Internet-Draft | CBOR Numbers | July 2024 |
Bormann | Expires 10 January 2025 | [Page] |
The Concise Binary Object Representation (CBOR), as defined in STD 94 (RFC 8949), is a data representation format whose design goals include the possibility of extremely small code size, fairly small message size, and extensibility without the need for version negotiation.¶
Among the kinds of data that a data representation format needs to be able to carry, numbers have a prominent role, but also have inherent complexity that needs attention from protocol designers and implementers of CBOR libraries and of the applications that use them.¶
This document gives an overview over number formats available in CBOR and some notable CBOR tags registered, and it attempts to provide information about opportunities and potential pitfalls of these number formats.¶
This is a rather drafty initial revision, pieced together from various components, so it has a higher level of redundancy than ultimately desired.¶
This note is to be removed before publishing as an RFC.¶
Status information for this document may be found at https://datatracker.ietf.org/doc/draft-bormann-cbor-numbers/.¶
Discussion of this document takes place on the Concise Binary Object Representation Maintenance and Extensions Working Group mailing list (mailto:cbor@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/cbor/. Subscribe at https://www.ietf.org/mailman/listinfo/cbor/.¶
Source for this draft and an issue tracker can be found at https://github.com/cabo/cbor-numbers.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 10 January 2025.¶
Copyright (c) 2024 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
The Concise Binary Object Representation (CBOR), as defined in RFC 8949 [STD94], is a data representation format whose design goals include the possibility of extremely small code size, fairly small message size, and extensibility without the need for version negotiation.¶
Among the kinds of data that a data representation format needs to be able to carry, numbers have a prominent role, but also have inherent complexity that needs attention from protocol designers and implementers of CBOR libraries and of the applications that use them.¶
This document gives an overview over number formats available in CBOR and some notable CBOR tags registered, and it attempts to provide information about opportunities and potential pitfalls of these number formats.¶
It discusses CBOR representation of numbers in four main Sections:¶
These sections will generally address considerations such as:¶
Encoding efficiency (number of bytes needed), possibly processing efficiency (CPU used in processing)¶
Preferred Serialization, Common Deterministic Encoding Profile (CDE, [I-D.ietf-cbor-cde], see also [I-D.bormann-cbor-det] for more background discussion)¶
Use by applications¶
Interoperability considerations, potential "dark corners"¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [BCP14] when, and only when, they appear in all capitals, as shown here.¶
Terms and definitions from [STD94], [RFC8610], and [IEEE754] apply.¶
CBOR provides representations of integer numbers in unsigned and negative forms:¶
Unsigned integers up to 2^{64}−1, major type 0¶
Negative integers down to −2^{64}, major type 1¶
Unsigned integers with no size limitations, tag 2 on a byte string¶
Negative integers with no size limitations, tag 3 on a byte string¶
The latter two forms are often called "bignums" for historical
reasons, the former "basic" integers. The Concise Data Definition
Language (CDDL) [RFC8610] has the types uint
,
nint
, and int
, for the ranges of values covered by major type 0,
major type 1, and either of them, respectively; biguint
, bignint
,
and bigint
for the range of value covered by tag 2, tag3, and either;
and unsigned
and integer
for a choice of either form (but
interestingly no negative
).
As the preferred encoding for an integer chooses between major type
0/1 and tag 2/3 automatically, in practice biguint
and unsigned
are the same type, as are bigint
and integer
.¶
The Major type 0 numbers come in five different encoding sizes, as indicated by their initial byte: immediate ("1+0") encoding (0..23), one-byte ("1+1") (0..255), two-byte ("1+2", 0..65535), four-byte, and eight-byte. The Preferred Serialization always uses the shortest of the major type 0 encodings available for an unsigned integer. The intention is that there is no semantic difference between the major type 0 encodings, and there also is no semantic difference between major type 0 and tag 2. This means that Preferred Serialization always uses major type 0 over tag 2 when possible, and the shortest encoding of these (and thus no leading zero bytes for the tagged encodings). Major type 1 and tag 3 are analogous.¶
Note that there is no "signed type" in CBOR: as any specific number to
be represented is either negative or not, it is represented as an
unsigned integer or as a negative integer.
