Merge branch 'CNugteren:master' into master

Author: EJainDev

CHANGELOG

@@ -4,6 +4,7 @@ Development version (next version)
 - Added tuned parameters for many devices (see doc/tuning.md)
 - Enabled parallel kernel compilation for faster kernel tuning (see doc/tuning.md)
 - Removed all generator scripts for the API
+- Added 3 new error codes for vector Z in routine XHAD
 
 Version 1.6.3
 - Fixed a bug in the GEMMK=1 kernel (with 2D register tiling) when MWG!=NWG

include/clblast.h

@@ -123,6 +123,9 @@ enum class StatusCode {
   kDatabaseError = -2041,             // Entry for the device was not found in the database
   kUnknownError = -2040,              // A catch-all error code representing an unspecified error
   kUnexpectedError = -2039,           // A catch-all error code representing an unexpected exception
+  kInvalidVectorZ = -2038,            // Vector Z is not a valid OpenCL buffer
+  kInvalidIncrementZ = -2037,         // Increment of vector Z cannot be zero
+  kInsufficientMemoryZ = -2036,       // Vector Z's OpenCL buffer is too small
 };
 
 // Matrix layout and transpose types

include/clblast_c.h

@@ -122,6 +122,9 @@ typedef enum CLBlastStatusCode_ {
   CLBlastDatabaseError = -2041,             // Entry for the device was not found in the database
   CLBlastUnknownError = -2040,              // A catch-all error code representing an unspecified error
   CLBlastUnexpectedError = -2039,           // A catch-all error code representing an unexpected exception
+  CLBlastInvalidVectorZ = -2038,            // Vector Z is not a valid OpenCL buffer
+  CLBlastInvalidIncrementZ = -2037,         // Increment of vector Z cannot be zero
+  CLBlastInsufficientMemoryZ = -2036,       // Vector Z's OpenCL buffer is too small
 } CLBlastStatusCode;
 
 // Matrix layout and transpose types

include/clblast_cuda.h

@@ -77,6 +77,9 @@ enum class StatusCode {
   kDatabaseError = -2041,             // Entry for the device was not found in the database
   kUnknownError = -2040,              // A catch-all error code representing an unspecified error
   kUnexpectedError = -2039,           // A catch-all error code representing an unexpected exception
+  kInvalidVectorZ = -2038,            // Vector Z is not a valid OpenCL buffer
+  kInvalidIncrementZ = -2037,         // Increment of vector Z cannot be zero
+  kInsufficientMemoryZ = -2036,       // Vector Z's OpenCL buffer is too small
 };
 
 // Matrix layout and transpose types

src/routines/levelx/xhad.cpp

@@ -41,7 +41,7 @@ void Xhad<T>::DoHad(const size_t n, const T alpha, const Buffer<T>& x_buffer, co
   // Tests the vectors for validity
   TestVectorX(n, x_buffer, x_offset, x_inc);
   TestVectorY(n, y_buffer, y_offset, y_inc);
-  TestVectorY(n, z_buffer, z_offset, z_inc);  // TODO: Make a TestVectorZ function with error codes
+  TestVectorZ(n, z_buffer, z_offset, z_inc);
 
   // Determines whether or not the fast-version can be used
   const auto use_faster_kernel = (x_offset == 0) && (x_inc == 1) && (y_offset == 0) && (y_inc == 1) &&

src/utilities/buffer_test.hpp

@@ -116,6 +116,22 @@ void TestVectorY(const size_t n, const Buffer<T>& buffer, const size_t offset, c
   }
 }
 
+// Tests vector 'Z' for validity
+template <typename T>
+void TestVectorZ(const size_t n, const Buffer<T>& buffer, const size_t offset, const size_t inc) {
+  if (inc == 0) {
+    throw BLASError(StatusCode::kInvalidIncrementZ);
+  }
+  try {
+    const auto required_size = ((n - 1) * inc + 1 + offset) * sizeof(T);
+    if (buffer.GetSize() < required_size) {
+      throw BLASError(StatusCode::kInsufficientMemoryZ);
+    }
+  } catch (const Error<std::runtime_error>& e) {
+    throw BLASError(StatusCode::kInvalidVectorZ, e.what());
+  }
+}
+
 // =================================================================================================
 
 // Tests vector 'scalar' for validity