i386: Auto vectorize sdot_prod, usdot_prod with VNNI instruction.

AVX512VNNI/AVXVNNI has vpdpwssd for HImode, vpdpbusd for QImode, so
Adjust HImode sdot_prod expander and add QImode usdot_prod expander
to enhance vectorization for dotprod.

gcc/ChangeLog:

	* config/i386/sse.md (VI2_AVX512VNNIBW): New mode iterator.
	(VI1_AVX512VNNI): Likewise.
	(SDOT_VPDP_SUF): New mode_attr.
	(VI1SI): Likewise.
	(vi1si): Likewise.
	(sdot_prod<mode>): Use VI2_AVX512F iterator, expand to
	vpdpwssd when VNNI targets available.
	(usdot_prod<mode>): New expander for vector QImode.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/vnni-auto-vectorize-1.c: New test.
	* gcc.target/i386/vnni-auto-vectorize-2.c: Ditto.

diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 2764a250229941d18ed3edc47dc116945f5bdae0..22435e5d036830af0b5393b6ea3132f7233024a6 100644 (file)
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -500,6 +500,9 @@
 (define_mode_iterator VI1_AVX512F
   [(V64QI "TARGET_AVX512F") (V32QI "TARGET_AVX") V16QI])
 
+(define_mode_iterator VI1_AVX512VNNI
+  [(V64QI "TARGET_AVX512VNNI") (V32QI "TARGET_AVX2") V16QI])
+
 (define_mode_iterator VI12_256_512_AVX512VL
   [V64QI (V32QI "TARGET_AVX512VL")
    V32HI (V16HI "TARGET_AVX512VL")])
@@ -510,6 +513,10 @@
 (define_mode_iterator VI2_AVX512F
   [(V32HI "TARGET_AVX512F") (V16HI "TARGET_AVX2") V8HI])
 
+(define_mode_iterator VI2_AVX512VNNIBW
+  [(V32HI "TARGET_AVX512BW || TARGET_AVX512VNNI")
+   (V16HI "TARGET_AVX2") V8HI])
+
 (define_mode_iterator VI4_AVX
   [(V8SI "TARGET_AVX") V4SI])
 
@@ -14798,19 +14805,37 @@
 (define_mode_attr SDOT_PMADD_SUF
   [(V32HI "512v32hi") (V16HI "") (V8HI "")])
 
+(define_mode_attr SDOT_VPDP_SUF
+  [(V32HI "v16si") (V16HI "v8si") (V8HI "v4si")])
+
 (define_expand "sdot_prod<mode>"
   [(match_operand:<sseunpackmode> 0 "register_operand")
-   (match_operand:VI2_AVX2 1 "register_operand")
-   (match_operand:VI2_AVX2 2 "register_operand")
+   (match_operand:VI2_AVX512VNNIBW 1 "register_operand")
+   (match_operand:VI2_AVX512VNNIBW 2 "register_operand")
    (match_operand:<sseunpackmode> 3 "register_operand")]
   "TARGET_SSE2"
 {
-  rtx t = gen_reg_rtx (<sseunpackmode>mode);
-  emit_insn (gen_<sse2_avx2>_pmaddwd<SDOT_PMADD_SUF> (t, operands[1], operands[2]));
-  emit_insn (gen_rtx_SET (operands[0],
-                         gen_rtx_PLUS (<sseunpackmode>mode,
-                                       operands[3], t)));
-  DONE;
+  /* Try with vnni instructions.  */
+  if ((<MODE_SIZE> == 64 && TARGET_AVX512VNNI)
+      || (<MODE_SIZE> < 64
+         && ((TARGET_AVX512VNNI && TARGET_AVX512VL) || TARGET_AVXVNNI)))
+    {
+      operands[1] = lowpart_subreg (<sseunpackmode>mode, operands[1], <MODE>mode);
+      operands[2] = lowpart_subreg (<sseunpackmode>mode, operands[2], <MODE>mode);
+      emit_insn (gen_rtx_SET (operands[0], operands[3]));
+      emit_insn (gen_vpdpwssd_<SDOT_VPDP_SUF> (operands[0], operands[3],
+                                              operands[1], operands[2]));
+    }
+    /* Otherwise use pmaddwd + paddd.  */
+    else
+    {
+      rtx t = gen_reg_rtx (<sseunpackmode>mode);
+      emit_insn (gen_<sse2_avx2>_pmaddwd<SDOT_PMADD_SUF> (t, operands[1], operands[2]));
+      emit_insn (gen_rtx_SET (operands[0],
+                             gen_rtx_PLUS (<sseunpackmode>mode,
+                                           operands[3], t)));
+    }
+    DONE;
 })
 
 ;; Normally we use widen_mul_even/odd, but combine can't quite get it all
@@ -27065,6 +27090,29 @@
    [(set_attr ("prefix") ("evex"))
    (set_attr "mode" "<sseinsnmode>")])
 
