change reference_relaxed_exp to use float type to calculate reference #2160
Conversation
|
|
anilavakundu007
force-pushed
the
main
branch
from
3c470fd to
71e418c
Compare
| double reference_relaxed_exp(double x) { return reference_exp(x); } | ||
| double reference_relaxed_exp(double x) | ||
| { | ||
| return reference_exp2(((float)x) * HEX_FLT(+, 1, 715476, +, 0)); |
There was a problem hiding this comment.
Hi, could you please explain why this test should use a float to compute the reference value in this case?
In case it helps other reviewers, HEX_FLT(+, 1, 715476, +, 0) expands to + 0x1.715476p+0, which is equal to 1.44269f, or M_LOG2E_F.
There was a problem hiding this comment.
the ulp error for exp according to the specs is ≤ 3 ulp for the full profile and ≤ 4 ulp for embedded profile. However, with -cl-unsafe-math-optimizations flag turned on the ulp requirements become ≤ 3 + floor(fabs(2 * x)) ulp for the full profile, and ≤ 4 ulp for the embedded profile.
As the ulp error is a bit relaxed with the unsafe math operations is there a reason for the same reference values being used for both the relaxed and normal implementation of the exp built-in?
There was a problem hiding this comment.
Altering the reference function doesn't feel like the right solution. unary_float.cpp already has logic to select the ULP threshold for exp and exp2 in relaxed mode:
if (strcmp(fname, "exp") == 0 || strcmp(fname, "exp2") == 0)
{
// For full profile, ULP depends on input value.
// For embedded profile, ULP comes from functionList.
if (!gIsEmbedded)
{
ulps = 3.0f + floor(fabs(2 * s[j]));
}
fail = !(fabsf(err) <= ulps);
}
Are you able to share any details on the values that are failing for you?
There was a problem hiding this comment.
There are a few failures but one example is the following
ERROR: exp: -4.058329 ulp error at 0x1.6005p+2 (0x40b00280): *0x1.e98882p+7 vs. 0x1.e9887ap+7
Our current understanding is that the use of the double input values makes the result of the reference calculation a bit 'overly' precise. There is a number of other reference_relaxed functions in this file where the input is truncated to float
use lower precision for reference values used for the relaxed implementation of exp math built-in.