Hi All,
This patch adds the instructions for Dot Product to ARM along
with the intrinsics and vectorizer pattern.
Armv8.2-a dot product supports 8-bit element values both
signed and unsigned.
Dot product is available from Armv8.2-a and onwards.
Regtested and bootstrapped on arm-none-eabi and no issues.
Ok for trunk?
gcc/
2017-09-01 Tamar Christina <tamar.christ...@arm.com>
* config/arm/arm-builtins.c (arm_unsigned_uternop_qualifiers): New.
(UTERNOP_QUALIFIERS, arm_umac_lane_qualifiers, UMAC_LANE_QUALIFIERS):
New.
* config/arm/arm_neon_builtins.def (sdot, udot, sdot_lane, udot_lane):
new.
* config/arm/iterators.md (DOTPROD, DOT_MODE, dot_mode): New.
(UNSPEC_DOT_S, UNSPEC_DOT_U, opsuffix): New.
* config/arm/neon.md (neon_<sup>dot<dot_mode>): New.
(neon_<sup>dot_lane<dot_mode>, <sup>dot_prod<dot_mode>): New.
* config/arm/types.md (neon_dot, neon_dot_q): New.
* config/arm/unspecs.md (UNSPEC_DOT_S, UNSPEC_DOT_U): New.
--
diff --git a/gcc/config/arm/arm-builtins.c b/gcc/config/arm/arm-builtins.c
index 7504ed581c63a657a0dff48442633704bd252b2e..467596c1dbefe62cc92a0ffb8a03ecaf950f3701 100644
--- a/gcc/config/arm/arm-builtins.c
+++ b/gcc/config/arm/arm-builtins.c
@@ -105,6 +105,13 @@ arm_ternop_qualifiers[SIMD_MAX_BUILTIN_ARGS]
= { qualifier_none, qualifier_none, qualifier_none, qualifier_none };
#define TERNOP_QUALIFIERS (arm_ternop_qualifiers)
+/* unsigned T (unsigned T, unsigned T, unsigned T). */
+static enum arm_type_qualifiers
+arm_unsigned_uternop_qualifiers[SIMD_MAX_BUILTIN_ARGS]
+ = { qualifier_unsigned, qualifier_unsigned, qualifier_unsigned,
+ qualifier_unsigned };
+#define UTERNOP_QUALIFIERS (arm_unsigned_uternop_qualifiers)
+
/* T (T, immediate). */
static enum arm_type_qualifiers
arm_binop_imm_qualifiers[SIMD_MAX_BUILTIN_ARGS]
@@ -131,6 +138,13 @@ arm_mac_lane_qualifiers[SIMD_MAX_BUILTIN_ARGS]
qualifier_none, qualifier_lane_index };
#define MAC_LANE_QUALIFIERS (arm_mac_lane_qualifiers)
+/* unsigned T (unsigned T, unsigned T, unsigend T, lane index). */
+static enum arm_type_qualifiers
+arm_umac_lane_qualifiers[SIMD_MAX_BUILTIN_ARGS]
+ = { qualifier_unsigned, qualifier_unsigned, qualifier_unsigned,
+ qualifier_unsigned, qualifier_lane_index };
+#define UMAC_LANE_QUALIFIERS (arm_umac_lane_qualifiers)
+
/* T (T, T, immediate). */
static enum arm_type_qualifiers
arm_ternop_imm_qualifiers[SIMD_MAX_BUILTIN_ARGS]
diff --git a/gcc/config/arm/arm_neon_builtins.def b/gcc/config/arm/arm_neon_builtins.def
index 07f0368343a0c940c1cc1848d31f28a47a587b6f..982eec810dafb5ec955273099853f8842020d104 100644
--- a/gcc/config/arm/arm_neon_builtins.def
+++ b/gcc/config/arm/arm_neon_builtins.def
@@ -331,3 +331,7 @@ VAR11 (STORE1, vst4,
v8qi, v4hi, v4hf, v2si, v2sf, di, v16qi, v8hi, v8hf, v4si, v4sf)
VAR9 (STORE1LANE, vst4_lane,
v8qi, v4hi, v4hf, v2si, v2sf, v8hi, v8hf, v4si, v4sf)
+VAR2 (TERNOP, sdot, v8qi, v16qi)
+VAR2 (UTERNOP, udot, v8qi, v16qi)
+VAR2 (MAC_LANE, sdot_lane, v8qi, v16qi)
+VAR2 (UMAC_LANE, udot_lane, v8qi, v16qi)
