ai.onnx.ArgMin / build /webgpu /test.json
Xenova's picture
Xenova HF Staff
sync 2e7068faf55e
7c6615c verified
Raw
History Blame
51.2 kB
{
"op": "ai.onnx.ArgMin",
"fixtureArrays": {
"onnx_backend_argmin_input_x": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564, 1.3608912229537964, 8.511932373046875, -8.579278945922852, -8.257413864135742, -9.59563159942627, 6.6523966789245605, 5.563135147094727, 7.400242805480957, 9.572366714477539, 5.983171463012695, -0.7704127430915833, 5.610583305358887],
"axis0_splitk_8192x32_ties_first_index_input_x": [2, -3, -3, 1, 2, -3, -3, 1, 2, -3, -3, 1, 2, -3, -3, 1, 2, -3, -3, 1, 2, -3, -3, 1, 2, -3, -3, 1, 2, -3, -3, 1, 2],
"subgroup_vec4_ties_first_index_input_x": [1, 5, 5, 2, 5, 0, 1, 0, 7, 7, 7, 7, 7, 7, 7, 7],
"axis0_splitk_i32_8192x16_precision_min_input_x": [16777217, 16777217, 16777217, 16777217, 16777217, 16777217, 16777217, 16777217, 16777216, 16777216, 16777216, 16777216, 16777216, 16777216, 16777216, 16777216, 16777216]
},
"cases": [
{
"name": "int16_last_axis_vec4_boundaries",
"attrs": { "axis": -1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "int16",
"shape": [2, 4],
"data": { "kind": "values", "values": [-32768, -1, 32767, 0, 32767, -32768, 32767, 1] }
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [2], "tolerance": 0, "data": { "kind": "values", "values": [0, 1] } }
}
},
{
"name": "onnx_default_axis0_omitted",
"provenance": {
"source": "https://onnx.ai/onnx/operators/onnx__ArgMin.html",
"test": "ArgMin axis default",
"notes": "Omitting axis must reduce the leading dimension (ONNX default 0), not axis 1. The rank-2 output shape and per-column indices distinguish the two semantics. ONNX int64 indices use the framework's representable uint32 slot."
},
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3],
"data": { "kind": "values", "values": [1.0, 9.0, 3.0, 4.0, 2.0, 8.0] }
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [1, 3], "tolerance": 0, "data": { "kind": "values", "values": [0, 1, 0] } }
}
},
{
"name": "f32_negative_subnormal_beats_zero_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": "ArgMin",
"notes": "A negative subnormal is strictly less than zero; index selection must not treat it as a tie with zero."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": { "dtype": "float32", "shape": [1, 3], "data": { "kind": "values", "values": [1e-40, 0.0, -1e-40] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [1], "tolerance": 0 } }
},
{
"name": "f32_negative_subnormal_beats_zero_select_last_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": "ArgMin_do_not_keepdims_2_select_last",
"notes": "select_last_index must only apply to true ties; a negative subnormal at index 0 is less than following zeros."
},
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": { "dtype": "float32", "shape": [1, 3], "data": { "kind": "values", "values": [-1e-40, 0.0, 0.0] } }
},
"outputs": {
"y": { "dtype": "uint32", "shape": [1], "tolerance": 0, "data": { "kind": "values", "values": [0] } }
}
},
{
"name": "f32_negative_subnormal_beats_zero_axis0_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": "ArgMin",
"notes": "Axis-0 companion for finite subnormal ordering: a negative subnormal is strictly less than zero and must not be tie-broken as flushed zero."
