Merge pull request #309 from CNugteren/CLBlast-306-omatcopy-conjugate

Fixes bug in conjugate transpose not being executed

6 files changed, 25 insertions, 5 deletions

diff --git a/CHANGELOG b/CHANGELOG
index f6d05df3..3134e7bf 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -2,6 +2,7 @@
 Development (next version)
 - Added support for shuffle instructions for NVIDIA GPUs (thanks to 'tyler-utah')
 - The tuners now check beforehand on invalid local thread sizes and skip those completely
+- Fixed an issue with conjugate transpose not being executed in certain cases for a.o. XOMATCOPY
 - Fixed an issue with AMD GPUs and the new GEMMK == 1 kernel
 - Various minor fixes and enhancements
 
diff --git a/src/routines/common.hpp b/src/routines/common.hpp
index c30a2e0e..c6db0152 100644
--- a/src/routines/common.hpp
+++ b/src/routines/common.hpp
@@ -76,6 +76,7 @@ void PadCopyTransposeMatrix(Queue &queue, const Device &device,
 
   // Determines the right kernel
   auto kernel_name = std::string{};
+  auto pad_kernel = false;
   if (do_transpose) {
     if (use_fast_kernel &&
         IsMultiple(src_ld, db["TRA_WPT"]) &&
@@ -85,7 +86,8 @@ void PadCopyTransposeMatrix(Queue &queue, const Device &device,
     }
     else {
       use_fast_kernel = false;
-      kernel_name = (do_pad) ? "TransposePadMatrix" : "TransposeMatrix";
+      pad_kernel = (do_pad || do_conjugate);
+      kernel_name = (pad_kernel) ? "TransposePadMatrix" : "TransposeMatrix";
     }
   }
   else {
@@ -97,7 +99,8 @@ void PadCopyTransposeMatrix(Queue &queue, const Device &device,
     }
     else {
       use_fast_kernel = false;
-      kernel_name = (do_pad) ? "CopyPadMatrix" : "CopyMatrix";
+      pad_kernel = do_pad;
+      kernel_name = (pad_kernel) ? "CopyPadMatrix" : "CopyMatrix";
     }
   }
 
@@ -123,7 +126,7 @@ void PadCopyTransposeMatrix(Queue &queue, const Device &device,
     kernel.SetArgument(8, static_cast<int>(dest_offset));
     kernel.SetArgument(9, dest());
     kernel.SetArgument(10, GetRealArg(alpha));
-    if (do_pad) {
+    if (pad_kernel) {
       kernel.SetArgument(11, static_cast<int>(do_conjugate));
     }
     else {
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>