On Mon, Jun 27, 2022 at 7:25 AM Tamar Christina via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> > -----Original Message-----
> > From: Richard Sandiford <richard.sandif...@arm.com>
> > Sent: Thursday, June 16, 2022 7:54 PM
> > To: Tamar Christina <tamar.christ...@arm.com>
> > Cc: gcc-patches@gcc.gnu.org; nd <n...@arm.com>; Richard Earnshaw
> > <richard.earns...@arm.com>; Marcus Shawcroft
> > <marcus.shawcr...@arm.com>; Kyrylo Tkachov <kyrylo.tkac...@arm.com>
> > Subject: Re: [PATCH 1/2]AArch64 Add fallback case using sdot for usdot
> >
> > Richard Sandiford via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> > > Tamar Christina <tamar.christ...@arm.com> writes:
> > >> Hi All,
> > >>
> > >> The usdot operation is common in video encoder and decoders including
> > >> some of the most widely used ones.
> > >>
> > >> This patch adds a +dotprod version of the optab as a fallback for
> > >> when you do have sdot but not usdot available.
> > >>
> > >> The fallback works by adding a bias to the unsigned argument to
> > >> convert it to a signed value and then correcting for the bias later on.
> > >>
> > >> Essentially it relies on (x - 128)y + 128y == xy where x is unsigned
> > >> and y is signed (assuming both are 8-bit values).  Because the range
> > >> of a signed byte is only to 127 we split the bias correction into:
> > >>
> > >>    (x - 128)y + 127y + y
> > >
> > > I bet you knew this question was coming, but: this technique isn't
> > > target-specific, so wouldn't it be better to handle it in
> > > tree-vect-patterns.cc instead?
>
> Ok, so after many hours of trying I don't know how to make this work.
> DOT_PROD_EXPR is a reduction, but emitting them as additional pattern
> statement doesn't work because they'll be marked as internal_def rather than
> reduction_def.  I tried marking the new vec_stmt_info that I create 
> explicitly as
> reduction_def but this gets overwritten during analysis.
>
> I then looked into getting it as a vectorizable_operation but has this 
> obvious problems
> In that it no longer treats it as a reduction and so tries to decompose into 
> hi/lo.
>
> I then looked into treating additional patterns from  a reduction as 
> reductions themselves
> but this is obviously wrong as non-reduction statements also get marked as 
> reductions.
>
> The conclusion is that I don't think the vectorizer allows additional 
> reductions to be
> emitted from patterns.

Indeed.  DOT_PROD is a weird beast and it doesn't define which lanes are
reduced to which so it's only usable when the result is reduced to a
single lane.

An SLP pattern might work if you use reduc-plus for the reduced lanes and keep
the multiply separate?

Richard.

