{
  "op": "ai.onnx.RNN",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256], "HIDDEN_BLOCK": [2, 4, 8] },
  "cases": [
    {
      "name": "small",
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 32 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [16, 1, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 }
        },
        "w": {
          "dtype": "float32",
          "shape": [1, 32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 }
        },
        "r": {
          "dtype": "float32",
          "shape": [1, 32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 }
        },
        "b": { "dtype": "float32", "shape": [1, 64], "data": { "kind": "constant", "value": 0.0 } }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [16, 1, 1, 32] },
        "y_h": { "dtype": "float32", "shape": [1, 1, 32] }
      },
      "bench": {
        "metrics": [
          {
            "type": "gflops",
            "value": "2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) * (dim(shapes.x, 2) + dim(shapes.r, 2))"
          },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "h256-in64-b8-seq32-vec4",
      "preset": "smoke",
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 256 },
      "vars": { "seq": 32, "batch": 8, "inputSize": 64, "hidden": 256 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [32, 8, 64], "dist": "normal", "seed": 441, "scale": 0.3 },
        "w": { "dtype": "float32", "shape": [1, 256, 64], "dist": "normal", "seed": 442, "scale": 0.05 },
        "r": { "dtype": "float32", "shape": [1, 256, 256], "dist": "normal", "seed": 443, "scale": 0.02 },
        "b": { "dtype": "float32", "shape": [1, 512], "dist": "normal", "seed": 444, "scale": 0.05 }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [32, 1, 8, 256] },
        "y_h": { "dtype": "float32", "shape": [1, 8, 256] }
      },
      "bench": {
        "primary": true,
        "metrics": [
          { "type": "gflops", "value": "args.seq * args.batch * 2 * args.hidden * (args.inputSize + args.hidden)" },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "h102-in80-b4-seq32-mixed-vector",
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 102 },
      "vars": { "seq": 32, "batch": 4, "inputSize": 80, "hidden": 102 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [32, 4, 80], "dist": "normal", "seed": 451, "scale": 0.3 },
        "w": { "dtype": "float32", "shape": [1, 102, 80], "dist": "normal", "seed": 452, "scale": 0.05 },
        "r": { "dtype": "float32", "shape": [1, 102, 102], "dist": "normal", "seed": 453, "scale": 0.05 },
        "b": { "dtype": "float32", "shape": [1, 204], "dist": "normal", "seed": 454, "scale": 0.05 }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [32, 1, 4, 102] },
        "y_h": { "dtype": "float32", "shape": [1, 4, 102] }
      },
      "bench": {
        "metrics": [
          { "type": "gflops", "value": "args.seq * args.batch * 2 * args.hidden * (args.inputSize + args.hidden)" },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "large-h512-in64-b2-seq4-vec4",
      "preset": "model",
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 512 },
      "vars": { "seq": 4, "batch": 2, "inputSize": 64, "hidden": 512 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [4, 2, 64], "dist": "normal", "seed": 501, "scale": 0.3 },
        "w": { "dtype": "float32", "shape": [1, 512, 64], "dist": "normal", "seed": 502, "scale": 0.05 },
        "r": { "dtype": "float32", "shape": [1, 512, 512], "dist": "normal", "seed": 503, "scale": 0.02 },
        "b": { "dtype": "float32", "shape": [1, 1024], "dist": "normal", "seed": 504, "scale": 0.05 }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [4, 1, 2, 512] },
        "y_h": { "dtype": "float32", "shape": [1, 2, 512] }
      },
      "bench": {
        "primary": false,
        "metrics": [
          { "type": "gflops", "value": "args.seq * args.batch * 2 * args.hidden * (args.inputSize + args.hidden)" },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "single_wg_global_hidden4096_b1_seq16_stress",
      "preset": "stress",
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 4096 },
      "vars": { "seq": 16, "batch": 1, "inputSize": 64, "hidden": 4096 },
      "tunableSpace": { "RECURRENT_WORKGROUP_SIZE": [1, 2, 4, 8, 16, 32, 64] },
      "inputs": {
        "x": { "dtype": "float32", "shape": [16, 1, 64], "dist": "normal", "seed": 771, "scale": 0.3 },
        "w": { "dtype": "float32", "shape": [1, 4096, 64], "dist": "normal", "seed": 772, "scale": 0.02 },
