Make tosa_verif_* tools assume they are in the ref_model repo with a
local build directory.
Update README with profile info, and updated commands to match
assumption of local build directory.
Add pytests for model file locations util.
Fix paths for non-build items with build_sub_dirs.

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I70cbcb64423b6782650166507016500268a2496d

And conformance generated folders.
And generated test numpy files.

Change-Id: I98b13d73d391e314744ea00cc35b61716c73a7d6
Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>

Ian Tayler Lessa authored Apr 24, 2025 and Jeremy Johnson committed Apr 28, 2025

OpRescale::checkTensorAttributes was missing checks related to the
relationship between the size of multiplier and shift arrays and the
value of the per_channel attribute.

This patch adds those checks and adds some tests to catch regressions in
the future.

Some cleanup of the ERROR_IF code was also done with no expected
change in behaviour.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I0b7a071af61c84fb733600a9f71f5ab7b1d754fc

cxxopts 2 is incompatible with CMake 4 so upgrade to 3.2.1

The latest version of doctest is still incompatible with CMake 4 so
patch the sources to make it compatible.
The patch command used is not pretty but it's pure CMake so doesn't rely
on any other tools being installed.

Signed-off-by: Michael Platings <michael.platings@arm.com>
Change-Id: I563d93004f73ef0e1cf7de740e7159c4aeaea1dc

Philip742 authored Mar 26, 2025 and Jeremy Johnson committed Apr 24, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: Icef432f4936ebe7059db198e409edc94f1736820

Philip742 authored Mar 26, 2025 and Jeremy Johnson committed Apr 24, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I85be7929d50772f6d5ee98c34fda98aa55912585

Philip742 authored Mar 25, 2025 and Jeremy Johnson committed Apr 24, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I560983fadd10c25e748df34b8a3d7a7c2829ad93

Philip742 authored Mar 18, 2025 and Jeremy Johnson committed Apr 24, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I0214f9d5f3ec9c2f8d9fc8700c96d0c67c07f1fb

  * Templatize Dtype to handle both InDtype and OutDtype as needed
  * move value_extend64 lambda as templated function
  * add unit test

Signed-off-by: Peng Sun <peng.sun@arm.com>
Change-Id: Ic482391efff1675744738880a495d19827817a59

Philip742 authored Mar 25, 2025 and Jeremy Johnson committed Apr 23, 2025

Platform specific file names and paths
Delete destination file before rename
Use shell to find Python interpreter in test case

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: Iec19b1533fa2bf1b982bdf0d250f438e57c48402

Change-Id: I6b8923b2047982190f71772f5d3ee8f35f450e8a
Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>

Add batch tests to CONV2D, CONV3D, DEPTHWISE_CONV2D,
TRANSPOSE_CONV2D

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: Ie9d3773ed41c03261152c7d187e22cf463091c0f

Remove exptected_return_code & expected_result_files from desc.json
Clean up readme about desc.json
Add "terminate_early" to desc.json schema and add override warning

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: Ib7f4c5365ccf58b3c6d1c7e11c4af6817931e6a2

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: Id02aead904bcd3cf8f920c7129bb0e6c4154a141

Loosen EXP compliance to have bounds calculated as:
* FP32: 3 + 2*abs(a)
* FP16/BF16: 1 + 2*abs(a)

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I921ce8630c5f113de9b2c4a39623337b7c0cc5a4

Michael Platings authored Apr 11, 2025 and Jeremy Johnson committed Apr 16, 2025

Updated to use new flatc build instruction and location

Signed-off-by: Michael Platings <michael.platings@arm.com>
Change-Id: I2958db32fec4070508dc61ee85decffee586be6f

Ian Tayler Lessa authored Apr 14, 2025 and Jeremy Johnson committed Apr 16, 2025

This makes it possible to determine whether a bounds file should be
passed to `tosa_verify` without knowledge of the details for how
different compliance modes work.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I3d7bac764ce13f636c7f82ce4a4aa595d82fb79c

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: Ic5f1fdebce54cfc246db5297776024df24e2fa89

Remove global random number test generation as not needed anymore,
now that all tests have migrated to the new RNG.

Fix concat pytest that broke due to removing global random number
gen.

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I95646ea799c07034cb730fa7be01392320c09044

Enable lazy_data_gen mode support for all CTS tests in
tosa_verif_conformance_generator

Add support for reading SHAPEs of [0] in ref model.

Change datagenerator to use serialization_lib save_npy to
write out data files in correct format.
Fix SHAPE to be output as int64 numpy format.

Clean up unused data generator functions in tosa_verif_build_tests

Add datagenerator tests for numpy format output.
Merge test_datagenerator.py tests into test_tosa_datagenerator.py

Disable cond_if_const test generation is lazy_data_gen is
requested as the work-around for loading const data late does not
work with these tests - some proper functionality to support lazy
data gen is needed.

