How to avoid floating point round off error in unit tests?

Question

I'm trying to write unit tests for some simple vector math functions that operate on arrays of single precision floating point numbers. The functions use SSE intrinsics and I'm getting false positives (at least I think) when running the tests on a 32-bit system (the tests pass on 64-bit). As the operation runs through the array, I accumulate more and more round off error. Here is a snippet of unit test code and output (my actual question(s) follow):

Test Setup:

static const int N = 1024;
static const float MSCALAR = 42.42f;

static void setup(void) {
    input = _mm_malloc(sizeof(*input) * N, 16);
    ainput = _mm_malloc(sizeof(*ainput) * N, 16);
    output = _mm_malloc(sizeof(*output) * N, 16);
    expected = _mm_malloc(sizeof(*expected) * N, 16);

    memset(output, 0, sizeof(*output) * N);

    for (int i = 0; i < N; i++) {
        input[i] = i * 0.4f;
        ainput[i] = i * 2.1f;
        expected[i] = (input[i] * MSCALAR) + ainput[i];
    }
}

My main test code then calls the function to be tested (which does the same calculation used to generate the expected array) and checks its output against the expected array generated above. The check is for closeness (within 0.0001) not equality.

Sample output:

0.000000    0.000000    delta: 0.000000
44.419998   44.419998   delta: 0.000000
...snip 100 or so lines...
2043.319946 2043.319946 delta: 0.000000
2087.739746 2087.739990 delta: 0.000244
...snip 100 or so lines...
4086.639893 4086.639893 delta: 0.000000
4131.059570 4131.060059 delta: 0.000488
4175.479492 4175.479980 delta: 0.000488
...etc, etc...

I know I have two problems:

1. On 32-bit machines, differences between 387 and SSE floating point arithmetic units. I believe 387 uses more bits for intermediate values.
2. Non-exact representation of my 42.42 value that I'm using to generate expected values.

So my question is, what is the proper way to write meaningful and portable unit tests for math operations on floating point data?

*By portable I mean should pass on both 32 and 64 bit architectures.

Answer 1

Per a comment, we see that the function being tested is essentially:

for (int i = 0; i < N; ++i)
    D[i] = A[i] * b + C[i];

where A[i], b, C[i], and D[i] all have type float. When referring to the data of a single iteration, I will use a, c, and d for A[i], C[i], and D[i].

Below is an analysis of what we could use for an error tolerance when testing this function. First, though, I want to point out that we can design the test so that there is no error. We can choose the values of A[i], b, C[i], and D[i] so that all the results, both final and intermediate results, are exactly representable and there is no rounding error. Obviously, this will not test the floating-point arithmetic, but that is not the goal. The goal is to test the code of the function: Does it execute instructions that compute the desired function? Simply choosing values that would reveal any failures to use the right data, to add, to multiply, or to store to the right location will suffice to reveal bugs in the function. We trust that the hardware performs floating-point correctly and are not testing that; we just want to test that the function was written correctly. To accomplish this, we could, for example, set b to a power of two, A[i] to various small integers, and C[i] to various small integers multiplied by b. I could detail limits on these values more precisely if desired. Then all results would be exact, and any need to allow for a tolerance in comparison would vanish.

That aside, let us proceed to error analysis.

The goal is to find bugs in the implementation of the function. To do this, we can ignore small errors in the floating-point arithmetic, because the kinds of bugs we are seeking almost always cause large errors: The wrong operation is used, the wrong data is used, or the result is not stored in the desired location, so the actual result is almost always very different from the expected result.

Now the question is how much error should we tolerate? Because bugs will generally cause large errors, we can set the tolerance quite high. However, in floating-point, “high” is still relative; an error of one million is small compared to values in the trillions, but it is too high to discover errors when the input values are in the ones. So we ought to do at least some analysis to decide the level.

The function being tested will use SSE intrinsics. This means it will, for each i in the loop above, either perform a floating-point multiply and a floating-point add or will perform a fused floating-point multiply-add. The potential errors in the latter are a subset of the former, so I will use the former. The floating-point operations for a*b+c do some rounding so that they calculate a result that is approximately a•b+c (interpreted as an exact mathematical expression, not floating-point). We can write the exact value calculated as (a•b•(1+e0)+c)•(1+e1) for some errors e0 and e1 with magnitudes at most 2^-24, provided all the values are in the normal range of the floating-point format. (2^-24 is the maximum relative error that can occur in any correctly rounded elementary floating-point operation in round-to-nearest mode in the IEEE-754 32-bit binary floating-point format. Rounding in round-to-nearest mode changes the mathematical value by at most half the value of the least significant bit in the significand, which is 23 bits below the most significant bit.)

Next, we consider what value the test program produces for its expected value. It uses the C code d = a*b + c;. (I have converted the long names in the question to shorter names.) Ideally, this would also calculate a multiply and an add in IEEE-754 32-bit binary floating-point. If it did, then the result would be identical to the function being tested, and there would be no need to allow for any tolerance in comparison. However, the C standard allows implementations some flexibility in performing floating-point arithmetic, and there are non-conforming implementations that take more liberties than the standard allows.

