def test_idcst2Random(self):
torch.manual_seed(10)
M = 4
N = 8
x = torch.empty(M, N, dtype=torch.int32).random_(0, 10).double()
print("2D x")
print(x)
golden_value = discrete_spectral_transform.idcst2(x).data.numpy()
print("2D golden_value")
print(golden_value)
# test cpu
#pdb.set_trace()
custom = dct.IDCST2()
dst_value = custom.forward(x)
print("dxt_value")
print(dst_value.data.numpy())
np.testing.assert_allclose(dst_value.data.numpy(),
golden_value,
atol=1e-14)
# test cpu
#pdb.set_trace()
custom = dct_lee.IDCST2()
dst_value = custom.forward(x)
print("dxt_value")
print(dst_value.data.numpy())
np.testing.assert_allclose(dst_value.data.numpy(),
golden_value,
atol=1e-14)
# test gpu
if torch.cuda.device_count():
custom = dct.IDCST2()
dst_value = custom.forward(x.cuda()).cpu()
print("dxt_value cuda")
print(dst_value.data.numpy())
np.testing.assert_allclose(dst_value.data.numpy(),
golden_value,
atol=1e-14)
# test gpu
if torch.cuda.device_count():
custom = dct_lee.IDCST2()
dst_value = custom.forward(x.cuda()).cpu()
print("dxt_value cuda")
print(dst_value.data.numpy())
np.testing.assert_allclose(dst_value.data.numpy(),
golden_value,
atol=1e-14)
def eval_runtime():
#x = torch.tensor([1, 2, 7, 9, 20, 31], dtype=torch.float64)
#print(dct_N(x))
N = 512
runs = 10
x = torch.empty(10, N, N, dtype=torch.float64).uniform_(0, 10.0).cuda()
perm = discrete_spectral_transform.get_perm(N, dtype=torch.int64, device=x.device)
expk = discrete_spectral_transform.get_expk(N, dtype=x.dtype, device=x.device)
#x_numpy = x.data.cpu().numpy()
#tt = time.time()
#for i in range(runs):
# y = fftpack.dct(fftpack.dct(x_numpy[i%10].T, norm=None).T, norm=None)
#print("scipy takes %.3f sec" % (time.time()-tt))
## 9s for 200 iterations 1024x1024 on GTX 1080
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_2N = dct2_2N(x[0], expk0=expk, expk1=expk)
#torch.cuda.synchronize()
##print(prof)
#print("dct_2N takes %.3f ms" % ((time.time()-tt)/runs*1000))
## 11s for 200 iterations 1024x1024 on GTX 1080
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_N = discrete_spectral_transform.dct2_N(x[i%10], perm0=perm, expk0=expk, perm1=perm, expk1=expk)
#torch.cuda.synchronize()
##print(prof)
#print("dct_N takes %.3f ms" % ((time.time()-tt)/runs*1000))
#dct2func = dct.DCT2(expk, expk, algorithm='2N')
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_N = dct2func.forward(x[0])
#torch.cuda.synchronize()
##print(prof)
#print("DCT2Function 2N takes %.3f ms" % ((time.time()-tt)/runs*1000))
#dct2func = dct.DCT2(expk, expk, algorithm='N')
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_N = dct2func.forward(x[0])
#torch.cuda.synchronize()
##print(prof)
#print("DCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000))
#dct2func = dct_lee.DCT2(expk, expk)
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_N = dct2func.forward(x[i%10])
#torch.cuda.synchronize()
##print(prof)
#print("DCT2Function lee takes %.3f ms" % ((time.time()-tt)/runs*1000))
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_N = discrete_spectral_transform.idct2_2N(x[i%10], expk0=expk, expk1=expk)
#torch.cuda.synchronize()
##print(prof)
#print("idct2_2N takes %.3f ms" % ((time.time()-tt)/runs*1000))
idct2func = dct.IDCT2(expk, expk, algorithm='2N')
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = idct2func.forward(x[i%10])
torch.cuda.synchronize()
#print(prof)
print("IDCT2Function 2N takes %.3f ms" % ((time.time()-tt)/runs*1000))
idct2func = dct.IDCT2(expk, expk, algorithm='N')
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = idct2func.forward(x[i%10])
torch.cuda.synchronize()
#print(prof)
print("IDCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000))
exit()
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_N = discrete_spectral_transform.idxt(x[i%10], 0, expk=expk)
#torch.cuda.synchronize()
##print(prof)
#print("idxt takes %.3f ms" % ((time.time()-tt)/runs*1000))
#idxct_func = dct.IDXCT(expk)
#torch.cuda.synchronize()
#tt = time.time()
##with torch.autograd.profiler.profile(use_cuda=True) as prof:
#for i in range(runs):
# y_N = idxct_func.forward(x[i%10])
#torch.cuda.synchronize()
##print(prof)
#print("IDXCTFunction takes %.3f ms" % ((time.time()-tt)/runs*1000))
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = torch.rfft(x[i%10].view([1, N, N]), signal_ndim=2, onesided=False)
torch.cuda.synchronize()
#print(prof)
