def prescan_test():
a = np.arange(2048).astype(np.int32)
reference = np.empty_like(a)
ref_sum = scan.exprefixsumNumba(a, reference)
a1 = np.arange(1024).astype(np.int32)
a2 = np.arange(1024, 2048).astype(np.int32)
ref1 = np.empty_like(a1)
ref2 = np.empty_like(a2)
ref_sum1 = scan.exprefixsumNumba(a1, ref1)
ref_sum2 = scan.exprefixsumNumba(a2, ref2)
dAux = cuda.device_array(2, dtype=np.int32)
dA = cuda.to_device(a)
sm_size = 1024 * a.dtype.itemsize
scan.prescan[2, 512, 0, sm_size](dA, dAux)
aux = dAux.copy_to_host()
a_gpu = dA.copy_to_host()
print "finish"
def prescan_test():
a = np.arange(2048).astype(np.int32)
reference = np.empty_like(a)
ref_sum = scan.exprefixsumNumba(a, reference)
a1 = np.arange(1024).astype(np.int32)
a2 = np.arange(1024, 2048).astype(np.int32)
ref1 = np.empty_like(a1)
ref2 = np.empty_like(a2)
ref_sum1 = scan.exprefixsumNumba(a1, ref1)
ref_sum2 = scan.exprefixsumNumba(a2, ref2)
dAux = cuda.device_array(2, dtype = np.int32)
dA = cuda.to_device(a)
sm_size = 1024 * a.dtype.itemsize
scan.prescan[2, 512, 0, sm_size](dA, dAux)
aux = dAux.copy_to_host()
a_gpu = dA.copy_to_host()
print "finish"
def get_new_graph(dest, weight, fe, od, mst, nod, nfe, ndest, nweight):
# first build the outDegree to get the first_edge
for e in range(mst.size):
edge = mst[e]
o_v = dest[edge] # destination
i_v = binaryEdgeIdSearch(edge, dest, fe, od)
if i_v == -1:
return -1
nod[o_v] += 1
nod[i_v] += 1
# get first edge from outDegree
exprefixsumNumba(nod, nfe, init = 0)
#get copy of newFirstEdge to serve as pointers for the newDest
top_edge = np.empty(nfe.size, dtype = np.int32)
for i in range(nfe.size):
top_edge[i] = nfe[i]
#top_edge = nfe.copy()
# go through all the mst edges again and write the new edges in the new arrays
for e in range(mst.size):
edge = mst[e]
o_v = dest[edge] # destination vertex
i_v = binaryEdgeIdSearch(edge, dest, fe, od)
if i_v == -1:
return -1
i_ptr = top_edge[i_v]
o_ptr = top_edge[o_v]
ndest[i_ptr] = o_v
ndest[o_ptr] = i_v
edge_w = weight[edge]
nweight[i_ptr] = edge_w
nweight[o_ptr] = edge_w
top_edge[i_v] += 1
top_edge[o_v] += 1
return 0
def last_block_test():
MAX_TPB = 512
n = 1024
a = np.arange(n).astype(np.int32)
reference = np.empty_like(a)
start = timer()
scan.exprefixsumNumba(a, reference, init=0)
end = timer()
auxidx = -1
elb = a.size
p2elb = np.int(np.ceil(np.log2(elb)))
telb = 2**p2elb
tlb = telb / 2
startIdx = 0
sm_size = telb * a.itemsize
aux = np.empty(1, dtype=np.int8)
trash = cuda.device_array(1)
e1, e2 = cuda.event(), cuda.event()
e1.record()
scan.last_scan[1, tlb, 0, sm_size](a, aux, -1, elb, startIdx)
e2.record()
print "CPU took: ", (end - start) * 1000, " ms"
print "Kernel took: ", cuda.event_elapsed_time(e1, e2), " ms"
print(a == reference).all()
def last_block_test():
MAX_TPB = 512
n = 1024
a = np.arange(n).astype(np.int32)
reference = np.empty_like(a)
start = timer()
scan.exprefixsumNumba(a, reference, init = 0)
end = timer()
auxidx = -1
elb = a.size
p2elb = np.int(np.ceil(np.log2(elb)))
telb = 2 ** p2elb
tlb = telb / 2
startIdx = 0
sm_size = telb * a.itemsize
aux = np.empty(1,dtype=np.int8)
trash = cuda.device_array(1)
e1, e2 = cuda.event(), cuda.event()
e1.record()
scan.last_scan[1, tlb, 0, sm_size](a, aux, -1, elb, startIdx)
e2.record()
print "CPU took: ", (end - start) * 1000, " ms"
print "Kernel took: ", cuda.event_elapsed_time(e1,e2), " ms"
print (a == reference).all()
def recursive_big_scan_test():
print "running recursive scan test"
MAX_TPB = 512
n = 2e6
n = int(n)
a = np.arange(n).astype(np.int32)
reference = np.empty_like(a)
start = timer()
sum_ref = scan.exprefixsumNumba(a, reference, init = 0)
end = timer()
dA = cuda.to_device(a)
# e1, e2 = cuda.event(), cuda.event()
# e1.record()
# e2.record()
start2 = timer()
total_sum = scan.scan_gpu(dA)
end2 = timer()
dA.copy_to_host(ary = a)
sum_gpu = total_sum.copy_to_host()
print "sum_ref = ", sum_ref
print "sum_gpu = ", sum_gpu
print "CPU took: ", (end - start) * 1000, " ms"
print "Kernel took: ", (end2 - start2) * 1000, " ms"
print (a == reference).all()
