def make_fp_tree():
#### Allocate host memory
offsets, transactions, num_transactions, all_items_in_transactions = readFile(
"data.txt")
print num_transactions, all_items_in_transactions
flist = np.zeros(MAX_UNIQUE_ITEMS, dtype=np.uint32)
#### Allocate and initialize GPU/Device memory
d_offsets = cuda.to_device(offsets)
d_transactions = cuda.to_device(transactions)
d_flist = cuda.to_device(flist)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(num_transactions / (1.0 * threads_per_block[0])) +
1, 1)
t1 = time()
makeFlistGPU[number_of_blocks,
threads_per_block](d_offsets, d_transactions, d_flist,
num_transactions,
all_items_in_transactions)
cuda.synchronize()
t2 = time()
d_flist.copy_to_host(flist)
cuda.synchronize()
#
# for i in range(0, MAX_UNIQUE_ITEMS):
# print i, flist[i]
t3 = time()
flist_cpu = makeFlist(transactions, all_items_in_transactions)
t4 = time()
#
match = 1
for i in range(1, MAX_UNIQUE_ITEMS):
if i not in flist_cpu and flist[i] == 0:
continue
#print i, flist[i], flist_cpu[i]
if flist[i] != flist_cpu[i]:
match = -1
break
if match == 1:
print "Test Passed"
else:
print "Test Failed"
print "Number of transactions = ", num_transactions
print "All items in transactions = ", all_items_in_transactions
print "GPU time = ", t2 - t1
print "CPU TIME = ", t4 - t3
def time_this(kernel, gridsz, blocksz, args):
timings = []
cuda.synchronize()
try:
for i in range(10): # best of 10
ts = timer()
kernel[gridsz, blocksz](*args)
cuda.synchronize()
te = timer()
timings.append(te - ts)
except cudadrv.error.CudaDriverError, e:
print 'exc suppressed', e
return -1
def time_this(kernel, gridsz, blocksz, args):
timings = []
cuda.synchronize()
try:
for i in range(10): # best of 10
ts = timer()
kernel[gridsz, blocksz](*args)
cuda.synchronize()
te = timer()
timings.append(te - ts)
except cudadrv.error.CudaDriverError, e:
print 'exc suppressed', e
return -1
def make_fp_tree():
#### Allocate host memory
offsets, transactions, num_transactions, all_items_in_transactions = readFile("data.txt")
print num_transactions, all_items_in_transactions
flist = np.zeros(MAX_UNIQUE_ITEMS, dtype=np.uint32)
#### Allocate and initialize GPU/Device memory
d_offsets = cuda.to_device(offsets)
d_transactions = cuda.to_device(transactions)
d_flist = cuda.to_device(flist)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(num_transactions / (1.0 * threads_per_block[0])) + 1, 1)
t1 = time()
makeFlistGPU [number_of_blocks, threads_per_block] (d_offsets, d_transactions, d_flist, num_transactions, all_items_in_transactions)
cuda.synchronize()
t2 = time()
d_flist.copy_to_host(flist)
cuda.synchronize()
#
# for i in range(0, MAX_UNIQUE_ITEMS):
# print i, flist[i]
t3 = time()
flist_cpu = makeFlist(transactions, all_items_in_transactions)
t4 = time()
#
match = 1
for i in range(1, MAX_UNIQUE_ITEMS):
if i not in flist_cpu and flist[i] == 0:
continue
#print i, flist[i], flist_cpu[i]
if flist[i] != flist_cpu[i]:
match = -1
break
if match == 1:
print "Test Passed"
else:
print "Test Failed"
print "Number of transactions = ", num_transactions
print "All items in transactions = ", all_items_in_transactions
print "GPU time = ", t2 - t1
print "CPU TIME = ", t4 - t3
def test_histogram():
#Allocate host memory
input_h = np.empty(NUM_ELEMENTS, dtype=np.uint32)
bins_h = np.zeros(BIN_SIZE, dtype=np.uint32)
myprint("Bin Size = " + str(bins_h.size))
## Initialize host memory
for i in range(0, NUM_ELEMENTS):
input_h[i] = randint(0, BIN_SIZE - 1)
## Allocate and initialize GPU/device memory
input_d = cuda.to_device(input_h)
bins_d = cuda.to_device(bins_h)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(NUM_ELEMENTS / (1.0 * threads_per_block[0]))), 1)#((NUM_ELEMENTS / threads_per_block[0]) + 1, 1)
t1 = time()
histogramGPU [number_of_blocks, threads_per_block] (input_d, bins_d, NUM_ELEMENTS)
cuda.synchronize()
t2 = time()
bins_d.copy_to_host(bins_h)
t3 = time()
bins_cpu = makeHist(input_h)
t4 = time()
# for i in range(0, BIN_SIZE):
# print i, bins_h[i], bins_cpu[i]
print "GPU time = ", t2 - t1
print "CPU TIME = ", t4 - t3
match = 1
for i in range(0, BIN_SIZE):
if bins_h[i] != bins_cpu[i]:
