Improve Windows gpu support

[illuzen]

Add comprehensive GPU detection support for AMD, NVIDIA, Intel, and Qualcomm

Overview

This PR significantly expands GPU detection and tier configuration to support a much wider range of graphics hardware. It addresses user reports of unrecognized GPUs (specifically Vega 8 APU and RX 560X) and proactively adds support for many other common GPUs that were previously falling back to conservative default settings.

What Changed

Bug Fixes

• Fixed critical pattern matching bug: The "mi" pattern for AMD Instinct was matching any GPU name containing "mi" (e.g., "Family", "Graphics"). Changed to specific patterns: "instinct mi", "mi100", "mi200", "mi250", "mi300"
• Fixed RX 560X misdetection: RX 560X was incorrectly matching the "5600" pattern (RDNA 1) instead of Polaris. Reordered detection to check Polaris patterns before RDNA 1

NVIDIA Additions

Tier	Models	Workgroup Formula
GTX 900 (Maxwell)	980, 970, 960, 950	max/18, min 768
GTX 700 (Kepler/Maxwell)	780, 770, 760, 750	max/20, min 512
GTX Legacy (Fermi/Kepler)	GTX 600/500/400 series	max/24, min 384
MX (Mobile)	MX550, MX450, MX350, etc.	max/24, min 384
GT (Entry-Level)	GeForce GT series	max/28, min 256
Professional	Added A100, H100, L4 datacenter GPUs	max/10, min 2560

Also added missing patterns: RTX 3060, 3050, GTX 1050, 1030

AMD Additions

Tier	Models	Workgroup Formula
Radeon 780M (RDNA 3 APU)	780M	max/12, min 2048
Radeon 7x0M (RDNA 3 APU)	760M, 740M	max/16, min 1024
Radeon 680M (RDNA 2 APU)	680M	max/16, min 1536
Radeon 6x0M (RDNA 2 APU)	660M, 610M	max/22, min 768
RX 6500/6400 (Entry RDNA 2)	6500 XT, 6400	max/22, min 512
Radeon VII (Vega 20)	Radeon VII	max/12, min 2048
Vega 64 (Discrete)	Vega 64	max/14, min 1536
Vega 56 (Discrete)	Vega 56	max/16, min 1280
Vega (APU)	Vega 8, Vega 11, etc.	max/28, min 384
R9 Fury/Nano (Fiji)	Fury X, Fury, Nano	max/16, min 1280
R9 (GCN)	390, 380, 290, 280	max/20, min 768
R7 (GCN)	370, 360, 270, 260	max/22, min 512
Radeon OEM	Radeon 600/700 (rebadged Polaris)	max/24, min 512
Radeon Graphics (APU)	Generic APU fallback	max/26, min 384

Also added: RX 590, RX 480/470/460, RX 6950/6750/6650 patterns

Intel Additions

Tier	Models	Workgroup Formula
Arc A5 (Desktop)	A580	max/14, min 1536
Arc A3 (Desktop)	A380, A310	max/18, min 768
Arc A7 Mobile	A770M, A730M	max/14, min 1536
Arc A5 Mobile	A550M, A570M	max/16, min 1024
Arc A3 Mobile	A370M, A350M	max/20, min 512
Iris Xe Max (Discrete)	DG1-based	max/20, min 512
Iris Pro	Haswell/Broadwell	max/26, min 256
Iris Plus	Ice Lake, etc.	max/26, min 320
UHD 700	Alder Lake+	max/26, min 320
UHD 600	Coffee Lake, Comet Lake	max/28, min 256
HD Graphics	Older generations	max/30, min 192

Qualcomm (New Vendor)

Tier	Models	Workgroup Formula
Adreno X1 (Snapdragon X)	X Elite, X Plus	max/14, min 1536
Adreno 700	Snapdragon 8 Gen 1/2/3	max/16, min 1024
Adreno 600	Snapdragon 800 series	max/20, min 512
Adreno 500	Older Snapdragon	max/24, min 384

Validation

• cargo check -p engine-gpu - passes
• cargo clippy -p engine-gpu -- -D warnings - passes, no warnings

Risks and Mitigations

• Risk: New patterns could potentially match unintended GPU names

  • Mitigation: Patterns are ordered from most specific to least specific; more conservative fallbacks are used for ambiguous cases

• Risk: Workgroup formulas for new tiers may not be optimal

  • Mitigation: Values are based on relative GPU compute capabilities; users can still override via CLI flags if needed; fallback detection still triggers a warning asking users to report unrecognized GPUs

Testing Notes

The originally reported GPUs should now be detected as:

• Radeon(TM) Vega 8 Graphics → AMD Vega (APU) tier
• Radeon RX 560X → AMD RX 500/400 (Polaris) tier (previously misdetected as RDNA 1)

---

Note

Medium Risk
Large heuristic-only change: mis-matched adapter name substrings could pick wrong workgroup limits (performance/stability), though users can still override via CLI and unknown GPUs still log fallback warnings.

