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.
Files changed (122) hide show
  1. package/CHANGELOG.md +49 -0
  2. package/README.md +140 -4
  3. package/benchmarks/compare.mjs +651 -0
  4. package/benchmarks/leak.mjs +255 -0
  5. package/benchmarks/results/before-spatial.json +391 -0
  6. package/benchmarks/results/before-spatial.md +47 -0
  7. package/benchmarks/results/int8-scan.json +2007 -0
  8. package/benchmarks/results/int8-scan.md +160 -0
  9. package/benchmarks/results/int8-w8a8-scan.json +2007 -0
  10. package/benchmarks/results/int8-w8a8-scan.md +160 -0
  11. package/benchmarks/results/int8-weight-channel.json +1413 -0
  12. package/benchmarks/results/int8-weight-channel.md +118 -0
  13. package/benchmarks/results/int8-weight-only.json +1437 -0
  14. package/benchmarks/results/int8-weight-only.md +118 -0
  15. package/benchmarks/results/kernel-webnn-final.json +1115 -0
  16. package/benchmarks/results/kernel-webnn-final.md +104 -0
  17. package/benchmarks/results/latest-optimized.json +375 -0
  18. package/benchmarks/results/latest-optimized.md +47 -0
  19. package/benchmarks/results/latest.json +391 -0
  20. package/benchmarks/results/latest.md +47 -0
  21. package/benchmarks/results/profile-baseline.json +331 -0
  22. package/benchmarks/results/profile-baseline.md +12 -0
  23. package/benchmarks/results/profile-conv2x.json +331 -0
  24. package/benchmarks/results/profile-conv2x.md +12 -0
  25. package/benchmarks/results/profile-fast-init.json +385 -0
  26. package/benchmarks/results/profile-fast-init.md +47 -0
  27. package/benchmarks/results/profile-fp16-fma.json +369 -0
  28. package/benchmarks/results/profile-fp16-fma.md +47 -0
  29. package/benchmarks/results/profile-fp16-tiled-decoder.json +369 -0
  30. package/benchmarks/results/profile-fp16-tiled-decoder.md +47 -0
  31. package/benchmarks/results/profile-fp16-tiled-encoder.json +369 -0
  32. package/benchmarks/results/profile-fp16-tiled-encoder.md +47 -0
  33. package/benchmarks/results/profile-fp16-unfused-pool.json +385 -0
  34. package/benchmarks/results/profile-fp16-unfused-pool.md +47 -0
  35. package/benchmarks/results/profile-input-major.json +347 -0
  36. package/benchmarks/results/profile-input-major.md +12 -0
  37. package/benchmarks/results/profile-k16.json +347 -0
  38. package/benchmarks/results/profile-k16.md +12 -0
  39. package/benchmarks/results/profile-k4.json +347 -0
  40. package/benchmarks/results/profile-k4.md +12 -0
  41. package/benchmarks/results/profile-pool-reuse.json +331 -0
  42. package/benchmarks/results/profile-pool-reuse.md +12 -0
  43. package/benchmarks/results/profile-precompiled.json +385 -0
  44. package/benchmarks/results/profile-precompiled.md +47 -0
  45. package/benchmarks/results/profile-static-channels.json +385 -0
  46. package/benchmarks/results/profile-static-channels.md +47 -0
  47. package/benchmarks/results/profile-static-io.json +385 -0
  48. package/benchmarks/results/profile-static-io.md +47 -0
  49. package/benchmarks/results/profile-tiled-conv.json +331 -0
  50. package/benchmarks/results/profile-tiled-conv.md +12 -0
  51. package/benchmarks/results/profile-tiled-decoder.json +347 -0
  52. package/benchmarks/results/profile-tiled-decoder.md +12 -0
  53. package/benchmarks/results/profile-tiled-matmul.json +331 -0
  54. package/benchmarks/results/profile-tiled-matmul.md +12 -0
  55. package/benchmarks/results/profile-unfused-decoder.json +379 -0
  56. package/benchmarks/results/profile-unfused-decoder.md +12 -0
  57. package/benchmarks/results/profile-unfused-pool.json +347 -0
  58. package/benchmarks/results/profile-unfused-pool.md +12 -0
  59. package/benchmarks/results/spatial-auto.json +575 -0
  60. package/benchmarks/results/spatial-auto.md +61 -0
  61. package/benchmarks/results/subgroup-smoke.json +1094 -0