match = -1
break
if match == 1:
print "Test Passed"
else:
print "Test Failed"
def test_scan():
in_h = np.empty(NUM_ELEMENTS, dtype=np.uint32)
out_h = np.zeros(NUM_ELEMENTS, dtype=np.uint32)
for i in range(0, NUM_ELEMENTS):
in_h[i] = NUM_ELEMENTS -i - 1#randint(0, 100)
tac1 = time()
in_d = cuda.to_device(in_h)
out_d = cuda.to_device(out_h)
cuda.synchronize()
tac2 = time()
tk1 = time()
for i in range(0, 32):
tk1 = time()
preScan(out_d, in_d, NUM_ELEMENTS)
cuda.synchronize()
tk2 = time()
print i, tk2 - tk1
tk2 = time()
th1 = time()
out_d.copy_to_host(out_h)
cuda.synchronize()
#print "Last = ", out_h[-1] + in_h[-1]
th2 = time()
def test_sort():
in_h = np.empty(NUM_ELEMENTS, dtype=np.uint32) #4, 7, 2, 6, 3, 5, 1, 0
out_h = np.empty(NUM_ELEMENTS, dtype=np.uint32)
for i in range(0, NUM_ELEMENTS):
in_h[i] = NUM_ELEMENTS - i - 1
in_d = cuda.to_device(in_h)
out_d = cuda.device_array(NUM_ELEMENTS, dtype=np.uint32)
temp_d = cuda.device_array(NUM_ELEMENTS, dtype=np.uint32)
tkg1 = time()
for bit_shift in range(0, 32):
tk1 = time()
#radix_sort(in_d, out_d, temp_d, in_h[NUM_ELEMENTS - 1], bit_shift)
preScan(out_d, in_d, NUM_ELEMENTS)
tk2 = time()
#print bit_shift, tk2 - tk1
in_d = out_d
out_d = temp_d
temp_d = in_d
tkg2 = time()
out_d.copy_to_host(out_h)
cuda.synchronize()
# line = ""
# for i in range(0, NUM_ELEMENTS):
# line += " " + str(out_h[i])
#
# print line
in_cpu = [NUM_ELEMENTS - i -1 for i in range(0, NUM_ELEMENTS)]
tc1 = time()
in_cpu.sort()
tc2 = time()
print "GPU Time = ", tkg2 - tkg1
print "CPU Time = ", tc2 - tc1
def test_sort():
in_h = np.empty(NUM_ELEMENTS, dtype=np.uint32) #4, 7, 2, 6, 3, 5, 1, 0
out_h = np.empty(NUM_ELEMENTS, dtype=np.uint32)
for i in range(0, NUM_ELEMENTS):
in_h[i] = NUM_ELEMENTS - i - 1
in_d = cuda.to_device(in_h)
out_d = cuda.device_array(NUM_ELEMENTS, dtype=np.uint32)
temp_d = cuda.device_array(NUM_ELEMENTS, dtype=np.uint32)
tkg1 = time()
for bit_shift in range(0, 32):
tk1 = time()
#radix_sort(in_d, out_d, temp_d, in_h[NUM_ELEMENTS - 1], bit_shift)
preScan(out_d, in_d, NUM_ELEMENTS)
tk2 = time()
#print bit_shift, tk2 - tk1
in_d = out_d
out_d = temp_d
temp_d = in_d
tkg2 = time()
out_d.copy_to_host(out_h)
cuda.synchronize()
# line = ""
# for i in range(0, NUM_ELEMENTS):
# line += " " + str(out_h[i])
#
# print line
in_cpu = [NUM_ELEMENTS - i - 1 for i in range(0, NUM_ELEMENTS)]
tc1 = time()
in_cpu.sort()
tc2 = time()
print "GPU Time = ", tkg2 - tkg1
print "CPU Time = ", tc2 - tc1
def radix_sort(in_d, out_d, out_scan_d, last_inp_element, bit_shift=0):
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(NUM_ELEMENTS / (1.0 * threads_per_block[0]))),
1)
################ Bit flip ########################
SplitGPU[number_of_blocks, threads_per_block](in_d, out_d, NUM_ELEMENTS,
bit_shift)
cuda.synchronize()
# out_d.copy_to_host(out_h)
# cuda.synchronize()
##################################################
t1 = time()
preScan(out_scan_d, out_d, NUM_ELEMENTS)
cuda.synchronize()
t2 = time()
#print "Time = ", t2 - t1
# out_scan_d.copy_to_host(out_h)
# cuda.synchronize()
#
###########################################################
IndexDefineGPU[number_of_blocks,
threads_per_block](out_scan_d, out_d, NUM_ELEMENTS,
last_inp_element)
cuda.synchronize()
# out_scan_d.copy_to_host(out_h)
# cuda.synchronize()
###########################################################
############################################################
ScatterElementGPU[number_of_blocks, threads_per_block](in_d, out_scan_d,
out_d, NUM_ELEMENTS)
cuda.synchronize()
def radix_sort(in_d, out_d, out_scan_d, last_inp_element, bit_shift=0):
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(NUM_ELEMENTS / (1.0 * threads_per_block[0]))), 1)
################ Bit flip ########################
SplitGPU [number_of_blocks, threads_per_block] (in_d, out_d, NUM_ELEMENTS, bit_shift)
cuda.synchronize()
# out_d.copy_to_host(out_h)
# cuda.synchronize()
##################################################
t1 = time()
preScan(out_scan_d, out_d, NUM_ELEMENTS)
cuda.synchronize()
t2 = time()
#print "Time = ", t2 - t1
