oidn-web 0.3.4 → 0.4.0

package/benchmarks/results/kernel-webnn-final.md

@@ -0,0 +1,104 @@
+# oidn-web benchmark
+
+- Current: `e6b8577b1c69cc8ce979022717984aec5b9ea881`
+- TFJS baseline: `136c00c34aa5cf4561b4137034dbd643177f0a5f`
+- Input: 512x512, fixed tile 512, 5 runs after 1 warmup(s)
+- Adapter: apple
+
+| Runtime | Init ms | Median ms | P95 ms | Speedup | Precision |
+| --- | ---: | ---: | ---: | ---: | --- |
+| TFJS (136c00c) | 9.40 | 35.20 | 35.40 | 1.00x | fp32 |
+| WGSL FP32 Auto | 39.20 | 36.30 | 36.50 | 0.97x | fp32 |
+| WGSL FP32 Direct | 37.60 | 63.30 | 63.40 | 0.56x | fp32 |
+| WGSL FP16 Direct | 27.60 | 27.80 | 28.00 | 1.27x | fp16 |
+| WGSL FP16 Implicit GEMM | 30.30 | 29.20 | 29.30 | 1.21x | fp16 |
+| WGSL FP16 Spatial (experimental) | 28.30 | 61.70 | 62.10 | 0.57x | fp16 |
+| WGSL FP16 Subgroup (experimental) | 29.20 | 29.50 | 29.60 | 1.19x | fp16 |
+| WebNN FP16 (experimental) | skipped | - | - | - | Error: WebNN FP16 WebGPU interop failed: NotSupportedError: Failed to execute 'createExportableTensor' on 'MLContext': WebGPU Interop is not supported. |
+
+## Output validation
+
+Compared against sampled TFJS FP32 output.
+
+| Runtime | Status | MAE | RMSE | Max error | Samples |
+| --- | --- | ---: | ---: | ---: | ---: |
+| WGSL FP32 Auto | pass | 1.11e-6 | 1.81e-6 | 1.36e-5 | 6144 |
+| WGSL FP32 Direct | pass | 1.11e-6 | 1.81e-6 | 1.36e-5 | 6144 |
+| WGSL FP16 Direct | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
+| WGSL FP16 Implicit GEMM | pass | 1.56e-3 | 2.55e-3 | 4.48e-2 | 6144 |
+| WGSL FP16 Spatial (experimental) | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
+| WGSL FP16 Subgroup (experimental) | pass | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
+
+## GPU hot layers
+
+### WGSL FP32 Auto
+
+Profiled GPU total: 35.24 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 6.89 |
+| dec_conv2a | 4.46 |
+| enc_conv1b | 3.82 |
+| dec_conv1b | 3.75 |
+| dec_conv3a | 2.34 |
+
+### WGSL FP32 Direct
+
+Profiled GPU total: 62.20 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 9.60 |
+| dec_conv1b | 8.66 |
+| enc_conv1b | 8.57 |
+| dec_conv2a | 6.76 |
+| dec_conv2b | 5.50 |
+
+### WGSL FP16 Direct
+
+Profiled GPU total: 27.07 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 4.51 |
+| enc_conv1b | 3.19 |
+| dec_conv1b | 3.10 |
+| dec_conv2a | 2.97 |
+| enc_conv2b | 2.00 |
+
+### WGSL FP16 Implicit GEMM
+
+Profiled GPU total: 28.77 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 5.85 |
+| dec_conv2a | 3.66 |
+| dec_conv1b | 2.93 |
+| enc_conv1b | 2.93 |
+| dec_conv3a | 1.90 |
+
+### WGSL FP16 Spatial (experimental)
+
+Profiled GPU total: 61.42 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 22.28 |
+| enc_conv1b | 5.43 |
+| dec_conv1b | 5.41 |
+| dec_conv2b | 4.41 |
+| enc_conv2b | 4.41 |
+
+### WGSL FP16 Subgroup (experimental)
+
+Profiled GPU total: 29.05 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 4.58 |
+| enc_conv1b | 3.47 |
+| dec_conv1b | 3.35 |
+| dec_conv2a | 3.11 |
+| enc_conv2b | 2.16 |

package/benchmarks/results/latest-optimized.json

@@ -0,0 +1,375 @@
+{
+  "generatedAt": "2026-08-20T03:59:27.859Z",
+  "currentCommit": "5d9c1c8771e2a03c90bfadafec4b5dd28fada446",
+  "baselineCommit": "136c00c34aa5cf4561b4137034dbd643177f0a5f",
+  "settings": {
+    "width": 512,
+    "height": 512,
+    "tileSize": 512,
+    "warmup": 1,
+    "runs": 5,
+    "model": "rt_hdr_calb_cnrm_large.tza",
+    "gpuQueueCompletionIncluded": true
+  },
