{
  "fixtureArrays": {
    "rank3_axis1_middle_no_keepdims_input_x": [1, 2, 3, 4, -1, -2, -3, -4, 0.5, 1.5, 2.5, 3.5, 10, 20, 30, 40, -10, -20, -30, -40, 2, 4, 6, 8],
    "rank3_axis0_no_keepdims_input_x": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]
  },
  "cases": [
    {
      "name": "all_axes_flat_packed_f16_boundary_8192",
      "provenance": {
        "notes": "Aligned half storage at the parallel threshold; cancellation distinguishes f32 lane accumulation from arithmetic performed in half."
      },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "float16", "shape": [8192], "data": { "kind": "cycle", "values": [2048.0, 1.0, -2048.0, 1.0] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [], "tolerance": 0, "relTolerance": 0 } }
    },
    {
      "name": "all_axes_flat_packed_f16_nine_partials_8200",
      "provenance": {
        "notes": "Aligned full reduction with nine partial workgroups and inactive lanes; the exact sum 4100 is representable in half."
      },
      "attrs": { "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 41, 100],
          "data": { "kind": "cycle", "values": [2048.0, 1.0, -2048.0, 1.0] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 1], "tolerance": 0, "relTolerance": 0 } }
    },
    {
      "name": "all_axes_flat_packed_f32_nine_partials_8196",
      "attrs": { "axes": [-1], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8196],
          "data": { "kind": "cycle", "values": [0.25, -0.125, 0.5, -0.0625] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1], "tolerance": 0, "relTolerance": 0 } }
    },
    {
      "name": "all_axes_flat_f16_three_element_tail_8195",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "float16", "shape": [8195], "data": { "kind": "cycle", "values": [0.25, -0.125, 0.5, -0.625] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [], "tolerance": 0, "relTolerance": 0 } }
    },
    {
      "name": "all_axes_flat_rank1_boundary_8192",
      "provenance": {
        "notes": "Exactly 8,192 rank-1 elements exercise the inclusive lower boundary of the parallel full reduction."
      },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [8192], "data": { "kind": "constant", "value": 1.0 } } },
      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0 } }
    },
    {
      "name": "all_axes_flat_fullreduce_32x32x32_keepdims",
      "attrs": { "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [32, 32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.5 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1], "tolerance": 0.01, "relTolerance": 0.0001 } }
    },
    {
      "name": "dispatch_cliff_rank3_axis1",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [4096, 1, 4097], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4096, 4097], "tolerance": 0.0001 } }
    },
    {
      "name": "axis0",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 4],
          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 0.000001 } }
    },
    {
      "name": "axis0_splitk_8192x32",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8192, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.0001 } }
    },
    {
      "name": "axis0_splitk_8192x48_keepdims",
      "attrs": { "axes": [0], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8192, 48],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 48], "tolerance": 0.0001 } }
    },
    {
      "name": "axis0_tiled_64x32",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [64, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.00001 } }
    },
    {
      "name": "f32_subnormal_axis0_tilecols_sum_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "An axis-0 reduction over 64 finite subnormal values produces a larger finite subnormal in each column."
      },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [64, 16], "data": { "kind": "constant", "value": 1e-40 } } },
      "outputs": {
        "y": { "dtype": "float32", "shape": [16], "tolerance": 5e-44, "data": { "kind": "constant", "value": 6.4e-39 } }
      }
    },
    {
      "name": "axis1",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 4],
          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
    },
    {
      "name": "f32_axis1_parallel_cancellation_order_gpu_gap",
      "skipGpu": {
        "category": "todo",
        "reason": "The current parallel reduction changes the fixture's required sequential evaluation order, so f32 rounding is not bit-exact. An order-preserving reduction route can implement this behavior."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "Serial float32 reduction of each [1e20, 1, -1e20, 0] block loses the 1 before cancellation; the parallel tree groups lanes differently and can leak the small terms."
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [1, 1024],
          "data": { "kind": "cycle", "values": [100000000000000000000.0, 1.0, -100000000000000000000.0, 0.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1], "tolerance": 0 } }
    },
    {
      "name": "f32_subnormal_axis1_sum_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "Summing finite subnormal values can still produce finite subnormal outputs; flushing the reduction input or accumulator loses the signal."
