From: Kyrill Tkachov <kyrylo.tkachov@foss.arm.com>
To: Tamar Christina <Tamar.Christina@arm.com>,
"gcc-patches@gcc.gnu.org" <gcc-patches@gcc.gnu.org>
Cc: nd <nd@arm.com>, James Greenhalgh <James.Greenhalgh@arm.com>,
Richard Earnshaw <Richard.Earnshaw@arm.com>,
Marcus Shawcroft <Marcus.Shawcroft@arm.com>
Subject: Re: [PATCH 3/9][GCC][AArch64] Add autovectorization support for Complex instructions
Date: Mon, 12 Nov 2018 12:17:00 -0000 [thread overview]
Message-ID: <5BE96F4D.50405@foss.arm.com> (raw)
In-Reply-To: <20181111102628.GA4529@arm.com>
Hi Tamar,
On 11/11/18 10:26, Tamar Christina wrote:
> Hi All,
>
> This patch adds the expander support for supporting autovectorization of complex number operations
> such as Complex addition with a rotation along the Argand plane. This also adds support for complex
> FMA.
>
> The instructions are described in the ArmARM [1] and are available from Armv8.3-a onwards.
>
> Concretely, this generates
>
> f90:
> mov x3, 0
> .p2align 3,,7
> .L2:
> ldr q0, [x0, x3]
> ldr q1, [x1, x3]
> fcadd v0.2d, v0.2d, v1.2d, #90
> str q0, [x2, x3]
> add x3, x3, 16
> cmp x3, 3200
> bne .L2
> ret
>
> now instead of
>
> f90:
> mov x4, x1
> mov x1, x2
> add x3, x4, 31
> add x2, x0, 31
> sub x3, x3, x1
> sub x2, x2, x1
> cmp x3, 62
> mov x3, 62
> ccmp x2, x3, 0, hi
> bls .L5
> mov x2, x4
> add x3, x0, 3200
> .p2align 3,,7
> .L3:
> ld2 {v2.2d - v3.2d}, [x0], 32
> ld2 {v4.2d - v5.2d}, [x2], 32
> cmp x0, x3
> fsub v0.2d, v2.2d, v5.2d
> fadd v1.2d, v4.2d, v3.2d
> st2 {v0.2d - v1.2d}, [x1], 32
> bne .L3
> ret
> .L5:
> add x6, x0, 8
> add x5, x4, 8
> add x2, x1, 8
> mov x3, 0
> .p2align 3,,7
> .L2:
> ldr d1, [x0, x3]
> ldr d3, [x5, x3]
> ldr d0, [x6, x3]
> ldr d2, [x4, x3]
> fsub d1, d1, d3
> fadd d0, d0, d2
> str d1, [x1, x3]
> str d0, [x2, x3]
> add x3, x3, 16
> cmp x3, 3200
> bne .L2
> ret
>
> For complex additions with a 90* rotation along the Argand plane.
>
> [1] https://developer.arm.com/docs/ddi0487/latest/arm-architecture-reference-manual-armv8-for-armv8-a-architecture-profile
>
> Bootstrap and Regtest on aarch64-none-linux-gnu, arm-none-gnueabihf and x86_64-pc-linux-gnu
> are still on going but previous patch showed no regressions.
>
> The instructions have also been tested on aarch64-none-elf and arm-none-eabi on a Armv8.3-a model
> and -march=Armv8.3-a+fp16 and all tests pass.
>
> Ok for trunk?
>
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> 2018-11-11 Tamar Christina <tamar.christina@arm.com>
>
> * config/aarch64/aarch64-simd.md (aarch64_fcadd<rot><mode>,
> fcadd<rot><mode>3, aarch64_fcmla<rot><mode>,
> fcmla<rot><mode>4): New.
> * config/aarch64/aarch64.h (TARGET_COMPLEX): New.
> * config/aarch64/iterators.md (UNSPEC_FCADD90, UNSPEC_FCADD270,
> UNSPEC_FCMLA, UNSPEC_FCMLA90, UNSPEC_FCMLA180, UNSPEC_FCMLA270): New.
> (FCADD, FCMLA): New.
> (rot, rotsplit1, rotsplit2): New.
> * config/arm/types.md (neon_fcadd, neon_fcmla): New.
>
> --
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index c4be3101fdec930707918106cd7c53cf7584553e..12a91183a98ea23015860c77a97955cb1b30bfbb 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -419,6 +419,63 @@
}
)
+;; The fcadd and fcmla patterns are made UNSPEC for the explicitly due to the
+;; fact that their usage need to guarantee that the source vectors are
+;; contiguous. It would be wrong to describe the operation without being able
+;; to describe the permute that is also required, but even if that is done
+;; the permute would have been created as a LOAD_LANES which means the values
+;; in the registers are in the wrong order.
