Skip to content

Commit 85b7b71

Browse files
hugomeilandcursoragent
hugomeiland
and
cursoragent
committed
Add U74 hybrid video production pack (no render).
English script, storyboard, branded HTML slides for screen capture, and VisionFive 2 terminal recording guide. Co-authored-by: Cursor <cursoragent@cursor.com>
1 parent 423458d commit 85b7b71

6 files changed

Lines changed: 452 additions & 0 deletions

File tree

‎video/u74/README.md‎

Lines changed: 35 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,35 @@
1+
# U74 progress video — hybrid production pack
2+
3+
**Status:** assets ready · **no render yet**
4+
**Language:** English
5+
**Format:** 16:9 · 1080p target · ~75 s
6+
**Style:** motion slides (screen-capture) + terminal clip(s)
7+
8+
## What’s in this folder
9+
10+
| File | Purpose |
11+
| ---- | ------- |
12+
| [`script.en.md`](script.en.md) | Voiceover with timestamps (~75 s) |
13+
| [`storyboard.md`](storyboard.md) | Shot list: motion vs terminal, edit notes |
14+
| [`terminal-recording.md`](terminal-recording.md) | Exact commands to record on VisionFive 2 |
15+
| [`slides/index.html`](slides/index.html) | Branded slide deck — open in browser, record with OBS/etc. |
16+
17+
## How to produce (when ready to render)
18+
19+
1. **Record motion (~45 s)** — open `slides/index.html` fullscreen (1920×1080). Advance with → or Space; optional auto-play via `?autoplay=1`. Screen-record at 60 fps.
20+
2. **Record terminal (~25 s)** — follow [`terminal-recording.md`](terminal-recording.md) on real VF2 hardware (or use placeholder if unavailable).
21+
3. **Record voiceover** — read [`script.en.md`](script.en.md) in a quiet room; export WAV/MP3.
22+
4. **Edit** — cut to [`storyboard.md`](storyboard.md); duck music under VO if used; end card = slide 6.
23+
5. **Export** — H.264 1080p, ~8 Mbps; optional 9:16 centre crop for social.
24+
25+
## Site links to cite on end card
26+
27+
- [opensolvers.com](https://www.opensolvers.com)
28+
- [VisionFive 2 board page](https://www.opensolvers.com/boards/VisionFive2.html)
29+
- [EESSI/docs#818](https://github.com/EESSI/docs/pull/818)
30+
- [easyconfigs#26436](https://github.com/easybuilders/easybuild-easyconfigs/pull/26436)
31+
32+
## Do not do yet
33+
34+
- Remotion / ffmpeg final render (explicitly deferred)
35+
- Upload to YouTube or embed on site until review

‎video/u74/script.en.md‎

Lines changed: 57 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,57 @@
1+
# Voiceover script — U74 on VisionFive 2 (English)
2+
3+
**Target duration:** 72–78 seconds
4+
**Pace:** ~140 words/minute · **Word count:** ~175
5+
6+
Record as one take per section if easier; align to [`storyboard.md`](storyboard.md).
7+
8+
---
9+
10+
## [0:00 – 0:14] Setup — SLIDE 1 → 2
11+
12+
OpenSolvers benchmarks open-source scientific software on real RISC-V hardware — through the EESSI stack, without rebuilding every application.
13+
14+
Our first board is the StarFive VisionFive 2: four SiFive U74 cores at one point five gigahertz. Scalar RV64GC — no vector extension. On this chip, our optimisation target is the **Scalar** path: tuned BLAS kernels, not RVV.
15+
16+
---
17+
18+
## [0:14 – 0:28] Problem — SLIDE 3
19+
20+
Out of the box, stock OpenBLAS on EESSI uses a generic C GEMM kernel.
21+
22+
High Performance Linpack on four cores reaches about **three point one three** gigaflops per second. The answer is correct — but we’re leaving throughput on the table.
23+
24+
---
25+
26+
## [0:28 – 0:52] Fix — SLIDE 4 + TERMINAL CLIP
27+
28+
We built an U74-tuned OpenBLAS with a hand-tuned **four-by-four DGEMM** micro-kernel — `TARGET=U74` — and swapped it in at runtime with FlexiBLAS.
29+
30+
Same HPL binary. No rebuild.
31+
32+
*(Terminal clip: module load, FlexiBLAS list, optional xhpl banner)*
33+
34+
Single-core DGEMM goes from about **one point four** to **one point eight** gigaflops per second. Four-core GEMM hits **six point three**. The kernel change propagates straight into Linpack.
35+
36+
---
37+
38+
## [0:52 – 1:15] Result — SLIDE 5 → 6
39+
40+
End-to-end HPL on the VisionFive 2: **three point one three** to **five point two eight** gigaflops per second — **one point seven times** faster. Wall time drops from **two hundred thirteen** seconds to **one hundred twenty-six**.
41+
42+
Same methodology we use on SpaceMiT X60 boards — change one backend, verify correctness first — but on U74 the story is **silicon-tuned scalar**, not fixing a broken vector path.
43+
44+
Everything is documented at **opensolvers dot com**.
45+
46+
---
47+
48+
## Pronunciation / on-screen numbers
49+
50+
| Spoken | On screen |
51+
| ------ | --------- |
52+
| three point one three | 3.13 GFLOP/s |
53+
| five point two eight | 5.28 GFLOP/s |
54+
| one point seven times | 1.69× |
55+
| one point four → one point eight | 1.4 → 1.77 GFLOP/s |
56+
| six point three | 6.31 GFLOP/s |
57+
| two hundred thirteen → one hundred twenty-six | 213 s → 126 s |

