def test_gpus_many_map():
N = 2048
Nproj = N
Ns = 2048
filter_type = 'None'
cor = N//2
interp_type = 'cubic'
# init random array
R = np.float32(np.sin(np.arange(0, Ns*Nproj*N)/float(Ns*Nproj)))
R = np.reshape(R, [Ns, Nproj, N])
# input parameters
tomo = R
reg_par = 0.001 # *np.max(tomo)
num_iter = 100
recon = np.zeros([Ns, N, N], dtype="float32")+1e-3
method = "em"
gpu_list = [0, 1, 2, 3]
# list of available methods for reconstruction
lpmethods_list = {
'fbp': lpmethods.fbp,
'grad': lpmethods.grad,
'cg': lpmethods.cg,
'tv': lpmethods.tv,
'em': lpmethods.em
}
ngpus = len(gpu_list)
# number of slices for simultaneous processing by 1 gpu
# (depends on gpu memory size, chosen for gpus with >= 8GB memory)
Nssimgpu = min(int(pow(2, 26)/float(N*N)), int(np.ceil(Ns/float(ngpus))))
tic()
# class lprec
lp = lpTransform.lpTransform(
N, Nproj, Nssimgpu, filter_type, cor, interp_type)
# if not fbp, precompute for the forward transform
lp.precompute(method != 'fbp')
print("Init time %f" % (toc()))
# list of slices sets for simultaneous processing b gpus
ids_list = [None]*int(np.ceil(Ns/float(Nssimgpu)))
for k in range(0, len(ids_list)):
ids_list[k] = range(k*Nssimgpu, min(Ns, (k+1)*Nssimgpu))
tic()
# init memory for each gpu
for igpu in range(0, ngpus):
gpu = gpu_list[igpu]
# if not fbp, allocate memory for the forward transform arrays
lp.initcmem(method != 'fbp', gpu)
# run reconstruciton on many gpus
with cf.ThreadPoolExecutor(ngpus) as e:
shift = 0
for reconi in e.map(partial(lpmultigpu, lp, lpmethods_list[method], recon, tomo, num_iter, reg_par, gpu_list), ids_list):
recon[np.arange(0, reconi.shape[0])+shift] = reconi
shift += reconi.shape[0]
print("Rec time %f" % (toc()))
def task_gen_photo_imagepath(root, query):
print('Get photo ids.');
db_helper = DBHelper();
db_helper.init(root);
photo_dao = PhotoDao(db_helper);
tic();
photo_ids = photo_dao.getClassPhotoIds(query, ''.join([query]));
toc();
print('Get photo path.');
imagepaths = get_photo_imagepath(root, query, photo_ids)
output_path = ''.join([db_helper.datasetDir, '/', query, '_imagepath.txt']);
file_io = FileIO();
file_io.write_strings_to_file(imagepaths, output_path);
h01 = psi1
h02 = psi2
for k in range(niter):
# registration
tic()
flow = dslv.registration_flow_batch(psi1,
data,
mmin,
mmax,
flow.copy(),
pars,
nproc=42)
t1 = toc()
tic()
# deformation subproblem
psi1 = dslv.cg_deform(data,
psi1,
flow,
4,
tslv.fwd_lam(u, theta) + lamd1 / rho1,
rho1,
nproc=42)
t2 = toc()
psi2 = tslv.solve_reg(u, lamd2, rho2, alpha)
# tomo subproblem
tic()
# u, psi, lamd, flow = interpdense(u,psi,lamd,flow)
# optical flow parameters
pars = [0.5, 1, w[il], 4, 5, 1.1, 4]
with tc.SolverTomo(theta, ntheta, nz, ne, pnz,
center / pow(2, binning), ngpus) as tslv:
with dc.SolverDeform(ntheta, nz, n, ptheta) as dslv:
rho = 0.5
h0 = psi
for k in range(niter[il]):
# registration
# print(np.linalg.norm(psi-data))
tic()
flow = dslv.registration_flow_batch(
psi, data, mmin, mmax, flow.copy(), pars, 40)
print(toc())
# Tpsi = dslv.apply_flow_gpu_batch(psi, flow)
# print(np.linalg.norm(Tpsi-data))
# deformation subproblem
tic()
psi = dslv.cg_deform_gpu_batch(
data, psi, flow, 4,
unpad(tslv.fwd_tomo_batch(u), ne, n) + lamd / rho, rho)
