Content

Example 1: Scalar

The scalar type scalar<T> with template parameter T denoting the underlying CPU floating point type (float or double) represents a single floating point value on the GPU. scalar<T> is designed to behave much like a scalar type on the CPU, but library users have to keep in mind that every operation on scalar<T> requires to launch the appropriate compute kernel on the GPU and is thus much slower then the CPU equivalent.

Be aware that operations between objects of type scalar<T> (e.g.~additions. comparisons) have large overhead. For every operation, a separate compute kernel launch is required.

Handling scalars in ViennaCL

typedef float        ScalarType;
//typedef double    ScalarType; //use this if your GPU supports double precision
 
// Define a few CPU scalars:
ScalarType s1 = 3.1415926;
ScalarType s2 = 2.71763;
ScalarType s3 = 42.0;
  
// ViennaCL scalars are defined in the same way:
viennacl::scalar<ScalarType> vcl_s1;
viennacl::scalar<ScalarType> vcl_s2 = 1.0;
viennacl::scalar<ScalarType> vcl_s3 = 1.0;
 
// CPU scalars can be transparently assigned to GPU scalars and vice versa:
vcl_s1 = s1;
s2 = vcl_s2;
vcl_s3 = s3;
  
// Operations between GPU scalars work just as for CPU scalars (but are much slower!)
s1 += s2;
vcl_s1 += vcl_s2;
 
s1 = s2 + s3;
vcl_s1 = vcl_s2 + vcl_s3;
  
s1 = s2 + s3 * s2 - s3 / s1;
vcl_s1 = vcl_s2 + vcl_s3 * vcl_s2 - vcl_s3 / vcl_s1;
  
// Operations can also be mixed:
vcl_s1 = s1 * vcl_s2 + s3 - vcl_s3;
  
// Output stream is overloaded as well:
std::cout << "CPU scalar s2: " << s2 << std::endl;
std::cout << "GPU scalar vcl_s2: " << vcl_s2 << std::endl;