{
  "cases": [
    {
      "name": "batch32_time128_hidden64",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [32, 128, 64] },
        "sequence_lens": { "dtype": "uint32", "shape": [32], "data": { "kind": "constant", "value": 128 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [32, 128, 64] } }
    },
    {
      "name": "batch64_time512_hidden512",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [64, 512, 512] },
        "sequence_lens": { "dtype": "uint32", "shape": [64], "data": { "kind": "constant", "value": 512 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [64, 512, 512] } },
      "preset": "smoke"
    },
    {
      "name": "reverse-seq-scalar-fallback-inner-not-mul4-f32-4x2048x1023",
      "preset": "stress",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 8380416 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 2048, 1023], "dist": "normal", "seed": 5101, "scale": 1 },
        "sequence_lens": {
          "dtype": "uint32",
          "shape": [4],
          "data": { "kind": "values", "values": [2048, 1024, 2048, 512] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 2048, 1023] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "reverse-seq-scalar-fallback-inner-is-time-f32-4x524288",
      "preset": "stress",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 2097152 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 524288] },
        "sequence_lens": {
          "dtype": "uint32",
          "shape": [4],
          "data": { "kind": "values", "values": [524288, 262144, 524288, 131072] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 524288] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x257x3",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 3084 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 257, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 257, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x257x6",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 6168 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 257, 6], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 257, 6] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x257x7",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 7196 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 257, 7], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 257, 7] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x257x3x4",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 257, 3, 4], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 257, 3, 4] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x257x4x3",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 257, 4, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 257, 4, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_257x4x3",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 3084 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [257, 4, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [257, 4, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_257x4x6",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 6168 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [257, 4, 6], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [257, 4, 6] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_257x4x7",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 7196 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [257, 4, 7], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [257, 4, 7] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_257x4x3x4",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [257, 4, 3, 4], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [257, 4, 3, 4] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_257x4x4x3",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [257, 4, 4, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [257, 4, 4, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x257x3",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 3084 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 257, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 257, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x257x6",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 6168 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 257, 6], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 257, 6] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x257x7",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 7196 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 257, 7], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 257, 7] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x257x3x4",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 257, 3, 4], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 257, 3, 4] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x257x4x3",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 257, 4, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 257, 4, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_257x4x3",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 3084 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [257, 4, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [257, 4, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_257x4x6",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 6168 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [257, 4, 6], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [257, 4, 6] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_257x4x7",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 7196 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [257, 4, 7], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [257, 4, 7] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_257x4x3x4",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [257, 4, 3, 4], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [257, 4, 3, 4] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_257x4x4x3",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 12336 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [257, 4, 4, 3], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 128, 257] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [257, 4, 4, 3] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x2048x1022",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 8372224 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 2048, 1022], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 2048, 1022] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x2048x1023",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 8380416 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 2048, 1023], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 2048, 1023] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x2048x1025",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float32", "count": 8396800 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 2048, 1025], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 2048, 1025] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis0_4x8193x65_wg8",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "tunables": { "WORKGROUP_SIZE": 8 },
      "vars": { "dtype": "float32", "count": 2130180 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [4, 8193, 65], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 4096, 8193] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 8193, 65] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_2048x4x1022",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 8372224 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [2048, 4, 1022], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2048, 4, 1022] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_2048x4x1023",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 8380416 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [2048, 4, 1023], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2048, 4, 1023] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_2048x4x1025",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float32", "count": 8396800 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [2048, 4, 1025], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2048, 4, 1025] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float32_batchaxis1_8193x4x65_wg8",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "tunables": { "WORKGROUP_SIZE": 8 },
      "vars": { "dtype": "float32", "count": 2130180 },
      "inputs": {
        "input": { "dtype": "float32", "shape": [8193, 4, 65], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 4096, 8193] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8193, 4, 65] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x2048x1022",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 8372224 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 2048, 1022], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 2048, 1022] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x2048x1023",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 8380416 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 2048, 1023], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 2048, 1023] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x2048x1025",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 8396800 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 2048, 1025], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 2048, 1025] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_4x8193x65_wg8",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "tunables": { "WORKGROUP_SIZE": 8 },
      "vars": { "dtype": "float16", "count": 2130180 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 8193, 65], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 4096, 8193] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 8193, 65] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_2048x4x1022",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 8372224 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [2048, 4, 1022], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2048, 4, 1022] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_2048x4x1023",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 8380416 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [2048, 4, 1023], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2048, 4, 1023] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_2048x4x1025",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 8396800 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [2048, 4, 1025], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 1024, 2048] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2048, 4, 1025] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_8193x4x65_wg8",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "tunables": { "WORKGROUP_SIZE": 8 },
      "vars": { "dtype": "float16", "count": 2130180 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [8193, 4, 65], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 1, 4096, 8193] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [8193, 4, 65] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis0_8x8193x65_wg8_packed_fold",
      "preset": "edge",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "tunables": { "WORKGROUP_SIZE": 8 },
      "vars": { "dtype": "float16", "count": 4260360 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [8, 8193, 65], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": {
          "dtype": "uint32",
          "shape": [8],
          "data": { "kind": "values", "values": [0, 1, 4096, 8193, 0, 1, 4096, 8193] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [8, 8193, 65] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_batchaxis1_8193x8x65_wg8_packed_fold",
      "preset": "edge",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "tunables": { "WORKGROUP_SIZE": 8 },
      "vars": { "dtype": "float16", "count": 4260360 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [8193, 8, 65], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": {
          "dtype": "uint32",
          "shape": [8],
          "data": { "kind": "values", "values": [0, 1, 4096, 8193, 0, 1, 4096, 8193] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [8193, 8, 65] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_rank2_large_batchaxis0",
      "preset": "stress",
      "attrs": { "batch_axis": 0, "time_axis": 1 },
      "vars": { "dtype": "float16", "count": 2097152 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [4, 524288], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": {
          "dtype": "uint32",
          "shape": [4],
          "data": { "kind": "values", "values": [0, 1, 262144, 524288] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4, 524288] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    },
    {
      "name": "address_float16_rank2_large_batchaxis1",
      "preset": "stress",
      "attrs": { "batch_axis": 1, "time_axis": 0 },
      "vars": { "dtype": "float16", "count": 2097152 },
      "inputs": {
        "input": { "dtype": "float16", "shape": [524288, 4], "dist": "normal", "seed": 3401, "scale": 1 },
        "sequence_lens": {
          "dtype": "uint32",
          "shape": [4],
          "data": { "kind": "values", "values": [0, 1, 262144, 524288] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [524288, 4] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    }
  ]
}