Major type 0 unsigned integers cover exactly the range of platform
types such as uint64_t
or u64
.
Signed platform types such as int64_t
or i64
can be represented in
the lower half of the unsigned space and the upper half of the
negative space.
Platforms typically have no nint64_t
type that could take all
negative numbers representable in major type 1; generic decoders will
therefore treat the lower half of the negative space in the same way
they will treat bignums that do not fit the signed platform type.
Similarly, generic encoders for a platform with u128
/i128
types
will choose between major type 0/1 and tag 2/3 just like they would
choose between the encoding sizes inside major type 0/1.¶
While additional representation of integers could be developed, the options already provided by [STD94] should be able to satisfy most applications.¶
While integer numbers are relatively easy to represent, floating point numbers as a realization of rational or real numbers are a much more varied subject. Many rational or real numbers require rounding until they can be encoded as a floating point number.¶
There are many choices that can be made when designing a machine representation for floating point numbers. After decades of vendor-specific formats, IEEE standardized [IEEE754], initially in 1985, updated in 2008 and then 2019 (IEC 559 is then mirroring IEEE 754). This standard is widely adopted in hardware and software, offering choices such as binary vs. decimal floating point numbers, and different representation sizes. Out of the large choice available, CBOR directly supports the three formats binary16, binary32, and binary64, i.e., the signed binary floating point formats in 16, 32, and 64 bits, colloquially known as half (16 bits), single (32 bits), and double (64 bits) precision. Most platforms that support floating point computation support at least single precision, except for the most constrained ones also double precision, while half precision is mostly used for storage and interchange only and may be software-supported only.¶
Mathematically speaking, integer numbers are a subset of the rational or real numbers from which floating point numbers are drawn. In many programming environments, however, integer numbers are clearly separated from floating point numbers (the most notable exception being the original JavaScript language, which only had one number type).¶
For specific applications, it may be desirable to represent all numbers that can be represented as integers as such, even if they are used where floating point numbers are used for non-integers. [I-D.mcnally-deterministic-cbor] defines a CDE application profile that enforces this for a certain subset of the integers.¶
Most CBOR applications so far have tended to get by with the kind of strong separation between the integer and floating point worlds that programming environments usually favor, so our focus will not be on approaches for intermingling them in this document.¶
IEEE754 distinguishes three kinds of floating point data item:¶
finite floating-point number: A finite number that is representable in a floating-point format. Note that these further divide into zero, subnormal, and normal; this distinction is usually not of interest in interchange, except that there are a few platforms with limited floating point support that may not support subnormal numbers.¶
infinite floating-point number: One of the two values −Infinite and (positive) Infinite. On many platforms, infinite numbers can be accessed via a floating point operation such as 1.0/0.0 (positive infinity) or −1.0/0.0 (negative infinity); they react to comparisons as one would expect.¶
NaN: a floating point datum that is not a number (NaN), used to represent computations that didn't lead to a numeric result, not even an infinity. A commonly implemented example for such a computation is 0.0/0.0. The formats provide a way to include additional information with a NaN, such as its sign bit, whether operations on the NaN are intended to fail immediately (signaling) or just return another NaN (quiet), and some remaining bits that may carry additional information (intended as diagnostic).¶
It can be surprising that according to [IEEE754], NaN values always compare as different even if they have the same NaN information (i.e., are identical). (There is also a totalorder relation that does give NaNs a defined place, depending on their sign bits; this only recently has been standardized as part of std::strong_order in C++20 [Cplusplus20].)¶
Not all platforms that can use IEEE 754 do provide all these kinds, e.g.,
Erlang only provides finite floating-point numbers.
Platforms that do provide them widely vary in the way they provide
access to non-finite numbers and NaNs beyond the floating point
operations given above.