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3],
          "data": { "kind": "values", "values": [1e-40, 1e-40, 1e-40, -1e-40, -1e-40, -1e-40] }
        }
      },
      "outputs": {
        "y": {
          "dtype": "float32",
          "shape": [2],
          "tolerance": 2e-45,
          "data": { "kind": "values", "values": [3e-40, -3e-40] }
        }
      }
    },
    {
      "name": "f32_subnormal_axis0_sum_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "An axis-0 reduction of finite subnormal columns must produce finite subnormal sums rather than zero."
      },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2],
          "data": { "kind": "values", "values": [1e-40, -1e-40, 1e-40, -1e-40, 1e-40, -1e-40] }
        }
      },
      "outputs": {
        "y": {
          "dtype": "float32",
          "shape": [2],
          "tolerance": 2e-45,
          "data": { "kind": "values", "values": [3e-40, -3e-40] }
        }
      }
    },
    {
      "name": "f32_many_subnormals_axis1_sum_to_normal_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "Many finite subnormal addends can reduce to a normal finite value; flushing inputs before accumulation loses a model-relevant low-magnitude signal."
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 1024], "data": { "kind": "constant", "value": 1e-39 } } },
      "outputs": { "y": { "dtype": "float32", "shape": [1], "tolerance": 1e-42 } }
    },
    {
      "name": "f32_subnormal_last_axis_vec4_sum_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "Finite subnormal rows should survive the vectorized last-axis reduction."
      },
      "attrs": { "axes": [-1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 4],
          "data": { "kind": "values", "values": [1e-40, 1e-40, 1e-40, 1e-40, -1e-40, -1e-40, -1e-40, -1e-40] }
        }
      },
      "outputs": {
        "y": {
          "dtype": "float32",
          "shape": [2],
          "tolerance": 3e-45,
          "data": { "kind": "values", "values": [4e-40, -4e-40] }
        }
      }
    },
    {
      "name": "f32_subnormal_last_axis_odd_sum_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "Finite subnormal rows should survive an odd-width last-axis reduction."
      },
      "attrs": { "axes": [-1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3],
          "data": { "kind": "values", "values": [1e-40, 1e-40, 1e-40, -1e-40, -1e-40, -1e-40] }
        }
      },
      "outputs": {
        "y": {
          "dtype": "float32",
          "shape": [2],
          "tolerance": 2e-45,
          "data": { "kind": "values", "values": [3e-40, -3e-40] }
        }
      }
    },
    {
      "name": "f32_subnormal_rank3_axis1_sum_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "A rank-3 axis-1 reduction must preserve finite subnormal column sums through middle-axis indexing."
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3, 2],
          "data": {
            "kind": "values",
            "values": [1e-40, -1e-40, 1e-40, -1e-40, 1e-40, -1e-40, -1e-40, 1e-40, -1e-40, 1e-40, -1e-40, 1e-40]
          }
        }
      },
      "outputs": {
        "y": {
          "dtype": "float32",
          "shape": [2, 2],
          "tolerance": 2e-45,
          "data": { "kind": "values", "values": [3e-40, -3e-40, -3e-40, 3e-40] }
        }
      }
    },
    {
      "name": "f32_subnormal_rank3_all_axes_sum_scalar_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_default_axes_do_not_keep_dims",
        "notes": "A rank-3 default-axes reduction should preserve the finite subnormal scalar total."
      },
      "attrs": { "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 3, 2], "data": { "kind": "constant", "value": 1e-40 } } },
      "outputs": {
        "y": { "dtype": "float32", "shape": [], "tolerance": 1e-44, "data": { "kind": "values", "values": [1.2e-39] } }
      }
    },
    {
      "name": "f32_subnormal_rank3_all_axes_keepdims_sum_gpu_gap",
      "skipGpu": {
        "category": "permanent",
        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero in f32; bit-exact subnormal preservation is unattainable on GPU."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_default_axes_keepdims",
        "notes": "A rank-3 default-axes reduction with keepdims should preserve the finite subnormal total in shape [1,1,1]."