# out_scan_d.copy_to_host(out_h)
# cuda.synchronize()
#
###########################################################
IndexDefineGPU [number_of_blocks, threads_per_block] (out_scan_d, out_d, NUM_ELEMENTS, last_inp_element)
cuda.synchronize()
# out_scan_d.copy_to_host(out_h)
# cuda.synchronize()
###########################################################
############################################################
ScatterElementGPU [number_of_blocks, threads_per_block] (in_d, out_scan_d, out_d, NUM_ELEMENTS)
cuda.synchronize()
def test_sort():
in_h = np.empty(NUM_ELEMENTS, dtype=np.uint32) #4, 7, 2, 6, 3, 5, 1, 0
#in_h = np.array([4, 7, 2, 6, 3, 5, 1, 0], dtype=np.uint32)
out_h = np.empty(NUM_ELEMENTS, dtype=np.uint32)
for i in range(0, NUM_ELEMENTS):
in_h[i] = randint(0, 100) #NUM_ELEMENTS - i - 1
#in_h = np.array([6, 44, 71, 79, 94, 92, 12, 56, 47, 17, 81, 98, 84, 9, 85, 99], dtype=np.uint32)
#in_h = np.array([85, 37, 50, 73, 51, 46, 62, 84, 65, 99, 76, 59, 73, 16, 27, 4, 75, 81, 80, 33, 73, 11, 29, 24, 81, 49, 27, 71, 74, 64, 60, 91], dtype=np.uint32)
print in_h
in_d = cuda.to_device(in_h)
out_d = cuda.device_array(NUM_ELEMENTS, dtype=np.uint32)
tkg1 = time()
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(
ceil(NUM_ELEMENTS / (2 * 1.0 * threads_per_block[0]))), 1)
RadixGPU[number_of_blocks, threads_per_block](in_d, out_d, NUM_ELEMENTS)
out_d.copy_to_host(out_h)
#print "Rad = ", list(out_h)
stride = 4
# while stride < NUM_ELEMENTS:
# number_of_blocks = (int(ceil(NUM_ELEMENTS / (stride * 1.0 * threads_per_block[0]))), 1)
# bitonicSort [number_of_blocks, threads_per_block] (out_d, NUM_ELEMENTS, stride)
# stride *= 2
# # number_of_blocks = (int(ceil(NUM_ELEMENTS / (2 * 1.0 * threads_per_block[0]))), 1)
# # RadixGPU [number_of_blocks, threads_per_block] (out_d, in_d, NUM_ELEMENTS)
# # out_d = in_d
# out_d.copy_to_host(out_h)
# print "Str = ", list(out_h)
# break
# # stride /= 2
# while stride >= 4:
# number_of_blocks = (int(ceil(NUM_ELEMENTS / (stride * 1.0 * threads_per_block[0]))), 1)
# bitonicSort [number_of_blocks, threads_per_block] (out_d, NUM_ELEMENTS, stride)
# stride /= 2
# cuda.synchronize()
#
# number_of_blocks = (int(ceil(NUM_ELEMENTS / (2 * 1.0 * threads_per_block[0]))), 1)
# RadixGPU [number_of_blocks, threads_per_block] (out_d, in_d, NUM_ELEMENTS)
# out_d = in_d
#
# out_d.copy_to_host(out_h)
# cuda.synchronize()
#
# line = ""
# for i in range(0, NUM_ELEMENTS):
# line += " " + str(out_h[i])
#
# print line
tkg2 = time()
out_d.copy_to_host(out_h)
cuda.synchronize()
#print "GPU = ", list(out_h)
# line = ""
# for i in range(0, NUM_ELEMENTS):
# line += " " + str(out_h[i])
#
# print line
in_cpu = list(in_h) #[NUM_ELEMENTS - i -1 for i in range(0, NUM_ELEMENTS)]
tc1 = time()
in_cpu.sort()
#print "CPU = ", in_cpu
tc2 = time()
print "GPU Time = ", tkg2 - tkg1
print "CPU Time = ", tc2 - tc1
print len(in_cpu)
# Prepare data on the GPU dA = cuda.to_device(A) dB = cuda.to_device(B) dC = cuda.to_device(C) # device_array_like(A) # Time numpy version s = timer() np_ans = np.dot(A, B) e = timer() t = e - s # Time the unoptimized version s = timer() cu_matmul[grid_dim, block_dim](dA, dB, dC, n) cuda.synchronize() e = timer() unopt_ans = dC.copy_to_host() tcuda_unopt = e - s # Time the shared memory version s = timer() cu_matmul_sm[grid_dim, block_dim](dA, dB, dC, n, tpb, bpg) cuda.synchronize() e = timer() opt_ans = dC.copy_to_host() tcuda_opt = e - s # Time for CuBLAS version s = timer() blas = cublas.Blas()
def test_apriori():
output_file = open("apriori_out.txt", "w")
offsets, transactions, num_transactions, num_elements = readFile("syncthetic_data.txt")
print "Offset = ", offsets[:num_transactions]
print "transactions = ", transactions[:num_elements]
print "Num transactions = ", num_transactions
print "Num elements = ", num_elements
min_support = MIN_SUPPORT
# to find number of max digits required to represent that many number of unique items
power = 1
while MAX_UNIQUE_ITEMS / (10 ** power) != 0:
power += 1
print "Power = ", power
t = [item for item in transactions.tolist()]
if num_elements > NUM_ELEMENTS:
print "Error: Elements exceeding NUM_ELEMENTS. Exiting..."
