Skip to content

X
x393
Commit 50aca841 authored by Andrey Filippov's avatar Andrey Filippov
Browse files
Options

Made LMA work with numpy libraries

parent 6989d804
Show whitespace changes
Inline Side-by-side
Showing with 513 additions and 26 deletions
py393/x393_lma.py
View file @ 50aca841
......@@ -28,7 +28,8 @@ __version__ = "3.0+"
__maintainer__ = "Andrey Filippov"
__email__ = "andrey@elphel.com"
__status__ = "Development"
import math
import numpy as np
"""
For each byte lane:
tSDQS delay ps/step (~1/5 of datasheet value) - 1
......@@ -78,7 +79,9 @@ PARAMETER_TYPES=(
{"name":"tDQS", "size":1, "units":"ps","description":"DQS delay (not adjusted)","en":0},
{"name":"tDQ", "size":8, "units":"ps","description":"DQi delay","en":1},
{"name":"tFDQS", "size":4, "units":"ps","description":"DQS fine delays (mod 5)","en":1}, #only 4 are independent, 5-th is -sum of 4
{"name":"tFDQ", "size":32,"units":"ps","description":"DQ fine delays (mod 5)","en":1})
{"name":"tFDQ", "size":32, "units":"ps","description":"DQ fine delays (mod 5)","en":1},
{"name":"anaScale","size":1, "dflt":20, "units":"ps","description":"Scale for non-binary measured results","en":0},
)
FINE_STEPS=5
DLY_STEPS =FINE_STEPS * 32 # =160
def make_repeat(value,nRep):
......@@ -88,9 +91,36 @@ def make_repeat(value,nRep):
return (value,)*nRep
class X393LMA(object):
lambdas={"initial":0.1,"current":0.1,"max":100.0}
maxNumSteps=25
finalDiffRMS=0.0001
parameters=None
parameterMask=None
# parameterMask={}
parameterMask={'tSDQS': True,
'tSDQ': [True, True, True, True, True, True, True, True],
'tDQSHL': True,
'tDQHL': [True, True, True, True, True, True, True, True],
'tDQS': False,
'tDQ': [True, True, True, True, True, True, True, True],
'tFDQS': [True, True, True, True],
'tFDQ': [True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True],
'anaScale': False
}
"""
parameterMask={'tSDQS': True,
'tSDQ': [True, True, True, True, True, True, True, True],
'tDQSHL': True, # False, # True,
'tDQHL': [True, True, True, True, True, True, True, True], # False, # [True, True, True, True, True, True, True, True], #OK
'tDQS': False,
'tDQ': [True, True, True, True, True, True, True, True], #BAD - without it 0 in JTbyJ for tFDQ
'tFDQS': [True, True, True, True], # False, # [True, True, True, True], # OK
'tFDQ': True, # False, # [True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True, True],
'anaScale': False
}
"""
parameterVector=None
clk_period=None
analog_scale=20 # ps when there is analog result -0.5...+0.5, multiply it by analog_scale and add to result
# hist_estimated=None # DQ/DQS delay period,
# # DQ-DQS shift (and number of periods later) for averaged and individual bits,
# # for each of 4 edge types
......@@ -102,10 +132,16 @@ class X393LMA(object):
data_set,
periods=None):
n=len(data_set)*32
y=[0]*n
w=[0]*n
# fx=np.zeros((DLY_STEPS*32,))
"""
use np.nan instead of the None data
np.isnan() test
, dtype=np.float
"""
y=np.zeros((n,), dtype=np.int) #[0]*n
w=np.zeros((n,)) #[0]*n
if not periods is None:
p=[0]*n
p=np.zeros((n), dtype=np.int)#[0]*n
for dly,data in enumerate(data_set):
if data:
data_lane=data[lane*8:(lane+1)*8]
......@@ -152,11 +188,64 @@ class X393LMA(object):
print("?",end=" ")
print()
def normalizeParameters(self,
parameters,
isMask=False):
"""
Convert single/lists as needed
"""