      },
      "attrs": { "keepdims": 1 },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 3, 2], "data": { "kind": "constant", "value": 1e-40 } } },
      "outputs": {
        "y": {
          "dtype": "float32",
          "shape": [1, 1, 1],
          "tolerance": 1e-44,
          "data": { "kind": "values", "values": [1.2e-39] }
        }
      }
    },
    {
      "name": "axis1_empty_cols_identity_zero",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 0], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0 } }
    },
    {
      "name": "axis0_empty_rows_identity_zero",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [0, 3], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0 } }
    },
    {
      "name": "axis1_zero_rows_noop",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [0, 3], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [0], "tolerance": 0 } }
    },
    {
      "name": "axis1_keepdims",
      "attrs": { "axes": [1], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 4],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, -1.0, -2.0, -3.0, -4.0, 0.5, 1.5, 2.5, 3.5] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 1], "tolerance": 0.000001 } }
    },
    {
      "name": "rank3_axis1_middle_no_keepdims",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3, 4],
          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis1_middle_no_keepdims_input_x" } }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 4], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_empty_rank3_middle_axis_keepdims",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.empty_set_ReduceSum_13"
      },
      "attrs": { "axes": [1], "keepdims": 1 },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 0, 4], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 1, 4], "tolerance": 0 } }
    },
    {
      "name": "rank3_two_axes_trailing_empty_axis_identity_zero",
      "provenance": {
        "notes": "Two axes are reduced and the trailing reduced axis is empty. The serial index walk must use its zero-stride safeguard and return the additive identity without dividing by zero."
      },
      "attrs": { "axes": [1, 2], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 3, 0], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0 } }
    },
    {
      "name": "rank4_axis1_channel_no_keepdims",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3, 2, 2],
          "data": {
            "kind": "values",
            "values": [1.0, -2.0, 3.0, -4.0, 10.0, 20.0, -30.0, -40.0, 0.25, -0.5, 0.75, -1.0, -5.0, 6.0, -7.0, 8.0, 0.0, 0.0, 1.5, -1.5, 100.0, -200.0, 300.0, -400.0]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 2, 2], "tolerance": 0.000001 } }
    },
    {
      "name": "rank1_axis0_scalar_output",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [5], "data": { "kind": "values", "values": [1.0, -2.0, 3.5, 0.25, -4.75] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_axis1_small_no_keepdims",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_do_not_keepdims"
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [1, 2, 2], "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 2], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_axis0_rank1_scalar",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_do_not_keepdims_2"
      },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1.0, 2.0, 3.0] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_rank0_scalar",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum0DTensor"
      },
      "inputs": { "x": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [2.0] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0 } }
    },
    {
      "name": "ort_axis1_rank3_keepdims",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_keepdims"
      },
      "attrs": { "axes": [1], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_axis2_rank3_no_keepdims",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_axes01"
      },
      "attrs": { "axes": [2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 2], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_axis1_rank3_no_keepdims",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_axes02"
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 2], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_mixed_infinities_axis1_nan_rows",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceInfSum"
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [6, 2],
          "data": {
            "kind": "values",
            "values": [1.0, "Infinity", "Infinity", 4.0, "Infinity", "-Infinity", "-Infinity", "Infinity", 1.0, "-Infinity", "-Infinity", 4.0]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [6], "tolerance": 0, "allowNaN": true } }
    },
    {
      "name": "onnx_backend_reduce_sum_do_not_keepdims_example",
      "attrs": { "keepdims": 0, "axes": [1] },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 2] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_do_not_keepdims_example",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "onnx_backend_reduce_sum_do_not_keepdims_random",
      "attrs": { "keepdims": 0, "axes": [1] },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": {
            "kind": "values",
            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 2] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_do_not_keepdims_random",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "onnx_backend_reduce_sum_empty_set",
      "attrs": { "keepdims": 1, "axes": [1] },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 0, 4], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 1, 4] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_empty_set",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "onnx_backend_reduce_sum_empty_set_non_reduced_axis_zero",
      "attrs": { "keepdims": 1, "axes": [2] },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 0, 4], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 0, 1] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_empty_set_non_reduced_axis_zero",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "onnx_backend_reduce_sum_keepdims_example",
      "attrs": { "keepdims": 1, "axes": [1] },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_keepdims_example",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "onnx_backend_reduce_sum_keepdims_random",
      "attrs": { "keepdims": 1, "axes": [1] },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": {
            "kind": "values",
            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_keepdims_random",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "onnx_backend_reduce_sum_negative_axes_keepdims_example",
      "attrs": { "keepdims": 1, "axes": [-2] },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_negative_axes_keepdims_example",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "onnx_backend_reduce_sum_negative_axes_keepdims_random",
      "attrs": { "keepdims": 1, "axes": [-2] },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": {
            "kind": "values",
            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_negative_axes_keepdims_random",
        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
      }
    },
    {
      "name": "ort_default_axes_rank3_no_keepdims_scalar",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_default_axes_do_not_keep_dims",
        "notes": "Default-axes reduction of every element to a rank-0 scalar with keepdims=0."
