Fix compilation errors

Co-authored-by: Wiebe van Breukelen <breukelen@astron.nl>

.pre-commit-config.yaml

@@ -7,5 +7,4 @@ repos:
     rev: v0.6.13
     hooks:
       - id: cmake-format
-      - id: cmake-lint
+      - id: cmake-lint
-        args: [--disabled-codes=C0301]

CMakeLists.txt

@@ -12,11 +12,6 @@ option(TRIGDX_BUILD_TESTS "Build tests" ON)
 option(TRIGDX_BUILD_BENCHMARKS "Build tests" ON)
 option(TRIGDX_BUILD_PYTHON "Build Python interface" ON)
 
-# Add compiler flags
-set(CMAKE_CXX_FLAGS
-    "${CMAKE_CXX_FLAGS} -Wall -Wnon-virtual-dtor -Wduplicated-branches -Wvla -Wpointer-arith -Wextra -Wno-unused-parameter"
-)
-
 list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
 configure_file(
   ${CMAKE_CURRENT_SOURCE_DIR}/cmake/trigdx_config.hpp.in

python/CMakeLists.txt

@@ -8,16 +8,5 @@ if(NOT pybind11_FOUND)
   FetchContent_MakeAvailable(pybind11)
 endif()
 
-# Needed to set ${Python_VERSION_MAJOR} and ${Python_VERSION_MINOR}
-find_package(Python REQUIRED)
-
 pybind11_add_module(pytrigdx bindings.cpp)
 target_link_libraries(pytrigdx PRIVATE trigdx)
-set_target_properties(pytrigdx PROPERTIES OUTPUT_NAME "trigdx")
-
-set(PYTHON_SITE_PACKAGES
-    "${CMAKE_INSTALL_LIBDIR}/python${Python_VERSION_MAJOR}.${Python_VERSION_MINOR}/site-packages/trigdx"
-)
-
-install(TARGETS pytrigdx DESTINATION ${PYTHON_SITE_PACKAGES})
-install(FILES __init__.py DESTINATION ${PYTHON_SITE_PACKAGES})

python/__init__.py

@@ -1,16 +0,0 @@
-from .trigdx import Reference, Lookup16K, Lookup32K, LookupAVX16K, LookupAVX32K
-
-try:
-    from .trigdx import MKL
-except ImportError:
-    pass
-
-try:
-    from .trigdx import GPU
-except ImportError:
-    pass
-
-try:
-    from .trigdx import LookupXSIMD16K, LookupXSIMD32K
-except ImportError:
-    pass

python/bindings.cpp

@@ -72,9 +72,7 @@ void bind_backend(py::module &m, const char *name) {
       .def("compute_sincosf", &compute_sincos<float>);
 }
 
-PYBIND11_MODULE(trigdx, m) {
+PYBIND11_MODULE(pytrigdx, m) {
-  m.doc() = "TrigDx python bindings";
-
   py::class_<Backend, std::shared_ptr<Backend>>(m, "Backend")
       .def("init", &Backend::init);
 
@@ -93,4 +91,4 @@ PYBIND11_MODULE(trigdx, m) {
   bind_backend<LookupXSIMDBackend<16384>>(m, "LookupXSIMD16K");
   bind_backend<LookupXSIMDBackend<32768>>(m, "LookupXSIMD32K");
 #endif
-}
+}

src/lookup_xsimd.cpp

@@ -20,8 +20,8 @@ template <std::size_t NR_SAMPLES> struct lookup_table {
       cos_values[i] = cosf(i * PI_FRAC);
     }
   }
-  std::array<float, NR_SAMPLES> sin_values;
   std::array<float, NR_SAMPLES> cos_values;
+  std::array<float, NR_SAMPLES> sin_values;
 };
 
 template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
@@ -33,6 +33,7 @@ template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
 
