7. Performance Considerations

GNAT for CUDA relies on the CUDA toolchain backend technology and provides a very similar compilation model. There’s no fundamental reason why the performance and resource consumption of a program in Ada would be essentially different than that of the same program in C. However, you may need to specify some non-default switches to get the best performance and your usage of some Ada features may negatively affect performance.

7.1. Compiler Switches

You should consider using the following switches:

  • -gnatp removes checks that are added by the Ada language by default. Removing those checks makes the application run faster.

  • we recommend at least the -O2 optimization level

  • -gnatn enables inlining.

7.2. Unconstrained Arrays

In Ada, unconstrained arrays are associated with two bounds per dimension, which are not known statically. This differs from their C equivalent where the index always starts at zero and which have known bounds. The consequence is that computing the offset of an Ada array normally requires a memory read. E.g.:

type Arr is array (Integer range <>) of Integer;
type Arr_Access is access all Arr;
V : Arr_Access := new Arr (5 .. 10);
begin
   V (6) := 0;

When computing the memory location of item 6 in the array, the code must first load the lower bound (5) from memory and then subtract this lower bound from the index to compute the memory offset from the start of the array. If you do this in a loop over the array, the compiler can move this operation out of the loop, making it considerably less expensive, but if you do only a single array operation, the computation of the starting position is significant.

To improve this performance, an Ada extension is available when compiling with -gnatX0. This extension indicates that the lower bound of the array is known at compile-time. If this lower bound is a constant, it will remove both the need for loading that lower bound and doing a subtraction when computing the offset. E.g.:

type Arr is array (Integer range 0 .. <>) of Integer;
type Arr_Access is access all Arr;
V : Arr_Access := new Arr (0 .. 5);
begin
   V (1) := 0;

The feature is fully described here.