Module numerics 

Device-neutral numerical acceptance cases for hardware backends.

Each case couples a stable TOSA artifact with exact input shapes and a numerical oracle. Host backends consume the same bytes and values, so a provider cannot quietly substitute a backend-specific graph while claiming cross-device equivalence.

Structs

Bfloat16Tensor

One immutable bfloat16 tensor represented by exact storage bits.

Float16Tensor

One immutable IEEE-754 binary16 tensor in a numerical acceptance case.

Float32Tensor

One immutable FP32 tensor in a numerical acceptance case.

Int32Tensor

One immutable INT32 tensor produced by an integer-profile acceptance case.

PackedTensor

One immutable packed low-precision tensor.

TosaBfloat16Case

A stable TOSA EXT-BF16 graph and bit-exact bfloat16 numerical oracle.

TosaFloat16Case

A stable TOSA graph and binary16 numerical oracle shared by host backends.

TosaFloat32Case

A stable TOSA graph and FP32 numerical oracle shared by host backends.

TosaFp8ToBfloat16Case

A stable explicit FP8 → BF16 CAST with a bit-exact output oracle.

TosaFp32OperatorCase

One FP32-tier operator case shared by host backends: a stable TOSA 1.0 graph over FP32 tensors with BOOL and INT32 auxiliaries, and a numerical oracle for its single output.

TosaInt8MatmulCase

A stable TOSA INT8 matrix multiplication with exact INT32 accumulation.

TosaInt32ToInt8RescaleCase

A stable TOSA INT32-to-INT8 RESCALE with exact fixed-point rounding and saturation.

TosaPackedCase

A stable TOSA graph and packed low-precision oracle shared by host backends.

TosaRawCase

One mixed-type TOSA operator case with a numerical output oracle.

Enums

Fp32TierTensor

Raw tensor storage used by the shared FP32-tier operator cases.

PackedDType

Packed scalar encoding used by a low-precision TOSA acceptance case.

RawTensor

Raw tensor storage used by mixed-type Hexagon operator-parity cases.

Constants

ADD_FP16

Binary16 broadcast addition over [2, 1] and [1, 3] inputs.

CAST_FP8E4M3_TO_BF16

Explicit FP8 E4M3 → BF16 CAST spanning signed zeros, subnormals, ordinary values, finite maximum, and NaN. All 256 byte encodings repeat four times across one XDNA conversion tile.

CAST_FP8E5M2_TO_BF16

Explicit FP8 E5M2 → BF16 CAST spanning signed zeros, subnormals, one, finite maximum, infinity, and NaN. All 256 byte encodings repeat four times across one XDNA conversion tile.

FP32_BINARY_CASES

FP32 broadcast binary operators over [2, 1] and [1, 3] inputs producing [2, 3].

FP32_LOGICAL_CASES

FP32 comparisons, BOOL logic, and selection.

FP32_MOVEMENT_CASES

Static FP32 constants and data-movement operators.

FP32_OPERATOR_CASE_GROUPS

Every FP32-tier operator case group, for backends that iterate the whole tier.

FP32_REDUCTION_CASES

FP32 reductions and INT32 argmax over a [2, 3] input along axis 1.

FP32_UNARY_CASES

FP32 unary and activation operators over [0.5, 1, 2, 4].

HEXAGON_LOGICAL_CASES

Mixed BOOL/FP16 comparison, logical, and selection cases.

HEXAGON_MOVEMENT_CASES

Static constants and FP16 data-movement cases.

HEXAGON_REDUCTION_CASES

FP16 reductions and INT32 argmax over a two-row input.

HEXAGON_UNARY_FP16_CASES

FP16 unary and activation cases supported by the QNN HTP operator package.

IDENTITY_EDGES_FP16

Binary16 identity over NaN, infinities, signed zeros, a subnormal, and finite values.

IDENTITY_EDGES_FP32

FP32 identity over NaN, infinities, signed zeros, a subnormal, and ordinary finite values.

IDENTITY_FP8E4M3

FP8 E4M3 identity over signed zeros, subnormals, ordinary values, finite maximum, and NaN.

IDENTITY_FP8E5M2

FP8 E5M2 identity over signed zeros, subnormals, one, finite maximum, infinity, and NaN.

IDENTITY_INT4

Packed INT4 identity spanning the TOSA-defined finite range.

IDENTITY_INT8

INT8 identity spanning negative, zero, and positive values.

LINEAR_TANH_FP32

Three-operator FP32 graph: tanh(x · w + bias) with constant zero points and a [1, 1, 2] bias broadcast over the [1, 2, 2] product. Inputs are the mock classifier’s features and weights; the oracle is evaluated in binary64 and rounded.

MATMUL_FP16

Binary16 batched matrix multiplication with non-square operands.

MATMUL_FP32

FP32 batched matrix multiplication with non-square operands.

MATMUL_INT8

Non-square INT8 batched matrix multiplication with nonzero zero points and INT32 accumulation.

MAXIMUM_FP16

Binary16 broadcast maximum over [2, 1] and [1, 3] inputs.

MAX_POOL2D_BF16

BF16 two-channel NHWC max pooling with a 2x2 kernel, stride two, zero padding, and an exact integer-valued oracle.

MAX_POOL2D_FP16

Binary16 two-channel NHWC max pooling with a 2x2 kernel and stride two.

MAX_POOL2D_FP32

FP32 two-channel NHWC max pooling with a 2x2 kernel and stride two.

MINIMUM_FP16

Binary16 broadcast minimum over [2, 1] and [1, 3] inputs.

MOCK_LINEAR_CLASSIFIER_FP16

Two-sample, three-feature FP16 linear classifier with a direct-bound 3x2 weight matrix.

MUL_FP16

Binary16 broadcast multiplication with a compile-time zero shift.

POW_FP16

Binary16 broadcast power over [2, 1] and [1, 3] inputs.

QUANTIZED_CLASSIFIER_INT8

Exact INT8 quantized linear classifier: two samples, three features, two classes, INT32 logits with unambiguous per-sample argmax winners.

RESCALE_INT32_TO_INT8

Signed scale32 RESCALE covering ties, negative values, saturation, and a nonzero output point.

SUB_FP16

Binary16 broadcast subtraction over [2, 1] and [1, 3] inputs.