arm: ACLE I8MM multiply-accumulate
authorDennis Zhang <dennis.zhang@arm.com>
Fri, 21 Feb 2020 15:36:13 +0000 (15:36 +0000)
committerDennis Zhang <dennis.zhang@arm.com>
Fri, 21 Feb 2020 15:36:13 +0000 (15:36 +0000)
This patch adds intrinsics for matrix multiply-accumulate instructions
including vmmlaq_s32, vmmlaq_u32, and vusmmlaq_s32.

gcc/ChangeLog:

2020-02-21  Dennis Zhang  <dennis.zhang@arm.com>

* config/arm/arm_neon.h (vmmlaq_s32, vmmlaq_u32, vusmmlaq_s32): New.
* config/arm/arm_neon_builtins.def (smmla, ummla, usmmla): New.
* config/arm/iterators.md (MATMUL): New iterator.
(sup): Add UNSPEC_MATMUL_S, UNSPEC_MATMUL_U, and UNSPEC_MATMUL_US.
(mmla_sfx): New attribute.
* config/arm/neon.md (neon_<sup>mmlav16qi): New.
* config/arm/unspecs.md (UNSPEC_MATMUL_S, UNSPEC_MATMUL_U): New.
(UNSPEC_MATMUL_US): New.

gcc/testsuite/ChangeLog:

2020-02-21  Dennis Zhang  <dennis.zhang@arm.com>

* gcc.target/arm/simd/vmmla_1.c: New test.

gcc/ChangeLog
gcc/config/arm/arm_neon.h
gcc/config/arm/arm_neon_builtins.def
gcc/config/arm/iterators.md
gcc/config/arm/neon.md
gcc/config/arm/unspecs.md
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.target/arm/simd/vmmla_1.c [new file with mode: 0644]

index 3d4d799ed5d14e83677e6a2ebec5dc21c6d9b616..833fa6a8383abfb893814ed218e4f6cd371664cd 100644 (file)
@@ -1,3 +1,14 @@
+2020-02-21  Dennis Zhang  <dennis.zhang@arm.com>
+
+       * config/arm/arm_neon.h (vmmlaq_s32, vmmlaq_u32, vusmmlaq_s32): New.
+       * config/arm/arm_neon_builtins.def (smmla, ummla, usmmla): New.
+       * config/arm/iterators.md (MATMUL): New iterator.
+       (sup): Add UNSPEC_MATMUL_S, UNSPEC_MATMUL_U, and UNSPEC_MATMUL_US.
+       (mmla_sfx): New attribute.
+       * config/arm/neon.md (neon_<sup>mmlav16qi): New.
+       * config/arm/unspecs.md (UNSPEC_MATMUL_S, UNSPEC_MATMUL_U): New.
+       (UNSPEC_MATMUL_US): New.
+
 2020-02-21  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
 
        * config/arm/arm.md: Prevent scalar shifts from being used when big
index 81f550988f7b987b4efe504ccc849e3f6fdc692a..e81681aa415437e61857e7edcc09b1a3ef4eeaa8 100644 (file)
@@ -18791,6 +18791,34 @@ vsudotq_lane_s32 (int32x4_t __r, int8x16_t __a,
 #pragma GCC pop_options
 #endif
 
+/* AdvSIMD 8-bit Integer Matrix Multiply (I8MM) intrinsics.  */
+
+#pragma GCC push_options
+#pragma GCC target ("arch=armv8.2-a+i8mm")
+
+__extension__ extern __inline int32x4_t
+__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+vmmlaq_s32 (int32x4_t __r, int8x16_t __a, int8x16_t __b)
+{
+  return __builtin_neon_smmlav16qi (__r, __a, __b);
+}
+
+__extension__ extern __inline uint32x4_t
+__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+vmmlaq_u32 (uint32x4_t __r, uint8x16_t __a, uint8x16_t __b)
+{
+  return __builtin_neon_ummlav16qi_uuuu (__r, __a, __b);
+}
+
+__extension__ extern __inline int32x4_t
+__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+vusmmlaq_s32 (int32x4_t __r, uint8x16_t __a, int8x16_t __b)
+{
+  return __builtin_neon_usmmlav16qi_ssus (__r, __a, __b);
+}
+
+#pragma GCC pop_options
+
 #ifdef __cplusplus
 }
 #endif
index b4537ff5de97d6a59435cc7fb7afd5c4c2aa4d44..f4a97fd764c3f92f850afeecfb1f997cb66bd207 100644 (file)
@@ -377,3 +377,7 @@ VAR2 (MAC_LANE_PAIR, vcmlaq_lane0, v4sf, v8hf)
 VAR2 (MAC_LANE_PAIR, vcmlaq_lane90, v4sf, v8hf)
 VAR2 (MAC_LANE_PAIR, vcmlaq_lane180, v4sf, v8hf)
 VAR2 (MAC_LANE_PAIR, vcmlaq_lane270, v4sf, v8hf)
+
+VAR1 (TERNOP, smmla, v16qi)
+VAR1 (UTERNOP, ummla, v16qi)
+VAR1 (USTERNOP, usmmla, v16qi)
index f8a76fb59cb8afcba762593d3617139490e3a091..136c45274aef61c7c41b553898e41b3838300a4c 100644 (file)
 (define_int_iterator VCADD [UNSPEC_VCADD90 UNSPEC_VCADD270])
 (define_int_iterator VCMLA [UNSPEC_VCMLA UNSPEC_VCMLA90 UNSPEC_VCMLA180 UNSPEC_VCMLA270])
 
