include_generator_kernels.h

namespace onnx_light

Alias that makes onnx-light headers compatible with code that references ONNX_LIGHT_NAMESPACE (the macro used in the standard onnx package).

Set to ONNX_LIGHT_NAMESPACE so both names resolve to the same namespace.

Symbol-visibility attribute for the public onnx-light C++ API.

Defined as empty because onnx-light does not require explicit __declspec(dllexport) or __attribute__((visibility("default"))) annotations — visibility is controlled at the shared-library level. The macro is provided so that vendored ONNX headers that decorate their declarations with ONNX_API compile without modification.

Namespace alias so that ONNX C++ code (and consumers such as onnxruntime) that refers to the literal onnx namespace — rather than the ONNX_NAMESPACE macro — resolves to the onnx-light namespace. The standard onnx package lives in namespace onnx; onnx-light uses onnx_light (via ONNX_LIGHT_NAMESPACE), so this alias keeps onnx-light a true drop-in. It is only introduced when the onnx-light namespace differs from onnx.

namespace onnx_kernels

namespace kernel

class Bernoulli : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX Bernoulli operator (since opset 15 in the ai.onnx domain). Draws binary samples y[i] from a Bernoulli distribution with probability input[i] (a value in [0, 1]); returns 1 with probability input[i] and 0 otherwise.

Bernoulli is non-deterministic; this reference implementation uses a std::mt19937 engine seeded either with the value of the optional seed attribute (interpreted by truncating to uint32_t) or, when the attribute is absent, with a fixed default seed so the kernel is reproducible for testing. Output dtype is controlled by the optional dtype attribute: when absent the output element type matches the input; when present it overrides the output element type.

Supported input dtypes are FLOAT, DOUBLE and FLOAT16; supported output dtypes are FLOAT, DOUBLE, FLOAT16, INT8, INT16, INT32, INT64, UINT8, UINT16, UINT32, UINT64 and BOOL (every type for which the produced 0/1 value has a natural representation).

Public Functions

Tensor operator()(const Tensor &input, int64_t seed = kNoSeed, int32_t dtype = 0) const

Draws Bernoulli samples for every element of input. seed is the value of the seed attribute when present (truncated to uint32_t); pass kNoSeed to use the kernel’s default seed. dtype is the value of the dtype attribute when present (a :cpp:class:DataType value); pass 0 to keep the output dtype identical to input.data_type.

void operator()(const Tensor &input, int64_t seed, int32_t dtype, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

The output buffer has the same byte size as the input only when dtype is identical to the input dtype, which we do not require; disable in-place support to keep the contract simple.

Public Static Attributes

static constexpr int64_t kNoSeed = -1

Sentinel value indicating the seed attribute is absent. Picked outside the uint32_t range so any 32-bit seed (including 0) is representable as a regular value.

class Constant : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Returns a copy of the value attribute of the Constant op.

Public Functions

Tensor operator()(const Tensor &value) const

void operator()(const Tensor &value, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

class ConstantOfShape : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX ConstantOfShape operator (since opset 9 in the ai.onnx domain). Produces an output tensor of the shape given by the 1-D int64 shape input, filled with the (single) scalar value of the value tensor.

When value is empty (data_type == 0 and data.empty()), the output defaults to a FLOAT tensor filled with zeros, per the schema.

Supported numeric value dtypes match the upstream test_constantofshape_int_zeros / test_constantofshape_float_ones node tests: every fixed-width whole-byte numeric type as well as BOOL. Other dtypes throw std::invalid_argument.

Public Functions

Tensor operator()(const Tensor &shape, const Tensor &value) const

shape must be a 1-D INT64 tensor whose entries describe the shape of the output. value is the (single-element) fill value taken from the operator’s value attribute; pass a default-constructed Tensor to use the schema default (FLOAT 0.0).

void operator()(const Tensor &shape, const Tensor &value, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

The output buffer has a different size than the inputs, so storage can not generally be shared.

class EyeLike : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of ONNX EyeLike (since opset 9): returns a matrix with ones on diagonal k and zeros elsewhere, with output shape copied from the 2-D input tensor.

Public Functions

Tensor operator()(const Tensor &input, int64_t k = 0, int32_t dtype = 0) const

void operator()(const Tensor &input, int64_t k, int32_t dtype, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

class Multinomial : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX Multinomial operator (since opset 7 in the ai.onnx domain). Draws sample_size samples from a multinomial distribution per batch row, where the per-row unnormalized log-probabilities are given by the 2-D input tensor of shape [batch_size, class_size]. The output is a 2-D tensor of shape [batch_size, sample_size] of integer class indices.

Multinomial is non-deterministic; this reference implementation uses a std::mt19937 engine seeded either with the value of the optional seed attribute (interpreted by truncating to uint32_t) or, when the attribute is absent, with a fixed default seed so the kernel is reproducible for testing.

Supported input dtypes are FLOAT, DOUBLE and FLOAT16. Supported output dtypes are INT32 (default) and INT64.

Public Functions

Tensor operator()(const Tensor &input, int64_t sample_size = 1, int64_t seed = kNoSeed, int32_t dtype = 0) const

Draws sample_size samples per batch row of input. seed is the value of the seed attribute when present (truncated to uint32_t); pass kNoSeed to use the kernel’s default seed. dtype is the value of the dtype attribute; pass 0 (or :cpp:enumerator:DataType::INT32) to produce INT32 output, or :cpp:enumerator:DataType::INT64 for INT64 output.

void operator()(const Tensor &input, int64_t sample_size, int64_t seed, int32_t dtype, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

The output buffer has a different shape than the input, so storage cannot generally be shared.

