oidn-web 0.4.0 → 0.5.0

package/benchmarks/results/profile-fp16-fma.json

@@ -1,369 +0,0 @@
-{
-  "generatedAt": "2026-08-20T04:03:41.799Z",
-  "currentCommit": "5d9c1c8771e2a03c90bfadafec4b5dd28fada446",
-  "baselineCommit": "136c00c34aa5cf4561b4137034dbd643177f0a5f",
-  "settings": {
-    "width": 512,
-    "height": 512,
-    "tileSize": 512,
-    "warmup": 1,
-    "runs": 3,
-    "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": 9.399999998509884,
-      "timesMs": [
-        34.79999999701977,
-        34.79999999701977,
-        34.80000000447035
-      ],
-      "supportsFP16": true,
-      "adapter": {
-        "vendor": "apple",
-        "architecture": "metal-3",
-        "device": "",
-        "description": ""
-      },
-      "summary": {
-        "minMs": 34.79999999701977,
-        "medianMs": 34.79999999701977,
-        "meanMs": 34.79999999950329,
-        "p95Ms": 34.80000000447035,
-        "maxMs": 34.80000000447035
-      },
-      "messages": []
-    },
-    {
-      "label": "WGSL FP32",
-      "bundle": "current",
-      "baseline": false,
-      "precision": "fp32",
-      "initializationMs": 350.1000000014901,
-      "timesMs": [
-        41.100000001490116,
-        41.29999999701977,
-        41.100000001490116
-      ],
-      "executionProfile": {
-        "totalMs": 40.44157,
-        "layers": [
-          {
-            "id": "input-pack",
-            "durationMs": 0.051163
-          },
-          {
-            "id": "enc_conv1a",
-            "durationMs": 0.956983
-          },
-          {
-            "id": "enc_conv1b",
-            "durationMs": 4.39719
-          },
-          {
-            "id": "pool1",
-            "durationMs": 0.416975
-          },
-          {
-            "id": "enc_conv2a",
-            "durationMs": 1.668606
-          },
-          {
-            "id": "enc_conv2b",
-            "durationMs": 2.496473
-          },
-          {
-            "id": "pool2",
-            "durationMs": 0.055329
-          },
-          {
-            "id": "enc_conv3a",
-            "durationMs": 0.880113
-          },
-          {
-            "id": "enc_conv3b",
-            "durationMs": 1.161971
-          },
-          {
-            "id": "pool3",
-            "durationMs": 0.018207
-          },
-          {
-            "id": "enc_conv4a",
-            "durationMs": 0.478053
-          },
-          {
-            "id": "enc_conv4b",
-            "durationMs": 0.716249
-          },
-          {
-            "id": "pool4",
-            "durationMs": 0.008457
-          },
-          {
-            "id": "enc_conv5a",
-            "durationMs": 0.293065
-          },
-          {
-            "id": "enc_conv5b",
-            "durationMs": 0.386643
-          },
-          {
-            "id": "dec_conv4a",
-            "durationMs": 1.462076
-          },
-          {
-            "id": "dec_conv4b",
-            "durationMs": 0.717873
-          },
-          {
-            "id": "dec_conv3a",
-            "durationMs": 2.680504
-          },
-          {
-            "id": "dec_conv3b",
-            "durationMs": 1.161471
-          },
-          {
-            "id": "dec_conv2a",
-            "durationMs": 5.161435
-          },
-          {
-            "id": "dec_conv2b",
-            "durationMs": 2.501847
-          },
-          {
-            "id": "dec_conv1a",
-            "durationMs": 7.830522
-          },
-          {
-            "id": "dec_conv1b",
-            "durationMs": 4.450562
-          },
-          {
-            "id": "dec_conv1c",
-            "durationMs": 0.489803
-          }
-        ]
-      },
-      "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": 42.100000001490116,
-          "tileTimeMs": {
-            "min": 42.100000001490116,
-            "median": 42.100000001490116,
-            "mean": 42.100000001490116,
-            "max": 42.100000001490116
-          }
-        }
-      },
-      "supportsFP16": true,
-      "adapter": {
-        "vendor": "apple",
-        "architecture": "metal-3",
-        "device": "",
-        "description": ""
-      },
-      "summary": {
-        "minMs": 41.100000001490116,
-        "medianMs": 41.100000001490116,
-        "meanMs": 41.166666666666664,
-        "p95Ms": 41.29999999701977,
-        "maxMs": 41.29999999701977
-      },
-      "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": 22.700000002980232,
-      "timesMs": [
-        30.200000002980232,
-        29.899999998509884,
-        30
-      ],
-      "executionProfile": {
-        "totalMs": 29.420870999999998,
-        "layers": [
-          {
-            "id": "input-pack",
-            "durationMs": 0.043788
-          },
-          {
-            "id": "enc_conv1a",
-            "durationMs": 0.719206
-          },
-          {
-            "id": "pool1",
-            "durationMs": 3.929344
-          },
-          {
-            "id": "enc_conv2a",
-            "durationMs": 1.447161
-          },
-          {
-            "id": "pool2",
-            "durationMs": 2.234155
-          },
-          {
-            "id": "enc_conv3a",
-            "durationMs": 0.742497
-          },
-          {
-            "id": "pool3",
-            "durationMs": 0.91986
-          },
-          {
-            "id": "enc_conv4a",
-            "durationMs": 0.34123
-          },
-          {
-            "id": "pool4",
-            "durationMs": 0.601713
-          },
-          {
-            "id": "enc_conv5a",
-            "durationMs": 0.189821
-          },
-          {
-            "id": "enc_conv5b",
-            "durationMs": 0.247027
-          },
-          {
-            "id": "dec_conv4a",
-            "durationMs": 0.960733
-          },
-          {
-            "id": "dec_conv4b",
-            "durationMs": 0.522011
-          },
-          {
-            "id": "dec_conv3a",
-            "durationMs": 1.701355
-          },
-          {
-            "id": "dec_conv3b",
-            "durationMs": 0.929527
-          },
-          {
-            "id": "dec_conv2a",
-            "durationMs": 3.002526
-          },
-          {
-            "id": "dec_conv2b",
-            "durationMs": 2.311275
-          },
-          {
-            "id": "dec_conv1a",
-            "durationMs": 4.435979
-          },
-          {
-            "id": "dec_conv1b",
-            "durationMs": 3.907137
-          },
-          {
-            "id": "dec_conv1c",
-            "durationMs": 0.234526
-          }
-        ]
-      },
-      "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": 31,
-          "tileTimeMs": {
-            "min": 31,
-            "median": 31,
-            "mean": 31,
-            "max": 31
-          }
-        }
-      },
-      "supportsFP16": true,
-      "adapter": {
-        "vendor": "apple",
-        "architecture": "metal-3",
-        "device": "",
-        "description": ""
-      },
-      "summary": {
-        "minMs": 29.899999998509884,
-        "medianMs": 30,
-        "meanMs": 30.03333333383004,
-        "p95Ms": 30.200000002980232,
-        "maxMs": 30.200000002980232
-      },
-      "messages": [],
-      "validation": {
-        "sampleCount": 6144,
-        "meanAbsoluteError": 0.0010210334021394374,
-        "rootMeanSquaredError": 0.001609680379854656,
-        "maxAbsoluteError": 0.012395620346069336,
-        "passed": true,
-        "thresholds": {
-          "maxMeanError": 0.005,
-          "maxAbsoluteError": 0.05
-        }
-      }
-    }
-  ]
-}

