PTX Bit Bucket Registers

Question

... are only mentioned in the PTX manual. There is no hint of what they are good for or how to use them.

Does anyone know more? Did I just miss the general concept?

Answer 1

Bart's commentary is mostly right. In more detail, as indicated in the PTX ISA 3.1 manual ,

For some instructions, the destination operand is optional. The bit-bucket operand indicated by the underscore ( _ ) can be used instead of the destination register.

In fact, there is only one instruction class specified in the 3.1 PTX specification for which _ is a valid destination: atom . Here is the semantics of atom :

Atomically loads the original value at location a into destination register d, performs the reduction operation with operand b and the value at location a, and saves the result of the specified operation at location a, overwriting the original value.

And there is a note for atom :

Simple abbreviations can be specified using the destination operand of the bit-bucket _ .

So we can build an example:

 atom.global.add.s32 _, [a], 4 

This will add 4 to the integer value in memory a and will not return the previous value of location a in the register. Therefore, if you do not need the previous value, you can use this. I assume that the compiler will generate this for this code

 atomicAdd(&a, 4); 

since the return value of atomicAdd is not stored in the variable.