SPV_INTEL_arbitrary_precision_fixed

Name Strings

SPV_INTEL_arbitrary_precision_fixed_point

Contact

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

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

Contributors

Ajaykumar Kannan, Intel
Shuo Niu, Intel
Daniel Zhang, Intel

Notice

Status

Final draft

Version

Last Modified Date

2023-03-21

Revision

Dependencies

This extension is written against the SPIR-V Specification Version 1.6, Revision 2.

While this extension does not require SPV_INTEL_arbitrary_precision_integers, the new operators it adds are significantly more useful when that extension is supported as the combination of the two extensions allows for more freedom in the width of arbitrary precision data types that can be represented (see the definition of the W parameter below).

This extension is built on Mentor Graphics ac_datatypes spec v3.9.2.

Overview

This extension introduces operations for arbitrary precision fixed point numbers called ac_fixed. The ac_fixed datatype is an industry standard for fixed point numbers and is published by Mentor Graphics at hlslibs.org. This datatype and its corresponding operations can be useful on targets that can take advantage of narrower representation such as FPGAs.

Data Representation

The ac_fixed datatype will be represented in SPIR-V as a pseudo type using OpTypeInt. It is characterized by three parameters: W, I, and S.

W is the total width of the datatype (including a sign bit, if required) and is encoded in the width of the OpTypeInt.
I determines the position of the decimal point.
- When I >= 0, I bits starting from the MSB of W are before the decimal point. If I is greater than W, then additional 0 bits are implied after the bits of W and before the decimal point.
- When I < 0, -I 0 bits are implied after the decimal point and before the MSB of W.
S determines if this is a signed or an unsigned number. Note that the support for signedness in OpTypeInt is not leveraged here. If the ac_fixed is signed, then the MSB (most significant bit) will contain the sign bit.

The datatype itself does not contain any information regarding I and S. Each operation will contain information about the input and result datatypes (including W, S, and I) where W is implicit from the size of the OpTypeInt.

A variable of ac_fixed type can contain both an integer component and a fractional component depending on the value of I. Based on its value, the number of bits allocated for the integer and fractional portions will change. Note that it is also possible that one of the two portions may have no bits.

Extension Name

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

OpExtension "SPV_INTEL_arbitrary_precision_fixed_point"

New Capabilities

This extension introduces a new capability:

ArbitraryPrecisionFixedPointINTEL

New Instructions

Instructions added under the ArbitraryPrecisionFixedPointINTEL capability:

OpFixedSqrtINTEL
OpFixedRecipINTEL
OpFixedRsqrtINTEL
OpFixedSinINTEL
OpFixedCosINTEL
OpFixedSinCosINTEL
OpFixedSinPiINTEL
OpFixedCosPiINTEL
OpFixedSinCosPiINTEL
OpFixedLogINTEL
OpFixedExpINTEL

Token Number Assignments

ArbitraryPrecisionFixedPointINTEL	5922
OpFixedSqrtINTEL	5923
OpFixedRecipINTEL	5924
OpFixedRsqrtINTEL	5925
OpFixedSinINTEL	5926
OpFixedCosINTEL	5927
OpFixedSinCosINTEL	5928
OpFixedSinPiINTEL	5929
OpFixedCosPiINTEL	5930
OpFixedSinCosPiINTEL	5931
OpFixedLogINTEL	5932
OpFixedExpINTEL	5933

Modifications to the SPIR-V Specification Version 1.6

After Section 3.16, add a new section "3.16b Quantization Modes" as follows:

Quantization Modes

Value	Mode	Behavior	Enabling Capabilities
0	TRN_INTEL	Truncate towards -Inf	ArbitraryPrecisionFixedPointINTEL
1	TRN_ZERO_INTEL	Truncate towards 0	ArbitraryPrecisionFixedPointINTEL
2	RND_INTEL	Round towards +Inf	ArbitraryPrecisionFixedPointINTEL
3	RND_ZERO_INTEL	Round towards 0	ArbitraryPrecisionFixedPointINTEL
4	RND_INF_INTEL	Round positive values toward +Inf and negative values toward -Inf	ArbitraryPrecisionFixedPointINTEL
5	RND_MIN_INF_INTEL	Round towards -Inf	ArbitraryPrecisionFixedPointINTEL
6	RND_CONV_INTEL	Round towards even	ArbitraryPrecisionFixedPointINTEL
7	RND_CONV_ODD_INTEL	Round towards odd	ArbitraryPrecisionFixedPointINTEL

Value

Mode

Behavior

Enabling Capabilities

TRN_INTEL

Truncate towards -Inf

ArbitraryPrecisionFixedPointINTEL

TRN_ZERO_INTEL

Truncate towards 0

ArbitraryPrecisionFixedPointINTEL

RND_INTEL

Round towards +Inf

ArbitraryPrecisionFixedPointINTEL

RND_ZERO_INTEL

Round towards 0

ArbitraryPrecisionFixedPointINTEL

RND_INF_INTEL

Round positive values toward +Inf and negative values toward -Inf

ArbitraryPrecisionFixedPointINTEL

RND_MIN_INF_INTEL

Round towards -Inf

ArbitraryPrecisionFixedPointINTEL

RND_CONV_INTEL

Round towards even

ArbitraryPrecisionFixedPointINTEL

RND_CONV_ODD_INTEL

Round towards odd

ArbitraryPrecisionFixedPointINTEL

After Section 3.16, add a new section "3.16c Overflow Modes" as follows:

Overflow Modes

Value	Mode	Behavior	Enabling Capabilities
0	WRAP_INTEL	Drop the bits to the left of the MSB	ArbitraryPrecisionFixedPointINTEL
1	SAT_INTEL	Saturate to the closest of MIN or MAX	ArbitraryPrecisionFixedPointINTEL
2	SAT_ZERO_INTEL	Set to 0 on overflow	ArbitraryPrecisionFixedPointINTEL
3	SAT_SYM_INTEL	For unsigned, treat as SAT_INTEL. For signed: a positive overflow will saturate at the maximum positive value, whereas a negative overflow will saturate to the negation of the maximum positive value, as opposed to the most negative value.	ArbitraryPrecisionFixedPointINTEL

Value

Mode

Behavior

Enabling Capabilities

WRAP_INTEL

Drop the bits to the left of the MSB

ArbitraryPrecisionFixedPointINTEL

SAT_INTEL

Saturate to the closest of MIN or MAX

ArbitraryPrecisionFixedPointINTEL

SAT_ZERO_INTEL

Set to 0 on overflow

ArbitraryPrecisionFixedPointINTEL

SAT_SYM_INTEL

For unsigned, treat as SAT_INTEL.

For signed: a positive overflow will saturate at the maximum positive value, whereas a negative overflow will saturate to the negation of the maximum positive value, as opposed to the most negative value.

ArbitraryPrecisionFixedPointINTEL

After Section 3.16, add a new section "3.16d Signedness Modes" as follows:

Signedness Modes

Value	Mode	Behavior	Enabling Capabilities
0	UNSIGNED_INTEL	Input and result types are unsigned	ArbitraryPrecisionFixedPointINTEL
1	SIGNED_INTEL	Input and result types are signed	ArbitraryPrecisionFixedPointINTEL

Value

Mode

Behavior

Enabling Capabilities

UNSIGNED_INTEL

Input and result types are unsigned

ArbitraryPrecisionFixedPointINTEL

SIGNED_INTEL

Input and result types are signed

ArbitraryPrecisionFixedPointINTEL

Capability

Modify Section 3.31, Capability, adding a row to the Capability table:

Capability	Implicitly Declares
5922	ArbitraryPrecisionFixedPointINTEL Enables arbitrary precision fixed-point math instructions.

Capability

Implicitly Declares

5922

ArbitraryPrecisionFixedPointINTEL

Enables arbitrary precision fixed-point math instructions.

Instructions

In Section 3.32.13, Arithmetic Instructions, add the following instructions:

OpFixedSqrtINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the square root of the value is returned in Result. The behavior of this function is undefined for input values < 0.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5923

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedRecipINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the reciprocal (1/Input) of the value is returned in Result.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5924

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedRsqrtINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the reciprocal square root (1/sqrt(Input)) of the value is returned in Result. The behavior of this function is undefined for input values < 0.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5925

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedSinINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the sine of the value is returned in Result. Note that the angles are measured in radians.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5926

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedCosINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the cosine of the value is returned in Result. Note that the angles are measured in radians.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5927

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedSinCosINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and both the sine and cosine of the value are returned in Result. Note that the angles are measured in radians.

Result Type must be a two-component vector of OpTypeInt. The first component of the Result contains the sine of the Input and is an arbitrary fixed point number. The second component of the Result contains the cosine of the Input and is also an arbitrary fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of each component of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5928

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedSinPiINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the sine of pi * Input is returned in Result. Note that the angles are measured in radians.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5929

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedCosPiINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the cosine of pi * Input is returned in Result. Note that the angles are measured in radians.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5930

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedSinCosPiINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and both the sine and cosine of pi * Input are returned in Result. Note that the angles are measured in radians.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of each component of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5931

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedLogINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the log of the value is returned in Result.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5932

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

OpFixedExpINTEL

An OpTypeInt representing an arbitrary precision fixed point number (ac_fixed) is passed in as the Input and the exp of the value is returned in Result.

Result Type must be OpTypeInt.

Result is the <id> of the operation’s result, which is an arbitrary precision fixed point number.

S is chosen from Table 3.16d that indicates the Signedness Mode of the input and output types.

I is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the input type.

rI is a signed 32-bit integer that refers to the location of the fixed-point relative to the MSB of the result type.

Q is a QuantizationMode enum chosen from Table 3.16b that indicates the Quantization Mode of this operation.

O is an OverflowMode enum chosen from Table 3.16c that indicates the Overflow Mode of this operation.

Capability: ArbitraryPrecisionFixedPointINTEL

5933

<id> Result Type

Result <id>

Input <id>

Signedness S

Literal I

Literal rI

QuatntizationMode Q

OverflowMode O

Validation Rules

None.

Issues

None.

Revision History

Rev	Date	Author	Changes
1	2023-03-21	Ajaykumar Kannan	Initial Public Release

Rev

Date

Author

Changes

2023-03-21

Ajaykumar Kannan

Initial Public Release