2020Q4
Date of Issue: 21st December 2020
The Vector Function Application Binary Interface Specification for AArch64 describes the application binary interface for vector functions generated by a compiler.
This document uses the OpenMP declare
simd
directive to classify the vector functions that can be
associated to a scalar function.
The following rules apply to a compiler that implements the OpenMP
directive #pragma omp declare simd
:
- The use of a
#pragma omp declare simd
construct for a function definition enables the creation of vector versions of the function from the scalar version of the function, that can be used to process multiple instances concurrently in a single invocation in a vector context (e.g. vectorized loops). - The use of a
#pragma omp declare simd
construct for a function declaration enables the compiler to know the exact list of available vector function implementations provided by a library that is based on the OpenMP pragmas found in the function's prototype of the library headers.
This Vector Function ABI defines a set of rules that the caller and the callee functions must obey.
The Vector Function ABI also describes how to use the declare
variant
directive introduced in OpenMP 5.0 to interface user-defined
vector functions with a compiler.
- SVE
- Scalable Vector Extension
- A64
- Instruction set of the ARMv8-A architecture
- AArch64
- 64-bit execution mode of the ARMv8-A architecture
- Advanced SIMD
- SIMD and floating point instructions of the A64 instruction set
- Qn register
- Quad-word (128-bit) floating point register
- Dn register
- Double-word (64-bit) floating point register
- Sn register
- Single-word (32-bit) floating point register
- Hn register
- Half-word (16-bit) floating point register
- Bn register
- Byte (8-bit) floating point register
- Vn register
- Advanced SIMD vector register
- Zn register
- SVE vector register
- ACLE
- ARM C Language Extensions
- SVE ACLE
- ARM C Language Extensions for SVE
- Vector function
- A function processing vector data through the SIMD registers
- Leaf function
- Function at the end of a call tree
- AAPCS
- ARM Architecture Procedure Call Standard
- AAELF64
- ELF for the Arm 64-bit Architecture
- OpenMP
- Open Multi-Processing standard
- Uniform parameter
- A function parameter marked with the OpenMP uniform clause
- Linear parameter
- A function parameter marked with the OpenMP linear clause
- LP64
- Data model in which Long and Pointers are 64-bit
- ILP32
- Data model in which Integer, Long and Pointers are 32-bit
- VLA
- Vector Length Agnostic
- VLS
- Vector Length Specific
MTV(P)
P
Maps To VectorPBV(P)
P
is Passed By ValueLS(P)
- Lane Size of
P
MAP(P)
- Mapping of
P
ADVSIMD_MAP(P)
- Mapping of
P
- Advanced SIMD specific rules. SVE_MAP(P)
- Mapping of
P
- SVE specific rules. NDS(f)
- Narrowest Data Size of
f
WDS(f)
- Widest Data Size of
f
Please check Application Binary Interface for the Arm® Architecture for the latest release of this document.
Please report defects in this specification to the issue tracker page on GitHub.
This work is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by-sa/4.0/ or send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.
Grant of Patent License. Subject to the terms and conditions of this license (both the Public License and this Patent License), each Licensor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Licensed Material, where such license applies only to those patent claims licensable by such Licensor that are necessarily infringed by their contribution(s) alone or by combination of their contribution(s) with the Licensed Material to which such contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Licensed Material or a contribution incorporated within the Licensed Material constitutes direct or contributory patent infringement, then any licenses granted to You under this license for that Licensed Material shall terminate as of the date such litigation is filed.
As identified more fully in the License section, this project is licensed under CC-BY-SA-4.0 along with an additional patent license. The language in the additional patent license is largely identical to that in Apache-2.0 (specifically, Section 3 of Apache-2.0 as reflected at https://www.apache.org/licenses/LICENSE-2.0) with two exceptions.
First, several changes were made related to the defined terms so as to reflect the fact that such defined terms need to align with the terminology in CC-BY-SA-4.0 rather than Apache-2.0 (e.g., changing “Work” to “Licensed Material”).
Second, the defensive termination clause was changed such that the scope of defensive termination applies to “any licenses granted to You” (rather than “any patent licenses granted to You”). This change is intended to help maintain a healthy ecosystem by providing additional protection to the community against patent litigation claims.
Contributions to this project are licensed under an inbound=outbound model such that any such contributions are licensed by the contributor under the same terms as those in the LICENSE file.
The text of and illustrations in this document are licensed by Arm under a Creative Commons Attribution–Share Alike 4.0 International license ("CC-BY-SA-4.0”), with an additional clause on patents. The Arm trademarks featured here are registered trademarks or trademarks of Arm Limited (or its subsidiaries) in the US and/or elsewhere. All rights reserved. Please visit https://www.arm.com/company/policies/trademarks for more information about Arm’s trademarks.
Copyright (c) 2018-2020, Arm Limited and its affiliates. All rights reserved.
Contents
- 1 Preamble
- 2 About this document
- 3 Definitions
- 4 Vector function signature
- 5 User defined vector functions
- 6 Masking
- 7 Additional examples
- 8 Footnotes
The following support level definitions are used by the Arm ABI specifications:
- Release
- Arm considers this specification to have enough implementations, which have received sufficient testing, to verify that it is correct. The details of these criteria are dependent on the scale and complexity of the change over previous versions: small, simple changes might only require one implementation, but more complex changes require multiple independent implementations, which have been rigorously tested for cross-compatibility. Arm anticipates that future changes to this specification will be limited to typographical corrections, clarifications and compatible extensions.
- Beta
- Arm considers this specification to be complete, but existing implementations do not meet the requirements for confidence in its release quality. Arm may need to make incompatible changes if issues emerge from its implementation.
- Alpha
- The content of this specification is a draft, and Arm considers the likelihood of future incompatible changes to be significant.
Unless otherwise indicated, all content in this document is at the Release quality level.
Issue | Date | Change |
---|---|---|
2Q2018 | 26th June 2018 | First public release. |
2019Q1 | 29th March 2019 | Fix broken link in License
section. Fix parameter
numbering for linear steps in
Vector function name mangling. Clarify the
behavior for structures like struct { int8_t R,
G, B; }; in
Parameter and return value mapping,
and relative RGB Example. |
2019Q1.1 | 30th April 2019 | Minor clarification on the definition of SVE unpacked vector. Refer to the original AAPCS and list the registers that are call-preserved and call-clobbered in the base convention (Vector Procedure Call Standard, no functional change). Add chapter on User defined vector functions via OpenMP 5.0. |
2019Q2 | 30th June 2019 | Fix the use of Add section on Dynamic linking for AAVPCS with new requirement for ELF platforms that support dynamic linking. Fix mangled name for function Non functional changes:
|
2019Q4 | 30th January 2020 | Github preview release with an open source license. Major changes:
Minor changes:
Several changes have been applied to the sources to fix the rendered page produced by github. In particular:
|
2020Q2 | 1st July 2020 | Clarify whether aarch64_vector_pcs is needed for SVE. Clarify the definition of complex type. |
AArch64 functions use the calling convention described in section 5 of the Procedure Call Standard for the ARM 64-bit Architecture (with SVE support), or AAPCS hereafter. The most recent version of the AAPCS can be found on developer.arm.com.
