Fixed a namespace clash with CUDA FP16 for the half-datatype

12 files changed, 14 insertions, 0 deletions

diff --git a/src/clblast_c.cpp b/src/clblast_c.cpp
index f5104bad..b6a64749 100644
--- a/src/clblast_c.cpp
+++ b/src/clblast_c.cpp
@@ -19,6 +19,7 @@
 #include "clblast.h"
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/copy_fast.cpp b/src/tuning/kernels/copy_fast.cpp
index 10ef864b..8d759a80 100644
--- a/src/tuning/kernels/copy_fast.cpp
+++ b/src/tuning/kernels/copy_fast.cpp
@@ -105,6 +105,7 @@ class TuneCopy {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/copy_pad.cpp b/src/tuning/kernels/copy_pad.cpp
index 1feb5683..2197e1c4 100644
--- a/src/tuning/kernels/copy_pad.cpp
+++ b/src/tuning/kernels/copy_pad.cpp
@@ -113,6 +113,7 @@ class TunePad {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/transpose_fast.cpp b/src/tuning/kernels/transpose_fast.cpp
index 433f9972..085cb791 100644
--- a/src/tuning/kernels/transpose_fast.cpp
+++ b/src/tuning/kernels/transpose_fast.cpp
@@ -110,6 +110,7 @@ class TuneTranspose {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/transpose_pad.cpp b/src/tuning/kernels/transpose_pad.cpp
index d7dc585d..a437e6da 100644
--- a/src/tuning/kernels/transpose_pad.cpp
+++ b/src/tuning/kernels/transpose_pad.cpp
@@ -117,6 +117,7 @@ class TunePadTranspose {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/xaxpy.cpp b/src/tuning/kernels/xaxpy.cpp
index 7984e184..a13e54f2 100644
--- a/src/tuning/kernels/xaxpy.cpp
+++ b/src/tuning/kernels/xaxpy.cpp
@@ -108,6 +108,7 @@ class TuneXaxpy {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/xdot.cpp b/src/tuning/kernels/xdot.cpp
index faf52089..3165eec6 100644
--- a/src/tuning/kernels/xdot.cpp
+++ b/src/tuning/kernels/xdot.cpp
@@ -114,6 +114,7 @@ class TuneXdot {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/xgemm.cpp b/src/tuning/kernels/xgemm.cpp
index d34035f4..9b4cea8b 100644
--- a/src/tuning/kernels/xgemm.cpp
+++ b/src/tuning/kernels/xgemm.cpp
@@ -173,6 +173,7 @@ class TuneXgemm {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/xgemm_direct.cpp b/src/tuning/kernels/xgemm_direct.cpp
index 5afcdd38..e7a35b93 100644
--- a/src/tuning/kernels/xgemm_direct.cpp
+++ b/src/tuning/kernels/xgemm_direct.cpp
@@ -172,6 +172,7 @@ class TuneXgemmDirect {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/xgemv.cpp b/src/tuning/kernels/xgemv.cpp
index c34e8a1c..9f6821e9 100644
--- a/src/tuning/kernels/xgemv.cpp
+++ b/src/tuning/kernels/xgemv.cpp
@@ -154,6 +154,7 @@ class TuneXgemv {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/tuning/kernels/xger.cpp b/src/tuning/kernels/xger.cpp
index c3fc243b..30efba48 100644
--- a/src/tuning/kernels/xger.cpp
+++ b/src/tuning/kernels/xger.cpp
@@ -113,6 +113,7 @@ class TuneXger {
 } // namespace clblast
 
 // Shortcuts to the clblast namespace
+using half = clblast::half;
 using float2 = clblast::float2;
 using double2 = clblast::double2;
 
diff --git a/src/utilities/utilities.hpp b/src/utilities/utilities.hpp
index b40ec541..09394cf3 100644
--- a/src/utilities/utilities.hpp
+++ b/src/utilities/utilities.hpp
@@ -31,6 +31,9 @@
 namespace clblast {
 // =================================================================================================
 
+// Shorthands for half-precision
+using half = cl_half; // based on the OpenCL type, which is actually an 'unsigned short'
+
 // Shorthands for complex data-types
 using float2 = std::complex<float>;
 using double2 = std::complex<double>;