+  "browser": "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome",
+  "adapter": {
+    "vendor": "apple",
+    "architecture": "metal-3",
+    "device": "",
+    "description": ""
+  },
+  "results": [
+    {
+      "label": "TFJS (136c00c)",
+      "bundle": "baseline",
+      "baseline": true,
+      "initializationMs": 9.100000001490116,
+      "timesMs": [
+        34.899999998509884,
+        35,
+        35.100000001490116,
+        34.5,
+        35.100000001490116
+      ],
+      "supportsFP16": true,
+      "adapter": {
+        "vendor": "apple",
+        "architecture": "metal-3",
+        "device": "",
+        "description": ""
+      },
+      "summary": {
+        "minMs": 34.5,
+        "medianMs": 35,
+        "meanMs": 34.920000000298025,
+        "p95Ms": 35.100000001490116,
+        "maxMs": 35.100000001490116
+      },
+      "messages": []
+    },
+    {
+      "label": "WGSL FP32",
+      "bundle": "current",
+      "baseline": false,
+      "precision": "fp32",
+      "initializationMs": 348.5,
+      "timesMs": [
+        41.100000001490116,
+        41.5,
+        41.29999999701977,
+        41.30000000447035,
+        41.29999999701977
+      ],
+      "executionProfile": {
+        "totalMs": 40.571729,
+        "layers": [
+          {
+            "id": "input-pack",
+            "durationMs": 0.051206
+          },
+          {
+            "id": "enc_conv1a",
+            "durationMs": 0.957123
+          },
+          {
+            "id": "enc_conv1b",
+            "durationMs": 4.522085
+          },
+          {
+            "id": "pool1",
+            "durationMs": 0.430064
+          },
+          {
+            "id": "enc_conv2a",
+            "durationMs": 1.667507
+          },
+          {
+            "id": "enc_conv2b",
+            "durationMs": 2.497385
+          },
+          {
+            "id": "pool2",
+            "durationMs": 0.054998
+          },
+          {
+            "id": "enc_conv3a",
+            "durationMs": 0.877502
+          },
+          {
+            "id": "enc_conv3b",
+            "durationMs": 1.16578
+          },
+          {
+            "id": "pool3",
+            "durationMs": 0.021374
+          },
+          {
+            "id": "enc_conv4a",
+            "durationMs": 0.478311
+          },
+          {
+            "id": "enc_conv4b",
+            "durationMs": 0.726759
+          },
+          {
+            "id": "pool4",
+            "durationMs": 0.011707
+          },
+          {
+            "id": "enc_conv5a",
+            "durationMs": 0.290404
+          },
+          {
+            "id": "enc_conv5b",
+            "durationMs": 0.384482
+          },
+          {
+            "id": "dec_conv4a",
+            "durationMs": 1.459184
+          },
+          {
+            "id": "dec_conv4b",
+            "durationMs": 0.718342
+          },
+          {
+            "id": "dec_conv3a",
+            "durationMs": 2.692501
+          },
+          {
+            "id": "dec_conv3b",
+            "durationMs": 1.162156
+          },
+          {
+            "id": "dec_conv2a",
+            "durationMs": 5.144015
+          },
+          {
+            "id": "dec_conv2b",
+            "durationMs": 2.493303
+          },
+          {
+            "id": "dec_conv1a",
+            "durationMs": 7.830059
+          },
+          {
+            "id": "dec_conv1b",
+            "durationMs": 4.44638
+          },
