oidn-web 0.3.4 → 0.4.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +49 -0
- package/README.md +140 -4
- package/benchmarks/compare.mjs +651 -0
- package/benchmarks/leak.mjs +255 -0
- package/benchmarks/results/before-spatial.json +391 -0
- package/benchmarks/results/before-spatial.md +47 -0
- package/benchmarks/results/int8-scan.json +2007 -0
- package/benchmarks/results/int8-scan.md +160 -0
- package/benchmarks/results/int8-w8a8-scan.json +2007 -0
- package/benchmarks/results/int8-w8a8-scan.md +160 -0
- package/benchmarks/results/int8-weight-channel.json +1413 -0
- package/benchmarks/results/int8-weight-channel.md +118 -0
- package/benchmarks/results/int8-weight-only.json +1437 -0
- package/benchmarks/results/int8-weight-only.md +118 -0
- package/benchmarks/results/kernel-webnn-final.json +1115 -0
- package/benchmarks/results/kernel-webnn-final.md +104 -0
- package/benchmarks/results/latest-optimized.json +375 -0
- package/benchmarks/results/latest-optimized.md +47 -0
- package/benchmarks/results/latest.json +391 -0
- package/benchmarks/results/latest.md +47 -0
- package/benchmarks/results/profile-baseline.json +331 -0
- package/benchmarks/results/profile-baseline.md +12 -0
- package/benchmarks/results/profile-conv2x.json +331 -0
- package/benchmarks/results/profile-conv2x.md +12 -0
- package/benchmarks/results/profile-fast-init.json +385 -0
- package/benchmarks/results/profile-fast-init.md +47 -0
- package/benchmarks/results/profile-fp16-fma.json +369 -0
- package/benchmarks/results/profile-fp16-fma.md +47 -0
- package/benchmarks/results/profile-fp16-tiled-decoder.json +369 -0
- package/benchmarks/results/profile-fp16-tiled-decoder.md +47 -0
- package/benchmarks/results/profile-fp16-tiled-encoder.json +369 -0
- package/benchmarks/results/profile-fp16-tiled-encoder.md +47 -0
- package/benchmarks/results/profile-fp16-unfused-pool.json +385 -0
- package/benchmarks/results/profile-fp16-unfused-pool.md +47 -0
- package/benchmarks/results/profile-input-major.json +347 -0
- package/benchmarks/results/profile-input-major.md +12 -0
- package/benchmarks/results/profile-k16.json +347 -0
- package/benchmarks/results/profile-k16.md +12 -0
- package/benchmarks/results/profile-k4.json +347 -0
- package/benchmarks/results/profile-k4.md +12 -0
- package/benchmarks/results/profile-pool-reuse.json +331 -0
- package/benchmarks/results/profile-pool-reuse.md +12 -0
- package/benchmarks/results/profile-precompiled.json +385 -0
- package/benchmarks/results/profile-precompiled.md +47 -0
- package/benchmarks/results/profile-static-channels.json +385 -0
- package/benchmarks/results/profile-static-channels.md +47 -0
- package/benchmarks/results/profile-static-io.json +385 -0
- package/benchmarks/results/profile-static-io.md +47 -0
- package/benchmarks/results/profile-tiled-conv.json +331 -0
- package/benchmarks/results/profile-tiled-conv.md +12 -0
- package/benchmarks/results/profile-tiled-decoder.json +347 -0
- package/benchmarks/results/profile-tiled-decoder.md +12 -0
- package/benchmarks/results/profile-tiled-matmul.json +331 -0
- package/benchmarks/results/profile-tiled-matmul.md +12 -0
- package/benchmarks/results/profile-unfused-decoder.json +379 -0
- package/benchmarks/results/profile-unfused-decoder.md +12 -0
- package/benchmarks/results/profile-unfused-pool.json +347 -0
- package/benchmarks/results/profile-unfused-pool.md +12 -0
- package/benchmarks/results/spatial-auto.json +575 -0
- package/benchmarks/results/spatial-auto.md +61 -0
