ds54e (@ds54e) header
ds54e (@ds54e) profile picture

Verilogであそびたかった(過去形)

Joined November 2019
Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)Photo by ds54e (@ds54e)
xezim 0.11.1 is out 🚀 An open-source SystemVerilog simulator, written in Rust. This release brings complete VPI, much leaner memory models, and a clause-by-clause IEEE 1800 conformance pass. 🔍 Conformance • We checked ~490 subclauses of IEEE 1800-2023 against a reference simulator, and fixed every high-severity gap: - procedural variable typing; - class property scoping and parameterized classes; - §13.5 argument passing (by-value copies, true ref -aliasing); - clocking-block skews (#1step, input/output skew); - disable from forked children, force/release on net bit-selects; - assertion local variables, procedural assertions, default disable iff. • A new 169-test conformance suite guards 409 subclauses. • The test suite runs with zero ignored tests. 🔌 VPI and DPI • Every IEEE 1800 clause 38 routine, every vpi_register_cb reason, and a full design walk. • Imported DPI tasks run on their own stacks, so concurrent callers suspend and resume independently. • The §35.9 disable protocol (svIsDisabledState) is supported. • export “DPI-C” works from any module instance. • Packed structs and multi-dimensional open arrays cross the DPI boundary. ⚡ Faster, leaner simulation • Large memories live in a packed arena by default: C906 CoreMark peaks at 0.6 GB instead of 2.6 GB. • A 64k-cell DRAM model needs about a tenth of the work, and a 256k-cell model fits in 2.1 GB instead of 7 GB. • Behavioural DDR4-style models run 4 to 6 times faster. • Startup is cheaper, and host instructions drop 7% on C906 CoreMark and 11% on a C910 memcpy run. 🧹 Cleanup • The experimental multikernel/PDES code is gone (−7,500 lines). • The release binary is 1.6% smaller. 🙏 Thank you to everyone who contributed to this release: Thomas Burg, Taichi Ishitani, Ganesh T S, eenky and Francesco Urbani. Release notes: github.com/aionhw/xezim/rele… Repo: github.com/aionhw/xezim Try it, break it, file issues, send PRs.
1
4
376
xezim 0.11.0 is out 🚀 xezim is an open-source SystemVerilog simulator written in Rust. This release was about one thing: closing the gap with commercial simulators. Nearly every behaviour change was checked against a commercial reference simulator before it landed, and more than 400 new regression tests pin the results. Here's what's new. 📊 Coverage • Code coverage is new: statement, branch and toggle coverage with --code-coverage (or the familiar +cover / -coverage). It's off by default and costs nothing when off. Results go to xezim_cov.json, per instance and per design unit. • Functional coverage grew up: – covergroups work in submodules, interfaces, packages and multi-top designs – crosses support binsof, intersect, with, and ignore/illegal bins – transition bins, bins x[N], iff sampling, start()/stop() and type-level get_coverage() all work, and cover sequence is new – illegal_bins hits are real errors • Coverage numbers match the reference simulator on every pinned case. 🎲 Constrained random • A new joint solver handles sum(), unique and ordering constraints together, so problems that used to fail now solve every time. • dist weights (:= and :/) and solve … before shape the distribution the way the standard describes. • Inline randomize() with {} binds names to the object first, then to the calling scope (§18.7). • Constraints on nested rand objects are solved jointly. 🧪 UVM • Built-in UVM DPI: regex and glob matching for config_db and the factory, uvm_hdl_* backdoor access, and the command-line processor, for both UVM 1.2 and 2020.3. UVM_NO_DPI is now optional. • Transcripts of five open-source UVM VIPs (axi4, apb, i3c, spi, axi4-lite) now match the reference simulator message for message. • UVM tables print correctly, $typename matches, and dozens of class, struct and interface fixes landed. ⚡ Performance • UVM class code: less than half the host instructions per transaction on our UVM stress bench (20.6 M → 9.0 M), and about a third fewer on the axi4 VIP. • RTL: a 2.5% gain on a RISC-V SoC CoreMark run. ⏱️ Timing and gate level • Specify-block timing checks with notifiers: $setup, $hold, $setuphold, $recovery, $removal, $recrem, $width, $period, $skew and more. • SDF back-annotation of timing checks. • Specify path delays honour the module timescale. 🛠️ Easier to adopt • Quiet transcripts by default; engine internals appear only with --verbose. • Errors show file:line:col with a source snippet. • Command lines written for other simulators run as-is: bare top names, -F, -do "run -all", -g/-G, -sv_seed, -sv_lib, -l. • The default time limit is now 100 ms, so long UVM tests finish. ✅ Conformance • sv-tests: 99.0% pass, up from 94.0%. • Scheduling order (time-0 activation, continuous-assign wakeups) now matches the reference, so racy testbenches behave the same. Opus 5.5 is good at cleaning up gaps and bugs with multiagents. Release notes: github.com/aionhw/xezim/rele… Repo: github.com/aionhw/xezim Coverage guide: github.com/aionhw/xezim/blob…
1
2
11
816
なぜ、情報を氷山に例えた際に、巨大かつ露見しない機密部分よりも、露見した公開部分が重要なのか?|pdfractal zenn.dev/pdfractal/articles/… #zenn
3
1
11
576
Strict SystemVerilog LRM compliance turns out to be necessary but not sufficient. A design can be perfectly legal SystemVerilog and still not run, because the flow around it assumes things no spec ever wrote down. Vendor behaviors. Undocumented command-line conventions. Twenty years of accumulated build assumptions. Most recent xezim fixes came out of exactly that gap. Real RTL goes in, it breaks where the standard is silent, and every commercial tool turns out to have quietly agreed on the same unwritten answer years ago. That's also the only kind of failure that moves confidence. Passing your own test suite proves self-consistency and nothing more. Real code fails, you understand why, you fix it, it never fails that way again. The fix goes into regression and the next project starts from a higher floor. Curated demos can't produce that. What would you prioritize: deeper LRM coverage, UVM/DPI/VPI compatibility, or compatibility with existing commercial tool flows?
