JIT: faster PPC_FP code #856
Conversation
|
I had a friend test this by using the first stage of Starfox Assault and she was able to complete the level without issues, but that's all the testing I've done. |
FioraAeterna
force-pushed
the
ppcfpopt
branch
7 times, most recently
from
3486830
to
49976c6
Compare
|
Some stats (first ~30 seconds of Sonic Colors) Double to Single: 378,150,740 floats sent to fast path, 18,092 to slow path |
| // A sneaky hack: floating-point zero is rather common and we don't want to confuse it for denormals and | ||
| // needlessly send it through the slow path. If we subtract 1 from the input, it turns float-zero into | ||
| // 0xffffffff (skipping the slow path). This results in a single non-denormal being sent through the | ||
| // slow path (0x00800000), but the effect of that should be negligible. |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
|
Nice PR. Just let me test this on one of the games which the original PPC_FP merge broke. |
The PPC_FP conversion code can be made a lot simpler with the observation that the only values that need to be sent through the slow x87 path are denormals. A whole bunch faster: 708->678 seconds on POV-RAY.
|
Tested Tak and the Power of Juju (GJUE78); works fine. |
|
Wow, super pro. This sounds awesome. |
|
4% boost in my CPU limited virtual console games. Affects Melee Fountain of Dreams by about 4 - 5% as well. Tested a ton more games, Everything seems to be working, including the games that were broke by the original (broken) PPC_FP merge. This looks good to me! Looking forward to the performance boosts. |
|
LGTM. I think this is ready to be merged. |
| // if it is. | ||
| MOVQ_xmm(R(RAX), src); | ||
| SHR(64, R(RAX), Imm8(55)); | ||
| // Exponents 0x369 <= x <= 0x380 are denormal. This code accepts the range 0x368 <= x <= 0x387 |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
| } | ||
|
|
||
| SetJumpTarget(dont_reset_qnan_bit); | ||
| MOVDDUP(dst, R(XMM0)); |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
The PPC_FP conversion code can be made a lot simpler with the observation
that the only values that need to be sent through the slow x87 path are
denormals.
A whole bunch faster: 708->678 seconds on POV-RAY.