Overview
Expands get_vendor_specific_dispatch in engine-gpu so more adapters get a named tier and tuned optimal_workgroups instead of generic fallbacks.

NVIDIA gains extra string matches (e.g. RTX 3060/3050, GTX 1050/1030) and new branches for GTX 900/700/legacy, MX mobile, GT entry-level, and datacenter names (A100/H100/L4).

AMD detection is reordered and broadened: Polaris is checked before RDNA 1 so names like RX 560X no longer hit 5600-style patterns; Instinct matching drops the loose "mi" substring in favor of explicit mi100/instinct mi-style patterns. New tiers cover RDNA APUs, RX 6500/6400, Vega (discrete and APU), GCN R9/R7, OEM rebadges, and generic Radeon Graphics APUs.

Intel splits Arc into finer desktop/mobile buckets and adds more integrated paths (Iris variants, UHD 700/600, HD). A new Qualcomm Adreno block (vendor ID + name patterns) covers Snapdragon X and Adreno 5/6/7 series on Windows ARM.

^{Reviewed by Cursor Bugbot for commit 911463c. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 911463c. Configure here.}

[n13]

Review: Improve Windows GPU support

Reviewed at head 3656655. Pulled the branch and ran the repo's validation checklist locally — all green:

• cargo test -p engine-gpu → 5 passed (the new gpu_tiers tests)
• cargo clippy -p engine-gpu --all-targets -- -D warnings → clean
• cargo fmt --all -- --check → clean
• cargo check --workspace --locked → clean

Verdict: Approve (with non-blocking follow-ups)

Solid net improvement that meets its stated goal. The big win is replacing the brittle String::contains heuristics with a table-driven, word-boundary regex approach that is much more maintainable and DRY (per-tier workgroup math is now one centralized expression instead of duplicated (max/N).max(M) everywhere).

The two Bugbot findings are resolved

Both inline comments were filed against the older contains() commit 911463c. The later refactor into gpu_tiers.rs fixes both, and there are now unit tests proving it:

• Polaris vs RDNA1: \b…\b boundaries + RDNA checked before Polaris → RX 5500/5600/5700 map to RDNA1 and RX 560X/580/550 map to Polaris. Verified by test_amd_rdna_vs_polaris.
• Arc mobile vs desktop: mobile tiers are checked first and \ba7[57]0\b won't match a770m. Verified by test_intel_arc_mobile_vs_desktop.

I spot-checked extra cases not covered by the tests: RTX A6000 → Professional (not swallowed by consumer tiers), the GT tier does not match GTX names, and RX 6500 XT → RDNA2 Entry. All correct.

Strong positive: fixes a real silent hang

The previous buffer-map path only set its flag on success and looped forever on a map error with no logging — a silent infinite hang, which violates the project's fail-early / always-log rule. The new AtomicU8 state (pending/success/error) + 30s poll timeout logs and bails out. Good catch.

Non-blocking follow-ups

1. unmap() on a non-mapped buffer (crates/engine-gpu/src/lib.rs ~L664 and ~L675): on the error path the buffer was never mapped, and on the timeout path the map is still pending. wgpu 27 flags this as a validation error (noisy, not a hard panic). Suggest guarding unmap() so it only runs when the buffer is actually mapped.
2. Persistent device-loss spins: returning NotFound { hash_count: 0 } is the right move for a transient failure, but on a genuinely lost device every batch will log-error and burn cycles at 0 H/s indefinitely. Consider tearing down / removing the dead context so the engine fails loudly, consistent with the repo's fail-early guidance.
3. Init "timeout" is post-hoc: init_start.elapsed() is checked after request_device().await returns (the comment acknowledges this), so a driver that truly hangs inside the await still blocks. It only guards slow-but-completed init. Fine as-is; if true hang-protection is the goal it needs to race the future against a timer / off-thread init.

Nits

• Qualcomm tiers use greedy adreno.*7 / adreno.*6, so e.g. "Adreno 627" buckets as 700-series. Tuning-only and Qualcomm-scoped — low impact.
• Duplicate "AMD RX 5000 (RDNA 1)" tier (\b5[56]00\b and rx 5\d{3}, identical params) could be merged into a single pattern.
• regex + once_cell are added as direct deps; once_cell::sync::Lazy could be std::sync::LazyLock (Rust ≥ 1.80) if you want to keep the dependency surface minimal.

None of the above blocks merge.