sys.exit(12)
input_h = np.array(t, dtype=np.int32)
print "Input transactions = ", list(input_h)
print "Size of transactions = ", input_h.size
ci_h = np.zeros(MAX_UNIQUE_ITEMS, dtype=np.int32)
li_h = np.empty(MAX_UNIQUE_ITEMS, dtype=np.int32)
input_d = cuda.to_device(input_h)
ci_d = cuda.to_device(ci_h)
li_d = cuda.device_array(MAX_UNIQUE_ITEMS, dtype=np.int32)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(NUM_ELEMENTS / (1.0 * threads_per_block[0]))), 1)#((NUM_ELEMENTS / threads_per_block[0]) + 1, 1)
histogramGPU [number_of_blocks, threads_per_block] (input_d, ci_d, num_elements)
#cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Ci_H Histogram result = ", ci_h # support count for each item
number_of_blocks = (int(ceil(MAX_UNIQUE_ITEMS / (1.0 * threads_per_block[0]))), 1)
pruneGPU [number_of_blocks, threads_per_block] (ci_d, MAX_UNIQUE_ITEMS, min_support)
cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Keys = ", [i for i in range(0, len(ci_h))]
print "Ci_H Pruning result = ", ci_h # support count for each item
# calculate concise list of items satisfying min support
l1_patterns = {}
k = 0 # number of items whose sup_count > min_support
for j in range(0, len(ci_h)):
if ci_h[j] != 0:
li_h[k] = j
l1_patterns[(j, )] = ci_h[j]
k += 1
print "\n=======================================================\n"
print "L1 = ", list(li_h)[:k] #items whose support_count > min_support
print "\n=======================================================\n"
output_file.write(createFormattedPatterns(l1_patterns, 1))
print "K(num_items_with_good_sup_count = ", k
#k = 102
ci_h = np.array([-1 for i in range(0, k ** 2)], dtype=np.int32)
ci_d = cuda.to_device(ci_h)
#li_h = np.array(sorted([randint(10, 99) for i in range(0, k)]), dtype=np.int32)
#tli_h = np.array([i for i in range(1, k + 1)], dtype=np.int32)
t1 = time()
li_d = cuda.to_device(li_h)
number_of_blocks = (int(ceil(k / (1.0 * MAX_ITEM_PER_SM))), 1)
print "Self join 2 number of blocks = ", number_of_blocks
print "K = ", k
print "Ci_H size = ", ci_h.size
print "LI_H size = ", li_h.size
selfJoinGPU [number_of_blocks, threads_per_block](li_d, ci_d, k, power)
cuda.synchronize()
li_d.copy_to_host(li_h)
ci_d.copy_to_host(ci_h)
t2 = time()
#sys.exit(0)
# f = open('join.txt', 'w')
#
# for i in range(0, k):
# line = ""
# for j in range(0, k):
# line += str(ci_h[k * i + j]) + " "
# f.write(line + "\n")
#
# f.close()
#ci_h = ci_h.reshape(k, k)
print "Initial Mask = ", ci_h.reshape(k, k)
print "Self joining time = ", (t2 - t1)
d_offsets = cuda.to_device(offsets)
d_transactions = cuda.to_device(transactions)
#number_of_blocks = (1, 1) #(int(num_transactions / (1.0 * threads_per_block[0])) + 1, 1)
number_of_blocks = (int(ceil(num_transactions / (1.0 * MAX_TRANSACTIONS_PER_SM))), 1)
print "Num blocks for findFrequency = ", number_of_blocks
print "Num transactions = ", num_transactions
print "Num patterns = ", k
print "index = ", list(li_h)[:k]
findFrequencyGPU [number_of_blocks, threads_per_block] (d_transactions, d_offsets, num_transactions, num_elements, li_d, ci_d, k)
cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Final Mask = ", ci_h.reshape(k, k)
d_transactions.copy_to_host(transactions)
threads_per_block = (BLOCK_SIZE, BLOCK_SIZE)
number_of_blocks = ((int(ceil(k / (1.0 * threads_per_block[0])))), (int(ceil(k / (1.0 * threads_per_block[0])))))
pruneMultipleGPU [number_of_blocks, threads_per_block] (ci_d, k, min_support) # prunes according to min_support
ci_d.copy_to_host(ci_h)
print "Outer Mask = ", ci_h.reshape(k, k)
ci_hn = np.zeros(k, dtype=np.int32)
ci_dn = cuda.to_device(ci_hn)
combinationsAvailable [threads_per_block, number_of_blocks] (ci_d, ci_dn, k) #Number of possible patterns in each row
ci_dn.copy_to_host(ci_hn)
print "Ci_hn = ", list(ci_hn)
ci_hnx = np.empty(k, dtype=np.int32)
ci_dnx = cuda.to_device(ci_hnx)
preScan(ci_dnx, ci_dn, k) # Prefix sum on patterns in each row
ci_dnx.copy_to_host(ci_hnx)
num_patterns = ci_hnx[-1]
print "Ci_hnx = ", list(ci_hnx)
sparseM_h = np.empty(ci_hnx[-1] * 3, dtype=np.uint32)
sparseM_d = cuda.to_device(sparseM_h)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(k / (1.0 * threads_per_block[0]))), 1)
convert2Sparse [threads_per_block, number_of_blocks] (ci_d, ci_dnx, sparseM_d, num_patterns, k)
