Stripped out OpenMP SIMD macros as they're not worth the effort.

Some testing on godbolt indicates that the compiler autovectorizes wherever it can anyway; the OpenMP macros don't change the compiler's decision much, at least not in the places that it's used in this codebase. So, we remove them to reduce the risk of incorrectness and the amount of testing we have to do.
libscran · Sep 2, 2024 · 95d4d20 · 95d4d20
1 parent f540b25
commit 95d4d20
Show file tree

Hide file tree

Showing 3 changed files with 0 additions and 24 deletions.
diff --git a/include/qdtsne/SPTree.hpp b/include/qdtsne/SPTree.hpp
@@ -288,18 +288,12 @@ class SPTree {
         Float_ mult = count * div;
         result_sum += mult;
         mult *= div;
-#ifdef _OPENMP
-        #pragma omp simd
-#endif
         for (int d = 0; d < num_dim_; ++d) {
             neg_f[d] += mult * (point[d] - center[d]);
         }
     }
 
     static void remove_self_from_center(const Float_* point, const std::array<Float_, num_dim_>& center, Float_ count, std::array<Float_, num_dim_>& temp) {
-#ifdef _OPENMP
-        #pragma omp simd
-#endif
         for (int d = 0; d < num_dim_; ++d) { 
             temp[d] = (center[d] * count - point[d]) / (count - 1);
         }

diff --git a/include/qdtsne/Status.hpp b/include/qdtsne/Status.hpp
@@ -185,9 +185,6 @@ class Status {
 
         // Update gains
         size_t ngains = my_gains.size(); 
-#ifdef _OPENMP
-        #pragma omp simd
-#endif
         for (size_t i = 0; i < ngains; ++i) {
             Float_& g = my_gains[i];
             constexpr Float_ lower_bound = 0.01;
@@ -197,9 +194,6 @@ class Status {
         }
 
         // Perform gradient update (with momentum and gains)
-#ifdef _OPENMP
-        #pragma omp simd
-#endif
         for (size_t i = 0; i < ngains; ++i) {
             my_uY[i] = momentum * my_uY[i] - my_options.eta * my_gains[i] * my_dY[i];
             Y[i] += my_uY[i];

diff --git a/include/qdtsne/gaussian.hpp b/include/qdtsne/gaussian.hpp
@@ -42,9 +42,6 @@ void compute_gaussian_perplexity(NeighborList<Index_, Float_>& neighbors, Float_
                 const Float_ first = current[0].second;
                 const Float_ first2 = first * first;
 
-#ifdef _OPENMP
-                #pragma omp simd
-#endif
                 for (int m = 1; m < K; ++m) {
                     Float_ dist = current[m].second;
                     Float_ squared_delta_dist_raw = dist * dist - first2; 
@@ -61,9 +58,6 @@ void compute_gaussian_perplexity(NeighborList<Index_, Float_>& neighbors, Float_
                 for (int iter = 0; iter < 200; ++iter) {
                     // We skip the first value because we know that squared_delta_dist[0] = 0
                     // (as we subtracted 'first') and thus output[0] = 1.
-#ifdef _OPENMP
-                    #pragma omp simd
-#endif
                     for (int m = 1; m < K; ++m) {
                         output[m] = std::exp(-beta * squared_delta_dist[m]); 
                     }
@@ -111,9 +105,6 @@ void compute_gaussian_perplexity(NeighborList<Index_, Float_>& neighbors, Float_
 
                     if (std::isinf(beta)) {
                         // Avoid propagation of NaNs via Inf * 0. 
-#ifdef _OPENMP
-                        #pragma omp simd
-#endif
                         for (int m = 1; m < K; ++m) {
                             output[m] = (squared_delta_dist[m] == 0);
                         }
@@ -122,9 +113,6 @@ void compute_gaussian_perplexity(NeighborList<Index_, Float_>& neighbors, Float_
                 }
 
                 // Row-normalize current row of P.
-#ifdef _OPENMP
-                #pragma omp simd
-#endif
                 for (int m = 0; m < K; ++m) {
                     current[m].second = output[m] / sum_P;
                 }