ta_cc_abcd: GFLOP = 10^9 FLOP, not 12^30 FLOP

ValeevGroup · Jul 22, 2024 · 269c59b · 269c59b
1 parent c9302ed
commit 269c59b
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 5 deletions.
diff --git a/examples/device/ta_cc_abcd_device.cpp b/examples/device/ta_cc_abcd_device.cpp
@@ -182,8 +182,8 @@ void cc_abcd(TA::World& world, const TA::TiledRange1& trange_occ,
   const double flops_per_fma =
       (complex_T ? 8 : 2);  // 1 multiply takes 6/1 flops for complex/real
                             // 1 add takes 2/1 flops for complex/real
-  const double n_gflop = flops_per_fma * std::pow(n_occ, 2) *
-                         std::pow(n_uocc, 4) / std::pow(1024., 3);
+  const double n_gflop =
+      flops_per_fma * std::pow(n_occ, 2) * std::pow(n_uocc, 4) / 1e9;
 
   using deviceTile =
       btas::Tensor<T, TA::Range, TiledArray::device_um_btas_varray<T>>;

diff --git a/examples/gemm/ta_cc_abcd.cpp b/examples/gemm/ta_cc_abcd.cpp
@@ -211,8 +211,8 @@ void cc_abcd(TA::World& world, const TA::TiledRange1& trange_occ,
   const double flops_per_fma =
       (complex_T ? 8 : 2);  // 1 multiply takes 6/1 flops for complex/real
                             // 1 add takes 2/1 flops for complex/real
-  const double gflops_per_call = flops_per_fma * std::pow(n_occ, 2) *
-                                 std::pow(n_uocc, 4) / std::pow(1024., 3);
+  const double gflops_per_call =
+      flops_per_fma * std::pow(n_occ, 2) * std::pow(n_uocc, 4) / 1e9;
 
   // Construct tensors
   TA::TArrayD t2(world, trange_oovv);

diff --git a/examples/gemm/ta_dense_asymm.cpp b/examples/gemm/ta_dense_asymm.cpp
@@ -148,10 +148,11 @@ int main(int argc, char** argv) {
       if (do_memtrace) {
         world.gop.fence();
         madness::print_meminfo(world.rank(), str);
+      } else {
+        world.gop.fence();
       }
 #ifdef TA_TENSOR_MEM_PROFILE
       {
-        world.gop.fence();
         std::cout
             << str << ": TA::Tensor allocated "
             << TA::hostEnv::instance()->host_allocator_getActualHighWatermark()