From e106285994f5fc1b7aaa4871b835da2d507c1587 Mon Sep 17 00:00:00 2001
From: Semen Tarlinskii <starlinskii@ydb.tech>
Date: Mon, 15 Jan 2024 14:33:26 +0300
Subject: [PATCH] test commit

---
 .../yql/utils/simd/exec/add_columns/main.cpp  | 72 +++++++++++++++++++
 .../yql/utils/simd/exec/add_columns/ya.make   | 13 ++++
 ydb/library/yql/utils/simd/exec/ya.make       |  1 +
 ydb/library/yql/utils/simd/simd_avx2.h        |  4 ++
 4 files changed, 90 insertions(+)
 create mode 100644 ydb/library/yql/utils/simd/exec/add_columns/main.cpp
 create mode 100644 ydb/library/yql/utils/simd/exec/add_columns/ya.make

diff --git a/ydb/library/yql/utils/simd/exec/add_columns/main.cpp b/ydb/library/yql/utils/simd/exec/add_columns/main.cpp
new file mode 100644
index 000000000000..f4155d9cc5f2
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/add_columns/main.cpp
@@ -0,0 +1,72 @@
+#include <util/generic/ptr.h>
+#include <util/system/cpu_id.h>
+#include <util/system/types.h>
+#include <util/stream/output.h>
+#include <util/generic/string.h>
+#include <vector>
+#include <immintrin.h>
+#include <avxintrin.h>
+#include <chrono>
+#include <ydb/library/yql/utils/simd/simd.h>
+
+
+const size_t size = 64e5;
+
+template <typename T>
+inline double GetSum(std::vector<std::vector<T>>& columns, std::vector<T>& result) {
+    const size_t SIZE_OF_TYPE = 256 / (sizeof(T) * 8);
+    const size_t align_size = columns[0].size();
+
+    std::chrono::steady_clock::time_point start = std::chrono::steady_clock::now();
+
+    for (size_t i = 0; i < align_size; i += SIZE_OF_TYPE) {
+         NSimd::NAVX2::TSimd8<T> final_register(&columns[0][i]);
+
+        for (size_t j = 1; j < columns.size(); ++j) {
+            final_register.Add64(&columns[j][i]);
+        }
+
+        final_register.Store(&result[i]);
+    }
+
+    std::chrono::steady_clock::time_point finish = std::chrono::steady_clock::now();
+
+    return std::chrono::duration_cast<std::chrono::microseconds>(finish - start).count();
+
+}
+
+double StandartAdding(std::vector<std::vector<ui64>>& columns, std::vector<ui64>& result) {
+    std::chrono::steady_clock::time_point start = std::chrono::steady_clock::now();
+
+    for (size_t j = 0; j < columns[0].size(); ++j) {
+
+        for (size_t i = 0; i < columns[i].size(); ++i) {
+            result[j] += columns[i][j];
+        }
+
+    }
+    std::chrono::steady_clock::time_point finish = std::chrono::steady_clock::now();
+
+    return std::chrono::duration_cast<std::chrono::microseconds>(finish - start).count();
+}
+
+int main() {
+    std::vector<std::vector<ui64>> vec1(10, std::vector<ui64>(size, 1e12 + 3));
+
+    std::vector<ui64> result1(size, 0);
+    std::vector<ui64> result2(size, 0);
+
+    double ans1 = GetSum(vec1, result1);
+    double ans2 = StandartAdding(vec1, result2);
+
+    for (size_t i = 0; i < result2.size(); ++i) {
+        if (result2[i] != result1[i]) {
+            Cerr << "something went wrong...";
+            return 0;
+        }
+    }
+
+    Cerr << "The results are the same. Let's compare times:\n";
+    Cerr << "Time, using AVX2: " << ans1 << " ms\n";
+    Cerr << "Time, using standart adding: " << ans2 << "ms";
+}
\ No newline at end of file
diff --git a/ydb/library/yql/utils/simd/exec/add_columns/ya.make b/ydb/library/yql/utils/simd/exec/add_columns/ya.make
new file mode 100644
index 000000000000..14b019febcc7
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/add_columns/ya.make
@@ -0,0 +1,13 @@
+OWNER(g:yql)
+
+PROGRAM(add_columns)
+
+SRCS(main.cpp)
+
+SIZE(MEDIUM)
+
+CFLAGS(-mavx2)
+
+PEERDIR(ydb/library/yql/utils/simd)
+
+END()
diff --git a/ydb/library/yql/utils/simd/exec/ya.make b/ydb/library/yql/utils/simd/exec/ya.make
index c36eb9197722..e86fb715d641 100644
--- a/ydb/library/yql/utils/simd/exec/ya.make
+++ b/ydb/library/yql/utils/simd/exec/ya.make
@@ -28,4 +28,5 @@ RECURSE(
     pack_tuple
     tuples_to_bucket
     stream_store
+    add_columns
 )
\ No newline at end of file
diff --git a/ydb/library/yql/utils/simd/simd_avx2.h b/ydb/library/yql/utils/simd/simd_avx2.h
index 319e201cdadf..27cbeb4a50ca 100644
--- a/ydb/library/yql/utils/simd/simd_avx2.h
+++ b/ydb/library/yql/utils/simd/simd_avx2.h
@@ -77,6 +77,10 @@ struct TSimd8 {
         crc = _mm_crc32_u64(crc, *((ui64*) &this->Value + 3));
         return crc;
     }
+    
+    inline void Add64(const TSimd8<T>& another) {
+        Value = _mm256_add_epi64(Value, another.Value);
+    }
 
     inline int ToBitMask() const {
         return _mm256_movemask_epi8(this->Value);