Sync HPL and BLAS from benchmarks repo; add LAPACK and ELPA.

Add scientific-libs pages from opensolvers/benchmarks, expand HPL A/B
tables for Orange Pi RV2, and add website.md to track site decisions.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: hugomeiland

_config.yml

@@ -21,3 +21,7 @@ navigation_apps:
 navigation_scientific_libs:
   - title: BLAS
     url: /scientific-libs/blas.html
+  - title: LAPACK
+    url: /scientific-libs/lapack.html
+  - title: ELPA
+    url: /scientific-libs/elpa.html

apps/hpl.md

@@ -1,19 +1,50 @@
 # HPL results overview
 
-Cross-board summary of **High Performance Linpack (HPL)** runs on consumer RISC-V hardware through the EESSI stack.
+Cross-board summary of **High Performance Linpack (HPL)** on consumer RISC-V hardware through the EESSI stack. Configs and an A/B runner (`run-hpl-ab.sh`) are in [opensolvers/benchmarks/hpl](https://github.com/opensolvers/benchmarks/tree/main/hpl).
 
-**Toolchain:** GCC 14.3.0, OpenBLAS 0.3.30, HPL 2.3.0. EESSI runs use `2025.06-001` on [`dev.eessi.io/riscv`](https://www.eessi.io/docs/repositories/dev.eessi.io-riscv/).
+**Toolchain:** GCC 14.3.0, OpenBLAS 0.3.30, HPL 2.3.0. EESSI `2025.06-001` on [`dev.eessi.io/riscv`](https://www.eessi.io/docs/repositories/dev.eessi.io-riscv/).
+
+## Cross-board summary
 
 | Board | Cores | Before | After |
 | ----- | ----- | ------ | ----- |
-| [StarFive VisionFive 2](../boards/VisionFive2.html) | 4× SiFive U74 | 3.13 GFLOP/s | **5.28 GFLOP/s** |
+| [VisionFive 2](../boards/VisionFive2.html) | 4× SiFive U74 | 3.13 GFLOP/s | **5.28 GFLOP/s** |
 | [Orange Pi RV2](../boards/RV2.html) | 8× SpacemiT X60 | FAILED (`nan`) | **10.53 GFLOP/s** |
 | [Banana Pi F3](../boards/F3.html) | 8× SpacemiT X60 | FAILED (`nan`) | — |
 
-**Before** — stock EESSI OpenBLAS 0.3.30 on each board.
+**Before** — stock EESSI OpenBLAS 0.3.30 (same `xhpl` binary throughout).
+
+**After** — fixed OpenBLAS via EasyBuild + FlexiBLAS backend swap (no HPL rebuild). See [BLAS overview](../scientific-libs/blas.html).
+
+## Orange Pi RV2 — detailed results
+
+Stock EESSI dispatches RVV `ZVL256B` on the X60, but the unpatched `gemv_n` kernel makes HPL report ~8.5 GFLOP/s while **failing** the residual check (`nan`). With the [easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444) fix, all runs below **PASSED**.
+
+### EESSI walkthrough (fixed backend)
+
+| Config | Grid | N | Result |
+| ------ | ---- | - | ------ |
+| Stock EESSI, default RVV | 1×8 | 8000 | ~8.5 GFLOP/s, **FAILED** (`nan`) |
+| Fixed RVV, peak | 2×4 | 20000 | **10.53 GFLOP/s**, PASSED |
+
+Walkthrough: [EESSI/docs#819](https://github.com/EESSI/docs/pull/819).
+
+### A/B — scalar vs patched RVV (`run-hpl-ab.sh`)
+
+Same `xhpl`, backend swapped via FlexiBLAS. All **PASSED** with the fixed vector library:
+
+| Config | Scalar (`RISCV64_GENERIC`) | Patched RVV (`ZVL256B`) | Speedup |
+| ------ | -------------------------- | ----------------------- | ------- |
+| `HPL.dat` (N=8000, 1×8) | 6.41 GFLOP/s | 11.55 GFLOP/s | **1.80×** |
+| `HPL_big.dat` (N=28672, 1×8) | 7.38 GFLOP/s | 13.41 GFLOP/s | **1.82×** |
+| `HPL-sweep.dat` (N=20000, 2×4) | — | ~10.5 GFLOP/s | — |
+
+A squarer **2×4** grid beats 1×8 for peak throughput. `HPL_big.dat` needs ~6.6 GB RAM — tight on 8 GB boards.
+
+## VisionFive 2
 