},
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": { "dtype": "float32", "shape": [3, 1], "data": { "kind": "values", "values": [1e-40, 0.0, -1e-40] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [1], "tolerance": 0 } }
},
{
"name": "dispatch_cliff_axis1_16777216x1",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": { "dtype": "float32", "shape": [16777216, 1], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [16777216], "tolerance": 0 } }
},
{
"name": "axis0",
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [3, 4],
"data": { "kind": "values", "values": [1.0, 9.0, 3.0, 4.0, -1.0, 2.0, 7.0, 8.0, 0.0, 5.0, -3.0, 6.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [4] } }
},
{
"name": "axis0_splitk_8192x32_ties_first_index",
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [8192, 32],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_8192x32_ties_first_index_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [32], "tolerance": 0 } }
},
{
"name": "axis0_splitk_8192x32_ties_select_last_index",
"attrs": { "axis": 0, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [8192, 32],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_8192x32_ties_first_index_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [32], "tolerance": 0 } }
},
{
"name": "axis0_splitk_8192x48_keepdims",
"attrs": { "axis": 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": "uint32", "shape": [1, 48], "tolerance": 0 } }
},
{
"name": "axis0_tiled_64x32_ties_first_index",
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [64, 32],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_8192x32_ties_first_index_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [32], "tolerance": 0 } }
},
{
"name": "axis0_tiled_64x32_ties_select_last_index",
"attrs": { "axis": 0, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [64, 32],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_8192x32_ties_first_index_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [32], "tolerance": 0 } }
},
{
"name": "axis1_tie_first_index",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 5],
"data": { "kind": "values", "values": [1.0, -3.0, -3.0, 2.0, -3.0, 1.0, 1.0, 2.0, 1.0, 3.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2] } }
},
{
"name": "ort_axis1_nan_first_incumbent_gpu_gap",
"skipGpu": {
"category": "todo",
"reason": "The parallel min/max reduction routes do not yet preserve the reference's first-element NaN incumbent semantics; explicit NaN and index tracking is implementable in WGSL."
},
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin",
"notes": "NaN extension: ORT seeds ArgMin from the first reduced element, so a leading NaN remains the selected incumbent while later NaNs are ignored after a finite incumbent."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [4, 4],
"data": {
"kind": "values",
"values": ["NaN", 1.0, -2.0, 0.0, 1.0, "NaN", -2.0, 0.0, 1.0, -2.0, "NaN", 0.0, "NaN", "NaN", "NaN", "NaN"]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [4], "tolerance": 0 } }
},
{
"name": "ort_axis1_nan_select_last_index_gpu_gap",
"skipGpu": {
"category": "todo",
"reason": "The parallel min/max reduction routes do not yet preserve the reference's first-element NaN incumbent semantics; explicit NaN and index tracking is implementable in WGSL."
},
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_int32_last_index_dups",
"notes": "NaN extension: select_last_index only changes equal finite ties. A leading NaN remains the incumbent because finite values are not less than NaN."
},
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [4, 4],
"data": {
"kind": "values",
"values": ["NaN", 1.0, -2.0, 0.0, 1.0, "NaN", -3.0, -3.0, 1.0, 2.0, "NaN", 3.0, "NaN", "NaN", "NaN", "NaN"]
}
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [0, 3, 0, 0] }, "tolerance": 0 }
}
},
{
"name": "axis1_f16",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float16",
"shape": [3, 4],
"data": { "kind": "values", "values": [1.0, 9.0, 3.0, 4.0, -1.0, 2.0, 7.0, 8.0, 0.0, 5.0, -3.0, 6.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [3] } }
},
{
"name": "axis1_select_last_index_ties",
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 5],
"data": { "kind": "values", "values": [1.0, -3.0, -3.0, 2.0, -3.0, 1.0, 1.0, 2.0, 1.0, 3.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2] } }
},
{
"name": "axis1_select_last_all_positive_infinity",
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 4],
"data": {
"kind": "values",
"values": ["Infinity", "Infinity", "Infinity", "Infinity", "Infinity", 5.0, "Infinity", 5.0]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "tolerance": 0 } }