      },
      "attrs": { "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": {
        "y": { "dtype": "float32", "shape": [], "tolerance": 0, "data": { "kind": "values", "values": [78.0] } }
      }
    },
    {
      "name": "onnx_backend_reduce_sum_default_axes_keepdims_example",
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_default_axes_keepdims_example"
      },
      "attrs": { "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1] } }
    },
    {
      "name": "ort_default_axes_keepdims_all_rank3",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_default_axes_keepdims"
      },
      "attrs": { "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1], "tolerance": 0.000001 } }
    },
    {
      "name": "onnx_backend_reduce_sum_default_axes_keepdims_random",
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_default_axes_keepdims_random"
      },
      "attrs": { "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": {
            "kind": "values",
            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1] } }
    },
    {
      "name": "ort_empty_default_axes_keepdims_all_rank3",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_EmptySet_DefaultAxes_KeepDims",
        "notes": "ORT uses an empty axes input to request reduction over all axes; this fixture represents that with omitted axis."
      },
      "attrs": { "keepdims": 1 },
      "inputs": { "x": { "dtype": "float32", "shape": [3, 0, 2], "data": { "kind": "values", "values": [] } } },
      "outputs": {
        "y": { "dtype": "float32", "shape": [1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } }
      }
    },
    {
      "name": "onnx_backend_reduce_sum_empty_axes_input_noop_example",
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_empty_axes_input_noop_example"
      },
      "attrs": { "keepdims": 1, "noop_with_empty_axes": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 2, 2] } }
    },
    {
      "name": "ort_missing_axes_noop_identity_keepdims0",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_missing_axes_input_noop_opset_13",
        "notes": "Missing optional axes plus noop_with_empty_axes=1 is represented as an empty axes input by the ORT validator."
      },
      "attrs": { "keepdims": 0, "noop_with_empty_axes": 1 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [1, 2, 2], "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 2], "tolerance": 0 } }
    },
    {
      "name": "onnx_backend_reduce_sum_empty_axes_input_noop_random",
      "provenance": {
        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_sum_empty_axes_input_noop"
      },
      "attrs": { "keepdims": 1, "noop_with_empty_axes": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": {
            "kind": "values",
            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 2, 2] } }
    },
    {
      "name": "subgroup_vec4_last_axis_2x256",
      "attrs": { "axes": [-1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 256],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0.0002, "relTolerance": 0.0001 } }
    },
    {
      "name": "subgroup_vec4_last_axis_2x1024",
      "attrs": { "axes": [-1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.25 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0.001, "relTolerance": 0.0001 } }
    },
    {
      "name": "subgroup_scalar_last_axis_2x65",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 65],
          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0.0002, "relTolerance": 0.0001 } }
    },
    {
      "name": "ort_int32_multi_axis_keepdims",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_int32"
      },
      "attrs": { "axes": [0, 2], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12] }
        }
      },
      "outputs": {
        "y": { "dtype": "int32", "shape": [1, 2, 1], "data": { "kind": "values", "values": [33, 45] }, "tolerance": 0 }
      }
    },
    {
      "name": "ort_int32_positive_overflow_saturates_gpu_gap",
      "skipGpu": {
        "category": "todo",
        "reason": "The current integer reduction route uses an i32 accumulator, so it cannot reproduce the fixture's widened intermediate arithmetic and final int32 saturation. A portable multiword accumulator can implement this behavior."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_int32_Overflow_Saturates"
      },
      "attrs": { "axes": [0], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [3],
          "data": { "kind": "values", "values": [1100000000, 1100000000, 1100000000] }
        }
      },
      "outputs": {
        "y": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2147483647] }, "tolerance": 0 }
      }
    },
    {
      "name": "ort_int32_negative_overflow_saturates_gpu_gap",
      "skipGpu": {
        "category": "todo",
        "reason": "The current integer reduction route uses an i32 accumulator, so it cannot reproduce the fixture's widened intermediate arithmetic and final int32 saturation. A portable multiword accumulator can implement this behavior."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
        "test": "ReductionOpTest.ReduceSum_int32_NegativeOverflow_Saturates"
      },
      "attrs": { "axes": [0], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [3],
          "data": { "kind": "values", "values": [-1100000000, -1100000000, -1100000000] }
        }
      },
      "outputs": {
        "y": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [-2147483648] }, "tolerance": 0 }
      }
    },
    {
      "name": "ort_float_multi_axis_keepdims",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/reduction/reduction_test_cases_generator.py",
        "test": "ReductionOpTest.ReduceSum",
        "notes": "Generated canonical reduction-table case materialized in reduction_ops_test.cc."
