Previously we were causing an internal error in allocation instead of
signlaling an ERROR in the graph.
Also fixes the shape given to the Eigen tensor used to represent shape_t
values of rank 0.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Signed-off-by: Peng Sun <peng.sun@arm.com>
Change-Id: I682fe0360b002b04af7b08e590d4f1acb4850d5c

- Introduced a new typedef: registration_callback_v2_t, which uses const char*
  instead of std::string for domain and operator names.
- Updated MasterRegistry::register_function to include a deprecated overload that forwards
  std::string parameters to the new const char* version.
- Updated loadSharedLibs() to support both v2 and legacy v1 custom operator libraries.
  It attempts to resolve the v2 interface first, and falls back to v1 if necessary.

Change-Id: I9c0d7b768c1b9be5daeaf9617c6fd46000a32758
Signed-off-by: Peng Sun <peng.sun@arm.com>

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: Ia29ed6c2f0cc41061c58997de1f5487c900196e4

Use distrubution classes to make random number generation behave
the same on Windows and Linux
Added generate unit test to check if values match on a given platform

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: Ibf5a452e12326cf4ca5194c7ab83a9e1bdb1ee66

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I3d54dfa61186a00aaee4d548041dbdac1d509788

Jerry Zhao authored Jun 09, 2025 and Ian Tayler Lessa committed Jun 17, 2025

- Implement ewise unary operator tests covering floating-point edge cases
- Covers ABS, FLOOR and CEIL
- Adjust ABS operator implementation to correctly handle sign of zero
  inputs.
- Register new elementwise unary op test file (ewise_unary_tests.cpp) in CMakeLists.

Signed-off-by: Jerry Zhao <Jerry.Zhao@arm.com>
Signed-off-by: Ian Tayler Lessa <Ian.TaylerLessa@arm.com>
Change-Id: I03a74877afd195125dcca03dce6cd0903c8a142e

- Apply element-wise `max(abs(x), normal_min)` to handle subnormal
values and ensure numerical stability in bounds mode

- Updated ops: CONV2D, CONV3D, DEPTHWISE_CONV2D, MATMUL,
TRANSPOSE_CONV2D, FFT2D, RFFT2D, AVG_POOL2D

Co-authored-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Signed-off-by: Peng Sun <peng.sun@arm.com>
Change-Id: I5a2c3a44be647c5acb08d67e045af8250f9380bc

* Paths unsupported in load library for python 3.9
* RESCALE min/max calculations not promoted to int64
* Utils product() calculation also not promoted to large enough storage
* Fixed color printing output
* Made conformance generator build ops in sorted order for consistence
* Updated default model file locations and tool help messages
* Fixed TransposePermsOutputShapeMismatch ERROR_IF

Change-Id: I18cfa49e7016ee5a5453b5669c39e587175a15f5
Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>

Yuvaraj Venkatesh authored Jun 09, 2025 and Jeremy Johnson committed Jun 12, 2025

- Previously, SYSTEM was incorrectly applied to normal includes as well
- Adjusted CMake to restrict SYSTEM usage to external dependencies (e.g., FlatBuffers)

Signed-off-by: Yuvaraj Venkatesh <yuvaraj.venkatesh@arm.com>
Change-Id: If484b7fb8a708d571b47fb1f17c456d3e6e9849f

Only produce any CPP files in lazy data gen mode.

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I32e25eb078722e16e2d2d3c2b5f7eb84290f695d

Due to incorrect tag generation for dynamic tests causing
tags to be corrupted on other tests.

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I00188d7a06f62c41e7fedba24d151dec6479522d

- Use find_package to link Python headers and libs for MSVC
- Replace VLA with std::vector and copy buffer to py::array_t for safe memory handling
- Update setup.py to detect and copy .so or .pyd extensions based on platform

Signed-off-by: Deeptanshu Sekhri <deeptanshu.sekhri@arm.com>
Change-Id: I64e8e85eb3664201cd59b277cc605983bc5adf1c

Improved clarity and structure for the ModelRunner
Python bindings section. Added build instructions with usage example.

