Vulkan-specific traits for abstract tensor data types. More...

Public Types
template<TensorDataType TDataType>
using	native_type = decltype(get_native_type<TDataType>())
	Type alias for the native Vulkan type corresponding to a TensorDataType.

Static Public Member Functions
static constexpr std::string_view	device_type_name ()
	Gets the Vulkan device type string for debugging.
template<TensorDataType TDataType>
static consteval bool	is_device_only ()
	Checks if the specified data type requires device-only access.
template<TensorDataType TDataType, typename T>
static constexpr auto	make_native_value (const T &value)
	Creates a native type value from a compatible input value.
template<TensorDataType TDataType>
static consteval std::string_view	preferred_allocation_strategy ()
	Gets the preferred memory allocation strategy for the data type.
template<TensorDataType TDataType>
static consteval std::string_view	required_extension ()
	Gets the required Vulkan extension name for the data type.
template<TensorDataType TDataType>
static consteval bool	requires_extensions ()
	Checks if the data type requires specific Vulkan device extensions.
template<TensorDataType TDataType>
static consteval uint32_t	spirv_bit_width ()
	Gets the SPIR-V OpTypeFloat/OpTypeInt bit width for the data type.
template<TensorDataType TDataType>
static consteval bool	supports ()
	Validates Vulkan support for abstract tensor data types.
template<TensorDataType TDataType>
static constexpr float	to_float (const native_type< TDataType > &value)
	Converts native Vulkan type values to float for host operations.

Static Private Member Functions
template<TensorDataType TDataType>
static consteval auto	get_native_type ()
	Helper function for compile-time native type deduction.

Detailed Description

Vulkan-specific traits for abstract tensor data types.

Provides device-specific characteristics and native type mappings for Vulkan-supported tensor data types, enabling type-safe value initialization and explicit graphics/compute operations with fine-grained synchronization control across different GPU architectures.

Vulkan's explicit design requires careful consideration of:

Memory heap selection and allocation strategy
Command buffer recording and submission
Pipeline barriers and synchronization primitives
Device-local vs host-visible memory trade-offs

Member Typedef Documentation

◆ native_type

template<TensorDataType TDataType>

using Mila::Dnn::VulkanTensorTraits::native_type = decltype(get_native_type<TDataType>())

Type alias for the native Vulkan type corresponding to a TensorDataType.

Provides compile-time mapping from abstract TensorDataType enumeration to concrete C++ types compatible with Vulkan's SPIR-V shader requirements and memory layout specifications.

Member Function Documentation

◆ device_type_name()

constexpr std::string_view Mila::Dnn::VulkanTensorTraits::device_type_name ( )

inlinestaticconstexpr

Gets the Vulkan device type string for debugging.

Returns: String identifier for Vulkan backend

◆ get_native_type()

template<TensorDataType TDataType>

consteval auto Mila::Dnn::VulkanTensorTraits::get_native_type ( )

inlinestaticconstevalprivate

Helper function for compile-time native type deduction.

Maps abstract tensor data types to Vulkan-compatible native types, using standard C++ types that align with SPIR-V and Vulkan shader data type requirements. Handles both available and unavailable Vulkan configurations with appropriate fallback types.

◆ is_device_only()

template<TensorDataType TDataType>

consteval bool Mila::Dnn::VulkanTensorTraits::is_device_only ( )

inlinestaticconsteval

Checks if the specified data type requires device-only access.

Determines whether a tensor data type can only be accessed from Vulkan compute shaders and cannot be directly manipulated from host code without explicit staging buffer transfers.

Template Parameters

TDataType Abstract tensor data type to check

Returns: true if device-only access required, false if host-accessible

Here is the caller graph for this function:

◆ make_native_value()

template<TensorDataType TDataType, typename T>

constexpr auto Mila::Dnn::VulkanTensorTraits::make_native_value ( const T & value )

inlinestaticconstexpr

Creates a native type value from a compatible input value.

