Ieee arithmetic
We will present some of the trade-offs for computation of IEEE arithmetic for REAL_64 and REAL_32 as implemented in EiffelStudio when enabling the Total Order on REALs option where NaN is not an unordered value but a value less than all the other values (note that in some other frameworks, we have seen it defined as the largest value). In other words:
- NaN = NaN yields True
- NaN < x for all x but NaN
To best show the trade-offs we will start by showing some benchmark results.
Note: This is an EiffelStudio specific extension to potentially help the ECMA standard committee in making a reasonable decision on the semantic of NaN.
Contents
- 1 Benchmarks
- 1.1 The code
- 1.2 The configuration
- 1.3 Equality Testing
- 1.3.1 Testing for non-NaN values which are always different
- 1.3.2 Testing for non-NaN values which are always the same
- 1.3.3 Testing for 100% NaN values
- 1.3.4 Testing for 50% NaN values and 50% non-NaN values
- 1.3.5 Testing for 25% NaN values and 75% non-NaN values
- 1.3.6 Testing for 10% NaN values and 90% non-NaN values
- 1.4 Less than Testing
- 1.4.1 Testing for non-NaN values which are always different
- 1.4.2 Testing for non-NaN values which are always the same
- 1.4.3 Testing for 100% NaN values
- 1.4.4 Testing for 50% NaN values and 50% non-NaN values
- 1.4.5 Testing for 25% NaN values and 75% non-NaN values
- 1.4.6 Testing for 10% NaN values and 90% non-NaN values
- 1.5 IsNaN Testing
- 1.5.1 Testing for non-NaN values which are always different
- 1.5.2 Testing for non-NaN values which are always the same
- 1.5.3 Testing for 100% NaN values
- 1.5.4 Testing for 50% NaN values and 50% non-NaN values
- 1.5.5 Testing for 25% NaN values and 75% non-NaN values
- 1.5.6 Testing for 10% NaN values and 90% non-NaN values
Benchmarks
The code
The code below defines an equality function as well as a comparison function. The test is divided in two parts, first the initialization and then the computation.
static EIF_NATURAL_64 to_raw_bits (EIF_REAL_64 d) { return *((EIF_NATURAL_64 *)&d); } static int eif_is_nan_bits (EIF_NATURAL_64 value) { /* Clear the sign mark. */ EIF_NATURAL_64 jvalue = (value & ~RTU64C(0x8000000000000000)); /* Ensure that it starts with 0x7ff and that the mantissa is not 0. */ return (jvalue > RTU64C(0x7ff0000000000000)); } static int eif_is_nan (EIF_REAL_64 v) { EIF_NATURAL_64 value = *((EIF_NATURAL_64 *)&v); value &= ~RTU64C(0x8000000000000000); return (value > RTU64C(0x7ff0000000000000)); } static int eif_is_nan_real_64 (EIF_REAL_64 v) { #ifdef NAN1 return v != v; #elif defined(NAN2) EIF_NATURAL_64 value = *((EIF_NATURAL_64 *)&v); value &= ~RTU64C(0x8000000000000000); return (value > RTU64C(0x7ff0000000000000)); #elif defined(NAN3) EIF_REAL_64 *l_v = &v; return ((*((EIF_NATURAL_64 *)(l_v)) & RTU64C(0x7FF0000000000000))==RTU64C(0x7FF0000000000000)) && (*((EIF_NATURAL_64 *)(l_v)) & RTU64C(0x000FFFFFFFFFFFFF)); #elif defined(NAN4) #ifdef _WIN32 return _isnan(v); #else return isnan(v); #endif #endif static int eif_equal_real_64 (EIF_REAL_64 d1, EIF_REAL_64 d2) { #ifdef METH1 /* Here the base comparison is IEEE arithmetic. */ return (d1 == d2); #elif defined(METH2) /* Conversion to perform comparison on the binary representation. */ EIF_NATURAL_64 f1 = to_raw_bits(d1); EIF_NATURAL_64 f2 = to_raw_bits(d2); return (f1 == f2 ? 1 : (eif_is_nan_bits (f1) && eif_is_nan_bits(f2))); #elif defined(METH3) /* Use IEEE arithmetic to compare and find out if we have NaNs. */ return (d1 == d2 ? 1 : ((d1 != d1) && (d2 != d2))); #elif defined (METH4) /* Pessimist case, we assume that we compare mostly NaNs. */ return (d1 == d1 ? d1 == d2 : d2 != d2); #elif defined(METH5) /* Use IEEE arithmetic to compare but use binary representation to * find out if we have NaNs. */ return (d1 == d2 ? 