TuplePackTest

more logs

add test

16 files changed, 375 insertions, 11 deletions

diff --git a/.mapping.json b/.mapping.json
index 808cc52993..027f520e1f 100644
--- a/.mapping.json
+++ b/.mapping.json
@@ -8790,6 +8790,11 @@
   "ydb/library/yql/utils/simd/CMakeLists.txt":"",
   "ydb/library/yql/utils/simd/CMakeLists.windows-x86_64.txt":"",
   "ydb/library/yql/utils/simd/exec/CMakeLists.txt":"",
+  "ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.darwin-x86_64.txt":"",
+  "ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-aarch64.txt":"",
+  "ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-x86_64.txt":"",
+  "ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.txt":"",
+  "ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.windows-x86_64.txt":"",
   "ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.darwin-x86_64.txt":"",
   "ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-aarch64.txt":"",
   "ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-x86_64.txt":"",
diff --git a/ydb/library/yql/utils/simd/exec/CMakeLists.txt b/ydb/library/yql/utils/simd/exec/CMakeLists.txt
index 1949327f1e..a3b4da0bea 100644
--- a/ydb/library/yql/utils/simd/exec/CMakeLists.txt
+++ b/ydb/library/yql/utils/simd/exec/CMakeLists.txt
@@ -6,4 +6,5 @@
 # original buildsystem will not be accepted.
 