-**After** — board-specific fixed OpenBLAS via EasyBuild, swapped in at runtime with FlexiBLAS (no HPL rebuild). Details in the [BLAS overview](../scientific-libs/blas.html).
+Scalar U74 — stock OpenBLAS uses a generic kernel; the U74-tuned build lifts HPL **3.13 → 5.28 GFLOP/s** (**1.69×**). Walkthrough: [EESSI/docs#818](https://github.com/EESSI/docs/pull/818).
 
-On **VisionFive 2** (scalar U74), stock OpenBLAS uses a generic kernel; the fix adds a tuned U74 micro-kernel ([easyconfigs#26436](https://github.com/easybuilders/easybuild-easyconfigs/pull/26436), [EESSI/docs#818](https://github.com/EESSI/docs/pull/818)).
+## Banana Pi F3
 
-On **X60 boards** (RVV 1.0, VLEN=256), stock OpenBLAS already dispatches RVV kernels but **0.3.30's `gemv_n` produces NaN**, so HPL fails its residual check despite reporting plausible GFLOP/s. The fix backports the upstream `gemv_n` correction ([easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444), [OpenBLAS#5408](https://github.com/OpenMathLib/OpenBLAS/pull/5408), [EESSI/docs#819](https://github.com/EESSI/docs/pull/819)). Banana Pi F3 uses the same K1/X60 SoC — the same fix applies; peak result not yet recorded.
+Same K1 / X60 SoC as the Orange Pi RV2 — same stock failure and same [easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444) fix applies. Peak HPL not yet recorded.

boards/RV2.md

@@ -13,6 +13,8 @@ End-to-end on real Orange Pi RV2 hardware using [EESSI](https://www.eessi.io/) `
 | | Before | After |
 | --- | ------ | ----- |
 | HPL (8 cores, N=20000, 2×4) | ~8.5 GFLOP/s, **FAILED** (`nan`) | **10.53 GFLOP/s**, **PASSED** |
+
+With the fixed backend, scalar-vs-RVV A/B ([benchmarks/hpl](https://github.com/opensolvers/benchmarks/tree/main/hpl)): **6.41 → 11.55 GFLOP/s** (N=8000, 1×8) and **7.38 → 13.41 GFLOP/s** (N=28672, 1×8).
 | Residual (N=8000, 1×8) | `nan` | 4.04e-03 |
 
 **Before** — stock EESSI OpenBLAS 0.3.30 (RVV `gemv_n` bug). **After** — fixed OpenBLAS built with `TARGET=RISCV64_ZVL256B` and a backported `gemv_n` patch ([easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444)), swapped in via FlexiBLAS — no HPL rebuild.

scientific-libs/blas.md

@@ -1,6 +1,6 @@
 # BLAS (OpenBLAS)
 
-Improvements to **OpenBLAS 0.3.30** on RISC-V boards in this project — built via EasyBuild, deployed through [EESSI](https://www.eessi.io/) with **FlexiBLAS** runtime swapping. See also the [HPL app](../apps/hpl.html) for end-to-end benchmark results.
+Improvements to **OpenBLAS 0.3.30** on RISC-V boards — built via EasyBuild, deployed through [EESSI](https://www.eessi.io/) with **FlexiBLAS** runtime swapping. Verification tools and A/B scripts live in the [opensolvers/benchmarks](https://github.com/opensolvers/benchmarks) repo (`dgemm/`, `numpy/`).
 
 **Base stack:** GCC 14.3.0, OpenBLAS 0.3.30, EESSI `2025.06-001` ([`dev.eessi.io/riscv`](https://www.eessi.io/docs/repositories/dev.eessi.io-riscv/)).
 