Usually there is an operation such as isnan()
in C, which is needed
as comparison to a NaN always yields inequality.¶
CBOR supports the interchange of all kinds of IEEE 754 data items, including non-finite numbers and non-numbers (NaNs). For an application developer that is already using IEEE 754 floating point, there is little additional consideration required: Both infinities and NaN are widely supported in IEEE-754 hardware and software by CPUs, OS’s and programming environments. CBOR protocol designs can generally rely on infinities and NaN as a concept being supported, but implementations may run into dark corners of their platforms when it comes to distinguishing and preserving NaN information in NaN values.¶
However, for a protocol that wants to achieve good interoperability over a wide variety of platforms, the fact that platforms differ in their support of non-finite numbers and NaNs becomes relevant. (See Section 3.2.2 below for reasons for such differences.) Protocol designs aiming for the widest possible platform support may want to implement replacements for infinite numbers and NaNs, or at least not rely on NaN information being successfully preserved during interchange.¶
Note that JSON supports neither infinite numbers nor NaN.
For protocols that are intended to work in both CBOR and JSON
representations and need an out-of-band indicator comparable to NaN, a
protocol developer might consider this (in CDDL, where float
is not
intended to be a NaN value):¶
float-with-null = float / null¶
Additional choices can be added for the infinities (e.g., false
and
true
, to stay within the CBOR simple values), if required.¶
Since null
, false
and true
have single-byte representations, the
replacement of NaN, −Infinity, and (positive) Infinity by these values can save
bytes even if JSON compatibility is not a consideration.¶
Applications that need to preserve the information in a NaN (sign bit,
quiet bit, payload) may want to replace null
with an
application-oriented representation of that information, or simply
with a (left-aligned, truncating trailing zero bytes) byte string
representing those bits:¶
float-with-nan-replacement = float / bytes¶
For JSON, the byte string can be base16- or base64-encoded, or it can be represented by an integer, preserving its left-aligned nature, or even as a (tagged) floating point value with a different exponent.¶
All floating-point numbers, including zeros and infinities, are signed. A NaN also carries a sign bit. Each of the three formats binary16, binary32, and binary64 define a fixed assignment of bits in the representation towards the sign bit, an exponent, and a "significand" (which represents the mantissa, with details sometimes depending on the specific exponent value).¶
Format | Sign bit | Exponent | Significand |
---|---|---|---|
binary16 | 1 | 5 | 10 |
binary32 | 1 | 8 | 23 |
binary64 | 1 | 11 | 52 |
Infinite numbers are represented in each format choice with a sign bit, the highest available exponent value (all ones) and all-zero significand. NaN values are represented with a sign bit, the highest available exponent value (all ones) and a non-zero significand, which carries a leading quiet bit with the rest of the bits allocated to the NaN payload.¶
To qualify as a generic encoder or decoder, a CBOR library needs to implement as much of [IEEE754] support as reasonably possible on the platform it addresses. What is reasonably possible depends on:¶
platform support for [IEEE754] numbers. If there is no such support, the generic decoder may need to resort to offering the interchanged value to the application, suitably tagged.¶
If there is partial support, it may be harder to find a good solution. This is specifically a problem for platform support that works well in most cases, but exhibits some dark corners. E.g., the implementation may support a single NaN value consistently, but not preserve NaN information present in the NaN values.¶
Where an implementation needs to convert between different floating point formats, e.g., because not all formats are fully supported by the platform, or to implement Preferred Serialization (as needed for Common Deterministic Encoding [I-D.ietf-cbor-cde]) in an encoder, conversion of NaNs in these formats is best done by operating on the bit patterns of the [IEEE754] number in the following way:¶
If the contraction is optional, e.g., for Preferred Serialization, do not perform the contraction if the removed bits in the significand truncation aren't all zero. If the contraction is required to fit into limited platform types (e.g., binary32 only), a failed truncation check indicates the loss of information and should be signaled to the application. We say a contraction "preserves the NaN information" if subsequent expansion to the original size format recreates the exact same NaN value.¶
