ptx optimization

Accelerated Computing CUDA CUDA Programming and Performance
1

I was looking through some PTX code generated by nvcc (with /O2 optimization) and it seemed like there were lots of redundant instructions loading immediate values into registers. For instance, in the C/CUDA code I had a float3 with an overloaded multiply operator, like this…
item = item * 0.5

and the PTX code listing looked like this …
mov.f32 %f95, 0f3f000000; // 0.5
mul.f32 %f43, %f43, %f95; //
mov.f32 %f96, 0f3f000000; // 0.5
mul.f32 %f45, %f45, %f96; //
mov.f32 %f97, 0f3f000000; // 0.5
mul.f32 %f47, %f47, %f97; //

So, there are 3 separate instructions loading 0.5 into a register. Is there is a reason the code is generated like this? Would the PTX compiler do any further optimization?

Kalle

2

PTX is not the final cubin code the processor runs. PTX is an intermediate form before any optimization.

To look at your real code, you may try decuda.

3

try this code for comparison

volatile float pointfive = 0.5f;

item *= pointfive;

Christian

4

Thanks, that’s a useful utility. Looks like those were turned into multiply instructions with 0.5f as the immediate value.