{
  "op": "ai.onnx.NonZero",
  "tunableSpace": {},
  "cases": [
    {
      "name": "f32_1024_capacity",
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 }
        }
      },
      "outputs": { "y": { "dtype": "uint32", "shape": [1, 1024] } }
    },
    {
      "name": "nonzero-rank2-u32-1m",
      "preset": "smoke",
      "inputs": { "x": { "dtype": "uint32", "shape": [1024, 1024], "dist": "linearMod", "mod": 4 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [2, 786432] } },
      "bench": {
        "primary": true,
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-exact-capacity-1024-f32",
      "preset": "smoke",
      "provenance": {
        "notes": "Legacy case name retained for longitudinal comparisons. At the measured crossover, 1024 elements now select the parallel scan. The traffic metric uses algorithm-independent logical I/O so an all-variant comparison weighs serial and parallel kernels fairly."
      },
      "inputs": { "x": { "dtype": "float32", "shape": [1024], "dist": "constant", "value": 1 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [1, 1024] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-parallel-1280-f32-healthy",
      "preset": "smoke",
      "inputs": { "x": { "dtype": "float32", "shape": [1280], "dist": "constant", "value": 1 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [1, 1280] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-rank2-f32-1m-dense-100pct",
      "preset": "smoke",
      "inputs": { "x": { "dtype": "float32", "shape": [1024, 1024], "dist": "constant", "value": 1 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [2, 1048576] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-rank2-f16-1m-dense",
      "preset": "smoke",
      "inputs": { "x": { "dtype": "float16", "shape": [1024, 1024], "dist": "constant", "value": 1 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [2, 1048576] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "2 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-dispatch-cliff-16m-f32-sparse",
      "preset": "smoke",
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [4096, 4096],
          "data": {
            "kind": "cycle",
            "values": [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
          }
        }
      },
      "outputs": { "y": { "dtype": "uint32", "shape": [2, 1048576] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-2dfold-boundary-16m-int32-exact",
      "preset": "stress",
      "inputs": { "x": { "dtype": "int32", "shape": [4096, 4097], "dist": "normal", "seed": 7, "scale": 2 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [2, 16781312] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-rank4-u32-1x32x512x512-exact-activation",
      "preset": "stress",
      "inputs": { "x": { "dtype": "uint32", "shape": [1, 32, 512, 512], "dist": "linearMod", "mod": 16 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [4, 7864320] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    },
    {
      "name": "nonzero-rank3-u32-32x512x512-exact-activation",
      "preset": "stress",
      "provenance": {
        "notes": "Rank-3 twin of the rank-4 activation case, same element count and density. The scatter writes one coordinate stream per rank, so it exists to place the items-per-thread threshold on measured rank-2/3/4 behaviour instead of extrapolating from two points."
      },
      "inputs": { "x": { "dtype": "uint32", "shape": [32, 512, 512], "dist": "linearMod", "mod": 16 } },
      "outputs": { "y": { "dtype": "uint32", "shape": [3, 7864320] } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "4 * numel(shapes.x) + 4 * numel(shapes.y) + 4 * (4 * ceilDiv(numel(shapes.x), 4) + 5 * ceilDiv(numel(shapes.x), 1024))"
          }
        ]
      }
    }
  ]
}