From: Uros Bizjak <ubizjak@gmail.com>
To: liuhongt <hongtao.liu@intel.com>
Cc: gcc-patches@gcc.gnu.org
Subject: Re: [PATCH] Support 64-bit vectorization for single-precision floating rounding operation.
Date: Tue, 20 Sep 2022 08:35:38 +0200 [thread overview]
Message-ID: <CAFULd4a9n0xV4XbNWQt9jo9uyREE6sr1d20GVyQGKus8KSnx8w@mail.gmail.com> (raw)
In-Reply-To: <20220920021235.2634748-1-hongtao.liu@intel.com>
On Tue, Sep 20, 2022 at 4:15 AM liuhongt via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Here's list the patch supported.
> rint/nearbyint/ceil/floor/trunc/lrint/lceil/lfloor/round/lround.
>
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> Ok for trunk?
>
> gcc/ChangeLog:
>
> PR target/106910
> * config/i386/mmx.md (nearbyintv2sf2): New expander.
> (rintv2sf2): Ditto.
> (ceilv2sf2): Ditto.
> (lceilv2sfv2si2): Ditto.
> (floorv2sf2): Ditto.
> (lfloorv2sfv2si2): Ditto.
> (btruncv2sf2): Ditto.
> (lrintv2sfv2si2): Ditto.
> (roundv2sf2): Ditto.
> (lroundv2sfv2si2): Ditto.
> (*mmx_roundv2sf2): New define_insn.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr106910-1.c: New test.
OK.
Thanks,
Uros.
> ---
> gcc/config/i386/mmx.md | 154 +++++++++++++++++++++
> gcc/testsuite/gcc.target/i386/pr106910-1.c | 77 +++++++++++
> 2 files changed, 231 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr106910-1.c
>
> diff --git a/gcc/config/i386/mmx.md b/gcc/config/i386/mmx.md
> index dda4b43f5c1..222a041de58 100644
> --- a/gcc/config/i386/mmx.md
> +++ b/gcc/config/i386/mmx.md
> @@ -1627,6 +1627,160 @@ (define_expand "vec_initv2sfsf"
> DONE;
> })
>
> +;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
> +;;
> +;; Parallel single-precision floating point rounding operations.
> +;;
> +;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
> +
> +(define_expand "nearbyintv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand")
> + (unspec:V2SF
> + [(match_operand:V2SF 1 "register_operand")
> + (match_dup 2)]
> + UNSPEC_ROUND))]
> + "TARGET_SSE4_1 && TARGET_MMX_WITH_SSE"
> + "operands[2] = GEN_INT (ROUND_MXCSR | ROUND_NO_EXC);")
> +
> +(define_expand "rintv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand")
> + (unspec:V2SF
> + [(match_operand:V2SF 1 "register_operand")
> + (match_dup 2)]
> + UNSPEC_ROUND))]
> + "TARGET_SSE4_1 && TARGET_MMX_WITH_SSE"
> + "operands[2] = GEN_INT (ROUND_MXCSR);")
> +
> +(define_expand "ceilv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand")
> + (unspec:V2SF
> + [(match_operand:V2SF 1 "register_operand")
> + (match_dup 2)]
> + UNSPEC_ROUND))]
> + "TARGET_SSE4_1 && !flag_trapping_math
> + && TARGET_MMX_WITH_SSE"
> + "operands[2] = GEN_INT (ROUND_CEIL | ROUND_NO_EXC);")
> +
> +(define_expand "lceilv2sfv2si2"
> + [(match_operand:V2SI 0 "register_operand")
> + (match_operand:V2SF 1 "register_operand")]
> + "TARGET_SSE4_1 && !flag_trapping_math
> + && TARGET_MMX_WITH_SSE"
> +{
> + rtx tmp = gen_reg_rtx (V2SFmode);
