This project is mirrored from https://review.mlplatform.org/tosa/specification.
Pull mirroring updated .
- Sep 25, 2024
-
-
Jerry-Ge authored- The existing complience requirement is too tight for BF16 - Add a lower bound of 20.0 ulp for the error_bound Signed-off-by:
Jerry Ge <jerry.ge@arm.com>
Change-Id: I74da351f82d3bef5b61bbf7cc6fe5df567b0aaaa
-
- Sep 23, 2024
-
-
Signed-off-by:Peng Sun <peng.sun@arm.com> Change-Id: Ieda96d00d47578174b77d7e284c56d9677a83140
-
bool is not a type TOSA defines. Signed-off-by:Kevin Petit <kevin.petit@arm.com> Change-Id: Ib2233a8bd0ae903f4c86bcd502e9619c709e294e
-
Update Sin and Cos error bounds: - Correct conversion to CalcAbsErrBound() - Allow small inputs to be flushed to zero Signed-off-by:Dominic Symes <dominic.symes@arm.com> Change-Id: I46e457678cfdb7f17b223224d54469421564e8e5
-
- If overflow is within the rounding error, then there are no accuracy requirements. - Corrected the conversion factor when scaling from accumulator precision to output precision in terms of the kernel size used for bias and variance bounds. - Clarified which data sets must pass Change-Id: I86f4f3fc2c3d24539111696e8222ee6755f01f1f Signed-off-by:
-
- Sep 20, 2024
-
-
Eric Kunze authored
The specification is currently XML and has a defined schema. These may both be changed in the future in a way that would affect tools reading the specification. Signed-off-by:Eric Kunze <eric.kunze@arm.com> Change-Id: I4d8a788a1151e9b600fde1e3c85c21e3cc796b75
-
For all calls to library functions. Change-Id: Ic7c4e9280ccaf719e182cd3c9d41eaad730f8355 Signed-off-by: -
This was the last instance. Change-Id: I2f9bc20771bccaa98386319519e8a758228ba25c Signed-off-by: -
Change-Id: I6040061c847dd1db295e1ecd5807f1e749630539 Signed-off-by: -
Change-Id: I54e25dcf9ee7404705517aac59624d2f0ad56638 Signed-off-by: -
In line with most other code. Change-Id: I825f97b05629ba231a753b381833030d5795cbbe Signed-off-by: -
As far as the formula of RFFT2D concerned, `sign_val` is superfluous. Signed-off-by:TatWai Chong <tatwai.chong@arm.com> Change-Id: I55a0a4aa675142f75e48b51b6083be049cdcce75
-
Escaped the `<=` symbol with a backslash (`\<=`) to ensure it renders correctly as "less than or equal to" instead of being interpreted as an arrow. Signed-off-by:
-
- Sep 18, 2024
-
-
Jerry-Ge authored
- For input names to be consistent within each operator category - Updated input names for TableOp and ReverseOp Signed-off-by:Jerry Ge <jerry.ge@arm.com>
Change-Id: I0385f2ececae649f9eed3950016e1f7457cd5458
-
- Sep 09, 2024
-
-
Eric Kunze authored
Signed-off-by: -
Eric Kunze authored
Signed-off-by: -
Eric Kunze authored
All other unary operators use input1 as the first input name. Align SIN and COS with those. Signed-off-by:
-
- Sep 06, 2024
-
-
Eric Kunze authored
To reduce variance errors, randomize 0.75 as +/- Adds a space to test set 3 as the parens were being lost when rendering Signed-off-by:
-
- Sep 04, 2024
-
-
Eric Kunze authored
acc_type_t is an enumeration that is used to define the acc_t type in pseudocode. Signed-off-by:
-
- Aug 14, 2024
-
-
An infinite bound is always met. Signed-off-by:
-
- Aug 03, 2024
-
-
Eric Kunze authored
FULLY_CONNECTED can be replaced by a CONV2D, including the bias addition. As a redundant operator it is not needed in TOSA. Signed-off-by: -
Eric Kunze authored
After working with the shape operators, it appears that their inclusion in TOSA is premature. We are removing them to avoid having them be a compatibility burden in the future. Signed-off-by: -
Eric Kunze authored
Arguments should be listed as inputs, attributes, outputs. Signed-off-by: -
Eric Kunze authored
The MAX_SCALE applies after the scale ratio has been reduced. Signed-off-by: -
Eric Kunze authored
List them as the underlying type instead of a tensor of rank 0 type, which should be easier for people to parse. Signed-off-by: -
Eric Kunze authored
Signed-off-by:
-
- Jul 24, 2024
-
-
Eric Kunze authored
There are places such as the shift attribute for MUL where a PRO-FP network needs to have integer and boolean constants. Signed-off-by: -
Change from BI to Integer (Int) and MI to Floating-Point (FP) Change-Id: I0d2aabdac1fa4736d6f2897cb35dfebd41fed758 Signed-off-by: -
Eric Kunze authored
Require that the logical shift amount is valid for the given data type. Signed-off-by: -
Run the verifier on the compilance data generated by the exporter Signed-off-by: -
Define a new extension `EXT-SHAPE` for (dynamic) shape inference. Includes tosa shape operators such as add_shape, concat_shape, dim, etc. except tosa `shape_t` type and `const_shape` operator particularly. Leave them in the core profiles for use by operands of shape_t types in static-shaped scenarios. Change-Id: I48c8c3faf9251354ff90a8f6c97a1f64c3c3b027 Signed-off-by: -
Verify the syntax correctness of generated compliance data Command line for verification: $tools/compliance_data_verifier.py --input out/gen/compliance.profile.meta $tools/compliance_data_verifier.py --input out/gen/compliance.extension.meta Signed-off-by: -
Parse tosa.xml and generate the required compliance metadata in the pre-defined format for profile and extension validation in MLIR Command line for generation: $tools/genspec.py --xml tosa.xml --outdir out/gen --profile Signed-off-by:
-
- Jul 11, 2024
-
-
Eric Kunze authored
Miscellaneous cleanup and clarifications across the specification. No functional changes intended in this commit. Signed-off-by: -
For maximum and minimum operations either sign of zero is permitted. Change-Id: If5ed01259a8c550f44599b2414f324a4a802e662 Signed-off-by:
-
- Jul 08, 2024
-
-
Eric Kunze authored
Signed-off-by: -
Update test set generator scaling factors so that different set values occur on calls to set_data() for each test S value. Signed-off-by: -
Signed-off-by:Thomas Preud'homme <thomas.preudhomme@arm.com> Change-Id: Ia26ffc9eae21bccdd1b48277ab494f1bc1b744c3
-
- Jun 17, 2024
-
-
TatWai Chong authored
Change-Id: Iec4e38a016d2ddeb142e9248f9c6ea2bbdedc67a Signed-off-by:
-
- Apr 19, 2024
-
-
Eric Kunze authored
Signed-off-by:
-
