AVX2 instead of AVX__2

I swear I had fixed this already mmh

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

benchmarks/benchmark_utils.hpp

@@ -26,9 +26,6 @@ static void benchmark_sinf(benchmark::State &state) {
       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)

include/trigdx/interface.hpp

@@ -16,7 +16,7 @@ public:
     return static_cast<void *>(new uint8_t[bytes]);
   };
 
-  virtual void free_memory(void *ptr) const { delete[] static_cast<uint8_t*>(ptr); };
+  virtual void free_memory(void *ptr) const { std::free(ptr); };
 
   // Compute sine for n elements
   virtual void compute_sinf(size_t n, const float *x, float *s) const = 0;

src/lookup_avx.cpp

@@ -89,6 +89,7 @@ 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) {
@@ -103,7 +104,7 @@ template <std::size_t NR_SAMPLES> struct LookupAVXBackend<NR_SAMPLES>::Impl {
 #else
       // fallback gather for AVX1
       float sin_tmp[VL];
-      int idx_a[VL];
+      int idx_a[VL], idxc_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

@@ -56,7 +56,7 @@ 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, dx2);
+      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);
@@ -78,7 +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 = dx2 * dx2;
+      const float dx4 = dx3 * dx;
       const float cosdx =
           1.0f - lookup_table_.TERM2 * dx2 + lookup_table_.TERM4 * dx4;
       const float sindx = dx - lookup_table_.TERM3 * dx3;
@@ -115,7 +115,7 @@ 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, dx2);
+      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);
@@ -138,7 +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 = dx2 * dx2;
+      const float dx4 = dx3 * dx;
       const float cosdx =
           1.0f - lookup_table_.TERM2 * dx2 + lookup_table_.TERM4 * dx4;
       const float sindx = dx - lookup_table_.TERM3 * dx3;
@@ -176,20 +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, dx2);
+      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);
 
       idx = xsimd::bitwise_and(idx, mask);
-      const b_type sinv_base = b_type::gather(lookup_table_.sin_values.data(), idx);
+      b_type sinv = b_type::gather(lookup_table_.sin_values.data(), idx);
-      const b_type cosv_base = b_type::gather(lookup_table_.cos_values.data(), idx);
+      b_type cosv = 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);
 
-      b_type sinv = xsimd::add(xsimd::mul(cosv_base, sindx), xsimd::mul(sinv_base, cosdx));
+      sinv = xsimd::add(xsimd::mul(cosv, sindx), xsimd::mul(sinv, cosdx));
-      b_type cosv = xsimd::sub(xsimd::mul(cosv_base, cosdx), xsimd::mul(sinv_base, sindx));
+      cosv = xsimd::sub(xsimd::mul(cosv, cosdx), xsimd::mul(sinv, sindx));
 
       sinv.store(s + i, Tag());
       cosv.store(c + i, Tag());
@@ -202,7 +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 = dx2 * dx2;
+      const float dx4 = dx3 * dx;
       const float cosdx =
           1.0f - lookup_table_.TERM2 * dx2 + lookup_table_.TERM4 * dx4;
       const float sindx = dx - lookup_table_.TERM3 * dx3;

src/reference.cpp

@@ -17,6 +17,7 @@ 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) {
-    sincosf(x[i], &s[i], &c[i]);
+    s[i] = sinf(x[i]);
+    c[i] = cosf(x[i]);
   }
 }