diff --git a/gcc/config/arm/iterators.md b/gcc/config/arm/iterators.md
index 7acbaf1bb40a4f270e75968804546508f7839e49..139e09fd929e17216ad9383505f1453a73d071fb 100644
--- a/gcc/config/arm/iterators.md
+++ b/gcc/config/arm/iterators.md
@@ -410,6 +410,8 @@
(define_int_iterator VFM_LANE_AS [UNSPEC_VFMA_LANE UNSPEC_VFMS_LANE])
+(define_int_iterator DOTPROD [UNSPEC_DOT_S UNSPEC_DOT_U])
+
;;----------------------------------------------------------------------------
;; Mode attributes
;;----------------------------------------------------------------------------
@@ -720,6 +722,10 @@
(define_mode_attr pf [(V8QI "p") (V16QI "p") (V2SF "f") (V4SF "f")])
+;; Mapping attribute for Dot Product input modes based on result mode.
+(define_mode_attr DOT_MODE [(V2SI "V8QI") (V4SI "V16QI")])
+(define_mode_attr dot_mode [(V2SI "v8qi") (V4SI "v16qi")])
+
;;----------------------------------------------------------------------------
;; Code attributes
;;----------------------------------------------------------------------------
@@ -816,6 +822,7 @@
(UNSPEC_VSRA_S_N "s") (UNSPEC_VSRA_U_N "u")
(UNSPEC_VRSRA_S_N "s") (UNSPEC_VRSRA_U_N "u")
(UNSPEC_VCVTH_S "s") (UNSPEC_VCVTH_U "u")
+ (UNSPEC_DOT_S "s") (UNSPEC_DOT_U "u")
])
(define_int_attr vcvth_op
@@ -1003,3 +1010,6 @@
(define_int_attr mrrc [(VUNSPEC_MRRC "mrrc") (VUNSPEC_MRRC2 "mrrc2")])
(define_int_attr MRRC [(VUNSPEC_MRRC "MRRC") (VUNSPEC_MRRC2 "MRRC2")])
+
+(define_int_attr opsuffix [(UNSPEC_DOT_S "s8")
+ (UNSPEC_DOT_U "u8")])
diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md
index 45b3bd18052dd4a33e4b9c10f3ca2ea7e2eed5ce..4c919f53757ada3ca40609b8244ed3b403be3329 100644
--- a/gcc/config/arm/neon.md
+++ b/gcc/config/arm/neon.md
@@ -3044,6 +3044,71 @@
DONE;
})
+;; These instructions map to the __builtins for the Dot Product operations.
+(define_insn "neon_<sup>dot<dot_mode>"
+ [(set (match_operand:VCVTI 0 "register_operand" "=w")
+ (unspec:VCVTI [(match_operand:VCVTI 1 "register_operand" "0")
+ (match_operand:<DOT_MODE> 2 "register_operand" "w")
+ (match_operand:<DOT_MODE> 3 "register_operand" "w")]
+ DOTPROD))]
+ "TARGET_DOTPROD"
+ "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>2, %<V_reg>3"
+ [(set_attr "type" "neon_dot")]
+)
+
+;; These instructions map to the __builtins for the Dot Product
+;; indexed operations.
+(define_insn "neon_<sup>dot_lane<dot_mode>"
+ [(set (match_operand:VCVTI 0 "register_operand" "=w")
+ (unspec:VCVTI [(match_operand:VCVTI 1 "register_operand" "0")
+ (match_operand:<DOT_MODE> 2 "register_operand" "w")
+ (match_operand:V8QI 3 "register_operand" "t")
+ (match_operand:SI 4 "immediate_operand" "i")]
+ DOTPROD))]
+ "TARGET_DOTPROD"
+ {
+ operands[4]
+ = GEN_INT (NEON_ENDIAN_LANE_N (V8QImode, INTVAL (operands[4])));
+ return "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>2, %P3[%c4]";
+ }
+ [(set_attr "type" "neon_dot")]
+)
+
+;; These expands map to the Dot Product optab the vectorizer checks for.
