FPU ops issues about rev HOT 9 CLOSED

Issue was confirmed and reproduced. Fix implemented and this test should be working in the latest devel branch

from rev.

We've seen it fixed on e285dff, but on latest devel it is still broken (FPU issues in general, not this particular test case).

from rev.

Can confirm we are still seeing that... root caused to an issue with FP MV, fix in place and will be included in devel shortly

from rev.

Additional FP op updates pushed to devel - test should be working

from rev.

This particular test case is fixed on (4414659), but we stil see FPU issues.
Example:

int main() {
        assert(fabs(-1) >= 1);
        assert(fabs(-1) < 1.01);
        return 0;
}

And something less trivial:

#include <cmath>
#define assert(x) if (!(x)) { asm(".byte 0x00"); asm(".byte 0x00"); asm(".byte 0x00"); asm(".byte 0x00"); }

template <typename T> void testPI()
{
        T pi = 0;
        T n = 1;

        for(int i = 0; i < 10000; i++, n += 2) {
                T tmp = 4.0 / n;
                if (i % 2)
                        pi -= tmp;
                else pi += tmp;
        }
        assert(pi > 3.1414);
        assert(pi < 3.1417);
}

int main() {
        testPI<float>();
}

Sqrt is a dedicated instruction, so we need a test for that as well:

#include <cmath>
#define assert(x) if (!(x)) { asm(".byte 0x00"); asm(".byte 0x00"); asm(".byte 0x00"); asm(".byte 0x00"); }

int main() {
        float fx, fy;
        double dx, dy;

        fx = 25;
        fy = sqrt(fx);
        assert(fy > 4.99);
        assert(fy < 5.01);

        dx = 25;
        dy = sqrt(dx);
        assert(dy > 4.99);
        assert(dy < 5.01);

        fx = 8;
        fy = sqrt(fx);
        assert(fy > 2.8);
        assert(fy < 2.9);

        dx = 8;
        dy = sqrt(dx);
        assert(dy > 2.8);
        assert(dy < 2.9);
        return 0;
}

from rev.

Tested on (581618e), out of the four test cases posted earlier, the last two are failing at this point.

from rev.

Seems that we have a regression again on devel branch and tests which were passing are now broken.
Example:

volatile float a = 0.199;
volatile float b = 0.2;
assert(a < b);
assert(a <= b);
assert(b > a);
assert(b >= a);
a = 0.2;
assert(a == b);

from rev.

Revisiting this issue... The test above appears to be working properly, however, the testPi() function appears to still fail. I believe the root cause of this is a double to single precision conversion that Rev is performing incorrectly - will address and have a fix ASAP

from rev.

This bug has become a container for multiple things. I believe it is not a good idea to track it this way. Original issue has been fixed so I am closing this bug and we'll create simpler self-contained issues with more specific problems.

from rev.