Ready for initial release

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I7ed2dc924494133f86182a2885046d761ac8d5fa

License must be redistributed with copies of the specification

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I89783150815edf574fba8b0609d2b84697b0549a

Dominic Symes authored May 16, 2025 and Eric Kunze committed May 16, 2025

For subnormal inputs, allow the error bound to be set from
the encoded/unnormalized subnormal exponent.

Signed-off-by: Dominic Symes <dominic.symes@arm.com>
Change-Id: I24be432f40804135376cbec1b907412e40eb8bd9

Adds a lower bound of 0.5 ulp relative to 1.0 to allow implementations a larger error for input values close to 1.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: If8e09717f78690fd422da639e218e6834b754b69

The existing error bound for SIN is stricter than some implementations allow

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I673dd39ea64e96734bce52621e044279778a91ab

Added text explaining that the precision requirements are subject to change
as additional implementations are tested.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I18f898475b73484a42c93f65abc328181433d5ff

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I3fb0f3326d1fe16b8695734a75e65166e10485fe

Inputs and outputs to a TOSA graph must be of tensor type.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: Ifc23fa61ca427d6046309a700c0d8de3180b771f

VARIABLE, VARIABLE_READ, VARIABLE_WRITE operations recently
had the 'uid' attribute updated to 'name'. This patch aligns
the pseudo-code accordingly.

Signed-off-by: Luke Hutton <luke.hutton@arm.com>
Change-Id: I2198233de1b9038c7802dce94cc3c206a7c2d7f2

By increasing the allowable error, this should allow the corresponding
GLSL functions to pass when run on GPUs.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I3ddea72631f88c081052fe80ea606a6a65a7bf66

This version allows the implementation to add `output_zp` using fp32,
while previously conformance could only be met by rounding and casting
to int32_t before output_zp was added.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: Ib6019db5b402dc328acb5ca0deb1961010364600

This allows for the precision defined in the Vulkan Environment for
SPIR-V.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I567dae274b1d99dd38b32c231f29ea93c731085f

Allows for easier use of user friendly variable names.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I9dfe7538a9f7e71652584a6f9391a1552f1a7551

Luke Hutton authored Apr 03, 2025 and Eric Kunze committed Apr 08, 2025

The description should come before the precision requirements.

Change-Id: I879661498f9c248049bf70c3278488cbbdca5527
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

Change-Id: Ia0cd6fcfa95593b7d6fe4a1ff047f01e4de7bc5f
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

Remove references to the "reference implementation" and instead
describe the reference values as that calculated by fp64_t arithmetic.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I44258c2c7d7ab1904b89b44c391e2c7bdeb20982

TatWai Chong authored Mar 17, 2025 and Eric Kunze committed Mar 31, 2025

The exporter does not take the attribute argument into account but
the `initial_value` attribule of VARIABLE is extension dependent.

Change-Id: I5ef458137e8477682e623c3124b48f070381069d
Signed-off-by: TatWai Chong <tatwai.chong@arm.com>

The previous version was hard to verify and failed if the output
was clipped.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I1050ae52dd37f258d23af670cfafc6a041a898f1

Non functional change fixing typos and punctuation.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: Icaaf8763c8284cac0cf7eedcb7e1a604aa608f01

Dominic Symes authored Feb 27, 2025 and Eric Kunze committed Mar 18, 2025

Change-Id: I4b884ddf1a644892b0e8643ace48e6cd65bf1f7f
Signed-off-by: Dominic Symes <dominic.symes@arm.com>

TatWai Chong authored Jan 13, 2025 and Eric Kunze committed Mar 18, 2025

The enumerated type information was moved from `tensor-element-type`
field to `type` field.

Signed-off-by: TatWai Chong <tatwai.chong@arm.com>
Change-Id: Ia20d80d3c7ac93d0447c1ac759c73673d37adf34

Remove cond_t type from COND_IF and WHILE_LOOP, they aren't used
anywhere else in the description of these operators

Clean up the types for VARIABLE ops
  VARIABLE and VARIABLE_WRITE change var_t to in_t which represents
    the real types
  VARIABLE_READ changes var_t to out_t for the same reason

Switch the order of index_t and in_out_t to make the Supported Data
Type table consistent between SCATTER and GATHER.

Fix a few typos

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I76295d6783b7531562cd730a12f189508579eeb2

Luke Hutton authored Feb 25, 2025 and Eric Kunze committed Mar 11, 2025

The rank range was already specified in the rank column of
the table, simply adding a level check for max rank based
on this.