sparseM_d.copy_to_host(sparseM_h)
# sparseM_h = sparseM_h.reshape(3, num_patterns)
print sparseM_h.reshape(3, num_patterns)
patterns = {}
for i in range(0, num_patterns):
item1 = sparseM_h[i]
item2 = sparseM_h[i + num_patterns]
support = sparseM_h[i + 2 * num_patterns]
patterns[tuple(sorted([li_h[item1], li_h[item2]]))] = support
print "\n=======================================================\n"
print "L2 = ", patterns
print "\n=======================================================\n"
output_file.write(createFormattedPatterns(patterns, 2))
new_modulo_map = {}
index_id = 1
actual_pattern_items = []
index_items_lookup = []
#patterns = {(2, 3, 5) : 1, (2, 3, 6) : 1, (2, 3, 7) : 1, (2, 4, 5) : 1, (2, 4, 7) : 1, (3, 5, 7) : 1}
for pattern in sorted(patterns.keys()):
if pattern[:-1] not in new_modulo_map:
new_modulo_map[pattern[:-1]] = index_id
prev_len = len(actual_pattern_items)
pattern_len = len(pattern[:-1])
actual_pattern_items += pattern[:-1]
index_items_lookup += [index_id, prev_len, pattern_len]
index_id += 1
if (pattern[-1],) not in new_modulo_map:
new_modulo_map[(pattern[-1],)] = index_id
prev_len = len(actual_pattern_items)
pattern_len = len([pattern[-1]])
actual_pattern_items += [pattern[-1]]
index_items_lookup += [index_id, prev_len, pattern_len]
index_id += 1
#print "Actual pattern items = ", actual_pattern_items
#print "Index lookup = ", index_items_lookup
print new_modulo_map
new_patterns = []
for pattern in patterns:
new_patterns.append((new_modulo_map[pattern[:-1]], new_modulo_map[(pattern[-1],)]))
print new_patterns
new_new_pattern = []
for pattern in new_patterns:
new_new_pattern.append(pattern[0] * 10 ** power + pattern[1])
new_new_pattern.sort()
print new_new_pattern
k = len(new_new_pattern)
li_h = np.array(new_new_pattern, dtype=np.int32)
ci_h = np.array([-1 for i in range(0, k ** 2)], dtype=np.int32)
ci_d = cuda.to_device(ci_h)
#li_h = np.array(sorted([randint(10, 99) for i in range(0, k)]), dtype=np.int32)
t1 = time()
li_d = cuda.to_device(li_h)
number_of_blocks = (int(ceil(k / (1.0 * MAX_ITEM_PER_SM))), 1)
selfJoinGPU [number_of_blocks, threads_per_block](li_d, ci_d, k, power)
li_d.copy_to_host(li_h)
ci_d.copy_to_host(ci_h)
api_h = np.array(actual_pattern_items, dtype=np.int32)
iil_h = np.array(index_items_lookup, dtype=np.int32)
api_d = cuda.to_device(api_h)
iil_d = cuda.to_device(iil_h)
print "Api_h = ", list(api_h), " Size = ", api_h.size
print "IIL_H = ", list(iil_h), " Size = ", iil_h.size
t2 = time()
print "LI_H = ", li_h
print "Initial Mask = ", ci_h.reshape(k, k)
#number_of_blocks = (1, 1) #(int(num_transactions / (1.0 * threads_per_block[0])) + 1, 1)
number_of_blocks = (int(ceil(num_transactions / (1.0 * MAX_TRANSACTIONS_PER_SM))), 1)
print "Num transactions = ", num_transactions
print "Num patterns = ", k
print "index = ", li_h
print "Size of api_d = ", api_h.size
print "Size of iil_h = ", iil_h.size
findHigherPatternFrequencyGPU [number_of_blocks, threads_per_block] (d_transactions, d_offsets, num_transactions, num_elements, li_d, ci_d, k, api_d, iil_d, power, api_h.size, iil_h.size)
cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Final Mask = ", ci_h.reshape(k, k)
#d_transactions.copy_to_host(transactions)
#print transactions[:num_elements]
threads_per_block = (BLOCK_SIZE, BLOCK_SIZE)
number_of_blocks = ((int(ceil(k / (1.0 * threads_per_block[0])))), (int(ceil(k / (1.0 * threads_per_block[0])))))
pruneMultipleGPU [number_of_blocks, threads_per_block] (ci_d, k, min_support)
ci_d.copy_to_host(ci_h)
print "Outer Mask = ", ci_h.reshape(k, k)
print "K = ", k
ci_hn = np.zeros(k, dtype=np.int32)
ci_dn = cuda.to_device(ci_hn)
combinationsAvailable [threads_per_block, number_of_blocks] (ci_d, ci_dn, k)
ci_dn.copy_to_host(ci_hn)
print "Ci_hn = ", list(ci_hn)
ci_hnx = np.empty(k, dtype=np.int32)
ci_dnx = cuda.to_device(ci_hnx)
preScan(ci_dnx, ci_dn, k)
ci_dnx.copy_to_host(ci_hnx)
num_patterns = ci_hnx[-1]
print list(ci_hnx)
sparseM_h = np.empty(ci_hnx[-1] * 3, dtype=np.uint32)
sparseM_d = cuda.to_device(sparseM_h)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(k / (1.0 * threads_per_block[0]))), 1)
print "K = ", k
convert2Sparse [threads_per_block, number_of_blocks] (ci_d, ci_dnx, sparseM_d, num_patterns, k)