package/benchmarks/results/profile-fp16-fma.md

@@ -1,47 +0,0 @@
-# oidn-web benchmark
-
-- Current: `5d9c1c8771e2a03c90bfadafec4b5dd28fada446`
-- TFJS baseline: `136c00c34aa5cf4561b4137034dbd643177f0a5f`
-- Input: 512x512, fixed tile 512, 3 runs after 1 warmup(s)
-- Adapter: apple
-
-| Runtime | Init ms | Median ms | P95 ms | Speedup | Precision |
-| --- | ---: | ---: | ---: | ---: | --- |
-| TFJS (136c00c) | 9.40 | 34.80 | 34.80 | 1.00x | fp32 |
-| WGSL FP32 | 350.10 | 41.10 | 41.30 | 0.85x | fp32 |
-| WGSL FP16 | 22.70 | 30.00 | 30.20 | 1.16x | 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 | 1.02e-3 | 1.61e-3 | 1.24e-2 | 6144 |
-
-## GPU hot layers
-
-### WGSL FP32
-
-Profiled GPU total: 40.44 ms
-
-| Node | GPU ms |
-| --- | ---: |
-| dec_conv1a | 7.83 |
-| dec_conv2a | 5.16 |
-| dec_conv1b | 4.45 |
-| enc_conv1b | 4.40 |
-| dec_conv3a | 2.68 |
-
-### WGSL FP16
-
-Profiled GPU total: 29.42 ms
-
-| Node | GPU ms |
-| --- | ---: |
-| dec_conv1a | 4.44 |
-| pool1 | 3.93 |
-| dec_conv1b | 3.91 |
-| dec_conv2a | 3.00 |
-| dec_conv2b | 2.31 |