Switched from separate sin and cos functions to sincos in CUDA, achieving around 30% speedup in local benchmark.

Author: CarloAndreaDeFilippo

src/cuda/rho1d_kernel.cu

@@ -17,7 +17,9 @@ __global__ void rho1D_kernel_add(const double* proj, size_t N,
 
   for (size_t i = 0; i < N; ++i) {
     double phase = proj[i] * q;
-    sum = cuCadd(sum, make_cuDoubleComplex(cos(phase), -sin(phase)));
+    double sPhase, cPhase;
+    sincos(phase, &sPhase, &cPhase);
+    sum = cuCadd(sum, make_cuDoubleComplex(sPhase, -cPhase));
   }
 
   rho[qq] = cuCadd(rho[qq], sum);

src/cuda/rho2d_kernel.cu

@@ -1,7 +1,8 @@
 #include <cuComplex.h>
 #include <math.h>
 
-extern "C" __global__ void rho2D_kernel_add(const double* proj1, const double* proj2, size_t N,
+extern "C" __global__ void rho2D_kernel_add(const double* proj1,
+                                            const double* proj2, size_t N,
                                             const double* q1vals, size_t M1,
                                             const double* q2vals, size_t M2,
                                             cuDoubleComplex* pos_pos,
@@ -24,8 +25,11 @@ extern "C" __global__ void rho2D_kernel_add(const double* proj1, const double* p
     double anglePP = phase1 + phase2;
     double anglePN = phase1 - phase2;
 
-    sumPP = cuCadd(sumPP, make_cuDoubleComplex(cos(anglePP), -sin(anglePP)));
-    sumPN = cuCadd(sumPN, make_cuDoubleComplex(cos(anglePN), -sin(anglePN)));
+    double sPP, cPP, sPN, cPN;
+    sincos(anglePP, &sPP, &cPP);
+    sincos(anglePN, &sPN, &cPN);
+    sumPP = cuCadd(sumPP, make_cuDoubleComplex(cPP, -sPP));
+    sumPN = cuCadd(sumPN, make_cuDoubleComplex(cPN, -sPN));
   }
 
   int idx = qq1 * M2 + qq2;