‎video/u74/slides/index.html‎

Lines changed: 236 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,236 @@
1+
<!DOCTYPE html>
2+
<html lang="en">
3+
<head>
4+
<meta charset="UTF-8">
5+
<meta name="viewport" content="width=device-width, initial-scale=1">
6+
<title>OpenSolvers — U74 progress (slides)</title>
7+
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
8+
<link href="https://fonts.googleapis.com/css2?family=Open+Sans:wght@400;600;700&display=swap" rel="stylesheet">
9+
<style>
10+
:root {
11+
--blue: #155799;
12+
--green: #159957;
13+
--teal: #157878;
14+
--link: #1e6bb8;
15+
--text: #606c71;
16+
--white: #fff;
17+
}
18+
* { box-sizing: border-box; margin: 0; padding: 0; }
19+
html, body { height: 100%; overflow: hidden; font-family: "Open Sans", sans-serif; background: #111; }
20+
.deck { height: 100%; position: relative; }
21+
.slide {
22+
position: absolute; inset: 0;
23+
display: flex; flex-direction: column; align-items: center; justify-content: center;
24+
padding: 3rem 4rem;
25+
opacity: 0; pointer-events: none;
26+
transition: opacity 0.45s ease;
27+
background: var(--white);
28+
}
29+
.slide.active { opacity: 1; pointer-events: auto; z-index: 1; }
30+
.slide--hero {
31+
background: linear-gradient(120deg, var(--blue), var(--green));
32+
color: var(--white); text-align: center;
33+
}
34+
.slide--hero h1 { font-size: 3.2rem; font-weight: 700; margin-bottom: 0.5rem; }
35+
.slide--hero p { font-size: 1.35rem; opacity: 0.85; max-width: 42rem; line-height: 1.5; }
36+
.slide h2 { color: var(--green); font-size: 2rem; font-weight: 600; margin-bottom: 1.5rem; text-align: center; }
37+
.slide p, .slide li { color: var(--text); font-size: 1.15rem; line-height: 1.55; }
38+
.row { display: flex; gap: 2.5rem; align-items: center; max-width: 1100px; width: 100%; }
39+
.row--center { justify-content: center; flex-wrap: wrap; }
40+
.board-img {
41+
width: 280px; border-radius: 0.5rem;
42+
box-shadow: 0 8px 32px rgba(21, 87, 153, 0.2);
43+
}
44+
.diagram { max-width: 560px; width: 100%; }
45+
.diagram img { width: 100%; height: auto; }
46+
.scalar-glow {
47+
outline: 3px solid var(--blue);
48+
outline-offset: 6px;
49+
border-radius: 12px;
50+
animation: pulse 2s ease-in-out infinite;
51+
}
52+
@keyframes pulse {
53+
0%, 100% { outline-color: var(--blue); box-shadow: 0 0 0 0 rgba(21, 87, 153, 0.4); }
54+
50% { outline-color: var(--link); box-shadow: 0 0 24px 4px rgba(30, 107, 184, 0.35); }
55+
}
56+
table {
57+
border-collapse: collapse; font-size: 1.1rem; margin: 0 auto;
58+
box-shadow: 0 4px 20px rgba(0,0,0,0.08); border-radius: 0.5rem; overflow: hidden;
59+
}
60+
th, td { padding: 0.75rem 1.25rem; border: 1px solid #e9ebec; text-align: left; }
61+
th { background: #f3f6fa; color: var(--green); font-weight: 600; }
62+
.badge {
63+
display: inline-block; padding: 0.35rem 0.9rem; border-radius: 0.35rem;
64+
font-weight: 700; font-size: 1.5rem; margin-top: 1rem;
65+
background: linear-gradient(120deg, var(--blue), var(--green));
66+
color: var(--white);
67+
}
68+
.stack {
69+
display: flex; flex-direction: column; align-items: center; gap: 0.5rem;
70+
font-size: 1.05rem; font-weight: 600;
71+
}
72+
.stack-box {
73+
padding: 0.6rem 1.5rem; border: 2px solid var(--blue); border-radius: 0.4rem;
74+