Appendix A.1 gives additional detailed considerations for implementations that aspire to provide full support for NaNs, preserving NaN information.¶
RFC 8949 [STD94] also defines tags 4 and 5 for a representation of decimal and binary floating point numbers that is not constrained by the types provided by IEEE 754. These tags are very flexible, but this flexibility comes with a choice of ways they could be integrated into a generic encoder. Because of this flexibility, tags 4 and 5 do not define a Preferred Serialization or a deterministic encoding.¶
Section 3.2 of [I-D.ietf-cbor-time-tag] uses representations derived from the tags 4 and 5 to represent timestamps. Section 6.1 of [I-D.ietf-cbor-time-tag] lists various other tags that can be used for representing numbers for advanced arithmetic, including rational numbers in fraction form (tag 30).¶
[RFC8746] defines tags for typed arrays, i.e., arrays of numbers that all are represented in the same way. The choices defined in the [RFC8746] are all based on traditional platform number representations (unsigned integers, signed integers, IEEE 754 floating point values) and even come in little-endian and big-endian variants, often removing the need to convert the numbers from an internal to an interchange form. As conversion for interchange is not envisioned, considerations for a preferred serialization are not applicable. As the recipient may need a conversion for ingestion of the arrays, some considerations from Section 3 may apply.¶
The general security considerations for representing data in common data representation formats apply, e.g., those in Section 10 of RFC 8949 [STD94].¶
(TODO)¶
(TODO:¶
Add nan'' registration when that is ready)¶
This check list employs [BCP14] keywords to indicate interoperability requirements on implementations.¶
The following considerations apply to encoding (emitting) floating point values in a generic encoder:¶
The length of the argument is encoded in the lower 5 bits of the first byte ("ai"), which indicates half precision (binary16, ai = 0x19), single precision (binary32, ai = 0x1a) and double precision (binary64, ai = 0x1b).¶
For preferred serialization: if multiple of these encodings preserve the precision of the value to be encoded, only the shortest form of these MUST be emitted. That implies that encoders MUST support half-precision and (if there is support for more than half precision on the platform) single-precision floating point. Positive and negative infinity and zero MUST be represented in half-precision floating point.¶
NaNs MUST be supported, for all values of NaN information allowed in [IEEE754].¶
As with all floating point numbers, NaNs with payloads MUST be contracted to the shortest of double, single or half precision that preserves the NaN information.¶
The reduction is performed by removing the rightmost N bits of the payload, where N is the difference in the number of bits in the significand (mantissa) between the original format and the reduced format. The reduction is performed only (preserves the value only) if all the rightmost bits removed are zero. (This will always reduce a double or single quiet NaN with an otherwise zero NaN payload, which is typically what is returned from an operation such as 0.0/0.0, to a half-precision quiet NaN encoded as 0xf9 7e00.)¶
The following considerations apply to decoding (ingesting) floating point values in a generic decoder that supports IEEE 754 floating-point numbers:¶
Half-precision values MUST be accepted.¶
Double- and single-precision values SHOULD be accepted; leaving these out is only foreseen for decoders that need to work in exceptionally constrained environments.¶
If double-precision values are accepted, single-precision values MUST be accepted.¶
NaNs, MUST be accepted, preserving the NaN information for use of the application.¶
An IEEE-754 data item has up to 52 bits in the significand. For a NaN, the first of these bits is used to indicate whether the NaN is signalling (0) or quiet (1). The up to 51 bits in the rest of the significand are called the "NAN payload".¶
The payload’s original purpose is diagnostic information to explain why a NaN was generated by a local computation. There is no standard for the contents of a NaN payload.¶
CBOR allows NaNs with non-zero payloads to be encoded. (Due to the way infinite numbers are encoded in [IEEE754], zero-payload NaN always must be quiet NaNs.)¶
As a result, if a protocol design does not use NaNs with non-zero payloads and is using preferred serialization then NaN must be encoded as a half-precision with the quiet bit set and the payload set as 0, specifically 0xF97E00. If a design does not use NaNs with non-zero payloads and preferred serialization is not used, then the single and double precision quiet NaNs, 0xFA7FC00000 and 0xFB7FF0000000000000, may also be used.¶
