class test_mean():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_mean(self):
"""Test mean function."""
a = self.rng.rand(129, 1025).astype(np.float32)
out_np = a.mean()
out_gp = self.gp.mean(a)
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(1025).astype(np.float32)
out_np = a.mean()
out_gp = self.gp.mean(a)
assert(np.allclose(out_np, out_gp))
def test_mean(self):
"""Test mean function along axes."""
a = self.rng.rand(129, 1025).astype(np.float32)
out_np = a.mean(axis=1)
out_gp = self.gp.mean(a, axis=1).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_np = a.mean(axis=0)
out_gp = self.gp.mean(a, axis=0).copy_to_host()
assert(np.allclose(out_np, out_gp))
class test_T():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_madd(self):
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(129, 1025).astype(np.float32)
out_np = a-b
out_gp = self.gp.sub(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp))
def test_madd_scale(self):
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(129, 1025).astype(np.float32)
alpha = .4
beta = -1.6
out_np = (alpha*a-beta*b).astype(np.float32)
out_gp = self.gp.sub(a,b,alpha = alpha, beta = beta).copy_to_host()
print out_np
print out_gp
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
def test_vadd(self):
a = self.rng.rand(1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
out_np = a-b
out_gp = self.gp.sub(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp))
def test_vadd_scale(self):
a = self.rng.rand(1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
alpha = .4
beta = -1.6
out_np = (alpha*a-beta*b).astype(np.float32)
out_gp = self.gp.sub(a,b,alpha = alpha, beta = beta).copy_to_host()
print out_np
print out_gp
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
class test_ones():
def setup(self):
self.gp = Gpupy()
def test_ones(self):
"""Create arrays of ones."""
out_np = np.ones(shape=(129,1025), dtype=np.float32)
out_gp = self.gp.ones(shape=(129,1025)).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_np = np.ones(shape=129, dtype=np.float32)
out_gp = self.gp.ones(shape=129).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_np = np.ones(129, dtype=np.float32)
out_gp = self.gp.ones(129).copy_to_host()
assert(np.allclose(out_np, out_gp))
class test_reshape():
def setup(self):
self.gp = Gpupy()
def test_reshape(self):
""" Reshapes array.
"""
a = np.array(np.arange(10), dtype=np.float32, order='F')
out_np = np.reshape(a, newshape=(2,5), order='F')
out_gp = self.gp.reshape(a, newshape=(2,5)).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = np.array(np.arange(10), dtype=np.float32, order='c')
out_np = np.reshape(a, newshape=(2,5), order='C')
out_gp = self.gp.reshape(a, newshape=(2,5), order='C').copy_to_host()
assert(np.allclose(out_np, out_gp))
class test_diag(): def setup(self): self.gp = Gpupy() self.rng = np.random.RandomState(0) def test_v2m(self): a = self.rng.rand(129).astype(np.float32) out_np = np.diag(a) out_gp = self.gp.diag(a).copy_to_host() assert(np.allclose(out_np, out_gp)) def test_m2v(self): a = self.rng.rand(129, 129).astype(np.float32) out_np = np.diag(a) out_gp = self.gp.diag(a).copy_to_host() assert(np.allclose(out_np, out_gp))
class test_zero_diag:
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_zero_diag(self):
a = self.rng.rand(129, 129).astype(np.float32)
out_np = a - np.diag(np.diag(a))
out_gp = self.gp.zero_diag(a).copy_to_host()
assert np.allclose(out_np, out_gp)
class test_relu():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_clip_m(self):
"""clip on matrices."""
a = self.rng.rand(129, 1025).astype(np.float32)
out_np = np.clip(a,-.1,.1)
out_gp = self.gp.clip(a,-.1,.1).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129, 1025).astype(np.float32)
out_np = np.clip(a,-.1,.1)
out_gp = cuda.to_device(a)
self.gp.clip(a, -.1, .1, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_clip_v(self):
"""clip on vectors."""
a = self.rng.rand(1025).astype(np.float32)
out_np = np.clip(a,-.1,.1)
out_gp = self.gp.clip(a,-.1,.1).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(1025).astype(np.float32)
out_np = np.clip(a,-.1,.1)
out_gp = cuda.to_device(a)
self.gp.clip(a, -.1, .1, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
class test_mult():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_matrix_multiply(self):
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(129, 1025).astype(np.float32)
out_np = a*b
out_gp = self.gp.mult(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp))
def test_vector_multiply(self):
a = self.rng.rand(1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
out_np = a*b
out_gp = self.gp.mult(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp))
class test_scal():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_mscal(self):
a = self.rng.rand(129, 1025).astype(np.float32)
alpha = 2.
out_np = (alpha*a).astype(np.float32)
out_gp = self.gp.scal(a, alpha).copy_to_host()
assert(np.allclose(out_np, out_gp))
def test_vscal(self):
a = self.rng.rand(1025).astype(np.float32)
alpha = 2.