> + emit_insn (gen_ceilv2sf2 (tmp, operands[1]));
> + emit_insn (gen_fix_truncv2sfv2si2 (operands[0], tmp));
> + DONE;
> +})
> +
> +(define_expand "floorv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand")
> + (unspec:V2SF
> + [(match_operand:V2SF 1 "vector_operand")
> + (match_dup 2)]
> + UNSPEC_ROUND))]
> + "TARGET_SSE4_1 && !flag_trapping_math
> + && TARGET_MMX_WITH_SSE"
> + "operands[2] = GEN_INT (ROUND_FLOOR | ROUND_NO_EXC);")
> +
> +(define_expand "lfloorv2sfv2si2"
> + [(match_operand:V2SI 0 "register_operand")
> + (match_operand:V2SF 1 "register_operand")]
> + "TARGET_SSE4_1 && !flag_trapping_math
> + && TARGET_MMX_WITH_SSE"
> +{
> + rtx tmp = gen_reg_rtx (V2SFmode);
> + emit_insn (gen_floorv2sf2 (tmp, operands[1]));
> + emit_insn (gen_fix_truncv2sfv2si2 (operands[0], tmp));
> + DONE;
> +})
> +
> +(define_expand "btruncv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand")
> + (unspec:V2SF
> + [(match_operand:V2SF 1 "register_operand")
> + (match_dup 2)]
> + UNSPEC_ROUND))]
> + "TARGET_SSE4_1 && !flag_trapping_math"
> + "operands[2] = GEN_INT (ROUND_TRUNC | ROUND_NO_EXC);")
> +
> +(define_insn "*mmx_roundv2sf2"
> + [(set (match_operand:V2SF 0 "register_operand" "=Yr,*x,v")
> + (unspec:V2SF
> + [(match_operand:V2SF 1 "register_operand" "Yr,x,v")
> + (match_operand:SI 2 "const_0_to_15_operand")]
> + UNSPEC_ROUND))]
> + "TARGET_SSE4_1 && TARGET_MMX_WITH_SSE"
> + "%vroundps\t{%2, %1, %0|%0, %1, %2}"
> + [(set_attr "isa" "noavx,noavx,avx")
> + (set_attr "type" "ssecvt")
> + (set_attr "prefix_data16" "1,1,*")
> + (set_attr "prefix_extra" "1")
> + (set_attr "length_immediate" "1")
> + (set_attr "prefix" "orig,orig,vex")
> + (set_attr "mode" "V4SF")])
> +
> +(define_insn "lrintv2sfv2si2"
> + [(set (match_operand:V2SI 0 "register_operand" "=v")
> + (unspec:V2SI
> + [(match_operand:V2SF 1 "register_operand" "v")]
> + UNSPEC_FIX_NOTRUNC))]
> + "TARGET_MMX_WITH_SSE"
> + "%vcvtps2dq\t{%1, %0|%0, %1}"
> + [(set_attr "type" "ssecvt")
> + (set (attr "prefix_data16")
> + (if_then_else
> + (match_test "TARGET_AVX")
> + (const_string "*")
> + (const_string "1")))
> + (set_attr "prefix" "maybe_vex")
> + (set_attr "mode" "TI")])
> +
> +(define_expand "roundv2sf2"
> + [(set (match_dup 3)
> + (plus:V2SF
> + (match_operand:V2SF 1 "register_operand")
> + (match_dup 2)))
> + (set (match_operand:V2SF 0 "register_operand")
> + (unspec:V2SF
> + [(match_dup 3) (match_dup 4)]
> + UNSPEC_ROUND))]
> + "TARGET_SSE4_1 && !flag_trapping_math
> + && TARGET_MMX_WITH_SSE"
> +{
> + const struct real_format *fmt;
> + REAL_VALUE_TYPE pred_half, half_minus_pred_half;
> + rtx half, vec_half;
> +
> + /* load nextafter (0.5, 0.0) */
> + fmt = REAL_MODE_FORMAT (SFmode);
> + real_2expN (&half_minus_pred_half, -(fmt->p) - 1, SFmode);
> + real_arithmetic (&pred_half, MINUS_EXPR, &dconsthalf, &half_minus_pred_half);
> + half = const_double_from_real_value (pred_half, SFmode);
> +
> + vec_half = ix86_build_const_vector (V2SFmode, true, half);
