Re: [PATCH RFC v2] target/arm: Implement SVE2 HISTCNT, HISTSEG

[Richard Henderson]

On 4/15/20 12:07 PM, Stephen Long wrote:
> +++ b/target/arm/sve.decode
> @@ -147,6 +147,7 @@
>                  &rprrr_esz rn=%reg_movprfx
>  @rdn_pg_rm_ra   ........ esz:2 . ra:5  ... pg:3 rm:5 rd:5 \
>                  &rprrr_esz rn=%reg_movprfx
> +@rd5_pg_rn_rm   ........ esz:2 . rm:5 ... pg:3 rn:5 rd:5       &rprr_esz

To retain the pattern this should drop the 5; there's no existing rd_pg_rn_rm
with which to conflict.


> +++ b/target/arm/sve_helper.c
> @@ -7016,3 +7016,68 @@ DO_PPZZ_MATCH(sve2_nmatch_ppzz_b, MO_8, true)
>  DO_PPZZ_MATCH(sve2_nmatch_ppzz_h, MO_16, true)
>  
>  #undef DO_PPZZ_MATCH
> +
> +#define DO_HISTCNT(NAME, TYPE, H)                                            
> \
> +void HELPER(NAME)(void *vd, void *vn, void *vm, void *vg, uint32_t desc)     
> \
> +{                                                                            
> \
> +    intptr_t i, j;                                                           
> \
> +    intptr_t opr_sz = simd_oprsz(desc) / 8;                                  
> \

Divide by sizeof(TYPE).

> +    TYPE *d = vd, *n = vn, *m = vm;                                          
> \
> +    uint8_t *pg = vg;                                                        
> \
> +    for (i = 0; i < opr_sz; ++i) {                                           
> \
> +        TYPE count = 0;                                                      
> \
> +        uint8_t pred = pg[H1(i)] >> ((i & 1) * 4);                           
> \

The indexing isn't correct.

You've got a mix of uint32_t and uint64_t here.
Indeed, it's probably simpler to not try to do both functions with a macro, but
just split them.

For uint64_t,

    uint8_t pred = pg[H1(i)];

for uint32_t,

    uint8_t pred = pg[H1(i >> 1)] >> ((i & 1) * 4);

or when i is in units of bytes instead of elements,

    uint8_t pred = pg[H1(i >> 3)] >> (i & 7);

> +        if (pred & 1) {                                                      
> \
> +            TYPE nn = n[H(i)];                                               
> \
> +            for (j = 0; j <= i; ++j) {                                       
> \
> +                uint8_t pred = pg[H1(j)] >> ((j & 1) * 4);                   
> \
> +                if (pred & 1 && nn == m[H(j)]) {                             
> \
> +                    ++count;                                                 
> \
> +                }                                                            
> \
> +            }                                                                
> \
> +        }                                                                    
> \
> +        d[H(i)] = count;                                                     
> \
> +    }                                                                        
> \
> +}
> +
> +DO_HISTCNT(sve2_histcnt_s, uint32_t, H1_2)

H4 for uint32_t when indexing by elements or H1_4 when indexing by bytes.

> +DO_HISTCNT(sve2_histcnt_d, uint64_t,     )
> +
> +#undef DO_HISTCNT
> +
> +static inline uint8_t get_count(uint8_t n, uint64_t m)
> +{
> +    int i;
> +    uint8_t count = 0;
> +
> +    for (i = 0; i < 64; i += 8) {
> +        if (n == extract64(m, i, 8)) {
> +            ++count;
> +        }
> +    }
> +    return count;
> +}

This can use the same logic as do_match2, except that you use ctpop64 at the
end to count the bits instead of turning into a boolean.

> +
> +void HELPER(sve2_histseg)(void *vd, void *vn, void *vm, uint32_t desc)
> +{
> +    intptr_t i, j;
> +    intptr_t opr_sz = simd_oprsz(desc);
> +    uint64_t *m = vm;
> +
> +    for (i = 0; i < opr_sz; i += 8) {
> +        uint64_t n = *(uint64_t *)(vn + i);
> +        uint64_t out = 0;
> +
> +        for (j = 0; j < 64; j += 8) {
> +            uint64_t m0 = *m;
> +            uint64_t m1 = *(m + 1);

The m values need to be loaded from the segment, just like MATCH.

> +
> +            uint8_t count = get_count(n >> j, m0) + get_count(n >> j, m1);
> +            out |= count << j;
> +
> +            m += 2;
You're adding 2 uint64_t per input byte, which is going to run well past the
end of the input vector.

> +static bool trans_HISTCNT(DisasContext *s, arg_rprr_esz *a)
> +{
> +    if (a->esz > 1) {
> +        return false;
> +    }

ESZ must be 2 or 3, not 0 or 1.  Note near the top:

    if size == '0x' then UNDEFINED;

thus only size == '1x' is valid, and thus size<0> does in fact select 2 vs 3
for S and D.

Missing the sve2 check.

> +    static gen_helper_gvec_4 * const fns[2] = {
> +        gen_helper_sve2_histcnt_s, gen_helper_sve2_histcnt_d
> +    };
> +    return do_zpzz_ool(s, a, fns[a->esz]);
> +}


r~