  62. package/benchmarks/results/subgroup-smoke.md +104 -0
  63. package/benchmarks/results/webnn-smoke.json +739 -0
  64. package/benchmarks/results/webnn-smoke.md +76 -0
  65. package/dist/oidn.js +4189 -22603
  66. package/dist/oidn.umd.cjs +784 -5807
  67. package/lib/UNet.d.ts +66 -15
  68. package/lib/UNet.js +162 -257
  69. package/lib/UNet.js.map +1 -1
  70. package/lib/WGPUComputePass.d.ts +1 -1
  71. package/lib/WGPUComputePass.js +6 -4
  72. package/lib/WGPUComputePass.js.map +1 -1
  73. package/lib/backend.d.ts +8 -4
  74. package/lib/backend.js +36 -44
  75. package/lib/backend.js.map +1 -1
  76. package/lib/graphOptimizer.d.ts +54 -0
  77. package/lib/graphOptimizer.js +216 -0
  78. package/lib/graphOptimizer.js.map +1 -0
  79. package/lib/main.d.ts +33 -10
  80. package/lib/main.js +5 -0
  81. package/lib/main.js.map +1 -1
  82. package/lib/modelSpec.d.ts +80 -0
  83. package/lib/modelSpec.js +270 -0
  84. package/lib/modelSpec.js.map +1 -0
  85. package/lib/nativeUNet.d.ts +67 -0
  86. package/lib/nativeUNet.js +1735 -0
  87. package/lib/nativeUNet.js.map +1 -0
  88. package/lib/process.js +38 -35
  89. package/lib/process.js.map +1 -1
  90. package/lib/resourceTracker.d.ts +26 -0
  91. package/lib/resourceTracker.js +65 -0
  92. package/lib/resourceTracker.js.map +1 -0
  93. package/lib/tileScheduler.d.ts +33 -0
  94. package/lib/tileScheduler.js +86 -0
  95. package/lib/tileScheduler.js.map +1 -0
  96. package/lib/webnnUNet.d.ts +52 -0
  97. package/lib/webnnUNet.js +535 -0
  98. package/lib/webnnUNet.js.map +1 -0
  99. package/package.json +9 -5
  100. package/scripts/inspect-model.mjs +64 -0
  101. package/src/UNet.ts +236 -339
  102. package/src/WGPUComputePass.ts +6 -4
  103. package/src/backend.ts +42 -55
  104. package/src/graphOptimizer.ts +301 -0
  105. package/src/main.ts +71 -11
  106. package/src/modelSpec.ts +414 -0
  107. package/src/nativeUNet.ts +2256 -0
  108. package/src/process.ts +38 -36
  109. package/src/resourceTracker.ts +94 -0
  110. package/src/tileScheduler.ts +138 -0
  111. package/src/webnnUNet.ts +812 -0
  112. package/tests/modelSpec.test.mjs +128 -0
  113. package/tests/resourceLifecycle.test.mjs +383 -0
  114. package/tests/tileScheduler.test.mjs +90 -0
  115. package/lib/helper.d.ts +0 -4
  116. package/lib/helper.js +0 -33
  117. package/lib/helper.js.map +0 -1
  118. package/lib/kernels.d.ts +0 -1
  119. package/lib/kernels.js +0 -26
  120. package/lib/kernels.js.map +0 -1
  121. package/src/helper.ts +0 -43
  122. package/src/kernels.ts +0 -31
package/benchmarks/results/int8-weight-only.md ADDED
@@ -0,0 +1,118 @@
1
+ # oidn-web benchmark
2
+
3
+ - Current: `4dc1797b24c64aa0d647eada0cb3389fb0d24705`
4
+ - TFJS baseline: `136c00c34aa5cf4561b4137034dbd643177f0a5f`
5
+ - Input: 512x512, fixed tile 512, 5 runs after 1 warmup(s)
6
+ - Adapter: apple
7
+
8
+ | Runtime | Init ms | Median ms | P95 ms | Speedup | Precision |
9
+ | --- | ---: | ---: | ---: | ---: | --- |
10
+ | TFJS (136c00c) | 9.30 | 36.70 | 37.10 | 1.00x | fp32 |
11
+ | WGSL FP32 Auto | 38.60 | 37.70 | 37.80 | 0.97x | fp32 |
12
+ | WGSL FP32 Direct | 37.60 | 66.50 | 67.20 | 0.55x | fp32 |
13
+ | WGSL FP16 Direct | 26.90 | 30.10 | 30.20 | 1.22x | fp16 |
14
+ | WGSL FP16 Implicit GEMM | 31.20 | 30.80 | 31.00 | 1.19x | fp16 |
15
+ | WGSL FP16 Spatial (experimental) | 28.60 | 65.10 | 65.70 | 0.56x | fp16 |
16
+ | WGSL FP16 Subgroup (experimental) | 28.50 | 31.30 | 31.50 | 1.17x | fp16 |
17
+ | WGSL INT8 weights / FP16 activations (experimental) | 152.10 | 101.30 | 101.60 | 0.36x | int8 |
18
+ | WebNN FP16 (experimental) | skipped | - | - | - | Error: WebNN FP16 WebGPU interop failed: NotSupportedError: Failed to execute 'createExportableTensor' on 'MLContext': WebGPU Interop is not supported. |
19
+
20
+ ## Output validation
21
+
22
+ Compared against sampled TFJS FP32 output.
