High-level initializer helpers (device-dispatching) for tensors. More...

#include <concepts>
#include <memory>
#include <span>
#include <type_traits>
#include <cstdint>
import Compute.DeviceType;
import Compute.ExecutionContext;
import Dnn.TensorDataTypeTraits;
import Dnn.TensorDataTypeMap;
import Dnn.TensorOps.Base;
import Dnn.TensorDataType;
import Dnn.Tensor;

Namespaces
namespace	Mila
	Mila main API namespace.

Functions
template<TensorDataType TDataType, typename TMemoryResource> requires isValidTensor<TDataType, TMemoryResource>
void	Mila::Dnn::fill (Tensor< TDataType, TMemoryResource > &tensor, host_value_t< TDataType > host_value, IExecutionContext *exec_context=nullptr)
	Fill a tensor with a scalar host value (device-dispatched) with optional ExecutionContext.
template<TensorDataType TDataType, typename TMemoryResource> requires isValidTensor<TDataType, TMemoryResource>
void	Mila::Dnn::fill (Tensor< TDataType, TMemoryResource > &tensor, std::span< const host_value_t< TDataType > > host_values, IExecutionContext *exec_context=nullptr)
	Copy host values into a tensor with device dispatch and optional ExecutionContext.

Detailed Description

High-level initializer helpers (device-dispatching) for tensors.

This partition provides the generic, device-agnostic entry points for tensor initialization operations (copying host-provided values into tensors and filling tensors with scalar values). The implementation forwards to the device-specific TensorOps<...> partitions (for example TensorOps<Compute::CpuComputeDeviceTag>::fill(...)).

The host representation for a logical tensor element is defined by TensorDataTypeTraits<TDataType>::host_type and is exposed here via the alias host_value_t<TDataType> so callers and implementations use a single, authoritative host-side type for conversions.

ExecutionContext handling:

Optional ExecutionContext parameter for stream control (borrowed, not owned)
When provided, operations use the context's stream (caller controls sync)
When null, operations use default stream and synchronize before returning
Raw pointer semantics ensure zero overhead

Usage:

Call fill(tensor, values) or fill(tensor, scalar) from user code.
Optionally provide ExecutionContext for explicit stream control: fill(tensor, values, ctx.get())
The host value type is selected automatically from the tensor data type (float for floating-point tensors, int32_t for integer tensors).

Preconditions:

The isValidTensor<TDataType,TMemoryResource> concept must hold for the tensor types used here (ensures memory resource compatibility and trait availability).
ExecutionContext (if provided) must outlive the function call.

Function Documentation

◆ fill() [1/2]

template<TensorDataType TDataType, typename TMemoryResource>
requires isValidTensor<TDataType, TMemoryResource>

void Mila::Dnn::fill	(	Tensor< TDataType, TMemoryResource > &	tensor,
		host_value_t< TDataType >	host_value,
		IExecutionContext *	exec_context = nullptr )

export

Fill a tensor with a scalar host value (device-dispatched) with optional ExecutionContext.

Forwards scalar fills to the device-specific TensorOps<Tag>::fill. Borrows execution context for stream control with zero overhead. The function signature enforces the expected host scalar representation for each abstract tensor data type via host_value_t<TDataType>.

Template Parameters

TDataType	Abstract tensor data type.
TMemoryResource	Memory resource type backing the tensor.

Parameters

tensor	Destination tensor to be filled. Must satisfy isValidTensor.
host_value	Scalar value in host representation to broadcast to the tensor.
exec_context	Optional execution context for stream control (borrowed, not owned)

Note: exec_context must outlive this function call; When exec_context provided, caller controls synchronization; When exec_context is null, uses default stream and synchronizes before returning; For CUDA tensors, use CudaExecutionContext; for CPU, parameter is ignored

Example:

// With explicit context (async)
auto ctx = std::make_unique<CudaExecutionContext>(0);
fill(float_tensor, 3.14f, ctx.get());
fill(int_tensor, 42, ctx.get());
ctx->synchronize();
 
// Without context (sync)
fill(float_tensor, 3.14f);  // Returns after completion
fill(int_tensor, 42);       // Returns after completion

◆ fill() [2/2]

template<TensorDataType TDataType, typename TMemoryResource>
requires isValidTensor<TDataType, TMemoryResource>

void Mila::Dnn::fill	(	Tensor< TDataType, TMemoryResource > &	tensor,
		std::span< const host_value_t< TDataType > >	host_values,
		IExecutionContext *	exec_context = nullptr )

export

Copy host values into a tensor with device dispatch and optional ExecutionContext.

Forwards the host->tensor copy operation (span form) to the device-specific implementation TensorOps<Tag>::fill. Borrows execution context for stream control with zero overhead. Falls back to default stream when no context provided.

The host element type is selected by host_value_t<TDataType> so callers must provide values in the expected host representation (float for floating-point tensor types, int32_t for integer tensor types). The device implementation performs any necessary conversion/quantization.