Data Types

Data Types#

MatX attempts to use the default C++ type system rules wherever possible; the inner types of operators should behave identically to normal C++ rules. For example, assigning a tensor of type double to a tensor of type float will have the same implications as assigning the simple data types to each other. For example:

auto t1 = make_tensor<float>();
auto t2 = make_tensor<double>();
(t1 = t2).run();

Follows the same rules as:

float t1;
double t2;
t1 = t2;

MatX tests the following types in unit tests:

  1. Integer * int8_t * uint8_t * int16_t * uint16_t * int32_t * uint32_t * int64_t * uint64_t

  2. Floating Point * matxFp16 (__half) * matxBf16 (__nv_bfloat16) * float * double

  3. Complex * matxfp16Complex * matxBf16Complex * cuda::std::complex<float> * cuda::std::complex<double>

Since MatX attempts to have parity for all functionality on both host and device, all types above are useable in both scenarios. While most types above are common C++ types, there are notable exceptions:

  • Native half precision types (__half/__nv_bfloat16) are swapped for matxFp16 and matxBf16. This is done because the native types do not provide the full set of operator overloads on both the host and device. The same concept applies to the complex versions matxfp16Complex and matxBf16Complex.

  • Complex float and double use the cuda::std versions rather than std:: since std::complex does not work in device code. libcudacxx is included with the CUDA toolkit.

User-defined types#

While the above types are tested, any type supporting the standard C++ operator overloading semantics should work, depending on the context used. For example, to sort a tensor the comparison and equality operators must be defined.