templating sd_add

1 files changed, 73 insertions, 0 deletions

diff --git a/vowpalwabbit/sparse_dense.h b/vowpalwabbit/sparse_dense.h
index 687488fe..b125a945 100644
--- a/vowpalwabbit/sparse_dense.h
+++ b/vowpalwabbit/sparse_dense.h
@@ -15,12 +15,85 @@ inline float trunc_weight(float w, float gravity){
   return (gravity < fabsf(w)) ? w - sign(w) * gravity : 0.f;
 }
 
+template <void (*T)(vw&,float,uint32_t,float&)>
+void sd_add(vw& all, feature* begin, feature* end, float& ret)
+{
+  for (feature* f = begin; f!= end; f++)
+    T(all, f->x, f->weight_index, ret);
+}
+
+inline void vec_add(vw& all, float fx, uint32_t fi, float& ret) {
+  ret += all.reg.weight_vectors[fi & all.weight_mask] * fx;
+}
+
+inline void vec_add_trunc(vw& all, float fx, uint32_t fi, float& ret) {
+  ret += trunc_weight(all.reg.weight_vectors[fi & all.weight_mask], all.sd->gravity) * fx;
+}
+
+inline void vec_add_rescale(vw& all, float fx, uint32_t fi, float& ret) {
+  weight* w = &all.reg.weight_vectors[fi & all.weight_mask];
+  float x_abs = fabs(fx);
+  if( x_abs > w[all.normalized_idx] ) {
+    if( w[all.normalized_idx] > 0. ) {
+      float rescale = (w[all.normalized_idx]/x_abs);
+      w[0] *= (all.adaptive ? rescale : rescale*rescale);
+    }
+    w[all.normalized_idx] = x_abs;
+  }
+  ret += w[0] * fx;
+}
+
+inline void vec_add_trunc_rescale(vw& all, float fx, uint32_t fi, float& ret) {
+  weight* w = &all.reg.weight_vectors[fi & all.weight_mask];
+  float x_abs = fabs(fx);
+  if( x_abs > w[all.normalized_idx] ) {
+    if( w[all.normalized_idx] > 0. ) {
+      float rescale = (w[all.normalized_idx]/x_abs);
+      w[0] *= (all.adaptive ? rescale : rescale*rescale);
+    }
+    w[all.normalized_idx] = x_abs;
+  }
+  ret += trunc_weight(w[0], all.sd->gravity) * fx;
+}
+
+inline void vec_add_rescale_general(vw& all, float fx, uint32_t fi, float& ret) {
+  weight* w = &all.reg.weight_vectors[fi & all.weight_mask];
+  float x_abs = fabs(fx);
+  float power_t_norm = 1.f - (all.adaptive ? all.power_t : 0.f);
+  if( x_abs > w[all.normalized_idx] ) {
+    if( w[all.normalized_idx] > 0. ) {
+      float rescale = (w[all.normalized_idx]/x_abs);
+      w[0] *= powf(rescale*rescale,power_t_norm);
+    }
+    w[all.normalized_idx] = x_abs;
+  }
+  ret += w[0] * fx;
+}
+
+inline void vec_add_trunc_rescale_general(vw& all, float fx, uint32_t fi, float& ret) {
+  weight* w = &all.reg.weight_vectors[fi & all.weight_mask];
+  float x_abs = fabs(fx);
+  float power_t_norm = 1.f - (all.adaptive ? all.power_t : 0.f);
+  if( x_abs > w[all.normalized_idx] ) {
+    if( w[all.normalized_idx] > 0. ) {
+      float rescale = (w[all.normalized_idx]/x_abs);
+      w[0] *= powf(rescale*rescale,power_t_norm);
+    }
+    w[all.normalized_idx] = x_abs;
+  }
+  ret += trunc_weight(w[0], all.sd->gravity) * fx;
+}
+
+
+
+/*
 float sd_add(weight* weights, size_t mask, feature* begin, feature* end);
 float sd_add_trunc(weight* weights, size_t mask, feature* begin, feature* end, float gravity);
 float sd_add_rescale(weight* weights, size_t mask, feature* begin, feature* end, bool is_adaptive, size_t idx_norm);
 float sd_add_trunc_rescale(weight* weights, size_t mask, feature* begin, feature* end, float gravity, bool is_adaptive, size_t idx_norm);
 float sd_add_rescale_general(weight* weights, size_t mask, feature* begin, feature* end, size_t idx_norm, float power_t_norm);
 float sd_add_trunc_rescale_general(weight* weights, size_t mask, feature* begin, feature* end, float gravity, size_t idx_norm, float power_t_norm);
+*/
 
 float sd_offset_add(weight* weights, size_t mask, feature* begin, feature* end, size_t offset);
 void sd_offset_update(weight* weights, size_t mask, feature* begin, feature* end, size_t offset, float update, float regularization);