Skip to content

Add floating point matrix multiply-accumulate widening intrinsics #418

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
vhscampos merged 2 commits into from
Oct 24, 2025
Merged

Add floating point matrix multiply-accumulate widening intrinsics #418

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5574e09
Add floating point matrix multiply-accumulate widening intrinsics
amilendra Sep 4, 2025
93e2348
Address review comments on the Neon FMMLA intrinsics
amilendra Oct 16, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
50 changes: 50 additions & 0 deletions main/acle.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -465,6 +465,8 @@ Armv8.4-A [[ARMARMv84]](#ARMARMv84). Support is added for the Dot Product intrin

* Added feature test macro for FEAT_SSVE_FEXPA.
* Added feature test macro for FEAT_CSSC.
* Added support for modal 8-bit floating point matrix multiply-accumulate widening intrinsics.
* Added support for 16-bit floating point matrix multiply-accumulate widening intrinsics.

### References

Expand Down Expand Up @@ -2346,6 +2348,26 @@ is hardware support for the SVE forms of these instructions and if the
associated ACLE intrinsics are available. This implies that
`__ARM_FEATURE_MATMUL_INT8` and `__ARM_FEATURE_SVE` are both nonzero.

##### Multiplication of modal 8-bit floating-point matrices

This section is in
[**Alpha** state](#current-status-and-anticipated-changes) and might change or be
extended in the future.

`__ARM_FEATURE_F8F16MM` is defined to `1` if there is hardware support
for the NEON and SVE modal 8-bit floating-point matrix multiply-accumulate to half-precision (FEAT_F8F16MM)
instructions and if the associated ACLE intrinsics are available.

`__ARM_FEATURE_F8F32MM` is defined to `1` if there is hardware support
for the NEON and SVE modal 8-bit floating-point matrix multiply-accumulate to single-precision (FEAT_F8F32MM)
instructions and if the associated ACLE intrinsics are available.

##### Multiplication of 16-bit floating-point matrices

`__ARM_FEATURE_SVE_F16F32MM` is defined to `1` if there is hardware support
for the SVE 16-bit floating-point to 32-bit floating-point matrix multiply and add
(FEAT_SVE_F16F32MM) instructions and if the associated ACLE intrinsics are available.