      },
      "attrs": { "axes": [0, 2], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 1], "tolerance": 0.000001 } }
    },
    {
      "name": "rank3_lastaxis_cols1024_tree_nosubgroup",
      "attrs": { "axes": [2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 2, 1024],
          "data": {
            "kind": "cycle",
            "values": [1.0, -2.0, 0.5, 3.25, -1.5, 2.0, -0.75, 4.0, -3.5, 1.25, 0.0, -2.25, 5.0, -4.0, 2.75, -1.0, 6.5]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 2], "tolerance": 0.00001 } }
    },
    {
      "name": "dispatch_cliff_axis0_cols_1048577",
      "requires": { "limits": { "maxBufferSize": 268435712, "maxStorageBufferBindingSize": 268435712 } },
      "provenance": {
        "notes": "With 1,048,577 columns, the axis-0 reduction requires 65,536 column tiles and therefore a two-row workgroup dispatch. The final tile verifies folded workgroup indexing and the over-dispatch bound. The 268,435,712-byte tensor requires the declared storage limits because it exceeds WebGPU's guaranteed minimum storage-binding size."
      },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [64, 1048577],
          "data": { "kind": "cycle", "values": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1048577], "tolerance": 0.0001 } }
    },
    {
      "name": "rank3_axis0_no_keepdims",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3, 4],
          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis0_no_keepdims_input_x" } }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 4], "tolerance": 0.000001 } }
    },
    {
      "name": "rank4_axis2_no_keepdims",
      "attrs": { "axes": [2], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [2, 3, 4, 5], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 5], "tolerance": 0.00001 } }
    },
    {
      "name": "rank4_axis0_no_keepdims",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "float32", "shape": [2, 3, 4, 5], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 4, 5], "tolerance": 0.00001 } }
    },
    {
      "name": "rank3_multi_axis02_no_keepdims",
      "attrs": { "axes": [0, 2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3, 4],
          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis0_no_keepdims_input_x" } }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
    },
    {
      "name": "rank4_multi_axis_12_keepdims",
      "attrs": { "axes": [1, 2], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3, 2, 2],
          "data": {
            "kind": "values",
            "values": [0.5, -1.0, 2.0, -0.25, 1.5, 0.75, -2.0, 1.0, 0.125, -0.5, 3.0, -1.5, 0.25, 2.5, -0.75, 1.25, -3.0, 0.5, 2.0, -1.0, 0.75, -0.25, 1.5, -2.5]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 1, 1, 2], "tolerance": 0.0001 } }
    },
    {
      "name": "int32_noop_with_empty_axes_identity_r3",
      "attrs": { "keepdims": 1, "noop_with_empty_axes": 1 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12] }
        }
      },
      "outputs": {
        "y": {
          "dtype": "int32",
          "shape": [3, 2, 2],
          "tolerance": 0,
          "data": { "kind": "values", "values": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12] }
        }
      }
    },
    {
      "name": "int32_lastaxis_serial_exact_signed_3x5",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [3, 5],
          "data": { "kind": "values", "values": [1, -2, 3, -4, 5, -6, 7, -8, 9, -10, 11, -12, 13, -14, 15] }
        }
      },
      "outputs": {
        "y": { "dtype": "int32", "shape": [3], "tolerance": 0, "data": { "kind": "values", "values": [3, -8, 13] } }
      }
    },
    {
      "name": "int32_lastaxis_tree_parallel_3x67",
      "provenance": {
        "notes": "A 67-column int32 row exercises parallel last-axis reduction. Values above 2^24 make any unintended f32 accumulation observable."
