math_brute_force: fix fdim to use device's rounding when converting result back to half.

[cycheng]

In the half-precision fdim test, the original code used CL_HALF_RTE to convert the float result back to half, causing a mismatch in computation results when the hardware uses RTZ. Some of the examples:

  fdim(0x365f, 0xdc63) = fdim( 0.398193f,  -280.75f)     =   281.148193f (RTE=0x5c65, RTZ=0x5c64)
  fdim(0xa4a3, 0xf0e9) = fdim(-0.018112f, 10056.0f)      = 10055.981445f (RTE=0x70e9, RTZ=0x70e8)
  fdim(0x1904, 0x9ab7) = fdim( 0.002449f,    -0.003279f) =     0.005728f (RTE=0x1dde, RTZ=0x1ddd)

Fixed this by using the hardware's default rounding mode when converting the result back to half.

[cycheng]

Thank you @svenvh, in the current change, we only set halfRoundingMode to CL_HALF_RTZ when both isFDim and gIsInRTZMode are true.

I can try to apply the change to all binary f16 operations and redo the CTS tests on my side.

[cycheng]

@svenvh
CC @AhmedAmraniAkdi , @paulfradgley

Change halfRoundingMode to CL_HALF_RTZ for all binary operations, I was seeing 4 test failures on our RTZ device:

• hypot

  • Case1:
    inputs = 0xfbe7, 0xfabb = -64736, -55136
    double r = std::hypot(-64736, -55136) = 85033.6875
    cl_half_from_float(r, CL_HALF_RTZ) = 0x7bff = 65504
    cl_half_from_float(r, CL_HALF_RTE) = 0x7c00 = inf
    ulp = -304.651367
  • Case2:
    inputs = 0xfbfe, 0x7387 = -65472, 15416
    double r = std::hypot(-65472, 15416) = 67262.4375
    cl_half_from_float(r, CL_HALF_RTZ) = 0x7bff = 65504
    cl_half_from_float(r, CL_HALF_RTE) = 0x7c00 = inf
    ulp = -26.975624

• ldexp

• Case 1:
  inputs = 0x77d5, 162
  fesetround(FE_TOWARDZERO)
  double r = std::ldexp(32080, 162) = 3.402823e+38
  cl_half_from_float(r, CL_HALF_RTZ) = 0x7bff = 65504
  cl_half_from_float(r, CL_HALF_RTE) = 0x7c00 = inf
  ulp = -2004.999877

• pow

  • Case 1:
    inputs = 0x3750, 0xcfc2
    double r = std::pow(0.45703125, -31.03125) = 3.567007e+10
    cl_half_from_float(r, CL_HALF_RTZ) = 0x7bff = 65504
    cl_half_from_float(r, CL_HALF_RTE) = 0x7c00 = inf
    ulp = -1063.048706

• powr

  • ..

[bashbaug]

LGTM

[bashbaug]

This has three approvals from three different companies, so no reason to wait to merge it.