+          {
+            "id": "dec_conv1c",
+            "durationMs": 0.489102
+          }
+        ]
+      },
+      "runtimeInfo": {
+        "configuredEngine": "wgsl",
+        "gpuEngine": "wgsl",
+        "precision": "fp32",
+        "model": "oidn-unet-large-v1",
+        "modelFamily": "oidn-unet-large",
+        "inputChannels": 9,
+        "dynamicTile": {
+          "enabled": false,
+          "currentTileSize": 512,
+          "minTileSize": 256,
+          "maxTileSize": 512,
+          "targetTileTimeMs": 16
+        },
+        "lastExecution": {
+          "width": 512,
+          "height": 512,
+          "tileWidth": 512,
+          "tileHeight": 512,
+          "tileCount": 1,
+          "durationMs": 42.20000000298023,
+          "tileTimeMs": {
+            "min": 42.20000000298023,
+            "median": 42.20000000298023,
+            "mean": 42.20000000298023,
+            "max": 42.20000000298023
+          }
+        }
+      },
+      "supportsFP16": true,
+      "adapter": {
+        "vendor": "apple",
+        "architecture": "metal-3",
+        "device": "",
+        "description": ""
+      },
+      "summary": {
+        "minMs": 41.100000001490116,
+        "medianMs": 41.29999999701977,
+        "meanMs": 41.3,
+        "p95Ms": 41.5,
+        "maxMs": 41.5
+      },
+      "messages": [],
+      "validation": {
+        "sampleCount": 6144,
+        "meanAbsoluteError": 0.0000011096847932397698,
+        "rootMeanSquaredError": 0.0000018071588418628667,
+        "maxAbsoluteError": 0.000013589859008789062,
+        "passed": true,
+        "thresholds": {
+          "maxMeanError": 0.0001,
+          "maxAbsoluteError": 0.001
+        }
+      }
+    },
+    {
+      "label": "WGSL FP16",
+      "bundle": "current",
+      "baseline": false,
+      "precision": "fp16",
+      "initializationMs": 23.5,
+      "timesMs": [
+        31.200000002980232,
+        31.299999997019768,
+        31.399999998509884,
+        31.399999998509884,
+        31.200000002980232
+      ],
+      "executionProfile": {
+        "totalMs": 30.727957000000004,
+        "layers": [
+          {
+            "id": "input-pack",
+            "durationMs": 0.04383
+          },
+          {
+            "id": "enc_conv1a",
+            "durationMs": 0.736455
+          },
+          {
+            "id": "pool1",
+            "durationMs": 4.136372
+          },
+          {
+            "id": "enc_conv2a",
+            "durationMs": 1.513117
+          },
+          {
+            "id": "pool2",
+            "durationMs": 2.205823
+          },
+          {
+            "id": "enc_conv3a",
+            "durationMs": 0.762245
+          },
+          {
+            "id": "pool3",
+            "durationMs": 1.01298
+          },
+          {
+            "id": "enc_conv4a",
+            "durationMs": 0.357895
+          },
+          {
+            "id": "pool4",
+            "durationMs": 0.588173
+          },
+          {
+            "id": "enc_conv5a",
+            "durationMs": 0.195989
+          },
+          {
+            "id": "enc_conv5b",
+            "durationMs": 0.252401
+          },
+          {
+            "id": "dec_conv4a",
+            "durationMs": 0.971607
+          },
+          {
+            "id": "dec_conv4b",
+            "durationMs": 0.542633
+          },
+          {
+            "id": "dec_conv3a",
+            "durationMs": 1.742476
+          },
+          {
