Skip to content

Update integrated-matrix.adoc#35

Merged
joseemoreira merged 19 commits intoriscv:integrated-matrix-extensionfrom
joseemoreira:integrated-matrix-extension
Apr 22, 2026
Merged

Update integrated-matrix.adoc#35
joseemoreira merged 19 commits intoriscv:integrated-matrix-extensionfrom
joseemoreira:integrated-matrix-extension

Conversation

@joseemoreira
Copy link
Copy Markdown
Collaborator

Update integrated-matrix.adoc
Revised the arithmetic considerations section for floating-point. Semantics is defined based on two implementation-defined parameters: G and psm. Those parameters must be defined (off-band) by the architecture as a table of (lambda, SEW, W) -> (G, psm) mappings.

This is a compromise between implementation freedom and complete architectural specification.

Please also check the "Specialized Extensions" section.

Update integrated-matrix.adoc
Revised the arithmetic considerations section for floating-point. Semantics is defined based on two implementation-defined parameters: G and psm. Those parameters must be defined (off-band) by the architecture as a table of (lambda, SEW, W) -> (G, psm) mappings.

This is a compromise between implementation freedom and complete architectural specification.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
stem:[C_{i,j} \leftarrow C_{i,j} + \sum_{k=0}^{\lambda W - 1} A_{i,k} \times B_{k,j}]

An implementation partitions the λ sub-dot-products for each output element into consecutive groups of `G` sub-dot-products.
The implementation-defined grouping factor `G` is applied separately within each `LMUL=1` step, so that a group shall not cross the boundary between two consecutive `LMUL=1` steps.
Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We didn't have this limitation before.

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What limitation you refer to?

Copy link
Copy Markdown
Collaborator

@ptomsich ptomsich left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Request dropping PSM as a normative, prescriptive specification. It is declaratory only.

Comment thread src/integrated-matrix.adoc
| `vmmacc.vv vd, vs1, vs2` | 1 | SEW | SEW
| `vwmmacc.vv vd, vs1, vs2` | 2 | SEW/2 | SEW
| `vqwmmacc.vv vd, vs1, vs2` | 4 | SEW/4 | SEW
| `vqmmacc.vv vd, vs1, vs2` | 4 | SEW/4 | SEW
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

As discussed: a pain, but ok.
We need to still mention somewhere that vqwmmacc is a valid alias, but vqmmacc is the canonical disassembly.

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi Philipp, could you please elaborate a bit? Why do we need the qw alias? I thought you found a precendet for using just q for quad-widening.

Comment thread src/integrated-matrix.adoc Outdated
Tightening language of arithmetic considerations a bit.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
Tightened language a bit.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
Updated description of `psm`.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
Tightened the description of dot-product across G groups.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
Tightened language a bit.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
Editing characters.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
typos

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
Editing characters.

Signed-off-by: Jose Moreira <jmoreira@us.ibm.com>
@joseemoreira joseemoreira merged commit 226e30e into riscv:integrated-matrix-extension Apr 22, 2026
2 of 3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants