Fix dx4 calculation in scalar remainder code and add null checks

- Fix dx4 computation in scalar remainder loops (should be dx2*dx2)
- Add missing null pointer check in benchmark_sinf for consistency

Co-authored-by: wvbbreu <185333235+wvbbreu@users.noreply.github.com>

.pre-commit-config.yaml

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

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.

benchmarks/benchmark_utils.hpp

@@ -2,13 +2,14 @@
 
 #include <chrono>
 #include <cmath>
+#include <stdexcept>
 #include <string>
 #include <vector>
 
 #include <benchmark/benchmark.h>
 
-void init_x(std::vector<float> &x) {
-  for (size_t i = 0; i < x.size(); ++i) {
+void init_x(float *x, size_t n) {
+  for (size_t i = 0; i < n; ++i) {
     x[i] = (i % 360) * 0.0174533f; // degrees to radians
   }
 }
@@ -16,24 +17,34 @@ void init_x(std::vector<float> &x) {
 template <typename Backend>
 static void benchmark_sinf(benchmark::State &state) {
   const size_t N = static_cast<size_t>(state.range(0));
-  std::vector<float> x(N), s(N);
-  init_x(x);
 
   Backend backend;
 
   auto start = std::chrono::high_resolution_clock::now();
   backend.init(N);
+  float *x =
+      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
+  float *s =
+      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
+  if (!x || !s) {
+    throw std::runtime_error("Buffer allocation failed");
+  }
   auto end = std::chrono::high_resolution_clock::now();
   state.counters["init_ms"] =
       std::chrono::duration_cast<std::chrono::microseconds>(end - start)
           .count() /
       1.e3;
 
+  init_x(x, N);
+
   for (auto _ : state) {
-    backend.compute_sinf(N, x.data(), s.data());
+    backend.compute_sinf(N, x, s);
     benchmark::DoNotOptimize(s);
   }
 
+  backend.free_memory(x);
+  backend.free_memory(s);
+
   state.SetItemsProcessed(static_cast<int64_t>(state.iterations()) *
                           static_cast<int64_t>(N));
 }
@@ -41,24 +52,35 @@ static void benchmark_sinf(benchmark::State &state) {
 template <typename Backend>
 static void benchmark_cosf(benchmark::State &state) {
   const size_t N = static_cast<size_t>(state.range(0));
-  std::vector<float> x(N), c(N);
-  init_x(x);
 
   Backend backend;
 
   auto start = std::chrono::high_resolution_clock::now();
   backend.init(N);
+  float *x =
+      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
+  float *c =
+      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
+
+  if (!x || !c) {
+    throw std::runtime_error("Buffer allocation failed");
+  }
   auto end = std::chrono::high_resolution_clock::now();
   state.counters["init_ms"] =
       std::chrono::duration_cast<std::chrono::microseconds>(end - start)
           .count() /
       1.e3;
 
+  init_x(x, N);
+
   for (auto _ : state) {
-    backend.compute_cosf(N, x.data(), c.data());
+    backend.compute_cosf(N, x, c);
     benchmark::DoNotOptimize(c);
   }
 
+  backend.free_memory(x);
+  backend.free_memory(c);
+
   state.SetItemsProcessed(static_cast<int64_t>(state.iterations()) *
                           static_cast<int64_t>(N));
 }
@@ -66,25 +88,38 @@ static void benchmark_cosf(benchmark::State &state) {
 template <typename Backend>
 static void benchmark_sincosf(benchmark::State &state) {
   const size_t N = static_cast<size_t>(state.range(0));
-  std::vector<float> x(N), s(N), c(N);
-  init_x(x);
 
   Backend backend;
 
   auto start = std::chrono::high_resolution_clock::now();
   backend.init(N);
+  float *x =
+      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
+  float *s =
+      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
+  float *c =
+      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
+  if (!x || !s || !c) {
+    throw std::runtime_error("Buffer allocation failed");
+  }
   auto end = std::chrono::high_resolution_clock::now();
   state.counters["init_ms"] =
       std::chrono::duration_cast<std::chrono::microseconds>(end - start)
           .count() /
       1.e3;
 
+  init_x(x, N);
+
   for (auto _ : state) {
-    backend.compute_sincosf(N, x.data(), s.data(), c.data());
+    backend.compute_sincosf(N, x, s, c);
     benchmark::DoNotOptimize(s);
     benchmark::DoNotOptimize(c);
   }
 
+  backend.free_memory(x);
+  backend.free_memory(s);
+  backend.free_memory(c);
+
   state.SetItemsProcessed(static_cast<int64_t>(state.iterations()) *
                           static_cast<int64_t>(N));
 }

include/trigdx/gpu.hpp

@@ -11,7 +11,8 @@ public:
   GPUBackend();
   ~GPUBackend() override;
 
-  void init(size_t n = 0) override;
+  void *allocate_memory(size_t bytes) const override;
+  void free_memory(void *ptr) const override;
   void compute_sinf(size_t n, const float *x, float *s) const override;
   void compute_cosf(size_t n, const float *x, float *c) const override;
   void compute_sincosf(size_t n, const float *x, float *s,

include/trigdx/interface.hpp

@@ -1,6 +1,8 @@
 #pragma once
 
 #include <cstddef>
+#include <cstdint>
+#include <cstdlib>
 
 // Base interface for all math backends
 class Backend {
@@ -10,6 +12,12 @@ public:
   // Optional initialization
   virtual void init(size_t n = 0) {}
 
+  virtual void *allocate_memory(size_t bytes) const {
+    return static_cast<void *>(new uint8_t[bytes]);
+  };
+
+  virtual void free_memory(void *ptr) const { delete[] static_cast<uint8_t*>(ptr); };
+
   // Compute sine for n elements
   virtual void compute_sinf(size_t n, const float *x, float *s) const = 0;
 

python/CMakeLists.txt

@@ -8,5 +8,16 @@ 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

@@ -0,0 +1,16 @@
+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

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

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/gpu.cpp

@@ -10,79 +10,63 @@
 
 struct GPUBackend::Impl {
 
-  ~Impl() {
-    if (h_x) {
-      cudaFreeHost(h_x);
-    }
-    if (h_s) {
-      cudaFreeHost(h_s);
-    }
-    if (h_c) {
-      cudaFreeHost(h_c);
-    }
-    if (d_x) {
-      cudaFree(d_x);
-    }
-    if (d_s) {
-      cudaFree(d_s);
-    }
-    if (d_c) {
-      cudaFree(d_c);
-    }
+  void *allocate_memory(size_t bytes) const {
+    void *ptr;
+    cudaMallocHost(&ptr, bytes);
+    return ptr;
   }
 
-  void init(size_t n) {
-    const size_t bytes = n * sizeof(float);
-    cudaMallocHost(&h_x, bytes);
-    cudaMallocHost(&h_s, bytes);
-    cudaMallocHost(&h_c, bytes);
-    cudaMalloc(&d_x, bytes);
-    cudaMalloc(&d_s, bytes);
-    cudaMalloc(&d_c, bytes);
-  }
+  void free_memory(void *ptr) const { cudaFreeHost(ptr); }
 
   void compute_sinf(size_t n, const float *x, float *s) const {
     const size_t bytes = n * sizeof(float);
-    std::memcpy(h_x, x, bytes);
-    cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
+    float *d_x, *d_s;
+    cudaMalloc(&d_x, bytes);
+    cudaMalloc(&d_s, bytes);
+    cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
     launch_sinf_kernel(d_x, d_s, n);
-    cudaMemcpy(h_s, d_s, bytes, cudaMemcpyDeviceToHost);
-    std::memcpy(s, h_s, bytes);
+    cudaMemcpy(s, d_s, bytes, cudaMemcpyDeviceToHost);
+    cudaFree(d_x);
+    cudaFree(d_s);
   }
 
