wouter/d4r
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Correct M attention precision and validate every temporal transformer block

Browse Source
This commit is contained in:
Codex committed 2026-09-30 21:30:05 +03:00
1 parent e15f6a73a6
commit b8d801ae4c
14 files changed
+623 -18
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

docs/windows-rdna4-port.md
+67 -4
View File
@@ -13,8 +13,12 @@ frames on Windows. M now passes all five synthetic and real-weight layers agains
NumPy and independent PTX, and four native CUDA frames. The standalone D3D12
harness now produces four K and four M frames using imported VRAM and shared
fences, with bit-exact RGB agreement against the corresponding CUDA harnesses.
Game command-list/queue integration remains in progress. The fully translated M
comparison is being rerun with the correct WMMA options and a writable JIT cache.
Game command-list/queue integration remains in progress. A full temporal M scan
found rare attention P/V accumulation mismatches that the earlier twelve-block
sampling missed. `SWIN_EXACT_PV` fixes the native baseline; ZLUDA patch `0013`
adds the corresponding opt-in f16 reference accumulation. All spatial blocks in
40 actual native M launches now exactly match NumPy. Four complete native M RGB
frames exactly match fully translated M with the corrected reference option.
Do not integrate NGX until the integer PTX workload is stable on the real GPU.
| Gate | Status |
@@ -25,11 +29,70 @@ Do not integrate NGX until the integer PTX workload is stable on the real GPU.
| M3: Windows NGX initialization and simple DLSS path | Init, SR capabilities, Create/Evaluate PASS; E, K and M CUDA harnesses execute |
| M4: D3D12 / HIP external memory and fence round trip | PASS with TheRock; stable 7.2 has mapped-view leak |
| M5: independently validated gfx12 WMMA backend | PASS: raw + legacy adapter + upstream layout, max abs/relative error 0 |
| M6: K layers, full transformer and image validation | CUDA harness PASS: 11 real-weight layers, native launches for all layers, four finite frames, 81.85–92.91 dB versus translated K; D3D12 integration pending |
| M7: M FP16-equivalent baseline and full transformer | CUDA harness PASS: 5 real-weight layers exactly match NumPy/PTX, 4 finite frames, 40 native Swin launches; D3D12 and full translated M comparison pending |
| M6: K layers, full transformer and image validation | PASS: 11 real-weight layers, 44 native launches, four D3D12 frames bit-exact against native CUDA, 81.85–92.91 dB versus translated K |
| M7: M FP16-equivalent baseline and full transformer | PASS: all blocks of 40 native launches exactly match NumPy; targeted independent PTX regression matches; four D3D12 frames bit-exact against corrected translated M |
| M8: standalone Windows D3D12 NGX harness | PASS at explicit submitted queue boundaries: K/M, four frames each, no shim CPU image copies, bit-exact RGB against CUDA harness |
| M9: OptiScaler integration, profiling, installation | Pending |
## M temporal precision regression (2026-09-30)
The earlier finite-frame and sampled-layer results did not cover every spatial
block. The second enc1 launch has a cancellation-sensitive P/V dot in block
`0,4`: the old native f32 WMMA route differed by up to 21 e4m3 code steps.
NumPy and the independent PTX interpreter agree exactly on that block.
`SWIN_EXACT_PV` stores each wave's sixteen P rows and V rows in LDS, sums sixteen
f16 products in f64, rounds directly to f16, then applies the existing e4m3
quantisation. Other GEMMs retain the gfx12 WMMA layout and FP16 widening.
The flag is enabled by `D4R_M_FP16_BASELINE`; upstream fast/Linux defaults are
preserved when that flag is absent. Intermediate snapshots are compiled only
with `D4R_SWIN_DIAGNOSTICS`; they are absent from packaged kernels.
The translated reference also differed: first frame, fourth enc3_tube launch,
block `0,0`, thirty stored codes, maximum fifteen e4m3 steps. Its inputs match
the corrected native launch exactly. Patch `0013` adds
`D4R_ZLUDA_WMMA_F16_REFERENCE=1` for gfx12 f16-accumulator MMA helpers, retaining
the established fragment gathers, pair fusion and conversions. It uses f64
sums and integer f16 RNE encoding, including subnormal bits. FP8 GEMMs continue
using the existing native WMMA widening. The default upstream helper behavior
remains available with this option unset or zero. Windows diagnostics select
the reference option by default. All d4r codegen options enter the JIT cache key.
This is a match to the independent mathematical NumPy/PTX baseline. NVIDIA's
[PTX floating-point MMA specification](https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-instructions-mma)
leaves accumulation order, rounding and subnormal handling unspecified; these
checks do not establish bit-exact agreement with NVIDIA hardware.
Recorded results on the RX 9070 XT / TheRock runtime:
* Public `cuda_mma_probe`: 512 synthetic matrix cases, 65,536 f16 outputs,
zero differing bits with reference accumulation. The old fast WMMA route
differs in 23,308 outputs on the same cancellation/subnormal stress fixture.
* `m-native-exact-pv-temporal-reference`: all spatial blocks of 40 actual native
launches, every stored e4m3 value identical to NumPy, max abs/relative error 0.
* `m-exact-pv-enc1-regression-ptx`: original problem block `0,4`, independent
PTX versus native and versus NumPy both max absolute/relative error 0.
* `m-exact-pv-complete-f16-reference.log`: four complete RGB frames against
corrected fully translated M, max absolute/relative error 0, PSNR infinite.
* `m-translated-f16-temporal-reference`: all blocks of all forty translated
M launches exactly match NumPy; no sampled-window exemption.
* `m-temporal-final-ctest.log`: 11/11 native Windows hardware tests PASS.
Pinned fresh-source application of all twelve ZLUDA patches PASS.
* `d3d12-m-exact-pv-final.zip`: four finite frames, forty native launches,
no shim CPU image copies, frame age zero. This remains the explicit queue
boundary harness; open game command-list support is the next gate.
All captures, extracted PTX and real weights remain in ignored private result
directories. To repeat the full temporal scan without editing sources:
```powershell
$env:PYTHONPATH="$PWD/.tools/python/vendor"
python tools/windows/m_temporal_validate.py --capture-dir test-results/ngx-m-exact-pv-capture/captures --output-dir test-results/m-temporal-reference --exact
```
The game selected for the next integration test is Silent Hill 2 at
`D:\Games\SILENT HILL 2\SHProto\Binaries\Win64\SHProto-Win64-Shipping.exe`.
No game files have been modified at this milestone.
## Source baseline (checked 2026-09-29)
* d4r: `ed1ab60`, main, 14 commits; https://github.com/countervolts/d4r.