Skip to content

R
riscv-gcc-1
Commit 84ae7213 by Przemyslaw Wirkus Committed by Kyrylo Tkachov
Options

[arm] Implement usadv16qi and ssadv16qi standard names 


This patch implements the usadv16qi and ssadv16qi standard names for arm.

The V16QImode variant is important as it is the most commonly used pattern:
reducing vectors of bytes into an int.
The midend expects the optab to compute the absolute differences of operands 1
and 2 and reduce them while widening along the way up to SImode. So the inputs
are V16QImode and the output is V4SImode.

I've based my solution on Aarch64 usadv16qi and ssadv16qi standard names
current implementation (r260437). This solution emits below sequence of
instructions:

        VABDL.u8        tmp, op1, op2   # op1, op2 lowpart
        VABAL.u8        tmp, op1, op2   # op1, op2 highpart
        VPADAL.u16      op3, tmp

So, for the code:

$ arm-none-linux-gnueabihf-gcc -S -O3 -march=armv8-a+simd -mfpu=auto -mfloat-abi=hard usadv16qi.c -dp

#define N 1024
unsigned char pix1[N];
unsigned char pix2[N];

int
foo (void)
{
  int i_sum = 0;
  int i;
  for (i = 0; i < N; i++)
    i_sum += __builtin_abs (pix1[i] - pix2[i]);
  return i_sum;
}

we now generate on arm:
foo:
        movw    r3, #:lower16:pix2      @ 57    [c=4 l=4]  *arm_movsi_vfp/3
        movt    r3, #:upper16:pix2      @ 58    [c=4 l=4]  *arm_movt/0
        vmov.i32        q9, #0  @ v4si  @ 3     [c=4 l=4]  *neon_movv4si/2
        movw    r2, #:lower16:pix1      @ 59    [c=4 l=4]  *arm_movsi_vfp/3
        movt    r2, #:upper16:pix1      @ 60    [c=4 l=4]  *arm_movt/0
        add     r1, r3, #1024   @ 8     [c=4 l=4]  *arm_addsi3/4
.L2:
        vld1.8  {q11}, [r3]!    @ 11    [c=8 l=4]  *movmisalignv16qi_neon_load
        vld1.8  {q10}, [r2]!    @ 10    [c=8 l=4]  *movmisalignv16qi_neon_load
        cmp     r1, r3  @ 21    [c=4 l=4]  *arm_cmpsi_insn/2
        vabdl.u8        q8, d20, d22    @ 12    [c=8 l=4]  neon_vabdluv8qi
        vabal.u8        q8, d21, d23    @ 15    [c=88 l=4]  neon_vabaluv8qi
        vpadal.u16      q9, q8  @ 16    [c=8 l=4]  neon_vpadaluv8hi
        bne     .L2             @ 22    [c=16 l=4]  arm_cond_branch
        vadd.i32        d18, d18, d19   @ 24    [c=120 l=4]  quad_halves_plusv4si
        vpadd.i32       d18, d18, d18   @ 25    [c=8 l=4]  neon_vpadd_internalv2si
        vmov.32 r0, d18[0]      @ 30    [c=12 l=4]  vec_extractv2sisi/1

instead of:
foo:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        movw    r3, #:lower16:pix1
        movt    r3, #:upper16:pix1
        vmov.i32        q9, #0  @ v4si
        movw    r2, #:lower16:pix2
        movt    r2, #:upper16:pix2
        add     r1, r3, #1024
.L2:
        vld1.8  {q8}, [r3]!
        vld1.8  {q11}, [r2]!
        vmovl.u8 q10, d16
        cmp     r1, r3
        vmovl.u8 q8, d17
        vmovl.u8 q12, d22
        vmovl.u8 q11, d23
        vsub.i16        q10, q10, q12
        vsub.i16        q8, q8, q11
        vabs.s16        q10, q10
        vabs.s16        q8, q8
        vaddw.s16       q9, q9, d20
        vaddw.s16       q9, q9, d21
        vaddw.s16       q9, q9, d16
        vaddw.s16       q9, q9, d17
        bne     .L2
        vadd.i32        d18, d18, d19
        vpadd.i32       d18, d18, d18
        vmov.32 r0, d18[0]

2019-06-12  Przemyslaw Wirkus  <przemyslaw.wirkus@arm.com>

        * config/arm/iterators.md (VABAL): New int iterator.
        * config/arm/neon.md (<sup>sadv16qi): New define_expand.
        * config/arm/unspecs.md ("unspec"): Define UNSPEC_VABAL_S, UNSPEC_VABAL_U
        values.

        * gcc.target/arm/ssadv16qi.c: New test.
        * gcc.target/arm/usadv16qi.c: Likewise.

