Question

How well does NVCC optimize device code? Does it do any sort of optimizations like constant folding and common subexpression elimination?

E.g, will it reduce the following:

float a = 1 / sqrtf(2 * M_PI);
float b = c / sqrtf(2 * M_PI);

to this:

float sqrt_2pi = sqrtf(2 * M_PI); // Compile time constant
float a = 1 / sqrt_2pi;
float b = c / sqrt_2pi;

What about more clever optimizations, involving knowing semantics of math functions:

float a = 1 / sqrtf(c * M_PI);
float b = c / sqrtf(M_PI);

to this:

float sqrt_pi = sqrtf(M_PI); // Compile time constant
float a = 1 / (sqrt_pi * sqrtf(c));
float b = c / sqrt_pi;

Answer 1

The compiler is way ahead of you. In your example:

float a = 1 / sqrtf(2 * M_PI);
float b = c / sqrtf(2 * M_PI);

nvopencc (Open64) will emit this:

    mov.f32         %f2, 0f40206c99;        // 2.50663
    div.full.f32    %f3, %f1, %f2;
    mov.f32         %f4, 0f3ecc422a;        // 0.398942

which is equivalent to

float b = c / 2.50663f;
float a = 0.398942f;

---

The second case gets compiled to this:

float a = 1 / sqrtf(c * 3.14159f); // 0f40490fdb
float b = c / 1.77245f; // 0f3fe2dfc5

I am guessing the expression for a generated by the compiler should be more accurate than your “optmized” version, but about the same speed.