Generate optimized overload of non-vectorized bitmap generator

This function can be selected on systems supporting BMI2, which introduces a shift instruction without modifying flags. This simplifes the uop pipeline and improves performance.

Author: alexbaden

omniscidb/ResultSet/BitmapGenerators.cpp

@@ -10,10 +10,15 @@
 
 #include <cstring>
 
-#include <iostream>
-
 #ifndef _WIN32
-size_t __attribute__((target("avx512f", "avx512bw"), optimize("no-tree-vectorize")))
+
+#if defined(__clang__)
+#define AVX512_TARGET target("avx512bw")
+#else
+#define AVX512_TARGET target("avx512f", "avx512bw")
+#endif
+
+size_t __attribute__((AVX512_TARGET, optimize("no-tree-vectorize")))
 gen_null_bitmap_8_impl(uint8_t* dst,
                        const uint8_t* src,
                        size_t size,
@@ -40,7 +45,7 @@ gen_null_bitmap_8_impl(uint8_t* dst,
   return null_count;
 }
 
-size_t __attribute__((target("avx512f", "avx512bw"), optimize("no-tree-vectorize")))
+size_t __attribute__((AVX512_TARGET, optimize("no-tree-vectorize")))
 gen_null_bitmap_16_impl(uint8_t* dst,
                         const uint16_t* src,
                         size_t size,
@@ -67,7 +72,7 @@ gen_null_bitmap_16_impl(uint8_t* dst,
   return null_count;
 }
 
-size_t __attribute__((target("avx512f", "avx512bw"), optimize("no-tree-vectorize")))
+size_t __attribute__((AVX512_TARGET, optimize("no-tree-vectorize")))
 gen_null_bitmap_32_impl(uint8_t* dst,
                         const uint32_t* src,
                         size_t size,
@@ -93,7 +98,7 @@ gen_null_bitmap_32_impl(uint8_t* dst,
   return null_count;
 }
 
-size_t __attribute__((target("avx512f", "avx512bw"), optimize("no-tree-vectorize")))
+size_t __attribute__((AVX512_TARGET, optimize("no-tree-vectorize")))
 gen_null_bitmap_64_impl(uint8_t* dst,
                         const uint64_t* src,
                         size_t size,
@@ -120,11 +125,15 @@ gen_null_bitmap_64_impl(uint8_t* dst,
 }
 #endif
 
+#if defined(_MSC_VER) || defined(__clang__)
+#define BITMAP_GEN_DEFAULT_TARGET_ATTRS
+#else
+#define BITMAP_GEN_DEFAULT_TARGET_ATTRS __attribute__((target_clones("bmi2", "default")))
+#endif
+
 template <typename TYPE>
-size_t gen_null_bitmap_default(uint8_t* dst,
-                               const TYPE* src,
-                               size_t size,
-                               const TYPE null_val) {
+size_t BITMAP_GEN_DEFAULT_TARGET_ATTRS
+gen_null_bitmap_default(uint8_t* dst, const TYPE* src, size_t size, const TYPE null_val) {
   size_t null_count = 0;
   TYPE loaded_data[8];
 
@@ -156,7 +165,6 @@ size_t BITMAP_GEN_TARGET_ATTRS gen_null_bitmap_8_impl(uint8_t* dst,
                                                       const uint8_t* src,
                                                       size_t size,
                                                       const uint8_t null_val) {
-  // std::cout << "Using default impl " << std::endl;
   return gen_null_bitmap_default<uint8_t>(dst, src, size, null_val);
 }
 

omniscidb/Tests/BitmapGeneratorBenchmark.cpp

@@ -45,8 +45,6 @@ static std::unique_ptr<T[]> gen_data(size_t size, const T null_val, float percen
   return data_buf;
 }
 
-#include <iostream>
-
 static auto alloc_bitmap(size_t input_sz) {
   // allocate extra space to ensure we can align
   size_t crt_sz = input_sz;
@@ -61,7 +59,6 @@ static auto alloc_bitmap(size_t input_sz) {
   return bitmap_buf_owned;
 }
 
-// TODO: get null count and verify
 static void null_bitmap_8(benchmark::State& state) {
   // generate a 4096 element buffer with 25% nulls (+ a few more since the null sentinel
   // will match the max size in the buffer)

omniscidb/Tests/CMakeLists.txt

@@ -58,8 +58,11 @@ endif()
 
 # Tests + Microbenchmarks
 add_executable(StringDictionaryBenchmark StringDictionaryBenchmark.cpp)
-add_executable(BitmapGeneratorBenchmark BitmapGeneratorBenchmark.cpp)
-target_link_libraries(BitmapGeneratorBenchmark ResultSet)
+if (NOT WIN32)
+  # aligned_alloc is missing on windows 
+  add_executable(BitmapGeneratorBenchmark BitmapGeneratorBenchmark.cpp)
+  target_link_libraries(BitmapGeneratorBenchmark ResultSet benchmark)
+endif()
 
 if(ENABLE_L0)
   add_executable(L0MgrExecuteTest L0MgrExecuteTest.cpp)
@@ -136,8 +139,6 @@ else()
   target_link_libraries(StringDictionaryBenchmark benchmark gtest StringDictionary Logger Utils $<$<AND:$<CXX_COMPILER_ID:GNU>,$<VERSION_LESS:$<CXX_COMPILER_VERSION>,9.0>>:stdc++fs> ${CMAKE_DL_LIBS} ${Boost_LIBRARIES} ${ZLIB_LIBRARIES})
 endif()
 
-target_link_libraries(BitmapGeneratorBenchmark benchmark)
-
 if(ENABLE_CUDA)
   target_link_libraries(GpuSharedMemoryTest gtest Logger QueryEngine)
 endif()