Merge pull request #204 from CNugteren/cuda_api

Cuda API to CLBlast

4 files changed, 19 insertions, 11 deletions

diff --git a/src/utilities/buffer_test.hpp b/src/utilities/buffer_test.hpp
index b5693181..fd071434 100644
--- a/src/utilities/buffer_test.hpp
+++ b/src/utilities/buffer_test.hpp
@@ -15,7 +15,7 @@
 #ifndef CLBLAST_BUFFER_TEST_H_
 #define CLBLAST_BUFFER_TEST_H_
 
-#include "clblast.h"
+#include "utilities/utilities.hpp"
 
 namespace clblast {
 // =================================================================================================
diff --git a/src/utilities/clblast_exceptions.hpp b/src/utilities/clblast_exceptions.hpp
index 0d0033b6..a790be9c 100644
--- a/src/utilities/clblast_exceptions.hpp
+++ b/src/utilities/clblast_exceptions.hpp
@@ -16,8 +16,7 @@
 #ifndef CLBLAST_EXCEPTIONS_H_
 #define CLBLAST_EXCEPTIONS_H_
 
-#include "clpp11.hpp"
-#include "clblast.h"
+#include "utilities/utilities.hpp"
 
 namespace clblast {
 // =================================================================================================
diff --git a/src/utilities/utilities.cpp b/src/utilities/utilities.cpp
index a5c1d45e..f2574104 100644
--- a/src/utilities/utilities.cpp
+++ b/src/utilities/utilities.cpp
@@ -413,13 +413,17 @@ std::string GetDeviceVendor(const Device& device) {
 // Mid-level info
 std::string GetDeviceArchitecture(const Device& device) {
   auto device_architecture = std::string{""};
-  if (device.HasExtension(kKhronosAttributesNVIDIA)) {
+  #ifdef CUDA_API
     device_architecture = device.NVIDIAComputeCapability();
-  }
-  else if (device.HasExtension(kKhronosAttributesAMD)) {
-    device_architecture = device.Name(); // Name is architecture for AMD APP and AMD ROCm
-  }
-  // Note: no else - 'device_architecture' might be the empty string
+  #else
+    if (device.HasExtension(kKhronosAttributesNVIDIA)) {
+      device_architecture = device.NVIDIAComputeCapability();
+    }
+    else if (device.HasExtension(kKhronosAttributesAMD)) {
+      device_architecture = device.Name(); // Name is architecture for AMD APP and AMD ROCm
+    }
+    // Note: no else - 'device_architecture' might be the empty string
+  #endif
 
   for (auto &find_and_replace : device_mapping::kArchitectureNames) { // replacing to common names
     if (device_architecture == find_and_replace.first) { device_architecture = find_and_replace.second; }
diff --git a/src/utilities/utilities.hpp b/src/utilities/utilities.hpp
index b2949c27..f56226be 100644
--- a/src/utilities/utilities.hpp
+++ b/src/utilities/utilities.hpp
@@ -21,8 +21,13 @@
 #include <complex>
 #include <random>
 
-#include "clpp11.hpp"
-#include "clblast.h"
+#ifdef OPENCL_API
+  #include "clpp11.hpp"
+  #include "clblast.h"
+#elif CUDA_API
+  #include "cupp11.hpp"
+  #include "clblast_cuda.h"
+#endif
 #include "clblast_half.h"
 #include "utilities/clblast_exceptions.hpp"
 #include "utilities/msvc.hpp"