{
  "cases": [
    {
      "name": "vector",
      "inputs": {
        "x": { "dtype": "float32", "shape": [32], "data": { "kind": "fillFloat32", "sinStep": 0.1, "cosStep": 0.2 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.000001 } }
    },
    {
      "name": "rank0_scalar",
      "inputs": { "x": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [-1.25] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
    },
    {
      "name": "saturation_and_special_values",
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [6],
          "data": { "kind": "values", "values": [100.0, -100.0, "Infinity", "-Infinity", "NaN", 0.0] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [6], "tolerance": 0.000001, "allowNaN": true } }
    },
    {
      "name": "f32_subnormal_identity_tail_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 (f32 and f16); the kernel cannot preserve denormal inputs/outputs bit-exactly."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
        "test": "ActivationOpTest.Tanh",
        "notes": "For tiny finite inputs tanh(x) rounds back to x in float32; zero-flushing erases the signed tail."
      },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [4],
          "data": { "kind": "values", "values": [1e-40, -1e-40, 1e-39, -1e-39] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 1e-43 } }
    },
    {
      "name": "f32_subnormal_identity_tail_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 (f32 and f16); the kernel cannot preserve denormal inputs/outputs bit-exactly."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
        "test": "ActivationOpTest.Tanh",
        "notes": "On the scalar path, tiny finite inputs should round back to their signed subnormal values."
      },
      "inputs": {
        "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [-1e-40, 0.0, 1e-40] } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 1e-43 } }
    },
    {
      "name": "ort_activation_extremes",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
        "test": "ActivationOpTest.Tanh",
        "notes": "ORT shared activation vector with infinities, float max, min, and subnormal values."
      },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [13],
          "data": {
            "kind": "values",
            "values": [-1.0, 0.0, 1.0, 100.0, -100.0, 1000.0, -1000.0, 1.1754943508222875e-38, 1.1754943508222876e-39, -1.1754943508222876e-39, 3.4028234663852886e+38, -3.4028234663852886e+38, "Infinity"]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [13], "tolerance": 0.000001 } }
    },
    {
      "name": "ort_fp16_activation_extremes",
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
        "test": "ActivationOpTest.Tanh_fp16"
      },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [13],
          "data": {
            "kind": "values",
            "values": [-1.0, 0.0, 1.0, 100.0, -100.0, 1000.0, -1000.0, 1.1754943508222875e-38, 1.1754943508222876e-39, -1.1754943508222876e-39, 3.4028234663852886e+38, -3.4028234663852886e+38, "Infinity"]
          }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [13], "tolerance": 0.001 } }
    },
    {
      "name": "f16_subnormal_identity_tail_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 (f32 and f16); the kernel cannot preserve denormal inputs/outputs bit-exactly."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
        "test": "ActivationOpTest.Tanh_fp16",
        "notes": "Exact half-subnormal inputs are valid f16 values; tanh(x) should round back to the same signed tail."
      },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [4],
          "data": {
            "kind": "values",
            "values": [5.960464477539063e-8, -5.960464477539063e-8, 1.1920928955078125e-7, -1.1920928955078125e-7]
          }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [4], "tolerance": 0 } }
    },
    {
      "name": "f16_subnormal_identity_tail_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 (f32 and f16); the kernel cannot preserve denormal inputs/outputs bit-exactly."
      },
      "provenance": {
        "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
        "test": "ActivationOpTest.Tanh_fp16",
        "notes": "The scalar path exercises exact float16 subnormal Tanh identity behavior."