min-width: 220px; text-align: center; color: var(--blue);
75+
}
76+
.stack-arrow { color: var(--text); font-size: 1.2rem; }
77+
.stack-box--highlight { background: rgba(21, 121, 87, 0.12); border-color: var(--green); color: var(--green); }
78+
.end-links { text-align: center; margin-top: 1.5rem; }
79+
.end-links a { color: var(--link); text-decoration: none; font-size: 1.05rem; display: block; margin: 0.35rem 0; }
80+
.end-links a:hover { text-decoration: underline; }
81+
.hint {
82+
position: fixed; bottom: 1rem; right: 1.25rem;
83+
font-size: 0.8rem; color: rgba(255,255,255,0.5);
84+
z-index: 10; pointer-events: none;
85+
}
86+
.slide:not(.slide--hero) .hint { color: #999; }
87+
.counter {
88+
position: fixed; bottom: 1rem; left: 1.25rem;
89+
font-size: 0.85rem; color: #999; z-index: 10;
90+
}
91+
.before { color: #819198; }
92+
.after { color: var(--green); font-weight: 700; }
93+
ul { margin: 1rem 0 0 1.5rem; }
94+
ul li { margin: 0.4rem 0; }
95+
</style>
96+
</head>
97+
<body>
98+
<div class="deck" id="deck">
99+
100+
<!-- Slide 1: Title -->
101+
<section class="slide slide--hero active" data-duration="6">
102+
<h1>OpenSolvers</h1>
103+
<p>U74 progress on the VisionFive 2 — scalar BLAS tuning through EESSI</p>
104+
<p style="margin-top:1.5rem;font-size:1rem;opacity:0.7">RISC-V learnings and fun</p>
105+
</section>
106+
107+
<!-- Slide 2: Board + Scalar -->
108+
<section class="slide" data-duration="8">
109+
<h2>VisionFive 2 — Scalar path (RV64GC)</h2>
110+
<div class="row row--center">
111+
<img class="board-img" src="../../boards/vf2.jpg" alt="StarFive VisionFive 2 board">
112+
<div>
113+
<ul>
114+
<li><strong>JH7110</strong> — 4× SiFive U74 @ 1.5 GHz</li>
115+
<li><strong>rv64gc</strong> — scalar, no RVV</li>
116+
<li>Optimisation target: tuned <strong>BLAS</strong>, not vector extensions</li>
117+
</ul>
118+
<div class="diagram scalar-glow" style="margin-top:1.25rem">
119+
<img src="../../assets/images/compute-backends.svg" alt="Compute paths — Scalar highlighted">
120+
</div>
121+
</div>
122+
</div>
123+
</section>
124+
125+
<!-- Slide 3: Before -->
126+
<section class="slide" data-duration="14">
127+
<h2>Before — stock OpenBLAS</h2>
128+
<p style="text-align:center;margin-bottom:1.25rem;max-width:36rem">
129+
Generic <code>RISCV64_GENERIC</code> C GEMM kernel — correct, but slow
130+
</p>
131+
<table>
132+
<thead><tr><th></th><th>HPL (4 cores, N=10000)</th></tr></thead>
133+
<tbody>
134+
<tr><td><strong>Before</strong></td><td class="before"><strong>3.13 GFLOP/s</strong> · 213 s</td></tr>
135+
<tr><td><strong>After</strong></td><td class="after">—</td></tr>
136+
</tbody>
137+
</table>
138+
</section>
139+
140+
<!-- Slide 4: Fix stack -->
141+
<section class="slide" data-duration="10">
142+
<h2>The fix — one backend swap</h2>
143+
<div class="row row--center">
144+
<div class="stack">
145+
<div class="stack-box">EESSI · HPL 2.3</div>
146+
<div class="stack-arrow">↓</div>
147+
<div class="stack-box">FlexiBLAS (runtime swap)</div>
148+
<div class="stack-arrow">↓</div>
149+
<div class="stack-box stack-box--highlight">OpenBLAS TARGET=U74<br><span style="font-weight:400;font-size:0.9rem">4×4 DGEMM micro-kernel</span></div>
150+
</div>
151+
<div style="max-width:22rem">
152+
<ul>
153+
<li>Same <strong>xhpl</strong> binary</li>