1 : (eif_is_nan (d1) && eif_is_nan(d2))); #endif } static int eif_is_less_real_64 (EIF_REAL_64 d1, EIF_REAL_64 d2) { #ifdef METH1 /* Here the base comparison is IEEE arithmetic. */ return d1 < d2; #elif defined(METH2) /* Use IEEE arithmetic to compare but use binary representation to * find out if we have NaNs. */ return (d1 < d2 ? 1 : eif_is_nan(d1) && !eif_is_nan(d2)); #elif defined(METH3) /* Use IEEE arithmetic to compare and find out if we have NaNs. */ return (d1 < d2 ? 1 : (d1 != d1) && (d2 == d2)); #elif defined(METH4) /* Pessimist case, we assume that we compare mostly NaNs. */ return (d1 == d1 ? d1 < d2 : d2 == d2); #elif defined(METH5) /* Variation on METH3 using a different order for comparison. */ return (eif_is_nan(d1) ? !eif_is_nan(d2) : d1 < d2); #endif } #define ARR_SIZE 100000 int main(void) { EIF_NATURAL_64 res, i; EIF_REAL_64 *d = (EIF_REAL_64 *) malloc (sizeof(EIF_REAL_64) * ARR_SIZE + 1); /* Initialization of `d'. */ ... for (i = 0; i <= 0x3FFFFFFF; i++) { /* Substitute comparison_function with what needs to be tested. */ res = res + comparison_function (d [i % ARR_SIZE], d[(i - 1) % ARR_SIZE]); } printf ("%d\n", res); }
The configuration
- On Windows XP 64-bit with a Intel Q9450 @ 3 GHz with VC++ 2005 using the following command line:
/O2 /GL /FD /MT /GS-
- On Linux Ubuntu 9.04 x64 with Intel Xeon E5420 @ 2.5 GHz with gcc 4.4.1 using the following command line:
-O3 -funroll-loops -lm
- On Solaris 10 x64 with AMD Opteron 248 @ 2.2 GHz with Sun C 5.9 using the following command line:
-xO5 -m64 -lm
Equality Testing
Testing for non-NaN values which are always different
The array is filled with non-NaN values which are always different, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 5.068s | 6.25s | 9.32s |
METH2 | 6.495s | 6.47s | 11.57s |
METH3 | 6.448s | 7.35s | 12.31s |
METH4 | 6.424s | 8.80s | 12.31s |
METH5 | 6.832s | 7.73s | 11.46s |
Testing for non-NaN values which are always the same
The array is filled with non-NaN values which are always the same, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 5.242s | 6.24s | 9.32s |
METH2 | 5.464s | 5.17s | 9.62s |
METH3 | 5.308s | 5.48s | 9.26s |
METH4 | 6.428s | 8.80s | 12.31s |
METH5 | 5.384s | 5.49s | 9.32s |
Testing for 100% NaN values
The array is filled with NaN values, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 4.777s | 6.24s | 9.25s |
METH2 | 5.440s | 5.17s | 9.92s |
METH3 | 6.266s | 7.32s | 16.48s |
METH4 | 5.560s | 8.80s | 12.44s |
METH5 | 7.413s | 8.34s | 15.74s |
Testing for 50% NaN values and 50% non-NaN values
The array is filled at 50% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 6.889s | 6.24s | 9.31s |
METH2 | 6.914s | 7.01s | 13.36s |
METH3 | 6.405s | 7.01s | 14.64s |
METH4 | 6.068s | 8.80s | 12.07s |
METH5 | 6.880s | 7.83s | 13.47s |
Testing for 25% NaN values and 75% non-NaN values
The array is filled at 25% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 7.553s | 6.24s | 9.31s |
METH2 | 6.672s | 6.73s | 12.25s |
METH3 | 8.997s | 7.23s | 13.47s |
METH4 | 6.250s | 8.80s | 12.00s |
METH5 | 8.063s | 7.76s | 12.55s |
Testing for 10% NaN values and 90% non-NaN values
The array is filled at 10% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 5.029s | 6.24s | 9.32s |
METH2 | 6.556s | 6.59s | 11.74s |
METH3 | 6.437s | 7.29s | 12.82s |
METH4 | 6.327s | 8.80s | 12.06s |
METH5 | 6.800s | 7.73s | 12.22s |
Less than Testing
Testing for non-NaN values which are always different
The array is filled with non-NaN values which are always different, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 4.749s | 4.42s | 8.60s |
