32: Set AVX and AVX2 flags using CMake checks (#34)

* 32: Set mavx and mavx2 based on CMake checks
* 32: Update flags for Intel compiler
* Fix: AVX2 instead of AVX__2

Co-authored-by: Bram Veenboer <bram.veenboer@gmail.com>
Co-authored-by: lukken <lukken@astron.nl>

CMakeLists.txt

@@ -12,6 +12,11 @@ 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

README.md

@@ -0,0 +1,54 @@
+# TrigDx
+
+High‑performance C++ library offering multiple implementations of transcendental trigonometric functions (e.g., sin, cos, tan and their variants), designed for numerical, signal‑processing, and real‑time systems where trading a small loss of accuracy for significantly higher throughput on modern CPUs (scalar and SIMD) and NVIDIA GPUs is acceptable.
+
+## Why TrigDx?
+Many applications use the standard library implementations, which prioritise correctness but are not always optimal for throughput on vectorized or GPU hardware. TrigDx gives you multiple implementations so you can:
+
+- Replace `std::sin` / `std::cos` calls with faster approximations when a small, bounded reduction in accuracy is acceptable.
+- Use SIMD/vectorized implementations and compact lookup tables for high throughput lookups.
+- Run massively parallel kernels that take advantage of a GPU's _Special Function Units_ (SFUs).
+
+## Requirements
+- A C++ compiler with at least C++17 support (GCC, Clang)
+- CMake 3.15+
+- Optional: NVIDIA CUDA Toolkit 11+ to build GPU kernels
+- Optional: GoogleTest (for unit tests) and GoogleBenchmark (for microbenchmarks)
+
+## Building
+```bash
+git clone https://github.com/astron-rd/TrigDx.git
+cd TrigDx
+mkdir build && cd build
+
+# CPU-only:
+cmake -DCMAKE_BUILD_TYPE=Release -DTRIGDX_USE_XSIMD=ON ..
+cmake --build . -j
+
+# Enable CUDA (if available):
+cmake -DCMAKE_BUILD_TYPE=Release -DTRIGDX_USE_GPU=ON ..
+cmake --build . -j
+
+# Run tests:
+ctest --output-on-failure -j
+```
+
+Common CMake options:
+- `TRIGDX_USE_GPU=ON/OFF` — build GPU support.
+- `TRIGDX_BUILD_TESTS=ON/OFF` — build tests.
+- `TRIGDX_BUILD_BENCHMARKS=ON/OFF` — build benchmarks.
+- `TRIGDX_BUILD_PYTHON` — build Python interface.
+
+## Contributing
+- Fork → create a feature branch → open a PR.
+- Include unit tests for correctness‑sensitive changes and benchmark results for performance changes.
+- Follow project style (clang‑format) and run tests locally before submitting.
+
+## Reporting issues
+When opening an issue for incorrect results or performance regressions, please include:
+- Platform and CPU/GPU model.
+- Compiler and version with exact compile flags.
+- Small reproducer (input data and the TrigDx implementation used).
+  
+## License
+See the LICENSE file in the repository for licensing details.

src/CMakeLists.txt

@@ -2,6 +2,24 @@ include(FetchContent)
 include(FindAVX)
 add_library(trigdx reference.cpp lookup.cpp)
 
+if(HAVE_AVX2)
+  target_compile_definitions(trigdx PUBLIC HAVE_AVX2)
+  if(CMAKE_CXX_COMPILER_ID STREQUAL "Intel" OR CMAKE_CXX_COMPILER_ID STREQUAL
+                                               "IntelLLVM")
+    target_compile_options(trigdx PUBLIC -xCORE-AVX2)
+  else()
+    target_compile_options(trigdx PUBLIC -mavx2)
+  endif()
+elseif(HAVE_AVX)
+  target_compile_definitions(trigdx PUBLIC HAVE_AVX)
+  if(CMAKE_CXX_COMPILER_ID STREQUAL "Intel" OR CMAKE_CXX_COMPILER_ID STREQUAL
+                                               "IntelLLVM")
+    target_compile_options(trigdx PUBLIC -xAVX)
+  else()
+    target_compile_options(trigdx PUBLIC -mavx)
+  endif()
+endif()
+
 target_include_directories(trigdx PUBLIC ${PROJECT_SOURCE_DIR}/include)
 
 if(HAVE_AVX)

src/lookup_avx.cpp

@@ -6,6 +6,16 @@
 
 #include "trigdx/lookup_avx.hpp"
 
+#if defined(HAVE_AVX) && !defined(__AVX__)
+static_assert(HAVE_AVX == 0, "__AVX__ should be defined when HAVE_AVX is "
+                             "defined");
+#endif
+
+#if defined(HAVE_AVX2) && !defined(__AVX2__)
+static_assert(HAVE_AVX2 == 0, "__AVX2__ should be defined when HAVE_AVX2 is "
+                              "defined");
+#endif
+
 template <std::size_t NR_SAMPLES> struct LookupAVXBackend<NR_SAMPLES>::Impl {
   std::vector<float> lookup;
   static constexpr std::size_t MASK = NR_SAMPLES - 1;

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> cos_values;
   std::array<float, NR_SAMPLES> sin_values;
+  std::array<float, NR_SAMPLES> cos_values;
 };
 
 template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
@@ -33,7 +33,6 @@ 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);
@@ -42,7 +41,7 @@ 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());
@@ -60,7 +59,7 @@ 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, term3);
+      const b_type t4 = xsimd::mul(dx4, term4);
 
       const b_type cosdx = xsimd::add(xsimd::sub(term1, t2), t4);
 
@@ -98,7 +97,6 @@ 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);
@@ -107,7 +105,7 @@ 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());
@@ -160,7 +158,6 @@ 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);
@@ -170,7 +167,6 @@ 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());