[NSE-927] Add macro __AVX512BW__ check for different CPU architecture (…

…#975) * Add __AVX512BW__ check * Fix cFormat
oap-project · Jun 20, 2022 · 261345c · 261345c
1 parent e2cfb29
commit 261345c
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 5 deletions.
diff --git a/native-sql-engine/cpp/src/CMakeLists.txt b/native-sql-engine/cpp/src/CMakeLists.txt
@@ -495,7 +495,7 @@ file(COPY codegen/common/hash_relation_number.h DESTINATION ${root_directory}/re
 
 add_definitions(-DNATIVESQL_SRC_PATH="${root_directory}/releases")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-deprecated-declarations -Wno-attributes")
-set(NATIVE_AVX512_FLAG "-march=icelake-server")
+set(NATIVE_AVX512_FLAG "-march=native")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${NATIVE_AVX512_FLAG}")
 set(SPARK_COLUMNAR_PLUGIN_SRCS
         jni/jni_wrapper.cc

diff --git a/native-sql-engine/cpp/src/operators/columnar_to_row_converter.cc b/native-sql-engine/cpp/src/operators/columnar_to_row_converter.cc
@@ -144,6 +144,7 @@ arrow::Status ColumnarToRowConverter::Init(
       int32_t j = 0;
       int32_t* length_data = lengths_.data();
 
+#ifdef __AVX512BW__
       if (ARROW_PREDICT_TRUE(support_avx512_)) {
         __m256i x7_8x = _mm256_load_si256((__m256i*)x_7);
         __m256i x8_8x = _mm256_load_si256((__m256i*)x_8);
@@ -172,6 +173,7 @@ arrow::Status ColumnarToRowConverter::Init(
           _mm_prefetch(&offsetarray[j + (128 + 128) / sizeof(offset_type)], _MM_HINT_T0);
         }
       }
+#endif
 
       for (j; j < num_rows_; j++) {
         offset_type length = offsetarray[j + 1] - offsetarray[j];
@@ -192,10 +194,13 @@ arrow::Status ColumnarToRowConverter::Init(
   // allocate one more cache line to ease avx operations
   if (buffer_ == nullptr || buffer_->capacity() < total_memory_size + 64) {
     ARROW_ASSIGN_OR_RAISE(buffer_, AllocateBuffer(total_memory_size + 64, memory_pool_));
+#ifdef __AVX512BW__
     if (ARROW_PREDICT_TRUE(support_avx512_)) {
       memset(buffer_->mutable_data() + total_memory_size, 0,
              buffer_->capacity() - total_memory_size);
-    } else {
+    } else
+#endif
+    {
       memset(buffer_->mutable_data(), 0, buffer_->capacity());
     }
   }
@@ -384,6 +389,7 @@ inline arrow::Status FillBuffer(int32_t& row_start, int32_t batch_rows,
                                 std::vector<uint8_t>& typewidth,
                                 std::vector<std::shared_ptr<arrow::Array>>& arrays,
                                 bool support_avx512) {
+#ifdef __AVX512BW__
   if (ARROW_PREDICT_TRUE(support_avx512)) {
     __m256i fill_0_8x;
     fill_0_8x = _mm256_xor_si256(fill_0_8x, fill_0_8x);
@@ -395,6 +401,7 @@ inline arrow::Status FillBuffer(int32_t& row_start, int32_t batch_rows,
       }
     }
   }
+#endif
 
   for (auto col_index = 0; col_index < num_cols; col_index++) {
     auto& array = arrays[col_index];
@@ -427,6 +434,7 @@ inline arrow::Status FillBuffer(int32_t& row_start, int32_t batch_rows,
             offset_type length = BinaryOffsets[j + 1] - BinaryOffsets[j];
             auto value = &dataptrs[col_index][2][BinaryOffsets[j]];
 
+#ifdef __AVX512BW__
             if (ARROW_PREDICT_TRUE(support_avx512)) {
               // write the variable value
               offset_type k;
@@ -440,7 +448,9 @@ inline arrow::Status FillBuffer(int32_t& row_start, int32_t batch_rows,
               __m256i v = _mm256_maskz_loadu_epi8(mask, value + k);
               _mm256_mask_storeu_epi8(buffer_address + offsets[j] + buffer_cursor[j] + k,
                                       mask, v);
-            } else {
+            } else
+#endif
+            {
               // write the variable value
               memcpy(buffer_address + offsets[j] + buffer_cursor[j], value, length);
             }
@@ -508,11 +518,14 @@ inline arrow::Status FillBuffer(int32_t& row_start, int32_t batch_rows,
           for (auto j = row_start; j < row_start + batch_rows; j++) {
             if (nullvec[col_index] || (!array->IsNull(j))) {
               const uint8_t* srcptr = dataptr + (j << shift);
+#ifdef __AVX512BW__
               if (ARROW_PREDICT_TRUE(support_avx512)) {
                 __m256i v = _mm256_maskz_loadu_epi8(mask, srcptr);
                 _mm256_mask_storeu_epi8(buffer_address_tmp + offsets[j], mask, v);
                 _mm_prefetch(srcptr + 64, _MM_HINT_T0);
-              } else {
+              } else
+#endif
+              {
                 memcpy(buffer_address_tmp + offsets[j], srcptr, typewidth[col_index]);
               }
             } else {

diff --git a/native-sql-engine/cpp/src/shuffle/splitter.cc b/native-sql-engine/cpp/src/shuffle/splitter.cc
@@ -1250,6 +1250,7 @@ arrow::Status Splitter::SplitBinaryType(const uint8_t* src_addr, const T* src_of
                   << " strlen = " << strlength << std::endl;
       }
       auto value_src_ptr = src_addr + src_offset_addr[src_offset];
+#ifdef __AVX512BW__
       if (ARROW_PREDICT_TRUE(support_avx512_)) {
         // write the variable value
         T k;
@@ -1260,7 +1261,9 @@ arrow::Status Splitter::SplitBinaryType(const uint8_t* src_addr, const T* src_of
         auto mask = (1L << (strlength - k)) - 1;
         __m256i v = _mm256_maskz_loadu_epi8(mask, value_src_ptr + k);
         _mm256_mask_storeu_epi8(dst_value_base + k, mask, v);
-      } else {
+      } else
+#endif
+      {
         memcpy(dst_value_base, value_src_ptr, strlength);
       }
       dst_value_base += strlength;