i965: Fix RNDZ and RNDE on Sandybridge and Ivybridge.

On gen4/5, the RNDZ and RNDE instructions return floor(x), but set special
"round increment bits" in the flag register; a predicated ADD (+1) fixes
the result.

The documentation still lists '.r' as existing, and says that the
predicated add is necessary, but it apparently lies.  According to the
simulator, BRW_CONDITIONAL_R (7) is not a valid conditional modifier
and the RNDZ and RNDE instructions simply produce the correct value.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
This commit is contained in:
Kenneth Graunke 2011-05-11 02:18:24 -07:00
parent 199a2f90ab
commit 344283de5d

View file

@ -784,6 +784,8 @@ struct brw_instruction *brw_##OP(struct brw_compile *p, \
* stores a rounded value (possibly the wrong way) in the dest register, but
* also sets a per-channel "increment bit" in the flag register. A predicated
* add of 1.0 fixes dest to contain the desired result.
*
* Sandybridge and later appear to round correctly without an ADD.
*/
#define ROUND(OP) \
void brw_##OP(struct brw_compile *p, \
@ -794,10 +796,13 @@ void brw_##OP(struct brw_compile *p, \
rnd = next_insn(p, BRW_OPCODE_##OP); \
brw_set_dest(p, rnd, dest); \
brw_set_src0(p, rnd, src); \
rnd->header.destreg__conditionalmod = 0x7; /* turn on round-increments */ \
\
add = brw_ADD(p, dest, dest, brw_imm_f(1.0f)); \
add->header.predicate_control = BRW_PREDICATE_NORMAL; \
if (p->brw->intel.gen < 6) { \
/* turn on round-increments */ \
rnd->header.destreg__conditionalmod = BRW_CONDITIONAL_R; \
add = brw_ADD(p, dest, dest, brw_imm_f(1.0f)); \
add->header.predicate_control = BRW_PREDICATE_NORMAL; \
} \
}