Vulkan.Extensions.VK_NV_cooperative_vector

vkGetPhysicalDeviceCooperativeVectorPropertiesNV - Returns properties describing what cooperative vector types are supported

Description

If pProperties is NULL, then the number of cooperative vector properties available is returned in pPropertyCount. Otherwise, pPropertyCount must point to a variable set by the user to the number of elements in the pProperties array, and on return the variable is overwritten with the number of structures actually written to pProperties. If pPropertyCount is less than the number of cooperative vector properties available, at most pPropertyCount structures will be written, and INCOMPLETE will be returned instead of SUCCESS, to indicate that not all the available cooperative vector properties were returned.

Valid Usage (Implicit)

• physicalDevice must be a valid PhysicalDevice handle

• pPropertyCount must be a valid pointer to a uint32_t value

• If the value referenced by pPropertyCount is not 0, and pProperties is not NULL, pProperties must be a valid pointer to an array of pPropertyCount CooperativeVectorPropertiesNV structures

Return Codes

Success

• INCOMPLETE

• SUCCESS

Failure

• ERROR_OUT_OF_DEVICE_MEMORY

• ERROR_OUT_OF_HOST_MEMORY

• ERROR_UNKNOWN

• ERROR_VALIDATION_FAILED

See Also

VK_NV_cooperative_vector, CooperativeVectorPropertiesNV, PhysicalDevice

vkConvertCooperativeVectorMatrixNV - Convert a cooperative vector matrix from one layout and type to another

Description

If pInfo->dstData is NULL, then the number of bytes required to store the converted matrix is returned in pDstSize. Otherwise, pInfo->pDstSize must point to a variable set by the user to the number of bytes in pInfo->dstData, and on return the variable is overwritten with the number of bytes actually written to pInfo->dstData. pInfo->srcData can be NULL when pInfo->dstData is NULL. If pInfo->pDstSize is less than the number of bytes required to store the converted matrix, no bytes will be written, and INCOMPLETE will be returned instead of SUCCESS, to indicate that not enough space was provided.

Valid Usage

• If pInfo->srcData.hostAddress is NULL, then pInfo->dstData.hostAddress must be NULL

• If pInfo->srcData.hostAddress is not NULL, then pInfo->srcSize must be large enough to contain the source matrix, based either on the standard matrix layout or based on the size filled out by this command

• If pInfo->dstData.hostAddress is not NULL, then the value pointed to by pInfo->pDstSize must be large enough to contain the destination matrix, based either on the standard matrix layout or based on the size filled out by this command

• If pInfo->dstData.hostAddress is not NULL, the source and destination memory ranges must not overlap

Valid Usage (Implicit)

• device must be a valid Device handle

• pInfo must be a valid pointer to a valid ConvertCooperativeVectorMatrixInfoNV structure

Return Codes

Success

• INCOMPLETE

• SUCCESS

Failure

• ERROR_OUT_OF_HOST_MEMORY

• ERROR_UNKNOWN

• ERROR_VALIDATION_FAILED

See Also

VK_NV_cooperative_vector, ConvertCooperativeVectorMatrixInfoNV, Device

vkCmdConvertCooperativeVectorMatrixNV - Convert a cooperative vector matrix from one layout and type to another

Description

This command does the same conversions as convertCooperativeVectorMatrixNV, but executes on the device. One conversion is performed for each of the infoCount elements of pInfos.

This command’s execution is synchronized using PIPELINE_STAGE_2_CONVERT_COOPERATIVE_VECTOR_MATRIX_BIT_NV.

Valid Usage

• For each element of pInfo, srcData.deviceAddress must be a valid DeviceAddress

• For each element of pInfo, dstData.deviceAddress must be a valid DeviceAddress

• For each element of pInfo, srcData.deviceAddress must be 64 byte aligned

• For each element of pInfo, dstData.deviceAddress must be 64 byte aligned

• For each element of pInfo, srcSize must be large enough to contain the source matrix, based either on the standard matrix layout or based on the size filled out by convertCooperativeVectorMatrixNV

• For each element of pInfo, the value pointed to by pDstSize must be large enough to contain the destination matrix, based either on the standard matrix layout or based on the size filled out by convertCooperativeVectorMatrixNV

• Memory accessed by the sources and destinations of all of the conversions must not overlap

Valid Usage (Implicit)

• commandBuffer must be a valid CommandBuffer handle

• pInfos must be a valid pointer to an array of infoCount valid ConvertCooperativeVectorMatrixInfoNV structures

• commandBuffer must be in the recording state

• The CommandPool that commandBuffer was allocated from must support QUEUE_COMPUTE_BIT, or QUEUE_GRAPHICS_BIT operations

• This command must only be called outside of a render pass instance

• This command must not be called between suspended render pass instances

• This command must only be called outside of a video coding scope

• infoCount must be greater than 0

Host Synchronization

• Host access to commandBuffer must be externally synchronized

• Host access to the CommandPool that commandBuffer was allocated from must be externally synchronized

