{
  "op": "ai.onnx.ReduceSumSquare",
  "cases": [
    {
      "name": "reducesumsquare-axis0-f32-1024x512",
      "preset": "smoke",
      "vars": { "rows": 1024, "cols": 512 },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 118, "scale": 0.2 } },
      "outputs": { "y": { "shape": [512], "dtype": "float32" } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
    },
    {
      "name": "1024x1024_axis1",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [1024, 1024] } },
      "outputs": { "y": { "dtype": "float32", "shape": [1024] } }
    },
    {
      "name": "reducesumsquare-lastaxis-f32-4096x4096",
      "preset": "smoke",
      "vars": { "rows": 4096, "cols": 4096 },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": { "x": { "shape": [4096, 4096], "dtype": "float32", "dist": "normal", "seed": 117, "scale": 0.2 } },
      "outputs": { "y": { "shape": [4096], "dtype": "float32" } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
    },
    {
      "name": "reducesumsquare-axis0-splitk-f32-8192x512",
      "vars": { "rows": 8192, "cols": 512 },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "shape": [8192, 512], "dtype": "float32", "dist": "normal", "seed": 211, "scale": 0.2 } },
      "outputs": { "y": { "shape": [512], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
    },
    {
      "name": "reducesumsquare-lastaxis-odd-f32-4096x4095",
      "vars": { "rows": 4096, "cols": 4095 },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": { "x": { "shape": [4096, 4095], "dtype": "float32", "dist": "normal", "seed": 212, "scale": 0.2 } },
      "outputs": { "y": { "shape": [4096], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
    },
    {
      "name": "reducesumsquare-rank4-axis1-f32-32x256x28x28",
      "vars": { "n": 32, "c": 256, "h": 28, "w": 28 },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": { "x": { "shape": [32, 256, 28, 28], "dtype": "float32", "dist": "normal", "seed": 213, "scale": 0.2 } },
      "outputs": { "y": { "shape": [32, 28, 28], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.n * args.c * args.h * args.w * 4" }] }
    },
    {
      "name": "reducesumsquare-fullreduce-serial-numel-not-mul4-r3",
      "preset": "stress",
      "attrs": { "keepdims": 0 },
      "inputs": { "x": { "shape": [1001, 1001, 1], "dtype": "float32", "dist": "normal", "seed": 701, "scale": 1 } },
      "outputs": { "y": { "shape": [], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "1001 * 1001 * 1 * 4" }] }
    },
    {
      "name": "reducesumsquare-int32-lastaxis-serial-4096x1024",
      "preset": "stress",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": { "shape": [4096, 1024], "dtype": "int32", "dist": "randint", "seed": 702, "min": -10, "max": 10 }
      },
      "outputs": { "y": { "shape": [4096], "dtype": "int32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "4096 * 1024 * 4" }] }
    },
    {
      "name": "reducesumsquare-rank3-spatial-axes12-f32-128x256x256-pathology",
      "preset": "stress",
      "provenance": {
        "source": "authored for branch coverage",
        "notes": "Rank-3 multi-axis reduction with 128 output lanes, each serially scanning a 256x256 plane."
      },
      "attrs": { "axes": [1, 2], "keepdims": 1 },
      "inputs": { "x": { "shape": [128, 256, 256], "dtype": "float32", "dist": "normal", "seed": 305, "scale": 0.2 } },
      "outputs": { "y": { "shape": [128, 1, 1], "dtype": "float32", "dist": "empty" } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "128 * 256 * 256 * 4" }] }
    },
    {
      "name": "reducesumsquare-spatial-axes23-f32-2x64x256x256-pathology",
      "preset": "stress",
      "provenance": {
        "source": "authored for branch coverage",
        "notes": "Rank-4 multi-axis reduction with 128 output lanes, each serially scanning a 256x256 plane."
      },
      "attrs": { "axes": [2, 3], "keepdims": 1 },
      "inputs": { "x": { "shape": [2, 64, 256, 256], "dtype": "float32", "dist": "normal", "seed": 304, "scale": 0.2 } },
      "outputs": { "y": { "shape": [2, 64, 1, 1], "dtype": "float32", "dist": "empty" } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "2 * 64 * 256 * 256 * 4" }] }
    },
    {
      "name": "reducesumsquare-axis0-f32-131072x15-narrow-splitk-boundary",
      "preset": "stress",
      "vars": { "rows": 131072, "cols": 15 },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "shape": [131072, 15], "dtype": "float32", "dist": "normal", "seed": 703, "scale": 0.1 } },
      "outputs": { "y": { "shape": [15], "dtype": "float32", "dist": "empty" } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
    }
  ],
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }
}