Note
The SVE-specific rules of the AAPCS are in beta version. The list of SVE call-clobbered and call-preserved registers in table AAVPCS Table will be updated when the final version of the AAPCS is published.
The procedural calling standard of the AAPCS requires that none of the 32 Advanced SIMD vector registers V0-V31 are treated as call-preserved (with the exception of the lower half of V8-V15, or D8-D15), thus requiring the caller to perform up to 32 vector stores before a call and up to 32 vector loads after it (see section 5.1.2 of AAPCS). For workloads with performance hot spots in leaf routines (an example of which are vector math functions), we find that a modified procedural calling standard for the vector units in AArch64 would be more efficient than the base procedural calling standard. Therefore, to efficiently support such vector routines, we define a modified version of the base procedural calling standard, called the Vector Procedure Call Standard for the Arm 64-bit Architecture (AAVPCS).
The list of parameter, result, call-preserved and call-clobbered registers for the AAVPCS are presented in the following table:
Extension | Parameter and Result registers | Call-clobbered registers | Call-preserved registers |
---|---|---|---|
Advanced SIMD | V0-V7 | V0-V7, V24-V31 | V8-V23 |
SVE | Z0-Z7 | See AAPCS |
The AAVPCS is implicit when a #pragma omp declare simd
clause is attached
to a function definition or declaration. For user-defined Advanced SIMD vector
functions, the same behavior can be obtained by adding the
aarch64_vector_pcs
function attribute to the function definition or
declaration as in the following examples. For user-defined SVE vector functions
the attribute is not required as AAPCS and AAVPCS are equivalent. Note that to
ensure the compiler produces ABI consistent code, the attribute must be
specified in every declaration and definition of the function.
/* function definition */
__attribute__((aarch64_vector_pcs))
uint64x2_t foo(uint32x2_t a, float32x2_t b) {
/* function body */
}
/* function declaration */
__attribute__((aarch64_vector_pcs)) float64x2_t bar(float64x2_t a);
On ELF platforms with dynamic linking support, symbol definitions
and references must be marked with the STO_AARCH64_VARIANT_PCS
flag set in their st_other
field if the following conditions hold:
- The binding for the symbol is not
STB_LOCAL
, or it is in the dynamic symbol table. - The symbol is associated with a function following the AAVPCS convention.
For more information on STO_AARCH64_VARIANT_PCS
, see AAELF64.
Note
Marking all functions that follow the AAVPCS convention is a valid way of implementing this requirement.
For the purposes of this specification, we define the following notational extensions for the Advanced SIMD vector types defined by the AAPCS64. These types are not made available to the user.
Padded short vectors extend the definition of short vectors and are
used as a notational convenience to describe vector types with a size
of less than 64 bits. These can be formed where the simdlen
clause
specified in an OpenMP declare simd
construct would force a
smaller vector than would meet the AAPCS definition of a short
vector. These have the form of a vector with <N>
elements of type
<T>
:
<T>x<N>_t
Where
sizeof(<T>) * <N> < 8
A padded short vector is represented as an 8-byte short vector type
with elements of type <T>
in which lanes <N>
and above have
unspecified values. For example, a padded short vector uint16x2_t
is represented as a uint16x4_t
in which lanes 2 and 3 have
unspecified values.
The contents of the 8-byte vector are arranged as though the whole
padded short vector were a single lane. For example, a uint16x2_t
is stored in the uint16x4_t
as though it were lane 0 in a
uint32x2_t
.
Note
When a padded short vector is transferred between registers and memory it is treated as an opaque object of the notional type. That is, a padded short vector is stored in memory as if it were stored with a single STR of an object of the size of the notional type of the padded short vector; a padded short vector is loaded from memory using the corresponding LDR instruction. On a little-endian system this means that element 0 will always contain the lowest addressed element of a padded short vector; on a big-endian system element 0 will contain the highest-addressed element of a padded short vector.
This is shown in the following table.
Padded short vector type | Short vector type | Little-endian | Big-endian |
---|---|---|---|
[u]int8x2_t |
[u]int8x8_t |
X|X|X|X|X|X|A[1]|A[0] |
X|X|X|X|X|X|A[0]|A[1] |
[u]int8x4_t |
[u]int8x8_t |
X|X|X|X|A[3]|...|A[0] |
X|X|X|X|A[0]|...|A[3] |
float16x2_t |
float16x4_t |
X|X|A[1]|A[0] |
X|X|A[0]|A[1] |
The set of padded short vector types, the short vector type they map to, and the appropriate store width for each type is given in the following table,
Padded short vector type | Short vector type |
|
---|---|---|
[u]int8x1_t |
[u]int8x8_t |
Bn |
[u]int8x2_t |
[u]int8x8_t |
Hn |
[u]int8x4_t |
[u]int8x8_t |
Sn |
[u]int16x1_t |
[u]int16x4_t |
Hn |
[u]int16x2_t |
[u]int16x4_t |
Sn |
float16x1_t |
float16x4_t |
Hn |
float16x2_t |
float16x4_t |
Sn |
float32x1_t |
float32x2_t |
Sn |
When using a padded short vector, the contents of the elements of the associated short vector that lie outside the padded short vector are undefined.
Where padded short vectors are used, this may cause the compiler to emit conservative, scalar code to process their content.
No language bindings are provided for padded short vectors. Padded short vectors are not generated for declare simd constructs with no simdlen clause.
Extended short vectors extend the AAPCS definition of short vectors and are used as a notational convenience to describe vector types with a size greater than 128 bits. These can be formed where the required vectorization factor would create a larger vector than would meet the AAPCS definition of a short vector. These have the form:
<T>x<N>_t
Where
sizeof(<T>) * <N> > 16
Extended short vectors are represented as a structure containing an array of short vectors of the appropriate type. These have the general form:
struct <T>x<NN>x<M>_t { <T>x<NN>_t val[<M>]; };
Where <NN>
is such that <N>=<NN> * <M>
.
A subset of the possible vector types are given in the following table.