print(toc())
# tomo subproblem
tic()
u = tslv.cg_tomo_batch(pad(psi - lamd / rho, ne, n), u, 4)
print(toc())
h = unpad(tslv.fwd_tomo_batch(u), ne, n)
def streaming():
"""
Main computational function, take data from pvdata ('2bmbSP1:Pva1:Image'),
reconstruct orthogonal slices and write the result to pvrec ('AdImage')
"""
##### init pvs ######
# init ca pvs
chscanDelta = pva.Channel('2bma:PSOFly2:scanDelta', pva.CA)
chrotangle = pva.Channel('2bma:m82', pva.CA)
chrotangleset = pva.Channel('2bma:m82.SET', pva.CA)
chrotanglestop = pva.Channel('2bma:m82.STOP', pva.CA)
chStreamX = pva.Channel('2bmS1:StreamX', pva.CA)
chStreamY = pva.Channel('2bmS1:StreamY', pva.CA)
chStreamZ = pva.Channel('2bmS1:StreamZ', pva.CA)
# init pva streaming pv for the detector
chdata = pva.Channel('2bmbSP1:Pva1:Image')
pvdata = chdata.get('')
# init pva streaming pv for reconstrucion with coping dictionary from pvdata
pvdict = pvdata.getStructureDict()
pvrec = pva.PvObject(pvdict)
# take dimensions
n = pvdata['dimension'][0]['size']
nz = pvdata['dimension'][1]['size']
# set dimensions for reconstruction
pvrec['dimension'] = [{
'size': 3 * n,
'fullSize': 3 * n,
'binning': 1
}, {
'size': n,
'fullSize': n,
'binning': 1
}]
##### run server for reconstruction pv #####
s = pva.PvaServer('AdImage', pvrec)
##### procedures before running fly #######
# 0) form circular buffer, whenever the angle goes higher than 180
# than corresponding projection is replacing the first one
scanDelta = chscanDelta.get('')['value']
ntheta = np.int(180 / scanDelta + 0.5)
databuffer = np.zeros([ntheta, nz * n], dtype='uint8')
thetabuffer = np.zeros(ntheta, dtype='float32')
# 1) stop rotation, replace rotation stage angle to a value in [0,360)
chrotanglestop.put(1)
time.sleep(3)
rotangle = chrotangle.get('')['value']
chrotangleset.put(1)
chrotangle.put(rotangle - rotangle // 360 * 360)
chrotangleset.put(0)
# 2) take flat field
flat = takeflat(chdata)
firstid = chdata.get('')['uniqueId']
# 3) create solver class on GPU, and copy flat field to gpu
slv = OrthoRec(ntheta, n, nz)
slv.set_flat(flat)
# 4) allocate memory for result slices
recall = np.zeros([n, 3 * n], dtype='float32')
# 5) start monitoring the detector pv for data collection
def addProjection(pv):
with mrwlock.w_locked():
curid = pv['uniqueId']
databuffer[np.mod(curid, ntheta)] = pv['value'][0]['ubyteValue']
thetabuffer[np.mod(curid, ntheta)] = (curid - firstid) * scanDelta
#print(firstid, curid)
chdata.monitor(addProjection, '')
##### start acquisition #######
start_fly()
##### streaming reconstruction ######
while (True): # infinite loop over angular partitions
with mrwlock.r_locked(): # lock buffer before reading
datap = databuffer.copy()
thetap = thetabuffer.copy()
# take 3 ortho slices ids
idx = chStreamX.get('')['value']
idy = chStreamY.get('')['value']
idz = chStreamZ.get('')['value']
# reconstruct on GPU
tic()
recx, recy, recz = slv.rec_ortho(datap, thetap * np.pi / 180, n // 2,
idx, idy, idz)
print('rec time:', toc())
# concatenate (supposing nz<n)
recall[:nz, :n] = recx
recall[:nz, n:2 * n] = recy
recall[:, 2 * n:] = recz
# write to pv
pvrec['value'] = ({'floatValue': recall.flatten()}, )
# reconstruction rate limit
time.sleep(0.1)