |
|---|---|---|
| st-gourichon-fid | Question about the FOMU. | 14:01 |
| st-gourichon-fid | Our application needs relatively fast non-integer computation. | 14:02 |
| st-gourichon-fid | Fixed point would be okay, maybe even fine. Floats would be ideal. | 14:02 |
| st-gourichon-fid | IIRC RiscV spec mentions fixed point but I could not find an implementation of RiscV that would fit float or even fixed point inside an Ice40. | 14:03 |
| st-gourichon-fid | We are considering putting another softcore, like Arm or Mips inside the Ice40, if it can fit fast calculus (float or fixed point). | 14:04 |
| xobs | You could definitely replace the core, but I don't know that it'll be fast. | 14:04 |
| st-gourichon-fid | But maybe that would be too big whatever the CPU architecture we choose? | 14:04 |
| xobs | You could certainly try. | 14:15 |
| xobs | Your options are to make a CSR-mapped (memory-mapped) accelerator, or add an extension to the CPU. | 14:15 |
| xobs | There is a fixed-point working group, but I don't think they've published anything yet. | 14:16 |
| *** TheJJ has quit IRC | 14:19 | |
| *** TheJJ has joined #tomu | 14:33 | |
| *** xkapastel has joined #tomu | 14:50 | |
| *** emeb has joined #tomu | 15:09 | |
| ovf | st-gourichon-fid: what kind of hardware support do you envision you will require for fixed point computations? | 15:13 |
| *** rohitksingh has joined #tomu | 15:44 | |
| *** rohitksingh has quit IRC | 15:45 | |
| *** rohitksingh has joined #tomu | 15:49 | |
| st-gourichon-fid | ovf, not sure I understand your question. Do you mean, like 16.16 bits? | 15:53 |
| *** daurnimator has quit IRC | 16:03 | |
| st-gourichon-fid | xobs, thansk for the suggestion of a memory-mapped accelerator. Indeed it would be simple to integrate. | 16:18 |
| st-gourichon-fid | But non-standard, have to use a modified toolchain. | 16:19 |
| tnt | well, is there a standard extension for fixed point ? | 16:20 |
| *** daurnimator has joined #tomu | 16:22 | |
| ovf | st-gourichon-fid: i meant, what specific non-standard functionality/instructions do you require? it is true that one can benefit from some specific types of saturation arithmetic, but generally the biggest draw of fixed point is its not needing any hardware support outside of conventional integer operations. | 16:28 |
| st-gourichon-fid | ovf, thanks for the extended question! | 16:29 |
| st-gourichon-fid | I have not mentioned saturation arithmetic yet it may be necessary. It remains one often overlooked benefit of usual floats, that they don't wrap. | 16:33 |
| st-gourichon-fid | All the details are not yet defined, yet it is desirable that users of our platform can just write code (ideally Python, else C or similar), without thinking much more than "real number with limited precision". | 16:37 |
| *** rohitksingh has quit IRC | 16:44 | |
| *** rohitksingh has joined #tomu | 16:46 | |
| *** rohitksingh has quit IRC | 16:51 | |
| ovf | it looks to me more of a tooling than hardware capacity problem, albeit far from a trivial one. there is a proposal for fixed-point types in c, N1169, which both gcc and clang purport to support to various degrees, but from what i know less than well. python implementations don't seem to have an option of built-in fixed point support. | 16:57 |
| *** xkapastel has quit IRC | 17:02 | |
| st-gourichon-fid | ovf, thanks for the reference. You confirm my intuitions. | 17:04 |
| *** rohitksingh has joined #tomu | 17:08 | |
| *** andi- has quit IRC | 17:43 | |
| *** andi- has joined #tomu | 17:48 | |
| *** andi- has quit IRC | 17:56 | |
| *** andi- has joined #tomu | 17:58 | |
| *** andi- has quit IRC | 18:17 | |
| *** andi- has joined #tomu | 18:21 | |
| *** rohitksingh has quit IRC | 18:36 | |
| *** rohitksingh has joined #tomu | 18:53 | |
| *** rohitksingh has quit IRC | 19:59 | |
| *** squelch has joined #tomu | 20:09 | |
| *** AnimaInvicta has joined #tomu | 21:09 | |
| *** CarlFK has quit IRC | 21:25 | |
| *** TheJJ has quit IRC | 21:39 | |
| *** squelch has quit IRC | 22:19 | |
| *** TheJJ has joined #tomu | 22:27 | |
| *** squelch has joined #tomu | 22:52 | |
| *** squelch has quit IRC | 22:58 | |
| *** AnimaInvicta has left #tomu | 23:34 | |
| *** emeb has quit IRC | 23:56 | |
Generated by irclog2html.py 2.13.1 by Marius Gedminas - find it at mg.pov.lt!