Added support for testing (performance and correctness) against a CPU BLAS library

2 files changed, 13 insertions, 6 deletions

diff --git a/include/internal/clpp11.h b/include/internal/clpp11.h
index aac66396..00905ef7 100644
--- a/include/internal/clpp11.h
+++ b/include/internal/clpp11.h
@@ -465,31 +465,33 @@ class Buffer {
   }
 
   // Copies from device to host: reading the device buffer a-synchronously
-  void ReadAsync(const Queue &queue, const size_t size, T* host, const size_t offset = 0) {
+  void ReadAsync(const Queue &queue, const size_t size, T* host, const size_t offset = 0) const {
     if (access_ == BufferAccess::kWriteOnly) { Error("reading from a write-only buffer"); }
     CheckError(clEnqueueReadBuffer(queue(), *buffer_, CL_FALSE, offset*sizeof(T), size*sizeof(T),
                                    host, 0, nullptr, nullptr));
   }
   void ReadAsync(const Queue &queue, const size_t size, std::vector<T> &host,
-                 const size_t offset = 0) {
+                 const size_t offset = 0) const {
     if (host.size() < size) { Error("target host buffer is too small"); }
     ReadAsync(queue, size, host.data(), offset);
   }
   void ReadAsync(const Queue &queue, const size_t size, BufferHost<T> &host,
-                 const size_t offset = 0) {
+                 const size_t offset = 0) const {
     if (host.size() < size) { Error("target host buffer is too small"); }
     ReadAsync(queue, size, host.data(), offset);
   }
 
   // Copies from device to host: reading the device buffer
-  void Read(const Queue &queue, const size_t size, T* host, const size_t offset = 0) {
+  void Read(const Queue &queue, const size_t size, T* host, const size_t offset = 0) const {
     ReadAsync(queue, size, host, offset);
     queue.Finish();
   }
-  void Read(const Queue &queue, const size_t size, std::vector<T> &host, const size_t offset = 0) {
+  void Read(const Queue &queue, const size_t size, std::vector<T> &host,
+            const size_t offset = 0) const {
     Read(queue, size, host.data(), offset);
   }
-  void Read(const Queue &queue, const size_t size, BufferHost<T> &host, const size_t offset = 0) {
+  void Read(const Queue &queue, const size_t size, BufferHost<T> &host,
+            const size_t offset = 0) const {
     Read(queue, size, host.data(), offset);
   }
 
diff --git a/include/internal/utilities.h b/include/internal/utilities.h
index 35f76722..6adc1d0a 100644
--- a/include/internal/utilities.h
+++ b/include/internal/utilities.h
@@ -35,6 +35,9 @@ using double2 = std::complex<double>;
 const std::string kKhronosHalfPrecision = "cl_khr_fp16";
 const std::string kKhronosDoublePrecision = "cl_khr_fp64";
 
+// Catched an unknown error
+constexpr auto kUnknownError = -999;
+
 // =================================================================================================
 
 // The routine-specific arguments in string form
@@ -70,6 +73,7 @@ constexpr auto kArgFraction = "fraction";
 
 // The client-specific arguments in string form
 constexpr auto kArgCompareclblas = "clblas";
+constexpr auto kArgComparecblas = "cblas";
 constexpr auto kArgStepSize = "step";
 constexpr auto kArgNumSteps = "num_steps";
 constexpr auto kArgNumRuns = "runs";
@@ -128,6 +132,7 @@ struct Arguments {
   double fraction = 1.0;
   // Client-specific arguments
   int compare_clblas = 1;
+  int compare_cblas = 1;
   size_t step = 1;
   size_t num_steps = 0;
   size_t num_runs = 10;