[Aarch64][SVE] Dot product support
Alejandro Martinez Vicente
Alejandro.MartinezVicente@arm.com
Mon Apr 29 08:40:00 GMT 2019
Hi Richard,
This is the updated patch with your comments. In addition to that, I removed
vectype_in from the build_vect_cond_expr call, since it wasn't really necessary.
Alejandro
> -----Original Message-----
> From: Richard Sandiford <richard.sandiford@arm.com>
> Sent: 26 April 2019 14:29
> To: Alejandro Martinez Vicente <Alejandro.MartinezVicente@arm.com>
> Cc: GCC Patches <gcc-patches@gcc.gnu.org>; nd <nd@arm.com>; Richard
> Biener <richard.guenther@gmail.com>
> Subject: Re: [Aarch64][SVE] Dot product support
>
> Alejandro Martinez Vicente <Alejandro.MartinezVicente@arm.com> writes:
> > Hi,
> >
> > This patch does two things. For the general vectorizer, it adds
> > support to perform fully masked reductions over expressions that don't
> support masking.
> > This is achieved by using VEC_COND_EXPR where possible. At the moment
> > this is implemented for DOT_PROD_EXPR only, but the framework is there
> > to extend it to other expressions.
> >
> > Related to that, this patch adds support to vectorize dot product
> > using SVE. It also uses the new functionality to ensure that the resulting
> loop is masked.
> >
> > Given this input code:
> >
> > uint32_t
> > dotprod (uint8_t *restrict x, uint8_t *restrict y, int n) {
> > uint32_t sum = 0;
> >
> > for (int i = 0; i < n; i++)
> > {
> > sum += x[i] * y[i];
> > }
> >
> > return sum;
> > }
> >
> > The resulting SVE code is:
> >
> > 0000000000000000 <dotprod>:
> > 0: 7100005f cmp w2, #0x0
> > 4: 5400024d b.le 4c <dotprod+0x4c>
> > 8: d2800003 mov x3, #0x0 // #0
> > c: 93407c42 sxtw x2, w2
> > 10: 2538c001 mov z1.b, #0
> > 14: 25221fe0 whilelo p0.b, xzr, x2
> > 18: 2538c003 mov z3.b, #0
> > 1c: d503201f nop
> > 20: a4034002 ld1b {z2.b}, p0/z, [x0, x3]
> > 24: a4034020 ld1b {z0.b}, p0/z, [x1, x3]
> > 28: 0430e3e3 incb x3
> > 2c: 0523c000 sel z0.b, p0, z0.b, z3.b
> > 30: 25221c60 whilelo p0.b, x3, x2
> > 34: 44820401 udot z1.s, z0.b, z2.b
> > 38: 54ffff41 b.ne 20 <dotprod+0x20> // b.any
> > 3c: 2598e3e0 ptrue p0.s
> > 40: 04812021 uaddv d1, p0, z1.s
> > 44: 1e260020 fmov w0, s1
> > 48: d65f03c0 ret
> > 4c: 1e2703e1 fmov s1, wzr
> > 50: 1e260020 fmov w0, s1
> > 54: d65f03c0 ret
> >
> > Notice how udot is used inside a fully masked loop.
> >
> > I tested this patch in an aarch64 machine bootstrapping the compiler
> > and running the checks.
> >
> > I admit it is too late to merge this into gcc 9, but I'm posting it
> > anyway so it can be considered for gcc 10.
> >
> > Alejandro
> >
> >
> > gcc/Changelog:
> >
> > 2019-01-31 Alejandro Martinez <alejandro.martinezvicente@arm.com>
> >
> > * config/aarch64/aarch64-sve.md (<sur>dot_prod<vsi2qi>): Taken
> from SVE
> > ACLE branch.
> > * config/aarch64/iterators.md: Copied Vetype_fourth, VSI2QI and
> vsi2qi from
> > SVE ACLE branch.
> > * tree-vect-loop.c (use_mask_by_cond_expr_p): New function to
> check if a
> > VEC_COND_EXPR be inserted to emulate a conditional internal
> function.
> > (build_vect_cond_expr): Emit the VEC_COND_EXPR.
> > (vectorizable_reduction): Use the functions above to vectorize in a
> > fully masked loop codes that don't have a conditional internal
> > function.
> >
> > gcc/testsuite/Changelog:
> >
> > 2019-01-31 Alejandro Martinez <alejandro.martinezvicente@arm.com>
> >
> > * gcc.target/aarch64/sve/dot_1.c: New test for dot product.
> >
> > diff --git a/gcc/config/aarch64/aarch64-sve.md
> > b/gcc/config/aarch64/aarch64-sve.md
> > index 5bb3422..2779a21 100644
> > --- a/gcc/config/aarch64/aarch64-sve.md
> > +++ b/gcc/config/aarch64/aarch64-sve.md
> > @@ -3128,3 +3128,17 @@
> > DONE;
> > }
> > )
> > +
> > +;; Unpredicated DOT product.
> > +(define_insn "<sur>dot_prod<vsi2qi>"
> > + [(set (match_operand:SVE_SDI 0 "register_operand" "=w, ?&w")
> > + (plus:SVE_SDI (unspec:SVE_SDI [(match_operand:<VSI2QI> 1
> "register_operand" "w, w")
> > + (match_operand:<VSI2QI> 2
> "register_operand" "w, w")]
> > + DOTPROD)
> > + (match_operand:SVE_SDI 3 "register_operand" "0, w")))]
> > + "TARGET_SVE"
> > + "@
> > + <sur>dot\\t%0.<Vetype>, %1.<Vetype_fourth>, %2.<Vetype_fourth>
> > +
> movprfx\t%0, %3\;<sur>dot\\t%0.<Vetype>, %1.<Vetype_fourth>, %2.<Vet
> ype_fourth>"
> > + [(set_attr "movprfx" "*,yes")]
> > +)
> > diff --git a/gcc/config/aarch64/iterators.md
> > b/gcc/config/aarch64/iterators.md index 85fa161..faf588a 100644
> > --- a/gcc/config/aarch64/iterators.md
> > +++ b/gcc/config/aarch64/iterators.md
> > @@ -663,6 +663,9 @@
> > (QI "b") (HI "h")
> > (SI "s") (DI "d")])
> >
> > +;; Mode to one fourth individual element type mapping used in instruction
> DOT.