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I9e18c960fbd636463ddf9e0bc2bf6f2ac9a760f5

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I62eee00683873a589393ef0bf413663a1a28042d

Ian Tayler Lessa authored Apr 08, 2025 and Eric Kunze committed Apr 10, 2025

Previously the error message reported the compute type internal to the
reference model as not matching, but the relevant dtype is the
serialization storage dtype. These two match in all cases except the
--precise_mode=1 reference model with floating point dtype.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I08e1e2cf5fa95b242d870af9364e635dda254c0e

Philip742 authored Apr 07, 2025 and Eric Kunze committed Apr 10, 2025

MSVC build showed up a problem in unit tests.
Correct hardcoded ulpmax value for float, not double, and use
nextafter function belonging to type.

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: If998abdff49a7b3b1098a7d82d98ce6c3e158591

udaya-ranga authored Apr 08, 2025 and udaya-ranga committed Apr 10, 2025

Coverity identified that without initialization,
`TOSAREF_TENSOR_ALLOCATE(1, float)` can potentially be problematic and
use stale data

Signed-off-by: Udaya Ranga <udaya.ranga@arm.com>
Change-Id: Ib7b3391ae5f33fcbce9fb3d98198e6adce9b44fc

Signed-off-by: Won Jeon <won.jeon@arm.com>
Change-Id: If81c415059c98f1f58be70a41d9c7cc02edd3e62

Instead use CMake's FetchContent feature.

Git submodules have many well-documented pain points.

A similar change has previously been attempted but was reverted by
revision 2d87391f because FetchContent's
git-clone was found to occasionally fail for no discernable reason.
The difference with this change is that it downloads source archives
instead of cloning Git repos. This is believed to cause less load and
therefore reduce the likelihood of getting rate-limited.
Should failures start to occur again then the sources should be
pre-downloaded and FETCHCONTENT_SOURCE_DIR_<uppercaseName> CMake
variables should be used to point at them.

Signed-off-by: Michael Platings <michael.platings@arm.com>
Change-Id: I12873b540b3fccece6e1e6bb885d574a586f0307

This patch removes references to DType_UINT8 and DType_UINT16

Signed-off-by: Tai Ly <tai.ly@arm.com>
Change-Id: Ia53def0c78f7fb17a1c32e110ee05fa7b0aa577b

Philip742 authored Mar 17, 2025 and Jeremy Johnson committed Apr 02, 2025

Large object files required to build
Multi threaded for speed

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I46a93f00a2532ba9c721f183eb0f5ca8839050fc

Philip742 authored Mar 17, 2025 and Jeremy Johnson committed Apr 02, 2025

Avoids multiple definition link error with MSVC

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I100daa4578b47890c7bfabf4808fab493dbdad8c

Philip742 authored Mar 17, 2025 and Jeremy Johnson committed Apr 02, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I1ea1306a25bbc4834271983375a91e9f3d6528a1

Philip742 authored Mar 17, 2025 and Jeremy Johnson committed Apr 02, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: Id764133242413c36966c4b82822fa20617965332

Philip742 authored Mar 17, 2025 and Jeremy Johnson committed Apr 02, 2025

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: I54a08a14c5d229ba8affc41bdef0a99d34889a9d

Philip742 authored Mar 17, 2025 and Jeremy Johnson committed Apr 02, 2025

Not portable

Signed-off-by: Philip Wilkinson <philip.wilkinson@arm.com>
Change-Id: Ie847772de825e43b7244579534356f2ce6e1ef89

openssf warning fixes incorrectly changed nextafter function for
ULPMax value.

Added extra unit tests to catch FP16/FP32 changes in future.

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I883d97ce1c71ae08226fe36ba545b391cb041b49

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

Yuvaraj Venkatesh authored Mar 26, 2025 and Eric Kunze committed Mar 31, 2025

Change-Id: If8b8a38bfe99cc0d4d7bca79f3df87c36f06ddb2
Signed-off-by: Yuvaraj Venkatesh <yuvaraj.venkatesh@arm.com>

This updates serialization_lib githash to the version
that removed Yield Op and Yield attribute

Signed-off-by: Tai Ly <tai.ly@arm.com>
Change-Id: I9e31e24ca93648e351f3a52604e2dce0492c1382

Ian Tayler Lessa authored Mar 27, 2025 and Eric Kunze committed Mar 28, 2025

The current version of the specification only allows output zero points
to be subtracted as integers, so the FP32 implementation was updated to
reflect that.

The error bound was fixed to now use the output before the zero point is
added to compute the error scale.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I5d2834b22180137880ffa905567ff1722d61c060

udaya-ranga authored Mar 26, 2025 and udaya-ranga committed Mar 27, 2025

The allocation was done to infer acc_type which can be inferred from the
given parameter.

Signed-off-by: Udaya Ranga <udaya.ranga@arm.com>
Change-Id: I41f1ec875172e6d0b69291dbeea6d25abcbce44e

Add ConvBiasShapeMismatch test to CONVOLUTION ops
Add ConcatNoInputList test to CONCAT
Re-enable StartSmallerZero, SizeSmallerEqualZero for SLICE
Add TileMultiplesOutputShapeMismatch test to TILE
Change to TransposePermsOutputShapeMismatch test to improve TRANSPOSE
Add RescalePerChannelRank0 test to RESCALE

Fix up RESCALE ERROR_IF tests to be generated without relying on
any unecessary extensions.

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I6a25881ffeb606dbf568cc017a8caa3736e04de1