> > Also, how about doing (x - 128)y + 64y + 64y instead, to reduce the number
> > of hoisted constants?
> >
> > Thanks,
> > Richard
> >
> > > Thanks,
> > > Richard
> > >
> > >> Concretely for:
> > >>
> > >> #define N 480
> > >> #define SIGNEDNESS_1 unsigned
> > >> #define SIGNEDNESS_2 signed
> > >> #define SIGNEDNESS_3 signed
> > >> #define SIGNEDNESS_4 unsigned
> > >>
> > >> SIGNEDNESS_1 int __attribute__ ((noipa)) f (SIGNEDNESS_1 int res,
> > >> SIGNEDNESS_3 char *restrict a,
> > >>    SIGNEDNESS_4 char *restrict b)
> > >> {
> > >>   for (__INTPTR_TYPE__ i = 0; i < N; ++i)
> > >>     {
> > >>       int av = a[i];
> > >>       int bv = b[i];
> > >>       SIGNEDNESS_2 short mult = av * bv;
> > >>       res += mult;
> > >>     }
> > >>   return res;
> > >> }
> > >>
> > >> we generate:
> > >>
> > >>         movi    v5.16b, 0x7f
> > >>         mov     x3, 0
> > >>         movi    v4.16b, 0x1
> > >>         movi    v3.16b, 0xffffffffffffff80
> > >>         movi    v0.4s, 0
> > >> .L2:
> > >>         ldr     q2, [x2, x3]
> > >>         ldr     q1, [x1, x3]
> > >>         add     x3, x3, 16
> > >>         sub     v2.16b, v2.16b, v3.16b
> > >>         sdot    v0.4s, v2.16b, v1.16b
> > >>         sdot    v0.4s, v5.16b, v1.16b
> > >>         sdot    v0.4s, v4.16b, v1.16b
> > >>         cmp     x3, 480
> > >>         bne     .L2
> > >>
> > >> instead of:
> > >>
> > >>         movi    v0.4s, 0
> > >>         mov     x3, 0
> > >> .L2:
> > >>         ldr     q2, [x1, x3]
> > >>         ldr     q1, [x2, x3]
> > >>         add     x3, x3, 16
> > >>         sxtl    v4.8h, v2.8b
> > >>         sxtl2   v3.8h, v2.16b
> > >>         uxtl    v2.8h, v1.8b
> > >>         uxtl2   v1.8h, v1.16b
> > >>         mul     v2.8h, v2.8h, v4.8h
> > >>         mul     v1.8h, v1.8h, v3.8h
> > >>         saddw   v0.4s, v0.4s, v2.4h
> > >>         saddw2  v0.4s, v0.4s, v2.8h
> > >>         saddw   v0.4s, v0.4s, v1.4h
> > >>         saddw2  v0.4s, v0.4s, v1.8h
> > >>         cmp     x3, 480
> > >>         bne     .L2
> > >>
> > >> The new sequence is significantly faster as the operations it uses
> > >> are well optimized.  Note that execution tests are already in the mid-end
> > testsuite.
> > >>
> > >> Thanks to James Greenhalgh for the tip-off.
> > >>
> > >> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> > >>
> > >> Ok for master?
> > >>
> > >> Thanks,
> > >> Tamar
> > >>
> > >> gcc/ChangeLog:
> > >>
> > >>    * config/aarch64/aarch64-simd.md (usdot_prod<vsi2qi>): Generate
> > fallback
> > >>    or call original isns ...
> > >>    (usdot_prod<vsi2qi>_insn): ...here.
> > >>
> > >> gcc/testsuite/ChangeLog:
> > >>
> > >>    * gcc.target/aarch64/simd/vusdot-autovec-2.c: New test.
> > >>
> > >> --- inline copy of patch --
> > >> diff --git a/gcc/config/aarch64/aarch64-simd.md
> > >> b/gcc/config/aarch64/aarch64-simd.md
> > >> index
> > >>
> > cf2f4badacc594df9ecf06de3f8ea570ef9e0ff2..235a6fa371e471816284e3383e8
> > >> 564e9cf643a74 100644
> > >> --- a/gcc/config/aarch64/aarch64-simd.md
> > >> +++ b/gcc/config/aarch64/aarch64-simd.md
> > >> @@ -623,7 +623,7 @@ (define_insn "<sur>dot_prod<vsi2qi>"
> > >>
> > >>  ;; These instructions map to the __builtins for the Armv8.6-a I8MM
> > >> usdot  ;; (vector) Dot Product operation and the vectorized optab.
> > >> -(define_insn "usdot_prod<vsi2qi>"
> > >> +(define_insn "usdot_prod<vsi2qi>_insn"
> > >>    [(set (match_operand:VS 0 "register_operand" "=w")