 
+add_subdirectory(pack_tuple)
 add_subdirectory(stream_store)
diff --git a/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.darwin-x86_64.txt b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.darwin-x86_64.txt
new file mode 100644
index 0000000000..ae045da3a7
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.darwin-x86_64.txt
@@ -0,0 +1,31 @@
+
+# This file was generated by the build system used internally in the Yandex monorepo.
+# Only simple modifications are allowed (adding source-files to targets, adding simple properties
+# like target_include_directories). These modifications will be ported to original
+# ya.make files by maintainers. Any complex modifications which can't be ported back to the
+# original buildsystem will not be accepted.
+
+
+
+add_executable(pack_tuple)
+target_compile_options(pack_tuple PRIVATE
+  -mavx2
+)
+target_link_libraries(pack_tuple PUBLIC
+  contrib-libs-cxxsupp
+  yutil
+  library-cpp-cpuid_check
+  yql-utils-simd
+)
+target_link_options(pack_tuple PRIVATE
+  -Wl,-platform_version,macos,11.0,11.0
+  -fPIC
+  -fPIC
+)
+target_sources(pack_tuple PRIVATE
+  ${CMAKE_SOURCE_DIR}/ydb/library/yql/utils/simd/exec/pack_tuple/main.cpp
+)
+target_allocator(pack_tuple
+  system_allocator
+)
+vcs_info(pack_tuple)
diff --git a/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-aarch64.txt b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-aarch64.txt
new file mode 100644
index 0000000000..2d36690667
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-aarch64.txt
@@ -0,0 +1,36 @@
+
+# This file was generated by the build system used internally in the Yandex monorepo.
+# Only simple modifications are allowed (adding source-files to targets, adding simple properties
+# like target_include_directories). These modifications will be ported to original
+# ya.make files by maintainers. Any complex modifications which can't be ported back to the
+# original buildsystem will not be accepted.
+
+
+
+add_executable(pack_tuple)
+target_compile_options(pack_tuple PRIVATE
+  -mavx2
+)
+target_link_libraries(pack_tuple PUBLIC
+  contrib-libs-linux-headers
+  contrib-libs-cxxsupp
+  yutil
+  yql-utils-simd
+)
+target_link_options(pack_tuple PRIVATE
+  -ldl
+  -lrt
+  -Wl,--no-as-needed
+  -fPIC
+  -fPIC
+  -lpthread
+  -lrt
+  -ldl
+)
+target_sources(pack_tuple PRIVATE
+  ${CMAKE_SOURCE_DIR}/ydb/library/yql/utils/simd/exec/pack_tuple/main.cpp
+)
+target_allocator(pack_tuple
+  cpp-malloc-jemalloc
+)
+vcs_info(pack_tuple)
diff --git a/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-x86_64.txt b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-x86_64.txt
new file mode 100644
index 0000000000..6299f9b65a
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.linux-x86_64.txt
@@ -0,0 +1,38 @@
+
+# This file was generated by the build system used internally in the Yandex monorepo.
+# Only simple modifications are allowed (adding source-files to targets, adding simple properties
+# like target_include_directories). These modifications will be ported to original
+# ya.make files by maintainers. Any complex modifications which can't be ported back to the
+# original buildsystem will not be accepted.
+
+
+
+add_executable(pack_tuple)
+target_compile_options(pack_tuple PRIVATE
+  -mavx2
+)
+target_link_libraries(pack_tuple PUBLIC
+  contrib-libs-linux-headers
+  contrib-libs-cxxsupp
+  yutil
+  library-cpp-cpuid_check
+  yql-utils-simd
+)
+target_link_options(pack_tuple PRIVATE
+  -ldl
+  -lrt
+  -Wl,--no-as-needed
+  -fPIC
+  -fPIC
+  -lpthread
+  -lrt
+  -ldl
+)
+target_sources(pack_tuple PRIVATE
+  ${CMAKE_SOURCE_DIR}/ydb/library/yql/utils/simd/exec/pack_tuple/main.cpp
+)
+target_allocator(pack_tuple
+  cpp-malloc-tcmalloc
+  libs-tcmalloc-no_percpu_cache
+)
+vcs_info(pack_tuple)
diff --git a/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.txt b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.txt
new file mode 100644
index 0000000000..f8b31df0c1
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.txt
@@ -0,0 +1,17 @@
+
+# This file was generated by the build system used internally in the Yandex monorepo.
+# Only simple modifications are allowed (adding source-files to targets, adding simple properties
+# like target_include_directories). These modifications will be ported to original
+# ya.make files by maintainers. Any complex modifications which can't be ported back to the
+# original buildsystem will not be accepted.
+
+
+if (CMAKE_SYSTEM_NAME STREQUAL "Linux" AND CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64" AND NOT HAVE_CUDA)
+  include(CMakeLists.linux-aarch64.txt)
+elseif (CMAKE_SYSTEM_NAME STREQUAL "Darwin" AND CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+  include(CMakeLists.darwin-x86_64.txt)
+elseif (WIN32 AND CMAKE_SYSTEM_PROCESSOR STREQUAL "AMD64" AND NOT HAVE_CUDA)
+  include(CMakeLists.windows-x86_64.txt)
+elseif (CMAKE_SYSTEM_NAME STREQUAL "Linux" AND CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" AND NOT HAVE_CUDA)
+  include(CMakeLists.linux-x86_64.txt)
+endif()
diff --git a/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.windows-x86_64.txt b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.windows-x86_64.txt
