[NSE-927] Enable AVX512 in Binary length calculation for native Colum…

…nartoRow (#924) * Fix AVX512 code of calculating binary length * Fix clang format * Enable NATIVE_AVX512 In C code
oap-project · May 24, 2022 · 10605fb · 10605fb
1 parent c81f8af
commit 10605fb
Show file tree

Hide file tree

Showing 2 changed files with 25 additions and 26 deletions.
diff --git a/native-sql-engine/cpp/src/CMakeLists.txt b/native-sql-engine/cpp/src/CMakeLists.txt
@@ -538,6 +538,7 @@ if(NATIVE_AVX512)
   list(APPEND SPARK_COLUMNAR_PLUGIN_SRCS operators/columnar_to_row_converter_avx512.cc)
   set(NATIVE_AVX512_FLAG "-march=native")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${NATIVE_AVX512_FLAG}")
+  add_definitions(-DNATIVE_AVX512)
 else()
   list(APPEND SPARK_COLUMNAR_PLUGIN_SRCS operators/columnar_to_row_converter.cc)
 endif()

diff --git a/native-sql-engine/cpp/src/operators/columnar_to_row_converter_avx512.cc b/native-sql-engine/cpp/src/operators/columnar_to_row_converter_avx512.cc
@@ -137,35 +137,33 @@ arrow::Status ColumnarToRowConverter::Init(
     if (arrow::is_binary_like(array->type_id())) {
       auto binary_array = std::static_pointer_cast<arrow::BinaryArray>(array);
       using offset_type = typename arrow::BinaryType::offset_type;
-      //      offset_type length;
       const offset_type* offsetarray = binary_array->raw_value_offsets();
-      //      __m256i x7_8x = _mm256_load_si256((__m256i*)x_7);
-      //      __m256i x8_8x = _mm256_load_si256((__m256i*)x_8);
+      __m256i x7_8x = _mm256_load_si256((__m256i*)x_7);
+      __m256i x8_8x = _mm256_load_si256((__m256i*)x_8);
       int32_t j = 0;
       int32_t* length_data = lengths_.data();
-      //
-      //      __m256i offsetarray_1_8x;
-      //      if (j + 16 < num_rows_)
-      //      {
-      //        offsetarray_1_8x = _mm256_load_si256((__m256i*)&offsetarray[j]);
-      //      }
-      //      for (j; j + 16 < num_rows_; j += 8) {
-      //        __m256i offsetarray_8x = offsetarray_1_8x;
-      //        offsetarray_1_8x = _mm256_load_si256((__m256i*)&offsetarray[j+8]);
-      //
-      //        __m256i length_8x =
-      //        _mm256_alignr_epi32(offsetarray_8x,offsetarray_1_8x,0x1); length_8x =
-      //        _mm256_sub_epi32(length_8x, offsetarray_8x);
-      //
-      //        __m256i reminder_8x = _mm256_and_si256(length_8x, x7_8x);
-      //        reminder_8x = _mm256_sub_epi32(x8_8x,reminder_8x);
-      //        reminder_8x = _mm256_and_si256(reminder_8x,x7_8x);
-      //        __m256i dst_length_8x = _mm256_loadu_si256((__m256i*)length_data);
-      //        dst_length_8x = _mm256_add_epi32(dst_length_8x, reminder_8x);
-      //        _mm256_storeu_si256((__m256i*)length_data,dst_length_8x);
-      //        length_data+=8;
-      //        _mm_prefetch(&offsetarray[j+(128+128)/sizeof(offset_type)],_MM_HINT_T0);
-      //      }
+
+      __m256i offsetarray_1_8x;
+      if (j + 16 < num_rows_) {
+        offsetarray_1_8x = _mm256_loadu_si256((__m256i*)&offsetarray[j]);
+      }
+      for (j; j + 16 < num_rows_; j += 8) {
+        __m256i offsetarray_8x = offsetarray_1_8x;
+        offsetarray_1_8x = _mm256_loadu_si256((__m256i*)&offsetarray[j + 8]);
+
+        __m256i length_8x = _mm256_alignr_epi32(offsetarray_1_8x, offsetarray_8x, 0x1);
+        length_8x = _mm256_sub_epi32(length_8x, offsetarray_8x);
+
+        __m256i reminder_8x = _mm256_and_si256(length_8x, x7_8x);
+        reminder_8x = _mm256_sub_epi32(x8_8x, reminder_8x);
+        reminder_8x = _mm256_and_si256(reminder_8x, x7_8x);
+        reminder_8x = _mm256_add_epi32(reminder_8x, length_8x);
+        __m256i dst_length_8x = _mm256_loadu_si256((__m256i*)length_data);
+        dst_length_8x = _mm256_add_epi32(dst_length_8x, reminder_8x);
+        _mm256_storeu_si256((__m256i*)length_data, dst_length_8x);
+        length_data += 8;
+        _mm_prefetch(&offsetarray[j + (128 + 128) / sizeof(offset_type)], _MM_HINT_T0);
+      }
 
       for (j; j < num_rows_; j++) {
         offset_type length = offsetarray[j + 1] - offsetarray[j];