> + vec_half = force_reg (V2SFmode, vec_half);
> +
> + operands[2] = gen_reg_rtx (V2SFmode);
> + emit_insn (gen_copysignv2sf3 (operands[2], vec_half, operands[1]));
> +
> + operands[3] = gen_reg_rtx (V2SFmode);
> + operands[4] = GEN_INT (ROUND_TRUNC);
> +})
> +
> +(define_expand "lroundv2sfv2si2"
> + [(match_operand:V2SI 0 "register_operand")
> + (match_operand:V2SF 1 "register_operand")]
> + "TARGET_SSE4_1 && !flag_trapping_math
> + && TARGET_MMX_WITH_SSE"
> +{
> + rtx tmp = gen_reg_rtx (V2SFmode);
> + emit_insn (gen_roundv2sf2 (tmp, operands[1]));
> + emit_insn (gen_fix_truncv2sfv2si2 (operands[0], tmp));
> + DONE;
> +})
> +
> +
> ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
> ;;
> ;; Parallel half-precision floating point arithmetic
> diff --git a/gcc/testsuite/gcc.target/i386/pr106910-1.c b/gcc/testsuite/gcc.target/i386/pr106910-1.c
> new file mode 100644
> index 00000000000..c7685a32183
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106910-1.c
> @@ -0,0 +1,77 @@
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-msse4.1 -O2 -Ofast" } */
> +/* { dg-final { scan-assembler-times "roundps" 9 } } */
> +/* { dg-final { scan-assembler-times "cvtps2dq" 1 } } */
> +/* { dg-final { scan-assembler-times "cvttps2dq" 3 } } */
> +
> +#include<math.h>
> +
> +void
> +foo (float* p, float* __restrict q)
> +{
> + p[0] = truncf (q[0]);
> + p[1] = truncf (q[1]);
> +}
> +
> +void
> +foo1 (float* p, float* __restrict q)
> +{
> + p[0] = floorf (q[0]);
> + p[1] = floorf (q[1]);
> +}
> +
> +void
> +foo1i (int* p, float* __restrict q)
> +{
> + p[0] = (int) floorf (q[0]);
> + p[1] = (int) floorf (q[1]);
> +}
> +
> +void
> +foo2 (float* p, float* __restrict q)
> +{
> + p[0] = ceilf (q[0]);
> + p[1] = ceilf (q[1]);
> +}
> +
> +void
> +foo2i (int* p, float* __restrict q)
> +{
> + p[0] = (int) ceilf (q[0]);
> + p[1] = (int) ceilf (q[1]);
> +}
> +
> +void
> +foo3 (float* p, float* __restrict q)
> +{
> + p[0] = rintf (q[0]);
> + p[1] = rintf (q[1]);
> +}
> +
> +void
> +foo3i (int* p, float* __restrict q)
> +{
> + p[0] = (int) rintf (q[0]);
> + p[1] = (int) rintf (q[1]);
> +}
> +
> +void
> +foo4 (float* p, float* __restrict q)
> +{
> + p[0] = nearbyintf (q[0]);
> + p[1] = nearbyintf (q[1]);
> +}
> +
> +void
> +foo5(float* p, float* __restrict q)
> +{
> + p[0] = roundf (q[0]);
> + p[1] = roundf (q[1]);
> +}
> +
> +void
> +foo5i(int* p, float* __restrict q)
> +{
> + p[0] = (int) roundf (q[0]);
> + p[1] = (int) roundf (q[1]);
> +}
> --
> 2.27.0
>
prev parent reply other threads:[~2022-09-20 6:35 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-09-20 2:12 liuhongt
2022-09-20 2:17 ` Hongtao Liu
2022-09-20 6:35 ` Uros Bizjak [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAFULd4a9n0xV4XbNWQt9jo9uyREE6sr1d20GVyQGKus8KSnx8w@mail.gmail.com \
--to=ubizjak@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hongtao.liu@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).