> > >>    (plus:VS
> > >>      (unspec:VS [(match_operand:<VSI2QI> 1 "register_operand" "w")
> > @@
> > >> -635,6 +635,43 @@ (define_insn "usdot_prod<vsi2qi>"
> > >>    [(set_attr "type" "neon_dot<q>")]
> > >>  )
> > >>
> > >> +;; usdot auto-vec fallback code
> > >> +(define_expand "usdot_prod<vsi2qi>"
> > >> +  [(set (match_operand:VS 0 "register_operand")
> > >> +  (plus:VS
> > >> +    (unspec:VS [(match_operand:<VSI2QI> 1 "register_operand")
> > >> +                (match_operand:<VSI2QI> 2 "register_operand")]
> > >> +    UNSPEC_USDOT)
> > >> +    (match_operand:VS 3 "register_operand")))]
> > >> +  "TARGET_DOTPROD || TARGET_I8MM"
> > >> +{
> > >> +  if (TARGET_I8MM)
> > >> +    {
> > >> +      emit_insn (gen_usdot_prod<vsi2qi>_insn (operands[0], operands[1],
> > >> +                                        operands[2], operands[3]));
> > >> +      DONE;
> > >> +    }
> > >> +
> > >> +  machine_mode elemmode = GET_MODE_INNER (<VSI2QI>mode);
> > >> +  HOST_WIDE_INT val = 1 << (GET_MODE_BITSIZE
> > (elemmode).to_constant
> > >> +() - 1);
> > >> +  rtx signbit = gen_int_mode (val, elemmode);
> > >> +  rtx t1 = gen_reg_rtx (<MODE>mode);
> > >> +  rtx t2 = gen_reg_rtx (<MODE>mode);
> > >> +  rtx tmp = gen_reg_rtx (<VSI2QI>mode);
> > >> +  rtx c1 = gen_const_vec_duplicate (<VSI2QI>mode,
> > >> +                              gen_int_mode (val - 1, elemmode));
> > >> +  rtx c2 = gen_const_vec_duplicate (<VSI2QI>mode, gen_int_mode (1,
> > >> +elemmode));
> > >> +  rtx dup = gen_const_vec_duplicate (<VSI2QI>mode, signbit);
> > >> +  c1 = force_reg (<VSI2QI>mode, c1);
> > >> +  c2 = force_reg (<VSI2QI>mode, c2);
> > >> +  dup = force_reg (<VSI2QI>mode, dup);
> > >> +  emit_insn (gen_sub<vsi2qi>3 (tmp, operands[1], dup));
> > >> +  emit_insn (gen_sdot_prod<vsi2qi> (t1, tmp, operands[2],
> > >> +operands[3]));
> > >> +  emit_insn (gen_sdot_prod<vsi2qi> (t2, c1, operands[2], t1));
> > >> +  emit_insn (gen_sdot_prod<vsi2qi> (operands[0], c2, operands[2],
> > >> +t2));
> > >> +  DONE;
> > >> +})
> > >> +
> > >>  ;; These instructions map to the __builtins for the Dot Product  ;;
> > >> indexed operations.
> > >>  (define_insn "aarch64_<sur>dot_lane<vsi2qi>"
> > >> diff --git a/gcc/testsuite/gcc.target/aarch64/simd/vusdot-autovec-2.c
> > >> b/gcc/testsuite/gcc.target/aarch64/simd/vusdot-autovec-2.c
> > >> new file mode 100644
> > >> index
> > >>
> > 0000000000000000000000000000000000000000..acd8e36209690386d021df72f1
> > 4
> > >> 67a696750ac3e
> > >> --- /dev/null
> > >> +++ b/gcc/testsuite/gcc.target/aarch64/simd/vusdot-autovec-2.c
> > >> @@ -0,0 +1,25 @@
> > >> +/* { dg-do compile } */
> > >> +/* { dg-options "-O3 -march=armv8.2-a+noi8mm+dotprod" } */
> > >> +
> > >> +#define N 480
> > >> +#define SIGNEDNESS_1 unsigned
> > >> +#define SIGNEDNESS_2 signed
> > >> +#define SIGNEDNESS_3 signed
> > >> +#define SIGNEDNESS_4 unsigned
> > >> +
> > >> +SIGNEDNESS_1 int __attribute__ ((noipa)) f (SIGNEDNESS_1 int res,
> > >> +SIGNEDNESS_3 char *restrict a,
> > >> +   SIGNEDNESS_4 char *restrict b)
> > >> +{
> > >> +  for (__INTPTR_TYPE__ i = 0; i < N; ++i)
> > >> +    {
> > >> +      int av = a[i];
> > >> +      int bv = b[i];
> > >> +      SIGNEDNESS_2 short mult = av * bv;
> > >> +      res += mult;
> > >> +    }
> > >> +  return res;
> > >> +}
> > >> +
> > >> +/* { dg-final { scan-assembler-not {\tusdot\t} } } */
> > >> +/* { dg-final { scan-assembler-times {\tsdot\t} 3 } } */

Reply via email to