Provides safe conversion from input values to the appropriate native Vulkan-compatible type, handling cross-platform type conversions and ensuring compatibility with SPIR-V data type requirements.

Template Parameters

TDataType	Target abstract tensor data type
T	Input value type

Parameters

value Input value to convert

Returns: Native Vulkan-compatible type value suitable for compute operations

◆ preferred_allocation_strategy()

template<TensorDataType TDataType>

consteval std::string_view Mila::Dnn::VulkanTensorTraits::preferred_allocation_strategy ( )

inlinestaticconsteval

Gets the preferred memory allocation strategy for the data type.

Returns the optimal Vulkan memory allocation strategy based on data type characteristics, device capabilities, and expected usage patterns.

Template Parameters

TDataType Abstract tensor data type

Returns: Recommended allocation strategy for optimal performance

Here is the call graph for this function:

◆ required_extension()

template<TensorDataType TDataType>

consteval std::string_view Mila::Dnn::VulkanTensorTraits::required_extension ( )

inlinestaticconsteval

Gets the required Vulkan extension name for the data type.

Returns the Vulkan extension string required for the specified data type, enabling proper device capability validation and extension enablement.

Template Parameters

TDataType Abstract tensor data type requiring extensions

Returns: String view of the required extension name

◆ requires_extensions()

template<TensorDataType TDataType>

consteval bool Mila::Dnn::VulkanTensorTraits::requires_extensions ( )

inlinestaticconsteval

Checks if the data type requires specific Vulkan device extensions.

Determines whether a tensor data type requires explicit Vulkan device extensions to be enabled, affecting device selection and capability validation.

Template Parameters

TDataType Abstract tensor data type to check

Returns: true if extensions required, false for core Vulkan types

◆ spirv_bit_width()

template<TensorDataType TDataType>

consteval uint32_t Mila::Dnn::VulkanTensorTraits::spirv_bit_width ( )

inlinestaticconsteval

Gets the SPIR-V OpTypeFloat/OpTypeInt bit width for the data type.

Returns the bit width used in SPIR-V instruction generation for the specified data type, essential for compute shader compilation and validation.

Template Parameters

TDataType Abstract tensor data type

Returns: Bit width for SPIR-V type instructions

◆ supports()

template<TensorDataType TDataType>

consteval bool Mila::Dnn::VulkanTensorTraits::supports ( )

inlinestaticconsteval

Validates Vulkan support for abstract tensor data types.

Determines whether Vulkan devices and memory resources support the specified abstract tensor data type, considering Vulkan's explicit memory model and compute pipeline requirements across different GPU vendors and architectures.

Template Parameters

TDataType Abstract tensor data type to validate

Returns: true if Vulkan supports the data type, false otherwise

◆ to_float()

template<TensorDataType TDataType>

constexpr float Mila::Dnn::VulkanTensorTraits::to_float ( const native_type< TDataType > & value )

inlinestaticconstexpr

Converts native Vulkan type values to float for host operations.

Provides conversion from Vulkan native types back to float values for host-side processing, validation, and cross-platform compatibility. Handles special cases like half-precision and 64-bit types that require explicit device extension support.

Template Parameters

TDataType Source abstract tensor data type

Parameters

value Native Vulkan type value to convert

Returns: Float representation of the value

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

/__w/Mila/Mila/Mila/Src/Dnn/Compute/Devices/Rocm/Tensors/RocmTensorTraits.ixx

Public Types

Static Public Member Functions

Static Private Member Functions

Detailed Description

Member Typedef Documentation

◆ native_type

Member Function Documentation

◆ device_type_name()

◆ get_native_type()

◆ is_device_only()

◆ make_native_value()

◆ preferred_allocation_strategy()

◆ required_extension()

◆ requires_extensions()

◆ spirv_bit_width()

◆ supports()

◆ to_float()