A common behavior is for an expression to be computed with more precision than its nominal type. Some compilers may calculate a*b + c using double or long double arithmetic. The C standard requires that results be converted to the nominal type in casts or assignments; extra precision must be discarded. If the C implementation is using extra precision, then the calculation proceeds: a*b is calculated with extra precision, yielding exactly a•b, because double and long double have enough precision to exactly represent the product of any two float values. A C implementation might then round this result to float. This is unlikely, but I allow for it anyway. However, I also dismiss it because it moves the expected result to be closer to the result of the function being tested, and we just need to know the maximum error that can occur. So I will continue, with the worse (more distant) case, that the result so far is a•b. Then c is added, yielding (a•b+c)•(1+e2) for some e2 with magnitude at most 2^-53 (the maximum relative error of normal numbers in the 64-bit binary format). Finally, this value is converted to float for assignment to d, yielding (a•b+c)•(1+e2)•(1+e3) for some e3 with magnitude at most 2^-24.

Now we have expressions for the exact result computed by a correctly operating function, (a•b•(1+e0)+c)•(1+e1), and for the exact result computed by the test code, (a•b+c)•(1+e2)•(1+e3), and we can calculate a bound on how much they can differ. Simple algebra tells us the exact difference is a•b•(e0+e1+e0•e1-e2-e3-e2•e3)+c•(e1-e2-e3-e2•e3). This is a simple function of e0, e1, e2, and e3, and we can see its extremes occur at endpoints of the potential values for e0, e1, e2, and e3. There are some complications due to interactions between possibilities for the signs of the values, but we can simply allow some extra error for the worst case. A bound on the maximum magnitude of the difference is |a•b|•(3•2^-24+2^-53+2^-48)+|c|•(2•2^-24+2^-53+2^-77).

Because we have plenty of room, we can simplify that, as long as we do it in the direction of making the values larger. E.g., it might be convenient to use |a•b|•3.001•2^-24+|c|•2.001•2^-24. This expression should suffice to allow for rounding in floating-point calculations while detecting nearly all implementation errors.

Note that the expression is not proportional to the final value, a*b+c, as calculated either by the function being tested or by the test program. This means that, in general, tests using a tolerance relative to the final values calculated by the function being tested or by the test program are wrong. The proper form of a test should be something like this:

double tolerance = fabs(input[i] * MSCALAR) * 0x3.001p-24 + fabs(ainput[i]) * 0x2.001p-24;
double difference = fabs(output[i] - expected[i]);
if (! (difference < tolerance))
   // Report error here.

In summary, this gives us a tolerance that is larger than any possible differences due to floating-point rounding, so it should never give us a false positive (report the test function is broken when it is not). However, it is very small compared to the errors caused by the bugs we want to detect, so it should rarely give us a false negative (fail to report an actual bug).

(Note that there are also rounding errors computing the tolerance, but they are smaller than the slop I have allowed for in using .001 in the coefficients, so we can ignore them.)

(Also note that ! (difference < tolerance) is not equivalent to difference >= tolerance. If the function produces a NaN, due to a bug, any comparison yields false: both difference < tolerance and difference >= tolerance yield false, but ! (difference < tolerance) yields true.)

Answer 2

On 32-bit machines, differences between 387 and SSE floating point arithmetic units. I believe 387 uses more bits for intermediate values.

If you are using GCC as 32-bit compiler, you can tell it to generate SSE2 code still with options -msse2 -mfpmath=sse. Clang can be told to do the same thing with one of the two options and ignores the other one (I forget which). In both cases the binary program should implement strict IEEE 754 semantics, and compute the same result as a 64-bit program that also uses SSE2 instructions to implement strict IEEE 754 semantics.

Non-exact representation of my 42.42 value that I'm using to generate expected values.

The C standard says that a literal such as 42.42f must be converted to either the floating-point number immediately above or immediately below the number represented in decimal. Moreover, if the literal is representable exactly as a floating-point number of the intended format, then this value must be used. However, a quality compiler (such as GCC) will give you(*) the nearest representable floating-point number, of which there is only one, so again, this is not a real portability issue as long as you are using a quality compiler (or at the very least, the same compiler).

Should this turn out to be a problem, a solution is to write an exact representation of the constants you intend. Such an exact representation can be very long in decimal format (up to 750 decimal digits for the exact representation of a double) but is always quite compact in C99's hexadecimal format: 0x1.535c28p+5 for the exact representation of the float nearest to 42.42. A recent version of the static analysis platform for C programs Frama-C can provide the hexadecimal representation of all inexact decimal floating-point constants with option -warn-decimal-float:all.

---

(*) barring a few conversion bugs in older GCC versions. See Rick Regan's blog for details.