add longest_common_prefix_avx2_8, but is slower than 64bit version

Author: koeppl

bench_lcp.cpp

@@ -90,6 +90,10 @@ BENCHMARK_F(LCP, sse, LCPFixture, 0, 10000)
 
 
 #ifdef __AVX2__
+BENCHMARK_F(LCP, avx2s, LCPFixture, 0, 10000)
+{
+   celero::DoNotOptimizeAway(longest_common_prefix_avx2_8(instance().m_stra, instance().m_strb, instance().m_length));
+}
 BENCHMARK_F(LCP, avx2, LCPFixture, 0, 10000)
 {
    celero::DoNotOptimizeAway(longest_common_prefix_avx2(instance().m_stra, instance().m_strb, instance().m_length));

lcp.cpp

@@ -136,7 +136,37 @@ size_t longest_common_prefix_sse(const char*const a, const char*const b, const s
 }
 #endif //SSE2
 
+
 #ifdef __AVX2__
+
+size_t longest_common_prefix_avx2_8(const char*const a, const char*const b, const size_t length) {
+   constexpr size_t register_size = 256/8;
+   size_t c_length = 0; // c_length in register_size units
+   for(; c_length < length / register_size; ++c_length) {
+      __m256i ma = _mm256_load_si256((__m256i*)(a+c_length*register_size)); 
+      __m256i mb = _mm256_load_si256((__m256i*)(b+c_length*register_size)); 
+      const unsigned int mask = _mm256_movemask_epi8(_mm256_cmpeq_epi8(ma, mb));
+      if(~mask == 0) {
+	 continue;
+      }
+      if(mask == 0) {
+	 return c_length == 0 ?  0 : (c_length-1)*register_size;
+      }
+      const size_t pos = __builtin_ctz(~mask);
+
+      const size_t ret = c_length*register_size + pos;
+      DCHECK_EQ(ret, longest_common_prefix_character(a, b, length));
+      return ret;
+   }
+   const size_t ret = c_length*register_size + 
+      longest_common_prefix_packed(
+	    a + c_length*register_size,
+	    b + c_length*register_size, 
+	    length - c_length*register_size);
+   DCHECK_EQ(ret, longest_common_prefix_character(a, b, length));
+   return ret;
+}
+
 size_t longest_common_prefix_avx2(const char*const a, const char*const b, const size_t length) {
    constexpr size_t register_size = 256/8;
    size_t c_length = 0; // c_length in register_size units

lcp.hpp

@@ -18,11 +18,18 @@ size_t longest_common_prefix(const uint64_t* a, const uint64_t* b, const size_t
 size_t longest_common_prefix_character(const uint64_t*const a, const uint64_t*const b, const size_t length);
 
 #ifdef __SSE3__
+/**
+ * uses _mm256_cmpeq_epi64 to compare 64-bit packed characters, and resorts to longest_common_prefix_packed for the final packed character
+ */
 size_t longest_common_prefix_sse(const uint64_t*const a, const uint64_t*const b, const size_t length);
 #endif
 
 #ifdef __AVX2__
 size_t longest_common_prefix_avx2(const uint64_t*const a, const uint64_t*const b, const size_t length);
+/**
+ * uses _mm256_cmpeq_epi8 to compare 8-bit characters in blocks of 256 bit
+ */
+size_t longest_common_prefix_avx2_8(const char*const a, const char*const b, const size_t length);
 #endif
 
 

precalc_lcp_common.cpp

@@ -106,6 +106,7 @@ const char*const lcp_name[] =
   , "sse"
 #endif
 #ifdef __AVX2__
+  , "avx2_8"
   , "avx2"
 #endif
 #ifdef __AVX512__
@@ -119,6 +120,7 @@ const char*const lcp_name[] =
     , packed::longest_common_prefix_sse
 #endif
 #ifdef __AVX2__
+    , packed::longest_common_prefix_avx2_8
     , packed::longest_common_prefix_avx2
 #endif
 #ifdef __AVX512__

test_lcp.cpp

@@ -99,6 +99,9 @@ TEST(LCP64_off, sse) {  test64_lcp_offset(longest_common_prefix_sse); }
 #endif
 
 #ifdef __AVX2__
+TEST(LCP, avx2_8) {  test_lcp(longest_common_prefix_avx2_8); }
+TEST(LCPeq, avx2_8) {  test_eq(longest_common_prefix_avx2_8); }
+
 TEST(LCP, avx2) {  test_lcp(longest_common_prefix_avx2); }
 TEST(LCPeq, avx2) {  test_eq(longest_common_prefix_avx2); }