@@ -9,8 +9,30 @@ Improvements to **OpenBLAS 0.3.30** on RISC-V boards in this project — built v
 | Board / CPU | Problem (stock 0.3.30) | Fix | Result |
 | ----------- | ---------------------- | --- | ------ |
 | [VisionFive 2](../boards/VisionFive2.html) — SiFive **U74** (scalar) | No U74 kernel; falls back to generic `RISCV64_GENERIC` C `2×2` GEMM | New **4×4 DGEMM micro-kernel** in RV64 assembly, `TARGET=U74` | Single-core DGEMM **~1.4 → 1.77 GFLOP/s**; 4-core DGEMM **6.31 GFLOP/s**; [HPL **1.69×**](../apps/hpl.html) |
-| [Orange Pi RV2](../boards/RV2.html) — SpacemiT **X60** (RVV VLEN=256) | RVV `gemv_n` zeroes an **uninitialized** vector register → `dgemv` returns NaN; HPL **fails** residual despite ~8.5 GFLOP/s | Backport upstream `gemv_n` fix; `TARGET=RISCV64_ZVL256B` | HPL **PASSED** (residual ~4e-03); peak **10.53 GFLOP/s** |
-| [Banana Pi F3](../boards/F3.html) — same K1 / X60 SoC | Same RVV `gemv_n` bug as RV2 | Same [easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444) fix | Fix applies; peak HPL not yet recorded |
+| [Orange Pi RV2](../boards/RV2.html) — SpacemiT **X60** (RVV VLEN=256) | RVV `gemv_n` zeroes an **uninitialized** vector register → `dgemv` returns NaN | Backport upstream `gemv_n` fix; `TARGET=RISCV64_ZVL256B` | Correct `dgemv`; see [HPL](../apps/hpl.html) and A/B tables below |
+| [Banana Pi F3](../boards/F3.html) — same K1 / X60 SoC | Same RVV `gemv_n` bug as RV2 | Same [easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444) fix | Fix applies; peak results not yet recorded |
+
+## X60 A/B — patched RVV vs scalar (Orange Pi RV2)
+
+All runs use a **fixed** RVV OpenBLAS (`gemv_n` patch) swapped in via FlexiBLAS. Source: [benchmarks/dgemm](https://github.com/opensolvers/benchmarks/tree/main/dgemm), [benchmarks/numpy](https://github.com/opensolvers/benchmarks/tree/main/numpy).
+
+| Probe | Scalar (`RISCV64_GENERIC`) | Patched RVV (`ZVL256B`) | Speedup |
+| ----- | -------------------------- | ----------------------- | ------- |
+| `bench_dgemm` N=2048, 1 core | 1.16 GFLOP/s | 2.62 GFLOP/s | **2.3×** |
+| NumPy `A @ B` N=4096, 8 threads | 4.77 GFLOP/s | 11.52 GFLOP/s | **2.4×** |
+| NumPy `eigvalsh` N=2048, 8 threads | 10.54 s | 6.72 s | **1.6×** |
+
+## X60 — isolating the `gemv_n` bug (`difftest`)
+
+The [benchmarks `difftest`](https://github.com/opensolvers/benchmarks/blob/main/dgemm/difftest.c) tool `dlopen`s a BLAS `.so` and reports NaN counts per routine:
+
+| Backend | `dgemv` NaN | `dgemm` NaN | `dgemv` sum |
+| ------- | ----------- | ----------- | ----------- |
+| Stock EESSI, default RVV | **192** | 0 | 198.94 (wrong) |
+| Stock EESSI, forced scalar | 0 | 0 | 42.06549 (reference) |
+| Patched RVV (`gemv_n` fix) | 0 | 0 | 42.06549 (matches) |
+
+`dgemm` looks fine on the broken build — which is why a raw GEMM benchmark can pass while [HPL](../apps/hpl.html) and [ELPA](elpa.html) fail.
 
 ## Packages
 
@@ -19,10 +41,10 @@ Improvements to **OpenBLAS 0.3.30** on RISC-V boards in this project — built v
 | SiFive U74 | [easyconfigs#26436](https://github.com/easybuilders/easybuild-easyconfigs/pull/26436) | [OpenBLAS#5903](https://github.com/OpenMathLib/OpenBLAS/pull/5903) | [EESSI/docs#818](https://github.com/EESSI/docs/pull/818) |
 | SpacemiT X60 | [easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444) | [OpenBLAS#5408](https://github.com/OpenMathLib/OpenBLAS/pull/5408), [#5476](https://github.com/OpenMathLib/OpenBLAS/pull/5476) | [EESSI/docs#819](https://github.com/EESSI/docs/pull/819) |
 
-Build from a PR with `eb --from-pr <num> --robot` into an **EESSI-extend** user install, then register with `flexiblas add` / `flexiblas default` — no need to rebuild downstream apps like HPL.
+Build with `eb --from-pr <num> --robot` into an **EESSI-extend** user install, then `flexiblas add` / `flexiblas default` — no downstream rebuild.
 
 ## Notes
 
 - **U74** — performance kernel; stock OpenBLAS works but leaves FP throughput on the table.
-- **X60** — correctness fix first: the stock EESSI `DYNAMIC_ARCH` build *does* dispatch RVV `ZVL256B` kernels, but 0.3.30's broken `gemv_n` makes HPL silently wrong. OpenBLAS **≥ 0.3.34** should carry the fix natively and improve RVV performance further.
-- On the experimental `dev.eessi.io/riscv` toolchain, pin a valid `-march` via `EASYBUILD_OPTARCH` when building (see board walkthroughs).
+- **X60** — correctness fix first. Stock EESSI `DYNAMIC_ARCH` *does* dispatch RVV, but 0.3.30's broken `gemv_n` corrupts BLAS-2 paths. OpenBLAS **≥ 0.3.34** should fix this natively.
+- Pin a valid `-march` via `EASYBUILD_OPTARCH` on the experimental `dev.eessi.io/riscv` toolchain (see board walkthroughs).

scientific-libs/elpa.md

@@ -0,0 +1,21 @@
+# ELPA
+
+[ELPA](https://elpa.mpcdf.mpg.de/) is a dense real-symmetric eigensolver used in DFT codes (CP2K, VASP, ELSI). It exercises a **mix of BLAS-2 and BLAS-3** (`dsymv`, `dsyr2k`, `dgemm`, `dtrmm`) — a more representative BLAS probe than raw `dgemm`.
+
+Benchmark source: [opensolvers/benchmarks/elpa](https://github.com/opensolvers/benchmarks/tree/main/elpa).
+
+## Orange Pi RV2 (SpaceMiT X60, 8 cores)
+
+`na=3000`, 1-stage solver, 8 threads, single MPI rank. Backends swapped via FlexiBLAS — no rebuild.
+
+| Backend | Time | Correctness |
+| ------- | ---- | ----------- |
+| Stock EESSI OpenBLAS 0.3.30, default RVV | 42.17 s | **`ev0=nan`, finite=0 — FAIL** |
+| Scalar (`RISCV64_GENERIC`) | 54.92 s | finite=1 |
+| Patched RVV (`gemv_n` fix) | **34.81 s** | finite=1 |
+
+The stock vector backend is faster than scalar but **wrong** — the same OpenBLAS 0.3.30 `gemv_n` NaN bug that breaks [HPL](../apps/hpl.html). With the patch, RVV is **1.58×** faster than scalar (54.92 / 34.81) and numerically correct.
+
+Speedup is smaller than pure `dgemm` (~2.3×) because tridiagonalization is latency-bound BLAS-2 — by design.
+
+See also the [BLAS overview](blas.html) and [LAPACK](lapack.html) NumPy probe.

scientific-libs/lapack.md

@@ -0,0 +1,18 @@
+# LAPACK
+
+LAPACK correctness and performance here are probed through **NumPy** (`eigvalsh` → `dsyevd`) and pure BLAS (`A @ B` → `dgemm`) — a real-application path for SciPy/NumPy stacks on EESSI.
+
+Benchmark source: [opensolvers/benchmarks/numpy](https://github.com/opensolvers/benchmarks/tree/main/numpy).
+
+## Orange Pi RV2 (SpaceMiT X60, 8 threads)
+
+Backends swapped via FlexiBLAS on a NumPy linked against the FlexiBLAS hub (`SciPy-bundle` module). Patched RVV OpenBLAS uses the `gemv_n` fix from [easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444).
+
+| Kernel | Scalar (`RISCV64_GENERIC`) | Patched RVV (`ZVL256B`) | Speedup |
+| ------ | -------------------------- | ----------------------- | ------- |
+| DGEMM (`A @ B`, N=4096) | 4.77 GFLOP/s | 11.52 GFLOP/s | **2.4×** |
+| EIGH (`eigvalsh`, N=2048) | 10.54 s | 6.72 s | **1.6×** |
+
+Both patched runs are **finite** (correct through real LAPACK). The eigensolver speedup is lower than DGEMM because `eigvalsh` mixes BLAS-3 with BLAS-2 tridiagonalization — the same pattern seen in [ELPA](elpa.html).
+
+Stock EESSI default RVV would return NaN on the eigensolver path; see the [BLAS `difftest` table](blas.html) for how the fault isolates to `dgemv`.

