[Trifocal+P2Pt] all aligned 64 improves vectorization

rfabbri · Nov 3, 2023 · ce7e756 · ce7e756
1 parent dc5f1d2
commit ce7e756
Show file tree

Hide file tree

Showing 4 changed files with 7 additions and 24 deletions.
diff --git a/minus/chicago14a-default-data.h b/minus/chicago14a-default-data.h
@@ -13,8 +13,8 @@ struct minus_data<chicago14a,F> {
   typedef minus_io_14a<chicago14a> io14;
   typedef std::complex<F> complex;
   static const complex start_sols_[M::nve*M::nsols];
-  static complex params_start_target_[2*M::f::nparams];
-  static complex default_params_start_target_gammified_[2*M::f::nparams];
+  alignas (64) static complex params_start_target_[2*M::f::nparams];
+  alignas (64) static complex default_params_start_target_gammified_[2*M::f::nparams];
   static const complex *params_;
   static F p_[io::pp::nviews][io::pp::npoints][io::ncoords2d];
   static const F p_correct_[io::pp::nviews][io::pp::npoints][io::ncoords2d];

diff --git a/minus/chicago14a-default-data.hxx b/minus/chicago14a-default-data.hxx
@@ -28,7 +28,7 @@ namespace MiNuS {
 // Internal note: m2: script t, variable sols
 // You can reinterpret in C as start_sols_[M::nsols][M::nve]
 template <typename F>
-alignas(sizeof(std::complex<F>)) const std::complex<F> minus_data<chicago14a,F>::
+alignas(64) const std::complex<F> minus_data<chicago14a,F>::
 start_sols_[M::nve*M::nsols] = {
   {-.59336028545681196,-.11013183013512155},
   {.11944671140724233,-.13633687755694085},
@@ -4729,7 +4729,7 @@ start_sols_[M::nve*M::nsols] = {
 //                                  to store target system params
 //                                  the latter M::nparams are trash
 template <typename F>
-alignas(sizeof(std::complex<F>)) std::complex<F> minus_data<chicago14a,F>::
+alignas(64) std::complex<F> minus_data<chicago14a,F>::
 params_start_target_[2*M::f::nparams] = {
   {.13016671344237549,-.36891394723672405},
   {.2649393534275909,-.23418132862391827},
@@ -4800,7 +4800,7 @@ params_start_target_[2*M::f::nparams] = {
 //
 // The point-tangent inputs giving rise to this are given below
 template <typename F>
-alignas(sizeof(std::complex<F>)) std::complex<F> minus_data<chicago14a,F>::
+alignas(64) std::complex<F> minus_data<chicago14a,F>::
 default_params_start_target_gammified_[2*M::f::nparams] = { // start-target param pairs, P01 in chicago.m2
   {.391195550619826,-.00262962533857666},
   {.310140709227333,+.169842562835882},

diff --git a/minus/chicago14a.hxx b/minus/chicago14a.hxx
@@ -3596,7 +3596,7 @@ minus<chicago14a, F>::solve(
     )
 {
   typedef minus_data<chicago14a,F> data;
-  alignas(sizeof(C<F>)) C<F> params[2*M::f::nparams];
+  alignas(64) C<F> params[2*M::f::nparams];
   memcpy(params, data::params_start_target_, M::f::nparams*sizeof(C<F>));
 
   constexpr int id_tgt0 = 0; constexpr int id_tgt1 = 1; // TODO: select the best / least degenerate directions

diff --git a/minus/minus.hxx b/minus/minus.hxx
@@ -103,24 +103,7 @@ track(const track_settings &s, const C<F> s_sols[f::nve*f::nsols], const C<F> pa
   asm("#------ ENDMUL!");
 
   asm("#------ Add to self!"); // not aligned
-    //  v::add_to_self(xt, dx4);
-    xt[0] += dx4[0];
-    xt[1] += dx4[1];
-    xt[2] += dx4[2];
-    xt[3] += dx4[3];
-    xt[4] += dx4[4];
-    xt[5] += dx4[5];
-    xt[6] += dx4[6];
-    xt[7] += dx4[7];
-    xt[8] += dx4[8];
-    xt[9] += dx4[9];
-    xt[10] += dx4[10];
-    xt[11] += dx4[11];
-    xt[12] += dx4[12];
-    xt[13] += dx4[13];
-//    xt[14] += dx4[14];
-//    xt[15] += dx4[15];
-//    xt[16] += dx4[16];
+    v::add_to_self(xt, dx4);
   asm("#------ end add to self!");
       v::multiply_scalar_to_self(dx4, 2.);
       xt[f::nve] += one_half_dt;  // t0+.5dt