##### Multiplication of 32-bit floating-point matrices

`__ARM_FEATURE_SVE_MATMUL_FP32` is defined to `1` if there is hardware support
Expand Down Expand Up @@ -2637,6 +2659,9 @@ be found in [[BA]](#BA).
| [`__ARM_FEATURE_SVE_BITS`](#scalable-vector-extension-sve) | The number of bits in an SVE vector, when known in advance | 256 |
| [`__ARM_FEATURE_SVE_MATMUL_FP32`](#multiplication-of-32-bit-floating-point-matrices) | 32-bit floating-point matrix multiply extension (FEAT_F32MM) | 1 |
| [`__ARM_FEATURE_SVE_MATMUL_FP64`](#multiplication-of-64-bit-floating-point-matrices) | 64-bit floating-point matrix multiply extension (FEAT_F64MM) | 1 |
| [`__ARM_FEATURE_F8F16MM`](#multiplication-of-modal-8-bit-floating-point-matrices) | Modal 8-bit floating-point matrix multiply-accumulate to half-precision extension (FEAT_F8F16MM) | 1 |
| [`__ARM_FEATURE_F8F32MM`](#multiplication-of-modal-8-bit-floating-point-matrices) | Modal 8-bit floating-point matrix multiply-accumulate to single-precision extension (FEAT_F8F32MM) | 1 |
| [`__ARM_FEATURE_SVE_F16F32MM`](#multiplication-of-16-bit-floating-point-matrices) | 16-bit floating-point matrix multiply-accumulate to single-precision extension (FEAT_SVE_F16F32MM) | 1 |
| [`__ARM_FEATURE_SVE_MATMUL_INT8`](#multiplication-of-8-bit-integer-matrices) | SVE support for the integer matrix multiply extension (FEAT_I8MM) | 1 |
| [`__ARM_FEATURE_SVE_PREDICATE_OPERATORS`](#scalable-vector-extension-sve) | Level of support for C and C++ operators on SVE vector types | 1 |
| [`__ARM_FEATURE_SVE_VECTOR_OPERATORS`](#scalable-vector-extension-sve) | Level of support for C and C++ operators on SVE predicate types | 1 |
Expand Down Expand Up @@ -9374,6 +9399,31 @@ BFloat16 floating-point multiply vectors.
uint64_t imm_idx);
```

### SVE2 floating-point matrix multiply-accumulate instructions.

#### FMMLA (widening, FP8 to FP16)

Modal 8-bit floating-point matrix multiply-accumulate to half-precision.
```c
// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_F8F16MM)
svfloat16_t svmmla[_f16_mf8]_fpm(svfloat16_t zda, svmfloat8_t zn, svmfloat8_t zm, fpm_t fpm);
```

#### FMMLA (widening, FP8 to FP32)

Modal 8-bit floating-point matrix multiply-accumulate to single-precision.
```c
// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_F8F32MM)
svfloat32_t svmmla[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn, svmfloat8_t zm, fpm_t fpm);
```
#### FMMLA (widening, FP16 to FP32)

16-bit floating-point matrix multiply-accumulate to single-precision.
```c
// Only if __ARM_FEATURE_SVE_F16F32MM
svfloat32_t svmmla[_f32_f16](svfloat32_t zda, svfloat16_t zn, svfloat16_t zm);
```

### SVE2.1 instruction intrinsics

The specification for SVE2.1 is in
Expand Down
11 changes: 11 additions & 0 deletions neon_intrinsics/advsimd.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6175,3 +6175,14 @@ The intrinsics in this section are guarded by the macro ``__ARM_NEON``.
| <code>float32x4_t <a href="https://developer.arm.com/architectures/instruction-sets/intrinsics/vmlalltbq_laneq_f32_mf8_fpm" target="_blank">vmlalltbq_laneq_f32_mf8_fpm</a>(<br>&nbsp;&nbsp;&nbsp;&nbsp; float32x4_t vd,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t vn,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t vm,<br>&nbsp;&nbsp;&nbsp;&nbsp; const int lane,<br>&nbsp;&nbsp;&nbsp;&nbsp; fpm_t fpm)</code> | `vd -> Vd.4S`<br>`vm -> Vn.16B`<br>`vm -> Vm.B`<br>`0 <= lane <= 15` | `FMLALLBB Vd.4S, Vn.16B, Vm.B[lane]` | `Vd.4S -> result` | `A64` |
| <code>float32x4_t <a href="https://developer.arm.com/architectures/instruction-sets/intrinsics/vmlallttq_lane_f32_mf8_fpm" target="_blank">vmlallttq_lane_f32_mf8_fpm</a>(<br>&nbsp;&nbsp;&nbsp;&nbsp; float32x4_t vd,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t vn,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x8_t vm,<br>&nbsp;&nbsp;&nbsp;&nbsp; const int lane,<br>&nbsp;&nbsp;&nbsp;&nbsp; fpm_t fpm)</code> | `vd -> Vd.4S`<br>`vm -> Vn.16B`<br>`vm -> Vm.B`<br>`0 <= lane <= 7` | `FMLALLBB Vd.4S, Vn.16B, Vm.B[lane]` | `Vd.4S -> result` | `A64` |
| <code>float32x4_t <a href="https://developer.arm.com/architectures/instruction-sets/intrinsics/vmlallttq_laneq_f32_mf8_fpm" target="_blank">vmlallttq_laneq_f32_mf8_fpm</a>(<br>&nbsp;&nbsp;&nbsp;&nbsp; float32x4_t vd,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t vn,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t vm,<br>&nbsp;&nbsp;&nbsp;&nbsp; const int lane,<br>&nbsp;&nbsp;&nbsp;&nbsp; fpm_t fpm)</code> | `vd -> Vd.4S`<br>`vm -> Vn.16B`<br>`vm -> Vm.B`<br>`0 <= lane <= 15` | `FMLALLBB Vd.4S, Vn.16B, Vm.B[lane]` | `Vd.4S -> result` | `A64` |

