Improve CPU reference accurary

The relative difference between contraction result and CPU reference is less than 0.1% after the improvement.
ROCm · Dec 6, 2023 · 89d5c93 · 89d5c93
1 parent 1d7aeee
commit 89d5c93
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 3 deletions.
diff --git a/library/src/contraction/contraction_cpu_reference_impl.hpp b/library/src/contraction/contraction_cpu_reference_impl.hpp
@@ -173,7 +173,8 @@ namespace hiptensor
                             arg.mOpB(valB, ((BDataType*)arg.mB)[indexB]);
 
                             // Mult / accum
-                            accum += ck::type_convert<float>(valA) * ck::type_convert<float>(valB);
+                            accum += ck::type_convert<float>(ck::type_convert<ComputeDataType>(
+                                ck::type_convert<float>(valA) * ck::type_convert<float>(valB)));
                         }
                     }
 

diff --git a/test/utils.hpp b/test/utils.hpp
@@ -140,7 +140,7 @@ template <typename DDataType>
 std::pair<bool, double> compareEqual(DDataType const* deviceD,
                                      DDataType const* hostD,
                                      std::size_t      elementsD,
-                                     double           tolerance = 0.005)
+                                     double           tolerance = 0.001)
 {
     bool   retval             = true;
     double max_relative_error = 0.0;
@@ -214,7 +214,7 @@ template <typename DDataType>
 std::pair<bool, double> compareEqualLaunchKernel(DDataType*  deviceD,
                                                  DDataType*  hostD,
                                                  std::size_t elementsD,
-                                                 double      tolerance = 0.005)
+                                                 double      tolerance = 0.001)
 {
     auto blockDim = dim3(1024, 1, 1);
     auto gridDim  = dim3(ceilDiv(elementsD, blockDim.x), 1, 1);