      },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [3],
          "data": { "kind": "values", "values": [-5.960464477539063e-8, 0.0, 5.960464477539063e-8] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [3], "tolerance": 0 } }
    },
    {
      "name": "onnx_backend_example",
      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_tanh_example" },
      "inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [-1.0, 0.0, 1.0] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
    },
    {
      "name": "onnx_backend_tanh",
      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_tanh" },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [3, 4, 5],
          "data": {
            "kind": "values",
            "values": [1.764052391052246, 0.40015721321105957, 0.978738009929657, 2.2408931255340576, 1.8675580024719238, -0.9772778749465942, 0.9500884413719177, -0.15135720372200012, -0.10321885347366333, 0.4105985164642334, 0.14404356479644775, 1.4542734622955322, 0.7610377073287964, 0.12167501449584961, 0.44386324286460876, 0.3336743414402008, 1.4940791130065918, -0.2051582634449005, 0.3130677044391632, -0.8540957570075989, -2.5529897212982178, 0.653618574142456, 0.8644362092018127, -0.7421650290489197, 2.269754648208618, -1.4543657302856445, 0.04575851559638977, -0.18718385696411133, 1.5327792167663574, 1.4693588018417358, 0.154947429895401, 0.37816253304481506, -0.8877857327461243, -1.980796456336975, -0.34791216254234314, 0.15634897351264954, 1.2302906513214111, 1.202379822731018, -0.38732680678367615, -0.302302747964859, -1.0485529899597168, -1.420017957687378, -1.7062702178955078, 1.950775384902954, -0.5096521973609924, -0.4380742907524109, -1.2527953386306763, 0.7774903774261475, -1.6138978004455566, -0.21274028718471527, -0.8954665660858154, 0.38690251111984253, -0.5108051300048828, -1.18063223361969, -0.02818222902715206, 0.4283318817615509, 0.06651721894741058, 0.30247190594673157, -0.6343221068382263, -0.3627411723136902]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [3, 4, 5], "tolerance": 0.00001 } }
    },
    {
      "name": "onnx_backend_tanh_example",
      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_tanh_example" },
      "inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [-1.0, 0.0, 1.0] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.00001 } }
    },
    {
      "name": "vec4_f16_lanes",
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [16],
          "data": {
            "kind": "values",
            "values": [-6.0, -4.0, -3.0, -2.0, -1.5, -1.0, -0.5, -0.25, 0.0, 0.25, 0.5, 1.0, 1.5, 2.0, 4.0, 6.0]
          }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [16], "tolerance": 0.001, "relTolerance": 0.002 } }
    },
    {
      "name": "vec4_f32_extremes",
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8],
          "data": { "kind": "values", "values": [100.0, -100.0, "Infinity", "-Infinity", "NaN", 0.0, 9.9, -9.9] }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8], "tolerance": 0.000001, "allowNaN": true } }
    },
    {
      "name": "empty_input_zero_dim",
      "inputs": { "x": { "dtype": "float32", "shape": [0], "data": { "kind": "values", "values": [] } } },
      "outputs": { "y": { "dtype": "float32", "shape": [0], "tolerance": 0 } }
    },
    {
      "name": "f16_scalar_fallback_saturation_and_tiny",
      "provenance": {
        "source": "synthetic",
        "notes": "Seven float16 elements exercise scalar processing. Exact float16 inputs cover saturation at magnitude 10 and the tiny-normal identity branch at 2^-13 without relying on subnormal preservation."
      },
      "inputs": {
        "x": {
          "dtype": "float16",
          "shape": [7],
          "data": { "kind": "values", "values": [10.0, -10.0, 0.0, 0.5, -0.5, 0.0001220703125, -0.0001220703125] }
        }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [7], "tolerance": 0.003 } }
    },
    {
      "name": "f16_scalar_fallback_midrange",
      "provenance": {
        "source": "synthetic",
        "notes": "Second odd-numel (5%4!=0) f16 scalar-fallback case over the steep mid-range of tanh where output precision matters most; complements the saturation/tiny case. f16-exact inputs (quarter-fractions)."
      },
      "inputs": {
        "x": { "dtype": "float16", "shape": [5], "data": { "kind": "values", "values": [-2.0, -0.75, 0.25, 1.0, 1.5] } }
      },
      "outputs": { "y": { "dtype": "float16", "shape": [5], "tolerance": 0.003 } }
    },
    {
      "name": "vec4_path_tiny_x_identity_branch",
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [8],
          "data": {
            "kind": "values",
            "values": [0.00001, -0.00001, 0.00005, -0.00005, 0.000099, -0.000099, 0.0001, 0.0]
          }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8], "tolerance": 1e-7 } }
    },
    {
      "name": "f32_vec4_sustained_4096",
      "provenance": { "notes": "A compact aligned float32 payload exercises vectorized Tanh across multiple lanes." },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [4096],
          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.007, "scale": 0.5 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4096], "tolerance": 0.000001 } }
    },
    {
      "name": "f32_scalar_tail_4095",
      "provenance": { "notes": "An odd float32 element count exercises scalar Tanh." },
      "inputs": {
        "x": {
          "dtype": "float32",
          "shape": [4095],
          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.011, "scale": 0.5 }
        }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4095], "tolerance": 0.000001 } }
    }
  ]
}