From-SVN: r272180
parent d134323b
Showing with 101 additions and 0 deletions
gcc/ChangeLog
2019-06-12 Przemyslaw Wirkus <przemyslaw.wirkus@arm.com>
* config/arm/iterators.md (VABAL): New int iterator.
* config/arm/neon.md (<sup>sadv16qi): New define_expand.
* config/arm/unspecs.md ("unspec"): Define UNSPEC_VABAL_S, UNSPEC_VABAL_U
values.
2019-06-12 Martin Liska <mliska@suse.cz>
* value-prof.c (stream_out_histogram_value): Only first value
......
gcc/config/arm/iterators.md
......@@ -341,6 +341,8 @@
(define_int_iterator VSUBHN [UNSPEC_VSUBHN UNSPEC_VRSUBHN])
(define_int_iterator VABAL [UNSPEC_VABAL_S UNSPEC_VABAL_U])
(define_int_iterator VABD [UNSPEC_VABD_S UNSPEC_VABD_U])
(define_int_iterator VABDL [UNSPEC_VABDL_S UNSPEC_VABDL_U])
......@@ -834,6 +836,7 @@
(UNSPEC_VSUBW_S "s") (UNSPEC_VSUBW_U "u")
(UNSPEC_VHSUB_S "s") (UNSPEC_VHSUB_U "u")
(UNSPEC_VQSUB_S "s") (UNSPEC_VQSUB_U "u")
(UNSPEC_VABAL_S "s") (UNSPEC_VABAL_U "u")
(UNSPEC_VABD_S "s") (UNSPEC_VABD_U "u")
(UNSPEC_VABDL_S "s") (UNSPEC_VABDL_U "u")
(UNSPEC_VMAX "s") (UNSPEC_VMAX_U "u")
......
gcc/config/arm/neon.md
......@@ -3256,6 +3256,32 @@
[(set_attr "type" "neon_arith_acc<q>")]
)
(define_expand "<sup>sadv16qi"
[(use (match_operand:V4SI 0 "register_operand"))
(unspec:V16QI [(use (match_operand:V16QI 1 "register_operand"))
(use (match_operand:V16QI 2 "register_operand"))] VABAL)
(use (match_operand:V4SI 3 "register_operand"))]
"TARGET_NEON"
{
rtx reduc = gen_reg_rtx (V8HImode);
rtx op1_highpart = gen_reg_rtx (V8QImode);
rtx op2_highpart = gen_reg_rtx (V8QImode);
emit_insn (gen_neon_vabdl<sup>v8qi (reduc,
gen_lowpart (V8QImode, operands[1]),
gen_lowpart (V8QImode, operands[2])));
emit_insn (gen_neon_vget_highv16qi (op1_highpart, operands[1]));
emit_insn (gen_neon_vget_highv16qi (op2_highpart, operands[2]));
emit_insn (gen_neon_vabal<sup>v8qi (reduc, reduc,
op1_highpart, op2_highpart));
emit_insn (gen_neon_vpadal<sup>v8hi (operands[3], operands[3], reduc));
emit_move_insn (operands[0], operands[3]);
DONE;
}
)
(define_insn "neon_v<maxmin><sup><mode>"
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
......
gcc/config/arm/unspecs.md
......@@ -200,6 +200,8 @@
UNSPEC_SHA256SU1
UNSPEC_VMULLP64
UNSPEC_LOAD_COUNT
UNSPEC_VABAL_S
UNSPEC_VABAL_U
UNSPEC_VABD_F
UNSPEC_VABD_S
UNSPEC_VABD_U
......
gcc/testsuite/ChangeLog
2019-06-12 Przemyslaw Wirkus <przemyslaw.wirkus@arm.com>
* gcc.target/arm/ssadv16qi.c: New test.
* gcc.target/arm/usadv16qi.c: Likewise.
2019-06-12 Jakub Jelinek <jakub@redhat.com>
PR c/90760
......
gcc/testsuite/gcc.target/arm/ssadv16qi.c 0 → 100644
/* { dg-do compile } */
/* { dg-additional-options "-O3 --save-temps" } */
/* { dg-require-effective-target arm_fp_ok } */
/* { dg-require-effective-target arm_neon_ok } */
/* { dg-add-options arm_neon } */
#define N 1024
signed char pix1[N], pix2[N];
int
foo (void)
{
int i_sum = 0;
int i;
for (i = 0; i < N; i++)
i_sum += __builtin_abs (pix1[i] - pix2[i]);
return i_sum;
}
/* { dg-final { scan-assembler {\tvabdl\.s8\t} } } */
/* { dg-final { scan-assembler {\tvabal\.s8\t} } } */
/* { dg-final { scan-assembler {\tvpadal\.s16\t} } } */
/* { dg-final { scan-assembler-not {\tvmovl} } } */
/* { dg-final { scan-assembler-not {\tvsub} } } */
/* { dg-final { scan-assembler-not {\tvabs} } } */
gcc/testsuite/gcc.target/arm/usadv16qi.c 0 → 100644
/* { dg-do compile } */
/* { dg-additional-options "-O3 --save-temps" } */
/* { dg-require-effective-target arm_fp_ok } */
/* { dg-require-effective-target arm_neon_ok } */
/* { dg-add-options arm_neon } */
#define N 1024
unsigned char pix1[N], pix2[N];
int
foo (void)
{
int i_sum = 0;
int i;
for (i = 0; i < N; i++)
i_sum += __builtin_abs (pix1[i] - pix2[i]);
return i_sum;
}
/* { dg-final { scan-assembler {\tvabdl\.u8\t} } } */
/* { dg-final { scan-assembler {\tvabal\.u8\t} } } */
/* { dg-final { scan-assembler {\tvpadal\.u16\t} } } */
/* { dg-final { scan-assembler-not {\tvmovl} } } */
/* { dg-final { scan-assembler-not {\tvsub} } } */
/* { dg-final { scan-assembler-not {\tvabs} } } */
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment