On Fri, Mar 31, 2023 at 11:38:26AM +0800, xuhaoyue (A) wrote:
>
>
> On 2023/3/30 21:01:20, Jason Gunthorpe wrote:
> > On Thu, Mar 30, 2023 at 08:57:41PM +0800, xuhaoyue (A) wrote:
> >>
> >>
> >> On 2023/3/27 20:55:59, Jason Gunthorpe wrote:
> >>> On Mon, Mar 27, 2023 at 08:53:35PM +0800, xuhaoyue (A) wrote:
> >>>
> >>>>>> static void hns_roce_write512(uint64_t *dest, uint64_t *val)
> >>>>>> {
> >>>>>> mmio_memcpy_x64(dest, val, sizeof(struct hns_roce_rc_sq_wqe));
> >>>>>> @@ -314,7 +319,10 @@ static void hns_roce_write_dwqe(struct hns_roce_qp *qp, void *wqe)
> >>>>>> hr_reg_write(rc_sq_wqe, RCWQE_DB_SL_H, qp->sl >> HNS_ROCE_SL_SHIFT);
> >>>>>> hr_reg_write(rc_sq_wqe, RCWQE_WQE_IDX, qp->sq.head);
> >>>>>>
> >>>>>> - hns_roce_write512(qp->sq.db_reg, wqe);
> >>>>>> + if (qp->flags & HNS_ROCE_QP_CAP_SVE_DIRECT_WQE)
> >>>>>
> >>>>> Why do you need a device flag here?
> >>>>
> >>>> Our CPU die can support NEON instructions and SVE instructions,
> >>>> but some CPU dies only have SVE instructions that can accelerate our direct WQE performance.
> >>>> Therefore, we need to add such a flag bit to distinguish.
> >>>
> >>> NEON vs SVE is available to userspace already, it shouldn't come
> >>> throuhg a driver flag. You need another reason to add this flag
> >>>
> >>> The userspace should detect the right instruction to use based on the
> >>> cpu flags using the attribute stuff I pointed you at
> >>>
> >>> Jason
> >>> .
> >>>
> >>
> >> We optimized direct wqe based on different instructions for
> >> different CPUs, but the architecture of the CPUs is the same and
> >> supports both SVE and NEON instructions. We plan to use cpuid to
> >> distinguish between them. Is this more reasonable?
> >
> > Uhh, do you mean certain CPUs won't work with SVE and others won't
> > work with NEON?
> >
> > That is quite horrible
> >
> > Jason
> > .
> >
>
> No, acctually for general scenarios, our CPU supports two types of instructions, SVE and NEON.
> However, for the CPU that requires high fp64 floating-point computing power, the SVE instruction is enhanced and the NEON instruction is weakened.
Ideally the decision of what CPU instruction to use will be made by
rdma-core, using the the various schemes for dynamic link time
selection
It should apply universally to all providers
Jason