154+
<li>No HPL rebuild</li>
155+
<li>easyconfigs#26436 · OpenBLAS#5903</li>
156+
</ul>
157+
</div>
158+
</div>
159+
</section>
160+
161+
<!-- Slide 5: After + numbers -->
162+
<section class="slide" data-duration="13">
163+
<h2>After — U74-tuned OpenBLAS</h2>
164+
<table style="margin-bottom:1rem">
165+
<thead><tr><th>Metric</th><th>Before</th><th>After</th></tr></thead>
166+
<tbody>
167+
<tr><td>HPL (4 cores)</td><td class="before">3.13 GFLOP/s</td><td class="after">5.28 GFLOP/s</td></tr>
168+
<tr><td>Wall time</td><td class="before">213 s</td><td class="after">126 s</td></tr>
169+
<tr><td>DGEMM (1 core)</td><td class="before">~1.4 GFLOP/s</td><td class="after">1.77 GFLOP/s</td></tr>
170+
<tr><td>DGEMM (4 cores)</td><td class="before">—</td><td class="after">6.31 GFLOP/s</td></tr>
171+
</tbody>
172+
</table>
173+
<div style="text-align:center"><span class="badge">1.69× faster HPL</span></div>
174+
<p style="text-align:center;margin-top:1rem;font-size:0.95rem;max-width:40rem">
175+
U74: silicon-tuned <strong>scalar</strong>. X60 boards: RVV + IME — same “one backend” methodology.
176+
</p>
177+
</section>
178+
179+
<!-- Slide 6: End card -->
180+
<section class="slide slide--hero" data-duration="10">
181+
<h1>opensolvers.com</h1>
182+
<p>Full walkthroughs, benchmarks, and board notes</p>
183+
<div class="end-links">
184+
<a href="https://www.opensolvers.com/boards/VisionFive2.html">VisionFive 2 · HPL 3.13 → 5.28 GFLOP/s</a>
185+
<a href="https://github.com/EESSI/docs/pull/818">EESSI/docs#818 — U74 via EESSI</a>
186+
<a href="https://github.com/easybuilders/easybuild-easyconfigs/pull/26436">easyconfigs#26436 — U74 OpenBLAS</a>
187+
</div>
188+
</section>
189+
190+
</div>
191+
192+
<div class="counter" id="counter">1 / 6</div>
193+
<div class="hint">← → or Space · <kbd>F</kbd> fullscreen · <kbd>?</kbd>autoplay</div>
194+
195+
<script>
196+
const slides = [...document.querySelectorAll(".slide")];
197+
const counter = document.getElementById("counter");
198+
let idx = 0;
199+
let autoplayTimer = null;
200+
201+
function show(i) {
202+
idx = (i + slides.length) % slides.length;
203+
slides.forEach((s, n) => s.classList.toggle("active", n === idx));
204+
counter.textContent = `${idx + 1} / ${slides.length}`;
205+
}
206+
207+
function next() { show(idx + 1); }
208+
function prev() { show(idx - 1); }
209+
210+
function startAutoplay() {
211+
stopAutoplay();
212+
const run = () => {
213+
const dur = (parseInt(slides[idx].dataset.duration, 10) || 8) * 1000;
214+
autoplayTimer = setTimeout(() => { next(); run(); }, dur);
215+
};
216+
run();
217+
}
218+
function stopAutoplay() {
219+
if (autoplayTimer) clearTimeout(autoplayTimer);
220+
autoplayTimer = null;
221+
}
222+
223+
document.addEventListener("keydown", (e) => {
224+
if (e.key === "ArrowRight" || e.key === " ") { e.preventDefault(); stopAutoplay(); next(); }
225+
if (e.key === "ArrowLeft") { e.preventDefault(); stopAutoplay(); prev(); }
226+
if (e.key === "f" || e.key === "F") {
227+
if (!document.fullscreenElement) document.documentElement.requestFullscreen();
228+
else document.exitFullscreen();
229+
}
230+
});
231+
232+
if (location.search.includes("autoplay=1")) startAutoplay();
233+
show(0);
234+
</script>
235+
</body>
236+
</html>

