vppinfra: fix x86 packs / packus wrappers

They both take signed value as input.

Type: fix
Change-Id: If3d8ec4e0b1c02d7d65262bdd9db49ff7fbfef39
Signed-off-by: Damjan Marion <damarion@cisco.com>

4 files changed, 52 insertions, 94 deletions

diff --git a/src/vppinfra/vector_avx2.h b/src/vppinfra/vector_avx2.h
index 584bd207b27..f38a3bdae73 100644
--- a/src/vppinfra/vector_avx2.h
+++ b/src/vppinfra/vector_avx2.h
@@ -105,17 +105,19 @@ _(u64x2, u64x4)
 #undef _
 /* *INDENT-ON* */
 
-always_inline u8x32
-u16x16_pack (u16x16 lo, u16x16 hi)
-{
-  return (u8x32) _mm256_packus_epi16 ((__m256i) lo, (__m256i) hi);
-}
+/* 256 bit packs. */
+#define _(f, t, fn)                                                           \
+  always_inline t t##_pack (f lo, f hi)                                       \
+  {                                                                           \
+    return (t) fn ((__m256i) lo, (__m256i) hi);                               \
+  }
+
+_ (i16x16, i8x32, _mm256_packs_epi16)
+_ (i16x16, u8x32, _mm256_packus_epi16)
+_ (i32x8, i16x16, _mm256_packs_epi32)
+_ (i32x8, u16x16, _mm256_packus_epi32)
 
-always_inline i8x32
-i16x16_pack (i16x16 lo, i16x16 hi)
-{
-  return (i8x32) _mm256_packs_epi16 ((__m256i) lo, (__m256i) hi);
-}
+#undef _
 
 static_always_inline u32
 u8x32_msb_mask (u8x32 v)
diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h
index 2f5763e3c92..3a01c1ed824 100644
--- a/src/vppinfra/vector_avx512.h
+++ b/src/vppinfra/vector_avx512.h
@@ -85,6 +85,19 @@ u16x32_msb_mask (u16x32 v)
   return (u32) _mm512_movepi16_mask ((__m512i) v);
 }
 
+/* 512-bit packs */
+#define _(f, t, fn)                                                           \
+  always_inline t t##_pack (f lo, f hi)                                       \
+  {                                                                           \
+    return (t) fn ((__m512i) lo, (__m512i) hi);                               \
+  }
+
+_ (i16x32, i8x64, _mm512_packs_epi16)
+_ (i16x32, u8x64, _mm512_packus_epi16)
+_ (i32x16, i16x32, _mm512_packs_epi32)
+_ (i32x16, u16x32, _mm512_packus_epi32)
+#undef _
+
 static_always_inline u32x16
 u32x16_byte_swap (u32x16 v)
 {
diff --git a/src/vppinfra/vector_funcs.h b/src/vppinfra/vector_funcs.h
index 2b02d9eb301..5c446a5d50d 100644
--- a/src/vppinfra/vector_funcs.h
+++ b/src/vppinfra/vector_funcs.h
@@ -27,37 +27,36 @@ clib_compare_u16_x64 (u16 v, u16 *a)
   u16x16u *av = (u16x16u *) a;
   i8x32 x;
 
-  x = i16x16_pack (v16 == av[0], v16 == av[1]);
+  x = i8x32_pack (v16 == av[0], v16 == av[1]);
   mask = i8x32_msb_mask ((i8x32) u64x4_permute (x, 0, 2, 1, 3));
-  x = i16x16_pack (v16 == av[2], v16 == av[3]);
+  x = i8x32_pack (v16 == av[2], v16 == av[3]);
   mask |= (u64) i8x32_msb_mask ((i8x32) u64x4_permute (x, 0, 2, 1, 3)) << 32;
 #elif defined(CLIB_HAVE_VEC128) && defined(__ARM_NEON)
-  u16x8 idx8 = u16x8_splat (v);
+  u16x8 v8 = u16x8_splat (v);
   u16x8 m = { 1, 2, 4, 8, 16, 32, 64, 128 };
   u16x8u *av = (u16x8u *) a;
 
-  /* compare each u16 elemment with idx8, result gives 0xffff in each element
+  /* compare each u16 elemment with v8, result gives 0xffff in each element
      of the resulting vector if comparison result is true.
      Bitwise AND with m will give us one bit set for true result and offset
      of that bit represend element index. Finally vaddvq_u16() gives us sum
      of all elements of the vector which will give us u8 bitmap. */
 
