Use aligned load/store in AVX2 Slide Hash.

Also test slidehash from 512 bytes, the minimum window size we use.

diff --git a/arch/x86/slide_hash_avx2.c b/arch/x86/slide_hash_avx2.c
index 853347323436eb43e14653a44fbee979e2ed8bb7..1000012392a773babaec4048d176bd639f504734 100644 (file)
--- a/arch/x86/slide_hash_avx2.c
+++ b/arch/x86/slide_hash_avx2.c
@@ -21,9 +21,9 @@ static inline void slide_hash_chain(Pos *table, uint32_t entries, const __m256i
     do {
         __m256i value, result;
 
-        value = _mm256_loadu_si256((__m256i *)table);
+        value = _mm256_load_si256((__m256i *)table);
         result = _mm256_subs_epu16(value, wsize);
-        _mm256_storeu_si256((__m256i *)table, result);
+        _mm256_store_si256((__m256i *)table, result);
 
         table -= 16;
         entries -= 16;

diff --git a/test/benchmarks/benchmark_slidehash.cc b/test/benchmarks/benchmark_slidehash.cc
index 9d98420b16724b620168496e6a5fa79901ff737c..53e9516817770141a7e4d03d53de7b15143d03cf 100644 (file)
--- a/test/benchmarks/benchmark_slidehash.cc
+++ b/test/benchmarks/benchmark_slidehash.cc
@@ -25,13 +25,13 @@ private:
 
 public:
     void SetUp(const ::benchmark::State& state) {
-        l0 = (uint16_t *)zng_alloc(HASH_SIZE * sizeof(uint16_t));
+        l0 = (uint16_t *)zng_alloc_aligned(HASH_SIZE * sizeof(uint16_t), 64);
 
         for (uint32_t i = 0; i < HASH_SIZE; i++) {
             l0[i] = rand();
         }
 
-        l1 = (uint16_t *)zng_alloc(MAX_RANDOM_INTS * sizeof(uint16_t));
+        l1 = (uint16_t *)zng_alloc_aligned(MAX_RANDOM_INTS * sizeof(uint16_t), 64);
 
         for (int32_t i = 0; i < MAX_RANDOM_INTS; i++) {
             l1[i] = rand();
@@ -53,8 +53,8 @@ public:
     }
 
     void TearDown(const ::benchmark::State& state) {
-        zng_free(l0);
-        zng_free(l1);
+        zng_free_aligned(l0);
+        zng_free_aligned(l1);
         free(s_g);
     }
 };
@@ -66,7 +66,7 @@ public:
         } \
         Bench(state, fptr); \
     } \
-    BENCHMARK_REGISTER_F(slide_hash, name)->RangeMultiplier(2)->Range(1024, MAX_RANDOM_INTS);
+    BENCHMARK_REGISTER_F(slide_hash, name)->RangeMultiplier(2)->Range(512, MAX_RANDOM_INTS);
 
 BENCHMARK_SLIDEHASH(c, slide_hash_c, 1);