Re: 回复:Re: 回复:[RISC-V] [tech-vector-ext] RISC-V Vector Spec version 1.0-rc1-20210608

Guy Lemieux

for zip, you don’t need to use vrgather. instead, use vector widening (eg vwaddu with x0) to double the element size (ensuring the MSBs are 0) for one half of the set.  
apply vslide1up and widening to the second set, then add the two sets (using original SEW).

for unzip, you can do the reverse. use narrowing for one half and vslide1dn (in SEW/2) + narrowing for the other half.

sorry, what is trn?


On Tue, Jun 15, 2021 at 12:08 AM Roger Ferrer Ibanez <roger.ferrer@...> wrote:


I agree that computing those indexes is not always trivial

Some ideas you can consider (not claiming these are the most

efficient ways)

  • reverse is not too complex: vid.v + vrsub.vx using vl as the

    scalar to subtract

  • zip is harder: start with the "halved indexes" vid +

    (0, 0, 1, 1, 2, 2, 3, 3, ...) then compute the "even elements"

    vector (0, 1, 0, 1, 0, 1, ...) and then multiply it (or if a

    power of two, shift) with the first index of the second vector

    (which maybe is vl/2 in your case). So you get 0, vl/2, 0, vl/2,

    0, vl/2, .... Then add this vector to the halved indexes so you

    get 0, vl/2, 1, 1+vl/2, 2, 2+vl/2, ...

  • unzip worst of the cases you can reverse what you did for zip

  • trn I don't have any ideas from the top of my head, /cc Romain

    Dolbeau who may recall how he worked around the cases in FFTW

    where he needed a trn-like operation

Hope this helps.

Kind regards,

On 15/6/21 8:14, Linjie Yu via wrote:

Dear  Craig and Roger,

     ​    ​Thanks a lot for providing  me  good

solutionI have tried them,  they are all good solutions of 

upsample application.

    ​    ​ But, when it comes to other applications, such  as zip/unzip,

trn,reverse and  so on.  The Index value  is 
still difficult to be initialized.

Best Regards

Damon Yu



发送时间:06/11/21 18:51:12

主题:Re: 回复:[RISC-V] [tech-vector-ext] RISC-V Vector

Spec version 1.0-rc1-20210608

Hi Linjie,

I'm not sure I understood your question. I think a vid.v (with a vl of your choice) that then you

(logical) shift right 1 bit (

would generate an index like the one you have now the "index[]" array.

This looks like it does not require to hardcode any size and

you don't have to load a materialised value from memory (you

compute it instead).

Hope this helps.

Kind regards,

On 11/6/21 9:22, Linjie Yu via wrote:

Hi, all

    I encountered a difficulty of applying

 "vrgather" instruction recently.  The details are shown blow:

    The date from source should be duplicated as pair

in a upsample application. 

     Eg:  src = [0, 1, 2, 3, 4, 5, 6, 7, 8]

           dst =  [0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5,

5, 6, 6, 7, 7, 8, 8]

     So, my relazation is:


  int inex[64] = {0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5

........,31,31};// to be compatibled of all the VLEN(128

~ 1024)

   vfloat32m2_t data = vundefined_f32m2();

    vfloat32m1_t  zero =

( vfloat32m1_t )vmv_v_i_i32m1(0);

   while(length >0)


         int gvl = vsetvl_e32m1(length);

        vuint32m2_t v_index = vle32_v_u32m2(index,


         vfloat32m1_t src_data = vle32_v_f32m1(src,


         data = vset_f32m2(data, src_data, 0);

         data  = vset_f32m2(data, zero, 1);

         vfloat32m2_t res =

vrrgather_vv_f32m2(data, v_index, gvl); 

         length -=gvl;

         src += gvl;

        vse32_v_f32m2(out, res, gvl);

        dst +=gvl;




    As shown before, the index data should be

 initialized as the max VLEN to make the code  compatibled.

So do all the

applications, that need a constant.

    I think it is

contrary to the idea of RISC-V, that one code can run

on all the RISC-V hardware.  Does anyone have a better

method ? 

Best Regards

Damon Yu



发送时间:06/09/21 14:46:30

主题:[RISC-V] [tech-vector-ext] RISC-V Vector Spec

version 1.0-rc1-20210608

I've just tagged the first release candidate for v1.0 of the vector

spec in github.  PDF attached below.

I've included the TG agreed updates and handled almost all of the

outstanding issues for v1.0.  Thanks for all the feedback.

I would appreciate if folks could read through the whole thing and

give comments over email and through Github issues.  Please also

submit PRs for typos and other wording improvements.

I'd like to try and settle most concerns over email if possible, and

assume it'll take a little while for everyone to go through the doc.

I'll tentatively schedule a vector TG meeting on Friday June 25 to go

over issues best dealt with on a call.  I'm hoping we can enter public

review at the same point in time.  Remember, we don't have to reach

agreement on all the issues before starting public review, just be OK

as a group with putting this out for public review.


-- Roger Ferrer Ibáñez - roger.ferrer@...Barcelona Supercomputing Center - Centro Nacional de Supercomputación

WARNING / LEGAL TEXT: This message is intended only for the use

of theindividual or entity to which it is addressed and may

containinformation which is privileged, confidential,

proprietary, or exemptfrom disclosure under applicable law. If

you are not the intendedrecipient or the person responsible for

delivering the message to theintended recipient, you are

strictly prohibited from disclosing,distributing, copying, or in

any way using this message. If you havereceived this

communication in error, please notify the sender anddestroy and

delete any copies you may have received.


Roger Ferrer Ibáñez - roger.ferrer@...

Barcelona Supercomputing Center - Centro Nacional de Supercomputación

Join to automatically receive all group messages.