{
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "gather-axis0-f32-4096x128-idx512",
      "preset": "smoke",
      "vars": { "rows": 4096, "cols": 128, "indexCount": 512 },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [4096, 128], "dtype": "float32", "dist": "normal", "seed": 121, "scale": 0.2 },
        "indices": { "shape": [512], "dtype": "int32", "dist": "linearMod", "seed": 121, "step": 17, "mod": 4096 }
      },
      "outputs": { "output": { "shape": [512, 128], "dtype": "float32" } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * 4 * 2" }] }
    },
    {
      "name": "gather-axis1-f32-1024x512-idx128",
      "preset": "smoke",
      "vars": { "rows": 1024, "cols": 512, "indexCount": 128 },
      "attrs": { "axis": 1 },
      "inputs": {
        "data": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 122, "scale": 0.2 },
        "indices": { "shape": [128], "dtype": "int32", "dist": "linearMod", "seed": 122, "step": 17, "mod": 512 }
      },
      "outputs": { "output": { "shape": [1024, 128], "dtype": "float32" } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.indexCount * 4 * 2" }] }
    },
    {
      "name": "gather-axis0-u32-bool-like-4096x128-idx512",
      "preset": "edge",
      "vars": { "rows": 4096, "cols": 128, "indexCount": 512, "dtype": "uint32" },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [4096, 128], "dtype": "uint32", "dist": "linearMod", "mod": 2 },
        "indices": { "shape": [512], "dtype": "int32", "dist": "linearMod", "seed": 123, "step": 17, "mod": 4096 }
      },
      "outputs": { "output": { "shape": [512, 128], "dtype": "uint32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "gather-embedding-f16-32000x512-idx1024",
      "preset": "smoke",
      "vars": { "rows": 32000, "cols": 512, "indexCount": 1024, "dtype": "float16" },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [32000, 512], "dtype": "float16", "dist": "normal", "seed": 124, "scale": 0.2 },
        "indices": { "shape": [1024], "dtype": "int32", "dist": "linearMod", "seed": 124, "step": 29, "mod": 32000 }
      },
      "outputs": { "output": { "shape": [1024, 512], "dtype": "float16" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "gather-embedding-f16-32000x513-idx1024-unaligned",
      "preset": "edge",
      "vars": { "rows": 32000, "cols": 513, "indexCount": 1024, "dtype": "float16" },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [32000, 513], "dtype": "float16", "dist": "normal", "seed": 131, "scale": 0.2 },
        "indices": { "shape": [1024], "dtype": "int32", "dist": "linearMod", "seed": 131, "step": 29, "mod": 32000 }
      },
      "outputs": { "output": { "shape": [1024, 513], "dtype": "float16" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "gather-axis0-f32-4096x127-idx512-unaligned",
      "preset": "edge",
      "vars": { "rows": 4096, "cols": 127, "indexCount": 512, "dtype": "float32" },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [4096, 127], "dtype": "float32", "dist": "normal", "seed": 132, "scale": 0.2 },
        "indices": { "shape": [512], "dtype": "int32", "dist": "linearMod", "seed": 132, "step": 17, "mod": 4096 }
      },
      "outputs": { "output": { "shape": [512, 127], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "gather-axis0-f32-50000x16-idx2048-narrow-rows",
      "preset": "edge",
      "vars": { "rows": 50000, "cols": 16, "indexCount": 2048, "dtype": "float32" },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [50000, 16], "dtype": "float32", "dist": "normal", "seed": 133, "scale": 0.2 },
        "indices": { "shape": [2048], "dtype": "int32", "dist": "linearMod", "seed": 133, "step": 29, "mod": 50000 }
      },
      "outputs": { "output": { "shape": [2048, 16], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "gather-axis0-f32-1000000x128-idx131072-saturated-grid",
      "preset": "edge",
      "vars": { "rows": 1000000, "cols": 128, "indexCount": 131072, "dtype": "float32" },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [1000000, 128], "dtype": "float32", "dist": "normal", "seed": 134, "scale": 0.2 },
        "indices": { "shape": [131072], "dtype": "int32", "dist": "linearMod", "seed": 134, "step": 7, "mod": 1000000 }
      },
      "outputs": { "output": { "shape": [131072, 128], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "gather-axis1-f32-8x32000x128-idx1024-pathology",
      "preset": "edge",
      "provenance": {
        "notes": "Axis-1 gather (per-head or per-expert row selection) over an outer extent of 8: the row-vec4 route requires axis 0, so the generic per-element kernel is the selected route. Its control gathers the same number of 128-wide rows along axis 0."
      },
      "vars": { "outer": 8, "rows": 32000, "cols": 128, "indexCount": 1024, "dtype": "float32" },
      "attrs": { "axis": 1 },
      "inputs": {
        "data": { "shape": [8, 32000, 128], "dtype": "float32", "dist": "normal", "seed": 140, "scale": 0.2 },
        "indices": { "shape": [1024], "dtype": "int32", "dist": "linearMod", "seed": 140, "step": 17, "mod": 32000 }
      },
      "outputs": { "output": { "shape": [8, 1024, 128], "dtype": "float32" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.outer * args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "gather-axis0-f32-32000x128-idx8192-control",
      "preset": "edge",
      "provenance": {
        "notes": "Axis-0 control for the axis-1 gather: 8192 distinct 128-wide rows through the row-vec4 route, moving the same output bytes."
      },
      "vars": { "rows": 32000, "cols": 128, "indexCount": 8192, "dtype": "float32" },
      "attrs": { "axis": 0 },
      "inputs": {
        "data": { "shape": [32000, 128], "dtype": "float32", "dist": "normal", "seed": 141, "scale": 0.2 },
        "indices": { "shape": [8192], "dtype": "int32", "dist": "linearMod", "seed": 141, "step": 17, "mod": 32000 }
      },
      "outputs": { "output": { "shape": [8192, 128], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    }
  ]
}