Public Static Attributes

static constexpr int64_t kNoSeed = -1

Sentinel meaning seed attribute is absent.

class RandomNormal : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX RandomNormal operator (since opset 1 in the ai.onnx domain). Produces an output tensor of the given shape filled with samples drawn from a normal distribution with mean mean and standard deviation scale.

RandomNormal is non-deterministic; this reference kernel uses the deterministic :cpp:func:Randn helper (SplitMix64 + Irwin-Hall) so expected outputs are bit-stable for testing. When seed is :cpp:var:kNoSeed the kernel’s default seed is used; otherwise the optional seed attribute (declared as FLOAT in the schema) is truncated to uint64_t and forwarded to :cpp:func:Randn.

Supported output dtypes are FLOAT (default) and DOUBLE; any other value of dtype triggers std::invalid_argument.

Public Functions

Tensor operator()(const std::vector<int64_t> &shape, double mean = 0.0, double scale = 1.0, int64_t seed = kNoSeed, int32_t dtype = 0) const

Produces a shape-shaped tensor of normal samples. mean and scale come from the operator attributes (defaults 0.0 / 1.0). dtype defaults to DataType::FLOAT to mirror the schema default. seed selects the RNG seed; pass kNoSeed to use the kernel’s default.

void operator()(const std::vector<int64_t> &shape, double mean, double scale, int64_t seed, int32_t dtype, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

No input tensor to alias with: in-place semantics are not relevant.

Public Static Attributes

static constexpr int64_t kNoSeed = -1

Sentinel meaning seed attribute is absent.

class RandomNormalLike : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX RandomNormalLike operator (since opset 1 in the ai.onnx domain). Produces an output tensor whose shape matches input filled with samples drawn from a normal distribution. When dtype is 0 the output dtype equals the input dtype (which must be one of the supported output dtypes); otherwise the value overrides the output dtype.

Public Functions

Tensor operator()(const Tensor &input, double mean = 0.0, double scale = 1.0, int64_t seed = kNoSeed, int32_t dtype = 0) const

void operator()(const Tensor &input, double mean, double scale, int64_t seed, int32_t dtype, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

Public Static Attributes

static constexpr int64_t kNoSeed = -1

class RandomUniform : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX RandomUniform operator (since opset 1 in the ai.onnx domain). Produces an output tensor of the given shape filled with samples drawn uniformly from the half-open interval [low, high).

Uses the deterministic :cpp:func:Rand helper (SplitMix64). Supported output dtypes are FLOAT (default) and DOUBLE; any other value of dtype triggers std::invalid_argument.

Public Functions

Tensor operator()(const std::vector<int64_t> &shape, double low = 0.0, double high = 1.0, int64_t seed = kNoSeed, int32_t dtype = 0) const

Produces a shape-shaped tensor of uniform samples in [low, high). low defaults to 0.0 and high to 1.0 to mirror the schema. dtype defaults to DataType::FLOAT. seed selects the RNG seed; pass kNoSeed to use the kernel’s default.

void operator()(const std::vector<int64_t> &shape, double low, double high, int64_t seed, int32_t dtype, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

Public Static Attributes

static constexpr int64_t kNoSeed = -1

Sentinel meaning seed attribute is absent.

class RandomUniformLike : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX RandomUniformLike operator (since opset 1 in the ai.onnx domain). Same semantics as :cpp:class:RandomUniform but the output shape is copied from input.

Public Functions

Tensor operator()(const Tensor &input, double low = 0.0, double high = 1.0, int64_t seed = kNoSeed, int32_t dtype = 0) const

void operator()(const Tensor &input, double low, double high, int64_t seed, int32_t dtype, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

Public Static Attributes

static constexpr int64_t kNoSeed = -1

class Range : public onnx_light::onnx_kernels::kernel::KernelBase

#include <include_generator_kernels.h>

Reference implementation of the ONNX Range operator (since opset 11 in the ai.onnx domain, with opset 27 widening the supported types to include FLOAT16/BFLOAT16). Generates a 1-D tensor with values [start, start + delta, start + 2*delta, ...] up to (but excluding) limit.

The three inputs start, limit and delta must all be scalar tensors of the same element type. Supported element types are FLOAT, DOUBLE, INT16, INT32, INT64, FLOAT16 and BFLOAT16 (matching the upstream Range schema’s T type-constraint at opset 27). For FLOAT16 and BFLOAT16 the loop accumulator runs in float (the stash_type = 1 default introduced at opset 27).

The number of output elements is max(ceil((limit - start) / delta), 0). When the result would be empty (e.g. start >= limit with a positive delta), the kernel returns an empty 1-D tensor of length 0.

Public Functions

Tensor operator()(const Tensor &start, const Tensor &limit, const Tensor &delta) const

void operator()(const Tensor &start, const Tensor &limit, const Tensor &delta, Tensor &output) const

inline explicit KernelBase(const KernelContext &ctx)

Public Static Functions

static inline constexpr bool CanRunInPlace() noexcept

The output buffer has a different size than the inputs, so storage cannot be shared.