def test_dot():
A1 = np.float32(np.random.rand(2,4))
A2 = np.float32(np.random.rand(4,2))
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.dot(B1,B2)
C = B3.tocpu()
t.assert_array_almost_equal(np.dot(A1,A2),C,4,"array.tocpu not equal to init array!")
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.empty((2,2))
gpu.dot(B1,B2,B3)
t.assert_array_almost_equal(np.dot(A1,A2),B3.tocpu(),4,"array.tocpu not equal to init array!")
def test_dot():
A1 = np.float32(np.random.rand(2, 4))
A2 = np.float32(np.random.rand(4, 2))
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.dot(B1, B2)
C = B3.tocpu()
t.assert_array_almost_equal(np.dot(A1, A2), C, 4,
"array.tocpu not equal to init array!")
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.empty((2, 2))
gpu.dot(B1, B2, B3)
t.assert_array_almost_equal(np.dot(A1, A2), B3.tocpu(), 4,
"array.tocpu not equal to init array!")
def test_timer():
if gpu.lib.pt_clusterNet == gpu.lib.pt_clusterNetCPU: return
t = gpu.Timer()
A = gpu.rand(100, 100)
B = gpu.rand(100, 100)
C = gpu.rand(100, 100)
time = 0
t.tick()
for i in range(10):
gpu.dot(A, B, C)
time = t.tock()
assert time > 0
time = 0
t.tick("Timer test")
gpu.dot(A, B, C)
time = t.tock("Timer test")
assert time > 0
accumulative_time = 0
for i in range(100):
t.tick('cumulative')
gpu.dot(A, B, C)
t.tick('cumulative')
accumulative_time = t.tock('cumulative')
assert accumulative_time > 5 * time
def test_timer():
if gpu.lib.pt_clusterNet == gpu.lib.pt_clusterNetCPU: return
t = gpu.Timer()
A = gpu.rand(100,100)
B = gpu.rand(100,100)
C = gpu.rand(100,100)
time = 0
t.tick()
for i in range(10):
gpu.dot(A,B,C)
time = t.tock()
assert time > 0
time = 0
t.tick("Timer test")
gpu.dot(A,B,C)
time = t.tock("Timer test")
assert time > 0
accumulative_time = 0
for i in range(100):
t.tick('cumulative')
gpu.dot(A,B,C)
t.tick('cumulative')
accumulative_time = t.tock('cumulative')
assert accumulative_time > 5*time
input = gpu.rand(dim_inner,dim1)
W = gpu.rand(dim_outer,dim_inner)
output = gpu.rand(dim_outer,dim1)
input2 = gpu2.random.rand(dim_inner,dim1)
W2 = gpu2.random.rand(dim_outer,dim_inner)
output2 = gpu2.random.rand(dim_outer,dim1)
mean_time = 0
for i in range(5):
iters = 100
#warmup
for j in range(1000):
if batch_first_mode:
gpu.dot(input,W,output)
else:
gpu.dot(W, input, output)
t.tick(str(dim_inner))
for j in range(iters):
if batch_first_mode:
gpu.dot(input,W,output)
else:
gpu.dot(W, input, output)
t.tick(str(dim_inner))
print t.tock(str(dim_inner))/5/iters
mean_time = 0
dim_inner = 32
dim_outer = 256
for i in range(1000):
dim_inner += 32
A = gpu.rand(dim1, dim_inner)
B = gpu.rand(dim_inner, dim_outer)
C = gpu.rand(dim1, dim_outer)
if dim_inner > 0: iters = 1000
if dim_inner > 100: iters = 100
if dim_inner > 1000: iters = 10
if dim_inner > 3000: iters = 4
#warmup
for j in range(2):
gpu.dot(A, B, C)
t.tick(str(dim_inner))
for j in range(iters):
gpu.dot(A, B, C)
sec = t.tock(str(dim_inner)) / 1000.
tilesA = (dim1 / 16) * ((dim_inner / 64) +
(1 if dim_inner % 64 > 0 else 0))
tilesB = ((dim_inner / 64) + (1 if dim_inner % 64 > 0 else 0)) * (
(dim_inner / 16) * ((dim_outer / 64) +
(1 if dim_outer % 64 > 0 else 0)))
memops = (tilesA + tilesB) * 16 * 64 + (dim_inner * dim_outer)
#print sec / (memops*iters)
#print (memops/sec)*4*(1024**-3)*iters
#print iters*(dim**3)/(sec*1000*1000*1000)
#print iters*(dim1*dim_inner*dim_outer)/(sec*1000*1000*1000)
else:
input = gpu.rand(dim_inner, dim1)
W = gpu.rand(dim_outer, dim_inner)
output = gpu.rand(dim_outer, dim1)
input2 = gpu2.random.rand(dim_inner, dim1)
W2 = gpu2.random.rand(dim_outer, dim_inner)
output2 = gpu2.random.rand(dim_outer, dim1)
mean_time = 0
for i in range(5):
iters = 100
#warmup
for j in range(1000):
if batch_first_mode:
gpu.dot(input, W, output)
else:
gpu.dot(W, input, output)
t.tick(str(dim_inner))
for j in range(iters):
if batch_first_mode:
gpu.dot(input, W, output)
else:
gpu.dot(W, input, output)
t.tick(str(dim_inner))
print t.tock(str(dim_inner)) / 5 / iters
mean_time = 0
for i in range(5):
iters = 100
dim_inner = 32
dim_outer = 256
for i in range(1000):
dim_inner += 32
A = gpu.rand(dim1,dim_inner)
B = gpu.rand(dim_inner,dim_outer)
C = gpu.rand(dim1,dim_outer)
if dim_inner > 0: iters = 1000
if dim_inner > 100: iters = 100
if dim_inner > 1000: iters = 10
if dim_inner > 3000: iters = 4
#warmup
for j in range(2):
gpu.dot(A,B,C)
t.tick(str(dim_inner))
for j in range(iters):
gpu.dot(A,B,C)
sec = t.tock(str(dim_inner))/1000.
tilesA = (dim1/16)*((dim_inner/64) + (1 if dim_inner % 64 > 0 else 0))
tilesB = ((dim_inner/64) + (1 if dim_inner % 64 > 0 else 0))*((dim_inner/16)*((dim_outer/64) + (1 if dim_outer % 64 > 0 else 0)))
memops = (tilesA+tilesB)*16*64 + (dim_inner*dim_outer)
#print sec / (memops*iters)
#print (memops/sec)*4*(1024**-3)*iters
#print iters*(dim**3)/(sec*1000*1000*1000)
#print iters*(dim1*dim_inner*dim_outer)/(sec*1000*1000*1000)
print iters*dim1*dim_inner*dim_outer/(6144.*1000*1000*1000)*24, sec
A2 = gpu.rand(dim1,dim_inner)