Signed-off-by: Deeptanshu Sekhri <deeptanshu.sekhri@arm.com>
Change-Id: I8176f83cc8a92260dea9b4cab1b5d5c6fa7ec0f0

Eirene Vlassi Pandi authored Apr 29, 2025 and Ian Tayler Lessa committed Jun 05, 2025

- Replace all instances of half_float::half with ct::float16
- Remove some special code in fpTrunc for handling bfloat16 as cfloat.h
  can do that special handling now
- Update serialization_lib submodule
- Update unit tests to match new serialization_lib version

Backport notes:
- A change to accomodate Location in unit tests was added because
  otherwise the serialization lib submodule cannot be updated

Signed-off-by: Eirene Vlassi Pandi <eirene.vlassipandi@arm.com>
Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I6374f2712e943a756322940eea753b7c2f57f943

Use a base bound value of 2 instead of 1 for SIN operations.

Refactor compliance meta data function in test generator to be
less operator specific for ABS_ERROR compliance

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I90182ce20387d5dda1d0e148f05a2edf830648fa

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: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: Id97f5c9802d7896bf10db6fef475b182ffed5c21

This patch changes to use TosaSerializationShape in
serialization_lib.
Lazy Data Generation mode for tests now embeds CONST_SHAPES
in the graph.

TOSA_MLIR_TRANSLATOR_REFSPEC: refs/changes/09/684209/28

Signed-off-by: Tai Ly <tai.ly@arm.com>
Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I754a6085b83bff04f0468e81c3d72f1221dcbde2

Yuvaraj Venkatesh authored May 27, 2025 and udaya-ranga committed May 30, 2025

CID: 3055367, Resource leak

Signed-off-by: Yuvaraj Venkatesh <yuvaraj.venkatesh@arm.com>
Change-Id: Ibfcd4146aadc930e3593bcace119543af70a1e43

Allow error bounds ranges, to support CAST ULP range of 1.0 to 0.5
Add unit tests for normal ULP with/without lower setting

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I5a8eefa39999911b61c8e04f2cde64921c53c69e

This warnings occurs only in v1.0 branch

Change-Id: I84471318fa7ecb452041b5fac9de1e2d77e51eaf
Signed-off-by: Yuvaraj Venkatesh <yuvaraj.venkatesh@arm.com>

TatWai Chong authored Apr 28, 2025 and Eric Kunze committed May 19, 2025

- update the version of tools and packages
- add high-level featue description
- create verif/README.md to introduce detailed CTS information

Change-Id: Id80810a700ddf8040954857813cce7464700f45b
Signed-off-by: TatWai Chong <tatwai.chong@arm.com>

Includes changes:
* 83c3a9c Fix warnings specific to GCC 9.4
* 43fdb52 Allow adding placeholders without data
* 3f7f423 Generate strict json by default when serializing

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: Ic3b9615197938a5ae55ff2932f95878d855031c1

Enable build tests to run in multi-processing mode
Refactor main loop of tosa_verify_conformance_generator
Improve logging for multiprocessing jobs
Change to use concurrent.futures from multiprocessing

Backport v1.0 Conflicts:
       verif/conformance/tosa_verif_conformance_generator.py

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I3b556a2d240acce4ff0e9faa8e815eba95489953

These warnings were not flagged by Clang but triggered by the GCC compiler.

GCC 9.4 throws relocation errors when linking with -pie unless all objects are compiled with -fpie.
To resolve this, replaced -pie with -fpie to ensure consistent position-independent code generation
during both compilation and linking.

Change-Id: Iba7ce1d0e84d8040dafe65c41fd553c342507cbc
Signed-off-by: Yuvaraj Venkatesh <yuvaraj.venkatesh@arm.com>

Philip742 authored May 12, 2025 and Jeremy Johnson committed May 15, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I135a43eb81592ac57f9f137117d003776b256775

Yuvaraj Venkatesh authored Apr 29, 2025 and Jeremy Johnson committed May 15, 2025

- Updated dumpTensor to use std::ostringstream instead of dynamic format strings
  with fprintf.
- Parsed g_func_config.fp_format (e.g., "0.5") into width and precision using istringstream.
- Applied std::fixed, std::setw, and std::setprecision to format the floating-point
  value safely and consistently.
- Fixed few other openssf warnings.
- Refactor dumpTensor function call.