Notional type | Parameter/Return type |
---|---|
int32x16_t |
struct int32x4x4_t { int32x4_t val[4]; }; |
float64x4_t |
struct float64x2x2_t { float64x2_t val[2]; }; |
int32x16_t |
struct int32x4x4_t { int32x4_t val[4]; }; |
No language bindings are provided for extended short vectors, though
some of these types are also defined by arm_neon.h
.
Let sv<T>_t
be an SVE ACLE vector type with lanes of type
<T>
. The vector is said to be unpacked if only the logical lanes
corresponding to the multiples of some power of 2 greater or equal
than 2 can be set active by a svbool_t
predicate. Conversely, the
vector is said to be packed if any lane can be active.
For example, 32-bit signed integers from a reference int32_t * A
can
be loaded into an unpacked svint32_t
vector at lanes 0, 2,
4,... and so on, effectively using only half of the lanes available in
the vector. In the following example, the resulting SVE packed vector
is shown together with two unpacked versions (X
is for undefined
content):
lane idx 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 [msb] ... | A[7] | A[6] | A[5] | A[4] | A[3] | A[2] | A[1] | A[0] [lsb] // packed [msb] ... | X | A[3] | X | A[2] | X | A[1] | X | A[0] [lsb] // unpacked 0, 2, 4, ... [msb] ... | X | X | X | A[1] | X | X | X | A[0] [lsb] // unpacked 0, 4, 8, ...
In this specification, the term complex type will be used to refer to the following language-dependent types:
- For C/C++, any of the complex types as defined in the C99 and C11 standards. For more information, see C complex numbers.
- For C++, any of the complex types as defined in the header file <complex>. For more information, see C++ complex numbers.
- For Fortran, any of the types defined using the COMPLEX statement.
This section describes how the scalar functions decorated with the
OpenMP declare simd
pragma are associated to vector function
signatures.
When vectorizing the following loop, whatever vectorization factor we
choose, we want to make sure that the compiler expects a vector
version of f
and g
that operates on the same number of
lanes.
float f(double);
double g(float);
float x[];
//...
for (int i = 0; i < 100; ++i)
x[i] = f(g(x[i]));
The rules given in this chapter guarantee that any #pragma omp
declare simd
attached to a function declaration or definition
generates a unique set of vector functions associated to the original
scalar function. This is done to make sure that library vendors can
provide a unique way to interface the routines of the library with a
compiler, by means of the declare simd
directive.
In all cases, the order of the vector function parameters reflects the ordering of the parameters of the original scalar function.
Throughout this chapter, f
is a function declaration or
definition decorated with an OpenMP declare simd
directive, <P>
is the return value or an input parameter of f
, and <T(P)>
is its associated type.
One or more vector functions F
are associated to the original
scalar function f
. The return value and each function parameter
is mapped to a unique return value or input parameter respectively,
named Mapping of P
, or MAP(P)
. The type of
these vector function return value and input parameters depends on the
following rules. Their order is the same as in the original scalar
function f
.
To each <P>
, a true
/ false
predicate "P
Maps To
Vector", or MTV(P)
hereafter, is associated as follows:
If
<P>
is an input parameter such that:<P>
is auniform
value, or<P>
is alinear
value and not a reference marked withval
or no linear modifiers,
then
MTV(P)
isfalse
.If
P
is avoid
return value, thenMTV(P)
isfalse
;In all other cases,
MTV(P)
istrue
.
When a scalar parameter maps to a vector, that vector sometimes
contains the values of the scalar parameters and sometimes contains
the addresses of the scalar parameters. The predicate Pass by
Value PBV(T)
is true
if the former case applies for scalar
parameters of type T
; it is false
if the latter case
applies. The predicate is defined as follows:
PBV(T)
istrue
if (a)T
is an integer, floating-point or pointer type and (b)sizeof(T)
is1
,2
,4
or8
.PBV(T)
istrue
ifT
is a complex type with components of typeT'
and ifPBV(T')
istrue
.- Otherwise
PBV(T)
is false.
When mapping the return value or an input parameter <P>
of the scalar
function to the corresponding MAP(P)
in the
vector function, the following rules apply:
If
MTV(P)
isfalse
, thenMAP(P)
isP
.Otherwise, if
MTV(P)
istrue
, thenMAP(P)
is target specific:- For Advanced SIMD,
MAP(P) = ADVSIMD_MAP(P)
, withADVSIMD_MAP(P)
defined in section Advanced SIMD-specific rules. - For SVE,
MAP(P) = SVE_MAP(P)
, withSVE_MAP(P)
defined in section SVE-specific rules.
- For Advanced SIMD,
In all cases, when
<P>
is the return value, and:MTV(P) = true
.PBV(P) = false
.MAP(P)
is a vector of pointers.
Then the return type of the associated vector function is
void
, andMAP(P)
becomes the first parameter of the vector function. The caller is responsible for allocating the memory associated with the pointers inMAP(P)
.
A set of vector lengths VLEN
is sometimes associated with the
generated vector function F
. When this is done, the algorithm for
selecting the value(s) of VLEN
is target dependent. The algorithm
makes use of the definitions in this section.
We then define the Lane Size of P, or LS(P)
, as follows.
- If
MTV(P)
isfalse
andP
is a pointer or reference to some typeT
for whichPBV(T)
istrue
,LS(P) = sizeof(T)
. - If
PBV(T(P))
istrue
,LS(P) = sizeof(P)
. - Otherwise
LS(P) = sizeof(uintptr_t)
.
For the function f
, we define the following concepts:
- The Narrowest Data Size of f, or
NDS(f)
, as the minumum of the lane sizeLS(P)
among all input parameters and return value<P>
off
. - The Widest Data Size of f, or
WDS(f)
, as the maximum of the lane sizeLS(P)
among all input parameters and return value<P>
off
.
Note that by definition the value of NDS(f)
and WDS(f)
can
only be 1, 2, 4, 8, and 16.
This section describes the Advanced SIMD-specific rules for mapping
<P>
to its corresponding vector parameter MAP(P)
when MTV(P)
= true
.
A VLEN
is always associated with the vector function. The rules to
generate the set of the available values are:
- If
simdlen(len)
is specified, then the compiler generates only one version withVLEN = len
. The value ofvlen
must be a power of 2. - If no
simdlen
is specified, the compiler generates multiple versions, according to the following rules:- if
NDS(f) = 1
, thenVLEN = 16, 8
; - if
NDS(f) = 2
, thenVLEN = 8, 4
; - if
NDS(f) = 4
, thenVLEN = 4, 2
; - if
NDS(f) = 8
orNDS(f) = 16
, thenVLEN = 2
.
