def test_bin_op(f):
for s in [(1,), (2,2), (128,128)]:
an = np.random.randn(*s)
bn = np.random.randn(*s)
ap = ph.Tensor(an).cuda()
bp = ph.Tensor(bn).cuda()
cp = f(ap, bp).cpu().np()
torch.testing.assert_allclose(cp, f(an, bn))
def test_broadcast():
for s in [(1,), (2,2), (128,128)]:
an = np.random.randn(1)
liken = np.random.randn(*s)
ap = ph.Tensor(an).cuda()
likep = ph.Tensor(liken).cuda()
cp = ap.broadcast_like(likep).cpu().np()
torch.testing.assert_allclose(cp, np.broadcast_to(an, s))
def f3(a, b):
return (a * b + a) * a + b
def f4(a, b):
c = a + b
for _ in range(100):
c = a * c
return c
funcs = [f0, f1, f2, f3, f4]
for func in funcs:
a = ph.Tensor(a_)
b = ph.Tensor(b_)
c = func(a, b)
g = c.grad(a)
a_grad = g(ph.Tensor(ones)).np()
a = torch.tensor(a_)
a.requires_grad = True
b = torch.tensor(b_)
c = func(a, b)
c.backward(torch.tensor(ones))
torch.testing.assert_allclose(a_grad, a.grad.numpy())
for _ in range(4):
a_np = np.random.randn(_, _).astype(np.float32)
import hma
import pyhma as ph
import numpy as np
hma.set_debug(True)
for A in [2, 5, 16, 27]:
for B in [2, 5, 16, 27]:
for C in [2, 5, 16, 27]:
anp = np.random.randn(A, B).astype(np.float32)
bnp = np.random.randn(B, C).astype(np.float32)
a = ph.Tensor(anp).cuda()
b = ph.Tensor(bnp).cuda()
c = ph.Tensor(hma.mm_nn([a.cTensor, b.cTensor])[0])
torch.testing.assert_allclose(c.cpu().np(), (anp) @ (bnp))
anp = np.random.randn(A, B).astype(np.float32)
bnp = np.random.randn(C, B).astype(np.float32)
a = ph.Tensor(anp).cuda()
b = ph.Tensor(bnp).cuda()
c = ph.Tensor(hma.mm_nt([a.cTensor, b.cTensor])[0])
torch.testing.assert_allclose(c.cpu().np(), (anp) @ (bnp.T))
anp = np.random.randn(B, A).astype(np.float32)
bnp = np.random.randn(B, C).astype(np.float32)
a = ph.Tensor(anp).cuda()
b = ph.Tensor(bnp).cuda()
c = ph.Tensor(hma.mm_tn([a.cTensor, b.cTensor])[0])
def test_sum():
for s in [(1,), (2,2), (128,128)]:
an = np.random.randn(*s)
ap = ph.Tensor(an).cuda()
cp = ap.sum().cpu().np()
torch.testing.assert_allclose(cp, an.sum())
def test_unary_op(f, *args):
for s in [(1,), (2,2), (128,128)]:
an = np.random.randn(*s)
ap = ph.Tensor(an).cuda()
cp = f(ap, *args).cpu().np()
torch.testing.assert_allclose(cp, f(an))
device = torch.device('cuda', 0)
N = 64
C_i = 128
C_o = 32
xnp = np.random.randn(N, C_i).astype(np.float32)
wnp = np.random.randn(C_o, C_i).astype(np.float32)
snp = np.random.randn(N, C_o).astype(np.float32)
xt = torch.tensor(xnp).to(device)
wt = torch.tensor(wnp).to(device)
st = torch.tensor(snp).to(device)
yt = torch.nn.functional.linear(xt, wt)
x = ph.Tensor(xnp).cuda()
w = ph.Tensor(wnp).cuda()
s = ph.Tensor(snp).cuda()
y = ph.Tensor(hma.mm_nt([x.cTensor, w.cTensor])[0])
ynp = y.cpu().np()
torch.testing.assert_allclose(y.cpu().np(), yt.cpu())
print("passed.")
t = time.perf_counter()
for i in range(10000):
_ = torch.nn.functional.linear(xt, wt)
print(_.cpu().numpy().shape)
print(time.perf_counter() - t)
import numpy as np import hma import pyhma as ph a = np.random.randn(16) a_ = ph.Tensor(a) b = hma.to_cuda([a_.cTensor])[0] c = hma.mul([b, b])[0] d = hma.to_cpu([c])[0] assert np.allclose(ph.Tensor(d).np(), a * a) a = np.random.randn(16) b = np.random.randn(16) c_ref = a * b a = ph.Tensor(a).cuda() b = ph.Tensor(b).cuda() c = a * b assert np.allclose(c.cpu().np(), c_ref) _ = np.random.randn(128) k = ph.Tensor(_).cuda().sum() print(k.cpu().np(), _.sum())
import numpy as np
import hma
import pyhma as ph
import torch
import time
iters = 10000
a = np.random.randn(1,1)
a_ph = ph.Tensor(a)
a_t = torch.tensor(a)
h2 = ph.Tensor(np.array(2.0)).broadcast_like(a_ph)
for _ in range(iters):
_ = (a_t + a_t) / 2
t = time.time()
for _ in range(iters):
a_t = (a_t + a_t) / 2
print(a_t.numpy())
print("PT: ", time.time() - t)
t = time.time()
for _ in range(iters):
a_ph = (a_ph + a_ph) / 2
print(a_ph.np())
print("PH: ", time.time() - t)
size = 128
iters = 20000
print()