Command Properties

'

Conditional Rendering

vkCmdConvertCooperativeVectorMatrixNV is not affected by conditional rendering

See Also

VK_NV_cooperative_vector, CommandBuffer, ConvertCooperativeVectorMatrixInfoNV

VkPhysicalDeviceCooperativeVectorFeaturesNV - Structure describing cooperative vector features that can be supported by an implementation

Members

This structure describes the following features:

Description

If the PhysicalDeviceCooperativeVectorFeaturesNV structure is included in the pNext chain of the PhysicalDeviceFeatures2 structure passed to getPhysicalDeviceFeatures2, it is filled in to indicate whether each corresponding feature is supported. If the application wishes to use a Device with any features described by PhysicalDeviceCooperativeVectorFeaturesNV, it must add an instance of the structure, with the desired feature members set to TRUE, to the pNext chain of DeviceCreateInfo when creating the Device.

Structure Chaining

Extends the structures

• DeviceCreateInfo

• PhysicalDeviceFeatures2

See Also

VK_NV_cooperative_vector, Bool32, StructureType

VkCooperativeVectorPropertiesNV - Structure specifying cooperative vector properties

Description

COMPONENT_TYPE_SINT8_PACKED_NV and COMPONENT_TYPE_UINT8_PACKED_NV must not be used for members other than inputInterpretation.

The following combinations must be supported (each row is a required combination):

Valid Usage (Implicit)

See Also

VK_NV_cooperative_vector, Bool32, ComponentTypeKHR, StructureType, getPhysicalDeviceCooperativeVectorPropertiesNV

VkPhysicalDeviceCooperativeVectorPropertiesNV - Structure describing cooperative vector properties supported by an implementation

Description

If the PhysicalDeviceCooperativeVectorPropertiesNV structure is included in the pNext chain of the PhysicalDeviceProperties2 structure passed to getPhysicalDeviceProperties2, it is filled in with each corresponding implementation-dependent property.

Structure Chaining

Extends the structure

• PhysicalDeviceProperties2

See Also

VK_NV_cooperative_vector, Bool32, ShaderStageFlags, StructureType

VkConvertCooperativeVectorMatrixInfoNV - Structure specifying a request to convert the layout and type of a cooperative vector matrix

Description

When called from cmdConvertCooperativeVectorMatrixNV, the deviceAddress members of srcData and dstData are used. When called from convertCooperativeVectorMatrixNV, the hostAddress members of srcData and dstData are used.

For each of the source and destination matrix, if the layout is not either COOPERATIVE_VECTOR_MATRIX_LAYOUT_ROW_MAJOR_NV or COOPERATIVE_VECTOR_MATRIX_LAYOUT_COLUMN_MAJOR_NV, then the corresponding stride parameter is ignored.

The size of the destination is only a function of the destination layout information, and does not depend on the source layout information.

Conversion can be used to convert between COMPONENT_TYPE_FLOAT32_KHR or COMPONENT_TYPE_FLOAT16_KHR and any supported lower-precision floating-point type. In this case, the conversion uses round-to-nearest-even rounding.

Valid Usage

• If srcLayout is row-major or column-major, then srcStride must be greater than the length of a row/column, and a multiple of the element size

• If dstLayout is row-major or column-major, then dstStride must be greater than the length of a row/column, and a multiple of the element size

• If srcComponentType is not a supported CooperativeVectorPropertiesNV::matrixInterpretation value as reported by getPhysicalDeviceCooperativeVectorPropertiesNV, then srcComponentType must be COMPONENT_TYPE_FLOAT32_KHR

• If dstComponentType is not a supported CooperativeVectorPropertiesNV::matrixInterpretation value as reported by getPhysicalDeviceCooperativeVectorPropertiesNV, then dstComponentType must be COMPONENT_TYPE_FLOAT32_KHR

• If srcComponentType and dstComponentType are not equal, then one must be COMPONENT_TYPE_FLOAT32_KHR or COMPONENT_TYPE_FLOAT16_KHR and the other must be a lower-precision floating-point type

• If dstComponentType is COMPONENT_TYPE_FLOAT_E4M3_NV or COMPONENT_TYPE_FLOAT_E5M2_NV, then dstLayout must be COOPERATIVE_VECTOR_MATRIX_LAYOUT_INFERENCING_OPTIMAL_NV or COOPERATIVE_VECTOR_MATRIX_LAYOUT_TRAINING_OPTIMAL_NV

Valid Usage (Implicit)

• sType must be STRUCTURE_TYPE_CONVERT_COOPERATIVE_VECTOR_MATRIX_INFO_NV

• pNext must be NULL

• srcData must be a valid DeviceOrHostAddressConstKHR union

• pDstSize must be a valid pointer to a size_t value

• dstData must be a valid DeviceOrHostAddressKHR union

• srcComponentType must be a valid ComponentTypeKHR value

• dstComponentType must be a valid ComponentTypeKHR value

• srcLayout must be a valid CooperativeVectorMatrixLayoutNV value

• dstLayout must be a valid CooperativeVectorMatrixLayoutNV value

See Also

VK_NV_cooperative_vector, ComponentTypeKHR, CooperativeVectorMatrixLayoutNV, DeviceOrHostAddressConstKHR, DeviceOrHostAddressKHR, StructureType, cmdConvertCooperativeVectorMatrixNV, convertCooperativeVectorMatrixNV

VkComponentTypeKHR - Specify SPIR-V cooperative matrix component type

Description

• COMPONENT_TYPE_FLOAT16_KHR corresponds to SPIR-V OpTypeFloat 16.