- if
For a value of VLEN
, the ADVSIMD_MAP(P)
is build as follows:
- If
PBV(T(P))
isfalse
,ADVSIMD_MAP(P)
is a vector ofVLEN
elements of typeuintptr_t
. - If
T(P)
is a complex type with components of typeT
,MAP(P)
is a vector of2*VLEN
elements of typeT
. - Otherwise
ADVSIMD_MAP(P)
is a vector ofVLEN
elements of typeT(P)
. - An optional
{not}inbrach
clause defines whether or not a vector mask parameter is added as the last input parameter ofF
, according to the rules in table 1 in chapter 4. The vector mask type is selected by building a vector ofVLEN
elements consisting of unsigned integers ofNDS(f)
bytes. The generation of the values in the mask parameter is described in section 4.1.
This section describes the SVE-specific rules for mapping <P>
to
its corresponding vector parameter MAP(P)
when MTV(P) = true
.
One vector function F
is associated to f
depending on its classification via the declare simd
directive.
The vector signatures that get generated are the same in all cases.
- If no
simdlen
clause is specified, a VLA vector version is associated. - When using a
simdlen(len)
clause, the compiler expects a VLS vector version of the function that is tuned for a specific implementation of SVE. The size of the implementation isWDS(f)* len * 8
.
Whether targeting VLA SVE or VLS SVE, the rules for mapping <P>
to
SVE_MAP(P)
are:
- If
PBV(T(P))
isfalse
,SVE_MAP(P)
is a scalable vector ofuintptr_t
. - If
T(P)
is a complex type with components of typeT
,SVE_MAP(P)
is a scalable vector ofT
. - Otherwise
SVE_MAP(P)
is a scalable vector ofT(P)
. - An additional
svbool_t
mask parameter is added as the last parameter ofF
. The generation of the mask values is described in section 4.2.
The vectors of the signature of F
are packed or unpacked
according to the following rules:
- if
LS(P) = WDS(f)
, then the vector is packed. - If
LS(P) < WDS(f)
, then the vector is unpacked.
Each element in the unpacked vector occupies the same number of bits as in the packed vector, and all elements are aligned to their least significant bits.
The following example shows the contents of an SVE vector consisting
of 1-byte lanes, unpacked and aligned with the 4-byte lanes of a
packed vector. The ??
characters indicate a byte whose value is
undefined.
Zn.b [msb] ... 0x??????03 0x??????02 0x??????01 0x??????00 [lsb] Zn.s [msb] ... 0x00000003 0x00000002 0x00000001 0x00000000 [lsb]
The rules of the mangling scheme for vector functions are summarized by Name mangling function.
With reference to Name mangling function, the rules for
building the <parameters>
group are:
- We generate one
<parameter>
token in the<parameters>
group for each of the input parameters of the scalar function. The tokens are in the same order as the input parameters. - The rules for choosing the
<parameter_kind>
are defined in the Description of the parameter_kind token. - The optional
"a" <X>
token represents the alignment value (in bytes) specified in thealigned
clause (for examplealigned(c:a)
).- When targeting Advanced SIMD, if the value
a
is missing, the default alignment value is 16 (128 bits), so that an aligned clause with no alignment is mangled asa16
. - When targeting SVE, the default value of an
aligned
clause is the alignment of the type pointed to by the corresponding parameter of the scalar signature. For example,aligned(x)
forT *x
defaults to the value_Alignof(typeof(T))
.
- When targeting Advanced SIMD, if the value
Name mangling grammar for vector functions.
<vector name> := <prefix> "_" <name> <name> := Assembly name of the function <prefix> := "_ZGV" <isa> <mask> <len> <parameters> <isa> := "n" (Advanced SIMD) | "s" (SVE) <mask> := "N" (No Mask) | "M" (Mask) <len> := VLEN (VLS SVE or Advanced SIMD) | "x" (VLA SVE) <parameters> := <parameter> { <parameter> } <parameter> := <parameter_kind> [ "a" <X> ] OpenMP version support (onwards) <parameter_kind> := "v" 4.0 | "l" | "l" <number> 4.0 | "R" | "R" <number> 4.5 | "L" | "L" <number> 4.5 | "U" | "U" <number> 4.5 | "ls" <pos> 4.5 | "Rs" <pos> 4.5 | "Ls" <pos> 4.5 | "Us" <pos> 4.5 | "u" 4.0 <number> := "n" <X> // "n" means negative | <Y> <pos> := <X> <X> := integral number greater than or equal to 1 <Y> := integral number greater than or equal to 2
"v"
- Vector parameter - default for no linear/uniform clause.
"u"
- Uniform parameter specified in the uniform clause. For example,
uniform(c)
.
"l" | "l" <number>
- Linear parameter
<P>
for which (a) the step is a compile-time constant, (b)MTV(P)=false
and (c) the linear clause has either a val modifier or no modifier.<number>
is the value of the constant linear step, or an empty string if the step is 1. For example,linear(i:2)
givesl2
andlinear(i:1)
givesl
when the type ofi
isinteger
. "R" | "R" <number>
- Linear parameter
<P>
for which (a) the step is a compile-time constant, and (b) the linear clause has a ref modifier.<number>
is the value of the constant linear step, or an empty string if the step is 1. For example,linear(ref(i):3)
givesR3
andlinear(ref(i):1)
givesR
when the type ofi
isinteger
. "L" | "L" <number>
- Linear parameter
<P>
for which (a) the step is a compile-time constant, (b)MTV(P)=true
and (c) the linear clause has either a val modifier or no modifier.<number>
is the value of the constant linear step, or an empty string if the step is 1. For example,linear(val(i):-3)
givesLn3
when the type ofi
isinteger
.
In the previous cases, when the parameter <P>
marked by the linear
clause is a pointer or an OpenMP integral reference to a type T
,
the step
of the linear clause must be multiplied by the size in
bytes of the pointee, so that <number>=sizeof(T) x step
.
"U" | "U" <number>
- Linear parameter
<P>
for which (a) the step is a compile-time constant and (b) the linear clause has a uval modifier.<number>
is the value of the constant linear step, or an empty string if the step is 1. For example,linear(uval(i):2)
givesU2
.
