vppinfra: AVX512 mask load/stores and compress store

Type: improvement
Change-Id: Id6be598aade072653e408cca465e62931d060233
Signed-off-by: Damjan Marion <damarion@cisco.com>

3 files changed, 55 insertions, 37 deletions

diff --git a/src/vppinfra/vector_avx2.h b/src/vppinfra/vector_avx2.h
index 69752275f08..c24ed728c3c 100644
--- a/src/vppinfra/vector_avx2.h
+++ b/src/vppinfra/vector_avx2.h
@@ -213,14 +213,6 @@ u16x16_mask_last (u16x16 v, u8 n_last)
   return v & masks[16 - n_last];
 }
 
-#ifdef __AVX512F__
-static_always_inline u8x32
-u8x32_mask_load (u8x32 a, void *p, u32 mask)
-{
-  return (u8x32) _mm256_mask_loadu_epi8 ((__m256i) a, mask, p);
-}
-#endif
-
 static_always_inline f32x8
 f32x8_from_u32x8 (u32x8 v)
 {
diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h
index 87999eee947..a688baec7fb 100644
--- a/src/vppinfra/vector_avx512.h
+++ b/src/vppinfra/vector_avx512.h
@@ -202,17 +202,43 @@ u32x16_sum_elts (u32x16 sum16)
   return sum8[0] + sum8[4];
 }
 
-static_always_inline u8x64
-u8x64_mask_load (u8x64 a, void *p, u64 mask)
-{
-  return (u8x64) _mm512_mask_loadu_epi8 ((__m512i) a, mask, p);
-}
+#define _(t, m, p, i, e)                                                      \
+  static_always_inline t t##_mask_load (t a, void *p, m mask)                 \
+  {                                                                           \
+    return (t) p##_mask_loadu_##e ((i) a, mask, p);                           \
+  }                                                                           \
+  static_always_inline t t##_mask_load_zero (void *p, m mask)                 \
+  {                                                                           \
+    return (t) p##_maskz_loadu_##e (mask, p);                                 \
+  }                                                                           \
+  static_always_inline void t##_mask_store (t a, void *p, m mask)             \
+  {                                                                           \
+    p##_mask_storeu_##e (p, mask, (i) a);                                     \
+  }
 
-static_always_inline void
-u8x64_mask_store (u8x64 a, void *p, u64 mask)
-{
-  _mm512_mask_storeu_epi8 (p, mask, (__m512i) a);
-}
+_ (u8x64, u64, _mm512, __m512i, epi8)
+_ (u8x32, u32, _mm256, __m256i, epi8)
+_ (u8x16, u16, _mm, __m128i, epi8)
+_ (u16x32, u32, _mm512, __m512i, epi16)
+_ (u16x16, u16, _mm256, __m256i, epi16)
+_ (u16x8, u8, _mm, __m128i, epi16)
+_ (u32x16, u16, _mm512, __m512i, epi32)
+_ (u32x8, u8, _mm256, __m256i, epi32)
+_ (u32x4, u8, _mm, __m128i, epi32)
+_ (u64x8, u8, _mm512, __m512i, epi64)
+_ (u64x4, u8, _mm256, __m256i, epi64)
+_ (u64x2, u8, _mm, __m128i, epi64)
+#undef _
+
+#ifdef CLIB_HAVE_VEC512
+#define CLIB_HAVE_VEC512_MASK_LOAD_STORE
+#endif
+#ifdef CLIB_HAVE_VEC256
+#define CLIB_HAVE_VEC256_MASK_LOAD_STORE
+#endif
+#ifdef CLIB_HAVE_VEC128
+#define CLIB_HAVE_VEC128_MASK_LOAD_STORE
+#endif
 
 static_always_inline u8x64
 u8x64_splat_u8x16 (u8x16 a)
@@ -267,25 +293,33 @@ _ (u32x8, u16x8, _mm256_cvtusepi32_epi16, __m256i)
 _ (u32x8, u64x8, _mm512_cvtepu32_epi64, __m256i)
 #undef _
 
-#define _(vt, mt, bits, epi)                                                  \
+#define _(vt, mt, p, it, epi)                                                 \
   static_always_inline vt vt##_compress (vt a, mt mask)                       \
   {                                                                           \
-    return (vt) _mm##bits##_maskz_compress_##epi (mask, (__m##bits##i) a);    \
+    return (vt) p##_maskz_compress_##epi (mask, (it) a);                      \
   }                                                                           \
   static_always_inline vt vt##_expand (vt a, mt mask)                         \
   {                                                                           \
-    return (vt) _mm##bits##_maskz_expand_##epi (mask, (__m##bits##i) a);      \
+    return (vt) p##_maskz_expand_##epi (mask, (it) a);                        \
+  }                                                                           \
+  static_always_inline void vt##_compress_store (vt v, mt mask, void *p)      \
+  {                                                                           \
+    p##_mask_compressstoreu_##epi (p, mask, (it) v);                          \
   }
 
-_ (u64x8, u8, 512, epi64)
-_ (u32x16, u16, 512, epi32)
-_ (u64x4, u8, 256, epi64)
-_ (u32x8, u8, 256, epi32)
+_ (u64x8, u8, _mm512, __m512i, epi64)
+_ (u32x16, u16, _mm512, __m512i, epi32)
+_ (u64x4, u8, _mm256, __m256i, epi64)
+_ (u32x8, u8, _mm256, __m256i, epi32)
+_ (u64x2, u8, _mm, __m128i, epi64)
+_ (u32x4, u8, _mm, __m128i, epi32)
 #ifdef __AVX512VBMI2__
-_ (u16x32, u32, 512, epi16)
-_ (u8x64, u64, 512, epi8)
-_ (u16x16, u16, 256, epi16)
-_ (u8x32, u32, 256, epi8)
+_ (u16x32, u32, _mm512, __m512i, epi16)
+_ (u8x64, u64, _mm512, __m512i, epi8)
+_ (u16x16, u16, _mm256, __m256i, epi16)
+_ (u8x32, u32, _mm256, __m256i, epi8)
+_ (u16x8, u8, _mm, __m128i, epi16)
+_ (u8x16, u16, _mm, __m128i, epi8)
 #endif
 #undef _
 
diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h
index effab3fe487..06952f117dd 100644
--- a/src/vppinfra/vector_sse42.h
+++ b/src/vppinfra/vector_sse42.h
@@ -764,14 +764,6 @@ u8x16_xor3 (u8x16 a, u8x16 b, u8x16 c)
   return a ^ b ^ c;
 }
 
-#ifdef __AVX512F__
-static_always_inline u8x16
-u8x16_mask_load (u8x16 a, void *p, u16 mask)
-{
-  return (u8x16) _mm_mask_loadu_epi8 ((__m128i) a, mask, p);
-}
-#endif
-
 #endif /* included_vector_sse2_h */
 
 /*