+(define_mode_attr VI1SI
+ [(V64QI "V16SI") (V32QI "V8SI") (V16QI "V4SI")])
+
+(define_mode_attr vi1si
+ [(V64QI "v16si") (V32QI "v8si") (V16QI "v4si")])
+
+(define_expand "usdot_prod<mode>"
+  [(match_operand:<VI1SI> 0 "register_operand")
+   (match_operand:VI1_AVX512VNNI 1 "register_operand")
+   (match_operand:VI1_AVX512VNNI 2 "register_operand")
+   (match_operand:<VI1SI> 3 "register_operand")]
+  "(<MODE_SIZE> == 64
+    ||((TARGET_AVX512VNNI && TARGET_AVX512VL)
+           || TARGET_AVXVNNI))"
+{
+  operands[1] = lowpart_subreg (<VI1SI>mode, operands[1], <MODE>mode);
+  operands[2] = lowpart_subreg (<VI1SI>mode, operands[2], <MODE>mode);
+  emit_insn (gen_rtx_SET (operands[0], operands[3]));
+  emit_insn (gen_vpdpbusd_<vi1si> (operands[0], operands[3],
+                                 operands[1], operands[2]));
+  DONE;
+})
+
 (define_insn "vpdpbusd_v16si"
   [(set (match_operand:V16SI 0 "register_operand" "=v")
        (unspec:V16SI

diff --git a/gcc/testsuite/gcc.target/i386/vnni-auto-vectorize-1.c b/gcc/testsuite/gcc.target/i386/vnni-auto-vectorize-1.c
new file mode 100644 (file)
index 0000000..844f37d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/vnni-auto-vectorize-1.c
@@ -0,0 +1,30 @@
+/* { dg-do compile } */                                     
+/* { dg-options "-mavx512f -mavx512vnni -mavx512vl -O2" } */
+
+/* { dg-final { scan-assembler "vpdpwssd\t" } } */
+/* { dg-final { scan-assembler "vpdpbusd\t" } } */
+/* { dg-final { scan-assembler-not "vpmaddwd\t" } } */
+
+int __attribute__((noinline, noclone, optimize("tree-vectorize")))
+sdot_prod_hi (short * restrict a, short * restrict b,
+             int c, int n)
+{
+  int i;
+  for (i = 0; i < n; i++)
+    {
+      c += ((int) a[i] * (int) b[i]);
+    }
+  return c;
+}
+
+int __attribute__((noinline, noclone, optimize("tree-vectorize")))
+usdot_prod_qi (unsigned char * restrict a, char *restrict b,
+              int c, int n)
+{
+  int i;
+  for (i = 0; i < n; i++)
+    {
+      c += ((int) a[i] * (int) b[i]);
+    }
+  return c;
+}

diff --git a/gcc/testsuite/gcc.target/i386/vnni-auto-vectorize-2.c b/gcc/testsuite/gcc.target/i386/vnni-auto-vectorize-2.c
new file mode 100644 (file)
index 0000000..dc8047d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/vnni-auto-vectorize-2.c
@@ -0,0 +1,72 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -mavx512vnni -mavx512vl" } */
+/* { dg-require-effective-target avx512vnni } */
+/* { dg-require-effective-target avx512vl } */
+
+static void vnni_test (void);
+#define DO_TEST vnni_test
+#define AVX512VNNI
+#define AVX512VL
+#include "avx512f-check.h"
+#include "vnni-auto-vectorize-1.c"
+
+#define N 256
+unsigned char a_u8[N];
+char b_i8[N];
+short a_i16[N], b_i16[N];
+int i8_exp, i8_ref, i16_exp, i16_ref;
+
+int __attribute__((noinline, noclone, optimize("no-tree-vectorize")))
+sdot_prod_hi_scalar (short * restrict a, short * restrict b,
+                    int c, int n)
+{
+  int i;
+  for (i = 0; i < n; i++)
+    {
+      c += ((int) a[i] * (int) b[i]);
+    }
+  return c;
+}
+
+int __attribute__((noinline, noclone, optimize("no-tree-vectorize")))
+usdot_prod_qi_scalar (unsigned char * restrict a, char *restrict b,
+              int c, int n)
+{
+  int i;
+  for (i = 0; i < n; i++)
+    {
+      c += ((int) a[i] * (int) b[i]);
+    }
+  return c;
+}
+
+void init()
+{
+  int i;
+
+  i8_exp = i8_ref = 127;
+  i16_exp = i16_ref = 65535;
+
+  for (i = 0; i < N; i++)
+    {
+      a_u8[i] = (i + 3) % 256;
+      b_i8[i] = (i + 1) % 128; 
+      a_i16[i] = i * 2;
+      b_i16[i] = -i + 2;
+    }
+}
+
+static void vnni_test()
+{
+  init ();
+  i16_exp = sdot_prod_hi (a_i16, b_i16, i16_exp, N);
+  i16_ref = sdot_prod_hi_scalar (a_i16, b_i16, i16_ref, N);
+  if (i16_exp != i16_ref)
+    abort ();
+
+  init ();
+  i8_exp = usdot_prod_qi (a_u8, b_i8, i8_exp, N);
+  i8_ref = usdot_prod_qi_scalar (a_u8, b_i8, i8_ref, N);
+  if (i8_exp != i8_ref)
+    abort ();
+}