moge3-webgpu / benchmarks /buffer-summary.json
xjodoin's picture
Publish subgroup FP32 kernels and validated multipass WebGPU inference
ecd48a0 verified
Raw History Blame Contribute Delete
8.46 kB
{
"date": "2026-10-30T18:12:19.631Z",
"device": "Apple M4 Pro 24 GB, Chrome 156 Metal-3",
"method": "Same production build; reuse disabled and enabled; group order counterbalanced by resolution. One warmup per configuration, three repeats per photo. Compute excludes loading, validation and rendering. Timestamp-profiling runs are separate and deliberately synchronized; GPU timestamps measure shader execution, ONNX linear timings include CPU dispatch and GPU completion. Baseline disabled retains the same capacity rounding to isolate reuse.",
"performanceDecision": "Reuse remains opt-in: allocation count falls, but observed latency does not improve consistently and retained buffers increase memory.",
"defaultReuse": false,
"profiles": [
{
"profile": "fast",
"refineSteps": 1,
"disabled": {
"computeMs": 1005.1150000095367,
"refinementMs": 423.8899999856949,
"createdBuffers": 97,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 2726940592,
"worstPointsNrmse": 0.0003349385405136117,
"maskMismatches": 0
},
"enabled": {
"computeMs": 1003.0149999856949,
"refinementMs": 422.2450000047684,
"createdBuffers": 34,
"reusedBuffers": 63,
"peakTrackedGpuBytes": 3003084640,
"worstPointsNrmse": 0.0003349385405136117,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": 0.20893131868711112
},
{
"profile": "fast",
"refineSteps": 2,
"disabled": {
"computeMs": 1408.0400000810623,
"refinementMs": 828.4400000572205,
"createdBuffers": 194,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 2726940592,
"worstPointsNrmse": 0.0007103017689260219,
"maskMismatches": 0
},
"enabled": {
"computeMs": 1416.7799999713898,
"refinementMs": 837.7699999809265,
"createdBuffers": 52,
"reusedBuffers": 142,
"peakTrackedGpuBytes": 3000371696,
"worstPointsNrmse": 0.0007103017689260219,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -0.6207209944194991
},
{
"profile": "fast",
"refineSteps": 3,
"disabled": {
"computeMs": 1822.5350000858307,
"refinementMs": 1245.210000038147,
"createdBuffers": 291,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 2726940592,
"worstPointsNrmse": 0.0009413059419531438,
"maskMismatches": 0
},
"enabled": {
"computeMs": 1829.3150000572205,
"refinementMs": 1231.8199999332428,
"createdBuffers": 74,
"reusedBuffers": 217,
"peakTrackedGpuBytes": 3002660880,
"worstPointsNrmse": 0.0009413059419531438,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -0.37200931510619206
},
{
"profile": "detail",
"refineSteps": 1,
"disabled": {
"computeMs": 1509.079999923706,
"refinementMs": 588.9099999666214,
"createdBuffers": 97,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 3266447328,
"worstPointsNrmse": 0.0002887960802053227,
"maskMismatches": 0
},
"enabled": {
"computeMs": 1522.6599999666214,
"refinementMs": 596.25,
"createdBuffers": 35,
"reusedBuffers": 62,
"peakTrackedGpuBytes": 3500665008,
"worstPointsNrmse": 0.0002887960802053227,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -0.8998860261617558
},
{
"profile": "detail",
"refineSteps": 2,
"disabled": {
"computeMs": 2068.399999976158,
"refinementMs": 1143.8300000429153,
"createdBuffers": 194,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 3266447328,
"worstPointsNrmse": 0.0008415597114219503,
