efd: add AVX2 vector lookup function
authorPablo de Lara <pablo.de.lara.guarch@intel.com>
Tue, 17 Jan 2017 22:23:52 +0000 (22:23 +0000)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Wed, 18 Jan 2017 19:53:45 +0000 (20:53 +0100)
Signed-off-by: Byron Marohn <byron.marohn@intel.com>
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Signed-off-by: Saikrishna Edupuganti <saikrishna.edupuganti@intel.com>
Acked-by: Christian Maciocco <christian.maciocco@intel.com>
lib/librte_efd/rte_efd.c
lib/librte_efd/rte_efd_x86.h [new file with mode: 0644]

index 2bcfd62..68e6dab 100644 (file)
@@ -52,6 +52,9 @@
 #include <rte_hash_crc.h>
 
 #include "rte_efd.h"
+#if defined(RTE_ARCH_X86)
+#include "rte_efd_x86.h"
+#endif
 
 #define EFD_KEY(key_idx, table) (table->keys + ((key_idx) * table->key_len))
 /** Hash function used to determine chunk_id and bin_id for a group */
@@ -100,6 +103,7 @@ allocated memory
 /* All different internal lookup functions */
 enum efd_lookup_internal_function {
        EFD_LOOKUP_SCALAR = 0,
+       EFD_LOOKUP_AVX2,
        EFD_LOOKUP_NUM
 };
 
@@ -662,7 +666,16 @@ rte_efd_create(const char *name, uint32_t max_num_rules, uint32_t key_len,
                }
        }
 
-       table->lookup_fn = EFD_LOOKUP_SCALAR;
+#if defined(RTE_ARCH_X86)
+       /*
+        * For less than 4 bits, scalar function performs better
+        * than vectorised version
+        */
+       if (RTE_EFD_VALUE_NUM_BITS > 3 && rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
+               table->lookup_fn = EFD_LOOKUP_AVX2;
+       else
+#endif
+               table->lookup_fn = EFD_LOOKUP_SCALAR;
 
        /*
         * Allocate the EFD table offline portion (with the actual rules
@@ -1253,6 +1266,13 @@ efd_lookup_internal(const struct efd_online_group_entry * const group,
 
        switch (lookup_fn) {
 
+#if defined(RTE_ARCH_X86)
+       case EFD_LOOKUP_AVX2:
+               return efd_lookup_internal_avx2(group->hash_idx,
+                                       group->lookup_table,
+                                       hash_val_a,
+                                       hash_val_b);
+#endif
        case EFD_LOOKUP_SCALAR:
        /* Fall-through */
        default:
diff --git a/lib/librte_efd/rte_efd_x86.h b/lib/librte_efd/rte_efd_x86.h
new file mode 100644 (file)
index 0000000..34f37d7
--- /dev/null
@@ -0,0 +1,86 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2016-2017 Intel Corporation. All rights reserved.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/* rte_efd_x86.h
+ * This file holds all x86 specific EFD functions
+ */
+#include <immintrin.h>
+
+#if (RTE_EFD_VALUE_NUM_BITS == 8 || RTE_EFD_VALUE_NUM_BITS == 16 || \
+       RTE_EFD_VALUE_NUM_BITS == 24 || RTE_EFD_VALUE_NUM_BITS == 32)
+#define EFD_LOAD_SI128(val) _mm_load_si128(val)
+#else
+#define EFD_LOAD_SI128(val) _mm_lddqu_si128(val)
+#endif
+
+static inline efd_value_t
+efd_lookup_internal_avx2(const efd_hashfunc_t *group_hash_idx,
+               const efd_lookuptbl_t *group_lookup_table,
+               const uint32_t hash_val_a, const uint32_t hash_val_b)
+{
+#ifdef RTE_MACHINE_CPUFLAG_AVX2
+       efd_value_t value = 0;
+       uint32_t i = 0;
+       __m256i vhash_val_a = _mm256_set1_epi32(hash_val_a);
+       __m256i vhash_val_b = _mm256_set1_epi32(hash_val_b);
+
+       for (; i < RTE_EFD_VALUE_NUM_BITS; i += 8) {
+               __m256i vhash_idx =
+                               _mm256_cvtepu16_epi32(EFD_LOAD_SI128(
+                               (__m128i const *) &group_hash_idx[i]));
+               __m256i vlookup_table = _mm256_cvtepu16_epi32(
+                               EFD_LOAD_SI128((__m128i const *)
+                               &group_lookup_table[i]));
+               __m256i vhash = _mm256_add_epi32(vhash_val_a,
+                               _mm256_mullo_epi32(vhash_idx, vhash_val_b));
+               __m256i vbucket_idx = _mm256_srli_epi32(vhash,
+                               EFD_LOOKUPTBL_SHIFT);
+               __m256i vresult = _mm256_srlv_epi32(vlookup_table,
+                               vbucket_idx);
+
+               value |= (_mm256_movemask_ps(
+                       (__m256) _mm256_slli_epi32(vresult, 31))
+                       & ((1 << (RTE_EFD_VALUE_NUM_BITS - i)) - 1)) << i;
+       }
+
+       return value;
+#else
+       RTE_SET_USED(group_hash_idx);
+       RTE_SET_USED(group_lookup_table);
+       RTE_SET_USED(hash_val_a);
+       RTE_SET_USED(hash_val_b);
+       /* Return dummy value, only to avoid compilation breakage */
+       return 0;
+#endif
+
+}