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [3, 67],
          "data": { "kind": "cycle", "values": [16777216, 1, -3, 7, 16777219, -11] }
        }
      },
      "outputs": { "y": { "dtype": "int32", "shape": [3], "tolerance": 0 } }
    },
    {
      "name": "int32_axis0_empty_rows_identity_zero",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "dtype": "int32", "shape": [0, 4], "data": { "kind": "values", "values": [] } } },
      "outputs": {
        "y": { "dtype": "int32", "shape": [4], "tolerance": 0, "data": { "kind": "values", "values": [0, 0, 0, 0] } }
      }
    },
    {
      "name": "int32_fullreduce_r3_single_lane_exact_4x3x2",
      "attrs": { "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [4, 3, 2],
          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis0_no_keepdims_input_x" } }
        }
      },
      "outputs": {
        "y": { "dtype": "int32", "shape": [], "tolerance": 0, "data": { "kind": "values", "values": [276] } }
      }
    },
    {
      "name": "all_axes_flat_fullreduce_101x103_nonmul4_keepdims",
      "provenance": {
        "notes": "A 10,403-element full reduction uses float32 vec4 groups plus a three-element scalar tail. The kept singleton dimensions and independently computed result make any dropped tail element observable."
      },
      "attrs": { "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [101, 103],
          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.5 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1], "tolerance": 0.01, "relTolerance": 0.0001 } }
    },
    {
      "name": "int32_axis0_splitk_8192x2",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "dtype": "int32", "shape": [8192, 2], "data": { "kind": "cycle", "values": [1, -1, 2, -2] } } },
      "outputs": { "y": { "dtype": "int32", "shape": [2], "tolerance": 0 } }
    },
    {
      "name": "int32_all_axes_flat_128x64",
      "attrs": { "keepdims": 0 },
      "inputs": { "x": { "dtype": "int32", "shape": [128, 64], "data": { "kind": "cycle", "values": [7000, -1, 2] } } },
      "outputs": { "y": { "dtype": "int32", "shape": [], "tolerance": 0 } },
      "provenance": {
        "notes": "A three-value int32 cycle over 8,192 elements sums to the odd value 19,119,729. The result is inside int32 range but not exactly representable as f32, exposing any unintended float accumulation."
      }
    },
    {
      "name": "axis0_narrow_f32_8192x3_splitk",
      "provenance": {
        "notes": "An 8,192-by-3 axis-0 reduction exercises split-K with a narrow output. Constant ones make the partial reduction and reassociation exact."
      },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": { "x": { "dtype": "float32", "shape": [8192, 3], "data": { "kind": "constant", "value": 1.0 } } },
      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0 } }
    },
    {
      "name": "contiguous_suffix_axes12_parallel",
      "provenance": {
        "notes": "Contiguous axes {1,2} exercise the shared cooperative suffix reduction instead of one serial lane per output."
      },
      "attrs": { "axes": [1, 2], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 16, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 1], "tolerance": 0.00001 } }
    },
    {
      "name": "noop_empty_axes_identity_rank2",
      "attrs": { "keepdims": 1, "noop_with_empty_axes": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 4],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 4] } }
    },
    {
      "name": "noop_empty_axes_identity_rank4",
      "attrs": { "keepdims": 1, "noop_with_empty_axes": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 3, 2, 2],
          "data": {
            "kind": "values",
            "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 2, 2] } }
    },
    {
      "name": "axis_split_rank3_axis1_2x8192x4",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 8192, 4],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 4], "tolerance": 0.0001 } }
    },
    {
      "name": "f16_axis_split_tiled_narrow_2x8192x4",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 8192, 4],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2, 4], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "axis_split_rank3_axis1_wide_2x8192x32",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 8192, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 32], "tolerance": 0.0001 } }
    },
    {
      "name": "f16_axis_split_wide_2x8192x32",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 8192, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2, 32], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "axis_split_rank3_axis1_keepdims_1x16384x8",
      "attrs": { "axes": [1], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [1, 16384, 8],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 8], "tolerance": 0.0001 } }
    },
    {
      "name": "f16_rank3_axis1_serial",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 3, 4],
          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis1_middle_no_keepdims_input_x" } }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2, 4], "tolerance": 0.05 } }
    },
    {
      "name": "f16_last_axis_2x65_serial_fallback",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 65],
          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2], "tolerance": 0.05, "relTolerance": 0.0001 } }
    },
    {
      "name": "f16_rank3_multi_axis02",
      "attrs": { "axes": [0, 2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 3, 4],
          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis0_no_keepdims_input_x" } }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [3], "tolerance": 0.05 } }
    },
    {
      "name": "f16_all_axes_keepdims",
      "attrs": { "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [3, 2, 2],
          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 1], "tolerance": 0.05 } }
    },
    {
      "name": "f16_axis0_splitk_8192x8",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [8192, 8],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 0.25 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.05, "relTolerance": 0.002 } },
      "provenance": {
        "notes": "A 8192-row float16 axis-0 sum. Values oscillate about 0.25 so each column total is proportional to the reduction length, making a dropped, duplicated or mis-strided partial observable at float16 resolution."