"maskMismatches": 0
},
"enabled": {
"computeMs": 2092.264999985695,
"refinementMs": 1168.905000090599,
"createdBuffers": 54,
"reusedBuffers": 140,
"peakTrackedGpuBytes": 3501008656,
"worstPointsNrmse": 0.0008415597114219503,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -1.1537903698420005
},
{
"profile": "detail",
"refineSteps": 3,
"disabled": {
"computeMs": 2620.1600000858307,
"refinementMs": 1694.1899999380112,
"createdBuffers": 291,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 3266447328,
"worstPointsNrmse": 0.0009430723759879896,
"maskMismatches": 0
},
"enabled": {
"computeMs": 2749.420000076294,
"refinementMs": 1756.0249999761581,
"createdBuffers": 80,
"reusedBuffers": 211,
"peakTrackedGpuBytes": 3501326768,
"worstPointsNrmse": 0.0009430723759879896,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -4.933286516328361
}
],
"gpuLayers": [
{
"profile": "detail",
"layer": "down_stages.4.0",
"medianGpuMsAcrossThreePasses": 181.27257600000002
},
{
"profile": "detail",
"layer": "bottleneck_stage.0",
"medianGpuMsAcrossThreePasses": 180.682752
},
{
"profile": "detail",
"layer": "down_stages.3.0",
"medianGpuMsAcrossThreePasses": 165.74054400000003
},
{
"profile": "detail",
"layer": "up_stages.0.0",
"medianGpuMsAcrossThreePasses": 164.298752
},
{
"profile": "detail",
"layer": "down_stages.2.0",
"medianGpuMsAcrossThreePasses": 142.147584
},
{
"profile": "detail",
"layer": "up_stages.1.0",
"medianGpuMsAcrossThreePasses": 141.55776
},
{
"profile": "fast",
"layer": "bottleneck_stage.0",
"medianGpuMsAcrossThreePasses": 131.858432
},
{
"profile": "fast",
"layer": "down_stages.4.0",
"medianGpuMsAcrossThreePasses": 130.351104
},
{
"profile": "fast",
"layer": "down_stages.3.0",
"medianGpuMsAcrossThreePasses": 118.35801599999999
},
{
"profile": "fast",
"layer": "up_stages.0.0",
"medianGpuMsAcrossThreePasses": 117.44051199999998
},
{
"profile": "detail",
"layer": "down_stages.1.0",
"medianGpuMsAcrossThreePasses": 116.65408000000001
},
{
"profile": "detail",
"layer": "up_stages.2.0",
"medianGpuMsAcrossThreePasses": 116.19532799999999
},
{
"profile": "fast",
"layer": "down_stages.2.0",
"medianGpuMsAcrossThreePasses": 102.23616
},
{
"profile": "fast",
"layer": "up_stages.1.0",
"medianGpuMsAcrossThreePasses": 101.58080000000001
},
{
"profile": "detail",
"layer": "up_stages.3.0",
"medianGpuMsAcrossThreePasses": 85.06572800000001
},
{
"profile": "detail",
"layer": "down_stages.0.0",
"medianGpuMsAcrossThreePasses": 82.051072
},
{
"profile": "fast",
"layer": "down_stages.1.0",
"medianGpuMsAcrossThreePasses": 80.543744
},
{
"profile": "fast",
"layer": "up_stages.2.0",
"medianGpuMsAcrossThreePasses": 80.216064
},
{
"profile": "fast",
"layer": "up_stages.3.0",
"medianGpuMsAcrossThreePasses": 55.96774399999999
},
{
"profile": "fast",
"layer": "down_stages.0.0",
"medianGpuMsAcrossThreePasses": 55.508992
},
{
"profile": "detail",
"layer": "gather",
"medianGpuMsAcrossThreePasses": 24.707072
},
{
"profile": "fast",
"layer": "gather",
"medianGpuMsAcrossThreePasses": 21.889024
},
{
"profile": "detail",
"layer": "upsampleAdd",
"medianGpuMsAcrossThreePasses": 10.944512
},
{
"profile": "fast",
"layer": "upsampleAdd",
"medianGpuMsAcrossThreePasses": 7.405568
},
{
"profile": "detail",
"layer": "pool",
"medianGpuMsAcrossThreePasses": 6.029312000000001
},
{
"profile": "fast",
"layer": "pool",
"medianGpuMsAcrossThreePasses": 3.5389440000000003
}
],
"passed": true
}