sparseM_d.copy_to_host(sparseM_h)
# sparseM_h = sparseM_h.reshape(3, num_patterns)
print sparseM_h.reshape(3, num_patterns)
patterns = {}
for i in range(0, num_patterns):
item1 = sparseM_h[i]
item2 = sparseM_h[i + num_patterns]
support = sparseM_h[i + 2 * num_patterns]
patterns[tuple(sorted([li_h[item1], li_h[item2]]))] = support
print patterns
actual_patterns = {}
for pattern in patterns:
v_common_pat = pattern[0] / (10 ** power)
vitem1 = pattern[0] % (10 ** power)
vitem2 = pattern[1] % (10 ** power)
item1 = actual_pattern_items[index_items_lookup[(vitem1-1) * 3 + 1]]
item2 = actual_pattern_items[index_items_lookup[(vitem2-1) * 3 + 1]]
common_pat_start = index_items_lookup[(v_common_pat-1) * 3 + 1]
common_pat_length = index_items_lookup[(v_common_pat-1) * 3 + 2]
common_pat_end = common_pat_start + common_pat_length
common_pattern = actual_pattern_items[common_pat_start:common_pat_end]
pattern_key = tuple(common_pattern) + tuple(sorted([item1, item2]))
actual_patterns[pattern_key] = patterns[pattern]
print "\n=======================================================\n"
print "L3 = ", actual_patterns
print "\n=======================================================\n"
output_file.write(createFormattedPatterns(actual_patterns, 3))
output_file.close()
def test_sort():
in_h = np.empty(NUM_ELEMENTS, dtype=np.uint32) #4, 7, 2, 6, 3, 5, 1, 0
#in_h = np.array([4, 7, 2, 6, 3, 5, 1, 0], dtype=np.uint32)
out_h = np.empty(NUM_ELEMENTS, dtype=np.uint32)
for i in range(0, NUM_ELEMENTS):
in_h[i] = randint(0, 100)#NUM_ELEMENTS - i - 1
#in_h = np.array([6, 44, 71, 79, 94, 92, 12, 56, 47, 17, 81, 98, 84, 9, 85, 99], dtype=np.uint32)
#in_h = np.array([85, 37, 50, 73, 51, 46, 62, 84, 65, 99, 76, 59, 73, 16, 27, 4, 75, 81, 80, 33, 73, 11, 29, 24, 81, 49, 27, 71, 74, 64, 60, 91], dtype=np.uint32)
print in_h
in_d = cuda.to_device(in_h)
out_d = cuda.device_array(NUM_ELEMENTS, dtype=np.uint32)
tkg1 = time()
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(NUM_ELEMENTS / (2 * 1.0 * threads_per_block[0]))), 1)
RadixGPU [number_of_blocks, threads_per_block] (in_d, out_d, NUM_ELEMENTS)
out_d.copy_to_host(out_h)
#print "Rad = ", list(out_h)
stride = 4
# while stride < NUM_ELEMENTS:
# number_of_blocks = (int(ceil(NUM_ELEMENTS / (stride * 1.0 * threads_per_block[0]))), 1)
# bitonicSort [number_of_blocks, threads_per_block] (out_d, NUM_ELEMENTS, stride)
# stride *= 2
# # number_of_blocks = (int(ceil(NUM_ELEMENTS / (2 * 1.0 * threads_per_block[0]))), 1)
# # RadixGPU [number_of_blocks, threads_per_block] (out_d, in_d, NUM_ELEMENTS)
# # out_d = in_d
# out_d.copy_to_host(out_h)
# print "Str = ", list(out_h)
# break
# # stride /= 2
# while stride >= 4:
# number_of_blocks = (int(ceil(NUM_ELEMENTS / (stride * 1.0 * threads_per_block[0]))), 1)
# bitonicSort [number_of_blocks, threads_per_block] (out_d, NUM_ELEMENTS, stride)
# stride /= 2
# cuda.synchronize()
#
# number_of_blocks = (int(ceil(NUM_ELEMENTS / (2 * 1.0 * threads_per_block[0]))), 1)
# RadixGPU [number_of_blocks, threads_per_block] (out_d, in_d, NUM_ELEMENTS)
# out_d = in_d
#
# out_d.copy_to_host(out_h)
# cuda.synchronize()
#
# line = ""
# for i in range(0, NUM_ELEMENTS):
# line += " " + str(out_h[i])
#
# print line
tkg2 = time()
out_d.copy_to_host(out_h)
cuda.synchronize()
#print "GPU = ", list(out_h)
# line = ""
# for i in range(0, NUM_ELEMENTS):
# line += " " + str(out_h[i])
#
# print line
in_cpu = list(in_h)#[NUM_ELEMENTS - i -1 for i in range(0, NUM_ELEMENTS)]
tc1 = time()
in_cpu.sort()
#print "CPU = ", in_cpu
tc2 = time()
print "GPU Time = ", tkg2 - tkg1
print "CPU Time = ", tc2 - tc1
print len(in_cpu)
def test_apriori():
output_file = open("apriori_out.txt", "w")
offsets, transactions, num_transactions, num_elements = readFile(
"syncthetic_data.txt")
print "Offset = ", offsets[:num_transactions]
print "transactions = ", transactions[:num_elements]
print "Num transactions = ", num_transactions
print "Num elements = ", num_elements
min_support = MIN_SUPPORT
# to find number of max digits required to represent that many number of unique items
power = 1
while MAX_UNIQUE_ITEMS / (10**power) != 0:
power += 1
print "Power = ", power
t = [item for item in transactions.tolist()]
if num_elements > NUM_ELEMENTS:
print "Error: Elements exceeding NUM_ELEMENTS. Exiting..."