Change-Id: I8ccbab5ae4afa38803fc053ba5f423840aba50ce
Signed-off-by: Yuvaraj Venkatesh <yuvaraj.venkatesh@arm.com>

This patch updates the serialization lib githash to include:

23f755a [serialization_lib] update tosa_serializer.py
30b9ca8 [serialization_lib] update README.md
199425b Set warnings as errors
a24fbe5 [serialization_lib] Add Op Info
9e3b662 Avoid supplying unsupported linker flags
877c7ff Use portable type with random engine
b22b01d Apply hardened flags to test/gtest subdir and fixed the warnings

Signed-off-by: Tai Ly <tai.ly@arm.com>
Change-Id: I15ca084c334595d42b7774c83222631b70d1e72a

This target was incorrectly using the `-z nodlopen`
linker option.

Change-Id: Ie111c7ac968b707eb1d23aa81117ac1895209314
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

This commit aims to reduce duplication by moving
constants that are commonly defined across the
code-base into a conventient 'config.h' header,
that can be included wherever needed.

Change-Id: Ibcf6979c37e0e664c9d9604116fa41d7619e5e43
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

Luke Hutton authored May 02, 2025 and Philip742 committed May 09, 2025

This commit aims to fix the MacOS build of the
reference model by, more generally, ensuring
unsupported linker flags are not supplied to the
selected linker.

Note: The "check_linker_flag" utility is used by
this change which requires cmake >= 3.18,
therefore, the minimum supported cmake version in
this project has been bumped from 3.16 to 3.18.

Change-Id: I17b1c176d328088554c231539ddfe3b057914547
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

Luke Hutton authored Apr 30, 2025 and Philip742 committed May 09, 2025

Previously we were drifting between LP64’s default (long) and a
hand-exported int64_t, which led to ABI/template mismatches
(“Eigen::Index == long” vs “int64_t”). This commit changes eigen's
default index type to int64_t to align with the reference models's
usage of eigen.

Change-Id: If88a6b137011c942044b6516e5a030e2db4bb92f
Signed-off-by: Luke Hutton <luke.hutton@arm.com>

- Added documentation on load_npy and save_npy behavior.
- Included usage examples showing support for both string and DType
 inputs.

Signed-off-by: Eirene Vlassi Pandi <eirene.vlassipandi@arm.com>
Change-Id: I84ec2fd385109857633e8da3eaddea03acab9d16

Fix library load
Set compatible preprocessor
Export lib entry function

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I64df73653b920cf614575519fed8fcf33e7839ea

FP_SPECIAL and RESCALE_INEXACT modes now fail with reasonable error
messages when no bounds file is provided. This follows the existing
behaviour of DOT_PRODUCT and ABS_ERROR.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: Ib5133bd6b5c1f7994d7da1f7d41f71065aa0f0d3

This improves the FetchContent cache which avoids some build issues when
changing branches and fixes builds for aarch64.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I7296559282ef93959a84bf48161b608da724255e

Michael Platings authored Apr 11, 2025 and Jeremy Johnson committed May 01, 2025

Signed-off-by: Michael Platings <michael.platings@arm.com>
Change-Id: I040377af437d3ab6546416867b0a4ef746e95918

Ian Tayler Lessa authored Apr 30, 2025 and Jeremy Johnson committed May 01, 2025

-fcf-protection=full is not compatible with aarch64 targets and
-D_FORTIFY_SOURCE=3 was introduced in GCC 12, while we need to support
older versions too.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I52f70bbbbf532672aae26b670b6f1ee8d0ebc90c

Yuvaraj Venkatesh authored Mar 28, 2025 and Jeremy Johnson committed May 01, 2025

Change-Id: I2dce52162184291fea1342b0d97301f4d5556f92
Signed-off-by: Yuvaraj Venkatesh <yuvaraj.venkatesh@arm.com>

Ian Tayler Lessa authored Apr 28, 2025 and udaya-ranga committed May 01, 2025

Update the bounds calculations to account for floating point rounding
when summing the output zero points in rescale.

Also add some unit tests for inexact rescale in the output_unsigned case
because the "signless" input value could easily be interpreted
incorrectly when converting to floating point.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: Ia2b1fba90162b78772ce15c2f4fc3a521967b4b4

Refactored:
* std::memcpy to std::copy_n
* snprintf usages to C++ std::string

Signed-off-by: Jerry Ge <jerry.ge@arm.com>
Change-Id: I6ebeea6d1be86e35d44e949d47bbfa01c51f8c8d