moge3-webgpu / benchmarks /buffer-summary-ac.json
xjodoin's picture
Publish subgroup FP32 kernels and validated multipass WebGPU inference
ecd48a0 verified
Raw History Blame Contribute Delete
8.53 kB
{
"powerSource": "AC verified throughout run",
"powerSamples": 75,
"date": "2026-10-30T19:06:25.627Z",
"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": 991.3949999809265,
"refinementMs": 418.1650000810623,
"createdBuffers": 97,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 2726940592,
"worstPointsNrmse": 0.0003349385405136117,
"maskMismatches": 0
},
"enabled": {
"computeMs": 999.914999961853,
"refinementMs": 417.710000038147,
"createdBuffers": 34,
"reusedBuffers": 63,
"peakTrackedGpuBytes": 3003084640,
"worstPointsNrmse": 0.0003349385405136117,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -0.8593950928833083
},
{
"profile": "fast",
"refineSteps": 2,
"disabled": {
"computeMs": 1497.2849999666214,
"refinementMs": 827.3049999475479,
"createdBuffers": 194,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 2726940592,
"worstPointsNrmse": 0.0007103017689260219,
"maskMismatches": 0
},
"enabled": {
"computeMs": 1383.9700000286102,
"refinementMs": 803.2599999904633,
"createdBuffers": 52,
"reusedBuffers": 142,
"peakTrackedGpuBytes": 3000371696,
"worstPointsNrmse": 0.0007103017689260219,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": 7.5680314663232
},
{
"profile": "fast",
"refineSteps": 3,
"disabled": {
"computeMs": 1774.6750000715256,
"refinementMs": 1199.5,
"createdBuffers": 291,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 2726940592,
"worstPointsNrmse": 0.0009413059419531438,
"maskMismatches": 0
},
"enabled": {
"computeMs": 1783.0049999952316,
"refinementMs": 1206.3849999904633,
"createdBuffers": 74,
"reusedBuffers": 217,
"peakTrackedGpuBytes": 3002660880,
"worstPointsNrmse": 0.0009413059419531438,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -0.4693817134613587
},
{
"profile": "detail",
"refineSteps": 1,
"disabled": {
"computeMs": 1481.835000038147,
"refinementMs": 569.2250000238419,
"createdBuffers": 97,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 3266447328,
"worstPointsNrmse": 0.0002887960802053227,
"maskMismatches": 0
},
"enabled": {
"computeMs": 1500.7949999570847,
"refinementMs": 575.8300000429153,
"createdBuffers": 35,
"reusedBuffers": 62,
"peakTrackedGpuBytes": 3500665008,
"worstPointsNrmse": 0.0002887960802053227,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -1.2794946750785074
},
{
"profile": "detail",
"refineSteps": 2,
"disabled": {
"computeMs": 2030.4500000476837,
"refinementMs": 1115.6749999523163,
"createdBuffers": 194,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 3266447328,
"worstPointsNrmse": 0.0008415597114219503,
"maskMismatches": 0
},
"enabled": {
"computeMs": 2045.3949999809265,
"refinementMs": 1134.045000076294,
"createdBuffers": 54,
"reusedBuffers": 140,
"peakTrackedGpuBytes": 3501008656,
"worstPointsNrmse": 0.0008415597114219503,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -0.7360437308425061
},
{
"profile": "detail",
"refineSteps": 3,
"disabled": {
"computeMs": 2681.0350000858307,
"refinementMs": 1685.335000038147,
"createdBuffers": 291,
"reusedBuffers": 0,
"peakTrackedGpuBytes": 3266447328,
"worstPointsNrmse": 0.0009430723759879896,
"maskMismatches": 0
},
"enabled": {
"computeMs": 2686.9649999141693,
"refinementMs": 1667.9150000810623,
"createdBuffers": 80,
"reusedBuffers": 211,
"peakTrackedGpuBytes": 3501326768,
"worstPointsNrmse": 0.0009430723759879896,
"maskMismatches": 0
},
"maxNrmseChangeAcrossReuseModes": 0,
"latencyReductionPercent": -0.22118323066087076
}
],
"gpuLayers": [
{
"profile": "detail",
"layer": "down_stages.4.0",
"medianGpuMsAcrossThreePasses": 181.338112
},
{
"profile": "detail",
"layer": "bottleneck_stage.0",
"medianGpuMsAcrossThreePasses": 180.09292799999997
},
{
"profile": "detail",
"layer": "down_stages.3.0",
"medianGpuMsAcrossThreePasses": 163.77446399999997
},
{
"profile": "detail",
"layer": "up_stages.0.0",
"medianGpuMsAcrossThreePasses": 163.05356799999998
},
{
"profile": "detail",
"layer": "down_stages.2.0",
"medianGpuMsAcrossThreePasses": 142.802944
},
{
"profile": "detail",
"layer": "up_stages.1.0",
"medianGpuMsAcrossThreePasses": 142.016512
},
{
"profile": "fast",
"layer": "down_stages.4.0",
"medianGpuMsAcrossThreePasses": 130.482176
},
{
"profile": "fast",
"layer": "bottleneck_stage.0",
"medianGpuMsAcrossThreePasses": 129.499136
},
{
"profile": "fast",
"layer": "down_stages.3.0",
"medianGpuMsAcrossThreePasses": 117.243904
},
{
"profile": "fast",
"layer": "up_stages.0.0",
"medianGpuMsAcrossThreePasses": 117.243904
},
{
"profile": "detail",
"layer": "down_stages.1.0",
"medianGpuMsAcrossThreePasses": 116.19532799999999
},
{
"profile": "detail",
"layer": "up_stages.2.0",
"medianGpuMsAcrossThreePasses": 114.884608
},
{
"profile": "fast",
"layer": "down_stages.2.0",
"medianGpuMsAcrossThreePasses": 101.384192
},
{
"profile": "fast",
"layer": "up_stages.1.0",
"medianGpuMsAcrossThreePasses": 100.27008000000001
},
{
"profile": "detail",
"layer": "up_stages.3.0",
"medianGpuMsAcrossThreePasses": 83.689472
},
{
"profile": "detail",
"layer": "down_stages.0.0",
"medianGpuMsAcrossThreePasses": 82.57536000000002
},
{
"profile": "fast",
"layer": "down_stages.1.0",
"medianGpuMsAcrossThreePasses": 80.412672
},
{
"profile": "fast",
"layer": "up_stages.2.0",
"medianGpuMsAcrossThreePasses": 79.95392000000001
},
{
"profile": "fast",
"layer": "up_stages.3.0",
"medianGpuMsAcrossThreePasses": 55.967743999999996
},
{
"profile": "fast",
"layer": "down_stages.0.0",
"medianGpuMsAcrossThreePasses": 55.508992
},
{
"profile": "detail",
"layer": "gather",
"medianGpuMsAcrossThreePasses": 30.867455999999997
},
{
"profile": "fast",
"layer": "gather",
"medianGpuMsAcrossThreePasses": 21.233664
},
{
"profile": "detail",
"layer": "upsampleAdd",
"medianGpuMsAcrossThreePasses": 10.616832
},
{
"profile": "fast",
"layer": "upsampleAdd",
"medianGpuMsAcrossThreePasses": 7.4711039999999995
},
{
"profile": "detail",
"layer": "pool",
"medianGpuMsAcrossThreePasses": 5.439487999999999
},
{
"profile": "fast",
"layer": "pool",
"medianGpuMsAcrossThreePasses": 3.5389440000000003
}
],
"passed": true
}