sys.exit(12)
input_h = np.array(t, dtype=np.int32)
print "Input transactions = ", list(input_h)
print "Size of transactions = ", input_h.size
ci_h = np.zeros(MAX_UNIQUE_ITEMS, dtype=np.int32)
li_h = np.empty(MAX_UNIQUE_ITEMS, dtype=np.int32)
input_d = cuda.to_device(input_h)
ci_d = cuda.to_device(ci_h)
li_d = cuda.device_array(MAX_UNIQUE_ITEMS, dtype=np.int32)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(NUM_ELEMENTS / (1.0 * threads_per_block[0]))),
1) #((NUM_ELEMENTS / threads_per_block[0]) + 1, 1)
histogramGPU[number_of_blocks, threads_per_block](input_d, ci_d,
num_elements)
#cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Ci_H Histogram result = ", ci_h # support count for each item
number_of_blocks = (int(
ceil(MAX_UNIQUE_ITEMS / (1.0 * threads_per_block[0]))), 1)
pruneGPU[number_of_blocks, threads_per_block](ci_d, MAX_UNIQUE_ITEMS,
min_support)
cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Keys = ", [i for i in range(0, len(ci_h))]
print "Ci_H Pruning result = ", ci_h # support count for each item
# calculate concise list of items satisfying min support
l1_patterns = {}
k = 0 # number of items whose sup_count > min_support
for j in range(0, len(ci_h)):
if ci_h[j] != 0:
li_h[k] = j
l1_patterns[(j, )] = ci_h[j]
k += 1
print "\n=======================================================\n"
print "L1 = ", list(li_h)[:k] #items whose support_count > min_support
print "\n=======================================================\n"
output_file.write(createFormattedPatterns(l1_patterns, 1))
print "K(num_items_with_good_sup_count = ", k
#k = 102
ci_h = np.array([-1 for i in range(0, k**2)], dtype=np.int32)
ci_d = cuda.to_device(ci_h)
#li_h = np.array(sorted([randint(10, 99) for i in range(0, k)]), dtype=np.int32)
#tli_h = np.array([i for i in range(1, k + 1)], dtype=np.int32)
t1 = time()
li_d = cuda.to_device(li_h)
number_of_blocks = (int(ceil(k / (1.0 * MAX_ITEM_PER_SM))), 1)
print "Self join 2 number of blocks = ", number_of_blocks
print "K = ", k
print "Ci_H size = ", ci_h.size
print "LI_H size = ", li_h.size
selfJoinGPU[number_of_blocks, threads_per_block](li_d, ci_d, k, power)
cuda.synchronize()
li_d.copy_to_host(li_h)
ci_d.copy_to_host(ci_h)
t2 = time()
#sys.exit(0)
# f = open('join.txt', 'w')
#
# for i in range(0, k):
# line = ""
# for j in range(0, k):
# line += str(ci_h[k * i + j]) + " "
# f.write(line + "\n")
#
# f.close()
#ci_h = ci_h.reshape(k, k)
print "Initial Mask = ", ci_h.reshape(k, k)
print "Self joining time = ", (t2 - t1)
d_offsets = cuda.to_device(offsets)
d_transactions = cuda.to_device(transactions)
#number_of_blocks = (1, 1) #(int(num_transactions / (1.0 * threads_per_block[0])) + 1, 1)
number_of_blocks = (int(
ceil(num_transactions / (1.0 * MAX_TRANSACTIONS_PER_SM))), 1)
print "Num blocks for findFrequency = ", number_of_blocks
print "Num transactions = ", num_transactions
print "Num patterns = ", k
print "index = ", list(li_h)[:k]
findFrequencyGPU[number_of_blocks,
threads_per_block](d_transactions, d_offsets,
num_transactions, num_elements, li_d,
ci_d, k)
cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Final Mask = ", ci_h.reshape(k, k)
d_transactions.copy_to_host(transactions)
threads_per_block = (BLOCK_SIZE, BLOCK_SIZE)
number_of_blocks = ((int(ceil(k / (1.0 * threads_per_block[0])))),
(int(ceil(k / (1.0 * threads_per_block[0])))))
pruneMultipleGPU[number_of_blocks, threads_per_block](
ci_d, k, min_support) # prunes according to min_support
ci_d.copy_to_host(ci_h)
print "Outer Mask = ", ci_h.reshape(k, k)
ci_hn = np.zeros(k, dtype=np.int32)
ci_dn = cuda.to_device(ci_hn)
combinationsAvailable[threads_per_block, number_of_blocks](
ci_d, ci_dn, k) #Number of possible patterns in each row
ci_dn.copy_to_host(ci_hn)
print "Ci_hn = ", list(ci_hn)
ci_hnx = np.empty(k, dtype=np.int32)
ci_dnx = cuda.to_device(ci_hnx)
preScan(ci_dnx, ci_dn, k) # Prefix sum on patterns in each row
ci_dnx.copy_to_host(ci_hnx)
num_patterns = ci_hnx[-1]
print "Ci_hnx = ", list(ci_hnx)
sparseM_h = np.empty(ci_hnx[-1] * 3, dtype=np.uint32)
sparseM_d = cuda.to_device(sparseM_h)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(k / (1.0 * threads_per_block[0]))), 1)
convert2Sparse[threads_per_block,
number_of_blocks](ci_d, ci_dnx, sparseM_d, num_patterns, k)
sparseM_d.copy_to_host(sparseM_h)
# sparseM_h = sparseM_h.reshape(3, num_patterns)
print sparseM_h.reshape(3, num_patterns)
patterns = {}
for i in range(0, num_patterns):
item1 = sparseM_h[i]
item2 = sparseM_h[i + num_patterns]
support = sparseM_h[i + 2 * num_patterns]
patterns[tuple(sorted([li_h[item1], li_h[item2]]))] = support
print "\n=======================================================\n"
print "L2 = ", patterns
print "\n=======================================================\n"
output_file.write(createFormattedPatterns(patterns, 2))
new_modulo_map = {}
index_id = 1
actual_pattern_items = []