-  mask = ((u64) vaddvq_u16 ((av[0] == idx8) & m) |
-	  (u64) vaddvq_u16 ((av[1] == idx8) & m) << 8 |
-	  (u64) vaddvq_u16 ((av[2] == idx8) & m) << 16 |
-	  (u64) vaddvq_u16 ((av[3] == idx8) & m) << 24 |
-	  (u64) vaddvq_u16 ((av[4] == idx8) & m) << 32 |
-	  (u64) vaddvq_u16 ((av[5] == idx8) & m) << 40 |
-	  (u64) vaddvq_u16 ((av[6] == idx8) & m) << 48 |
-	  (u64) vaddvq_u16 ((av[7] == idx8) & m) << 56);
+  mask = ((u64) vaddvq_u16 ((av[0] == v8) & m) |
+	  (u64) vaddvq_u16 ((av[1] == v8) & m) << 8 |
+	  (u64) vaddvq_u16 ((av[2] == v8) & m) << 16 |
+	  (u64) vaddvq_u16 ((av[3] == v8) & m) << 24 |
+	  (u64) vaddvq_u16 ((av[4] == v8) & m) << 32 |
+	  (u64) vaddvq_u16 ((av[5] == v8) & m) << 40 |
+	  (u64) vaddvq_u16 ((av[6] == v8) & m) << 48 |
+	  (u64) vaddvq_u16 ((av[7] == v8) & m) << 56);
 #elif defined(CLIB_HAVE_VEC128) && defined(CLIB_HAVE_VEC128_MSB_MASK)
-  u16x8 idx8 = u16x8_splat (v);
+  u16x8 v8 = u16x8_splat (v);
   u16x8u *av = (u16x8u *) a;
-  mask =
-    ((u64) i8x16_msb_mask (i16x8_pack (idx8 == av[0], idx8 == av[1])) |
-     (u64) i8x16_msb_mask (i16x8_pack (idx8 == av[2], idx8 == av[3])) << 16 |
-     (u64) i8x16_msb_mask (i16x8_pack (idx8 == av[4], idx8 == av[5])) << 32 |
-     (u64) i8x16_msb_mask (i16x8_pack (idx8 == av[6], idx8 == av[7])) << 48);
+  mask = ((u64) i8x16_msb_mask (i8x16_pack (v8 == av[0], v8 == av[1])) |
+	  (u64) i8x16_msb_mask (i8x16_pack (v8 == av[2], v8 == av[3])) << 16 |
+	  (u64) i8x16_msb_mask (i8x16_pack (v8 == av[4], v8 == av[5])) << 32 |
+	  (u64) i8x16_msb_mask (i8x16_pack (v8 == av[6], v8 == av[7])) << 48);
 #else
   for (int i = 0; i < 64; i++)
     if (a[i] == v)
diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h
index f86fad39b02..1bdb34b866e 100644
--- a/src/vppinfra/vector_sse42.h
+++ b/src/vppinfra/vector_sse42.h
@@ -184,74 +184,18 @@ u32x2_interleave_lo (u32x2 a, u32x2 b)
 }
 
 /* 128 bit packs. */
-always_inline u8x16
-u16x8_pack (u16x8 lo, u16x8 hi)
-{
-  return (u8x16) _mm_packus_epi16 ((__m128i) lo, (__m128i) hi);
-}
-
-always_inline i8x16
-i16x8_pack (i16x8 lo, i16x8 hi)
-{
-  return (i8x16) _mm_packs_epi16 ((__m128i) lo, (__m128i) hi);
-}
-
-always_inline u16x8
-u32x4_pack (u32x4 lo, u32x4 hi)
-{
-  return (u16x8) _mm_packs_epi32 ((__m128i) lo, (__m128i) hi);
-}
-
-/* 64 bit packs. */
-always_inline u8x8
-u16x4_pack (u16x4 lo, u16x4 hi)
-{
-  return (u8x8) _m_packuswb ((__m64) lo, (__m64) hi);
-}
-
-always_inline i8x8
-i16x4_pack (i16x4 lo, i16x4 hi)
-{
-  return (i8x8) _m_packsswb ((__m64) lo, (__m64) hi);
-}
-
-always_inline u16x4
-u32x2_pack (u32x2 lo, u32x2 hi)
-{
-  return (u16x4) _m_packssdw ((__m64) lo, (__m64) hi);
-}
-
-always_inline i16x4
-i32x2_pack (i32x2 lo, i32x2 hi)
-{
-  return (i16x4) _m_packssdw ((__m64) lo, (__m64) hi);
-}
-
-#ifndef __ICC
-always_inline u64x2
-u64x2_read_lo (u64x2 x, u64 * a)
-{
-  return (u64x2) _mm_loadl_pi ((__m128) x, (__m64 *) a);
-}
-
-always_inline u64x2
-u64x2_read_hi (u64x2 x, u64 * a)
-{
-  return (u64x2) _mm_loadh_pi ((__m128) x, (__m64 *) a);
-}
+#define _(f, t, fn)                                                           \
+  always_inline t t##_pack (f lo, f hi)                                       \
+  {                                                                           \
+    return (t) fn ((__m128i) lo, (__m128i) hi);                               \
+  }
 
-always_inline void
-u64x2_write_lo (u64x2 x, u64 * a)
-{
-  _mm_storel_pi ((__m64 *) a, (__m128) x);
-}
+_ (i16x8, i8x16, _mm_packs_epi16)
+_ (i16x8, u8x16, _mm_packus_epi16)
+_ (i32x4, i16x8, _mm_packs_epi32)
+_ (i32x4, u16x8, _mm_packus_epi32)
 
-always_inline void
-u64x2_write_hi (u64x2 x, u64 * a)
-{
-  _mm_storeh_pi ((__m64 *) a, (__m128) x);
-}
-#endif
+#undef _
 
 #define _signed_binop(n,m,f,g)                                         \
   /* Unsigned */                                                       \