     constexpr uint_fast32_t VL = b_type::size;
     const uint_fast32_t VS = n - n % VL;
+    const uint_fast32_t Q_PI = NR_SAMPLES / 4U;
     const b_type scale = b_type::broadcast(lookup_table_.SCALE);
     const b_type pi_frac = b_type::broadcast(lookup_table_.PI_FRAC);
     const m_type mask = m_type::broadcast(lookup_table_.MASK);
@@ -41,7 +42,7 @@ template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
     const b_type term2 = b_type::broadcast(lookup_table_.TERM2); // 1/2!
     const b_type term3 = b_type::broadcast(lookup_table_.TERM3); // 1/3!
     const b_type term4 = b_type::broadcast(lookup_table_.TERM4); // 1/4!
-
+    const m_type quarter_pi = m_type::broadcast(Q_PI);
     uint_fast32_t i;
     for (i = 0; i < VS; i += VL) {
       const b_type vx = b_type::load(a + i, Tag());
@@ -59,7 +60,7 @@ template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
       const b_type dx4 = xsimd::mul(dx2, dx);
       const b_type t2 = xsimd::mul(dx2, term2);
       const b_type t3 = xsimd::mul(dx3, term3);
-      const b_type t4 = xsimd::mul(dx4, term4);
+      const b_type t4 = xsimd::mul(dx4, term3);
 
       const b_type cosdx = xsimd::add(xsimd::sub(term1, t2), t4);
 
@@ -97,6 +98,7 @@ template <std::size_t NR_SAMPLES> struct sinf_dispatcher {
 
     constexpr uint_fast32_t VL = b_type::size;
     const uint_fast32_t VS = n - n % VL;
+    const uint_fast32_t Q_PI = NR_SAMPLES / 4U;
     const b_type scale = b_type::broadcast(lookup_table_.SCALE);
     const b_type pi_frac = b_type::broadcast(lookup_table_.PI_FRAC);
     const m_type mask = m_type::broadcast(lookup_table_.MASK);
@@ -105,7 +107,7 @@ template <std::size_t NR_SAMPLES> struct sinf_dispatcher {
     const b_type term2 = b_type::broadcast(lookup_table_.TERM2); // 1/2!
     const b_type term3 = b_type::broadcast(lookup_table_.TERM3); // 1/3!
     const b_type term4 = b_type::broadcast(lookup_table_.TERM4); // 1/4!
-
+    const m_type quarter_pi = m_type::broadcast(Q_PI);
     uint_fast32_t i;
     for (i = 0; i < VS; i += VL) {
       const b_type vx = b_type::load(a + i, Tag());
@@ -118,7 +120,7 @@ template <std::size_t NR_SAMPLES> struct sinf_dispatcher {
       const b_type dx4 = xsimd::mul(dx2, dx);
       const b_type t2 = xsimd::mul(dx2, term2);
       const b_type t3 = xsimd::mul(dx3, term3);
-      const b_type t4 = xsimd::mul(dx4, term4);
+      const b_type t4 = xsimd::mul(dx4, term3);
 
       const b_type cosdx = xsimd::add(xsimd::sub(term1, t2), t4);
       const b_type sindx = xsimd::sub(dx, t3);
@@ -158,6 +160,7 @@ template <std::size_t NR_SAMPLES> struct sin_cosf_dispatcher {
 
     constexpr uint_fast32_t VL = b_type::size;
     const uint_fast32_t VS = n - n % VL;
+    const uint_fast32_t Q_PI = NR_SAMPLES / 4U;
     const b_type scale = b_type::broadcast(lookup_table_.SCALE);
     const m_type mask = m_type::broadcast(lookup_table_.MASK);
     const b_type pi_frac = b_type::broadcast(lookup_table_.PI_FRAC);
@@ -167,6 +170,7 @@ template <std::size_t NR_SAMPLES> struct sin_cosf_dispatcher {
     const b_type term3 = b_type::broadcast(lookup_table_.TERM3); // 1/3!
     const b_type term4 = b_type::broadcast(lookup_table_.TERM4); // 1/4!
 
+    const m_type quarter_pi = m_type::broadcast(Q_PI);
     uint_fast32_t i;
     for (i = 0; i < VS; i += VL) {
       const b_type vx = b_type::load(a + i, Tag());
@@ -179,7 +183,7 @@ template <std::size_t NR_SAMPLES> struct sin_cosf_dispatcher {
       const b_type dx4 = xsimd::mul(dx2, dx);
       const b_type t2 = xsimd::mul(dx2, term2);
       const b_type t3 = xsimd::mul(dx3, term3);
-      const b_type t4 = xsimd::mul(dx4, term4);
+      const b_type t4 = xsimd::mul(dx4, term3);
 
       idx = xsimd::bitwise_and(idx, mask);
       b_type sinv = b_type::gather(lookup_table_.sin_values.data(), idx);