ization on a simple OpenGL fragment shader
The NVIDIA driver has a miscompilation on a simple fragment shader program. The program is as follows:
uniform int t;
layout(location = 0) out vec4 fragColor;
void main() {
uint c = 0u;
for (int i = -1; i > -2; i -= t) {
c = uint(1);
}
fragColor = vec4(float(c));
}
During runtime, the uniform input variable t is set to 1. So the loop will be executed once, and variable c will receive the value of 1, causing the fragColor to be set to vec4(1.0) (white color). However, the NVIDIA driver incorrectly optimizes the loop, causing the fragment shader to return vec4(0.0) (black color) instead.
Further investigation shows that any of the following modifications will make the bug disappear:
- Remove the explicit initialization of
c(i.e.,uint c;). - Change the loop condition to
i = 0; i < 1; i += t. - Change the loop condition to
i = -1; i >= -2; i -= 1, i.e., substitutetwith a constant value.
Steps to reproduce:
I prepare a minimal OpenGL application that loads the fragment shader to draw a 256x256 white square. The application source code and the fragment shader that triggers the error are given here. The package also contains a detailed README file describing system information and steps to run the code. It should take no more than a minute to reproduce the issue.