SPV_KHR_16bit

Name Strings

SPV_KHR_16bit_storage

Contact

To report problems with this extension, please open a new issue at:

https://github.com/KhronosGroup/SPIRV-Registry

Contributors

Alexander Galazin, ARM
Jan-Harald Fredriksen, ARM
Joerg Wagner, ARM
Neil Henning, Codeplay
Jeff Bolz, Nvidia
David Neto, Google
John Kessenich, Google

Notice

Status

Complete
Approved by the SPIR Working Group: 2017-01-11
Approved by the Khronos Board of Promoters: 2017-02-24

Version

Last Modified Date

2018-06-11

Revision

Dependencies

This extension is written against the SPIR-V Specification, Version 1.1 Revision 1.

This extension requires SPIR-V 1.0.

This extension interacts with SPV_KHR_storage_buffer_storage_class.

Overview

This extension adds new StorageUniformBufferBlock16, StorageUniform16 capabilities that allow to access 16-bit data in objects in Uniform storage class with BufferBlock and Block decorations.

If the SPV_KHR_storage_buffer_storage_class extension is supported, it also allows use of the StorageBuffer16BitAccess and the UniformAndStorageBuffer16BitAccess capabilities, providing the same functionality as the StorageUniformBufferBlock16 and the StorageUniform16 capabilities.

It also adds new StoragePushConstant16 capability that allows access to 16-bit data in objects in PushConstant storage class.

Finally, this extension adds new StorageInputOutput16 capability that allows to access to 16-bit data in objects in Input and Output storage classes.

To facilitate stores of 32-bit data to 16-bit storage this extensions enables use of the FP Rounding Mode decoration in graphics shaders.

Extension Name

To use this extension within a SPIR-V module, the following OpExtension must be present in the module:

OpExtension "SPV_KHR_16bit_storage"

New Capabilities

This extension introduces new capabilities:

StorageUniformBufferBlock16
StorageUniform16
StoragePushConstant16
StorageInputOutput16

If the SPV_KHR_storage_buffer_storage_class extension is supported, the following capabilities are also supported:

StorageBuffer16BitAccess
UniformAndStorageBuffer16BitAccess

New Builtins

None.

New Instructions

None.

Token Number Assignments

StorageUniformBufferBlock16	4433
StorageBuffer16BitAccess	4433
StorageUniform16	4434
UniformAndStorageBuffer16BitAccess	4434
StoragePushConstant16	4435
StorageInputOutput16	4436

Modifications to the SPIR-V Specification, Version 1.1

Modify Section 3.16, FP Rounding Mode, amend the Enabling Capabilities column to say:

Any of Kernel, StorageUniformBufferBlock16, StorageUniform16, StoragePushConstant16, StorageInputOutput16.

Modify Section 3.20, Decoration, amend the Enabling Capabilities column for the FPRoundingMode decoration to say:

Any of Kernel, StorageUniformBufferBlock16, StorageUniform16, StoragePushConstant16, StorageInputOutput16.

Modify Section 3.31, Capability, adding the following rows to the Capability table:

Capability		Depends On
4433	StorageUniformBufferBlock16 Allows 16-bit OpTypeFloat and OpTypeInt instructions for creating scalar, vector, and composite types that become members of a block residing in the Uniform Storage Class. A type that is or contains such a 16-bit type can be used only as an operand of an OpTypePointer instruction. The block must be decorated with BufferBlock. Other uses of 16-bit types are not enabled by this capability.
4434	StorageUniform16 Allows 16-bit OpTypeFloat and OpTypeInt instructions for creating scalar, vector, and composite types that become members of a block residing in the Uniform Storage Class. A type that is, or contains, such a 16-bit type can be used only as an operand of an OpTypePointer instruction. The block can have any supported decoration, including BufferBlock. Other uses of 16-bit types are not enabled by this capability.	StorageUniformBufferBlock16
4435	StoragePushConstant16 Allows 16-bit OpTypeFloat and OpTypeInt instructions for creating scalar, vector, and composite types that become members of a block residing in the PushConstant Storage Class. A type that is, or contains, such a 16-bit type can be used only as an operand of an OpTypePointer instruction. Other uses of 16-bit types are not enabled by this capability.
4436	StorageInputOutput16 Allows 16-bit OpTypeFloat and OpTypeInt instructions for creating scalar, vector, and composite types that become members of a block residing in the Output Storage Class. A type that is, or contains, such a 16-bit type can be used only as an operand of an OpTypePointer instruction. Other uses of 16-bit types are not enabled by this capability.

Validation Rules

If the StorageUniformBufferBlock16, StorageUniform16, StoragePushConstant16, or StorageInputOutput16 capability is declared:

An OpTypePointer pointing to a 16-bit scalar, a 16-bit vector, or a composite containing a 16-bit member can be used as the result type of an OpVariable, OpAccessChain, or OpInBoundsAccessChain.
OpLoad can only load 16-bit scalars, 16-bit vectors, and 16-bit matrices.
OpStore can only store 16-bit scalars, 16-bit vectors, and 16-bit matrices.
OpCopyObject can be used for 16-bit scalars or composites containing 16-bit members.
16-bit scalars or 16-bit vectors can be used as operands to a width-only conversion instruction to a 32-bit type (OpFConvert, OpSConvert, or OpUConvert), and can be produced as results of a width-only conversion instruction from a 32-bit type.
A structure containing a 16-bit member can be an operand to OpArrayLength.
Any other instructions not explicitly listed by the capabilities or allowed by the validations rules cannot operate on variables with 16-bit scalar, 16-bit vector, or 16-bit composite types.

A FPRoundingMode decoration can be applied only to:

a width-only conversion instruction that is used as the object argument of an OpStore storing through a pointer to a 16-bit floating-point object in Uniform, or PushConstant, or Input, or Output Storage Classes.

Interactions with SPV_KHR_storage_buffer_storage_class

If SPV_KHR_uniform_buffer_storage_class is supported,

modify the description of the StorageUniformBufferBlock16 capability, adding the following sentence to the first paragraph of the description:

The object can also be in the StorageBuffer Storage Class and have any decorations supported for this Storage Class.

modify the description of the StorageUniform16 capability, adding the following sentence to the first paragraph of the description:

The object can also be in the StorageBuffer Storage Class and have any decorations supported for this Storage Class.

Modify Section 3.31, Capability, adding the following rows to the Capability table:

Capability		Depends On
4433	StorageBuffer16BitAccess Same as StorageUniformBufferBlock16
4434	UniformAndStorageBuffer16BitAccess Same as StorageUniform16	StorageBuffer16BitAccess

Issues

Revision History

Rev	Date	Author	Changes
1	2016-11-22	Alexander Galazin	Initial revision
2	2016-11-28	Alexander Galazin	Address first round of feedback
3	2016-12-01	Alexander Galazin	Removed combined Load/Store and Convert instructions. Renamed capabilities and described them in terms of storage classes.
4	2016-12-08	David Neto	Assigned token numbers
5	2016-12-14	Alexander Galazin	Renamed the extension. Removed changes to the default rounding modes. Made StorageUniform16 dependent on StorageUniformBufferBlock16
6	2017-02-22	JohnK	Clarified that conversions for changing width can only change the width, not the fundamental type domain.
7	2017-03-15	Alexander Galazin	Clarified that FP Rounding mode can be used only if the capabilities from this extension are enabled
8	2017-03-23	Alexander Galazin	Added interactions with SPV_KHR_uniform_buffer_storage_class
9	2018-06-11	Alexander Galazin	Added clarifications for SPIR-V issue 319