"ls" <pos>
- Linear parameter
<P>
for which (a) the step is a loop-independent runtime invariant, (b)MTV(P)=false
and (c) the linear clause has either a val modifier or no modifier.<pos>
is the position (starting from 0) of the step parameter specified in the uniform clause (required by the OpenMP specs). For example,linear(i:c) uniform(c)
withc
being the third parameter givesls2
. "Rs" <pos>
- Linear parameter
<P>
for which (a) the step is a loop-independent runtime invariant and (b) the linear clause has a ref modifier.<pos>
is the position of the step parameter (starting from 0) specified in the uniform clause (required by the OpenMP specs). For example,linear(ref(i):c) uniform(c)
withc
being the third parameter givesRs2
. "Ls" <pos>
- Linear parameter
<P>
for which (a) the step is a loop-independent runtime invariant, (b)MTV(P)=true
and (c) the linear clause has either a val modifier or no modifier.<pos>
is the position of the step parameter (starting from 0) specified in the uniform clause (required by the OpenMP specs). For example,linear(val(i):c) uniform(c)
withc
being the first parameter, givesLs0
. "Us" <pos>
- Linear parameter
<P>
for which (a) the step is a loop-independent runtime invariant and (b) the linear clause has a uval modifier.<pos>
is the position of the step parameter (starting from 0) specified in the uniform clause (required by the OpenMP specs). For example,linear(uval(i):c) uniform(c)
withc
being the third parameter, givesUs2
.
The following example shows which vector versions are provided when no
simdlen
clause is attached to the declare simd
directive of a
function declaration.
#pragma omp declare simd
float f(double x);
#pragma omp declare simd
double g(float x);
In this case, the vector versions of f
and g
operate on
vectors consisting of 2 and 4 lanes, both with and without an
additional lane masking parameter.
For the example, the available (unmasked) signatures associated to
f
and g
are:
float32x2_t _ZGVnN2v_f(float64x2_t vx);
2-lanef
;float64x2_t _ZGVnN2v_g(float32x2_t vx);
2-laneg
;float32x4_t _ZGVnN4v_f(float64x4_t vx);
4-lanef
;float64x4_t _ZGVnN4v_g(float32x4_t vx);
4-laneg
;
It is possible to tune the number of lanes using the simdlen(N)
clause, where N=2^k for k>=0. No other values of
simdlen
are allowed.
#pragma omp declare simd simdlen(2)
short foo(int64_t x, uint32_t y , int8_t z);
// 2-lane version.
int16x2_t _ZGVnN2vvv_foo(int64x2_t vx, uint32x2_t vy, int8x2_t vz);
#pragma omp declare simd simdlen(4)
short foo(int64_t x, uint32_t y, int8_t z);
// 4-lane version.
int16x4_t _ZGVnN4vvv_foo(int64x4_t vx, uint32x4_t vy, int8x4_t vz);
Note
Because AArch64 Advanced SIMD uses the first 8 SIMD
registers for passing parameters and returning values, it is
recommended that the value passed to simdlen
is such
that the signature of the vector function does not use more
than 8 input registers, or more than 8 return registers.
In case of the functions float f(double)
, double g(float)
and short foo(int64_t, int32_t, int8_t)
, the use of
#pragma omp declare simd
will generate the following function
signatures:
svfloat32_t _ZGVsMxv_f(svfloat64_t, svbool_t)
VLA signature for the vector version off
;svfloat64_t _ZGVsMxv_g(svfloat32_t, svbool_t)
VLA signature for the vector version ofg
;svint16_t _ZGVsMxvvv_foo(svint64_t, svint32_t, svint8_t, svbool_t)
VLA signature for the vector version offoo
.
Note that the svbool_t
parameter is described in SVE masking.
// Example with explicit `simdlen` for SVE.
#pragma omp declare simd simdlen(10) notinbranch
#pragma omp declare simd simdlen(16) notinbranch
int32_t foo(int32_t x);
// No 10-lane version generated because ten 4-byte lanes do not
// fit an SVE register.
// SVE 512-bit - widest type is 4 bytes -> 16 lanes
svint32_t _ZGVsM16v_foo(svint32_t vx, svbool_t vmask);
#pragma omp declare simd simdlen(8)
float bar(double x, double y);
// widest type is 8 bytes
// SVE 512-bit -> 8 lanes
svfloat32_t _ZGVsM8vv_bar(svfloat64_t vx, svfloat64_t vy,
svbool_t vmask);
Input parameters marked with a linear
clause need special
handling. In particular, the linear clause specifies an implicit
vector of values or addresses, depending on the type of the clause.
linear
clause when x
is an integral parameter.
Clause | MAP(x) |
Mangled parameter name when s is: |
Constraints at lane
i of the
implicit vector |
|
---|---|---|---|---|
Compile time constant | uniform parameter |
|||
linear(x:s) |
x |
"l" + s |
"ls" + pos(s) |
x_i = x + i * s |
linear(val(x):s) |
||||
linear(uval(x):s) |
n/a | n/a | n/a | n/a |
linear(ref(x):s) |
linear
clause when x
is a pointer.
Clause | MAP(x) |
Mangled parameter name when s is: |
Constraints at lane i of
the implicit vector |
|
---|---|---|---|---|
Compile time constant | uniform parameter |
|||
linear(x:s) |
x |
"l" + s * sizeof(*x) |
"ls" + pos(s) |
x_i = x + i * s |
linear(val(x):s) |
||||
linear(uval(x):s) |
n/a | n/a | n/a | n/a |
linear(ref(x):s) |
linear
clause when x
is an integral reference (C++ and Fortran dummy parameters only).
Clause | MAP(x) |
Mangled parameter name when s is: |
Constraints at lane i of the
implicit vector |
|
---|---|---|---|---|
Compile time constant | uniform parameter |
|||
linear(x:s) |
[&x_0, &x_1, ..., &x_i, ...] |
"L" + s |
"Ls" + pos(s) |
x_i = x + s * i |
linear(val(x):s) |
||||
linear(uval(x):s) |
x |
"U" + s |
"Us" + pos(s) |
x_i = x + s * i and &x_i = &x |
linear(ref(x):s) |
x |
"R" + s * sizeof(x) |
"Rs" + pos(s) |
&x_i = &x + s * i |
// C examples for the ``linear`` clause.
// The same rules apply to dummy arguments passed by value in
// Fortran. Note that the function signatures for the ``val`` modifier
// are the same as when no modifier is present.
// Advanced SIMD
#pragma omp declare simd linear(i)
float bar(int32_t i);
// 2-lane version
float32x2_t _ZGVnN2l_bar(int32_t);
// 4-lane version
float32x4_t _ZGVnN4l_bar(int32_t);
#pragma omp declare simd linear(x)
float foo(double *x);
// 2-lane version
float32x2_t _ZGVnN2l8_foo(double *);
// 4-lane version
float32x4_t _ZGVnN4l8_foo(double *);
// SVE
#pragma omp declare simd linear(i)
float bax(int32_t i);
// VLA version
svfloat32_t _ZGVsMxl_bax(int32_t, svbool_t);
#pragma omp declare simd linear(x)
float bax(double *x);
// VLA version with signature
svfloat32_t _ZGVsMxl8_bax(double *, svbool_t);
// C++ examples for ``linear`` clause when using reference parameters.