23
+
24
+ | Runtime | Status | MAE | RMSE | Max error | Samples |
25
+ | --- | --- | ---: | ---: | ---: | ---: |
26
+ | WGSL FP32 Auto | pass | 1.11e-6 | 1.81e-6 | 1.36e-5 | 6144 |
27
+ | WGSL FP32 Direct | pass | 1.11e-6 | 1.81e-6 | 1.36e-5 | 6144 |
28
+ | WGSL FP16 Direct | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
29
+ | WGSL FP16 Implicit GEMM | pass | 1.56e-3 | 2.55e-3 | 4.48e-2 | 6144 |
30
+ | WGSL FP16 Spatial (experimental) | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
31
+ | WGSL FP16 Subgroup (experimental) | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
32
+ | WGSL INT8 weights / FP16 activations (experimental) | FAIL | 5.26e-1 | 7.73e-1 | 2.58e+0 | 6144 |
33
+
34
+ ## GPU hot layers
35
+
36
+ ### WGSL FP32 Auto
37
+
38
+ Profiled GPU total: 37.10 ms
39
+
40
+ | Node | GPU ms |
41
+ | --- | ---: |
42
+ | dec_conv1a | 7.13 |
43
+ | dec_conv2a | 4.59 |
44
+ | enc_conv1b | 4.25 |
45
+ | dec_conv1b | 3.80 |
46
+ | dec_conv3a | 2.60 |
47
+
48
+ ### WGSL FP32 Direct
49
+
50
+ Profiled GPU total: 65.77 ms
51
+
52
+ | Node | GPU ms |
53
+ | --- | ---: |
54
+ | dec_conv1a | 10.51 |
55
+ | enc_conv1b | 9.26 |
56
+ | dec_conv1b | 8.71 |
57
+ | dec_conv2a | 6.99 |
58
+ | enc_conv2b | 5.62 |
59
+
60
+ ### WGSL FP16 Direct
61
+
62
+ Profiled GPU total: 29.51 ms
63
+
64
+ | Node | GPU ms |
65
+ | --- | ---: |
66
+ | dec_conv1a | 4.42 |
67
+ | dec_conv1b | 3.82 |
68
+ | enc_conv1b | 3.12 |
69
+ | dec_conv2a | 2.97 |
70
+ | dec_conv2b | 2.66 |
71
+
72
+ ### WGSL FP16 Implicit GEMM
73
+
74
+ Profiled GPU total: 29.74 ms
75
+
76
+ | Node | GPU ms |
77
+ | --- | ---: |
78
+ | dec_conv1a | 6.22 |
79
+ | dec_conv2a | 4.07 |
80
+ | dec_conv1b | 2.94 |
81
+ | enc_conv1b | 2.94 |
82
+ | dec_conv3a | 1.90 |
83
+
84
+ ### WGSL FP16 Spatial (experimental)
85
+
86
+ Profiled GPU total: 64.84 ms
87
+
88
+ | Node | GPU ms |
89
+ | --- | ---: |
90
+ | dec_conv1a | 23.73 |
91
+ | dec_conv1b | 5.72 |
92
+ | enc_conv1b | 5.44 |
93
+ | enc_conv2b | 4.76 |
94
+ | dec_conv2b | 4.30 |
95
+
96
+ ### WGSL FP16 Subgroup (experimental)
97
+
98
+ Profiled GPU total: 30.59 ms
99
+
100
+ | Node | GPU ms |
101
+ | --- | ---: |
102
+ | dec_conv1a | 4.38 |
103
+ | dec_conv1b | 4.07 |
104
+ | enc_conv1b | 4.02 |
105
+ | dec_conv2a | 3.04 |
106
+ | dec_conv2b | 2.55 |
107
+
108
+ ### WGSL INT8 weights / FP16 activations (experimental)
109
+
110
+ Profiled GPU total: 100.74 ms
111
+
112
+ | Node | GPU ms |
113
+ | --- | ---: |
114
+ | dec_conv1a | 19.51 |
115
+ | dec_conv2a | 12.72 |
116
+ | dec_conv1b | 11.86 |
117
+ | enc_conv1b | 11.63 |
118
+ | dec_conv2b | 6.59 |