Signed-off-by: Luke Hutton <luke.hutton@arm.com>
Change-Id: Id463dcee66ccef0c15ccc4991db9800dd7a1f889

New rows can be added within a major version, but existing rows
cannot be removed within a major version.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I2c938c416fed2ec24ab712ea95010fda1d0d392e

The main inference precision section no longer exists. Add notes
for rounding modes when using the round_to_nearest_float and
round_to_nearest_int.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I5a2f1226e6eb61d782ce1778332c4354e553f89d

It was unclear whether it referred to the input or output type being
i32_t. It now is specific to the input type being i32_t.

This aligns with the ERROR_IFs in the pseudocode.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I348db4af96d4c2b1b2b34c8ba8791cb2e818ab45

When we added the signless behavior of RESCALE, the available
conversions was made too wide. These new ERROR_IF checks bring
back the original intent, restricting to a small set of allowed
unsigned conversions.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I83d00d9b71c85bd0e8f58ac15ace665a0ba87e02

Matches wording in REDUCE_MAX and REDUCE_MIN.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I679bc9ddb2163d1d65c955da439f07417feb761d

It is an elementwise unary operation.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I21d176152a5a582b2f4d4230692d608a0cbc35a5

Enum values currently start with 0, which differs from the value
used in the serialization library. While the values aren't
mandatory, it is less confusing to have the TOSA provided values
align.

Rename two enumerations for consistency:
nan_propagation_t to nan_propagation_mode_t
rounding_t to rounding_mode_t

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: If3c8e958351a77973b910e061f270c6cf8a6dae8

The existing variance magnitude bound for dot product operations is too strict.
Loosen the general case, and allow for further refinement by type in a later commit.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I8f97572e4e5b51b488b188b99671479799f63f3c

Kevin Petit authored Dec 18, 2024 and Eric Kunze committed Jan 13, 2025

- The shape_t and tensor_list_t types are defined globally and cannot be used as
template type names.
- Neither CUSTOM nor CONST_SHAPE require a template type be used and a single
type support entry without a template type can be used.
- Also use monospace font for type names in section that defines shape_t and
tensor_size_t.

Change-Id: Ida88451d167fedf966894f3b5b5f2945f7633860
Signed-off-by: Kevin Petit <kevin.petit@arm.com>

Kevin Petit authored Feb 08, 2024 and Eric Kunze committed Jan 13, 2025

Change-Id: I9a4042c6e9154ebdc640e2d61407bbf2cf4dda4b
Signed-off-by: Kevin Petit <kevin.petit@arm.com>

In the floating point compliance section in the spec,
there were some subtle discrepancies between the
introductory paragraphs and the pseudocode below it.
The changes in this commit aim to fix these. More
specifically:
- "fp64 arithmetic" was defined but not used. This
  phrase has now been used in the bullet below and
  in the pseudocode.
- "operation_fp64" and "operation_imp" functions
  were defined but not used elsewhere in the spec.
  These have been amended to the same name as the
  functions called in the pseudocode.
- An unecessary "." in the `output_bnd` defintion
  was removed from the pseudocode.

Signed-off-by: Luke Hutton <luke.hutton@arm.com>
Change-Id: I229bc81a7b3b577cb6c609dbaa102a77f1a8d604

Used when constructing the offsets into a tensor, these portions of
a tensor_shape are defined to be of type tensor_size_t.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I9d8d43f2ae89efe6cb9f529d3a76de073dee06ff

In the removed operators section, 'have' was
duplicated.

Change-Id: I523852af6a897d6ec89473d5c443ae2fa246571e
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

The zero points were in a different order than the other CONV ops.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: Ie5ef3a8a5726e94eeed6ce5aa0aa4ace9c7c5e1d

Also remove rank specification for some non-tensor arguments.

Signed-off-by: Eric Kunze <eric.kunze@arm.com>
Change-Id: I316213b0345b2c50e15fdb462fa210a89e36737a

Similarly to RFFT2D, this commit adds a short
description detailing how the output can be validated
for compliance.

Change-Id: Ief435be34892ef1e1b3922386a47df7d8ffb78bb
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

Won Jong Jeon authored Nov 08, 2024 and Eric Kunze committed Nov 20, 2024

- Remove isNaN and change isNaN to existing is_a_NaN
- Add missing parenthesis for normal_min and normal_frac

Signed-off-by: Won Jeon <won.jeon@arm.com>
Change-Id: I31c8974d0c5bd8f84a8f0d4ae703313649bd6f0c