Commits · 4e6f2d4fc2b2c3ab68941215a0b8578f5d77089c · Kleidi / KleidiAI

Jan 10, 2025

Emil Ohlsson authored Jan 10, 2025



Describe changes since 1.1.0

Signed-off-by: Emil Ohlsson <emil.ohlsson@arm.com>

Reviewed-by: Emil Ohlsson <emil.ohlsson@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

095d7d20

Dec 30, 2024

Update CHANGELOG for v1.1.0 · 83fce470

Jakub Sujak authored Dec 30, 2024



Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

83fce470

Dec 24, 2024

Add 16x4 GEMM dotprod microkernel · c607bfcf

Michael Kozlov authored Dec 24, 2024 and

Jakub Sujak committed Dec 24, 2024



- Add unit test

Signed-off-by: Michael Kozlov <michael.kozlov@arm.com>

Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Anitha Raj <anitha.raj@arm.com>
Reviewed-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Reviewed-by: Emil Ohlsson <emil.ohlsson@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

c607bfcf

Optimize F32 <- QAI8DXP (LHS) x QSI4CXP (RHS) for SME · 3a9a21ca

Anitha Raj authored Dec 24, 2024 and

Jakub Sujak committed Dec 24, 2024



* GEMM and GEMV micro-kernels to compute the matrix multiplication of dynamically quantized 8-bit integer (QAI8DX) LHS matrix and quantized 4-bit integer (QSI4CX) RHS matrix and the accumulation of the result into a single-precision (F32) output, optimized for SME2 technology.

Signed-off-by: Mohamad Najem <mohamad.najem@arm.com>

Signed-off-by: Anitha Raj <anitha.raj@arm.com>

Signed-off-by: Michael Kozlov <michael.kozlov@arm.com>

Signed-off-by: Thomas Bamelis <thomas.bamelis@arm.com>

Reviewed-by: Anitha Raj <anitha.raj@arm.com>
Reviewed-by: Anton Bondarenko <anton.bondarenko@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

3a9a21ca

Dec 13, 2024

Bump version to 1.0.0 · 2d160cf6

Emil Ohlsson authored Dec 13, 2024



Changes reported release version, and updates changelog since last
release

Signed-off-by: Emil Ohlsson <emil.ohlsson@arm.com>

Reviewed-by: Emil Ohlsson <emil.ohlsson@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

2d160cf6

Update changelog · 6dd0a4f8

Emil Ohlsson authored Dec 13, 2024 and

Felix Johnny Thomasmathibalan committed Dec 13, 2024

There are a few items missing in the changelog since last relase. This
commit updates the list with recent changes.

Signed-off-by: Emil Ohlsson <emil.ohlsson@arm.com>

Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

6dd0a4f8

Optimize scalar RHS packing function NxK F32 <- QAI8DXP x QSU4C32 · de8ce1bb

Gian Marco Iodice authored Dec 13, 2024 and

Anton Bondarenko committed Dec 13, 2024



- Optimize the generic RHS packing NxK. The performance improvement is
around ~1.5x

Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

Reviewed-by: Anton Bondarenko <anton.bondarenko@arm.com>
Approved-by: Anton Bondarenko <anton.bondarenko@arm.com>

de8ce1bb

Optimize QAI8DXP LHS quant and pack micro-kernel using Arm® Neon™ · 354ebc9f

Anitha Raj authored Dec 13, 2024 and

Jakub Sujak committed Dec 13, 2024



Signed-off-by: Anitha Raj <anitha.raj@arm.com>

Reviewed-by: Anitha Raj <anitha.raj@arm.com>
Reviewed-by: Emil Ohlsson <emil.ohlsson@arm.com>
Reviewed-by: alankelly <alankelly@google.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

354ebc9f

Dec 11, 2024

Replace usage of __fp16 type · 93a2ed93

Jakub Sujak authored Dec 11, 2024

The `__fp16` type is known to be troublesome for certain build environments. Instead use type `float` in the interface, and in the implementation use types `float16_t` (an alias for `__fp16`) for casting or `uint16_t` for pointer arithmetic.

Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

93a2ed93

Dec 02, 2024

Optimize F32 <- QSI8D32 (LHS) x QSI4C32 (RHS) for SME · ba20def9

Gian Marco Iodice authored Dec 02, 2024 and

Jakub Sujak committed Dec 02, 2024



- GEMM and GEMV Micro-kernels to compute the matrix multiplication of dynamically quantized symmetric signed 8-bit integer with per-block quantization (QSI8D32) LHS matrix and quantized symmetric 4-bit signed integer with per-block quantization (QSI4C32) RHS matrix and the accumulation of the result into a single-precision (F32) output, optimized for SME2 technology.

Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

Signed-off-by: Anitha Raj <anitha.raj@arm.com>

Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Anitha Raj <anitha.raj@arm.com>
Reviewed-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Reviewed-by: Anton Bondarenko <anton.bondarenko@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

ba20def9

Nov 29, 2024
- Update CHANGELOG for v0.5.0 · d41854ec
  Jakub Sujak authored Nov 29, 2024 and Anton Bondarenko committed Nov 29, 2024
```
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Approved-by: Anton Bondarenko <anton.bondarenko@arm.com>
```
  v0.5.0-rc1
  
  d41854ec
Nov 28, 2024

Matmul int8 micro-kernels for F32 <- QA8DX (LHS) x QS8ICX (RHS) · f51cd7ec

Gian Marco Iodice authored Nov 28, 2024 and

Jakub Sujak committed Nov 28, 2024



- Add GeMM-like micro-kernels
- Add GeMV-like micro-kernels

Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

f51cd7ec

Nov 27, 2024

Matmul int4 micro-kernel variants 16x4 and 1x4 for QSI8D32 (LHS) x QSI4C32 (RHS) -> F32 · 39ab4d69

Michael Kozlov authored Nov 27, 2024 and

Felix Johnny Thomasmathibalan committed Nov 27, 2024



Signed-off-by: Michael Kozlov <michael.kozlov@arm.com>

Signed-off-by: Anitha Raj <anitha.raj@arm.com>

Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Anitha Raj <anitha.raj@arm.com>
Reviewed-by: Anton Bondarenko <anton.bondarenko@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>
Approved-by: Anton Bondarenko <anton.bondarenko@arm.com>

39ab4d69

Nov 04, 2024

Add list of v0.4.0 changes to CHANGELOG · 7985c263

Jakub Sujak authored Nov 04, 2024

Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Reviewed-by: Anton Bondarenko <anton.bondarenko@arm.com>
Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

7985c263

Oct 16, 2024

Add SME F32 GEMV micro-kernel · 5d265fa1

Jakub Sujak authored Oct 16, 2024

Compute the Vector-Matrix multiply of F32 inputs to produce an F32 matrix, optimized using SME instructions.

Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

5d265fa1

Oct 14, 2024
- Add int4 matrix multiplication ukernels with dotprod extension · 10930855
  Anitha Raj authored Oct 14, 2024 and Felix Johnny Thomasmathibalan committed Oct 14, 2024
```
Signed-off-by: Anitha Raj <anitha.raj@arm.com>

Reviewed-by: Anitha Raj <anitha.raj@arm.com>
Approved-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
```
  10930855
Sep 27, 2024
- Mention removal of FP16 GEMV in CHANGELOG · 35e156d6
  Jakub Sujak authored Sep 27, 2024 and Felix Johnny Thomasmathibalan committed Sep 27, 2024
```
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
```
  v0.3.0-rc2
  