NaN payloads have been in the IEEE-754 standard since 2008, but programming environments often still do not provide facilities (e.g., APIs) to make use of them. For example, in C there is the isnan() API to check if a value is a NaN, but there are no APIs to construct or access the NaN payload. The typical way to work with a NaN payload is to reinterpret the floating-point value as an unsigned integer and then use shifts and masks to unpack the IEEE-754 representation.¶
This section is primarily for CBOR library implementors.¶
CBOR attempts to limit the MUSTs about CBOR implementations in order to allow its use in a large variety of constrained use cases. For example, support for integers is not required because a protocol might need only strings. Similarly, there is no MUST that requires support of NaN and NaNs with non-zero payloads, but the recommendation here is that any generic CBOR library that supports floating-point support NaNs, preferably also with non-zero NaN payloads.¶
In most environments, there is little extra work to do to support NaN without payloads if floating-point is supported. NaNs will usually flow through as any other floating-point value.¶
Generic CBOR libraries are expected to support preferred serialization of floating-point including NaNs. For NaNs with zero payloads, this requires reducing to a half-precision NaN without a payload. This requires a few explicit extra lines of code. See the sample half-precision implementation in Appendix D of RFC 8949.¶
The implementation of preferred serialization of NaN payloads needs a few more additional lines. As with preferred serialization, NaN payloads must be reduced but only if they can be reduced without the loss of any non-zero payload bits. Programming platform provided floating-point hardware and software may or may not do this correctly for double to single conversion. The sample half-precision implementation in Appendix D of RFC 8949 only supports NaNs without payloads.¶
A double precision NaN payload contains 51 bits, a single 22 bits and a half 9 bits, in each case all but the first bit of the significand. A double precision NaN can be reduced to a single precision NaN only if the right-most 29 payload bits are zero. A single precision NaN can be reduced to a half precision NaN only if the right-most 13 payload bits are zero. A double NaN can be reduced to a half precision NaN only if the right-most 42 payload bits are zero. Note that the exponent is always all-ones for NaN, so this is simpler than the equivalent contraction of regular, non-NAN, floating-point values.¶
To implement the above, most CBOR libraries will have to reinterpret the floating point value as an unsigned integer and use shifts and masks, based in the internal representation defined in [IEEE754].¶
Testing on some CPUs has shown them to do this correctly for conversion between single and double. However, it may not be very useful to rely on platform libraries for the following reasons. First, they may provide no support at all for half-precision and half-precision is required for preferred serialization. Second, NaN payloads are a relatively recent and very specialist feature that is not usually used in interchange.¶
If platform implementation is relied upon, NaN payload reduction should be tested on each platform. Open source libraries intended to run on multiple platforms may be better off not relying on the platform.¶
The IEEE-754 numbers are given as a 64-bit (binary64) or 32-bit (binary32) unsigned integer in hex to show the bits that make up the floating-point value. All of the following are NaNs.¶
IEEE-754 Number | CBOR Preferred Serialization | Comment |
---|---|---|
0x7ff8000000000000 | 0xf97e00 | qNaN contracted from double to half |
0x7ff8000000000001 | 0xfb7ff8000000000001 | Can't be contracted because of bit set in right-side part of payload |
0x7ffffc0000000000 | 0xf97fff | 10-bit payload that can be contracted to half |
0x7ff80000000003ff | 0xfb7ff80000000003ff | right-justified payload can't be contracted |
0x7fffffffe0000000 | 0xfa7fffffff | 23-bit payload that reduces to single |
0x7ffffffff0000000 | 0xfb7ffffffff0000000 | 24-bit payload that can't be contracted |
0x7fffffffffffffff | 0xfb7fffffffffffffff | All payload bits set, can't be contracted |
0x7fc00000 | 0xf97e00 | qNaN contracted from single to half |
0x7fffe000 | 0xf97fff | single 10-bit payload that can be contracted |
0x7fbff000 | 0xfa7fbff000 | single payload that can't be contracted to 10 bits |
Laurence wrote much of the initial text about NaN processing.¶