- package/benchmarks/results/subgroup-smoke.json +1094 -0
- package/benchmarks/results/subgroup-smoke.md +104 -0
- package/benchmarks/results/webnn-smoke.json +739 -0
- package/benchmarks/results/webnn-smoke.md +76 -0
- package/dist/oidn.js +4189 -22603
- package/dist/oidn.umd.cjs +784 -5807
- package/lib/UNet.d.ts +66 -15
- package/lib/UNet.js +162 -257
- package/lib/UNet.js.map +1 -1
- package/lib/WGPUComputePass.d.ts +1 -1
- package/lib/WGPUComputePass.js +6 -4
- package/lib/WGPUComputePass.js.map +1 -1
- package/lib/backend.d.ts +8 -4
- package/lib/backend.js +36 -44
- package/lib/backend.js.map +1 -1
- package/lib/graphOptimizer.d.ts +54 -0
- package/lib/graphOptimizer.js +216 -0
- package/lib/graphOptimizer.js.map +1 -0
- package/lib/main.d.ts +33 -10
- package/lib/main.js +5 -0
- package/lib/main.js.map +1 -1
- package/lib/modelSpec.d.ts +80 -0
- package/lib/modelSpec.js +270 -0
- package/lib/modelSpec.js.map +1 -0
- package/lib/nativeUNet.d.ts +67 -0
- package/lib/nativeUNet.js +1735 -0
- package/lib/nativeUNet.js.map +1 -0
- package/lib/process.js +38 -35
- package/lib/process.js.map +1 -1
- package/lib/resourceTracker.d.ts +26 -0
- package/lib/resourceTracker.js +65 -0
- package/lib/resourceTracker.js.map +1 -0
- package/lib/tileScheduler.d.ts +33 -0
- package/lib/tileScheduler.js +86 -0
- package/lib/tileScheduler.js.map +1 -0
- package/lib/webnnUNet.d.ts +52 -0
- package/lib/webnnUNet.js +535 -0
- package/lib/webnnUNet.js.map +1 -0
- package/package.json +9 -5
- package/scripts/inspect-model.mjs +64 -0
- package/src/UNet.ts +236 -339
- package/src/WGPUComputePass.ts +6 -4
- package/src/backend.ts +42 -55
- package/src/graphOptimizer.ts +301 -0
- package/src/main.ts +71 -11
- package/src/modelSpec.ts +414 -0
- package/src/nativeUNet.ts +2256 -0
- package/src/process.ts +38 -36
- package/src/resourceTracker.ts +94 -0
- package/src/tileScheduler.ts +138 -0
- package/src/webnnUNet.ts +812 -0
- package/tests/modelSpec.test.mjs +128 -0
- package/tests/resourceLifecycle.test.mjs +383 -0
- package/tests/tileScheduler.test.mjs +90 -0
- package/lib/helper.d.ts +0 -4
- package/lib/helper.js +0 -33
- package/lib/helper.js.map +0 -1
- package/lib/kernels.d.ts +0 -1
- package/lib/kernels.js +0 -26
- package/lib/kernels.js.map +0 -1
- package/src/helper.ts +0 -43
- package/src/kernels.ts +0 -31
|
@@ -0,0 +1,76 @@
|
|
|
1
|
+
# oidn-web benchmark
|
|
2
|
+
|
|
3
|
+
- Current: `e6b8577b1c69cc8ce979022717984aec5b9ea881`
|
|
4
|
+
- TFJS baseline: `136c00c34aa5cf4561b4137034dbd643177f0a5f`
|
|
5
|
+
- Input: 512x512, fixed tile 512, 1 runs after 0 warmup(s)
|
|
6
|
+
- Adapter: apple
|
|
7
|
+
|
|
8
|
+
| Runtime | Init ms | Median ms | P95 ms | Speedup | Precision |
|
|
9
|
+
| --- | ---: | ---: | ---: | ---: | --- |
|
|
10
|
+
| TFJS (136c00c) | 9.70 | 486.20 | 486.20 | 1.00x | fp32 |
|
|
11
|
+
| WGSL FP32 | 38.50 | 43.10 | 43.10 | 11.28x | fp32 |
|
|
12
|
+
| WGSL FP16 Direct | 28.50 | 34.10 | 34.10 | 14.26x | fp16 |
|
|
13
|
+
| WGSL FP16 Implicit GEMM | 29.90 | 34.80 | 34.80 | 13.97x | fp16 |
|
|
14
|
+
| WGSL FP16 Spatial (experimental) | 32.30 | 70.10 | 70.10 | 6.94x | fp16 |
|
|
15
|
+
| WebNN FP16 (experimental) | skipped | - | - | - | Error: WebNN FP16 WebGPU interop failed: NotSupportedError: Failed to execute 'createExportableTensor' on 'MLContext': WebGPU Interop is not supported. |
|
|
16
|
+
|
|
17
|
+
## Output validation
|
|
18
|
+
|
|
19
|
+
Compared against sampled TFJS FP32 output.
