{
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "uint8_1m_scalar",
      "preset": "smoke",
      "vars": { "inputDtype": "uint8", "outputDtype": "float32", "count": 1048576 },
      "inputs": {
        "x": { "dtype": "uint8", "shape": [1024, 1024], "data": { "kind": "constant", "value": 127 } },
        "x_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } },
        "x_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1024, 1024] } },
      "bench": {
        "primary": true,
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "uint8_1m_no_zero_point",
      "preset": "smoke",
      "vars": { "inputDtype": "uint8", "outputDtype": "float32", "count": 1048576 },
      "inputs": {
        "x": { "dtype": "uint8", "shape": [1024, 1024], "data": { "kind": "constant", "value": 127 } },
        "x_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1024, 1024] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "uint8_1m_plus3_scalar",
      "preset": "smoke",
      "vars": { "inputDtype": "uint8", "outputDtype": "float32", "count": 1048579 },
      "inputs": {
        "x": { "dtype": "uint8", "shape": [1048579], "data": { "kind": "constant", "value": 127 } },
        "x_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } },
        "x_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1048579] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "int8_1m_plus3_no_zero_point",
      "preset": "smoke",
      "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 1048579 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [1048579], "data": { "kind": "constant", "value": -42 } },
        "x_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1048579] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "int8_1m_f16",
      "preset": "smoke",
      "vars": { "inputDtype": "int8", "outputDtype": "float16", "count": 1048576 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [1024, 1024], "data": { "kind": "constant", "value": -42 } },
        "x_scale": { "dtype": "float16", "shape": [1], "data": { "kind": "values", "values": [0.125] } },
        "x_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [1024, 1024] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "int8_per_axis0_512x2048",
      "preset": "smoke",
      "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 1048576 },
      "attrs": { "axis": 0 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [512, 2048], "data": { "kind": "constant", "value": -42 } },
        "x_scale": {
          "dtype": "float32",
          "shape": [512],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
        },
        "x_zero_point": { "dtype": "int8", "shape": [512], "data": { "kind": "constant", "value": 3 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [512, 2048] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "per_channel_innermost_axis1_scalar_fallback_2048x2048",
      "preset": "smoke",
      "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 4194304 },
      "attrs": { "axis": 1 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [2048, 2048], "data": { "kind": "constant", "value": -42 } },
        "x_scale": {
          "dtype": "float32",
          "shape": [2048],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
        },
        "x_zero_point": { "dtype": "int8", "shape": [2048], "data": { "kind": "constant", "value": 3 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2048, 2048] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "per_channel_axis0_vec4_healthy_2048x2048",
      "preset": "smoke",
      "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 4194304 },
      "attrs": { "axis": 0 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [2048, 2048], "data": { "kind": "constant", "value": -42 } },
        "x_scale": {
          "dtype": "float32",
          "shape": [2048],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
        },
        "x_zero_point": { "dtype": "int8", "shape": [2048], "data": { "kind": "constant", "value": 3 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2048, 2048] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "per_channel_axis0_unaligned_inner2050_scalar_cliff",
      "preset": "smoke",
      "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 1049600 },
      "attrs": { "axis": 0 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [512, 2050], "data": { "kind": "constant", "value": -42 } },
        "x_scale": {
          "dtype": "float32",
          "shape": [512],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
        },
        "x_zero_point": { "dtype": "int8", "shape": [512], "data": { "kind": "constant", "value": 3 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [512, 2050] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "scalar_above_67M_vec4_ceiling_fallback",
      "preset": "stress",
      "provenance": {
        "notes": "Stress-only capacity case: uint8 elements use 4-byte WebGPU storage, bringing the declared footprint to 512 MiB."
      },
      "vars": { "inputDtype": "uint8", "outputDtype": "float32", "count": 67108864 },
      "attrs": { "axis": 0 },
      "inputs": {
        "x": { "dtype": "uint8", "shape": [8192, 8192], "data": { "kind": "constant", "value": 127 } },
        "x_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } },
        "x_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8192, 8192] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "f16_tail_scalar_x4_1m_plus3",
      "preset": "smoke",
      "vars": { "inputDtype": "int8", "outputDtype": "float16", "count": 1048579 },
      "attrs": { "axis": 0 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [1048579], "data": { "kind": "constant", "value": -42 } },
        "x_scale": { "dtype": "float16", "shape": [1], "data": { "kind": "values", "values": [0.125] } },
        "x_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [1048579] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "blocked_rank4_axis1_bs32_stress_4m",
      "preset": "stress",
      "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 4194304 },
      "attrs": { "axis": 1, "block_size": 32 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [4, 512, 32, 64], "dist": "randint", "seed": 101, "min": -128, "max": 127 },
        "x_scale": { "dtype": "float32", "shape": [4, 16, 32, 64], "dist": "normal", "seed": 102, "scale": 0.02 },
        "x_zero_point": {
          "dtype": "int8",
          "shape": [4, 16, 32, 64],
          "dist": "randint",
          "seed": 103,
          "min": -8,
          "max": 8
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4, 512, 32, 64], "dist": "empty" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    },
    {
      "name": "per_axis_mid_axis1_rank3_inner_not_mult4_scalar_fallback_stress",
      "preset": "stress",
      "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 8396800 },
      "attrs": { "axis": 1 },
      "inputs": {
        "x": { "dtype": "int8", "shape": [8, 512, 2050], "dist": "randint", "seed": 201, "min": -128, "max": 127 },
        "x_scale": { "dtype": "float32", "shape": [512], "dist": "normal", "seed": 202, "scale": 0.02 },
        "x_zero_point": { "dtype": "int8", "shape": [512], "dist": "randint", "seed": 203, "min": -8, "max": 8 }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8, 512, 2050], "dist": "empty" } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
        ]
      }
    }
  ]
}