        "r": { "dtype": "float32", "shape": [1, 4096, 4096], "dist": "normal", "seed": 773, "scale": 0.005 },
        "b": { "dtype": "float32", "shape": [1, 8192], "dist": "normal", "seed": 774, "scale": 0.02 }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [16, 1, 1, 4096], "dist": "empty" },
        "y_h": { "dtype": "float32", "shape": [1, 1, 4096], "dist": "empty" }
      },
      "bench": {
        "primary": false,
        "metrics": [
          { "type": "gflops", "value": "args.seq * args.batch * 2 * args.hidden * (args.inputSize + args.hidden)" },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "single_wg_batch1_hidden2048_seq64_serial_recurrence_stress",
      "preset": "stress",
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 2048 },
      "vars": { "seq": 64, "batch": 1, "inputSize": 256, "hidden": 2048 },
      "tunableSpace": { "RECURRENT_WORKGROUP_SIZE": [1, 2, 4, 8, 16, 32, 64] },
      "inputs": {
        "x": { "dtype": "float32", "shape": [64, 1, 256], "dist": "normal", "seed": 781, "scale": 0.3 },
        "w": { "dtype": "float32", "shape": [1, 2048, 256], "dist": "normal", "seed": 782, "scale": 0.03 },
        "r": { "dtype": "float32", "shape": [1, 2048, 2048], "dist": "normal", "seed": 783, "scale": 0.01 },
        "b": { "dtype": "float32", "shape": [1, 4096], "dist": "normal", "seed": 784, "scale": 0.03 }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [64, 1, 1, 2048], "dist": "empty" },
        "y_h": { "dtype": "float32", "shape": [1, 1, 2048], "dist": "empty" }
      },
      "bench": {
        "primary": false,
        "metrics": [
          { "type": "gflops", "value": "args.seq * args.batch * 2 * args.hidden * (args.inputSize + args.hidden)" },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "speech-encoder-b1-seq128-in256-h1024-single-wg-pathology",
      "preset": "model",
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 1024 },
      "vars": { "seq": 128, "batch": 1, "inputSize": 256, "hidden": 1024 },
      "tunableSpace": { "RECURRENT_WORKGROUP_SIZE_MEDIUM_ROWS": [1, 2, 4, 8, 16, 32, 64] },
      "inputs": {
        "x": { "dtype": "float32", "shape": [128, 1, 256], "dist": "normal", "seed": 791, "scale": 0.3 },
        "w": { "dtype": "float32", "shape": [1, 1024, 256], "dist": "normal", "seed": 792, "scale": 0.03 },
        "r": { "dtype": "float32", "shape": [1, 1024, 1024], "dist": "normal", "seed": 793, "scale": 0.01 },
        "b": { "dtype": "float32", "shape": [1, 2048], "dist": "normal", "seed": 794, "scale": 0.03 }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [128, 1, 1, 1024], "dist": "empty" },
        "y_h": { "dtype": "float32", "shape": [1, 1, 1024], "dist": "empty" }
      },
      "bench": {
        "metrics": [
          { "type": "gflops", "value": "args.seq * args.batch * 2 * args.hidden * (args.inputSize + args.hidden)" },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "manifest-gap-sequence-lens-b8-h256-in256-seq64-general-path",
      "preset": "stress",
      "provenance": {
        "source": "authored for boundary coverage",
        "notes": "Full-length sequence_lens case that verifies the shared post-mask route keeps the projected parallel RNN path eligible."
      },
      "attrs": { "layout": 0, "direction": "forward", "hidden_size": 256 },
      "vars": { "seq": 64, "batch": 8, "inputSize": 256, "hidden": 256 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [64, 8, 256], "dist": "normal", "seed": 9181, "scale": 0.3 },
        "w": { "dtype": "float32", "shape": [1, 256, 256], "dist": "normal", "seed": 9182, "scale": 0.05 },
        "r": { "dtype": "float32", "shape": [1, 256, 256], "dist": "normal", "seed": 9183, "scale": 0.02 },
        "b": { "dtype": "float32", "shape": [1, 512], "dist": "normal", "seed": 9184, "scale": 0.05 },
        "sequence_lens": {
          "dtype": "int32",
          "shape": [8],
          "data": { "kind": "values", "values": [64, 64, 64, 64, 64, 64, 64, 64] }
        }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [64, 1, 8, 256], "dist": "empty" },
        "y_h": { "dtype": "float32", "shape": [1, 8, 256], "dist": "empty" }
      },
      "bench": {
        "metrics": [
          { "type": "gflops", "value": "args.seq * args.batch * 2 * args.hidden * (args.inputSize + args.hidden)" },
          {
            "type": "bandwidth",
            "name": "recurrent lower-bound BW",
            "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))"
          }
        ]
      }
    }
  ]
}