• COMPONENT_TYPE_FLOAT32_KHR corresponds to SPIR-V OpTypeFloat 32.

• COMPONENT_TYPE_FLOAT64_KHR corresponds to SPIR-V OpTypeFloat 64.

• COMPONENT_TYPE_SINT8_KHR corresponds to SPIR-V OpTypeInt 8 0/1.

• COMPONENT_TYPE_SINT16_KHR corresponds to SPIR-V OpTypeInt 16 0/1.

• COMPONENT_TYPE_SINT32_KHR corresponds to SPIR-V OpTypeInt 32 0/1.

• COMPONENT_TYPE_SINT64_KHR corresponds to SPIR-V OpTypeInt 64 0/1.

• COMPONENT_TYPE_UINT8_KHR corresponds to SPIR-V OpTypeInt 8 0/1.

• COMPONENT_TYPE_UINT16_KHR corresponds to SPIR-V OpTypeInt 16 0/1.

• COMPONENT_TYPE_UINT32_KHR corresponds to SPIR-V OpTypeInt 32 0/1.

• COMPONENT_TYPE_UINT64_KHR corresponds to SPIR-V OpTypeInt 64 0/1.

• COMPONENT_TYPE_BFLOAT16_KHR corresponds to SPIR-V OpTypeFloat 16 BFloat16KHR.

• COMPONENT_TYPE_SINT8_PACKED_NV corresponds to four 8-bit signed integers packed in a 32-bit unsigned integer.

• COMPONENT_TYPE_UINT8_PACKED_NV corresponds to four 8-bit unsigned integers packed in a 32-bit unsigned integer.

• COMPONENT_TYPE_FLOAT_E4M3_NV corresponds to a floating-point type with a sign bit in the most significant bit, followed by four exponent bits, followed by three mantissa bits.

• COMPONENT_TYPE_FLOAT_E5M2_NV corresponds to a floating-point type with a sign bit in the most significant bit, followed by five exponent bits, followed by two mantissa bits.

• COMPONENT_TYPE_FLOAT8_E4M3_EXT corresponds to SPIR-V OpTypeFloat 8 Float8E4M3EXT.

• COMPONENT_TYPE_FLOAT8_E5M2_EXT corresponds to SPIR-V OpTypeFloat 8 Float8E5M2EXT.

See Also

VK_KHR_cooperative_matrix, VK_NV_cooperative_matrix, VK_NV_cooperative_vector, ConvertCooperativeVectorMatrixInfoNV, CooperativeMatrixFlexibleDimensionsPropertiesNV, CooperativeMatrixPropertiesKHR, CooperativeMatrixPropertiesNV, CooperativeVectorPropertiesNV

VkCooperativeVectorMatrixLayoutNV - Specify cooperative vector matrix layout

Description

• COOPERATIVE_VECTOR_MATRIX_LAYOUT_ROW_MAJOR_NV corresponds to SPIR-V RowMajorNV layout.

• COOPERATIVE_VECTOR_MATRIX_LAYOUT_COLUMN_MAJOR_NV corresponds to SPIR-V ColumnMajorNV layout.

• COOPERATIVE_VECTOR_MATRIX_LAYOUT_INFERENCING_OPTIMAL_NV corresponds to SPIR-V InferencingOptimalNV layout.

• COOPERATIVE_VECTOR_MATRIX_LAYOUT_TRAINING_OPTIMAL_NV corresponds to SPIR-V TrainingOptimalNV layout.

All enum values match the corresponding SPIR-V value.

Row-major layout has elements of each row stored consecutively in memory, with a controllable stride from the start of one row to the start of the next row. Column-major layout has elements of each column stored consecutively in memory, with a controllable stride from the start of one column to the start of the next column. Inferencing-optimal and Training-optimal layouts are implementation-dependent, and the application can convert a matrix to those layouts using convertCooperativeVectorMatrixNV or cmdConvertCooperativeVectorMatrixNV. Training-optimal layout with COMPONENT_TYPE_FLOAT16_KHR or COMPONENT_TYPE_FLOAT32_KHR type has the additional guarantee that the application can reinterpret the data as an array of elements and perform element-wise operations on the data, and finite values in any padding elements do not affect the result of a matrix-vector multiply (inf/NaN values may still cause NaN values in the result).

See Also

VK_NV_cooperative_vector, ConvertCooperativeVectorMatrixInfoNV