Steven Robertson
b938c320a8
Last touchups before ripping out the DSL
2010-09-13 12:22:08 -04:00
Steven Robertson
e4aac6993f
A few touchups
2010-09-13 00:20:15 -04:00
Steven Robertson
e0b218feba
A new (somewhat experimental) approach to fusing
2010-09-12 23:45:38 -04:00
Steven Robertson
5a5fcf5bb9
Fix the unbelieveably stupid bug I've been chasing for days.
2010-09-12 18:42:52 -04:00
Steven Robertson
2f48d01aa9
Fix linear variation typo
2010-09-12 17:38:51 -04:00
Steven Robertson
5c5122e8c8
Optimization doubles performance... but breaks the output (even more)
2010-09-12 17:17:08 -04:00
Steven Robertson
d01de61952
Simultaneous occupancy microbenchmark
2010-09-12 16:23:24 -04:00
Steven Robertson
3e4e1d88a2
Allow device call exceptions to propagate after cleanup
2010-09-12 16:22:56 -04:00
Steven Robertson
70ca6d7729
Fix RNG test
2010-09-12 16:22:22 -04:00
Steven Robertson
a6141f492d
A byte is *8* bits
2010-09-12 15:48:31 -04:00
Steven Robertson
7ef0d334ca
...except I missed the file that actually contained the new method
2010-09-12 14:06:07 -04:00
Steven Robertson
6ed8907fcb
LaunchContext.get_per_thread
2010-09-12 13:45:55 -04:00
Steven Robertson
3265982fec
Change 'ctx.threads' to 'ctx.nthreads', as it should have been from the start
2010-09-12 11:13:53 -04:00
Steven Robertson
a439bf671d
Fix occupancy issues (1 block/SM when shuffle was on).
...
There are 16 bar.sync() registers available per *chip*, not per block, and I
was using number 8 in the shuffle code. Evidently the driver rewrites them per
SM, but does not compact their range. Good to know.
2010-09-12 11:09:47 -04:00
Steven Robertson
ce0571deaf
A fake log filter stage while I work on other stuff
2010-09-12 02:32:03 -04:00
Steven Robertson
c13f6a06cf
Experiments with larger CTAs for IterThread
2010-09-12 02:01:03 -04:00
Steven Robertson
e2b1c161cf
More readable memory allocations
2010-09-12 01:13:22 -04:00
Steven Robertson
802ca1d585
Allow swapping out store methods for easier testing of performance
2010-09-12 01:09:04 -04:00
Steven Robertson
f368a99a16
Shuffle points between threads of a CTA
2010-09-12 00:17:18 -04:00
Steven Robertson
40a5ceafde
Use a somewhat better writeback mechanism for now
2010-09-12 00:16:35 -04:00
Steven Robertson
aa688564f1
Add Timeouter, for timing out infinite loops so data can be recovered.
2010-09-11 13:18:40 -04:00
Steven Robertson
a5d7c2cc1a
Use variations. This works, but is still fragile.
2010-09-11 13:15:36 -04:00
Steven Robertson
860d7b2fad
Add xforms and variations.
2010-09-11 13:10:41 -04:00
Steven Robertson
383c0f1f9a
Fixed bench.py, with the help of Device Assertions™!*
...
* Not actually a trademark
2010-09-11 00:16:43 -04:00
Steven Robertson
56404b629f
Add device assertions to standard library.
2010-09-11 00:12:02 -04:00
Steven Robertson
3932412539
Test to make sure floating point numbers were in the right range.
2010-09-10 19:36:39 -04:00
Steven Robertson
e71a8422e5
Make store_per_thread reuse gtid in multiple calls when possible
2010-09-10 18:45:32 -04:00
Steven Robertson
943e92b80c
Use pycuda SourceModule to work around crashes, and a few invocation touchups.
2010-09-10 18:02:37 -04:00
Steven Robertson
c3d12d07c2
Fix MWCRNGTest.
2010-09-10 18:01:50 -04:00
Steven Robertson
36f1c1c056
Rename "cuburnlib" (stupid) to "cuburn" (stupid but shorter)
...
--HG--
rename : cuburnlib/__init__.py => cuburn/__init__.py
rename : cuburnlib/cuda.py => cuburn/cuda.py
rename : cuburnlib/device_code.py => cuburn/device_code.py
rename : cuburnlib/ptx.py => cuburn/ptx.py
rename : cuburnlib/render.py => cuburn/render.py
2010-09-10 14:48:34 -04:00
Steven Robertson
4552589b35
Refactor call() to be more elegant
2010-09-10 14:43:20 -04:00
Steven Robertson
fb4e5b75e9
Add support for writing float literals in store_per_thread
2010-09-10 14:33:56 -04:00
Steven Robertson
6eaa80be7a
Added property ctx.warps_per_cta
2010-09-10 12:53:40 -04:00
Steven Robertson
2f3ac42153
Improved DataStream record format
2010-09-10 12:53:20 -04:00
Steven Robertson
7e0d36af7d
Add performance tuning (maxnregs) to entries
2010-09-10 12:52:47 -04:00
Steven Robertson
086e4e4fb4
Lots-o-stuff.
2010-09-09 11:36:14 -04:00
Steven Robertson
1f7b00b61e
instmethod decorator: another hack (to get around ctx.ptx.instances[])
2010-09-08 13:12:46 -04:00
Steven Robertson
094890c324
Use shared memory for iter_count and have each CP processed by only one CTA.
...
Slower, but the code is a bit simpler conceptually, and the difference will be
more than accounted for by better scheduling towards the end of the process.
2010-09-07 14:54:50 -04:00
Steven Robertson
aa065dc25d
Add the first of many microbenchmarks
2010-09-07 12:44:12 -04:00
Steven Robertson
db72a7d496
Allow register local name rebinding
2010-09-06 16:50:54 -04:00
Steven Robertson
e03f20392d
Switch from to_inject() to object insertion. One less kludge to deal with.
2010-09-06 16:09:37 -04:00
Steven Robertson
ada0fe20c7
Random floats (I think)
2010-09-06 14:19:06 -04:00
Steven Robertson
f3298e0bed
Finally runs again
2010-09-06 11:18:20 -04:00
Steven Robertson
27e7fd82a3
Time to go have nightmares about this code again (no really)
2010-09-03 00:52:27 -04:00
Steven Robertson
2c26ff9ab6
* Fix deptrace typos
...
* Add predicate support to DeviceStream fetches, making them even uglier
* Add `store_per_thread` to PTX stdlib
2010-09-03 00:51:23 -04:00
Steven Robertson
a68fc064a1
Added TODO, because even in my nightmares I underestimate how much is left
2010-09-03 00:08:58 -04:00
Steven Robertson
7e7fbda2cc
Start of render module
2010-09-02 17:26:16 -04:00
Steven Robertson
a23a493d68
Formatter improvements
2010-09-02 16:12:22 -04:00
Steven Robertson
731c637f80
DataStream. Completely untested. I want to see the bugfixes in the log.
2010-09-02 16:11:44 -04:00
Steven Robertson
bf79dc7fa0
Adjust softjoin and vec to be easier to format
2010-09-02 15:05:24 -04:00
