Tags

TOSA Specification 1.0.0 rc0 release notes

The first release candidate for the TOSA specification

Changed the way profiles are defined. Profiles are no longer supersets of functionality. Each profile contains enough operators to define a useful set of functionality.

Added the concept of extensions. Each extension provides additional functionality to an underlying profile.
  Profiles included in the current version are:
    EXT-INT16, EXT-INT4, EXT-BF16, EXT-FP8E4M3, EXT-FP8E5M2, EXT-FFT, EXT-VARIABLE

Added SIN and COS to the elementwise operators.

Modified convolution operators to have the accumulator type specified as an attribute.
  The output types were also adjusted such that a FP16 convolution can have either FP16
  or FP32 accumulator, with the output type always as FP16. These now work similar to AVG_POOL2D.

Removed the Main Training profile.
  It did not have any operators of its own, and thus wasn't possible to justify in its current state.

FP8 types now use non-saturating rules when CAST is used to cast from a wider data type.

Added a new limit on the number of tensors allowed in a tensor list: MAX_TENSOR_LIST_SIZE

New non-normative appendix added to carry information about why choices in TOSA were made.

Cleanup and update of TOSA status noting that this is the first release candidate for TOSA 1.0

Added a new field to the XML to note what version of the specification an operator was added in. (Nothing before 1.0 is included)

Error bound fixes

Bugfixes, pseudocode clarifications

TOSA specification v0.90.0 Release Notes

Add shape operators to track shape operations for future shape inference
Continued pseudocode cleanup
Conformance testing improvements
Specify accumulator type for REDUCE_SUM
Remove redundant out_shape argument from TRANSPOSE_CONV2D

Start of development for 0.80 TOSA

Start of work on TOSA v0.90

TOSA specification v0.80.0 Release Notes

Change specification to refer to signless types consistently.
    RESCALE takes new arguments to determine whether inputs and outputs are unsigned
Add ERF activation function
Improve conformance definitions for floating-point operations
    New flag for convolution operators to determine accuracy requirements
Add stateful operators - VARIABLE / VARIABLE_READ / VARIABLE_WRITE
Add DIM operator to allow for improved shape inference calculations
Allow bias tensors to be broadcast
Various other cleanup, see the git revision history for complete details

TOSA Specification v0.70.0 release tag

Formalize the description of arguments in tosa.xml
Generate more from the tosa.xml file.
	Rank requirements for arguments
	Maximums for each TOSA level.
Updated Main Inference floating-point conformance testing
Add enumerations to TOSA specification
	Accumulator size
	Resize type
Replace uint6_t with int8_t
	Stick to more common types.
	Where uint6_t was used has range restrictions so existing implementations do not require changes
Improve description of the execution model for TOSA
Minor cleanup

TOSA 0.60.0 Specification release tag

- Add definitions for main inference conformance. Defines requirements for floating-point accuracy for TOSA operations.
- Fix rank limitations on data layout operations
- Add accumulator size for AVG_POOL2D
- Use XML as source for specification version
- Misc. cleanups

TOSA 0.50.0 release tag

- Add a new definition for TOSA levels. Levels define operator parameter
  ranges that an implementation shall support. This is distinct from a profile
  that defines the operations and data-types supported
- Tensor coordinates are clarified to be 32-bit integer values
- Operator clarifications: SLICE, RESIZE, RESCALE
- Add the ability to CAST between floating-point types

See the git log for all changes.

Specification v0.40.0 release tag

- Some of the table data for the specification has been moved into an XML
  base document. In the specification build, we use the XML to generate
  asciidoc tables to be included into the specification document. The XML
  can be used in other projects to stay in sync with changes to the
  specification.
- Convolution weights are now noted as inputs for all profiles. Previously,
  they were attributes to the convolution functions for the Base Inference
  and Main Inference profiles. Implementations may choose to optimize the
  weights if they are known at compile time, but it is no longer required to
  be compliant.
- RESIZE scale is now defined as a ratio of integers. This allows the output
  dimensions to be derived from the input dimensions without rounding.
- Remove rank limits on RESCALE and POW. This brings them in line with other
  similar operations in the specification.
- Inverted the definition of out_pad for TRANSPOSE_CONV2D. We have inverted
  the sign of the output pad values such that increasing out_pad values
  leads to increasing pad. This change also specifies a pad limit range of
  output pad values that give an output dependent on the whole input.

Tag for the release of the TOSA v0.30.0 specification

Changes in this version:
• The FFT2D and RFFT2D operators are added to the Main Inference and Main
  Training profiles for floating point operations.
• Unsigned 16-bit integer (uint16) support added. The only operator
  supporting uint16 is the RESCALE operator. It is expected that networks
  consuming uint16 input immediately use a RESCALE to convert to signed 16-bit
  values (int16).
• Floating point data types have been defined. Instead of a generic
  floating-point format, fp16, fp32, and bf16 data types have been introduced.
• Chapter one has been reorganized to try to match the overall specification
  better, including an updated description of integer operations/quantization
  within TOSA.
• The specification now describes the status of each profile separately, to
  better convey the stability level of the profile.
• Floating point edge-case behavior definition has been improved.
• In the pseudocode and data type tables for operators, if the input and
  output data types are the same, an in_out_t type is used to attempt to make
  this easier to read.
• Specifies the output formulae to define the output dimensions for tensor
  operators.
• Restricted the slope between adjacent 16-bit TABLE entries.
• Removed support for -1 dimensions in the RESHAPE operator so that all TOSA
  shapes are fully determined.
• Miscellaneous other clarifications and corrections. (See the git log for
  details)

V0.23.0 release tag

This release is primarily clarifications and improvements in error handling
and reporting.

The REQUIRE and ERROR_IF macros were introduced into the pseudocude to
describe the action when an operators required conditions are not met.
See section 3.1 of the specification for details on the new validation helpers.

Operator functionality changes:

Removed IDENTITYN, PLACEHOLDER, and RELUN. RELUN can be implmented with the
existing CLAMP operator. IDENTITYN and PLACEHOLDER can be handled with
appropriate graph actions

INTDIV - The integer division operator was renamed from DIV to clarify that it only operates on integer values.

PAD - The PAD operator now takes a value to be used as the pad constant.

Miscellaneous other clarification and cleanups were made to text and pseudocode.

v0.22.0 specification release tag

Functional changes

Add batch dimension to MATMUL.
CONCAT now accepts a variable number of input tensors.
Add 8-bit to 8-bit TABLE operation.
Add 32-bit integer DIV operation.
Removed the aint8_t data type.

Descriptive changes

Many changes to improve consistency within the specification,
including to pseudocode and argument types.
Note limits on axis where appropriate.

Tag TOSA specification for 0.21.0 release

Functional changes:
- MUL, add right shift on 32x32 bit multiply
- ARITHMETIC_RIGHT_SHIFT: add rounding
- RESIZE: add float type
- Add int48 to supported number formats
- Add SCATTER operator

Descriptive changes:
- Pseudocode cleanup, usage clarifications
- Fix various rank limitations
- Rewrite the introduction for improved clarity
- Improve the documentation for attributes and inputs
for training profile vs. inference profiles.

Tag TOSA specification for 0.20.0 release