},
{
"name": "axis_minus_one",
"attrs": { "axis": -1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 4],
"data": { "kind": "values", "values": [1.0, -2.0, 9.0, 4.0, 5.0, 8.0, -7.0, 6.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2] } }
},
{
"name": "axis1_keepdims",
"attrs": { "axis": 1, "keepdims": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 4],
"data": { "kind": "values", "values": [4.0, -3.0, -3.0, 1.0, 2.0, 2.0, -5.0, -5.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1] } }
},
{
"name": "rank3_axis0",
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 2, 3],
"data": { "kind": "values", "values": [4.0, -3.0, 5.0, 1.0, 2.0, -5.0, 3.0, -4.0, 6.0, 0.0, 3.0, -6.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 3], "tolerance": 0 } }
},
{
"name": "rank4_axis1_channel_select_last_ties",
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [1, 3, 2, 2],
"data": { "kind": "values", "values": [1.0, -5.0, 3.0, -5.0, -2.0, -5.0, 3.0, -7.0, -2.0, -4.0, 0.0, -7.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 2, 2], "tolerance": 0 } }
},
{
"name": "rank4_axis2_spatial_keepdims_select_last_ties",
"attrs": { "axis": 2, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [1, 2, 3, 2],
"data": { "kind": "values", "values": [1.0, -4.0, -3.0, -4.0, -3.0, -2.0, 1.0, -5.0, 0.0, -5.0, 0.0, -3.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 2, 1, 2], "tolerance": 0 } }
},
{
"name": "rank4_last_axis_nhwc_keepdims_select_last",
"attrs": { "axis": -1, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [1, 2, 2, 3],
"data": { "kind": "values", "values": [1.0, -3.0, -3.0, 5.0, 4.0, -7.0, -3.0, -3.0, 0.0, 5.0, -7.0, -7.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 2, 2, 1], "tolerance": 0 } }
},
{
"name": "rank1_axis0_keepdims_select_last",
"attrs": { "axis": 0, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [6],
"data": { "kind": "values", "values": [4.0, -8.0, 2.0, -8.0, 0.0, -1.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1], "tolerance": 0 } }
},
{
"name": "ort_uint8_axis0_no_keepdims",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_uint8"
},
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "uint8",
"shape": [3, 2, 2],
"data": { "kind": "values", "values": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 2], "tolerance": 0 } }
},
{
"name": "ort_int8_axis0_no_keepdims",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_int8"
},
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "int8",
"shape": [3, 2, 2],
"data": { "kind": "values", "values": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 2], "tolerance": 0 } }
},
{
"name": "ort_int32_axis0_select_last_dups",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_int32_select_last"
},
"attrs": { "axis": 0, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [3, 2, 2],
"data": { "kind": "values", "values": [1, 2, 3, 4, 1, 6, 7, 8, 9, 10, 11, 12] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 2], "tolerance": 0 } }
},
{
"name": "ort_int32_axis0_no_keepdims",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_int32"
},
"attrs": { "axis": 0, "keepdims": 0 },
"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": "uint32", "shape": [2, 2], "tolerance": 0 } }
},
{
"name": "ort_axis0_keepdims_rank3_f32",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin"
},
"attrs": { "axis": 0, "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": "uint32", "shape": [1, 2, 2], "tolerance": 0 } }
},
{
"name": "ort_axis0_no_keepdims_rank3_f32",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_do_not_keepdims"
},
"attrs": { "axis": 0, "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": "uint32", "shape": [2, 2], "tolerance": 0 } }
},
{
"name": "ort_axis0_no_keepdims_rank1_scalar_f32",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_do_not_keepdims_2"
},
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1.0, 2.0, 3.0] } } },
"outputs": { "y": { "dtype": "uint32", "shape": [], "tolerance": 0 } }
},
{
"name": "ort_axis0_no_keepdims_rank1_scalar_select_last_f32",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_do_not_keepdims_2_select_last"
},
"attrs": { "axis": 0, "keepdims": 0, "select_last_index": 1 },
"inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1.0, 2.0, 3.0] } } },
"outputs": { "y": { "dtype": "uint32", "shape": [], "tolerance": 0 } }
},
{
"name": "ort_float_first_index_negative_infinity_deterministic",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_float_first_index_random",
"notes": "Deterministic compact projection of ORT's random negative-infinity first-index test."