+(define_int_iterator MATMUL [UNSPEC_MATMUL_S UNSPEC_MATMUL_U UNSPEC_MATMUL_US])
+
 ;;----------------------------------------------------------------------------
 ;; Mode attributes
 ;;----------------------------------------------------------------------------
   (UNSPEC_DOT_S "s") (UNSPEC_DOT_U "u")
   (UNSPEC_DOT_US "us") (UNSPEC_DOT_SU "su")
   (UNSPEC_SSAT16 "s") (UNSPEC_USAT16 "u")
+  (UNSPEC_MATMUL_S "s") (UNSPEC_MATMUL_U "u") (UNSPEC_MATMUL_US "us")
 ])
 
 (define_int_attr vfml_half
                            (UNSPEC_SMUADX "smuadx") (UNSPEC_SSAT16 "ssat16")
                            (UNSPEC_USAT16 "usat16")])
 
+(define_int_attr mmla_sfx [(UNSPEC_MATMUL_S "s8") (UNSPEC_MATMUL_U "u8")
+                          (UNSPEC_MATMUL_US "s8")])
+
 ;; Both kinds of return insn.
 (define_code_iterator RETURNS [return simple_return])
 (define_code_attr return_str [(return "") (simple_return "simple_")])
index 5d085dc6f613f6ab4ce3c4302d6021b01446327f..039cd90c3da140f48db764d51dda80036d8a7c76 100644 (file)
@@ -6585,3 +6585,14 @@ if (BYTES_BIG_ENDIAN)
  "vabd.<V_if_elem> %<V_reg>0, %<V_reg>1, %<V_reg>2"
  [(set_attr "type" "neon_fp_abd_s<q>")]
 )
+
+(define_insn "neon_<sup>mmlav16qi"
+  [(set (match_operand:V4SI 0 "register_operand" "=w")
+       (plus:V4SI
+        (unspec:V4SI [(match_operand:V16QI 2 "register_operand" "w")
+                      (match_operand:V16QI 3 "register_operand" "w")] MATMUL)
+        (match_operand:V4SI 1 "register_operand" "0")))]
+  "TARGET_I8MM"
+  "v<sup>mmla.<mmla_sfx>\t%q0, %q2, %q3"
+  [(set_attr "type" "neon_mla_s_q")]
+)
index 2dd62467fe6e221ad569c7186454b1460436efef..c8f3f95c7f3176cae1ad9d9db9647b961078965d 100644 (file)
   UNSPEC_VCMLA90
   UNSPEC_VCMLA180
   UNSPEC_VCMLA270
+  UNSPEC_MATMUL_S
+  UNSPEC_MATMUL_U
+  UNSPEC_MATMUL_US
 ])
index 01da6fb33ce036f5e4818afa10dd07010e674c6b..b608a1aa83b9eaf6bbebc5004628d28f09bf14dc 100644 (file)
@@ -1,3 +1,7 @@
+2020-02-21  Dennis Zhang  <dennis.zhang@arm.com>
+
+       * gcc.target/arm/simd/vmmla_1.c: New test.
+
 2020-02-21  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>
 
        * gcc.target/arm/armv8_1m-shift-imm-1.c: Add MVE target checks.
diff --git a/gcc/testsuite/gcc.target/arm/simd/vmmla_1.c b/gcc/testsuite/gcc.target/arm/simd/vmmla_1.c
new file mode 100644 (file)
index 0000000..b766a91
--- /dev/null
@@ -0,0 +1,28 @@
+/* { dg-do assemble } */
+/* { dg-require-effective-target arm_v8_2a_i8mm_ok } */
+/* { dg-options "-save-temps -O2" } */
+/* { dg-additional-options "-march=armv8.2-a+i8mm" } */
+
+#include "arm_neon.h"
+
+int32x4_t
+test_vmmlaq_s32 (int32x4_t r, int8x16_t a, int8x16_t b)
+{
+  return vmmlaq_s32 (r, a, b);
+}
+
+uint32x4_t
+test_vmmlaq_u32 (uint32x4_t r, uint8x16_t a, uint8x16_t b)
+{
+  return vmmlaq_u32 (r, a, b);
+}
+
+int32x4_t
+test_vusmmlaq_s32 (int32x4_t r, uint8x16_t a, int8x16_t b)
+{
+  return vusmmlaq_s32 (r, a, b);
+}
+
+/* { dg-final { scan-assembler-times {\tvsmmla.s8\tq[0-9]+, q[0-9]+, q[0-9]+} 1 } } */
+/* { dg-final { scan-assembler-times {\tvummla.u8\tq[0-9]+, q[0-9]+, q[0-9]+} 1 } } */
+/* { dg-final { scan-assembler-times {\tvusmmla.s8\tq[0-9]+, q[0-9]+, q[0-9]+} 1 } } */