print("fft2 takes %.3f ms" % ((time.time()-tt)/runs*1000))
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = discrete_spectral_transform.idcct2(x[i%10], expk_0=expk, expk_1=expk)
torch.cuda.synchronize()
#print(prof)
print("idcct2 takes %.3f ms" % ((time.time()-tt)/runs*1000))
func = dct.IDCCT2(expk, expk)
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = func.forward(x[i%10])
torch.cuda.synchronize()
#print(prof)
print("IDCCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000))
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = discrete_spectral_transform.idcst2(x[i%10], expk_0=expk, expk_1=expk)
torch.cuda.synchronize()
#print(prof)
print("idcst2 takes %.3f ms" % ((time.time()-tt)/runs*1000))
func = dct.IDCST2(expk, expk)
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = func.forward(x[i%10])
torch.cuda.synchronize()
#print(prof)
print("IDCST2Function takes %.3f ms" % ((time.time()-tt)/runs*1000))
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = discrete_spectral_transform.idsct2(x[i%10], expk_0=expk, expk_1=expk)
torch.cuda.synchronize()
#print(prof)
print("idsct2 takes %.3f ms" % ((time.time()-tt)/runs*1000))
func = dct.IDSCT2(expk, expk)
torch.cuda.synchronize()
tt = time.time()
#with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = func.forward(x[i%10])
torch.cuda.synchronize()
#print(prof)
print("IDSCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000))
print(ratio)
# compute field xi
auv_by_wu2_plus_wv2_wu = auv.mul(wu_by_wu2_plus_wv2)
ratio = auv_by_wu2_plus_wv2_wu.numpy() / ex_in_map
print("auv_by_wu2_plus_wv2_wu / ex_in_map")
print(ratio)
field_map_x = discrete_spectral_transform.idsct2(auv_by_wu2_plus_wv2_wu, expk_M, expk_N)
ratio = field_map_x.numpy() / ex_out_map
print("field_map_x/ex_out_map")
print(ratio)
pdb.set_trace()
auv_by_wu2_plus_wv2_wv = auv.mul(wv_by_wu2_plus_wv2)
ratio = auv_by_wu2_plus_wv2_wv.numpy() / ey_in_map
print("auv_by_wu2_plus_wv2_wv / ey_in_map")
print(ratio)
field_map_y = discrete_spectral_transform.idcst2(auv_by_wu2_plus_wv2_wv, expk_M, expk_N)
ratio = field_map_y.numpy() / ey_out_map
print("field_map_y/ey_out_map")
print(ratio)
pdb.set_trace()
exit()
"""
my_auv_map = discrete_spectral_transform.dct2(torch.from_numpy(density_map)) * (1.0/density_map.size)
my_auv_map[0, :].mul_(0.5)
my_auv_map[:, 0].mul_(0.5)
ratio = my_auv_map.numpy() / auv_map
print("my_auv_map/auv_map")
print(ratio.min())
print(ratio.max())
def eval_others(x, expk0, expk1, expkM, expkN, runs):
y_N = discrete_spectral_transform.idcct2(x, expk_0=expk0, expk_1=expk1)
torch.cuda.synchronize()
tt = time.time()
# with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = discrete_spectral_transform.idcct2(x, expk_0=expk0, expk_1=expk1)
torch.cuda.synchronize()
# print(prof)
print("idcct2 takes %.7f ms" % ((time.time()-tt)/runs*1000))
func = dct.IDCCT2(expk0, expk1)
y_N = func.forward(x)
torch.cuda.synchronize()
tt = time.time()
# with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = func.forward(x)
torch.cuda.synchronize()
# print(prof)
print("IDCCT2 Function takes %.7f ms" % ((time.time()-tt)/runs*1000))
y_N = discrete_spectral_transform.idcst2(x, expk_0=expk0, expk_1=expk1)
torch.cuda.synchronize()
tt = time.time()
# with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = discrete_spectral_transform.idcst2(x, expk_0=expk0, expk_1=expk1)
torch.cuda.synchronize()
# print(prof)
print("idcst2 takes %.7f ms" % ((time.time()-tt)/runs*1000))
func = dct.IDCST2(expk0, expk1)
y_N = func.forward(x)
torch.cuda.synchronize()
tt = time.time()
# with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = func.forward(x)
torch.cuda.synchronize()
# print(prof)
print("IDCST2 Function takes %.7f ms" % ((time.time()-tt)/runs*1000))
y_N = discrete_spectral_transform.idsct2(x, expk_0=expk0, expk_1=expk1)
torch.cuda.synchronize()
tt = time.time()
# with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = discrete_spectral_transform.idsct2(x, expk_0=expk0, expk_1=expk1)
torch.cuda.synchronize()
# print(prof)
print("idsct2 takes %.7f ms" % ((time.time()-tt)/runs*1000))
func = dct.IDSCT2(expk0, expk1)
y_N = func.forward(x)
torch.cuda.synchronize()
tt = time.time()
# with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(runs):
y_N = func.forward(x)
torch.cuda.synchronize()
# print(prof)
print("IDSCT2 Function takes %.7f ms" % ((time.time()-tt)/runs*1000))
print("")