},
"attrs": { "axis": 0, "keepdims": 1, "select_last_index": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [16],
"data": {
"kind": "values",
"values": [0.0, 1.0, "-Infinity", -5.0, "-Infinity", 3.0, -2.0, "-Infinity", -4.0, 1.0, -3.0, "-Infinity", 0.0, -2.0, -1.0, 0.0]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1], "tolerance": 0 } }
},
{
"name": "ort_negative_axis_int32_no_keepdims",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_int32_neg_axis"
},
"attrs": { "axis": -3, "keepdims": 0 },
"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": "uint32", "shape": [2, 2], "tolerance": 0 } }
},
{
"name": "int32_axis1_exact_above_float24",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [2, 4],
"data": { "kind": "values", "values": [16777217, 16777216, 5, 9, -16777216, -16777217, 0, 3] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_default_axis_example",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_default_axis_example",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "keepdims": 1, "axis": 0 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 1.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 2], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_default_axis_example_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_default_axis_example_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "keepdims": 1, "select_last_index": 1, "axis": 0 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 2.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 2], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_default_axis_random",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_default_axis_random",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "keepdims": 1, "axis": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 3, 4], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_default_axis_random_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_default_axis_random_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "keepdims": 1, "select_last_index": 1, "axis": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 3, 4], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_keepdims_example",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_keepdims_example",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 1 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 1.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_keepdims_example_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_keepdims_example_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 2.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_keepdims_random",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_keepdims_random",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1, 4], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_keepdims_random_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_keepdims_random_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1, 4], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_negative_axis_keepdims_example",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_negative_axis_keepdims_example",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": -1, "keepdims": 1 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 1.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_negative_axis_keepdims_example_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_negative_axis_keepdims_example_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": -1, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 2.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_negative_axis_keepdims_random",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_negative_axis_keepdims_random",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": -1, "keepdims": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 3, 1], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_negative_axis_keepdims_random_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_negative_axis_keepdims_random_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": -1, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 3, 1], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_no_keepdims_example",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_no_keepdims_example",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 1.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_no_keepdims_example_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_no_keepdims_example_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [2.0, 2.0, 3.0, 10.0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_no_keepdims_random",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_no_keepdims_random",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 4], "tolerance": 0 } }
},
{
"name": "onnx_backend_argmin_no_keepdims_random_select_last_index",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_no_keepdims_random_select_last_index",
"notes": "ONNX arg-reduction outputs are int64; this framework uses uint32 index storage."
},
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 3, 4],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_argmin_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 4], "tolerance": 0 } }
},
{
"name": "subgroup_vec4_ties_first_index",
"attrs": { "axis": -1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 8],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/subgroup_vec4_ties_first_index_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [5, 0] } } }
},
{
"name": "subgroup_vec4_ties_select_last_index",
"attrs": { "axis": -1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 8],
"data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/subgroup_vec4_ties_first_index_input_x" } }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [7, 7] } } }
},
{
"name": "subgroup_scalar_f16_2x5",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float16",
"shape": [2, 5],
"data": { "kind": "values", "values": [1.5, 2.5, 2.5, 0.5, 1.0, -1.0, -2.0, -0.5, -0.5, -3.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [3, 4] } } }
},
{
"name": "subgroup_vec4_uint8_full_range",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": { "dtype": "uint8", "shape": [2, 4], "data": { "kind": "values", "values": [255, 0, 254, 1, 2, 2, 1, 0] } }
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [1, 3] } } }
},
{
"name": "subgroup_vec4_long_row_2x256",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 256],
"data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29 }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2] } }
},
{
"name": "subgroup_min_vec4_rows128x256_ties_last",
"provenance": {
"notes": "Route-locks the one-minimum-subgroup vec4 path on variable-width adapters and verifies last-index tie ordering without a workgroup-memory combine."
},
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": { "x": { "dtype": "float32", "shape": [128, 256], "data": { "kind": "constant", "value": -7.0 } } },
"outputs": {
"y": { "dtype": "uint32", "shape": [128], "data": { "kind": "constant", "value": 255 }, "tolerance": 0 }
}
},
{
"name": "subgroup_min_scalar_rows128x257_ties_first",
"provenance": {
"notes": "Route-locks the unaligned one-minimum-subgroup path and verifies first-index ordering across all lane-local chunks."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": { "x": { "dtype": "float32", "shape": [128, 257], "data": { "kind": "constant", "value": -7.0 } } },
"outputs": {
"y": { "dtype": "uint32", "shape": [128], "data": { "kind": "constant", "value": 0 }, "tolerance": 0 }
}
},
{
"name": "subgroup_scalar_long_row_2x65",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 65],
"data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2] } }
},
{
"name": "empty_zero_dim",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": { "x": { "dtype": "float32", "shape": [0, 4], "data": { "kind": "values", "values": [] } } },
"outputs": { "y": { "dtype": "uint32", "shape": [0], "tolerance": 0 } }
},
{
"name": "empty_zero_dim_f16",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": { "x": { "dtype": "float16", "shape": [0, 4], "data": { "kind": "values", "values": [] } } },
"outputs": { "y": { "dtype": "uint32", "shape": [0], "tolerance": 0 } }
},
{
"name": "empty_zero_dim_int32",
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": { "x": { "dtype": "int32", "shape": [0, 4], "data": { "kind": "values", "values": [] } } },
"outputs": { "y": { "dtype": "uint32", "shape": [0], "tolerance": 0 } }
},
{
"name": "rank4_axis0_batch_argmin_int32",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_int32 (extended to rank4 batch axis)",
"notes": "Reduces the leading batch axis of a rank-4 NCHW int32 tensor."