out_np = (alpha*a).astype(np.float32)
out_gp = self.gp.scal(a, alpha).copy_to_host()
assert(np.allclose(out_np, out_gp))
class test_zeros():
def setup(self):
self.gp = Gpupy()
def test_zeros(self):
"""Create array of zeros."""
out_np = np.zeros(shape=(129,1025), dtype=np.float32)
out_gp = self.gp.zeros(shape=(129,1025)).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_np = np.zeros(shape=(129,1025), dtype=np.float32)
out_gp = cuda.to_device(np.ones(shape=(129,1025), dtype=np.float32))
self.gp.zeros(shape=(129,1025), out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
out_np = np.zeros(shape=129, dtype=np.float32)
out_gp = self.gp.zeros(shape=129).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_np = np.zeros(129, dtype=np.float32)
out_gp = cuda.to_device(np.ones(129, dtype=np.float32))
self.gp.zeros(shape=129, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def setup(self):
self.gp = Gpupy()
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
class test_add():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_madd(self):
"""Add two matrices."""
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(129, 1025).astype(np.float32)
out_np = a+b
out_gp = self.gp.add(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_gp = self.gp.zeros(shape=(129, 1025))
self.gp.add(a,b, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_madd_scale(self):
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(129, 1025).astype(np.float32)
alpha = .4
beta = -1.6
out_np = (alpha*a+beta*b).astype(np.float32)
out_gp = self.gp.add(a, b, alpha = alpha, beta = beta).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=(129, 1025))
self.gp.add(a,b, alpha = alpha, beta = beta, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host(), atol=1.e-5))
def test_vadd(self):
a = self.rng.rand(1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
out_np = a+b
out_gp = self.gp.add(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_gp = self.gp.zeros(shape=1025)
self.gp.add(a,b, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_vadd_scale(self):
a = self.rng.rand(1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
alpha = .4
beta = -1.6
out_np = (alpha*a+beta*b).astype(np.float32)
out_gp = self.gp.add(a, b, alpha = alpha, beta = beta).copy_to_host()
assert(np.allclose(out_np, out_gp, atol=1.e-5))
out_gp = self.gp.zeros(shape=1025)
self.gp.add(a, b, alpha = alpha, beta = beta, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_ms_scale(self):
a = self.rng.rand(129,1025).astype(np.float32)
b = 1.
alpha = .4
beta = -1.6
out_np = (alpha*a+beta*b).astype(np.float32)
out_gp = self.gp.add(a,b,alpha = alpha, beta = beta).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=(129, 1025))
self.gp.add(a, b, alpha = alpha, beta = beta, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host(), atol = 1.e-5))
def test_ms(self):
a = self.rng.rand(129,1025).astype(np.float32)
b = 1.
out_np = (a+b).astype(np.float32)
out_gp = self.gp.add(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=(129, 1025))
self.gp.add(a,b, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host(), atol = 1.e-5))
def test_vs_scale(self):
a = self.rng.rand(1025).astype(np.float32)
b = 1.
alpha = .4
beta = -1.6
out_np = (alpha*a+beta*b).astype(np.float32)
out_gp = self.gp.add(a,b,alpha = alpha, beta = beta).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=1025)
self.gp.add(a, b, alpha = alpha, beta = beta, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host(), atol = 1.e-5))
def test_vs(self):
""" Add a vector and a scalar."""
a = self.rng.rand(1025).astype(np.float32)
b = 1.
out_np = (a+b).astype(np.float32)
out_gp = self.gp.add(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=1025)
self.gp.add(a,b, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_mv_scale(self):
""" Add a matrix to a vector and scale the result."""
a = self.rng.rand(129,1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
alpha = .4
beta = -1.6
out_np = (alpha*a+beta*b).astype(np.float32)
out_gp = self.gp.add(a, b, alpha = alpha, beta = beta).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=(129, 1025))
self.gp.add(a, b, alpha = alpha, beta = beta, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host(), atol = 1.e-5))
def test_mv(self):
""" Add a matrix to a vector."""
a = self.rng.rand(129,1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
out_np = (a+b).astype(np.float32)
out_gp = self.gp.add(a,b).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=(129, 1025))
self.gp.add(a,b, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_ma_scale(self):
""" Add a matrix to a vector with np.newaxis and scale the result."""
a = self.rng.rand(1025, 1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
alpha = .4
beta = -1.6
out_np = (alpha*a+beta*b[np.newaxis,:]).astype(np.float32)
out_gp = self.gp.add(a,b[np.newaxis,:],alpha = alpha, beta = beta).copy_to_host()
assert(np.allclose(out_np, out_gp, atol=1.e-5))
out_gp = self.gp.zeros(shape=(1025, 1025))
self.gp.add(a, b[np.newaxis,:], alpha = alpha, beta = beta, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host(), atol=1.e-5))
out_np = (alpha*a+beta*b[:,np.newaxis]).astype(np.float32)
out_gp = self.gp.add(a, b[:,np.newaxis], alpha = alpha, beta = beta).copy_to_host()
assert(np.allclose(out_np, out_gp, atol=1.e-5))
out_gp = self.gp.zeros(shape=(1025, 1025))
self.gp.add(a, b[:,np.newaxis], alpha = alpha, beta = beta, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host(), atol=1.e-5))
def test_ma(self):
"""Add a matrix to a vector with a np.newaxis."""
