{
  "op": "ai.onnx.LogSoftmax",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "attention_scores",
      "inputs": { "x": { "dtype": "float32", "shape": [32, 128] } },
      "outputs": { "y": { "dtype": "float32", "shape": [32, 128] } }
    },
    {
      "name": "logsoftmax-f32-1024x2048",
      "preset": "smoke",
      "vars": { "dtype": "float32", "rows": 1024, "cols": 2048 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [1024, 2048], "dtype": "float32", "dist": "normal", "seed": 211, "scale": 2 } },
      "outputs": { "y": { "shape": [1024, 2048], "dtype": "float32" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "logsoftmax-f32-16384x512",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 16384, "cols": 512 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [16384, 512], "dtype": "float32", "dist": "normal", "seed": 212, "scale": 2 } },
      "outputs": { "y": { "shape": [16384, 512], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "logsoftmax-logits-1x262144",
      "preset": "model",
      "vars": { "dtype": "float32", "rows": 1, "cols": 262144 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [1, 262144], "dtype": "float32", "dist": "normal", "seed": 213, "scale": 4 } },
      "outputs": { "y": { "shape": [1, 262144], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "logsoftmax-f32-strided-axis1-64x1024x64",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 64, "mid": 1024, "inner": 64 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [64, 1024, 64], "dtype": "float32", "dist": "normal", "seed": 311, "scale": 2 } },
      "outputs": { "y": { "shape": [64, 1024, 64], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.rows * args.mid * args.inner * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "logsoftmax-f16-1024x2048-axis1",
      "preset": "stress",
      "vars": { "dtype": "float16", "rows": 1024, "cols": 2048 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [1024, 2048], "dtype": "float16", "dist": "normal", "seed": 312, "scale": 2 } },
      "outputs": { "y": { "shape": [1024, 2048], "dtype": "float16" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "logsoftmax-f32-align-cliff-1024x2046",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 1024, "cols": 2046 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [1024, 2046], "dtype": "float32", "dist": "normal", "seed": 313, "scale": 2 } },
      "outputs": { "y": { "shape": [1024, 2046], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "logsoftmax-f32-tallwide-70000x2048",
      "preset": "model",
      "vars": { "dtype": "float32", "rows": 70000, "cols": 2048 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [70000, 2048], "dtype": "float32", "dist": "normal", "seed": 314, "scale": 2 } },
      "outputs": { "y": { "shape": [70000, 2048], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "logsoftmax-f32-nosg-3pass-occupancy-131072x32",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 131072, "cols": 32 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [131072, 32], "dtype": "float32", "dist": "normal", "seed": 411, "scale": 2 } },
      "outputs": { "y": { "shape": [131072, 32], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "logsoftmax-f32-strided-axis1-64x2048x130-packed4-control",
      "preset": "stress",
      "provenance": {
        "source": "strided-axis selector crossover",
        "notes": "Below-crossover control for the large-inner tail pathology. At 8,320 independent strided rows the cooperative packed-4 reducer still beats the serial-lane capacity path."
      },
      "vars": { "dtype": "float32", "outer": 64, "axisDim": 2048, "inner": 130 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [64, 2048, 130], "dtype": "float32", "dist": "normal", "seed": 414, "scale": 2 } },
      "outputs": { "y": { "shape": [64, 2048, 130], "dtype": "float32", "dist": "empty" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.outer * args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "logsoftmax-f32-strided-axis1-96x2048x130-crossover",
      "preset": "stress",
      "provenance": {
        "source": "strided-axis selector crossover",
        "notes": "Above-crossover probe for the large-inner tail pathology. At 12,480 independent strided rows the serial-lane capacity path overtakes the cooperative packed-4 reducer."
      },
      "vars": { "dtype": "float32", "outer": 96, "axisDim": 2048, "inner": 130 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [96, 2048, 130], "dtype": "float32", "dist": "normal", "seed": 413, "scale": 2 } },
      "outputs": { "y": { "shape": [96, 2048, 130], "dtype": "float32", "dist": "empty" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.outer * args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "logsoftmax-f32-strided-axis1-120x2048x130-large-inner",
      "preset": "stress",
      "provenance": {
        "source": "authored for variant coverage",
        "notes": "Inner=130 exercises the two-element tail of the scalar-aligned packed-4 online reducer while sharing one workgroup across four adjacent rows."
      },
      "vars": { "dtype": "float32", "outer": 120, "axisDim": 2048, "inner": 130 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [120, 2048, 130], "dtype": "float32", "dist": "normal", "seed": 412, "scale": 2 } },
      "outputs": { "y": { "shape": [120, 2048, 130], "dtype": "float32", "dist": "empty" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.outer * args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "logsoftmax-f32-8192x512-midwidth",
      "provenance": {
        "notes": "Covers the 512-column midpoint between the 128- and 2048-column cases and probes the width term in online_wg_vec4's selector."
      },
      "preset": "smoke",
      "vars": { "dtype": "float32", "rows": 8192, "cols": 512 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [8192, 512], "dtype": "float32", "dist": "normal", "seed": 215, "scale": 2 } },
      "outputs": { "y": { "shape": [8192, 512], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    }
  ]
}