index_items_lookup = []
#patterns = {(2, 3, 5) : 1, (2, 3, 6) : 1, (2, 3, 7) : 1, (2, 4, 5) : 1, (2, 4, 7) : 1, (3, 5, 7) : 1}
for pattern in sorted(patterns.keys()):
if pattern[:-1] not in new_modulo_map:
new_modulo_map[pattern[:-1]] = index_id
prev_len = len(actual_pattern_items)
pattern_len = len(pattern[:-1])
actual_pattern_items += pattern[:-1]
index_items_lookup += [index_id, prev_len, pattern_len]
index_id += 1
if (pattern[-1], ) not in new_modulo_map:
new_modulo_map[(pattern[-1], )] = index_id
prev_len = len(actual_pattern_items)
pattern_len = len([pattern[-1]])
actual_pattern_items += [pattern[-1]]
index_items_lookup += [index_id, prev_len, pattern_len]
index_id += 1
#print "Actual pattern items = ", actual_pattern_items
#print "Index lookup = ", index_items_lookup
print new_modulo_map
new_patterns = []
for pattern in patterns:
new_patterns.append(
(new_modulo_map[pattern[:-1]], new_modulo_map[(pattern[-1], )]))
print new_patterns
new_new_pattern = []
for pattern in new_patterns:
new_new_pattern.append(pattern[0] * 10**power + pattern[1])
new_new_pattern.sort()
print new_new_pattern
k = len(new_new_pattern)
li_h = np.array(new_new_pattern, dtype=np.int32)
ci_h = np.array([-1 for i in range(0, k**2)], dtype=np.int32)
ci_d = cuda.to_device(ci_h)
#li_h = np.array(sorted([randint(10, 99) for i in range(0, k)]), dtype=np.int32)
t1 = time()
li_d = cuda.to_device(li_h)
number_of_blocks = (int(ceil(k / (1.0 * MAX_ITEM_PER_SM))), 1)
selfJoinGPU[number_of_blocks, threads_per_block](li_d, ci_d, k, power)
li_d.copy_to_host(li_h)
ci_d.copy_to_host(ci_h)
api_h = np.array(actual_pattern_items, dtype=np.int32)
iil_h = np.array(index_items_lookup, dtype=np.int32)
api_d = cuda.to_device(api_h)
iil_d = cuda.to_device(iil_h)
print "Api_h = ", list(api_h), " Size = ", api_h.size
print "IIL_H = ", list(iil_h), " Size = ", iil_h.size
t2 = time()
print "LI_H = ", li_h
print "Initial Mask = ", ci_h.reshape(k, k)
#number_of_blocks = (1, 1) #(int(num_transactions / (1.0 * threads_per_block[0])) + 1, 1)
number_of_blocks = (int(
ceil(num_transactions / (1.0 * MAX_TRANSACTIONS_PER_SM))), 1)
print "Num transactions = ", num_transactions
print "Num patterns = ", k
print "index = ", li_h
print "Size of api_d = ", api_h.size
print "Size of iil_h = ", iil_h.size
findHigherPatternFrequencyGPU[number_of_blocks,
threads_per_block](d_transactions, d_offsets,
num_transactions,
num_elements, li_d, ci_d,
k, api_d, iil_d, power,
api_h.size, iil_h.size)
cuda.synchronize()
ci_d.copy_to_host(ci_h)
print "Final Mask = ", ci_h.reshape(k, k)
#d_transactions.copy_to_host(transactions)
#print transactions[:num_elements]
threads_per_block = (BLOCK_SIZE, BLOCK_SIZE)
number_of_blocks = ((int(ceil(k / (1.0 * threads_per_block[0])))),
(int(ceil(k / (1.0 * threads_per_block[0])))))
pruneMultipleGPU[number_of_blocks, threads_per_block](ci_d, k, min_support)
ci_d.copy_to_host(ci_h)
print "Outer Mask = ", ci_h.reshape(k, k)
print "K = ", k
ci_hn = np.zeros(k, dtype=np.int32)
ci_dn = cuda.to_device(ci_hn)
combinationsAvailable[threads_per_block, number_of_blocks](ci_d, ci_dn, k)
ci_dn.copy_to_host(ci_hn)
print "Ci_hn = ", list(ci_hn)
ci_hnx = np.empty(k, dtype=np.int32)
ci_dnx = cuda.to_device(ci_hnx)
preScan(ci_dnx, ci_dn, k)
ci_dnx.copy_to_host(ci_hnx)
num_patterns = ci_hnx[-1]
print list(ci_hnx)
sparseM_h = np.empty(ci_hnx[-1] * 3, dtype=np.uint32)
sparseM_d = cuda.to_device(sparseM_h)
threads_per_block = (BLOCK_SIZE, 1)
number_of_blocks = (int(ceil(k / (1.0 * threads_per_block[0]))), 1)
print "K = ", k
convert2Sparse[threads_per_block,
number_of_blocks](ci_d, ci_dnx, sparseM_d, num_patterns, k)
sparseM_d.copy_to_host(sparseM_h)
# sparseM_h = sparseM_h.reshape(3, num_patterns)
print sparseM_h.reshape(3, num_patterns)
patterns = {}
for i in range(0, num_patterns):
item1 = sparseM_h[i]
item2 = sparseM_h[i + num_patterns]
support = sparseM_h[i + 2 * num_patterns]
patterns[tuple(sorted([li_h[item1], li_h[item2]]))] = support
print patterns
actual_patterns = {}
for pattern in patterns:
v_common_pat = pattern[0] / (10**power)
vitem1 = pattern[0] % (10**power)
vitem2 = pattern[1] % (10**power)
item1 = actual_pattern_items[index_items_lookup[(vitem1 - 1) * 3 + 1]]
item2 = actual_pattern_items[index_items_lookup[(vitem2 - 1) * 3 + 1]]
common_pat_start = index_items_lookup[(v_common_pat - 1) * 3 + 1]
common_pat_length = index_items_lookup[(v_common_pat - 1) * 3 + 2]
common_pat_end = common_pat_start + common_pat_length
common_pattern = actual_pattern_items[common_pat_start:common_pat_end]
pattern_key = tuple(common_pattern) + tuple(sorted([item1, item2]))
actual_patterns[pattern_key] = patterns[pattern]
print "\n=======================================================\n"
print "L3 = ", actual_patterns
print "\n=======================================================\n"
output_file.write(createFormattedPatterns(actual_patterns, 3))
output_file.close()