a = self.rng.rand(1025, 1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
out_np = (a+b[np.newaxis,:]).astype(np.float32)
out_gp = self.gp.add(a,b[np.newaxis,:]).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=(1025, 1025))
self.gp.add(a, b[np.newaxis,:], out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
out_np = (a+b[:,np.newaxis]).astype(np.float32)
out_gp = self.gp.add(a,b[:,np.newaxis]).copy_to_host()
assert(np.allclose(out_np, out_gp, atol = 1.e-5))
out_gp = self.gp.zeros(shape=(1025, 1025))
self.gp.add(a,b[:,np.newaxis], out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
class test_relu():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_relu_m(self):
"""relu on matrices."""
a = self.rng.rand(129, 1025).astype(np.float32)
t = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = 0.
out_gp = self.gp.relu(a, thresh=t).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129, 1025).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = val
out_gp = self.gp.relu(a, thresh=t, set_val=val).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129, 1025).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = val
out_gp = cuda.to_device(a)
self.gp.relu(a, thresh=t, set_val=val, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_relu_m_v(self):
"""relu on matrices with vectors for thresholds."""
a = self.rng.rand(129, 1025).astype(np.float32)
t = self.rng.rand(1025).astype(np.float32)
out_np = copy.deepcopy(a)
out_np[out_np<t] = 0.
out_gp = self.gp.relu(a, thresh=t).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129, 1025).astype(np.float32)
t = self.rng.rand(129).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = val
out_gp = self.gp.relu(a, thresh=t, set_val=val).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129, 1025).astype(np.float32)
t = self.rng.rand(129).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = val
out_gp = cuda.to_device(a)
self.gp.relu(a, thresh=t, set_val=val, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_relu_m_v(self):
"""relu on matrices with newaxis vectors for thresholds."""
a = self.rng.rand(129, 1025).astype(np.float32)
t = self.rng.rand(1025).astype(np.float32)
out_np = copy.deepcopy(a)
out_np[out_np<t[np.newaxis,:]] = 0.
out_gp = self.gp.relu(a, thresh=t[np.newaxis,:]).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129, 1025).astype(np.float32)
t = self.rng.rand(129).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t[:,np.newaxis]] = val
out_gp = self.gp.relu(a, thresh=t[:,np.newaxis], set_val=val).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129, 1025).astype(np.float32)
t = self.rng.rand(129).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t[:,np.newaxis]] = val
out_gp = cuda.to_device(a)
self.gp.relu(a, thresh=t[:,np.newaxis], set_val=val, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_relu_v(self):
"""relu on vectors"""
a = self.rng.rand(129).astype(np.float32)
t = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = 0.
out_gp = self.gp.relu(a, thresh=t).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = val
out_gp = self.gp.relu(a, thresh=t, set_val=val).copy_to_host()
assert(np.allclose(out_np, out_gp))
a = self.rng.rand(129).astype(np.float32)
val = .5
out_np = copy.deepcopy(a)
out_np[out_np<t] = val
out_gp = cuda.to_device(a)
self.gp.relu(a, thresh=t, set_val=val, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
class test_dot():
def setup(self):
self.gp = Gpupy()
self.rng = np.random.RandomState(0)
def test_gemm(self):
""" Test matrix-matrix dot product."""
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(1025, 257).astype(np.float32)
out_np = np.dot(a, b)
out_gp = self.gp.dot(a, b).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_gp = self.gp.zeros(shape=(129, 257))
self.gp.dot(a, b, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_gemv(self):
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(1025).astype(np.float32)
out_np = np.dot(a, b)
out_gp = self.gp.dot(a, b).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_gp = self.gp.zeros(shape=129)
self.gp.dot(a, b, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_gevm(self):
a = self.rng.rand(129, 1025).astype(np.float32)
b = self.rng.rand(129).astype(np.float32)
out_np = np.dot(b,a)
out_gp = self.gp.dot(b,a).copy_to_host()
assert(np.allclose(out_np, out_gp))
out_gp = self.gp.zeros(shape=1025)
self.gp.dot(b, a, out=out_gp)
assert(np.allclose(out_np, out_gp.copy_to_host()))
def test_vvdot(self):
a = self.rng.rand(129).astype(np.float32)
b = self.rng.rand(129).astype(np.float32)
out_np = np.dot(a,b)
out_gp = self.gp.dot(a,b)
assert(np.allclose(out_np, out_gp))