if parameters is None:
parameters = self.parameters
for par in PARAMETER_TYPES:
name=par['name']
size=par["size"]
try:
v=parameters[name]
except:
if isMask:
v=par['en']
else:
try:
v=par['dflt']
except:
raise Exception("parameter['%s'] is not defined and PARAMETER_TYPES['%s'] does not provide default value"%(name,name))
if size == 1:
if isinstance(v,(list,tuple)):
v=v[0]
if isMask:
if v:
v=True
else:
v=False
else:
if isinstance(v,tuple):
v=list(v)
elif not isinstance(v,list):
v=[v]*size
if isMask:
for i in range(size):
if v[i]:
v[i]=True
else:
v[i]=False
parameters[name]=v
return parameters
def copyParameters(self,
parameters):
newPars={}
for k,v in parameters.items():
if isinstance(v,(list,tuple)):
newPars[k]=list(v)
else:
newPars[k]=v
return newPars
def createParameterVector(self,
parameters=None,
parameterMask=None):
global PARAMETER_TYPES
# global PARAMETER_TYPES
if parameters is None:
parameters = self.parameters
if parameterMask is None:
......@@ -181,7 +270,49 @@ class X393LMA(object):
for m,p in zip(mask, parVal):
if m:
vector.append(p)
return vector
return np.array(vector)
def createParameterIndex(self,
parameters=None,
parameterMask=None):
"""
create dict as parameters, but instead of values - index in the parameter vector, or -1
"""
if parameters is None:
parameters = self.parameters
if parameterMask is None:
parameterMask = self.parameterMask
indices={}
parIndex=0
for par in PARAMETER_TYPES:
name=par['name']
size=par["size"]
if par['en']:
try:
mask=parameterMask[name]
except:
mask=True
if mask:
if size==1:
indices[name]=parIndex
parIndex += 1
else:
if not isinstance(mask,(list,tuple)):
mask=[mask]*size
indices[name]=[]
for m in mask:
if m:
indices[name].append(parIndex)
parIndex += 1
else:
indices[name].append(-1)
if not name in indices:
if size==1:
indices[name]=-1
else:
indices[name]=[-1]*size
indices['numPars']=parIndex # extra key with total number of parameters
return indices
def getParametersFromVector(self,
vector=None,
......@@ -543,6 +674,8 @@ class X393LMA(object):
@data_set measured data set
@quiet reduce output
"""
self.clk_period=clk_period
hist_estimated=self.estimate_from_histograms(lane, # byte lane
bin_size,
clk_period,
......@@ -550,6 +683,7 @@ class X393LMA(object):
primary_set,
data_set,
quiet)
if quiet < 3:
print ("hist_estimated=%s"%(str(hist_estimated)))
data_periods_map=self.get_periods_map(lane,
data_set,
......@@ -560,8 +694,10 @@ class X393LMA(object):
data_set,
data_periods_map)
# print("ywp=%s"%(str(ywp)))
if quiet < 2:
print("\nY-vector:")
self.showYOrVector(ywp)
if quiet < 2:
print("\nperiods map:")
self.showYOrVector(ywp,ywp['p'])
......@@ -591,9 +727,93 @@ class X393LMA(object):
"tDQS": 0.0,
"tDQ": tDQ,
"tFDQS": (0.0,)*4,
"tFDQ": (0.0,)*32
"tFDQ": (0.0,)*32#,
# "anaScale":self.analog_scale
}
print ("parameters=%s"%(str(parameters)))
self.normalizeParameters(parameters) #isMask=False)
print ("normalized parameters=%s"%(str(parameters)))
"""
both ways work:
self.parameterMask={}
self.normalizeParameters(self.parameterMask,isMask=True)
and
"""
# self.parameterMask=self.normalizeParameters({},isMask=True)
self.parameterMask=self.normalizeParameters(self.parameterMask,isMask=True)
print ("parameters mask=%s"%(str(self.parameterMask)))
create_jacobian=True
fxj= self.createFxAndJacobian(parameters,
ywp, # keep in self.variable?