new file mode 100644
index 0000000000..37c13c73e8
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/pack_tuple/CMakeLists.windows-x86_64.txt
@@ -0,0 +1,26 @@
+
+# This file was generated by the build system used internally in the Yandex monorepo.
+# Only simple modifications are allowed (adding source-files to targets, adding simple properties
+# like target_include_directories). These modifications will be ported to original
+# ya.make files by maintainers. Any complex modifications which can't be ported back to the
+# original buildsystem will not be accepted.
+
+
+
+add_executable(pack_tuple)
+target_compile_options(pack_tuple PRIVATE
+  -mavx2
+)
+target_link_libraries(pack_tuple PUBLIC
+  contrib-libs-cxxsupp
+  yutil
+  library-cpp-cpuid_check
+  yql-utils-simd
+)
+target_sources(pack_tuple PRIVATE
+  ${CMAKE_SOURCE_DIR}/ydb/library/yql/utils/simd/exec/pack_tuple/main.cpp
+)
+target_allocator(pack_tuple
+  system_allocator
+)
+vcs_info(pack_tuple)
diff --git a/ydb/library/yql/utils/simd/exec/pack_tuple/main.cpp b/ydb/library/yql/utils/simd/exec/pack_tuple/main.cpp
new file mode 100644
index 0000000000..ed9a4e0444
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/pack_tuple/main.cpp
@@ -0,0 +1,175 @@
+#include <util/generic/ptr.h>
+#include <util/system/cpu_id.h>
+#include <util/system/types.h>
+
+#include <ydb/library/yql/utils/simd/simd.h>
+
+struct TPerfomancer {
+    TPerfomancer() = default;
+
+    struct TWrapWorker {
+        virtual int PackTuple(bool log) = 0;
+        virtual ~TWrapWorker() = default;
+    };
+
+    template<typename TTraits>
+    struct TWorker : TWrapWorker {
+        template<typename T>
+        using TSimd = typename TTraits::template TSimd8<T>;
+        TWorker() = default;
+
+        ui8* ShuffleMask(ui32 v[8]) {
+            ui8* det = new ui8[32];
+            for (size_t i = 0; i < 32; i += 1) {
+                det[i] = v[i / 4] + i % 4;
+            }
+            return det;
+        }
+
+        int PackTupleImpl(bool log = true) {
+            if (TTraits::Size != 32)
+                return 1;
+            const ui64 NTuples = 32 << 18;
+            const ui64 TupleSize =  sizeof(ui32) + sizeof(ui64);
+
+            ui32 *arrUi32 __attribute__((aligned(32))) = new ui32[NTuples];
+            ui64 *arrUi64 __attribute__((aligned(32))) = new ui64[NTuples];
+
+            for (ui32 i = 0; i < NTuples; i++) {
+                arrUi32[i] = 2 * i;
+            }
+
+            for (ui32 i = 0; i < NTuples; i++) {
+                arrUi64[i] = 2 * i + 1;
+            }
+
+            TSimd<ui8> readReg1, readReg2, readReg1Fwd;
+
+            TSimd<ui8> permReg11, permReg21;
+            TSimd<ui8> permReg12, permReg22;
+
+            TSimd<ui8> permIdx11(ShuffleMask((ui32[8]) {0, 0, 0, 0, 1, 0, 0, 0}));
+            TSimd<ui8> permIdx12(ShuffleMask((ui32[8]) {0, 0, 0, 0, 3, 0, 0, 2}));
+            TSimd<ui8> permIdx1f(ShuffleMask((ui32[8]) {7, 7, 7, 7, 7, 6, 5, 4}));
+
+            TSimd<ui8> permIdx21(ShuffleMask((ui32[8]) {0, 0, 3, 2, 0, 1, 0, 0}));
+            TSimd<ui8> permIdx22(ShuffleMask((ui32[8]) {0, 0, 7, 6, 0, 5, 4, 0}));
+
+            ui32 val1[8], val2[8]; // val3[8];
+
+            using TReg = typename TTraits::TRegister;
+            TSimd<ui8> blended1, blended2;
+
+            TReg *addr1 = (TReg*) arrUi32;
+            TReg *addr2 = (TReg*) arrUi64;
+
+            std::chrono::steady_clock::time_point begin01 =
+                std::chrono::steady_clock::now();
+
+            ui64 accum1 = 0;
+            ui64 accum2 = 0;
+            ui64 accum3 = 0;
+            ui64 accum4 = 0;
+
+            const int blendMask = 0b00110110;
+
+            ui32 hash1 = 0;
+            ui32 hash2 = 0;
+            ui32 hash3 = 0;
+            ui32 hash4 = 0;
+
+            for (ui32 i = 0; i < NTuples; i += 8) {
+                readReg1 = TSimd<ui8>((ui8*) addr1);
+                for (ui32 j = 0; j < 2; j++) {
+
+                    permReg11 = readReg1.Shuffle(permIdx11);
+                    readReg2 = TSimd<ui8>((ui8*) addr2);
+                    addr2++;
+                    permReg21 = readReg2.Shuffle(permIdx21);
+                    blended1 = permReg11.template Blend16<blendMask>(permReg21);
+                    blended1.Store((ui8*) val1);
+
+                    hash1 = TSimd<ui8>::CRC32u32(0, val1[0]);
+                    hash2 = TSimd<ui8>::CRC32u32(0, val1[3]);
+
+                    accum1 += hash1;
+                    accum2 += hash2;
+
+                    permReg12 = readReg1.Shuffle(permIdx12);
+                    permReg22 = readReg2.Shuffle(permIdx22);
+                    blended2 = permReg12.template Blend16<blendMask>(permReg12);
+                    blended2.Store((ui8*) val2);
+
+                    hash3 = TSimd<ui8>::CRC32u32(0, val2[0]);
+                    hash4 = TSimd<ui8>::CRC32u32(0, val2[3]);
+
+                    accum3 += hash3;
+                    accum4 += hash4;
+
+                    readReg1Fwd = readReg1.Shuffle(permIdx1f);
+                    readReg1Fwd.Store((ui8*) &readReg1.Value);
+
+                }
