Are max(a, b) and min(a, b) divergent?

Accelerated Computing CUDA CUDA Programming and Performance
1

It a simple question, but one which I don’t see answered anywhere in the programming guide.

The standard CPU implementation seems to be:

(b<a) ? a : b;

which is clearly divergent, but I’d like to know if CUDA does anything clever to get around it.

Also when doing something like

a = max(a,0);

will the compiler reduce that to

a *= (a>0)

to prevent divergence (assuming that max is divergent in the first place)?

EDIT: I particularly care about when a and b are floats, but a more general answer may be helpful for others. I hope there is someone who knows!

2

It a simple question, but one which I don’t see answered anywhere in the programming guide.

The standard CPU implementation seems to be:

(b<a) ? a : b;

which is clearly divergent, but I’d like to know if CUDA does anything clever to get around it.

Also when doing something like

a = max(a,0);

will the compiler reduce that to

a *= (a>0)

to prevent divergence (assuming that max is divergent in the first place)?

EDIT: I particularly care about when a and b are floats, but a more general answer may be helpful for others. I hope there is someone who knows!

3

The “standard” C source-level idioms to compute min() and max() do not usually result in divergent code for most scalar types, since the compiler translates them into predicated execution or select instructions. You can use cuobdump to disassemble the machine code if you need to know for sure.

The min() / max() functionality is directly supported by hardware instructions for many scalar types (e.g. int, float, double) on both sm_1x and sm_2x devices. The various min/max instructions will be readily noticeable in disassembled machine code by their names.

4

The “standard” C source-level idioms to compute min() and max() do not usually result in divergent code for most scalar types, since the compiler translates them into predicated execution or select instructions. You can use cuobdump to disassemble the machine code if you need to know for sure.

The min() / max() functionality is directly supported by hardware instructions for many scalar types (e.g. int, float, double) on both sm_1x and sm_2x devices. The various min/max instructions will be readily noticeable in disassembled machine code by their names.

5

Thank you for the quick response!

6

Thank you for the quick response!

7

You are welcome. I notice belatedly that I should have been clearer in the second paragraph. By “min() / max() functionality” I was refering to the overloaded min() and max() functions that CUDA makes available in device code, as opposed to discrete source level constructs such as macros that accomplish the same thing. The functions are what gets translated into min/max instructions for common scalar types; the discrete constructs will typically map to predicated execution or select instructions for common scalar types.

The overall message is that programmers should not worry about divergence due to min/max computations. In general the direct use of built-in min() and max() functions will result in somewhat higher performance than the use of discrete equivalents, as the dynamic instruction count will be minimized.

8

You are welcome. I notice belatedly that I should have been clearer in the second paragraph. By “min() / max() functionality” I was refering to the overloaded min() and max() functions that CUDA makes available in device code, as opposed to discrete source level constructs such as macros that accomplish the same thing. The functions are what gets translated into min/max instructions for common scalar types; the discrete constructs will typically map to predicated execution or select instructions for common scalar types.

The overall message is that programmers should not worry about divergence due to min/max computations. In general the direct use of built-in min() and max() functions will result in somewhat higher performance than the use of discrete equivalents, as the dynamic instruction count will be minimized.