},
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [3, 2, 2, 2],
"data": {
"kind": "values",
"values": [5, 9, 4, 8, 7, 2, 6, 3, 3, 1, 9, 2, 0, 8, 5, 7, 8, 4, 1, 6, 9, 3, 2, 1]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 2, 2], "tolerance": 0 } }
},
{
"name": "rank4_axis0_batch_argmin_f32_keepdims",
"provenance": {
"source": "cmake/external/onnx/onnx/backend/test/data/node/test_argmin_keepdims_random (extended to rank4 batch axis)",
"notes": "Same rank4 axis=0 selection hole on the default float32 dtype with keepdims=1, to prove the gap is dtype-independent (not an integer-only branch)."
},
"attrs": { "axis": 0, "keepdims": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [3, 2, 2, 2],
"data": {
"kind": "values",
"values": [4.0, -3.0, 5.0, 1.0, 2.0, -5.0, 3.0, -4.0, 6.0, 0.0, 3.0, -6.0, -1.0, 7.0, 2.0, 8.0, -9.0, 4.0, 5.0, 1.0, -2.0, 3.0, 0.0, 6.0]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [1, 2, 2, 2], "tolerance": 0 } }
},
{
"name": "rank4_axis1_channel_int8_nchw",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin_int8 (rank4 NCHW channel argmin)",
"notes": "int8 channel-axis argmin on a rank4 NCHW tensor. rank4_axis1 gates dtype only via f16Ok(T) (true for int8) so this is a WORKING-but-untested integer dtype on the rank4 selection path."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "int8",
"shape": [2, 4, 3, 3],
"data": {
"kind": "values",
"values": [3, 10, 17, 24, 31, 38, 45, 52, 59, 66, 73, 80, 87, 94, 101, 108, 115, 122, 127, -120, -113, -106, -99, -92, -85, -78, -71, -64, -57, -50, -43, -36, -29, -22, -15, -8, -1, 6, 13, 20, 27, 34, 41, 48, 55, 62, 69, 76, 83, 90, 97, 104, 111, 118, 125, -124, -117, -110, -103, -96, -89, -82, -75, -68, -61, -54, -47, -40, -33, -26, -19, -12]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 3, 3], "tolerance": 0 } }
},
{
"name": "rank3_axis1_f16_keepdims_select_last",
"provenance": {
"source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
"test": "ReductionOpTest.ArgMin (rank3 middle-axis f16 select_last)",
"notes": "f16 middle-axis argmin on rank3 with keepdims=1 and select_last_index=1. rank3_axis1 admits f16 via f16Ok(T) on an f16 device; no existing rank3/rank4 case is f16."