primary_set,
jacobian=create_jacobian,
parMask=None,
quiet=1)
if create_jacobian:
fx=fxj['fx']
else:
fx=fxj
if quiet < 2:
print("\nfx:")
self.showYOrVector(ywp,fx)
SX=0.0
SX2=0.0
S0=0.0
for d,w in zip(fx,ywp['w']):
if w>0:
S0+=w
SX+=w*d
SX2+=w*d*d
avg= SX/S0
rms= math.sqrt(SX2/S0)
print ("average(fx)= %fps, rms(fx)=%fps"%(avg,rms))
jByJT=np.dot(fxj['jacob'],np.transpose(fxj['jacob']))
if quiet < 3:
print("\njByJT:")
for i,l in enumerate(jByJT):
print ("%d"%(i),end=" ")
for d in l:
print ("%f"%(d),end=" ")
print()
self.lambdas ['current']=self.lambdas ['initial']
for _ in range(self.maxNumSteps):
OK,finished=self.LMA_step(parameters,
ywp, # keep in self.variable?
primary_set, # prima
None, # parMask= None,
self.lambdas,
self.finalDiffRMS,
quiet)
if OK:
print ("parameters=%s"%(str(parameters)))
if finished:
break
fx= self.createFxAndJacobian(parameters,
ywp, # keep in self.variable?
primary_set,
False,
parMask=None,
quiet=1)
if quiet < 3:
print("\nfx:")
self.showYOrVector(ywp,fx)
# print("delta=%s"%(str(delta)))
# for i,d in enumerate(delta):
# print ("%d %f"%(i,d))
"""
ir = ir0 - s/4 + d/4 # ir - convert to ps from steps
if = if0 + s/4 - d/4
......@@ -603,6 +823,273 @@ class X393LMA(object):
(s+d)/2=of-or
s=if-ir+of-or
d=ir-if+of-or
"""
def createFxAndJacobian(self,
parameters,
y_data, # keep in self.variable?
primary_set, # prima
jacobian=False, # create jacobian, False - only fx
parMask=None,
quiet=1):
def pythIsNone(obj):
return obj is None
isNone=pythIsNone # swithch to np.isnan
y_vector = y_data['y']
periods_vector=y_data['p']
period=self.clk_period
try:
y_fractions = y_data['f']
except:
y_fractions = None
try:
w_vector = y_data['w']
except:
w_vector = None
anaScale = parameters['anaScale']
if y_fractions is None:
anaScale = 0
elif isinstance(y_fractions,np.ndarray):
isNone=np.isnan
# fx=[0.0]*DLY_STEPS*32
fx=np.zeros((DLY_STEPS*32,))
#self.clk_period
tFDQS5=list(parameters['tFDQS'])
tFDQS5.append(-tFDQS5[0]-tFDQS5[1]-tFDQS5[2]-tFDQS5[3])
tFDQ=[]
for b in range(8):
tFDQi=list(parameters['tFDQ'][4*b:4*(b+1)])
tFDQi.append(-tFDQi[0]-tFDQi[1]-tFDQi[2]-tFDQi[3])
tFDQ.append(tFDQi)
tSDQS=parameters['tSDQS']
tSDQ= parameters['tSDQ'] # list
tDQS =parameters['tDQS']#single value
tDQ= parameters['tDQ'] # list
tDQSHL =parameters['tDQSHL']#single value
tDQHL= parameters['tDQHL'] # list
for dly in range(DLY_STEPS):
tdqs=dly * tSDQS - tDQS - tFDQS5[dly % FINE_STEPS] # t - time from DQS pad to internal DQS clock with zero setup/hold times to DQ FFs
tdqs_r = tdqs + 0.25 * tDQSHL # sign opposite from: ir = ir0 - s/4 + d/4; or = or0 - s/4 - d/4
tdqs_f = tdqs - 0.25 * tDQSHL # sign opposite from: if = if0 + s/4 - d/4; of = of0 + s/4 + d/4
tdqs_rf=(tdqs_r, tdqs_f)
#correct for DQS edge type
for b in range(8): # use all 4 variants
for t in range(4):
indx=32*dly+t*8+b
if (w_vector is None) or (w_vector[indx] > 0):
tdq=y_vector[indx] * tSDQ[b] - tDQ[b] - tFDQ[b][y_vector[indx] % FINE_STEPS]
# correct for periods
tdq -= period*periods_vector[indx] # or should it be minus here?
