One kernel, three views

The same addition at each level of vector_add

TIRx-lite source

txl.ptx.add.f32(x, x, y)

vector_add

TIRx IR

T.ptx.add(
    buffer, buffer,
    buffer_1, ...)

vector_add.func,
read by the analyses

Generated CUDA

asm("add.f32 %0, %1, %2;"
    : "=f"(__d)
    : "f"(__a), "f"(__b));

vector_add.source()

Output abridged.

CUDA → ptxas → Machine code