Safe Haskell	Safe-Inferred
Language	Haskell2010

Vulkan.Extensions.VK_KHR_shader_float_controls

Description

Name

VK_KHR_shader_float_controls - device extension

VK_KHR_shader_float_controls

Name String: VK_KHR_shader_float_controls

Extension Type

Device extension

Registered Extension Number

198

Revision

4

Ratification Status

Ratified

Extension and Version Dependencies

VK_KHR_get_physical_device_properties2

Deprecation State

Promoted to Vulkan 1.2

Contact

Alexander Galazin @alegal-arm%0A*Here describe the issue or question you have about the VK_KHR_shader_float_controls extension*

Other Extension Metadata

Last Modified Date: 2018-09-11

Interactions and External Dependencies

Promoted to Vulkan 1.2 Core
This extension requires SPV_KHR_float_controls

IP Status

No known IP claims.

Contributors

Alexander Galazin, Arm
Jan-Harald Fredriksen, Arm
Jeff Bolz, NVIDIA
Graeme Leese, Broadcom
Daniel Rakos, AMD

Description

The VK_KHR_shader_float_controls extension enables efficient use of floating-point computations through the ability to query and override the implementation’s default behavior for rounding modes, denormals, signed zero, and infinity.

Promotion to Vulkan 1.2

All functionality in this extension is included in core Vulkan 1.2, with the KHR suffix omitted. The original type, enum and command names are still available as aliases of the core functionality.

New Structures

Extending PhysicalDeviceProperties2:
- PhysicalDeviceFloatControlsPropertiesKHR

New Enums

ShaderFloatControlsIndependenceKHR

New Enum Constants

KHR_SHADER_FLOAT_CONTROLS_EXTENSION_NAME

New SPIR-V Capabilities

DenormPreserve

Issues

1) Which instructions must flush denorms?

RESOLVED: Only floating-point conversion, floating-point arithmetic, floating-point relational (except OpIsNaN, OpIsInf), and floating-point GLSL.std.450 extended instructions must flush denormals.

2) What is the denorm behavior for intermediate results?

RESOLVED: When a SPIR-V instruction is implemented as a sequence of other instructions:

in the DenormFlushToZero execution mode, the intermediate instructions may flush denormals, the final result of the sequence must not be denormal.
in the DenormPreserve execution mode, denormals must be preserved throughout the whole sequence.

3) Do denorm and rounding mode controls apply to OpSpecConstantOp?

RESOLVED: Yes, except when the opcode is OpQuantizeToF16.

4) The SPIR-V specification says that OpConvertFToU and OpConvertFToS unconditionally round towards zero. Do the rounding mode controls specified through the execution modes apply to them?

RESOLVED: No, these instructions unconditionally round towards zero.

5) Do any of the “Pack” GLSL.std.450 instructions count as conversion instructions and have the rounding mode applied?

RESOLVED: No, only instructions listed in “section 3.32.11. Conversion Instructions” of the SPIR-V specification count as conversion instructions.

6) When using inf/nan-ignore mode, what is expected of OpIsNan and OpIsInf?

RESOLVED: These instructions must always accurately detect inf/nan if it is passed to them.

Version 4 API Incompatibility

The original versions of VK_KHR_shader_float_controls shipped with booleans named “separateDenormSettings” and “separateRoundingModeSettings”, which at first glance could have indicated “they can all be set independently, or not”. However the spec language as written indicated that the 32-bit value could always be set independently, and only the 16- and 64-bit controls needed to be the same if these values were FALSE.

As a result of this slight disparity, and lack of test coverage for this facet of the extension, we ended up with two different behaviors in the wild, where some implementations worked as written, and others worked based on the naming. As these are hard limits in hardware with reasons for exposure as written, it was not possible to standardise on a single way to make this work within the existing API.

No known users of this part of the extension exist in the wild, and as such the Vulkan WG took the unusual step of retroactively changing the once boolean value into a tri-state enum, breaking source compatibility. This was however done in such a way as to retain ABI compatibility, in case any code using this did exist; with the numerical values 0 and 1 retaining their original specified meaning, and a new value signifying the additional “all need to be set together” state. If any applications exist today, compiled binaries will continue to work as written in most cases, but will need changes before the code can be recompiled.