vppinfra: add support for mask_compare for u64

Type: improvement

Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
Change-Id: Ie323a8445f8540cd26d92eb61807afad7f7a0b74

2 files changed, 103 insertions, 0 deletions

diff --git a/src/vppinfra/vector/mask_compare.h b/src/vppinfra/vector/mask_compare.h
index cac48a31f47..92d5ca35474 100644
--- a/src/vppinfra/vector/mask_compare.h
+++ b/src/vppinfra/vector/mask_compare.h
@@ -163,4 +163,63 @@ clib_mask_compare_u32 (u32 v, u32 *a, u64 *bitmap, u32 n_elts)
   bitmap[0] = clib_mask_compare_u32_x64 (v, a, n_elts) & pow2_mask (n_elts);
 }
 
+static_always_inline u64
+clib_mask_compare_u64_x64 (u64 v, u64 *a, u32 n_elts)
+{
+  u64 mask = 0;
+#if defined(CLIB_HAVE_VEC512)
+  u64x8 v8 = u64x8_splat (v);
+  u64x8u *av = (u64x8u *) a;
+  mask = ((u64) u64x8_is_equal_mask (av[0], v8) |
+	  (u64) u64x8_is_equal_mask (av[1], v8) << 8 |
+	  (u64) u64x8_is_equal_mask (av[2], v8) << 16 |
+	  (u64) u64x8_is_equal_mask (av[3], v8) << 24 |
+	  (u64) u64x8_is_equal_mask (av[4], v8) << 32 |
+	  (u64) u64x8_is_equal_mask (av[5], v8) << 40 |
+	  (u64) u64x8_is_equal_mask (av[6], v8) << 48 |
+	  (u64) u64x8_is_equal_mask (av[7], v8) << 56);
+
+#elif defined(CLIB_HAVE_VEC256) && defined(__BMI2__)
+  u64x4 v4 = u64x4_splat (v);
+  u64x4u *av = (u64x4u *) a;
+
+  for (int i = 0; i < 16; i += 2)
+    {
+      u64 l = u8x32_msb_mask (v4 == av[i]);
+      u64 h = u8x32_msb_mask (v4 == av[i + 1]);
+      mask |= _pext_u64 (l | h << 32, 0x0101010101010101) << (i * 4);
+    }
+#else
+  for (int i = 0; i < n_elts; i++)
+    if (a[i] == v)
+      mask |= 1ULL << i;
+#endif
+  return mask;
+}
+
+/** \brief Compare 64-bit elemments with provied value and return bitmap
+
+    @param v value to compare elements with
+    @param a array of u64 elements
+    @param mask array of u64 where reuslting mask will be stored
+    @param n_elts number of elements in the array
+    @return none
+*/
+
+static_always_inline void
+clib_mask_compare_u64 (u64 v, u64 *a, u64 *bitmap, u32 n_elts)
+{
+  while (n_elts >= 64)
+    {
+      bitmap++[0] = clib_mask_compare_u64_x64 (v, a, 64);
+      n_elts -= 64;
+      a += 64;
+    }
+
+  if (PREDICT_TRUE (n_elts == 0))
+    return;
+
+  bitmap[0] = clib_mask_compare_u64_x64 (v, a, n_elts) & pow2_mask (n_elts);
+}
+
 #endif
diff --git a/src/vppinfra/vector/test/mask_compare.c b/src/vppinfra/vector/test/mask_compare.c
index 40cac1b6ce3..70608e11215 100644
--- a/src/vppinfra/vector/test/mask_compare.c
+++ b/src/vppinfra/vector/test/mask_compare.c
@@ -18,6 +18,12 @@ clib_mask_compare_u32_wrapper (u32 v, u32 *a, u64 *mask, u32 n_elts)
   clib_mask_compare_u32 (v, a, mask, n_elts);
 }
 
+__test_funct_fn void
+clib_mask_compare_u64_wrapper (u64 v, u64 *a, u64 *mask, u64 n_elts)
+{
+  clib_mask_compare_u64 (v, a, mask, n_elts);
+}
+
 static clib_error_t *
 test_clib_mask_compare_u16 (clib_error_t *err)
 {
@@ -93,3 +99,41 @@ REGISTER_TEST (clib_mask_compare_u32) = {
   .name = "clib_mask_compare_u32",
   .fn = test_clib_mask_compare_u32,
 };
+
+static clib_error_t *
+test_clib_mask_compare_u64 (clib_error_t *err)
+{
+  u64 array[513];
+  u64 mask[10];
+  u32 i, j;
+
+  for (i = 0; i < ARRAY_LEN (array); i++)
+    array[i] = i;
+
+  for (i = 0; i < ARRAY_LEN (array); i++)
+    {
+      for (j = 0; j < ARRAY_LEN (mask); j++)
+	mask[j] = 0xa5a5a5a5a5a5a5a5;
+
+      clib_mask_compare_u64_wrapper (i, array, mask, i + 1);
+
+      for (j = 0; j < (i >> 6); j++)
+	{
+	  if (mask[j])
+	    return clib_error_return (err, "mask at position %u not zero", j);
+	}
+      if (mask[j] != 1ULL << (i & 0x3f))
+	return clib_error_return (err,
+				  "mask at position %u is %lx, expected %lx",
+				  j, mask[j], 1ULL << (i % 64));
+
+      if (mask[j + 1] != 0xa5a5a5a5a5a5a5a5)
+	return clib_error_return (err, "mask overrun at position %u", j + 1);
+    }
+  return err;
+}
+
+REGISTER_TEST (clib_mask_compare_u64) = {
+  .name = "clib_mask_compare_u64",
+  .fn = test_clib_mask_compare_u64,
+};