Use __cplusplus version check, and minor improve on the bitmask. (#239)

* use __cplusplus version checking instead of feature macros * use bit operations for count_true * add lib implementation test * clang-format * minor --------- Co-authored-by: Samuel Li <shaomeng@cisl-vapor> Co-authored-by: Samuel Li <Sam@Navada>
NCAR · Apr 30, 2024 · fe9df24 · fe9df24
1 parent ddb02cd
commit fe9df24
Show file tree

Hide file tree

Showing 8 changed files with 37 additions and 29 deletions.
diff --git a/include/Bitmask.h b/include/Bitmask.h
@@ -63,7 +63,7 @@ class Bitmask {
   auto view_buffer() const -> const std::vector<uint64_t>&;
   void use_bitstream(const void* p);
 
-#if defined __cpp_lib_three_way_comparison && defined __cpp_impl_three_way_comparison
+#if __cplusplus >= 201907L && defined __cpp_lib_three_way_comparison
   auto operator<=>(const Bitmask& rhs) const noexcept;
   auto operator==(const Bitmask& rhs) const noexcept -> bool;
 #endif

diff --git a/src/Bitmask.cpp b/src/Bitmask.cpp
@@ -4,6 +4,10 @@
 #include <cassert>
 #include <limits>
 
+#if __cplusplus >= 201907L
+#include <bit>
+#endif
+
 sperr::Bitmask::Bitmask(size_t nbits)
 {
   if (nbits > 0) {
@@ -48,11 +52,13 @@ auto sperr::Bitmask::has_true(size_t start, size_t len) const -> int64_t
 {
   auto long_idx = start / 64;
   auto processed_bits = int64_t{0};
-
-  // Collect the remaining bits from the start long.
   auto word = m_buf[long_idx];
   auto answer = uint64_t{0};
-  for (auto i = start - long_idx * 64; i < 64 && processed_bits < len; i++) {
+
+  // Collect the remaining bits from the start long.
+  auto begin_idx = start - long_idx * 64;
+  auto nbits = std::min(size_t{64}, begin_idx + len);
+  for (auto i = begin_idx; i < nbits; i++) {
     answer |= word & (uint64_t{1} << i);
     if constexpr (Position) {
       if (answer != 0)
@@ -70,9 +76,14 @@ auto sperr::Bitmask::has_true(size_t start, size_t len) const -> int64_t
     word = m_buf[++long_idx];
     if (word) {
       if constexpr (Position) {
+#if __cplusplus >= 201907L
+        int64_t i = std::countr_zero(word);
+        return processed_bits + i;
+#else
         for (int64_t i = 0; i < 64; i++)
           if (word & (uint64_t{1} << i))
             return processed_bits + i;
+#endif
       }
       else
         return 1;
@@ -82,7 +93,7 @@ auto sperr::Bitmask::has_true(size_t start, size_t len) const -> int64_t
 
   // Examine the remaining bits
   if (processed_bits < len) {
-    auto nbits = len - processed_bits;
+    nbits = len - processed_bits;
     assert(nbits < 64);
     word = m_buf[++long_idx];
     answer = 0;
@@ -112,11 +123,15 @@ auto sperr::Bitmask::count_true() const -> size_t
 
   // Note that unused bits in the last long are not guaranteed to be all 0's.
   for (size_t i = 0; i < m_buf.size() - 1; i++) {
-    const auto val = m_buf[i];
+    auto val = m_buf[i];
+#if __cplusplus >= 201907L
+    counter += std::popcount(val);
+#else
     if (val != 0) {
       for (size_t j = 0; j < 64; j++)
         counter += ((val >> j) & uint64_t{1});
     }
+#endif
   }
   const auto val = m_buf.back();
   if (val != 0) {
@@ -186,7 +201,7 @@ void sperr::Bitmask::use_bitstream(const void* p)
   std::copy(pu64, pu64 + m_buf.size(), m_buf.begin());
 }
 
-#if defined __cpp_lib_three_way_comparison && defined __cpp_impl_three_way_comparison
+#if __cplusplus >= 201907L && defined __cpp_lib_three_way_comparison
 auto sperr::Bitmask::operator<=>(const Bitmask& rhs) const noexcept
 {
   auto cmp = m_num_bits <=> rhs.m_num_bits;

diff --git a/src/SPECK1D_INT_DEC.cpp b/src/SPECK1D_INT_DEC.cpp
@@ -5,7 +5,7 @@
 #include <cstring>  // std::memcpy()
 #include <numeric>
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
 #include <bit>
 #endif
 
@@ -19,11 +19,10 @@ void sperr::SPECK1D_INT_DEC<T>::m_sorting_pass()
   for (size_t i = 0; i < bits_x64; i += 64) {
     auto value = m_LIP_mask.rlong(i);
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
     while (value) {
       size_t j = std::countr_zero(value);
       m_process_P(i + j, j, true);
-
       value &= value - 1;
     }
 #else

diff --git a/src/SPECK1D_INT_ENC.cpp b/src/SPECK1D_INT_ENC.cpp
@@ -5,7 +5,7 @@
 #include <cstring>  // std::memcpy()
 #include <numeric>
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
 #include <bit>
 #endif
 