+                addr1++;
+            }
+
+            Cerr << "Loaded col1 ";
+            readReg1.template Log<ui32>(Cerr);
+            Cerr << "Loaded col2 ";
+            readReg2.template Log<ui32>(Cerr);;
+            Cerr << "Permuted col1 ";
+            permReg11.template Log<ui32>(Cerr);;
+            Cerr << "Permuted col2 ";
+            permReg21.template Log<ui32>(Cerr);
+            Cerr << "Blended ";
+            blended1.template Log<ui32>(Cerr);
+
+
+            std::chrono::steady_clock::time_point end01 =
+                std::chrono::steady_clock::now();
+
+            ui64 microseconds =
+                std::chrono::duration_cast<std::chrono::microseconds>(end01 - begin01).count();
+            if (log) {
+                Cerr << "Accum 1 2 hash: " << accum1 << " " << accum2 << " "  << accum3 << " " << accum4 << " "
+                << hash1 << " " << hash2 << " " << hash3 << " " << hash4 << Endl;
+                Cerr << "Time for stream load = " << microseconds << "[microseconds]"
+                    << Endl;
+                Cerr << "Data size =  " << ((NTuples * TupleSize) / (1024 * 1024))
+                    << " [MB]" << Endl;
+                Cerr << "Stream load/save/accum speed = "
+                    << (NTuples * TupleSize * 1000 * 1000) /
+                            (1024 * 1024 * (microseconds + 1))
+                    << " MB/sec" << Endl;
+                Cerr << Endl;
+            }
+            delete[] arrUi32;
+            delete[] arrUi64;
+
+            return 1;
+        }
+
+        int PackTuple(bool log = true) override {
+            return PackTupleImpl(log);
+        }
+
+        ~TWorker() = default;
+    };
+
+    template<typename TTraits>
+    THolder<TWrapWorker> Create() const {
+        return MakeHolder<TWorker<TTraits>>();
+    };
+};
+
+int main() {
+    if (!NX86::HaveAVX2())
+        return 0;
+
+    TPerfomancer tp;
+    auto worker = tp.Create<NSimd::TSimdAVX2Traits>();
+
+    bool fine = true;
+    fine &= worker->PackTuple(true);
+    return !fine;
+}
\ No newline at end of file
diff --git a/ydb/library/yql/utils/simd/exec/pack_tuple/ya.make b/ydb/library/yql/utils/simd/exec/pack_tuple/ya.make
new file mode 100644
index 0000000000..61c6742c2a
--- /dev/null
+++ b/ydb/library/yql/utils/simd/exec/pack_tuple/ya.make
@@ -0,0 +1,9 @@
+PROGRAM()
+
+SRCS(main.cpp)
+
+CFLAGS(-mavx2)
+
+PEERDIR(ydb/library/yql/utils/simd)
+
+END()
\ No newline at end of file
diff --git a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.darwin-x86_64.txt b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.darwin-x86_64.txt
index a99ac27d6f..e5df3ab627 100644
--- a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.darwin-x86_64.txt
+++ b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.darwin-x86_64.txt
@@ -12,6 +12,7 @@ target_link_libraries(stream_store PUBLIC
   contrib-libs-cxxsupp
   yutil
   library-cpp-cpuid_check
+  yql-utils-simd
 )
 target_link_options(stream_store PRIVATE
   -Wl,-platform_version,macos,11.0,11.0
diff --git a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-aarch64.txt b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-aarch64.txt
index 049d3baab7..d4dd5985c4 100644
--- a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-aarch64.txt
+++ b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-aarch64.txt
@@ -12,6 +12,7 @@ target_link_libraries(stream_store PUBLIC
   contrib-libs-linux-headers
   contrib-libs-cxxsupp
   yutil
+  yql-utils-simd
 )
 target_link_options(stream_store PRIVATE
   -ldl
diff --git a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-x86_64.txt b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-x86_64.txt
index cbf6c43e02..2dbcba4a9e 100644
--- a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-x86_64.txt
+++ b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.linux-x86_64.txt
@@ -13,6 +13,7 @@ target_link_libraries(stream_store PUBLIC
   contrib-libs-cxxsupp
   yutil
   library-cpp-cpuid_check
+  yql-utils-simd
 )
 target_link_options(stream_store PRIVATE
   -ldl
diff --git a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.windows-x86_64.txt b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.windows-x86_64.txt
index a00f4d06e7..c5757a9974 100644
--- a/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.windows-x86_64.txt
+++ b/ydb/library/yql/utils/simd/exec/stream_store/CMakeLists.windows-x86_64.txt
@@ -12,6 +12,7 @@ target_link_libraries(stream_store PUBLIC
   contrib-libs-cxxsupp
   yutil
   library-cpp-cpuid_check
+  yql-utils-simd
 )
 target_sources(stream_store PRIVATE
   ${CMAKE_SOURCE_DIR}/ydb/library/yql/utils/simd/exec/stream_store/main.cpp
diff --git a/ydb/library/yql/utils/simd/exec/stream_store/main.cpp b/ydb/library/yql/utils/simd/exec/stream_store/main.cpp
index 4a25a86852..b879bf6fd7 100644
--- a/ydb/library/yql/utils/simd/exec/stream_store/main.cpp
+++ b/ydb/library/yql/utils/simd/exec/stream_store/main.cpp
@@ -1,4 +1,3 @@
-
 #include <util/generic/ptr.h>
 #include <util/system/cpu_id.h>
 #include <util/system/types.h>
@@ -9,7 +8,7 @@ struct TPerfomancer {
     TPerfomancer() = default;
 
