faster segment coo cpu implementation

Author: rusty1s

cpu/segment.cpp

@@ -180,6 +180,7 @@ segment_coo(at::Tensor src, at::Tensor index, at::Tensor out,
     arg_out_data = arg_out.value().DATA_PTR<int64_t>();
   }
 
+  auto E = index.numel();
   auto E_1 = index.numel() / src.size(reduce_dim);
   auto E_2 = src.size(reduce_dim);
   auto K = src.numel() / index.numel();
@@ -191,41 +192,48 @@ segment_coo(at::Tensor src, at::Tensor index, at::Tensor out,
     auto src_data = src.DATA_PTR<scalar_t>();
     auto out_data = out.DATA_PTR<scalar_t>();
 
-    scalar_t val;
-    int64_t idx, next_idx, row_start, arg;
+    scalar_t vals[K];
+    int64_t idx, next_idx, row_start, args[K];
     AT_DISPATCH_REDUCTION_TYPES(reduce, [&] {
       for (int e_1 = 0; e_1 < E_1; e_1++) {
         int offset = IndexToOffset<int64_t>::get(e_1 * E_2, index_info);
+        idx = index_info.data[offset];
+        row_start = 0;
 
         for (int k = 0; k < K; k++) {
-          idx = index_info.data[offset];
-          row_start = 0;
-          val = out_data[e_1 * N * K + k];
+          vals[k] = out_data[e_1 * N * K + k];
+        }
 
-          for (int e_2 = 0; e_2 < E_2; e_2++) {
+        for (int e_2 = 0; e_2 < E_2; e_2++) {
+
+          for (int k = 0; k < K; k++) {
             Reducer<scalar_t, REDUCE>::update(
-                &val, src_data[e_1 * E_2 * K + e_2 * K + k], &arg, e_2);
+                &vals[k], src_data[e_1 * E_2 * K + e_2 * K + k], &args[k], e_2);
+          }
 
-            if (e_2 == E_2 - 1) {
+          if (e_2 == E_2 - 1) {
+            for (int k = 0; k < K; k++) {
               Reducer<scalar_t, REDUCE>::write(
-                  out_data + e_1 * N * K + idx * K + k, val,
-                  arg_out_data + e_1 * N * K + idx * K + k, arg,
+                  out_data + e_1 * N * K + idx * K + k, vals[k],
+                  arg_out_data + e_1 * N * K + idx * K + k, args[k],
                   e_2 + 1 - row_start);
-            } else {
-              next_idx = index_info.data[offset + (e_2 + 1) * stride];
+            }
+          } else {
+            next_idx = index_info.data[offset + (e_2 + 1) * stride];
 
-              if (idx != next_idx) {
+            if (idx != next_idx) {
+              for (int k = 0; k < K; k++) {
                 Reducer<scalar_t, REDUCE>::write(
-                    out_data + e_1 * N * K + idx * K + k, val,
-                    arg_out_data + e_1 * N * K + idx * K + k, arg,
+                    out_data + e_1 * N * K + idx * K + k, vals[k],
+                    arg_out_data + e_1 * N * K + idx * K + k, args[k],
                     e_2 + 1 - row_start);
 
-                row_start = e_2 + 1;
-                val = out_data[e_1 * N * K + next_idx * K + k];
+                vals[k] = out_data[e_1 * N * K + next_idx * K + k];
               }
-
-              idx = next_idx;
+              row_start = e_2 + 1;
             }
+
+            idx = next_idx;
           }
         }
       }