      }
    },
    {
      "name": "f16_last_axis_vec4_8x1024",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [8, 1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "f16_last_axis_scalar_8x1023",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [8, 1023],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "f16_all_axes_flat_65543",
      "attrs": { "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [65543],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "f16_suffix_vec4_4x8x128",
      "attrs": { "axes": [1, 2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [4, 8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "f16_suffix_scalar_4x7x37",
      "attrs": { "axes": [1, 2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [4, 7, 37],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "f16_axis0_tilecols_4096x64",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [4096, 64],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 0.25 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [64], "tolerance": 0.05, "relTolerance": 0.002 } },
      "provenance": {
        "notes": "A 4096-row float16 axis-0 sum. Values oscillate about 0.25 so each column total is proportional to the reduction length, making a dropped, duplicated or mis-strided partial observable at float16 resolution."
      }
    },
    {
      "name": "int32_axis0_tiled_64x32",
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [64, 32],
          "data": { "kind": "cycle", "values": [16777217, 3, -5, 16777219, 7, -11, 2] }
        }
      },
      "outputs": { "y": { "dtype": "int32", "shape": [32], "tolerance": 0 } }
    },
    {
      "name": "subgroup_rows_last_axis_f32_96x256",
      "attrs": { "axes": [-1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [96, 256],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [96], "tolerance": 0.0002, "relTolerance": 0.0001 } }
    },
    {
      "name": "subgroup_rows_last_axis_f32_rank3_2x40x1024",
      "attrs": { "axes": [-1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 40, 1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.25 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 40], "tolerance": 0.001, "relTolerance": 0.0001 } }
    },
    {
      "name": "subgroup_rows_last_axis_f16_80x1024",
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [80, 1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [80], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "int32_lastaxis_subgroup_rows_64x1024",
      "provenance": {
        "notes": "Sixty-four rows of 1,024 int32 values exercise subgroup-per-row reduction in the output type. A five-value cycle shifts phase every row so outputs differ while matching tree accumulation exactly."
      },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": { "dtype": "int32", "shape": [64, 1024], "data": { "kind": "cycle", "values": [3, 4, 0, -2, 1] } }
      },
      "outputs": { "y": { "dtype": "int32", "shape": [64], "tolerance": 0 } }
    },
    {
      "name": "f32_lastaxis_rowserial_band_subgroup_rows_8192x256",
      "provenance": { "notes": "An 8192-by-256 float32 input checks independent last-axis sums over many rows." },
      "attrs": { "axes": [1], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8192, 256],
          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.029, "scale": 1.0 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8192], "tolerance": 0.001, "relTolerance": 0.0001 } }
    },
    {
      "name": "rank4_axes023_noncontiguous_multi_axis_2x4x3x3",
      "provenance": {
        "notes": "A rank-4 reduction over axes {0,2,3} keeps the middle channel axis. The reduced elements are not a contiguous suffix, so each output uses a serial fold."
      },
      "attrs": { "axes": [0, 2, 3], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 4, 3, 3],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 0.0001 } }
    },
    {
      "name": "multi_axis_rank4_coop_channel_reduce_axes023",
      "provenance": {
        "notes": "A rank-4 reduction over axes {0,2,3} leaves one output per channel and 512 reduced elements per output, exercising cooperative tree accumulation. The tolerance allows its f32 reassociation."