  35e156d6
- Extend qsi4cxp RHS packing kernels to support signed int4 inputs · 49f0ca8b
  Anitha Raj authored Sep 27, 2024 and Felix Johnny Thomasmathibalan committed Sep 27, 2024
```
Signed-off-by: Anitha Raj <anitha.raj@arm.com>

Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
```
  49f0ca8b
- Update changelog for 0.3.0 release · fab361a5
  Felix Johnny Thomasmathibalan authored Sep 27, 2024
```
Signed-off-by: Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

Approved-by: Jakub Sujak <jakub.sujak@arm.com>
```
  fab361a5
Sep 25, 2024

Revert "Add FP32 GEMV micro kernel" · 33ba646c

Jakub Sujak authored Sep 25, 2024 and

Felix Johnny Thomasmathibalan committed Sep 25, 2024

This reverts commit f4f59599

The existing FP32 GEMM micro-kernel (matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla) has a dedicated path for M=1 (a "GEMV" operation).

Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

33ba646c

Sep 12, 2024

Matmul int4 micro-kernels for QAI8DX (LHS) x QSI4C32 (RHS) -> F32 · ae4b17a6

Gian Marco Iodice authored Sep 12, 2024 and

Viet-Hoa Do committed Sep 12, 2024



- The LHS matrix is Quantized (Q) Asymmetric (A) Signed 8-bit (I8) with per-row
(DX) quantization parameters
- The RHS matrix is quantized (Q) Symmetric (S) Signed 4-bit (I4) with
per-block quantization
- The destination is F32
- Implement micro-kernels to perform the matrix multiplication
- Implement a micro-kernel to pack the RHS matrix

Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

Signed-off-by: Anitha Raj <anitha.raj@arm.com>

Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>

Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Anitha Raj <anitha.raj@arm.com>
Reviewed-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Max Ren <maxren@meta.com>
Approved-by: Viet-Hoa Do <viet-hoa.do@arm.com>

ae4b17a6

Aug 30, 2024

Add FP32 GEMV micro kernel · f4f59599

Felix Johnny Thomasmathibalan authored Aug 30, 2024 and

Jakub Sujak committed Aug 30, 2024



Signed-off-by: Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

Approved-by: Jakub Sujak <jakub.sujak@arm.com>

f4f59599

Aug 19, 2024

Bump version to v0.2.0 · 1ea3667f

Jakub Sujak authored Aug 19, 2024 and

Felix Johnny Thomasmathibalan committed Aug 19, 2024



Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

1ea3667f

Aug 16, 2024

Matmul int4 micro-kernels for QSI8D32 (LHS) x QSI4C32 (RHS) -> F32 · cf999441

Gian Marco Iodice authored Aug 16, 2024 and

Jakub Sujak committed Aug 16, 2024



* The LHS matrix is Quantized (Q) Symmetric (S) Signed 8-bit (I8) with per-block quantization (D32) quantization parameters
* The RHS matrix is Quantized (Q) Symmetric (S) Signed 4-bit (I4) with per-block quantization(C32) F16 scale factors,
* The destination is F32
* Implement micro-kernels to perform the matrix multiplication
* Implement a micro-kernel to pack the LHS and RHS matrices
* Added unit tests

Signed-off-by: Gian Marco <Iodice gianmarco.iodice@arm.com>

Signed-off-by: Anitha <Raj Anitha.Raj@arm.com>

Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>

Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Approved-by: Jakub Sujak <jakub.sujak@arm.com>

cf999441

Jul 15, 2024

Update release log for v0.1.0 · 8fda0bd9

Felix Johnny Thomasmathibalan authored Jul 15, 2024



Signed-off-by: Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

Approved-by: Anton Bondarenko <anton.bondarenko@arm.com>

8fda0bd9

Jul 05, 2024

Minor fixes · 547f7506

Jakub Sujak authored Jul 05, 2024



Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Approved-by: Anton Bondarenko <anton.bondarenko@arm.com>

547f7506

Jul 04, 2024

Add CHANGELOG.md file · 0dd5b868

Jakub Sujak authored Jul 04, 2024

Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>

Reviewed-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Approved-by: Felix Johnny Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>

0dd5b868