},
"attrs": { "axis": 1, "keepdims": 1, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float16",
"shape": [2, 5, 4],
"data": {
"kind": "values",
"values": [-1.5, -1.0, -0.5, 0.0, 0.5, 1.0, 1.5, -1.5, -1.0, -0.5, 0.0, 0.5, 1.0, 1.5, -1.5, -1.0, -0.5, 0.0, 0.5, 1.0, 1.5, -1.5, -1.0, -0.5, 0.0, 0.5, 1.0, 1.5, -1.5, -1.0, -0.5, 0.0, 0.5, 1.0, 1.5, -1.5, -1.0, -0.5, 0.0, 0.5]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 1, 4], "tolerance": 0 } }
},
{
"name": "axis1_serial_fallback_rows32768_cols5_select_last",
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [32768, 5],
"data": { "kind": "cycle", "values": [2.0, -3.0, -3.0, 1.0, -3.0, 2.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [32768], "tolerance": 0 } }
},
{
"name": "axis1_serial_fallback_rows32768_cols5_first_index",
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [32768, 5],
"data": { "kind": "cycle", "values": [2.0, -3.0, -3.0, 1.0, -3.0, 2.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [32768], "tolerance": 0 } }
},
{
"name": "axis0_fallback_rows4_int32_select_last_ties",
"attrs": { "axis": 0, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [4, 4],
"data": { "kind": "values", "values": [5, 1, 3, 2, 5, 7, 3, 2, 4, 7, 9, 2, 4, 1, 9, 8] }
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [4], "tolerance": 0, "data": { "kind": "values", "values": [3, 3, 1, 2] } }
}
},
{
"name": "axis1_int8_serial_fallback_negative_min_select_last",
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "int8",
"shape": [2, 8],
"data": {
"kind": "values",
"values": [-5, -5, -128, -128, 3, 3, 127, 127, 0, -128, 0, -127, -100, -100, -100, -50]
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2], "tolerance": 0 } }
},
{
"name": "axis0_splitk_i32_8192x16_precision_min",
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [8192, 16],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_i32_8192x16_precision_min_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [16], "tolerance": 0 } }
},
{
"name": "axis0_splitk_u32_8192x16_precision_min",
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "uint32",
"shape": [8192, 16],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_i32_8192x16_precision_min_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [16], "tolerance": 0 } }
},
{
"name": "axis0_tilecols_i32_64x16_precision_min",
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [64, 16],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_i32_8192x16_precision_min_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [16], "tolerance": 0 } }
},
{
"name": "axis0_tilecols_i32_64x16_precision_select_last_min",
"attrs": { "axis": 0, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [64, 16],
"data": {
"kind": "cycle",
"values": { "$ref": "#/fixtureArrays/axis0_splitk_i32_8192x16_precision_min_input_x" }
}
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [16], "tolerance": 0 } }
},
{
"name": "axis0_narrow_splitk_16384x8_f32",
"provenance": {
"notes": "Tall narrow axis-0 (16384 rows, 8 cols) routes to the split-and-tiled kernel: all 256 lanes cooperate across eight columns instead of leaving 248 lanes idle. Cross-checks the tiled partials and combine pass against the reference."
},
"attrs": { "axis": 0, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [16384, 8],
"data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29 }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [8], "tolerance": 0 } }
},
{
"name": "last_axis_split_vec4_logits_ties_first",
"provenance": {
"notes": "A single 32768-element row route-locks the cooperative split-row vec4 reduction. Equal minima span many split segments and must merge to the first global index."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": { "dtype": "float32", "shape": [1, 32768], "data": { "kind": "cycle", "values": [-7.0, 3.0, -7.0, 1.0] } }
},
"outputs": {
"y": { "dtype": "uint32", "shape": [1], "data": { "kind": "constant", "value": 0 }, "tolerance": 0 }
}
},
{
"name": "last_axis_split_scalar_logits_ties_last",
"provenance": {
"notes": "The unaligned 32769-element companion route-locks the scalar split-row reduction and verifies select_last_index across split boundaries."
},
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": { "x": { "dtype": "float32", "shape": [1, 32769], "data": { "kind": "constant", "value": -5.0 } } },
"outputs": {
"y": { "dtype": "uint32", "shape": [1], "data": { "kind": "constant", "value": 32768 }, "tolerance": 0 }
}
},
{
"name": "last_axis_split_vec4_i32_positive_rows",
"provenance": {
"notes": "int32 on the cooperative split-row vec4 reduction (32768 columns, two rows). Every value in the row is positive, so a lane identity of 0 instead of INT_MAX would beat the whole row and the merge would return the sentinel; this directly exercises the i32 identity and native signed compare. The period-7 cycle does not divide 32768, so the two rows are phase-shifted and their answers differ — a kernel that reduced row 0 for both rows cannot pass. ONNX ArgMin admits int32 via T and breaks ties on the first index."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": { "dtype": "int32", "shape": [2, 32768], "data": { "kind": "cycle", "values": [9, 5, 3, 1, 2, 4, 6] } }
},
"outputs": {
"y": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [3, 2] }, "tolerance": 0 }
}
},
{
"name": "last_axis_split_scalar_u32_above_int32_max",
"provenance": {
"notes": "uint32 on the unaligned (32769-column) scalar split-row reduction. Six of the seven cycle values exceed 2^31, so a signed compare would rank them negative and answer [4, 2] instead of [0, 5]. This directly exercises the unsigned identity 0xffffffffu and the partial store that must write the value through without bitcasting it. ONNX ArgMin admits uint32 via T; first index on ties."
},
"attrs": { "axis": 1, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "uint32",
"shape": [2, 32769],
"data": {
"kind": "cycle",
"values": [2147483000, 4200000000, 3000000000, 4294967290, 2500000000, 3500000000, 4000000000]
}
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [0, 5] }, "tolerance": 0 }
}
},
{
"name": "last_axis_split_vec4_f16_rows",
"provenance": {
"notes": "float16 on the vec4 split-row reduction: each vec4 component must be widened to f32 before comparison, and the shader must enable f16. Rank 3 (1, 2, 32768) reduces the last axis exactly as the rank-2 logits case does while keeping the reference on its hoisted-decode path. All seven cycle values are exactly representable in float16, so the answer does not depend on rounding. ONNX ArgMin admits float16 via T; first index on ties."
},
"attrs": { "axis": 2, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float16",
"shape": [1, 2, 32768],
"data": { "kind": "cycle", "values": [1.5, -2.0, 0.5, 3.5, -1.0, 2.25, -3.75] }
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [1, 2], "data": { "kind": "values", "values": [6, 5] }, "tolerance": 0 }
}
},
{
"name": "last_axis_split_scalar_f16_rows",
"provenance": {
"notes": "The unaligned (32769-column) companion: float16 on the scalar split-row reduction, where the widening happens on the plain load rather than on a vec4 component. Rank 3 keeps the reference on its hoisted-decode path; all cycle values are exact in float16. ONNX ArgMin admits float16 via T; first index on ties."
},
"attrs": { "axis": 2, "keepdims": 0 },
"inputs": {
"x": {
"dtype": "float16",
"shape": [1, 2, 32769],
"data": { "kind": "cycle", "values": [1.5, -2.0, 0.5, 3.5, -1.0, 2.25, -3.75] }
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [1, 2], "data": { "kind": "values", "values": [6, 4] }, "tolerance": 0 }
}
},
{
"name": "last_axis_split_vec4_ties_select_last_index",
"provenance": {
"notes": "select_last_index on the vec4 split-row route checks the last-index tie rule in both the segment reduction and combine. The minimum -7.0 recurs at cycle offsets 0, 2 and 5, so every segment reports a tie; the two phase-shifted rows must return 32767 and 32766."
},
"attrs": { "axis": 1, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 32768],
"data": { "kind": "cycle", "values": [-7.0, -3.0, -7.0, -1.0, -5.0, -7.0, -2.0] }
}
},
"outputs": {
"y": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [32767, 32766] }, "tolerance": 0 }
}
},
{
"name": "rank5_middle_axis_generic_geometry",
"provenance": {
"notes": "Proves the rank-independent (outer, axis, inner) fallback covers ranks beyond the former rank-4 manifest specializations."
},
"attrs": { "axis": 2, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2, 2, 3, 2, 2],
"data": { "kind": "cycle", "values": [3.0, -5.0, -1.0, -5.0, -2.0, 1.0] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [2, 2, 2, 2], "tolerance": 0 } }
},
{
"name": "axis0_splitk_select_last_index_ties_8192x16",
"provenance": {
"notes": "select_last_index=1 on the split-tiled route, which no case had ever rendered: every existing case for that geometry leaves the attribute at its default. The 17-value cycle is coprime with the 16 columns, so each column sees every value about 482 times and the minimum is a 482-way tie whose winning index differs per column -- exactly the tie-break the two arms disagree about, and one a per-lane fold can get wrong even when its lane-to-index map is monotone."
},
"attrs": { "axis": 0, "keepdims": 0, "select_last_index": 1 },
"inputs": {
"x": {
"dtype": "int32",
"shape": [8192, 16],
"data": { "kind": "cycle", "values": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16] }
}
},
"outputs": { "y": { "dtype": "uint32", "shape": [16], "tolerance": 0 } }
}
]
}