{
  "op": "ai.onnx.Concat",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "concat-f32-axis1",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 1048576 },
      "attrs": { "axis": 1 },
      "inputs": {
        "a": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 950, "scale": 1 },
        "b": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 951, "scale": 1 }
      },
      "outputs": { "c": { "shape": [1024, 1024], "dtype": "float32" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "concat-last-axis-f32-512x768-plus-256",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 524288 },
      "attrs": { "axis": -1 },
      "tunableSpace": { "LAST_AXIS_COOPERATIVE_WORKGROUP_SIZE": [32, 64, 128, 256] },
      "inputs": {
        "a": { "shape": [512, 768], "dtype": "float32", "dist": "normal", "seed": 360, "scale": 0.5 },
        "b": { "shape": [512, 256], "dtype": "float32", "dist": "normal", "seed": 361, "scale": 0.5 }
      },
      "outputs": { "c": { "shape": [512, 1024], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "concat-block-axis0-f32-8m",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 8388608 },
      "attrs": { "axis": 0 },
      "inputs": {
        "a": { "shape": [4096, 1024], "dtype": "float32", "dist": "normal", "seed": 952, "scale": 1 },
        "b": { "shape": [4096, 1024], "dtype": "float32", "dist": "normal", "seed": 953, "scale": 1 }
      },
      "outputs": { "c": { "shape": [8192, 1024], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "concat-block-axis1-f16-4m",
      "preset": "smoke",
      "vars": { "dtype": "float16", "count": 4194304 },
      "attrs": { "axis": 1 },
      "inputs": {
        "a": { "shape": [1024, 1024], "dtype": "float16", "dist": "normal", "seed": 954, "scale": 1 },
        "b": { "shape": [1024, 3072], "dtype": "float16", "dist": "normal", "seed": 955, "scale": 1 }
      },
      "outputs": { "c": { "shape": [1024, 4096], "dtype": "float16" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "concat-f32-axis1-unaligned",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 1048576 },
      "attrs": { "axis": 1 },
      "inputs": {
        "a": { "shape": [1024, 513], "dtype": "float32", "dist": "normal", "seed": 960, "scale": 1 },
        "b": { "shape": [1024, 511], "dtype": "float32", "dist": "normal", "seed": 961, "scale": 1 }
      },
      "outputs": { "c": { "shape": [1024, 1024], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "concat-f32-axis-last-inner1-odd",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 8388608 },
      "attrs": { "axis": 1 },
      "inputs": {
        "a": { "shape": [8192, 1], "dtype": "float32", "dist": "normal", "seed": 962, "scale": 1 },
        "b": { "shape": [8192, 1023], "dtype": "float32", "dist": "normal", "seed": 963, "scale": 1 }
      },
      "outputs": { "c": { "shape": [8192, 1024], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "concat-f32-axis1-nine-inputs",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 4718592 },
      "attrs": { "axis": 1 },
      "inputs": {
        "a": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 970, "scale": 1 },
        "b": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 971, "scale": 1 },
        "d": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 972, "scale": 1 },
        "e": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 973, "scale": 1 },
        "f": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 974, "scale": 1 },
        "g": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 975, "scale": 1 },
        "h": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 976, "scale": 1 },
        "i": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 977, "scale": 1 },
        "j": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 978, "scale": 1 }
      },
      "outputs": { "c": { "shape": [1024, 4608], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "concat-three-inputs-lastaxis-odd-scalar-stress",
      "preset": "stress",
      "vars": { "dtype": "float32", "count": 3145728 },
      "attrs": { "axis": -1 },
      "inputs": {
        "a": { "shape": [3072, 341], "dtype": "float32", "dist": "normal", "seed": 981, "scale": 1 },
        "b": { "shape": [3072, 342], "dtype": "float32", "dist": "normal", "seed": 982, "scale": 1 },
        "d": { "shape": [3072, 341], "dtype": "float32", "dist": "normal", "seed": 983, "scale": 1 }
      },
      "outputs": { "c": { "shape": [3072, 1024], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "concat-two-inputs-rank5-lastaxis-odd-scalar-stress",
      "preset": "stress",
      "vars": { "dtype": "float32", "count": 1048576 },
      "attrs": { "axis": -1 },
      "inputs": {
        "a": { "shape": [16, 16, 16, 64, 3], "dtype": "float32", "dist": "normal", "seed": 985, "scale": 1 },
        "b": { "shape": [16, 16, 16, 64, 1], "dtype": "float32", "dist": "normal", "seed": 986, "scale": 1 }
      },
      "outputs": { "c": { "shape": [16, 16, 16, 64, 4], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    }
  ]
}