# correct for edge types
if (t == 0) or (t == 3):
tdq -= 0.25*tDQHL[b]
else:
tdq += 0.25*tDQHL[b]
if anaScale:
# if y_fractions[indx] is None:
if isNone(y_fractions[indx]):
tdq+=2.5
else:
tdq+=anaScale*y_fractions[indx]
if (t ^ primary_set) & 2:
tdq -= 0.5*period
fx[indx] = tdq - tdqs_rf[t & 1] # odd are falling DQS, even are rising DQS
if not jacobian:
return fx
if parMask is None:
parMask=self.normalizeParameters(self.parameterMask,isMask=True)
# pv= self.createParameterVector(parameters,parMask)
# numPars=len(pv)
# print("pv=%s"%(str(pv)))
parInd=self.createParameterIndex(parameters,parMask)
print("parInd=%s"%(str(parInd)))
numPars=parInd['numPars']
jacob=np.zeros((numPars,DLY_STEPS*32))
fineM5=((1.0, 0.0, 0.0, 0.0, -0.25),
(0.0, 1.0, 0.0, 0.0, -0.25),
(0.0, 0.0, 1.0, 0.0, -0.25),
(0.0, 0.0, 0.0, 1.0, -0.25))
dqs_finedelay_en=parInd['tFDQS']
for e in dqs_finedelay_en:
if e>=0:
break
else:
dqs_finedelay_en=None
dq_finedelay_en=[None]*8
for b in range(8):
dq_finedelay_en[b]=parInd['tFDQ'][4*b:4*(b+1)]
for e in dq_finedelay_en[b]:
if e>=0:
break
else:
dq_finedelay_en[b]=None
for dly in range(DLY_STEPS):
dlyMod5=dly % FINE_STEPS
dtdqs_dtSDQS = dly
dtdqs_dtDQS = -1.0
dtdqs_dtFDQS = (-fineM5[0][dlyMod5],-fineM5[1][dlyMod5],-fineM5[2][dlyMod5],-fineM5[3][dlyMod5])
dtdqs_dtDQSHL_rf=(0.25,-0.25)
#correct for DQS edge type
for b in range(8): # use all 4 variants
for t in range(4):
indx=32*dly+t*8+b
if (w_vector is None) or (w_vector[indx] > 0):
#dependencies of DQS delays
if parInd['tSDQS'] >= 0:
jacob[parInd['tSDQS'],indx]=-dtdqs_dtSDQS
if parInd['tDQS'] >= 0:
jacob[parInd['tDQS'],indx]=-dtdqs_dtDQS
if dqs_finedelay_en:
for i,pIndx in enumerate (dqs_finedelay_en):
if pIndx >= 0:
jacob[pIndx,indx]=-dtdqs_dtFDQS[i]
if parInd['tDQSHL'] >= 0:
jacob[parInd['tDQSHL'],indx]=-dtdqs_dtDQSHL_rf[t & 1]
#dependencies of DQ delays
# tdq=y_vector[indx] * tSDQ[b] - tDQ[b] - tFDQ[b][y_vector[indx] % FINE_STEPS]
if parInd['tSDQ'][b] >= 0:
jacob[parInd['tSDQ'][b],indx]=y_vector[indx]
if parInd['tDQ'][b] >= 0:
jacob[parInd['tDQ'][b],indx] = -1
if dq_finedelay_en[b]:
yMod5=y_vector[indx] % FINE_STEPS
dtdq_dtFDQ = (-fineM5[0][yMod5],-fineM5[1][yMod5],-fineM5[2][yMod5],-fineM5[3][yMod5])
for i,pIndx in enumerate (dq_finedelay_en[b]):
if pIndx >= 0:
jacob[pIndx,indx]=dtdq_dtFDQ[i]
if parInd['tDQHL'][b] >= 0:
if (t == 0) or (t == 3):
jacob[parInd['tDQHL'][b],indx]=-0.25
else:
jacob[parInd['tDQHL'][b],indx]=+0.25
if parInd['anaScale'] >= 0:
if anaScale and not isNone(y_fractions[indx]):
jacob[parInd['anaScale'],indx]=y_fractions[indx]
return {'fx':fx,'jacob':jacob}
def getParAvgRMS(self,
parameters,
ywp,
primary_set, # prima
quiet=1):
fx= self.createFxAndJacobian(parameters,
ywp, # keep in self.variable?