> > +(define_mode_attr Vetype_fourth [(VNx4SI "b") (VNx2DI "h")])
>
> Maybe better as:
>
> ;; Like Vetype, but map to types that are a quarter of the element size.
>
> > diff --git a/gcc/testsuite/gcc.target/aarch64/sve/dot_1.c
> > b/gcc/testsuite/gcc.target/aarch64/sve/dot_1.c
> > new file mode 100644
> > index 0000000..8ff6671
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/aarch64/sve/dot_1.c
> > @@ -0,0 +1,39 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O2 -ftree-vectorize" } */
> > +
> > +#include <stdint.h>
> > +
> > +#define DEF_DOT(TYPE1, TYPE2)
> \
> > +TYPE1 __attribute__ ((noinline, noclone)) \
> > +dot_##TYPE1##_##TYPE2 (TYPE2 *restrict x, TYPE2 *restrict y, int n) \
> > +{ \
> > + TYPE1 sum = 0; \
> > + for (int i = 0; i < n; i++) \
> > + { \
> > + sum += x[i] * y[i]; \
> > + } \
> > + return sum;
> \
> > +}
> > +
> > +DEF_DOT(uint32_t, uint8_t)
> > +DEF_DOT(int32_t, int8_t)
> > +DEF_DOT(int64_t, int16_t)
> > +
> > +/* The uint16_t->uint64_t dot product requires a casting to satisfy the C
> > + language rules. */
> > +uint64_t __attribute__ ((noinline, noclone)) dot_uint64_t_uint16_t
> > +(uint16_t *restrict x, uint16_t *restrict y, int n) {
> > + uint64_t sum = 0;
> > + for (int i = 0; i < n; i++)
> > + {
> > + sum += (unsigned int)x[i] * y[i];
> > + }
> > + return sum;
> > +}
> > +
> > +/* { dg-final { scan-assembler-times {\tudot\tz[0-9]+\.s, z[0-9]+\.b,
> > +z[0-9]+\.b\n} 1 } } */
> > +/* { dg-final { scan-assembler-times {\tsdot\tz[0-9]+\.s, z[0-9]+\.b,
> > +z[0-9]+\.b\n} 1 } } */
> > +/* { dg-final { scan-assembler-times {\tudot\tz[0-9]+\.d, z[0-9]+\.h,
> > +z[0-9]+\.h\n} 1 } } */
> > +/* { dg-final { scan-assembler-times {\tsdot\tz[0-9]+\.d, z[0-9]+\.h,
> > +z[0-9]+\.h\n} 1 } } */
> > +/* { dg-final { scan-assembler-times {\twhilelo\t} 8 } } */
> > diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c index
> > 202cab9..2278e8b 100644
> > --- a/gcc/tree-vect-loop.c
> > +++ b/gcc/tree-vect-loop.c
> > @@ -5885,6 +5885,52 @@ is_nonwrapping_integer_induction
> (stmt_vec_info stmt_vinfo, struct loop *loop)
> > <= TYPE_PRECISION (lhs_type));
> > }
> >
> > +/* If there is no conditional internal function availabe to vectorize
> > +CODE, can
>
> available
>
> > + a VEC_COND_EXPR be inserted to emulate one? */ static bool
> > +use_mask_by_cond_expr_p (enum tree_code code, internal_fn cond_fn,
> > + tree vectype_in)
>
> Need to document the COND_FN and VECTYPE_IN parameters too.
>
> > +{
> > + if (cond_fn != IFN_LAST
> > + && direct_internal_fn_supported_p (cond_fn, vectype_in,
> > + OPTIMIZE_FOR_SPEED))
> > + return false;
> > +
> > + switch (code)
> > + {
> > + case DOT_PROD_EXPR:
> > + return true;
> > +
> > + default:
> > + return false;
> > + }
> > +}
> > +
> > +/* Create a vectorized conditional version of code by using a
> > +non-conditional
>
> ...of CODE by...
>
> > + vector expression plus any VEC_COND_EXPR that may be needed. */
> > +static void build_vect_cond_expr (enum tree_code code, tree vop[3],
> > +tree mask,
> > + tree vectype_in, gimple_stmt_iterator * gsi)
>
> Need to document the parameters here too. Should be no space after
> pointer "*". (Yeah, the code's not very consistent about this.)
>
> > +{
> > + switch (code)
> > + {
> > + case DOT_PROD_EXPR:
> > + {
> > + tree zero = build_int_cst (TREE_TYPE (vectype_in), 0);
> > + zero = build_vector_from_val (vectype_in, zero);
>
> Should reduce to:
>
> tree zero = build_zero_cst (vectype_in);
>
> > + tree new_op1 = make_temp_ssa_name (vectype_in, NULL,
> "new_op1");
>
> Maybe "masked_" rather than "new_"? This name will persist in debugging
> dumps for later passes.
>
> Thanks,
> Richard
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dot_v2.patch
Type: application/octet-stream
Size: 5344 bytes
Desc: dot_v2.patch
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20190429/32364082/attachment.obj>
More information about the Gcc-patches
mailing list