   void compute_cosf(size_t n, const float *x, float *c) const {
     const size_t bytes = n * sizeof(float);
-    std::memcpy(h_x, x, bytes);
-    cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
+    float *d_x, *d_c;
+    cudaMalloc(&d_x, bytes);
+    cudaMalloc(&d_c, bytes);
+    cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
     launch_cosf_kernel(d_x, d_c, n);
-    cudaMemcpy(h_c, d_c, bytes, cudaMemcpyDeviceToHost);
-    std::memcpy(c, h_c, bytes);
+    cudaMemcpy(c, d_c, bytes, cudaMemcpyDeviceToHost);
+    cudaFree(d_x);
+    cudaFree(d_c);
   }
 
   void compute_sincosf(size_t n, const float *x, float *s, float *c) const {
     const size_t bytes = n * sizeof(float);
-    std::memcpy(h_x, x, bytes);
-    cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
+    float *d_x, *d_s, *d_c;
+    cudaMalloc(&d_x, bytes);
+    cudaMalloc(&d_s, bytes);
+    cudaMalloc(&d_c, bytes);
+    cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
     launch_sincosf_kernel(d_x, d_s, d_c, n);
-    cudaMemcpy(h_s, d_s, bytes, cudaMemcpyDeviceToHost);
-    cudaMemcpy(h_c, d_c, bytes, cudaMemcpyDeviceToHost);
-    std::memcpy(s, h_s, bytes);
-    std::memcpy(c, h_c, bytes);
+    cudaMemcpy(s, d_s, bytes, cudaMemcpyDeviceToHost);
+    cudaMemcpy(c, d_c, bytes, cudaMemcpyDeviceToHost);
+    cudaFree(d_x);
+    cudaFree(d_s);
+    cudaFree(d_c);
   }
-
-  float *h_x = nullptr;
-  float *h_s = nullptr;
-  float *h_c = nullptr;
-  float *d_x = nullptr;
-  float *d_s = nullptr;
-  float *d_c = nullptr;
 };
 
 GPUBackend::GPUBackend() : impl(std::make_unique<Impl>()) {}
 
 GPUBackend::~GPUBackend() = default;
 
-void GPUBackend::init(size_t n) { impl->init(n); }
+void *GPUBackend::allocate_memory(size_t bytes) const {
+  return impl->allocate_memory(bytes);
+}
+
+void GPUBackend::free_memory(void *ptr) const { impl->free_memory(ptr); }
 
 void GPUBackend::compute_sinf(size_t n, const float *x, float *s) const {
   impl->compute_sinf(n, x, s);

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;
@@ -79,7 +89,6 @@ template <std::size_t NR_SAMPLES> struct LookupAVXBackend<NR_SAMPLES>::Impl {
     constexpr std::size_t VL = 8; // AVX processes 8 floats
     const __m256 scale = _mm256_set1_ps(SCALE);
     const __m256i mask = _mm256_set1_epi32(MASK);
-    const __m256i quarter_pi = _mm256_set1_epi32(NR_SAMPLES / 4);
 
     std::size_t i = 0;
     for (; i + VL <= n; i += VL) {
@@ -94,7 +103,7 @@ template <std::size_t NR_SAMPLES> struct LookupAVXBackend<NR_SAMPLES>::Impl {
 #else
       // fallback gather for AVX1
       float sin_tmp[VL];
-      int idx_a[VL], idxc_a[VL];
+      int idx_a[VL];
       _mm256_store_si256((__m256i *)idx_a, idx);
       for (std::size_t k = 0; k < VL; ++k) {
         sin_tmp[k] = lookup[idx_a[k]];

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());
@@ -57,10 +56,10 @@ template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
       const b_type dx = xsimd::sub(vx, xsimd::mul(f_idx, pi_frac));
       const b_type dx2 = xsimd::mul(dx, dx);
       const b_type dx3 = xsimd::mul(dx2, dx);
-      const b_type dx4 = xsimd::mul(dx2, dx);
+      const b_type dx4 = xsimd::mul(dx2, dx2);
       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);
 
