[WebGPU] optimizations for audio models (DFT, STFT, ConvTranspose, LSTM, GRU) (#33978)
### Description
This introduces important optimizations for audio models, which tends to
rely heavily on nodes such as DFT, STFT, ConvTranspose and for some LSTM
& GRU.
### Motivation and Context
WebGPU didn't have those node implemented (such as STFT) or they were
not really optimized (DFT only optimized for sizes <=512, slow
ConvTranspose, LSTM & GRU).
The speed gain was tested across 4 popular audio model architectures:
HTDemucs, UVR, SCNet, MP-SENet on 2 different hardware accelerators:
NVIDIA RTX 3090, and Apple M3.
The measured speed gain before and after these changes:
NVIDIA RTX 3090 on Windows 11 (CPU: Intel 12700K):
Model | WebGPU before | WebGPU after | Gain
-- | -- | -- | -- |
HTDemucs | 1132 ms | 265 ms | 4.3x
SCNet | 12440 ms | 525 ms | 23.7x
UVR | 457 ms | 342 ms | 1.34x
MP-SENet | 203 ms | 41 ms | 4.95x
Apple M3 on macOS 14.7:
Model | WebGPU before | WebGPU after | Gain | Output gap before vs after
-- | -- | -- | -- | --
HTDemucs | 3308 ms | 1436 ms | 2.30× | rel_l2 1.2e-6
SCNet | 38043 ms | 3114 ms | 12.2× | rel_l2 7.7e-6
UVR | 2130 ms | 2014 ms | 1.06× | rel_l2 4.7e-7
MP-SENet | 489 ms | 261 ms | 1.87× | rel_l2 1.4e-6
Op usage:
Model | DFT | STFT | ConvTranspose | LSTM| GRU
-- | -- | -- | -- | -- | --
HTDemucs | ✅ | ✅ | ✅ | –| –
SCNet | ✅ | ✅ | ✅ | ✅| –
UVR | – | – | – | ✅| –
MP-SENet | – | – | – | -| ✅
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>