TensorHostTypeMap.ixx File Reference

Device-agnostic host type mapping for abstract TensorDataType enumeration. More...

#include <cstdint>
#include <type_traits>
import Dnn.TensorDataType;

Classes
struct	Mila::Dnn::TensorHostTypeMap< TDataType >
	Maps abstract TensorDataType to host-compatible C++ type and TensorDataType. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::BF16 >
	Host type for 16-bit brain floating point. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::FP16 >
	Host type for 16-bit half precision floating point. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::FP32 >
	Host type for 32-bit IEEE 754 floating point. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::FP8_E4M3 >
	Host type for 8-bit floating point with E4M3 format. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::FP8_E5M2 >
	Host type for 8-bit floating point with E5M2 format. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::INT16 >
	Host type for 16-bit signed integer. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::INT32 >
	Host type for 32-bit signed integer. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::INT8 >
	Host type for 8-bit signed integer. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::UINT16 >
	Host type for 16-bit unsigned integer. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::UINT32 >
	Host type for 32-bit unsigned integer. More...
struct	Mila::Dnn::TensorHostTypeMap< TensorDataType::UINT8 >
	Host type for 8-bit unsigned integer. More...

Namespaces
namespace	Mila
	Mila main API namespace.

Typedefs
template<TensorDataType TDataType>
using	Mila::Dnn::host_type_t = typename TensorHostTypeMap<TDataType>::host_type
	Convenience alias for accessing host type mapping.

Variables
template<TensorDataType TDataType>
constexpr bool	Mila::Dnn::is_host_float_type = std::is_floating_point_v<host_type_t<TDataType>>
	Checks if a TensorDataType maps to a floating-point host type.
template<TensorDataType TDataType>
constexpr bool	Mila::Dnn::is_host_integer_type = std::is_integral_v<host_type_t<TDataType>>
	Checks if a TensorDataType maps to an integer host type.

Detailed Description

Device-agnostic host type mapping for abstract TensorDataType enumeration.

This module provides centralized mapping from abstract tensor data types to appropriate host-compatible C++ types for conversion and host-side processing. These mappings are device-agnostic and define the standard host representation for each abstract tensor data type across all compute backends.

Key architectural principles:

• Device-agnostic: Host types are consistent across all compute devices
• Conversion-oriented: Optimized for host-device data transfers
• Direct type mappings: Preserves precision and signedness where possible
• Single source of truth: Eliminates duplication across device-specific files

Usage patterns:

• Host-device data transfers and conversions
• Fill operations using host-provided values
• Debugging and inspection of tensor values
• External API integration requiring standard C++ types

Design rationale:

• Direct mappings preserve type characteristics (precision, signedness)
• Floating-point types use appropriate host representations
• Integer types maintain their original bit width and signedness
• Maps are closed via explicit specializations with clear error messages

Typedef Documentation

host_type_t

template<TensorDataType TDataType>

using Mila::Dnn::host_type_t = typename TensorHostTypeMap<TDataType>::host_type

export

Convenience alias for accessing host type mapping.

Provides a more concise way to access the host type for a given abstract tensor data type, following modern C++ alias template patterns.

Template Parameters

TDataType

Abstract tensor data type

Example usage:

using HostType = host_type_t<TensorDataType::FP16>;  // -> float
using IntHostType = host_type_t<TensorDataType::INT8>;  // -> std::int8_t

Variable Documentation

is_host_float_type

template<TensorDataType TDataType>

bool Mila::Dnn::is_host_float_type = std::is_floating_point_v<host_type_t<TDataType>>

constexprexport

Checks if a TensorDataType maps to a floating-point host type.

Compile-time utility to determine if the host representation of an abstract tensor data type is a floating-point type.

Template Parameters

TDataType

Abstract tensor data type to check

is_host_integer_type

template<TensorDataType TDataType>

bool Mila::Dnn::is_host_integer_type = std::is_integral_v<host_type_t<TDataType>>

constexprexport

Checks if a TensorDataType maps to an integer host type.

Compile-time utility to determine if the host representation of an abstract tensor data type is an integer type.

Template Parameters

TDataType

Abstract tensor data type to check