+            "id": "dec_conv3b",
+            "durationMs": 0.985274
+          },
+          {
+            "id": "dec_conv2a",
+            "durationMs": 3.19568
+          },
+          {
+            "id": "dec_conv2b",
+            "durationMs": 2.386271
+          },
+          {
+            "id": "dec_conv1a",
+            "durationMs": 4.708088
+          },
+          {
+            "id": "dec_conv1b",
+            "durationMs": 4.141663
+          },
+          {
+            "id": "dec_conv1c",
+            "durationMs": 0.248985
+          }
+        ]
+      },
+      "runtimeInfo": {
+        "configuredEngine": "wgsl",
+        "gpuEngine": "wgsl",
+        "precision": "fp16",
+        "model": "oidn-unet-large-v1",
+        "modelFamily": "oidn-unet-large",
+        "inputChannels": 9,
+        "dynamicTile": {
+          "enabled": false,
+          "currentTileSize": 512,
+          "minTileSize": 256,
+          "maxTileSize": 512,
+          "targetTileTimeMs": 16
+        },
+        "lastExecution": {
+          "width": 512,
+          "height": 512,
+          "tileWidth": 512,
+          "tileHeight": 512,
+          "tileCount": 1,
+          "durationMs": 32.29999999701977,
+          "tileTimeMs": {
+            "min": 32.29999999701977,
+            "median": 32.29999999701977,
+            "mean": 32.29999999701977,
+            "max": 32.29999999701977
+          }
+        }
+      },
+      "supportsFP16": true,
+      "adapter": {
+        "vendor": "apple",
+        "architecture": "metal-3",
+        "device": "",
+        "description": ""
+      },
+      "summary": {
+        "minMs": 31.200000002980232,
+        "medianMs": 31.299999997019768,
+        "meanMs": 31.3,
+        "p95Ms": 31.399999998509884,
+        "maxMs": 31.399999998509884
+      },
+      "messages": [],
+      "validation": {
+        "sampleCount": 6144,
+        "meanAbsoluteError": 0.0006845713796792552,
+        "rootMeanSquaredError": 0.001048555265853176,
+        "maxAbsoluteError": 0.00777888298034668,
+        "passed": true,
+        "thresholds": {
+          "maxMeanError": 0.005,
+          "maxAbsoluteError": 0.05
+        }
+      }
+    }
+  ]
+}

package/benchmarks/results/latest-optimized.md

@@ -0,0 +1,47 @@
+# oidn-web benchmark
+
+- Current: `5d9c1c8771e2a03c90bfadafec4b5dd28fada446`
+- TFJS baseline: `136c00c34aa5cf4561b4137034dbd643177f0a5f`
+- Input: 512x512, fixed tile 512, 5 runs after 1 warmup(s)
+- Adapter: apple
+
+| Runtime | Init ms | Median ms | P95 ms | Speedup | Precision |
+| --- | ---: | ---: | ---: | ---: | --- |
+| TFJS (136c00c) | 9.10 | 35.00 | 35.10 | 1.00x | fp32 |
+| WGSL FP32 | 348.50 | 41.30 | 41.50 | 0.85x | fp32 |
+| WGSL FP16 | 23.50 | 31.30 | 31.40 | 1.12x | fp16 |
+
+## Output validation
+
+Compared against sampled TFJS FP32 output.
+
+| Runtime | Status | MAE | RMSE | Max error | Samples |
+| --- | --- | ---: | ---: | ---: | ---: |
+| WGSL FP32 | pass | 1.11e-6 | 1.81e-6 | 1.36e-5 | 6144 |
+| WGSL FP16 | pass | 6.85e-4 | 1.05e-3 | 7.78e-3 | 6144 |
+
+## GPU hot layers
+
+### WGSL FP32
+
+Profiled GPU total: 40.57 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 7.83 |
+| dec_conv2a | 5.14 |
+| enc_conv1b | 4.52 |
+| dec_conv1b | 4.45 |
+| dec_conv3a | 2.69 |
+
+### WGSL FP16
+
+Profiled GPU total: 30.73 ms
+
+| Node | GPU ms |
+| --- | ---: |
+| dec_conv1a | 4.71 |
+| dec_conv1b | 4.14 |
+| pool1 | 4.14 |
+| dec_conv2a | 3.20 |
+| dec_conv2b | 2.39 |