|
|
20
|
+
|
|
21
|
+
| Runtime | Status | MAE | RMSE | Max error | Samples |
|
|
22
|
+
| --- | --- | ---: | ---: | ---: | ---: |
|
|
23
|
+
| WGSL FP32 | pass | 1.11e-6 | 1.81e-6 | 1.36e-5 | 6144 |
|
|
24
|
+
| WGSL FP16 Direct | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
|
|
25
|
+
| WGSL FP16 Implicit GEMM | pass | 1.56e-3 | 2.55e-3 | 4.48e-2 | 6144 |
|
|
26
|
+
| WGSL FP16 Spatial (experimental) | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
|
|
27
|
+
|
|
28
|
+
## GPU hot layers
|
|
29
|
+
|
|
30
|
+
### WGSL FP32
|
|
31
|
+
|
|
32
|
+
Profiled GPU total: 36.56 ms
|
|
33
|
+
|
|
34
|
+
| Node | GPU ms |
|
|
35
|
+
| --- | ---: |
|
|
36
|
+
| dec_conv1a | 6.87 |
|
|
37
|
+
| dec_conv2a | 4.79 |
|
|
38
|
+
| dec_conv1b | 4.17 |
|
|
39
|
+
| enc_conv1b | 4.16 |
|
|
40
|
+
| dec_conv3a | 2.32 |
|
|
41
|
+
|
|
42
|
+
### WGSL FP16 Direct
|
|
43
|
+
|
|
44
|
+
Profiled GPU total: 28.70 ms
|
|
45
|
+
|
|
46
|
+
| Node | GPU ms |
|
|
47
|
+
| --- | ---: |
|
|
48
|
+
| dec_conv1a | 4.40 |
|
|
49
|
+
| dec_conv1b | 3.43 |
|
|
50
|
+
| enc_conv1b | 3.13 |
|
|
51
|
+
| dec_conv2a | 2.99 |
|
|
52
|
+
| dec_conv3a | 2.24 |
|
|
53
|
+
|
|
54
|
+
### WGSL FP16 Implicit GEMM
|
|
55
|
+
|
|
56
|
+
Profiled GPU total: 30.75 ms
|
|
57
|
+
|
|
58
|
+
| Node | GPU ms |
|
|
59
|
+
| --- | ---: |
|
|
60
|
+
| dec_conv1a | 6.28 |
|
|
61
|
+
| dec_conv2a | 4.14 |
|
|
62
|
+
| enc_conv1b | 3.09 |
|
|
63
|
+
| dec_conv1b | 2.98 |
|
|
64
|
+
| dec_conv3a | 1.90 |
|
|
65
|
+
|
|
66
|
+
### WGSL FP16 Spatial (experimental)
|
|
67
|
+
|
|
68
|
+
Profiled GPU total: 64.85 ms
|
|
69
|
+
|
|
70
|
+
| Node | GPU ms |
|
|
71
|
+
| --- | ---: |
|
|
72
|
+
| dec_conv1a | 23.57 |
|
|
73
|
+
| dec_conv1b | 5.90 |
|
|
74
|
+
| enc_conv1b | 5.86 |
|
|
75
|
+
| dec_conv2b | 4.89 |
|
|
76
|
+
| enc_conv2b | 4.34 |
|