Merge branch 'master' into convgemm_multi_kernel

5 files changed, 31 insertions, 3 deletions

diff --git a/test/correctness/misc/preprocessor.cpp b/test/correctness/misc/preprocessor.cpp
index c5d115d3..4d6fa114 100644
--- a/test/correctness/misc/preprocessor.cpp
+++ b/test/correctness/misc/preprocessor.cpp
@@ -221,7 +221,7 @@ size_t RunPreprocessor(int argc, char *argv[], const bool silent, const Precisio
   ;
   if (TestKernel(device, context, "TransposePadMatrix", transpose_pad_sources, precision)) { passed++; } else { errors++; }
 
-  // GEMM (in-direct)
+  // GEMM (in-direct) GEMMK==0
   const auto gemm_sources =
     "#define KWI 2\n"
     "#define MWG 16\n"
@@ -234,6 +234,18 @@ size_t RunPreprocessor(int argc, char *argv[], const bool silent, const Precisio
   ;
   if (TestKernel(device, context, "Xgemm", gemm_sources, precision)) { passed++; } else { errors++; }
 
+  // GEMM (in-direct) GEMMK==1
+  const auto gemm_sources_gemmk1 =
+    "#define MWG 16\n"
+    "#define NWG 16\n"
+    "#define GEMMK 1\n"
+    #include "../src/kernels/level3/xgemm_part1.opencl"
+    #include "../src/kernels/level3/xgemm_part2.opencl"
+    #include "../src/kernels/level3/xgemm_part3.opencl"
+    #include "../src/kernels/level3/xgemm_part4.opencl"
+  ;
+  if (TestKernel(device, context, "Xgemm", gemm_sources_gemmk1, precision)) { passed++; } else { errors++; }
+
   // GEMM (direct)
   const auto gemm_direct_sources =
     "#define KWID 2\n"
diff --git a/test/correctness/testblas.cpp b/test/correctness/testblas.cpp
index 3c92565e..d28aba40 100644
--- a/test/correctness/testblas.cpp
+++ b/test/correctness/testblas.cpp
@@ -239,7 +239,7 @@ void TestBlas<T,U>::TestRegular(std::vector<Arguments<U>> &test_vector, const st
     }
 
     // Tests the error count (should be zero)
-    TestErrorCount(errors, get_id1_(args)*get_id2_(args), args);
+    TestErrorCount(errors, get_id1_(args)*get_id2_(args) + kCanarySize, args);
   }
   TestEnd();
 }
diff --git a/test/routines/levelx/xomatcopy.hpp b/test/routines/levelx/xomatcopy.hpp
index ea35dbe2..4a93b29d 100644
--- a/test/routines/levelx/xomatcopy.hpp
+++ b/test/routines/levelx/xomatcopy.hpp
@@ -45,7 +45,9 @@ StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host)
       const auto b_two = (b_rotated) ? id1 : id2;
       const auto a_index = a_two * args.a_ld + a_one + args.a_offset;
       const auto b_index = b_two * args.b_ld + b_one + args.b_offset;
-      buffers_host.b_mat[b_index] = args.alpha * buffers_host.a_mat[a_index];
+      auto a_value = buffers_host.a_mat[a_index];
+      if (args.a_transpose == Transpose::kConjugate) { a_value = ComplexConjugate(a_value); }
+      buffers_host.b_mat[b_index] = args.alpha * a_value;
     }
   }
   return StatusCode::kSuccess;
diff --git a/test/test_utilities.cpp b/test/test_utilities.cpp
index 59ec949d..c43200b9 100644
--- a/test/test_utilities.cpp
+++ b/test/test_utilities.cpp
@@ -31,6 +31,16 @@ template <> bool IsCloseToZero(const double2 value) { return IsCloseToZero(value
 
 // =================================================================================================
 
+// Performs a complex conjugate if complex
+template <typename T> T ComplexConjugate(const T value) { return value; }
+template half ComplexConjugate(const half);
+template float ComplexConjugate(const float);
+template double ComplexConjugate(const double);
+template <> float2 ComplexConjugate(const float2 value) { return float2{value.real(), -value.imag()}; }
+template <> double2 ComplexConjugate(const double2 value) { return double2{value.real(), -value.imag()}; }
+
+// =================================================================================================
+
 template <typename T, typename U>
 void DeviceToHost(const Arguments<U> &args, Buffers<T> &buffers, BuffersHost<T> &buffers_host,
                   Queue &queue, const std::vector<std::string> &names) {
diff --git a/test/test_utilities.hpp b/test/test_utilities.hpp
index 42660bdb..7bf5e65f 100644
--- a/test/test_utilities.hpp
+++ b/test/test_utilities.hpp
@@ -70,6 +70,10 @@ struct BuffersHost {
 
 // =================================================================================================
 
+template <typename T> T ComplexConjugate(const T value);
+
+// =================================================================================================
+
 // Converts a value (e.g. an integer) to a string. This also covers special cases for CLBlast
 // data-types such as the Layout and Transpose data-types.
 template <typename T>