on 2023/12/6 13:09, Michael Meissner wrote:
> On Wed, Dec 06, 2023 at 10:22:57AM +0800, Kewen.Lin wrote:
>> I'd expect you use UNSPEC_MMA_EXTRACT to extract V16QI from the result of
>> lxvp,
>> the current define_insn_and_split "*vsx_disassemble_pair" should be able to
>> take
>> care of it further (eg: reg and regoff).
>>
>> BR,
>> Kewen
>
> With Peter's subreg patch, UNSPEC_MMA_EXTRACT would produce two move with
> eSUBREGs:
With the below details, I think you meant that even with Peter's subreg patch
which was intended to get rid of UNSPEC_MMA_EXTRACT for OOmode, we could still
have sub-optimal moves?
The proposed subreg and the current UNSPEC_MMA_EXTRACT unspec are alternatives
to extract the component from the result of lxvp. Since the latest trunk still
adopts UNSPEC_MMA_EXTRACT, I replied to Ajit with it.
>
> For a FMA type loop such as:
>
> union vector_hack2 {
> vector unsigned char vuc[2];
> vector double v[2];
> };
>
> static void
> use_mma_ld_st_normal_no_unroll (double * __restrict__ r,
> const double * __restrict__ a,
> const double * __restrict__ b,
> size_t num)
> {
> __vector_pair * __restrict__ v_r = ( __vector_pair * __restrict__) r;
> const __vector_pair * __restrict__ v_a = (const __vector_pair *
> __restrict__) a;
> const __vector_pair * __restrict__ v_b = (const __vector_pair *
> __restrict__) b;
> size_t num_vector = num / (2 * (sizeof (vector double) / sizeof (double)));
> size_t num_scalar = num % (2 * (sizeof (vector double) / sizeof (double)));
> size_t i;
> union vector_hack2 a_union;
> union vector_hack2 b_union;
> union vector_hack2 r_union;
> vector double a_hi, a_lo;
> vector double b_hi, b_lo;
> vector double r_hi, r_lo;
> union vector_hack result_hi, result_lo;
>
> #pragma GCC unroll 0
> for (i = 0; i < num_vector; i++)
> {
> __builtin_vsx_disassemble_pair (&a_union.vuc, &v_a[i]);
> __builtin_vsx_disassemble_pair (&b_union.vuc, &v_b[i]);
> __builtin_vsx_disassemble_pair (&r_union.vuc, &v_r[i]);
>
> a_hi = a_union.v[0];
> b_hi = b_union.v[0];
> r_hi = r_union.v[0];
>
> a_lo = a_union.v[1];
> b_lo = b_union.v[1];
> r_lo = r_union.v[1];
>
> result_hi.v = (a_hi * b_hi) + r_hi;
> result_lo.v = (a_lo * b_lo) + r_lo;
>
> __builtin_vsx_build_pair (&v_r[i], result_hi.vuc, result_lo.vuc);
> }
>
> if (num_scalar)
> {
> r += num_vector * (2 * (sizeof (vector double) / sizeof (double)));
> a += num_vector * (2 * (sizeof (vector double) / sizeof (double)));
> b += num_vector * (2 * (sizeof (vector double) / sizeof (double)));
>
> #pragma GCC unroll 0
> for (i = 0; i < num_scalar; i++)
> r[i] += (a[i] * b[i]);
> }
>
> return;
> }
>
> Peter's code would produce the following in the inner loop:
>
> (insn 16 15 19 4 (set (reg:OO 133 [ _43 ])
> (mem:OO (plus:DI (reg/v/f:DI 150 [ a ])
> (reg:DI 143 [ ivtmp.1088 ])) [6 MEM[(__vector_pair *)a_30(D)
> + ivtmp.1088_88 * 1]+0 S32 A128])) "p10-fma.h":3285:1 2181 {*movoo}
> (nil))
> (insn 19 16 22 4 (set (reg:OO 136 [ _48 ])
> (mem:OO (plus:DI (reg/v/f:DI 151 [ b ])
> (reg:DI 143 [ ivtmp.1088 ])) [6 MEM[(__vector_pair *)b_31(D)
> + ivtmp.1088_88 * 1]+0 S32 A128])) "p10-fma.h":3285:1 2181 {*movoo}
> (nil))
> (insn 22 19 25 4 (set (reg:OO 139 [ _53 ])
> (mem:OO (plus:DI (reg/v/f:DI 149 [ r ])
> (reg:DI 143 [ ivtmp.1088 ])) [6 MEM[(__vector_pair *)r_29(D)
> + ivtmp.1088_88 * 1]+0 S32 A128])) "p10-fma.h":3285:1 2181 {*movoo}
> (nil))
> (insn 25 22 26 4 (set (reg:V2DF 117 [ _6 ])
> (fma:V2DF (subreg:V2DF (reg:OO 136 [ _48 ]) 16)
> (subreg:V2DF (reg:OO 133 [ _43 ]) 16)
> (subreg:V2DF (reg:OO 139 [ _53 ]) 16))) "p10-fma.h":3319:35 1265
> {*vsx_fmav2df4}
> (nil))
> (insn 26 25 27 4 (set (reg:V2DF 118 [ _8 ])
> (fma:V2DF (subreg:V2DF (reg:OO 136 [ _48 ]) 0)
> (subreg:V2DF (reg:OO 133 [ _43 ]) 0)
> (subreg:V2DF (reg:OO 139 [ _53 ]) 0))) "p10-fma.h":3320:35 1265
> {*vsx_fmav2df4}
> (expr_list:REG_DEAD (reg:OO 139 [ _53 ])
> (expr_list:REG_DEAD (reg:OO 136 [ _48 ])
> (expr_list:REG_DEAD (reg:OO 133 [ _43 ])
> (nil)))))
> (insn 27 26 28 4 (set (reg:OO 142 [ _59 ])
> (unspec:OO [
> (subreg:V16QI (reg:V2DF 117 [ _6 ]) 0)
> (subreg:V16QI (reg:V2DF 118 [ _8 ]) 0)
> ] UNSPEC_VSX_ASSEMBLE)) 2183 {*vsx_assemble_pair}
> (expr_list:REG_DEAD (reg:V2DF 118 [ _8 ])
> (expr_list:REG_DEAD (reg:V2DF 117 [ _6 ])
> (nil))))
>
> Now in theory you could get ride of the UNSPEC_VSX_ASSEMBLE also using
> SUBREG's.
Agree, it looks doable, this comment seems more for Peter's subreg patch. :)
BR,
Kewen