‎video/u74/storyboard.md‎

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
# Storyboard — hybrid (motion + terminal)
2+
3+
**Total runtime:** ~75 s · **Aspect:** 16:9 · **No render yet**
4+
5+
Legend: **M** = motion (record `slides/index.html`) · **T** = terminal (record on VF2)
6+
7+
| # | Time | Type | Visual | Audio (script section) | Notes |
8+
| - | ---- | ---- | ------ | ---------------------- | ----- |
9+
| 1 | 0:00–0:06 | **M** | Slide 1 — title + tagline | “OpenSolvers benchmarks…” | Fade in from black |
10+
| 2 | 0:06–0:14 | **M** | Slide 2 — board + Scalar highlight | “…VisionFive 2… Scalar path” | Pulse highlight on Scalar box |
11+
| 3 | 0:14–0:28 | **M** | Slide 3 — Before 3.13 GFLOP/s | “Out of the box… three point one three” | Animate “Before” row |
12+
| 4 | 0:28–0:38 | **M** | Slide 4 — stack diagram | “We built U74-tuned OpenBLAS…” | Hold while VO explains FlexiBLAS |
13+
| 5 | 0:38–0:52 | **T** | Terminal: EESSI + FlexiBLAS + xhpl snippet | “Same HPL binary… six point three” | See `terminal-recording.md`; cut to best 12–14 s |
14+
| 6 | 0:52–1:05 | **M** | Slide 5 — After 5.28 + speedup table | “End-to-end HPL… one hundred twenty-six” | Animate After row + 1.69× badge |
15+
| 7 | 1:05–1:15 | **M** | Slide 6 — end card + links | “Same methodology… opensolvers dot com” | Hold 3 s after VO ends |
16+
17+
## Edit rhythm
18+
19+
- **Crossfade** M↔T at shot 4→5 and 5→6 (0.3 s).
20+
- **Lower third** optional on terminal clip: `VisionFive 2 · EESSI 2025.06 · FlexiBLAS swap`.
21+
- **Music:** optional subtle bed (−18 dB under VO); none in terminal clip.
22+
- **Captions:** burn-in key numbers (3.13 → 5.28 GFLOP/s) on slide 5.
23+
24+
## Asset checklist before recording
25+
26+
- [ ] VisionFive 2 powered, EESSI/CVMFS working
27+
- [ ] U74 OpenBLAS module or FlexiBLAS backend registered
28+
- [ ] Terminal font ≥ 18 pt, dark theme, 1920×1080 display scaling 100%
29+
- [ ] `slides/index.html` tested fullscreen in Chrome/Safari
30+
- [ ] Mic test; room tone sample for noise reduction
31+
32+
## Deferred (not in this pass)
33+
34+
- Remotion project / automated render
35+
- Background music file
36+
- YouTube thumbnail

0 commit comments

Comments
 (0)