+(define_insn "aarch64_fcadd<rot><mode>"
+ [(set (match_operand:VHSDF 0 "register_operand" "=w")
+ (unspec:VHSDF [(match_operand:VHSDF 1 "register_operand" "w")
+ (match_operand:VHSDF 2 "register_operand" "w")]
+ FCADD))]
+ "TARGET_COMPLEX"
+ "fcadd\t%0.<Vtype>, %1.<Vtype>, %2.<Vtype>, #<rot>"
+ [(set_attr "type" "neon_fcadd")]
+)
+
+(define_expand "fcadd<rot><mode>3"
+ [(set (match_operand:VHSDF 0 "register_operand")
+ (unspec:VHSDF [(match_operand:VHSDF 1 "register_operand")
+ (match_operand:VHSDF 2 "register_operand")]
+ FCADD))]
+ "TARGET_COMPLEX"
+{
+ emit_insn (gen_aarch64_fcadd<rot><mode> (operands[0], operands[1],
+ operands[2]));
+ DONE;
+})
+
+(define_insn "aarch64_fcmla<rot><mode>"
+ [(set (match_operand:VHSDF 0 "register_operand" "=w")
+ (plus:VHSDF (match_operand:VHSDF 1 "register_operand" "0")
+ (unspec:VHSDF [(match_operand:VHSDF 2 "register_operand" "w")
+ (match_operand:VHSDF 3 "register_operand" "w")]
+ FCMLA)))]
+ "TARGET_COMPLEX"
+ "fcmla\t%0.<Vtype>, %2.<Vtype>, %3.<Vtype>, #<rot>"
+ [(set_attr "type" "neon_fcmla")]
+)
+
+;; The complex mla operations always need to expand to two instructions.
+;; The first operation does half the computation and the second does the
+;; remainder. Because of this, expand early.
+(define_expand "fcmla<rot><mode>4"
+ [(set (match_operand:VHSDF 0 "register_operand")
+ (plus:VHSDF (match_operand:VHSDF 1 "register_operand")
+ (unspec:VHSDF [(match_operand:VHSDF 2 "register_operand")
+ (match_operand:VHSDF 3 "register_operand")]
+ FCMLA)))]
+ "TARGET_COMPLEX"
+{
+ emit_insn (gen_aarch64_fcmla<rotsplit1><mode> (operands[0], operands[1],
+ operands[2], operands[3]));
+ emit_insn (gen_aarch64_fcmla<rotsplit2><mode> (operands[0], operands[0],
+ operands[2], operands[3]));
+ DONE;
+})
+
;; These instructions map to the __builtins for the Dot Product operations.
(define_insn "aarch64_<sur>dot<vsi2qi>"
[(set (match_operand:VS 0 "register_operand" "=w")
@@ -3026,22 +3083,22 @@
operands[2] = aarch64_endian_lane_rtx (<MODE>mode, INTVAL (operands[2]));
return "smov\\t%<GPI:w>0, %1.<VDQQH:Vetype>[%2]";
}
- [(set_attr "type" "neon_to_gp<q>")]
-)
-
-(define_insn "*aarch64_get_lane_zero_extend<GPI:mode><VDQQH:mode>"
- [(set (match_operand:GPI 0 "register_operand" "=r")
- (zero_extend:GPI
- (vec_select:<VEL>
- (match_operand:VDQQH 1 "register_operand" "w")
- (parallel [(match_operand:SI 2 "immediate_operand" "i")]))))]
- "TARGET_SIMD"
- {
- operands[2] = aarch64_endian_lane_rtx (<VDQQH:MODE>mode,
- INTVAL (operands[2]));
- return "umov\\t%w0, %1.<Vetype>[%2]";
- }
- [(set_attr "type" "neon_to_gp<q>")]
+ [(set_attr "type" "neon_to_gp<q>")]
+)
+
+(define_insn "*aarch64_get_lane_zero_extend<GPI:mode><VDQQH:mode>"
+ [(set (match_operand:GPI 0 "register_operand" "=r")
+ (zero_extend:GPI
+ (vec_select:<VEL>
+ (match_operand:VDQQH 1 "register_operand" "w")
+ (parallel [(match_operand:SI 2 "immediate_operand" "i")]))))]
+ "TARGET_SIMD"
+ {
+ operands[2] = aarch64_endian_lane_rtx (<VDQQH:MODE>mode,
+ INTVAL (operands[2]));
+ return "umov\\t%w0, %1.<Vetype>[%2]";
+ }
+ [(set_attr "type" "neon_to_gp<q>")]
)
Is this change intended? Is it just reformatting?
I guess that's okay, but please mention it in the ChangeLog if so.
Thanks,
Kyrill
next prev parent reply other threads:[~2018-11-12 12:17 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2018-11-11 10:26 Tamar Christina
2018-11-12 12:17 ` Kyrill Tkachov [this message]
2018-11-12 12:31 ` Tamar Christina
2018-11-28 17:16 ` James Greenhalgh
2022-11-16 10:50 ` Ramana Radhakrishnan
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=5BE96F4D.50405@foss.arm.com \
--to=kyrylo.tkachov@foss.arm.com \
--cc=James.Greenhalgh@arm.com \
--cc=Marcus.Shawcroft@arm.com \
--cc=Richard.Earnshaw@arm.com \
--cc=Tamar.Christina@arm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=nd@arm.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).