Better half
This MR adds an option to compile with CPU: CPU_USE_REAL_HALF
. Toggling it makes the CPU build switch between uint16_t
and float
for its backend datatype for half_t
.
In addition, the UT event model has been cleaned up. ut pre decode
now uses dev_ut_pre_decoded_hits_t
for its output datatype. This allows making UT::Hits_t
a single purpose type, and enables UT::Consolidated::Hits_t
to inherit from UT::Hits_t
, simplifying UT datatypes.
- The intrinsics instructions
_mm_cvtps_ph
and_mm_cvtph_ps
are used if: the code was compiled with__F16C__
(ie.-march=ivybridge
enables this), and if the CPU has the capability to run it (checked at runtime). - By default,
CPU_USE_REAL_HALF
is turned off. Note that the result doesn't change if it is turned on, since the conversion is done nevertheless.
Edited by Daniel Hugo Campora Perez