+;; The auto-vectorizer expects a dot product builtin that also does an
+;; accumulation into the provided register.
+;; Given the following pattern
+;;
+;; for (i=0; i<len; i++) {
+;; c = a[i] * b[i];
+;; r += c;
+;; }
+;; return result;
+;;
+;; This can be auto-vectorized to
+;; r = a[0]*b[0] + a[1]*b[1] + a[2]*b[2] + a[3]*b[3];
+;;
+;; given enough iterations. However the vectorizer can keep unrolling the loop
+;; r += a[4]*b[4] + a[5]*b[5] + a[6]*b[6] + a[7]*b[7];
+;; r += a[8]*b[8] + a[9]*b[9] + a[10]*b[10] + a[11]*b[11];
+;; ...
+;;
+;; and so the vectorizer provides r, in which the result has to be accumulated.
+(define_expand "<sup>dot_prod<dot_mode>"
+ [(set (match_operand:VCVTI 0 "register_operand")
+ (unspec:VCVTI [(match_operand:<DOT_MODE> 1 "register_operand")
+ (match_operand:<DOT_MODE> 2 "register_operand")
+ (match_operand:VCVTI 3 "register_operand")]
+ DOTPROD))]
+ "TARGET_DOTPROD"
+{
+ emit_insn (
+ gen_neon_<sup>dot<dot_mode> (operands[3], operands[3], operands[1],
+ operands[2]));
+ emit_insn (gen_rtx_SET (operands[0], operands[3]));
+ DONE;
+})
+
(define_expand "neon_copysignf<mode>"
[(match_operand:VCVTF 0 "register_operand")
(match_operand:VCVTF 1 "register_operand")
diff --git a/gcc/config/arm/types.md b/gcc/config/arm/types.md
index 253f4968d12a09fe69f3a0d82f896115b0a92c47..27398dffd9e71ee851d5d2ed5759d7c8a85df320 100644
--- a/gcc/config/arm/types.md
+++ b/gcc/config/arm/types.md
@@ -316,6 +316,8 @@
; neon_cls_q
; neon_cnt
; neon_cnt_q
+; neon_dot
+; neon_dot_q
; neon_ext
; neon_ext_q
; neon_rbit
@@ -764,6 +766,8 @@
\
neon_abs,\
neon_abs_q,\
+ neon_dot,\
+ neon_dot_q,\
neon_neg,\
neon_neg_q,\
neon_qneg,\
@@ -1110,8 +1114,8 @@
neon_sub, neon_sub_q, neon_sub_widen, neon_sub_long, neon_qsub,\
neon_qsub_q, neon_sub_halve, neon_sub_halve_q,\
neon_sub_halve_narrow_q,\
- neon_abs, neon_abs_q, neon_neg, neon_neg_q, neon_qneg,\
- neon_qneg_q, neon_qabs, neon_qabs_q, neon_abd, neon_abd_q,\
+ neon_abs, neon_abs_q, neon_dot, neon_dot_q, neon_neg, neon_neg_q,\
+ neon_qneg, neon_qneg_q, neon_qabs, neon_qabs_q, neon_abd, neon_abd_q,\
neon_abd_long, neon_minmax, neon_minmax_q, neon_compare,\
neon_compare_q, neon_compare_zero, neon_compare_zero_q,\
neon_arith_acc, neon_arith_acc_q, neon_reduc_add,\
diff --git a/gcc/config/arm/unspecs.md b/gcc/config/arm/unspecs.md
index 99cfa41b08dad24a85e78f069331e83c03c8bce1..c474f4bb5db995b60f464f098e478f0398ce15f9 100644
--- a/gcc/config/arm/unspecs.md
+++ b/gcc/config/arm/unspecs.md
@@ -410,4 +410,6 @@
UNSPEC_VRNDN
UNSPEC_VRNDP
UNSPEC_VRNDX
+ UNSPEC_DOT_S
+ UNSPEC_DOT_U
])