@@ -79,7 +78,7 @@ template <std::size_t NR_SAMPLES> struct cosf_dispatcher {
       const float dx = a[i] - idx * lookup_table_.PI_FRAC;
       const float dx2 = dx * dx;
       const float dx3 = dx2 * dx;
-      const float dx4 = dx3 * dx;
+      const float dx4 = dx2 * dx2;
       const float cosdx =
           1.0f - lookup_table_.TERM2 * dx2 + lookup_table_.TERM4 * dx4;
       const float sindx = dx - lookup_table_.TERM3 * dx3;
@@ -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());
@@ -117,10 +115,10 @@ template <std::size_t NR_SAMPLES> struct sinf_dispatcher {
       const b_type dx = xsimd::sub(vx, xsimd::mul(f_idx, pi_frac));
       const b_type dx2 = xsimd::mul(dx, dx);
       const b_type dx3 = xsimd::mul(dx2, dx);
-      const b_type dx4 = xsimd::mul(dx2, dx);
+      const b_type dx4 = xsimd::mul(dx2, dx2);
       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);
       const b_type sindx = xsimd::sub(dx, t3);
@@ -140,7 +138,7 @@ template <std::size_t NR_SAMPLES> struct sinf_dispatcher {
       const float dx = a[i] - idx * lookup_table_.PI_FRAC;
       const float dx2 = dx * dx;
       const float dx3 = dx2 * dx;
-      const float dx4 = dx3 * dx;
+      const float dx4 = dx2 * dx2;
       const float cosdx =
           1.0f - lookup_table_.TERM2 * dx2 + lookup_table_.TERM4 * dx4;
       const float sindx = dx - lookup_table_.TERM3 * dx3;
@@ -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());
@@ -180,20 +176,20 @@ template <std::size_t NR_SAMPLES> struct sin_cosf_dispatcher {
       const b_type dx = xsimd::sub(vx, xsimd::mul(f_idx, pi_frac));
       const b_type dx2 = xsimd::mul(dx, dx);
       const b_type dx3 = xsimd::mul(dx2, dx);
-      const b_type dx4 = xsimd::mul(dx2, dx);
+      const b_type dx4 = xsimd::mul(dx2, dx2);
       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);
 
       idx = xsimd::bitwise_and(idx, mask);
-      b_type sinv = b_type::gather(lookup_table_.sin_values.data(), idx);
-      b_type cosv = b_type::gather(lookup_table_.cos_values.data(), idx);
+      const b_type sinv_base = b_type::gather(lookup_table_.sin_values.data(), idx);
+      const b_type cosv_base = b_type::gather(lookup_table_.cos_values.data(), idx);
 
       const b_type cosdx = xsimd::add(xsimd::sub(term1, t2), t4);
       const b_type sindx = xsimd::sub(dx, t3);
 
-      sinv = xsimd::add(xsimd::mul(cosv, sindx), xsimd::mul(sinv, cosdx));
-      cosv = xsimd::sub(xsimd::mul(cosv, cosdx), xsimd::mul(sinv, sindx));
+      b_type sinv = xsimd::add(xsimd::mul(cosv_base, sindx), xsimd::mul(sinv_base, cosdx));
+      b_type cosv = xsimd::sub(xsimd::mul(cosv_base, cosdx), xsimd::mul(sinv_base, sindx));
 
       sinv.store(s + i, Tag());
       cosv.store(c + i, Tag());
@@ -206,7 +202,7 @@ template <std::size_t NR_SAMPLES> struct sin_cosf_dispatcher {
       const float dx = a[i] - idx * lookup_table_.PI_FRAC;
       const float dx2 = dx * dx;
       const float dx3 = dx2 * dx;
-      const float dx4 = dx3 * dx;
+      const float dx4 = dx2 * dx2;
       const float cosdx =
           1.0f - lookup_table_.TERM2 * dx2 + lookup_table_.TERM4 * dx4;
       const float sindx = dx - lookup_table_.TERM3 * dx3;

src/reference.cpp

@@ -17,7 +17,6 @@ void ReferenceBackend::compute_cosf(size_t n, const float *x, float *c) const {
 void ReferenceBackend::compute_sincosf(size_t n, const float *x, float *s,
                                        float *c) const {
   for (size_t i = 0; i < n; ++i) {
-    s[i] = sinf(x[i]);
-    c[i] = cosf(x[i]);
+    sincosf(x[i], &s[i], &c[i]);
   }
 }