// The same function signature is generated for dummy arguments
// passed by reference in Fortran. For simplicity, the masked version
// for Advanced SIMD is not shown.
#pragma omp declare simd linear(ref(x))
int32_t g_ref(int32_t &x); // The vector version holds a pointer to x
// Advanced SIMD - 2-lane version
int32x2_t _ZGVnN2R4_g_ref(int32_t *);
// Advanced SIMD - 4-lane version
int32x4_t _ZGVnN4R4_g_ref(int32_t *);
// SVE - VLA version
svint32_t _ZGVsMxR4_g_ref(int32_t *, svbool_t);
#pragma omp declare simd linear(val(x))
int32_t g_val(int32_t &x); // vector of integral values
// Advanced SIMD - 2-lane version
int32x2_t _ZGVnN2L4_g_val(uint64x2_t vxp);
// Advanced SIMD - 4-lane version
int32x4_t _ZGVnN4L4_g_val(uint64x4_t vxp);
// SVE - VLA version
svint32_t _ZGVsMxL4_g_val(svuint64_t vxp , svbool_t);
#pragma omp declare simd linear(uval(x))
int32_t g_uval(int32_t &x); // scalar, used to produce a vector of integral values from x
// Advanced SIMD - 2-lane version
int32x2_t _ZGVsN2U4_g_uval(int32_t *);
// Advanced SIMD - 4-lane version
int32x4_t _ZGVsN4U4_g_uval(int32_t *);
// SVE - VLA version
svint32_t _ZGVsMxU4_g_uval(int32_t *, svbool_t);
Warning
The context of this chapter is at Beta level. See Current status and anticipated changes. Any feedback should be provided via the issue tracker page on GitHub.
It is possible to map a scalar function f
to a user-defined
vector function F
by using the directive #pragma omp declare
variant
. This pragma was introduced in version 5.0 of the OpenMP
standard.
The following table shows the traits introduced by this Vector Function ABI.
Trait set | Trait value | Notes |
---|---|---|
device |
isa("simd") |
Advanced SIMD call. |
device |
isa("sve") |
SVE call. |
device |
arch("march-list") |
Used to match
-march=march-list
from the compiler. |
The scalar function f
that is decorated with a declare
variant
directive with a simd
trait in the construct
set is
mapped to the vector function F
according to the following rules:
- The signature of
F
must be the same as that obtained byf
when decorated with adeclare simd
directive that matches thesimd
construct specified in thedeclare variant
directive, according to the rules specified in Vector function signature. - The
device
traits defined in table AArch64 Variant Traits must be used to narrow the context for matching purposes:isa("simd")
targets Advanced SIMD function signatures.isa("sve")
targets SVE function signatures.- Either
isa("simd")
orisa("sve")
must be specified. - The
arch
traits of thedevice
set is optional, and it accepts any value that can be passed to the compiler via the command line option-march
.
- The
extension("scalable")
trait of theimplementation
set informs the compiler that thesimdlen
clause of thesimd
construct must be omitted to target all vector lengths. Its use in adeclare variant
directive is equivalent to having no simdlen on#pragma omp declare simd
when targeting SVE. - Using
extension("scalable")
when usingisa("simd")
is invalid.
Note
Decorating a scalar function f
with the pragma does not
automatically make the vector function F
use the vector
calling conventions in Vector Procedure Call Standard. The
vector function will only use the vector calling conventions
if it is marked with the aarch64_vector_pcs
attribute. The vector function does not need to use the
vector calling conventions, although it is recommended in
general.
// User defined `cosine` function for Advanced SIMD.
#pragma omp declare variant(UserCos) \
match(construct={simd(simdlen(2), notinbranch)}, device={isa("simd")})
double cos(double x);
float64x2_t UserCos(float64x2_t vx);
// User defined `sincosf` function for VLA SVE.
#pragma omp declare variant(UserSinCos) \
match(construct={simd(notinbranch, linear(sin, cos))}, \
device={isa("sve")}, implementation={extension("scalable")})
void sincosf(float in, float *sin, float *cos);
void UserSinCos(svfloat32_t vin, float *sin, float *cos, svbool_t vmask);
// Advanced SIMD function in an SVE context.
#pragma omp declare variant(F) \
match(construct={simd(simdlen(4), inbranch)}, \
device={isa("simd")})
double f(int x);
float64x4_t F(int32x4_t vx, uint32x4_t vmask);
// VLS version targeting SVE.
#pragma omp declare variant(F) \
match(construct={simd(simdlen(6), inbranch)}, \
device={isa("sve")})
double f(int x);
svfloat64_t F(svint32_t vx, svbool_t vmask);
// Matching via `-march`.
#pragma omp declare variant(H) \
match(construct={simd(notinbranch)}, \
implementation={extension("scalable")}, \
device={isa("sve"), arch("armv8.2-a+sve")})
int h(int x);
svint32_t H(svint32_t vx, svbool_t vmask);
// Invalid use. This vector signature cannot be derived from the scalar
// function by means of `#pragma omp declare simd`.
#pragma omp declare variant(G) \
match(construct={simd(simdlen(2),notinbranch)}, device={isa("sve")})
char g(double x);
svuint8_t G(float64x2_t vx);
The inbranch
and notinbranch
clauses define whether or not a
vector function should accept a masking parameter.
In all cases, the masking parameter is added to the vector function signature as the last parameter. The following table summarizes the behavior.
Notice that for SVE, masking is present regardless of whether
inbranch
or notinbranch
is used. [1]
Masked signature generation for [not]inbranch
clause.
Advanced SIMD | SVE | |||
---|---|---|---|---|
Masked | Unmasked | Masked | Unmasked | |
#pragma omp declare simd |
Yes | Yes | Yes | No |
#pragma omp declare simd inbranch |
Yes | No | Yes | No |
#pragma omp declare simd notinbranch |
No | Yes | Yes | No |
In a masked vector function, the contents of the inactive lanes of the input parameters and the inactive lanes of the return value are undefined.
For Advanced SIMD, the type of the mask is generated using
uint[NDS(f)*8]_t
-based vectors.
All bits are set to one for active lanes, and all bits are set to zero for inactive lanes.
Note
The narrowest vector input parameter is chosen over the widest one because masking is often intended for lane masking, and not for bit masking of the vector lanes. Using the narrowest vector input parameter also limits the number of parameter registers needed to pass the mask.
