Mila::Dnn::Compute::Cuda::RandomOps Struct Reference

Inheritance diagram for Mila::Dnn::Compute::Cuda::RandomOps:

Static Public Member Functions
template<TensorDataType TDataType, typename TMemoryResource> requires isValidTensor<TDataType, TMemoryResource> && TensorDataTypeTraits<TDataType>::is_float_type
static void	fill_normal (Tensor< TDataType, TMemoryResource > &tensor, float mean, float stddev, IExecutionContext *exec_context=nullptr)
	Fill a float tensor with values drawn from N(mean, stddev^2) using cuRAND.
template<TensorDataType TDataType, typename TMemoryResource> requires isValidTensor<TDataType, TMemoryResource> && TensorDataTypeTraits<TDataType>::is_float_type
static void	fill_uniform (Tensor< TDataType, TMemoryResource > &tensor, float min_val, float max_val, IExecutionContext *exec_context=nullptr)
	Fill a float tensor with uniform values in [min_val, max_val) using cuRAND.

Static Private Member Functions
static curandGenerator_t	make_temp_generator_ (cudaStream_t stream)

Member Function Documentation

fill_normal()

template<TensorDataType TDataType, typename TMemoryResource>
requires isValidTensor<TDataType, TMemoryResource> && TensorDataTypeTraits<TDataType>::is_float_type

void Mila::Dnn::Compute::Cuda::RandomOps::fill_normal ( Tensor< TDataType, TMemoryResource > & tensor, float mean, float stddev, IExecutionContext * exec_context = nullptr )

inlinestatic

Fill a float tensor with values drawn from N(mean, stddev^2) using cuRAND.

When exec_context is provided, the cached generator from CudaExecutionContext is reused. Without a context, a temporary generator is created and destroyed per call, seeded from Core::RandomGenerator.

curandGenerateNormal requires an even element count (Box-Muller pairs). If the tensor has an odd element count, a temporary device buffer of size n+1 is allocated, the full even count is generated into it, and n elements are copied to the tensor. This is zero-overhead for even-sized tensors.

Template Parameters

TDataType	Floating-point tensor data type.
TMemoryResource	CUDA memory resource type.

Parameters

tensor	Destination tensor (pre-allocated).
mean	Mean of the normal distribution.
stddev	Standard deviation of the normal distribution.
exec_context	Optional execution context for stream and generator reuse (borrowed, not owned).

Note

Only FP32 native tensors are currently supported. FP16/BF16 require a temporary float buffer with a subsequent conversion pass (not yet implemented).

Exceptions

std::runtime_error

On cuRAND or CUDA failure.

Here is the call graph for this function:

fill_uniform()

template<TensorDataType TDataType, typename TMemoryResource>
requires isValidTensor<TDataType, TMemoryResource> && TensorDataTypeTraits<TDataType>::is_float_type

void Mila::Dnn::Compute::Cuda::RandomOps::fill_uniform ( Tensor< TDataType, TMemoryResource > & tensor, float min_val, float max_val, IExecutionContext * exec_context = nullptr )

inlinestatic

Fill a float tensor with uniform values in [min_val, max_val) using cuRAND.

cuRAND generates values in [0, 1), which are scaled and shifted to [min_val, max_val) by a device kernel. When exec_context is provided, the cached generator is reused. Without a context, a temporary generator is created and destroyed per call.

Template Parameters

TDataType	Floating-point tensor data type.
TMemoryResource	CUDA memory resource type.

Parameters

tensor	Destination tensor (pre-allocated).
min_val	Lower bound of the uniform range (inclusive).
max_val	Upper bound of the uniform range (exclusive).
exec_context	Optional execution context for stream and generator reuse (borrowed, not owned).

Note

Only FP32 native tensors are currently supported.

Exceptions

std::runtime_error

On cuRAND or CUDA failure.

Here is the call graph for this function:

make_temp_generator_()

curandGenerator_t Mila::Dnn::Compute::Cuda::RandomOps::make_temp_generator_ ( cudaStream_t stream )

inlinestaticprivate

Here is the call graph for this function:

Here is the caller graph for this function:

---

The documentation for this struct was generated from the following file:

• /__w/Mila/Mila/Mila/Src/Dnn/Compute/Devices/Cuda/Tensors/Operations/CudaTensorOps.Random.ixx