1
1
116
Shipped a release of xezim — a SystemVerilog simulator written in Rust. Quick background: EDA tools (the software chip designers use to write and test silicon) usually cost millions and take big teams years to build. xezim is a small experiment in the other direction — what can a tiny crew plus AI actually pull off? Result on real RISC-V chips: • T-Head C910 hello: 95s → 73s (1.30× faster) • T-Head C910 memcpy: 216s → 166s (1.30×) • T-Head C910 CoreMark: 87 min → 73 min (1.19×) • T-Head C906 memcpy: 99s → 88s (1.13×) • T-Head C906 CoreMark: 714s → 587s (1.22×) Every run produces bit-identical results to baseline. CoreMark cycle counts match. Simulated timestamps match. TEST PASSED, same as before — just faster. It also runs real UVM 1.2 end-to-end. Hooked up riscv-dv (the standard tool for RISC-V random testing), asked it for 10 random programs, got 10 back that all assemble cleanly with riscv64-unknown-elf-as. That was satisfying. Open source. Rust-native. Built incrementally. Code: github.com/aionhw/xezim If you work on simulators, verification, or just enjoy watching expensive tools get cheaper — try it, break it, and tell me what's missing. #OpenSource #Rust #ChipDesign #RISCV #EDA #SystemVerilog #UVM #Verification
2
2
253
ds54e retweeted
若者に教えるオブジェクト指向、オブジェクト同士でどのように通信するかを考えるのは有用だけど、そこについてくるclass継承完全に邪魔だからお兄ちゃんどいてそいつ殺せない
4
56
3
418
48,727
ようやく投稿できました オブジェクト指向の継承がなぜ問題を発生させるかを解説しています(私見なので異論歓迎です) youtu.be/YuMBCWbXtuw
3
8
1,182
Design Verificationについては現場からの感想としては・・・UVMを爆破して一から作り直してほしいです。
2
5
598
趣味で勉強する分には楽しいけど業務で来週までに理解しといてね〜と言われると脳が拒否するなDesign Verification
3
1
24
4,181
Explore hundreds of videos covering the latest in design and verification standards, technical presentations, and working group insights — all available anytime, on demand. Subscribe to Accellera’s YouTube channel and dive into what interests you: bit.ly/44fLBHT #EDA
1
2
412
シミュレーション限定ながら、半導体人材の強化につながるKindleを4月上旬に出せるぞ! (2025年の覇権本を目指す)
半導体人材の強化(学科新設、トレーニング所開設)があちこちで進んでいるけど、新しい本とか出たのかな。あんまり聞かないけどな。まじめにウオッチしてないけど。
2
1
10
1,581
UVM-MS 1.0 Now Available! Accellera's Board has approved the Universal Verification Methodology for Mixed-Signal (UVM-MS) 1.0 standard for immediate download fee free. bit.ly/3Q1vpSj More info bit.ly/3WQbsl1 or visit bit.ly/3Ej5dQg #IEEE #UVM #AMS
3
1
3
879
お!SSTCのZVS条件の解析用の式あってそう!まだs領域だからここからが重要なんだけども。 ちなみに適当に逆ラプラス変換したら600kBのでかすぎる式が出てきて全く解析的ではなかった。何かうまい方法を見つけないといけなそう
1
2
7
722
UVM-Mixed Signal draft standard Available for Public Review! The objective is to standardize methods for driving & monitoring mixed-signal nets within UVM. Download bit.ly/3tdwMpe. Give feedback via the UVM-MS community forum: bit.ly/46Vlu8j #UVM #MixedSignal
1
4
238
ep3できた…長かった…
1
2