mirror-github-bspeice/cuburn
1
0
Fork 0
mirror of https://github.com/stevenrobertson/cuburn.git synced 2025-02-05 11:40:04 -05:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix missing control points in async version.

Browse Source 
The allocation pool was reallocating the same frame as soon as it had
left the current scope, before it had been copied. We just reallocate
the same chunks. I don't think this has any real performance impact but
this can be verified.
This commit is contained in:
Steven Robertson 2011-10-11 20:54:33 -04:00
parent b081bc9378
commit 9b03f557c2
2 changed files with 7 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
cuburn/code/filtering.py
View File

@ -241,7 +241,6 @@ void density_est(float4 *pixbuf, float4 *outbuf,
de_add(si, jj, -ii, scaled); de_add(si, jj, -ii, scaled);
iif += 1; iif += 1;
// TODO: validate that the above avoids bank conflicts
} }
} }
} }

17
cuburn/render.py
View File

@ -176,12 +176,10 @@ class Animation(object):
""" """
# Don't see this changing, but empirical tests could prove me wrong # Don't see this changing, but empirical tests could prove me wrong
NRENDERERS = 2 NRENDERERS = 2
# This could be shared too?
pool = pycuda.tools.PageLockedMemoryPool()
# TODO: under a slightly modified sequencing, certain buffers can be # TODO: under a slightly modified sequencing, certain buffers can be
# shared (though this may be unimportant if a good AA technique which # shared (though this may be unimportant if a good AA technique which
# doesn't require full SS can be found) # doesn't require full SS can be found)
rdrs = [_AnimRenderer(self, pool) for i in range(NRENDERERS)] rdrs = [_AnimRenderer(self) for i in range(NRENDERERS)]
# Zip up each genome with an alternating renderer, plus enough empty # Zip up each genome with an alternating renderer, plus enough empty
# genomes at the end to flush all pending tasks # genomes at the end to flush all pending tasks
@ -196,8 +194,6 @@ class Animation(object):
def _interp(self, time, cp): def _interp(self, time, cp):
flam3_interpolate(self._g_arr, len(self._g_arr), time, 0, byref(cp)) flam3_interpolate(self._g_arr, len(self._g_arr), time, 0, byref(cp))
class _AnimRenderer(object): class _AnimRenderer(object):
# Large launches lock the display for a considerable period and may be # Large launches lock the display for a considerable period and may be
# killed due to a device timeout; small launches are harder to load-balance # killed due to a device timeout; small launches are harder to load-balance
@ -214,9 +210,8 @@ class _AnimRenderer(object):
PAL_HEIGHT = 16 PAL_HEIGHT = 16
def __init__(self, anim, pool): def __init__(self, anim):
self.anim = anim self.anim = anim
self.pool = pool
self.pending = False self.pending = False
self.stream = cuda.Stream() self.stream = cuda.Stream()
@ -235,7 +230,9 @@ class _AnimRenderer(object):
self.nbins = anim.features.acc_height * anim.features.acc_stride self.nbins = anim.features.acc_height * anim.features.acc_stride
self.d_accum = cuda.mem_alloc(16 * self.nbins) self.d_accum = cuda.mem_alloc(16 * self.nbins)
self.d_out = cuda.mem_alloc(16 * self.nbins) self.d_out = cuda.mem_alloc(16 * self.nbins)
self.d_infos = cuda.mem_alloc(anim._iter.packer.align * self.ncps)
info_size = anim._iter.packer.align * self.ncps
self.d_infos = cuda.mem_alloc(info_size)
# Defer generation of seeds until they're first needed # Defer generation of seeds until they're first needed
self.d_seeds = None self.d_seeds = None
@ -286,7 +283,7 @@ class _AnimRenderer(object):
if not d_seeds: if not d_seeds:
seeds = mwc.MWC.make_seeds(iter.IterCode.NTHREADS * seeds = mwc.MWC.make_seeds(iter.IterCode.NTHREADS *
self.cps_per_block) self.cps_per_block)
h_seeds = self.pool.allocate(seeds.shape, seeds.dtype) h_seeds = cuda.pagelocked_empty(seeds.shape, seeds.dtype)
h_seeds[:] = seeds h_seeds[:] = seeds
size = seeds.dtype.itemsize * seeds.size size = seeds.dtype.itemsize * seeds.size
d_seeds = cuda.mem_alloc(size) d_seeds = cuda.mem_alloc(size)
@ -315,7 +312,7 @@ class _AnimRenderer(object):
bkgd += np.array(a.genomes[0].background) * len(block_times) bkgd += np.array(a.genomes[0].background) * len(block_times)
infos = np.concatenate(infos) infos = np.concatenate(infos)
h_infos = self.pool.allocate(infos.shape, infos.dtype) h_infos = cuda.pagelocked_empty(infos.shape, infos.dtype)
h_infos[:] = infos h_infos[:] = infos
offset = b * packer.align * self.cps_per_block offset = b * packer.align * self.cps_per_block
# TODO: portable across 32/64-bit arches? # TODO: portable across 32/64-bit arches?