On 12/6/20 10:46 PM, LIU Zhiwei wrote:
> For SIMD fcmla(by element), if the number of elements is less than
> the number of elements within one segment，i.e. 4H arrangement,
> we should not calculate the entire segment.
>
> Signed-off-by: LIU Zhiwei <[email protected]>
> ---
> target/arm/vec_helper.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
> index 7174030377..44b8165323 100644
> --- a/target/arm/vec_helper.c
> +++ b/target/arm/vec_helper.c
> @@ -544,6 +544,10 @@ void HELPER(gvec_fcmlah_idx)(void *vd, void *vn, void
> *vm,
> neg_real <<= 15;
> neg_imag <<= 15;
>
> + /* Adjust eltspersegment for simd 4H */
> + if (eltspersegment > elements) {
> + eltspersegment = elements;
> + }
Ok. Maybe better to fold this back to the initialization using MIN.
> for (i = 0; i < elements; i += eltspersegment) {
> float16 mr = m[H2(i + 2 * index + 0)];
> float16 mi = m[H2(i + 2 * index + 1)];
> @@ -610,6 +614,10 @@ void HELPER(gvec_fcmlas_idx)(void *vd, void *vn, void
> *vm,
> neg_real <<= 31;
> neg_imag <<= 31;
>
> + /* Adjust eltspersegment for simd 4H */
> + if (eltspersegment > elements) {
> + eltspersegment = elements;
> + }
Incorrect: this function only computes 4S.
> for (i = 0; i < elements; i += eltspersegment) {
> float32 mr = m[H4(i + 2 * index + 0)];
> float32 mi = m[H4(i + 2 * index + 1)];
>
r~