     struct TWrapWorker {
-        virtual int StreamLoad(bool log) = 0;
+        virtual int StoreStream(bool log) = 0;
         virtual ~TWrapWorker() = default;
     };
 
@@ -29,7 +28,7 @@ struct TPerfomancer {
             }
         }
 
-        int StreamLoad(bool log = true) override {
+        int StoreStream(bool log = true) override {
             const size_t batch = 32 / TTraits::Size;
             const size_t batch_size = TTraits::Size / 8;
             size_t log_batch_size = 0;
@@ -41,8 +40,15 @@ struct TPerfomancer {
                 log_batch_size = 2;
             }
 
-            size_t size = (32LL << 21);
-            i64* buf __attribute__((aligned(32))) = new i64[size];
+            const size_t size = (32LL << 21);
+            const size_t arrSize = size / 8;
+
+            i64* buf __attribute__((aligned(32))) = new i64[arrSize];
+
+            for (size_t i = 0; i < arrSize; i += 1) {
+                buf[i] = 0;
+            }
+
 
             i64 tmp[4];
             for (size_t i = 0; i < 4; i += 1) {
@@ -56,21 +62,18 @@ struct TPerfomancer {
             std::chrono::steady_clock::time_point begin01 =
                     std::chrono::steady_clock::now();
 
-            const size_t size_loop = size / 8;
-
-            for (size_t i = 0; i < size_loop; i += 4) {
+            for (size_t i = 0; i < arrSize; i += 4) {
                 for (size_t j = 0; j < batch; j += 1) {
                     tmpSimd[j].StoreStream((i8*)(buf + i + j * batch_size));
                 }
             }
 
-
             std::chrono::steady_clock::time_point end01 =
                 std::chrono::steady_clock::now();
 
             bool is_ok = true;
 
-            for (size_t i = 0; i < size_loop; i += 1) {
+            for (size_t i = 0; i < arrSize; i += 1) {
                 if (buf[i] != i % 4) {
                     is_ok = false;
                 }
@@ -94,6 +97,14 @@ struct TPerfomancer {
             return is_ok;
         }
 
+        ui8* ShuffleMask(ui32 v[8]) {
+            ui8* det = new ui8[32];
+            for (size_t i = 0; i < 32; i += 1) {
+                det[i] = v[i / 4] + i % 4;
+            }
+            return det;
+        }
+
         ~TWorker() = default;
     };
 
@@ -106,5 +117,8 @@ struct TPerfomancer {
 int main() {
     TPerfomancer tp;
     auto worker = NSimd::SelectSimdTraits(tp);
-    return !worker->StreamLoad(false);
+
+    bool fine = true;
+    fine &= worker->StoreStream(true);
+    return !fine;
 }
\ No newline at end of file
diff --git a/ydb/library/yql/utils/simd/exec/stream_store/ya.make b/ydb/library/yql/utils/simd/exec/stream_store/ya.make
index 631001cbdc..704a7b6c0a 100644
--- a/ydb/library/yql/utils/simd/exec/stream_store/ya.make
+++ b/ydb/library/yql/utils/simd/exec/stream_store/ya.make
@@ -2,4 +2,6 @@ PROGRAM()
 
 SRCS(main.cpp)
 
+PEERDIR(ydb/library/yql/utils/simd)
+
 END()
\ No newline at end of file
diff --git a/ydb/library/yql/utils/simd/exec/ya.make b/ydb/library/yql/utils/simd/exec/ya.make
index 3fcf4d03e0..5d18536908 100644
--- a/ydb/library/yql/utils/simd/exec/ya.make
+++ b/ydb/library/yql/utils/simd/exec/ya.make
@@ -4,8 +4,13 @@ RUN(
     stream_store
 )
 
+RUN(
+    pack_tuple
+)
+
 DEPENDS(
     ydb/library/yql/utils/simd/exec/stream_store
+    ydb/library/yql/utils/simd/exec/pack_tuple
 )
 
 PEERDIR(
@@ -15,5 +20,6 @@ PEERDIR(
 END()
 
 RECURSE(
+    pack_tuple
     stream_store
 )
\ No newline at end of file