website.md

@@ -0,0 +1,51 @@
+# OpenSolvers website — decisions log
+
+Context file for structural and content decisions on [opensolvers.com](https://www.opensolvers.com) (`opensolvers/opensolvers.github.io`).
+
+**Workflow:** commit and deploy directly to `main` via PR (no need to ask first).
+
+---
+
+## Site structure
+
+| Area | Path | Purpose |
+|------|------|---------|
+| Homepage | `README.md` → `/` | Intro, board summaries |
+| Boards | `boards/` | Per-board hardware + benchmark notes |
+| Apps | `apps/` | End-to-end application benchmarks (e.g. HPL) |
+| Scientific libs | `scientific-libs/` | Library-level probes (BLAS, LAPACK, ELPA) |
+
+## Navigation groups
+
+1. **Home**
+2. **Boards** — VisionFive 2, OrangePi RV2, BananaPi F3
+3. **Apps** — HPL
+4. **Scientific libs** — BLAS, LAPACK, ELPA
+
+Nav config: `_config.yml` (`navigation`, `navigation_boards`, `navigation_apps`, `navigation_scientific_libs`). Rendered in `_includes/header.html`. Cayman theme requires `_layouts/default.html` override to include the header.
+
+## Content sources
+
+| Source | Used for |
+|--------|----------|
+| [opensolvers/benchmarks](https://github.com/opensolvers/benchmarks) | BLAS verification (`dgemm/`), HPL configs (`hpl/`), ELPA (`elpa/`), NumPy/LAPACK (`numpy/`) |
+| [EESSI/docs#818](https://github.com/EESSI/docs/pull/818) | VisionFive 2 / U74 OpenBLAS + HPL |
+| [EESSI/docs#819](https://github.com/EESSI/docs/pull/819) | Orange Pi RV2 / X60 RVV `gemv_n` fix + HPL |
+| [easyconfigs#26436](https://github.com/easybuilders/easybuild-easyconfigs/pull/26436) | U74 OpenBLAS package |
+| [easyconfigs#26444](https://github.com/easybuilders/easybuild-easyconfigs/pull/26444) | X60 OpenBLAS package |
+
+## Key numbering conventions
+
+- **Before / After** tables — stock EESSI OpenBLAS 0.3.30 vs fixed EasyBuild module (FlexiBLAS swap, no app rebuild).
+- **Orange Pi RV2 HPL:** stock EESSI = FAILED (`nan`) ~8.5 GFLOP/s; fixed peak = **10.53 GFLOP/s** (N=20000, 2×4). Scalar A/B baselines from `run-hpl-ab.sh`: **6.41** and **7.38 GFLOP/s** — these are *scalar* backends, not “after” results.
+- **X60 DNF/FAIL** — OpenBLAS 0.3.30 RVV `gemv_n` bug; fixed in ≥ 0.3.31, EESSI bump ≥ 0.3.34 preferred.
+
+## Changelog
+
+| Date | Decision |
+|------|----------|
+| 2026-07-12 | Add `website.md`; sync HPL/BLAS from `opensolvers/benchmarks`; add scientific libs LAPACK + ELPA |
+| 2026-07-12 | Split nav: Boards / Apps / Scientific libs; move HPL to `apps/` |
+| 2026-07-12 | Fix homepage RV2 numbers (FAILED → 10.53 GFLOP/s, not 7.38 native) |
+| 2026-07-11 | Override Cayman `_layouts/default.html` so `_includes/header.html` renders |
+| 2026-07-11 | Site focus: RISC-V learnings; tagline “RISC-V learnings and fun” |