summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorLeyi Rong <leyi.rong@intel.com>2023-03-08 13:34:56 +0800
committerDamjan Marion <dmarion@0xa5.net>2023-03-10 15:57:09 +0000
commit767a9ab4b1f77760a633ba3de30a0dcad44cb4b1 (patch)
treedc742984919960e2dda931734ae2650ce2b5f451
parent7b0d48fbb3ceebcafb18177fba02ab1a9ad4c24e (diff)
vlib: 512-bit SIMD version of vlib_buffer_free
Process 8 packets perf batch in vlib_buffer_free_inline() when CLIB_HAVE_VEC512 is enabled. Type: improvement Signed-off-by: Leyi Rong <leyi.rong@intel.com> Change-Id: I78b8a525bce25ee355c9bf0e0f651698a8c45bda
-rw-r--r--src/vlib/buffer_funcs.h70
-rw-r--r--src/vppinfra/vector/index_to_ptr.h5
2 files changed, 73 insertions, 2 deletions
diff --git a/src/vlib/buffer_funcs.h b/src/vlib/buffer_funcs.h
index 708539957c9..017a70f4fbf 100644
--- a/src/vlib/buffer_funcs.h
+++ b/src/vlib/buffer_funcs.h
@@ -788,9 +788,16 @@ vlib_buffer_free_inline (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
vlib_buffer_t *b[8];
u32 bi, sum = 0, flags, next;
+#if defined(CLIB_HAVE_VEC512)
+ if (n_buffers < 8)
+#else
if (n_buffers < 4)
+#endif
goto one_by_one;
+#if defined(CLIB_HAVE_VEC512)
+ vlib_get_buffers (vm, buffers, b, 8);
+#else
vlib_get_buffers (vm, buffers, b, 4);
if (n_buffers >= 12)
@@ -801,8 +808,33 @@ vlib_buffer_free_inline (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
vlib_prefetch_buffer_header (b[6], LOAD);
vlib_prefetch_buffer_header (b[7], LOAD);
}
+#endif
-#if defined(CLIB_HAVE_VEC128)
+#if defined(CLIB_HAVE_VEC512)
+ u8x16 p0, p1, p2, p3, p4, p5, p6, p7, r;
+ p0 = u8x16_load_unaligned (b[0]);
+ p1 = u8x16_load_unaligned (b[1]);
+ p2 = u8x16_load_unaligned (b[2]);
+ p3 = u8x16_load_unaligned (b[3]);
+ p4 = u8x16_load_unaligned (b[4]);
+ p5 = u8x16_load_unaligned (b[5]);
+ p6 = u8x16_load_unaligned (b[6]);
+ p7 = u8x16_load_unaligned (b[7]);
+
+ r = p0 ^ bpi_vec.as_u8x16[0];
+ r |= p1 ^ bpi_vec.as_u8x16[0];
+ r |= p2 ^ bpi_vec.as_u8x16[0];
+ r |= p3 ^ bpi_vec.as_u8x16[0];
+ r |= p4 ^ bpi_vec.as_u8x16[0];
+ r |= p5 ^ bpi_vec.as_u8x16[0];
+ r |= p6 ^ bpi_vec.as_u8x16[0];
+ r |= p7 ^ bpi_vec.as_u8x16[0];
+ r &= bpi_mask.as_u8x16[0];
+ r |=
+ (p0 | p1 | p2 | p3 | p4 | p5 | p6 | p7) & flags_refs_mask.as_u8x16[0];
+
+ sum = !u8x16_is_all_zero (r);
+#elif defined(CLIB_HAVE_VEC128)
u8x16 p0, p1, p2, p3, r;
p0 = u8x16_load_unaligned (b[0]);
p1 = u8x16_load_unaligned (b[1]);
@@ -836,6 +868,36 @@ vlib_buffer_free_inline (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
if (sum)
goto one_by_one;
+#if defined(CLIB_HAVE_VEC512)
+ vlib_buffer_copy_indices (queue + n_queue, buffers, 8);
+ vlib_buffer_copy_template (b[0], &bt);
+ vlib_buffer_copy_template (b[1], &bt);
+ vlib_buffer_copy_template (b[2], &bt);
+ vlib_buffer_copy_template (b[3], &bt);
+ vlib_buffer_copy_template (b[4], &bt);
+ vlib_buffer_copy_template (b[5], &bt);
+ vlib_buffer_copy_template (b[6], &bt);
+ vlib_buffer_copy_template (b[7], &bt);
+ n_queue += 8;
+
+ vlib_buffer_validate (vm, b[0]);
+ vlib_buffer_validate (vm, b[1]);
+ vlib_buffer_validate (vm, b[2]);
+ vlib_buffer_validate (vm, b[3]);
+ vlib_buffer_validate (vm, b[4]);
+ vlib_buffer_validate (vm, b[5]);
+ vlib_buffer_validate (vm, b[6]);
+ vlib_buffer_validate (vm, b[7]);
+
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[0]);
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[1]);
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[2]);
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[3]);
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[4]);
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[5]);
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[6]);
+ VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[7]);
+#else
vlib_buffer_copy_indices (queue + n_queue, buffers, 4);
vlib_buffer_copy_template (b[0], &bt);
vlib_buffer_copy_template (b[1], &bt);
@@ -852,14 +914,20 @@ vlib_buffer_free_inline (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[1]);
VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[2]);
VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[3]);
+#endif
if (n_queue >= queue_size)
{
vlib_buffer_pool_put (vm, buffer_pool_index, queue, n_queue);
n_queue = 0;
}
+#if defined(CLIB_HAVE_VEC512)
+ buffers += 8;
+ n_buffers -= 8;
+#else
buffers += 4;
n_buffers -= 4;
+#endif
continue;
one_by_one:
diff --git a/src/vppinfra/vector/index_to_ptr.h b/src/vppinfra/vector/index_to_ptr.h
index 91de3546439..3985b757d54 100644
--- a/src/vppinfra/vector/index_to_ptr.h
+++ b/src/vppinfra/vector/index_to_ptr.h
@@ -92,7 +92,7 @@ clib_index_to_ptr_u32 (u32 *indices, void *base, u8 shift, void **ptrs,
indices += 16;
n_elts -= 16;
}
- if (n_elts > 8)
+ if (n_elts >= 8)
{
b0 = u64x8_from_u32x8 (u32x8_load_unaligned (indices));
u64x8_store_unaligned ((b0 << shift) + off, ptrs);
@@ -101,6 +101,9 @@ clib_index_to_ptr_u32 (u32 *indices, void *base, u8 shift, void **ptrs,
n_elts -= 8;
}
+ if (n_elts == 0)
+ return;
+
b0 = u64x8_from_u32x8 (u32x8_load_unaligned (indices + n_elts - 8));
u64x8_store_unaligned ((b0 << shift) + off, ptrs + n_elts - 8);
}