Hi all,
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:
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]
Bootstrapped and tested on arm-none-linux-gnueabihf.
Ok for trunk?
Thanks,
Przemyslaw
2019-05-29 Przemyslaw Wirkus <przemyslaw.wir...@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.
2019-05-29 Przemyslaw Wirkus <przemyslaw.wir...@arm.com>
* gcc.target/arm/ssadv16qi.c: New test.
* gcc.target/arm/usadv16qi.c: Likewise.diff --git a/gcc/config/arm/iterators.md b/gcc/config/arm/iterators.md
index
eb07c5b90c1b1905d35d7b480bdbe7d7a45ab7ba..2462b8c87ea7dbe60ba50d22b1e494bb4fe905c2
100644
--- a/gcc/config/arm/iterators.md
+++ b/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")
diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md
index
de9ae43849038b3cf75feceec36429d5c40c63f2..51ed11abc519ea9d4f9e31751ac6d26a3d1ae5cd
100644
--- a/gcc/config/arm/neon.md
+++ b/gcc/config/arm/neon.md
@@ -3255,6 +3255,32 @@
[(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")
diff --git a/gcc/config/arm/unspecs.md b/gcc/config/arm/unspecs.md
index
174bcc5e3d5e1123cb1c1a595f5003884840aea8..41068bac90aa0ce6fef531789a38e5f7b3b27dff
100644
--- a/gcc/config/arm/unspecs.md
+++ b/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
diff --git a/gcc/testsuite/gcc.target/arm/ssadv16qi.c
b/gcc/testsuite/gcc.target/arm/ssadv16qi.c
new file mode 100644
index
0000000000000000000000000000000000000000..dba5ef4f6b9c0b75463a08549e989edc9c22a9d7
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/ssadv16qi.c
@@ -0,0 +1,29 @@
+/* { 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} } } */
diff --git a/gcc/testsuite/gcc.target/arm/usadv16qi.c
b/gcc/testsuite/gcc.target/arm/usadv16qi.c
new file mode 100644
index
0000000000000000000000000000000000000000..d744bcbab575850de84b7038a2b65e46461c0185
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/usadv16qi.c
@@ -0,0 +1,29 @@
+/* { 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} } } */