      },
      "attrs": { "axes": [0, 2, 3], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 8, 16, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.029, "scale": 1.5 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8], "tolerance": 0.0005, "relTolerance": 0.0001 } }
    },
    {
      "name": "multi_axis_rank4_coop_channel_reduce_axes023_keepdims",
      "provenance": {
        "notes": "A rank-4 reduction over axes {0,2,3} with keepdims leaves one output per channel and 512 reduced elements per output, exercising cooperative tree accumulation. The tolerance allows its f32 reassociation."
      },
      "attrs": { "axes": [0, 2, 3], "keepdims": 1 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [2, 8, 16, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.029, "scale": 1.5 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 1, 1], "tolerance": 0.0005, "relTolerance": 0.0001 } }
    },
    {
      "name": "multi_axis_rank3_coop_axes02",
      "provenance": {
        "notes": "A rank-3 reduction over axes {0,2} leaves one output per channel and 512 reduced elements per output, exercising cooperative tree accumulation. The tolerance allows its f32 reassociation."
      },
      "attrs": { "axes": [0, 2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8, 6, 64],
          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.029, "scale": 1.5 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [6], "tolerance": 0.0005, "relTolerance": 0.0001 } }
    },
    {
      "name": "multi_axis_rank4_coop_channel_reduce_axes023_f16",
      "provenance": {
        "notes": "A cooperative multi-axis channel reduction widens f16 storage to f32 for accumulation."
      },
      "attrs": { "axes": [0, 2, 3], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 8, 16, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.029, "scale": 1.0 }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.05, "relTolerance": 0.002 } }
    },
    {
      "name": "multi_axis_rank4_coop_channel_reduce_axes023_i32",
      "provenance": {
        "notes": "A cooperative multi-axis channel reduction accumulates int32 values in the output type."
      },
      "attrs": { "axes": [0, 2, 3], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "int32",
          "shape": [2, 8, 16, 16],
          "data": { "kind": "cycle", "values": [3, -1, 4, -1, 5, -9, 2] }
        }
      },
      "outputs": { "y": { "dtype": "int32", "shape": [8], "tolerance": 0 } }
    },
    {
      "name": "coop_and_serial_one_large_many_small",
      "provenance": {
        "notes": "Each channel contains one 1e8 value and 2,047 ones. Cooperative reassociation at this reduction size must remain within the declared relative tolerance of the sequential sum."
      },
      "attrs": { "axes": [0, 2, 3], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8, 64, 16, 16],
          "data": { "kind": "cycle", "values": [100000000.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [64], "tolerance": 1, "relTolerance": 0.00001 } }
    },
    {
      "name": "packed_columns_f32_axis0_compact",
      "provenance": {
        "notes": "A 32-column, 8,192-row exact sum checks every column over a long reduced axis on different subgroup widths."
      },
      "tunables": { "WORKGROUP_SIZE": 16 },
      "attrs": { "axes": [0], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8192, 32],
          "data": { "kind": "cycle", "values": [-0.25, 0.125, 0.5, -0.0625] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0, "relTolerance": 0 } }
    },
    {
      "name": "packed_columns_f32_axis0_tail_negative_axis",
      "provenance": {
        "notes": "Reducing 8193 rows (axis -2, i.e. axis 0) down to 128 output columns, twice 64; binary-fraction values (powers of two) keep the sum exact across any reduction order, including the row past 8192."
      },
      "tunables": { "WORKGROUP_SIZE": 64 },
      "attrs": { "axes": [-2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8193, 128],
          "data": { "kind": "cycle", "values": [-0.25, 0.125, 0.5, -0.0625] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [128], "tolerance": 0, "relTolerance": 0 } }
    },
    {
      "name": "packed_columns_f16_middle_axis",
      "provenance": {
        "notes": "Reducing the middle axis (size 8192, axis -2) of a float16 tensor down to 64 output columns per batch; binary-fraction values (powers of two) keep the sum exact across any reduction order."
      },
      "tunables": { "WORKGROUP_SIZE": 64 },
      "attrs": { "axes": [-2], "keepdims": 0 },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [2, 8192, 64],
          "data": { "kind": "cycle", "values": [-0.25, 0.125, 0.5, -0.0625] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [2, 64], "tolerance": 0, "relTolerance": 0 } }
    }
  ]
}