Name Strings

SPV_KHR_float_controls

Contact

See Issues list in the Khronos SPIRV-Registry repository: https://github.com/KhronosGroup/SPIRV-Registry

Contributors

  • Alexander Galazin, Arm

  • Neil Henning, AMD

  • Jeff Bolz, NVIDIA

  • John Kessenich, Google

  • Ruihao Zhang, Qualcomm

Notice

Copyright (c) 2016-2018 The Khronos Group Inc. Copyright terms at http://www.khronos.org/registry/speccopyright.html

Status

  • Complete

  • Approved by the SPIR Working Group: 2018-09-05

  • Ratified by Khronos: 2018-10-26

Version

Last Modified Date

2018-09-06

Revision

1

Dependencies

This extension is written against the SPIR-V Specification, Version 1.3 Revision 1.

This extension requires SPIR-V 1.0.

Overview

This extension provides new execution modes to control floating-point computations by overriding an implementation’s default behavior for rounding modes, denormals, signed zero, and infinities.

Extension Name

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

OpExtension "SPV_KHR_float_controls"

Modifications to the SPIR-V Specification, Version 1.3

Modify Section 3.6, Execution Mode, adding the following rows to the Execution Mode table:
Execution mode Enabling Capabilities Extra Operands

4459

DenormPreserve
Any denormalized value input into a shader or potentially generated by any instruction in a shader must be preserved. Denormalized values obtained via unpacking an integer into a vector of values with smaller bit width and interpreting those values as floating-point numbers must be preserved. Only affects instructions operating on a scalar or a composite type derived from a scalar floating-point type defined by the extra operand of OpExecutionMode.

DenormPreserve

Literal Number
Scalar Floating-Point Type Bit Width

4460

DenormFlushToZero
Any denormalized value input into a shader or potentially generated by any instruction in a shader must be flushed to zero. Denormalized values obtained via unpacking an integer into a vector of values with smaller bit width and interpreting those values as floating-point numbers must be flushed. Only affects instructions operating on a scalar or a composite type derived from a scalar floating-point type defined by the extra operand of OpExecutionMode.

DenormFlushToZero

Literal Number
Scalar Floating-Point Type Bit Width

4461

SignedZeroInfNanPreserve
The implementation must not perform optimizations on floating-point instructions that do not preserve sign of a zero, or assume that operands and results are not NaNs or infinities. Bit patterns for NaNs might not be preserved. Only affects instructions operating on a scalar or a composite type derived from a scalar floating-point type defined by the extra operand of OpExecutionMode.

SignedZeroInfNanPreserve

Literal Number
Scalar Floating-Point Type Bit Width

4462

RoundingModeRTE
The default rounding mode for floating-point arithmetic and conversions instructions must be round-to-nearest-even. Only affects instructions operating on a scalar or a composite type derived from a scalar floating-point type defined by the extra operand of OpExecutionMode. If an instruction is decorated with FPRoundingMode or defines a rounding mode in its description, that rounding mode is applied and RoundingModeRTE is ignored.

RoundingModeRTE

Literal Number
Scalar Floating-Point Type Bit Width

4463

RoundingModeRTZ
The default rounding mode for floating-point arithmetic and conversions instructions must be round-towards-zero. Only affects instructions operating on a scalar or a composite type derived from a scalar floating-point type defined by the extra operand of OpExecutionMode. If an instruction is decorated with FPRoundingMode or defines a rounding mode in its description, that rounding mode is applied and RoundingModeRTZ is ignored.

RoundingModeRTZ

Literal Number
Scalar Floating-Point Type Bit Width
Modify Section 3.31, Capability, adding the following rows to the Capability table:
Capability Implicitly Declares Enabled by Extension

4464

DenormPreserve
Uses the DenormPreserve Execution Mode

SPV_KHR_float_controls

4465

DenormFlushToZero
Uses the DenormFlushToZero Execution Mode

SPV_KHR_float_controls

4466

SignedZeroInfNanPreserve
Uses the SignedZeroInfNanPreserve Execution Mode

SPV_KHR_float_controls

4467

RoundingModeRTE
Uses the RoundingModeRTE Execution Mode

SPV_KHR_float_controls

4468

RoundingModeRTZ
Uses the RoundingModeRTZ Execution Mode

SPV_KHR_float_controls
Modify Section 2.16.1. Universal Validation Rules, adding the following sub-items to the "Entry point and execution model" item:

  • An OpEntryPoint can set at most one of the DenormFlushToZero, DenormPreserve Execution Modes for the given Floating-Point Type Bit Width.

  • An OpEntryPoint can set at most one of the RoundingModeRTE, RoundingModeRTZ Execution Modes for the given Floating-Point Type Bit Width.

Interactions with the FP Rounding Mode

The RoundingModeRTE and the RoundingModeRTZ are applied globally per entry point. The FPRoundingMode is applied on per-instruction basis, which takes precedence over the RoundingModeRTE and the RoundingModeRTZ execution modes.

Interactions with the FP Fast Math Mode

The FPFastMathMode decoration is limited to the Kernel capability. There are no interactions with this extension.

Default execution modes

Default execution modes are expected to be documented by client APIs.

Issues

  1. Do we need DenormUnspecified, SignedZeroInfNanIgnore, RoundingModeUnspecified?

    RESOLVED: No. These are assumed to be the default execution modes. This must be explicitly stated in client API execution environment specifications.

  2. Why not reuse FPRoundingMode?

    RESOLVED: We would like to have an ability to specify the rounding mode on the global scope and also for a larger set of instructions than the FPRoundingMode decoration allows.

Revision History

Rev Date Author Changes

1

2018-09-06

Alexander Galazin

Initial revision