{
  "op": "ai.onnx.Softmax",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "softmax-f32-256x1024",
      "preset": "smoke",
      "vars": { "dtype": "float32", "rows": 256, "cols": 1024 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [256, 1024], "dtype": "float32", "dist": "normal", "seed": 107, "scale": 2 } },
      "outputs": { "y": { "shape": [256, 1024], "dtype": "float32" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "softmax-f32-4096x128-max-negative",
      "preset": "smoke",
      "vars": { "dtype": "float32", "rows": 4096, "cols": 128 },
      "attrs": { "axis": 1 },
      "inputs": {
        "x": { "shape": [4096, 128], "dtype": "float32", "dist": "constant", "value": -3.4028234663852886e+38 }
      },
      "outputs": { "y": { "shape": [4096, 128], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-f32-1024x2048",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 1024, "cols": 2048 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [1024, 2048], "dtype": "float32", "dist": "normal", "seed": 108, "scale": 2 } },
      "outputs": { "y": { "shape": [1024, 2048], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "gemma4-decode-logits-1x262144",
      "preset": "model",
      "vars": { "dtype": "float32", "rows": 1, "cols": 262144 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [1, 262144], "dtype": "float32", "dist": "normal", "seed": 380, "scale": 4 } },
      "outputs": { "y": { "shape": [1, 262144], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "llama-prefill-attn-scores-16384x512",
      "preset": "model",
      "vars": { "dtype": "float32", "rows": 16384, "cols": 512 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [16384, 512], "dtype": "float32", "dist": "normal", "seed": 381, "scale": 2 } },
      "outputs": { "y": { "shape": [16384, 512], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-f32-8192x1023-odd-cols",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 8192, "cols": 1023 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [8192, 1023], "dtype": "float32", "dist": "normal", "seed": 109, "scale": 2 } },
      "outputs": { "y": { "shape": [8192, 1023], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "llama-prefill-attn-scores-16384x512-f16",
      "preset": "model",
      "vars": { "dtype": "float16", "rows": 16384, "cols": 512 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [16384, 512], "dtype": "float16", "dist": "normal", "seed": 382, "scale": 2 } },
      "outputs": { "y": { "shape": [16384, 512], "dtype": "float16" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-f32-2048x8192",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 2048, "cols": 8192 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [2048, 8192], "dtype": "float32", "dist": "normal", "seed": 410, "scale": 2 } },
      "outputs": { "y": { "shape": [2048, 8192], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-f32-2048x8191-odd-cols",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 2048, "cols": 8191 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [2048, 8191], "dtype": "float32", "dist": "normal", "seed": 411, "scale": 2 } },
      "outputs": { "y": { "shape": [2048, 8191], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-strided-rank3-16x2048x128-axis1",
      "preset": "stress",
      "vars": { "dtype": "float32", "outer": 16, "axisDim": 2048, "inner": 128 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [16, 2048, 128], "dtype": "float32", "dist": "normal", "seed": 412, "scale": 2 } },
      "outputs": { "y": { "shape": [16, 2048, 128], "dtype": "float32" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.outer * args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "softmax-f16-2048x8192",
      "preset": "stress",
      "vars": { "dtype": "float16", "rows": 2048, "cols": 8192 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [2048, 8192], "dtype": "float16", "dist": "normal", "seed": 413, "scale": 2 } },
      "outputs": { "y": { "shape": [2048, 8192], "dtype": "float16" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-f32-packed-rows-131072x32-healthy",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 131072, "cols": 32 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [131072, 32], "dtype": "float32", "dist": "normal", "seed": 700, "scale": 2 } },
      "outputs": { "y": { "shape": [131072, 32], "dtype": "float32", "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-f32-nosg-2097152x32-stable3pass-launchbound",
      "preset": "stress",
      "vars": { "dtype": "float32", "rows": 2097152, "cols": 32 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [2097152, 32], "dtype": "float32", "dist": "normal", "seed": 701, "scale": 2 } },
      "outputs": { "y": { "shape": [2097152, 32], "dtype": "float32", "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "softmax-f16-strided-rank3-256x2048x130-axis1-uncoalesced",
      "preset": "stress",
      "vars": { "dtype": "float16", "outer": 256, "axisDim": 2048, "inner": 130 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [256, 2048, 130], "dtype": "float16", "dist": "normal", "seed": 702, "scale": 2 } },
      "outputs": { "y": { "shape": [256, 2048, 130], "dtype": "float16", "dist": "empty" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.outer * args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "softmax-f32-strided-rank3-256x2048x130-axis1-inner-not-mod4",
      "preset": "stress",
      "vars": { "dtype": "float32", "outer": 256, "axisDim": 2048, "inner": 130 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [256, 2048, 130], "dtype": "float32", "dist": "normal", "seed": 703, "scale": 2 } },
      "outputs": { "y": { "shape": [256, 2048, 130], "dtype": "float32", "dist": "empty" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.outer * args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "softmax-f32-64x262144-longrow-rows",
      "preset": "smoke",
      "vars": { "dtype": "float32", "rows": 2048, "cols": 8192 },
      "attrs": { "axis": 1 },
      "inputs": { "x": { "shape": [64, 262144], "dtype": "float32", "dist": "normal", "seed": 410, "scale": 2 } },
      "outputs": { "y": { "shape": [64, 262144], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] }
    }
  ]
}