Note
Because the masking is done using SIMD data registers, to avoid performance degradation it is recommended that the addition of the mask parameter does not overflow the maximum number of 8 vector input registers.
#pragma omp declare simd simdlen(2) inbranch
float f(double);
// 2-lane masked version
float32x2_t _ZGVnM2v_f(float64x2_t, uint32x2_t);
#pragma omp declare simd simdlen(2) inbranch
double g(float);
// 2-lane masked version
float64x2_t _ZGVnM2v_g(float32x2_t, uint32x2_t);
#pragma omp declare simd inbranch
float f(double); // -> float32x2_t(float64x2_t, uint32x2_t)
// 2 and 4-lane masked version
float32x2_t _ZGVnM2v_f(float64x2_t, uint32x2_t);
float32x4_t _ZGVnM4v_f(float64x4_t, uint32x4_t);
#pragma omp declare simd inbranch
double g(float);
// 2 and 4-lane masked version
float64x2_t _ZGVnM2v_g(float32x2_t, uint32x2_t);
float64x4_t _ZGVnM4v_g(float32x4_t, uint32x4_t);
#pragma omp declare simd simdlen(8) inbranch
float f(double);
// 8-lane masked version
float32x8_t _ZGVnM8v_f(float64x8_t, uint32x8_t);
#pragma omp declare simd simdlen(8) inbranch
double g(float);
// 8-lane masked version
float64x8_t _ZGVnM8v_g(float32x8_t, uint32x8_t);
Note
Using a mask parameter in AArch64 Advanced SIMD is not generally recommended for functions that operate on scalars of different widths, as widening of the input mask for wider types might require using call-preserved temporary registers (V8-V23).
Example of mask parameters for complex values.
#pragma omp declare simd inbranch
int32_t foo(_Complex double x);
// Advanced SIMD, 2-lane versions.
// Each logical lane of the mask is a 4 byte sequence,
// either 0x00000000 or 0xffffffff.
int32x2_t _ZGVnM2v_foo(float64x4_t vx, uint32x2_t vmask);
#pragma omp declare simd inbranch
float complex baz(double complex x);
// Double precision complex value -> 16 byte structure
// 2-lane Advanced SIMD.
// The narrowest type is an 8 byte structure, so mask
// is uint64x2_t
float32x4_t _ZGVnM2v_baz(float64x4_t vx, uint64x2_t vmask);
#pragma omp declare simd inbranch
double complex bar(float x, float y);
// Advanced SIMD, 2, and 4-lane.
float64x4_t _ZGVnM2vv_bar(float32x2_t vx, float32x2_t vy, uint32x2_t vmask);
float64x8_t _ZGVnM4vv_bar(float32x4_t vx, float32x4_t vy, uint32x4_t vmask);
For SVE vector functions, whether length-agnostic or length-specific,
masked signatures are generated by adding a svbool_t
mask (or
predicate in SVE terms) as the last parameter.
#pragma omp declare simd
#pragma omp declare simd inbranch
#pragma omp declare simd notinbranch
float f(double);
// SVE - VLA
// Notice that the default behavior is not affected by `inbranch`
// or `notinbranch`.
svfloat32_t _ZGVsMxv_f(svfloat64_t, svbool_t);
#pragma omp declare simd
double g(float);
// SVE - VLA
svfloat64_t _ZGVsMxv_f(svfloat32_t, svbool_t);
#pragma omp declare simd simdlen(4)
float f(double);
// SVE - VLS - > implies a 256-bit implementation
svfloat32_t _ZGVsM4v_f(svfloat64_t, svbool_t);
#pragma omp declare simd simdlen(4)
double g(float);
// SVE - VLS - > implies a 256-bit implementation
svfloat64_t _ZGVsM4v_g(svfloat32_t, svbool_t);
The logical lane subdivision of the predicate corresponds to the lane subdivision of the vector data type generated for the widest data type, with one bit in the predicate lane for each byte of the data lane. Active logical lanes of the predicate have the least significant bit set to 1, and the rest set to zero. The bits of the inactive logical lanes of the predicate are set to zero. This method ensures that:
- The inactive lanes of unpacked vectors do not get treated
erroneously as active (see example
foo
). - The correct predicate can be generated programmatically from the input predicate for those types of the scalar signature whose layout requires more than 1 bit per active lane.
In the function foo
of the following example, the widest data
type subdivision selects 8-byte wide lanes. Therefore,
the active lanes in the predicate will be represented by the 8-bit
sequence 00000001
. The original input predicate works for all the
types in the signature but not for the vy
parameter. The callee must
generate a new predicate for it, that carries the bit sequence
00010001
for the active lanes, so that the additional bytes of the
logical lane associated to the complex type are correctly marked as
active.
#pragma omp declare simd
double foo(double x, _Complex float y);
// VLA SVE
svfloat64_t _ZGVsMxv_foo(svfloat64_t vx, svfloat32_t vy,
svbool_t vmask);
// vmask active lane value: 00000001
// vy active lane value: 00010001
Throughout the following examples, for a given function f
, we
define NDS(f) and WDS(f) as the Narrowest (and respectively,
Widest) Size of f as the size in bytes of the narrowest (and
respectively, the widest) among the input parameter types and the
return type of the function signature.
The NDS and WDS values are placed next to the vector signature to
explain the choice of the vector length of the function. As a
reminder, the former is used to select the vector length when
targeting Advanced SIMD vectorization, the latter to select the vector
length when targeting VLS SVE functions by using
the simdlen
clause.
// Name mangling example for the SIMD directives with no
// decorations.
#pragma omp declare simd
int32_t foo(int32_t x);
// Advanced SIMD - NDS(foo) = 4 -> 2 and 4 lanes
int32x2_t _ZGVnN2v_foo(int32x2_t vx);
int32x2_t _ZGVnM2v_foo(int32x2_t vx, uint32x2_t vmask);
int32x4_t _ZGVnN4v_foo(int32x4_t vx);
int32x4_t _ZGVnM4v_foo(int32x4_t vx, uint32x4_t vmask);
// VLA SVE
svint32_t _ZGVsMxv_foo(svint32_t vx, svbool_t vmask);
// Example mangling for a function with `uniform` and `linear`
// clause, with `val` modifier. The `inbranch` clause generates only
// the masked version for Advanced SIMD.