## Matrix multiplication intrinsics from Armv9.6-A

### Vector arithmetic

#### Matrix multiply

| Intrinsic | Argument preparation | AArch64 Instruction | Result | Supported architectures |
|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------|-------------------------------|-------------------|---------------------------|
| <code>float16x8_t <a href="https://developer.arm.com/architectures/instruction-sets/intrinsics/vmmlaq_f16_mf8" target="_blank">vmmlaq_f16_mf8</a>(<br>&nbsp;&nbsp;&nbsp;&nbsp; float16x8_t r,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t a,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t b,<br>&nbsp;&nbsp;&nbsp;&nbsp; fpm_t fpm)</code> | `r -> Vd.4H`<br>`a -> Vn.16B`<br>`b -> Vm.16B` | `FMMLA Vd.4H, Vn.16B, Vm.16B` | `Vd.4H -> result` | `A64` |
| <code>float32x4_t <a href="https://developer.arm.com/architectures/instruction-sets/intrinsics/vmmlaq_f32_mf8" target="_blank">vmmlaq_f32_mf8</a>(<br>&nbsp;&nbsp;&nbsp;&nbsp; float32x4_t r,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t a,<br>&nbsp;&nbsp;&nbsp;&nbsp; mfloat8x16_t b,<br>&nbsp;&nbsp;&nbsp;&nbsp; fpm_t fpm)</code> | `r -> Vd.4S`<br>`a -> Vn.16B`<br>`b -> Vm.16B` | `FMMLA Vd.4S, Vn.16B, Vm.16B` | `Vd.4S -> result` | `A64` |
4 changes: 4 additions & 0 deletions tools/intrinsic_db/advsimd.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4810,3 +4810,7 @@ float32x4_t vmlalltbq_lane_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x
float32x4_t vmlalltbq_laneq_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm) vd -> Vd.4S;vm -> Vn.16B; vm -> Vm.B; 0 <= lane <= 15 FMLALLBB Vd.4S, Vn.16B, Vm.B[lane] Vd.4S -> result A64
float32x4_t vmlallttq_lane_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x8_t vm, __builtin_constant_p(lane), fpm_t fpm) vd -> Vd.4S;vm -> Vn.16B; vm -> Vm.B; 0 <= lane <= 7 FMLALLBB Vd.4S, Vn.16B, Vm.B[lane] Vd.4S -> result A64
float32x4_t vmlallttq_laneq_f32_mf8_fpm(float32x4_t vd, mfloat8x16_t vn, mfloat8x16_t vm, __builtin_constant_p(lane), fpm_t fpm) vd -> Vd.4S;vm -> Vn.16B; vm -> Vm.B; 0 <= lane <= 15 FMLALLBB Vd.4S, Vn.16B, Vm.B[lane] Vd.4S -> result A64

<SECTION> Matrix multiplication intrinsics from Armv9.6-A
float16x8_t vmmlaq_f16_mf8(float16x8_t r, mfloat8x16_t a, mfloat8x16_t b, fpm_t fpm) r -> Vd.4H;a -> Vn.16B;b -> Vm.16B FMMLA Vd.4H, Vn.16B, Vm.16B Vd.4H -> result A64
Copy link
Contributor

@AlfieRichardsArm AlfieRichardsArm Oct 28, 2025

Choose a reason for hiding this comment

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

@jthackray following on from
#409 (comment)

That was merged early accidentally, then reverted.

This is the genuine PR and this issue seems to have been fixed.

Copy link
Contributor

@Lukacma Lukacma Oct 29, 2025

Choose a reason for hiding this comment

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

It is missing _fpm suffix here though. No matter how many times we look at it we always miss smth in the ACLE :(.

Copy link
Contributor

@AlfieRichardsArm AlfieRichardsArm Oct 29, 2025

Choose a reason for hiding this comment

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

Ahh yeah agreed :(
@amilendra are you able to make a fix? I am happy to if youre busy

Copy link
Contributor Author

@amilendra amilendra Oct 29, 2025

Choose a reason for hiding this comment

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

Feel free to fix it. Thanks

Copy link
Contributor

@Lukacma Lukacma Oct 29, 2025

Choose a reason for hiding this comment

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

Suggested change
float16x8_t vmmlaq_f16_mf8(float16x8_t r, mfloat8x16_t a, mfloat8x16_t b, fpm_t fpm) r -> Vd.4H;a -> Vn.16B;b -> Vm.16B FMMLA Vd.4H, Vn.16B, Vm.16B Vd.4H -> result A64
float16x8_t vmmlaq_f16_mf8(float16x8_t r, mfloat8x16_t a, mfloat8x16_t b, fpm_t fpm) r -> Vd.8H;a -> Vn.16B;b -> Vm.16B FMMLA Vd.8H, Vn.16B, Vm.16B Vd.4H -> result A64

Also this

float32x4_t vmmlaq_f32_mf8(float32x4_t r, mfloat8x16_t a, mfloat8x16_t b, fpm_t fpm) r -> Vd.4S;a -> Vn.16B;b -> Vm.16B FMMLA Vd.4S, Vn.16B, Vm.16B Vd.4S -> result A64
2 changes: 2 additions & 0 deletions tools/intrinsic_db/advsimd_classification.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4697,3 +4697,5 @@ vmlalltbq_lane_f32_mf8_fpm Vector arithmetic|Multiply|Multiply-accumulate and wi
vmlalltbq_laneq_f32_mf8_fpm Vector arithmetic|Multiply|Multiply-accumulate and widen
vmlallttq_lane_f32_mf8_fpm Vector arithmetic|Multiply|Multiply-accumulate and widen
vmlallttq_laneq_f32_mf8_fpm Vector arithmetic|Multiply|Multiply-accumulate and widen
vmmlaq_f16_mf8 Vector arithmetic|Matrix multiply
vmmlaq_f32_mf8 Vector arithmetic|Matrix multiply