@@ -19,11 +19,10 @@ void sperr::SPECK1D_INT_ENC<T>::m_sorting_pass()
   for (size_t i = 0; i < bits_x64; i += 64) {
     auto value = m_LIP_mask.rlong(i);
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
     while (value) {
       size_t j = std::countr_zero(value);
       m_process_P(i + j, SigType::Dunno, j, true);
-
       value &= value - 1;
     }
 #else

diff --git a/src/SPECK2D_INT.cpp b/src/SPECK2D_INT.cpp
@@ -3,7 +3,7 @@
 #include <algorithm>
 #include <cassert>
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
 #include <bit>
 #endif
 
@@ -17,11 +17,10 @@ void sperr::SPECK2D_INT<T>::m_sorting_pass()
   for (size_t i = 0; i < bits_x64; i += 64) {
     auto value = m_LIP_mask.rlong(i);
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
     while (value) {
       size_t j = std::countr_zero(value);
       m_process_P(i + j, j, true);
-
       value &= value - 1;
     }
 #else

diff --git a/src/SPECK3D_INT.cpp b/src/SPECK3D_INT.cpp
@@ -5,7 +5,7 @@
 #include <cstring>
 #include <numeric>
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
 #include <bit>
 #endif
 
@@ -106,11 +106,10 @@ void sperr::SPECK3D_INT<T>::m_sorting_pass()
   for (size_t i = 0; i < bits_x64; i += 64) {
     auto value = m_LIP_mask.rlong(i);
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
     while (value) {
       auto j = std::countr_zero(value);
       m_process_P_lite(i + j);
-
       value &= value - 1;
     }
 #else

diff --git a/src/SPECK_INT.cpp b/src/SPECK_INT.cpp
@@ -5,7 +5,7 @@
 #include <cstring>
 #include <numeric>
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
 #include <bit>
 #endif
 
@@ -317,13 +317,12 @@ void sperr::SPECK_INT<T>::m_refinement_pass_encode()
   for (size_t i = 0; i < bits_x64; i += 64) {  // Evaluate 64 bits at a time.
     auto value = m_LSP_mask.rlong(i);
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
     while (value) {
       auto j = std::countr_zero(value);
       const bool o1 = m_coeff_buf[i + j] >= m_threshold;
       m_coeff_buf[i + j] -= tmp1[o1];
       m_bit_buffer.wbit(o1);
-
       value &= value - 1;
     }
 #else
@@ -375,7 +374,7 @@ void sperr::SPECK_INT<T>::m_refinement_pass_decode()
     for (size_t i = 0; i < bits_x64; i += 64) {  // <-- Point 2
       auto value = m_LSP_mask.rlong(i);
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
       while (value) {
         auto j = std::countr_zero(value);
         if (m_bit_buffer.rbit())
@@ -384,7 +383,6 @@ void sperr::SPECK_INT<T>::m_refinement_pass_decode()
           m_coeff_buf[i + j] -= half_t;
         if (++read_pos == m_avail_bits)              // <-- Point 3
           goto INITIALIZE_NEWLY_FOUND_POINTS_LABEL;  // <-- Point 4
-
         value &= value - 1;
       }
 #else
@@ -417,14 +415,13 @@ void sperr::SPECK_INT<T>::m_refinement_pass_decode()
     for (size_t i = 0; i < bits_x64; i += 64) {
       auto value = m_LSP_mask.rlong(i);
 
-#if defined __cpp_lib_bitops
+#if __cplusplus >= 201907L
       while (value) {
         auto j = std::countr_zero(value);
         if (m_bit_buffer.rbit())
           ++(m_coeff_buf[i + j]);
         if (++read_pos == m_avail_bits)
           goto INITIALIZE_NEWLY_FOUND_POINTS_LABEL;
-
         value &= value - 1;
       }
 #else

diff --git a/test_scripts/bitstream_unit_test.cpp b/test_scripts/bitstream_unit_test.cpp
@@ -362,7 +362,7 @@ TEST(Bitmask, BufferTransfer)
 
 TEST(Bitmask, has_true)
 {
-  const size_t mask_size = 130;
+  const size_t mask_size = 210;
 
   // Loop over all positions
   for (size_t idx = 0; idx < mask_size; idx++) {
@@ -390,7 +390,7 @@ TEST(Bitmask, has_true)
 
 TEST(Bitmask, has_true_position)
 {
-  const size_t mask_size = 130;
+  const size_t mask_size = 210;
 
   // Loop over all positions
   for (size_t idx = 0; idx < mask_size; idx++) {
@@ -421,7 +421,7 @@ TEST(Bitmask, has_true_position)
   {}
 }
 
-#if defined __cpp_lib_three_way_comparison && defined __cpp_impl_three_way_comparison
+#if __cplusplus >= 201907L && defined __cpp_lib_three_way_comparison
 TEST(Bitmask, spaceship)
 {
   auto src = Mask(60);