#pragma omp declare simd inbranch uniform(x) linear(val(i):4)
int32_t foo(int32_t *x, int32_t i);
// Advanced SIMD - NDS(foo) = 4 -> 2 and 4 lanes
int32x2_t _ZGVnM2ul4_foo(int32_t *x, int32_t i, uint32x2_t vmask);
int32x4_t _ZGVnM4ul4_foo(int32_t *x, int32_t i, uint32x4_t vmask);
// VLA SVE
svint32_t _ZGVsMxul4_foo(int32_t *x, int32_t i, svbool_t vmask);
// Example of function name mangling when a runtime linear step is
// specified in the `linear` clause.
#pragma omp declare simd inbranch uniform(x,c) linear(i:c)
int32_t foo(int32_t *x, int32_t i, uint8_t c);
// Advanced SIMD - NDS(foo) = 1 -> 8 and 16 lanes
int32x4x2_t _ZGVnM8uls2u_foo(int32_t *x, int32_t i, uint8_t c, uint32x8_t vmask);
int32x4x4_t _ZGVnM16uls2u_foo(int32_t *x, int32_t i, uint8_t c, uint32x16_t vmask);
// VLA SVE
svint32_t _ZGVsMxuls2u_foo(int32_t *x, int32_t i, uint8_t c, svbool_t vmask);
// Example of vector function name generation from a fixed length
// simd declaration.
#pragma omp declare simd simdlen(4)
int32_t foo(int32_t x, float y);
// Advanced SIMD - NDS(foo) = 4 -> 4 lanes
int32x4_t _ZGVnN4vv_foo(int32x4_t vx, float32x4_t vy);
int32x4_t _ZGVnM4vv_foo(int32x4_t vx, float32x4_t vy, uint32x4_t vmask);
// SVE 128-bit - WDS(foo) = 4 -> 4 lanes
svint32_t _ZGVsM4vv_foo(svint32_t vx, svfloat32_t vy, svbool_t vmask);
// Example with output size bigger than input size.
#pragma omp declare simd
double foo(float x)
// Advanced SIMD - NDS(foo) = 4 -> 2 and 4 lanes
float64x2_t _ZGVnN2v_foo(float32x2_t vx);
float64x2_t _ZGVnM2v_foo(float32x2_t vx, uint32x2_t vmask);
float64x4_t _ZGVnN4v_foo(float32x4_t vx);
float64x4_t _ZGVnM4v_foo(float32x4_t vx, uint32x4_t vmask);
// VLA SVE - input in unpacked
svfloat64_t _ZGVsMxv_foo(svfloat32_t vx, svbool_t vmask);
Example with explicit alignment.
#pragma omp declare simd linear(x) aligned(x:16) simdlen(4)
int32_t foo(int32_t *x, float y);
// Advanced SIMD - NDS(foo) = 4 -> 4 lanes
int32x4_t _ZGVnN4la16v_foo(int32_t *x, float32x4_t vy);
int32x4_t _ZGVnM4la16v_foo(int32_t *x, float32x4_t vy, uint32x4_t vmask);
// SVE 128-bit - WDS(foo) = 4 -> 4 lanes
svint32_t _ZGVsM4la16v_foo(int32_t *x, svfloat32_t vy, svbool_t vmask);
The following example shows how to handle types that do not map directly to integers, floating-point types or complex types. In this specific case, the rules give the following:
MTV(P) = true
by rule 3 of Maps To Vector.PBV(P) = false
by rule 3 pf Pass By Value.- Because
MTV(P)
istrue
, rule 2 of Parameter and return value mapping applies. - Because
PBV(P)
isfalse
andMTV(P)
istrue
, rule 3 of Lane Size of a function parameter / return value applies and thereforeLS(P)
issizeof(uintptr_t)
. - The vector of pointers to the output values is passed as the first parameter, as specified in rule 3 of Parameter and return value mapping.
// Example with generic types. In this case, the rules lead to
// mapping each concurrent object to pointers.
struct S { uint8_t R,G,B; };
#pragma omp declare simd notinbranch
S DoRGB(S x);
// Advanced SIMD - NDS(DoRGB) = 8 (LP64 data model)
void _ZGVnN2vv_DoRGB(uint64x2_t out, uint64x2_t vx); // 2-lane
// Advanced SIMD - NDS(DoRGB) = 4 (ILP32 data model)
void _ZGVnN2vv_DoRGB(uint32x2_t out, uint32x2_t vx); // 2-lane
void _ZGVnN4vv_DoRGB(uint32x4_t out, uint32x4_t vx); // 4-lane
// VLA SVE - WDS(DoRGB) = 8 (LP64 data model)
void _ZGVsMxvv_DoRGB(svuint64_t out, svint64_t vx, svbool_t vmask);
// VLA SVE - WDS(DoRGB) = 4 (ILP32 data model)
void _ZGVsMxvv_DoRGB(svuint32_t out, svint32_t vx, svbool_t vmask);
// Example mangling for a function with `uniform` and `linear`
// clause, for corner case values.
#pragma omp declare simd linear(x:y) uniform(y) linear(z) linear(ref(k):-1) notinbranch
uint32_t foo(int32_t x, int32_t y, int32_t z, int32_t &k) {
// Advanced SIMD - NDS(foo) = 4 -> 2 and 4 lanes
uint32x2_t _ZGVnN2ls1ulRn4_foo(int32_t x, int32_t y, int32_t z, int32_t *k)
uint32x4_t _ZGVnN4ls1ulRn4_foo(int32_t x, int32_t y, int32_t z, int32_t *k)
// VLA SVE
svuint32_t _ZGVsMxls1ulRn4_foo(int32_t x, int32_t y, int32_t z, int32_t *k, svbool_t vmask);
// Example mangling for default alignment values (assuming LP64).
typedef struct D { double a[2];} D_ty;
#pragma omp declare simd \
aligned(x) aligned(y) aligned(z) aligned(S) \
linear(x) linear(y) linear(z) linear(S) notinbranch
int32_t foo(int32_t *x, double *y, uint8_t *z, D_ty * S);
// Advanced SIMD - NDS(foo) = 4 -> 2 and 4 lanes (showing only the 2 lanes one)
int32x2_t _ZGVnN2l4a16l8a16la16l16a16_foo(int32_t *x, double *y, uint8_t *z, D_ty * S)
// VLA SVE (VLS would have the same aligment tokens)
svint32_t _ZGVsMxl4a4l8a8la1l16a16_foo(int32_t *x, double *y, uint8_t *z, D_ty * S, svbool_t)
[1] | The reason for using predication by default in SVE is
to avoid a scalar tail loop when auto-vectorizing loops. The
reason for using predication even for notinbranch is to avoid
the performance degradation that would occur when porting code
that uses functions not guarded by conditional branches that could
have been marked as notinbranch . |