def recursive_big_scan_test():
print "running recursive scan test"
MAX_TPB = 512
n = 2e6
n = int(n)
a = np.arange(n).astype(np.int32)
reference = np.empty_like(a)
start = timer()
sum_ref = scan.exprefixsumNumba(a, reference, init=0)
end = timer()
dA = cuda.to_device(a)
# e1, e2 = cuda.event(), cuda.event()
# e1.record()
# e2.record()
start2 = timer()
total_sum = scan.scan_gpu(dA)
end2 = timer()
dA.copy_to_host(ary=a)
sum_gpu = total_sum.copy_to_host()
print "sum_ref = ", sum_ref
print "sum_gpu = ", sum_gpu
print "CPU took: ", (end - start) * 1000, " ms"
print "Kernel took: ", (end2 - start2) * 1000, " ms"
print(a == reference).all()
def recursive_step_by_step():
## setup
MAX_TPB = 512
n = 5000
a = np.arange(n).astype(np.int32)
reference = np.empty_like(a)
start = timer()
sum_ref = scan.exprefixsumNumba(a, reference, init=0)
end = timer()
dA = cuda.to_device(a)
# e1, e2 = cuda.event(), cuda.event()
# e1.record()
# e2.record()
## scan
in_ary = dA
epb = MAX_TPB * 2
whole_blocks = n // epb
el_last_block = n % epb
n_scans = whole_blocks
if el_last_block != 0:
n_scans += 1
## prescan
dAux = cuda.device_array(shape=n_scans, dtype=np.int32)
sm_size = epb * in_ary.dtype.itemsize
scan.prescan[whole_blocks, MAX_TPB, 0, sm_size](in_ary, dAux)
# tIn = in_ary.copy_to_host()
# tAux = dAux.copy_to_host()
p2elb = np.int(np.ceil(np.log2(el_last_block)))
p2_el_last_block = 2**p2elb # the smallest number of elements that is power of 2
tlb = p2_el_last_block >> 1 # number of threads in last block
sm_size = p2_el_last_block * in_ary.dtype.itemsize
startIdx = n - el_last_block
auxIdx = n_scans - 1
scan.last_scan[1, tlb, 0, sm_size](in_ary, dAux, auxIdx, el_last_block,
startIdx)
in_ary2 = dAux
n2 = in_ary2.size
if n2 < MAX_TPB << 1:
el_last_block2 = n2
p2elb2 = np.int(np.ceil(np.log2(el_last_block2)))
p2_el_last_block2 = 2**p2elb # the smallest number of elements that is power of 2
tlb2 = p2_el_last_block2 >> 1 # number of threads in last block
total_sum = cuda.device_array(shape=1, dtype=np.int32)
sm_size2 = p2_el_last_block2 * in_ary2.dtype.itemsize
startIdx2 = 0
auxIdx2 = 0
scan.last_scan[1, tlb2, 0, sm_size2](in_ary2, total_sum, auxIdx2,
el_last_block2, startIdx2)
scan.scan_sum[n_scans, tlb](in_ary, dAux)
tIn = in_ary.copy_to_host()
tAux = dAux.copy_to_host()
tSum = total_sum.copy_to_host()
print "finish"
def recursive_step_by_step():
## setup
MAX_TPB = 512
n = 5000
a = np.arange(n).astype(np.int32)
reference = np.empty_like(a)
start = timer()
sum_ref = scan.exprefixsumNumba(a, reference, init = 0)
end = timer()
dA = cuda.to_device(a)
# e1, e2 = cuda.event(), cuda.event()
# e1.record()
# e2.record()
## scan
in_ary = dA
epb = MAX_TPB * 2
whole_blocks = n // epb
el_last_block = n % epb
n_scans = whole_blocks
if el_last_block != 0:
n_scans += 1
## prescan
dAux = cuda.device_array(shape = n_scans, dtype = np.int32)
sm_size = epb * in_ary.dtype.itemsize
scan.prescan[whole_blocks, MAX_TPB, 0, sm_size](in_ary, dAux)
# tIn = in_ary.copy_to_host()
# tAux = dAux.copy_to_host()
p2elb = np.int(np.ceil(np.log2(el_last_block)))
p2_el_last_block = 2 ** p2elb # the smallest number of elements that is power of 2
tlb = p2_el_last_block >> 1 # number of threads in last block
sm_size = p2_el_last_block * in_ary.dtype.itemsize
startIdx = n - el_last_block
auxIdx = n_scans - 1
scan.last_scan[1, tlb, 0, sm_size](in_ary, dAux, auxIdx, el_last_block, startIdx)
in_ary2 = dAux
n2 = in_ary2.size
if n2 < MAX_TPB << 1:
el_last_block2 = n2
p2elb2 = np.int(np.ceil(np.log2(el_last_block2)))
p2_el_last_block2 = 2 ** p2elb # the smallest number of elements that is power of 2
tlb2 = p2_el_last_block2 >> 1 # number of threads in last block
total_sum = cuda.device_array(shape = 1, dtype = np.int32)
sm_size2 = p2_el_last_block2 * in_ary2.dtype.itemsize
startIdx2 = 0
auxIdx2 = 0
scan.last_scan[1, tlb2, 0, sm_size2](in_ary2, total_sum, auxIdx2, el_last_block2, startIdx2)
scan.scan_sum[n_scans, tlb](in_ary, dAux)
tIn = in_ary.copy_to_host()
tAux = dAux.copy_to_host()
tSum = total_sum.copy_to_host()
print "finish"