chore: Improve the implementation of simd based packing (#577)

Signed-off-by: Roman Gershman <roman@dragonflydb.io>

src/core/compact_object.cc

@@ -674,7 +674,7 @@ void CompactObj::SetString(std::string_view str) {
     }
 
     tl.tmp_buf.resize(encode_len);
-    detail::ascii_pack_simd(str.data(), str.size(), tl.tmp_buf.data());
+    detail::ascii_pack_simd2(str.data(), str.size(), tl.tmp_buf.data());
     encoded = string_view{reinterpret_cast<char*>(tl.tmp_buf.data()), encode_len};
 
     if (encoded.size() <= kInlineLen) {

src/core/compact_object_test.cc

@@ -204,7 +204,7 @@ TEST_F(CompactObjectTest, AsciiUtil) {
   }
   string act_str(data3.size(), 'y');
   std::vector<uint8_t> binvec(detail::binpacked_len(data3.size()));
-  detail::ascii_pack_simd(data3.data(), data3.size(), binvec.data());
+  detail::ascii_pack_simd2(data3.data(), data3.size(), binvec.data());
   detail::ascii_unpack_simd(binvec.data(), data3.size(), act_str.data());
 
   ASSERT_EQ(data3, act_str);
@@ -546,6 +546,16 @@ static void BM_PackSimd(benchmark::State& state) {
 }
 BENCHMARK(BM_PackSimd);
 
+static void BM_PackSimd2(benchmark::State& state) {
+  string val(1024, 'a');
+  uint8_t buf[1024];
+
+  while (state.KeepRunning()) {
+    detail::ascii_pack_simd2(val.data(), val.size(), buf);
+  }
+}
+BENCHMARK(BM_PackSimd2);
+
 static void BM_UnpackNaive(benchmark::State& state) {
   string val(1024, 'a');
   uint8_t buf[1024];

src/core/detail/bitpacking.cc

@@ -141,6 +141,44 @@ void ascii_pack_simd(const char* ascii, size_t len, uint8_t* bin) {
   ascii_pack(ascii, end - ascii, bin);
 }
 
+void ascii_pack_simd2(const char* ascii, size_t len, uint8_t* bin) {
+  // I leave out 16 bytes in addition to 16 that we load in the loop
+  // because we store into bin full 16 bytes instead of 14. To prevent data
+  // overwrite we finish loop one iteration earlier.
+  const char* end = ascii + len - 32;
+
+  // Skips 8th byte (indexc 7) in the lower 8-byte part.
+  const __m128i control = _mm_set_epi8(-1, -1, 14, 13, 12, 11, 10, 9, 8, 6, 5, 4, 3, 2, 1, 0);
+
+  __m128i val, rpart, lpart;
+
+  // Based on the question I asked here: https://stackoverflow.com/q/74831843/2280111
+  while (ascii <= end) {
+    val = _mm_loadu_si128(reinterpret_cast<const __m128i*>(ascii));
+
+    /*
+    x = ((x & 0x7F007F007F007F00) >> 1) | (x & 0x007F007F007F007F);
+    x = ((x & 0x3FFF00003FFF0000) >> 2) | (x & 0x00003FFF00003FFF);
+    x = ((x & 0x0FFFFFFF00000000) >> 4) | (x & 0x000000000FFFFFFF);
+    */
+    val = _mm_maddubs_epi16(_mm_set1_epi16(0x8001), val);
+    val = _mm_madd_epi16(_mm_set1_epi32(0x40000001), val);
+
+    rpart = _mm_and_si128(val, _mm_set1_epi64x(0x000000000FFFFFFF));
+    lpart = _mm_and_si128(val, _mm_set1_epi64x(0x0FFFFFFF00000000));
+    val = _mm_or_si128(_mm_srli_epi64(lpart, 4), rpart);
+
+    val = _mm_shuffle_epi8(val, control);
+    _mm_storeu_si128(reinterpret_cast<__m128i*>(bin), val);
+    bin += 14;
+    ascii += 16;
+  }
+
+  end += 32;  // Bring back end.
+  DCHECK(ascii < end);
+  ascii_pack(ascii, end - ascii, bin);
+}
+
 // unpacks 8->7 encoded blob back to ascii.
 // generally, we can not unpack inplace because ascii (dest) buffer is 8/7 bigger than
 // the source buffer.

src/core/detail/bitpacking.h

@@ -26,6 +26,8 @@ void ascii_pack(const char* ascii, size_t len, uint8_t* bin);
 void ascii_pack2(const char* ascii, size_t len, uint8_t* bin);
 
 void ascii_pack_simd(const char* ascii, size_t len, uint8_t* bin);
+void ascii_pack_simd2(const char* ascii, size_t len, uint8_t* bin);
+
 bool compare_packed(const uint8_t* packed, const char* ascii, size_t ascii_len);
 
 // maps ascii len to 7-bit packed length. Each 8 bytes are converted to 7 bytes.