METH2 | 6.358s | 7.30s | 10.85s |
METH3 | 6.109s | 6.82s | 11.45s |
METH4 | 6.112s | 6.83s | 10.48s |
METH5 | 6.355s | 8.56s | 10.90s |
Testing for non-NaN values which are always the same
The array is filled with non-NaN values which are always the same, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 4.757s | 4.40s | 8.50s |
METH2 | 4.736s | 7.28s | 10.84s |
METH3 | 6.114s | 6.80s | 11.21s |
METH4 | 6.102s | 7.10s | 10.48s |
METH5 | 6.344s | 8.57s | 10.72s |
Testing for 100% NaN values
The array is filled with NaN values, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 4.753s | 4.41s | 8.60s |
METH2 | 7.366s | 8.34s | 12.13s |
METH3 | 6.284s | 7.31s | 15.18s |
METH4 | 5.570s | 6.82s | 14.08s |
METH5 | 6.993s | 8.56s | 11.03s |
Testing for 50% NaN values and 50% non-NaN values
The array is filled at 50% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 4.745s | 4.41s | 8.50s |
METH2 | 6.949s | 7.82s | 11.57s |
METH3 | 6.346s | 7.09s | 13.10s |
METH4 | 6.023s | 7.11s | 11.88s |
METH5 | 6.838s | 8.56s | 10.72s |
Testing for 25% NaN values and 75% non-NaN values
The array is filled at 25% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 4.812s | 4.41s | 8.50s |
METH2 | 6.633s | 7.54s | 11.20s |
METH3 | 7.324s | 6.96s | 12.07s |
METH4 | 6.050s | 7.10s | 11.04s |
METH5 | 6.595s | 8.55s | 11.03s |
Testing for 10% NaN values and 90% non-NaN values
The array is filled at 10% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
METH1 | 4.761s | 4.41s | 8.50s |
METH2 | 6.472s | 7.37s | 10.99s |
METH3 | 6.159s | 6.83s | 11.59s |
METH4 | 6.084s | 7.10s | 10.76s |
METH5 | 6.440s | 8.56s | 10.71s |
IsNaN Testing
We are testing eif_is_nan_real_64 which has 4 different implementations: NAN1, NAN2, NAN3 and NAN4. As we can see in the results below, it is always best to check for NaN using the `x != x' yielding True pattern.
Testing for non-NaN values which are always different
The array is filled with non-NaN values which are always different, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
NAN1 | 3.671s | 3.81s | 6.56s |
NAN2 | 4.071s | 3.98s | 6.11s |
NAN3 | 4.514s | 4.12s | 6.38s |
NAN4 | 4.881s | 8.28s | 10.96s |
Testing for non-NaN values which are always the same
The array is filled with non-NaN values which are always the same, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
NAN1 | 3.665s | 3.73s | 6.48s |
NAN2 | 4.071s | 4.17s | 6.12s |
NAN3 | 4.517s | 4.12s | 6.18s |
NAN4 | 4.886s | 8.28s | 10.71s |
Testing for 100% NaN values
The array is filled with NaN values, which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
NAN1 | 3.362s | 3.76s | 6.55s |
NAN2 | 4.071s | 3.99s | 6.18s |
NAN3 | 5.176s | 5.66s | 8.82s |
NAN4 | 4.880s | 8.28s | 10.96s |
Testing for 50% NaN values and 50% non-NaN values
The array is filled at 50% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
NAN1 | 3.527s | 3.74s | 6.48s |
NAN2 | 4.072s | 3.99s | 6.09s |
NAN3 | 5.011s | 4.87s | 7.78s |
NAN4 | 4.894s | 8.28s | 10.71s |
Testing for 25% NaN values and 75% non-NaN values
The array is filled at 25% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
NAN1 | 3.568s | 3.75s | 6.48s |
NAN2 | 4.073s | 3.99s | 6.09s |
NAN3 | 4.752s | 4.46s | 6.74s |
NAN4 | 4.880s | 8.27s | 10.83s |
Testing for 10% NaN values and 90% non-NaN values
The array is filled at 10% with NaN values and the rest with different values which gives the following results:
Method used | VC++ 2005 x64 | gcc 4.4.1 x64 | Sun cc 5.9 x64 |
NAN1 | 3.633s | 3.77s | 6.47s |
NAN2 | 4.084s | 3.98s | 6.09s |
NAN3 | 4.621s | 4.29s | 7.20s |
NAN4 | 4.879s | 8.29s | 10.89s |