primary_set,
False, # jacobian
None,
quiet)
SX=0.0
SX2=0.0
S0=0.0
for d,w in zip(fx,ywp['w']):
if w>0:
S0+=w
SX+=w*d
SX2+=w*d*d
avg= SX/S0
rms= math.sqrt(SX2/S0)
return {"avg":avg,"rms":rms}
def LMA_step(self,
parameters,
ywp, # keep in self.variable?
primary_set, # prima
parMask,
lambdas, #single-element list to update value
finalDiffRMS,
quiet= 1):
parVector0=self.createParameterVector(parameters, parMask) # initial parameter vector
arms0 = self.getParAvgRMS(parameters,
ywp,
primary_set, # prima
quiet+1)
if quiet < 3:
print ("LMA_step <start>: average(fx)= %fps, rms(fx)=%fps"%(arms0['avg'],arms0['rms']))
delta=self.LMA_solve(parameters,
ywp, # keep in self.variable?
primary_set, # prima
parMask,
lambdas["current"],
quiet)
parVector= parVector0+delta
# print ("\nparVector0=%s"%(str(parVector0)))
# print ("\ndelta=%s"%(str(delta)))
# print ("\nparVector=%s"%(str(parVector)))
# newPars = {}.update(parameters) # so fixed parameters will appear in the newPars
newPars = self.copyParameters(parameters) # so fixed parameters will appear in the newPars
# newPars = self.getParametersFromVector(parVector,
if quiet < 2:
print ("\nparameters=%s"%(str(parameters)))
# print ("\n1: newPars=%s"%(str(newPars)))
self.getParametersFromVector(parVector,
parMask,
newPars) # parameters=None):# if not None, will be updated
if quiet < 2:
print ("\n2: newPars=%s"%(str(newPars)))
print ("\nparameters=%s"%(str(parameters)))
arms1 = self.getParAvgRMS(newPars,
ywp,
primary_set, # prima
quiet+1)
finished=False
if arms1['rms'] < arms0['rms']:
parameters.update(newPars)
lambdas["current"]*=.5
success=True
if (arms0['rms'] - arms1['rms']) < finalDiffRMS:
finished=True
else:
lambdas["current"]*=8.0
success=False
if lambdas["current"] > lambdas["max"]:
finished=True
if quiet < 3:
print ("LMA_step %s: average(fx)= %fps, rms(fx)=%fps, lambda=%f"%(('FAILURE','SUCCESS')[success],arms1['avg'],arms1['rms'],lambdas["current"]))
return (success,finished)
def LMA_solve(self,
parameters,
ywp, # keep in self.variable?
primary_set, # prima
parMask= None,
lmbda= 0.001,
quiet= 1):
fxj= self.createFxAndJacobian(parameters,
ywp, # keep in self.variable?
primary_set,
True, # jacobian
parMask,
quiet)
JT=np.transpose(fxj['jacob'])
jByJT=np.dot(fxj['jacob'],JT)
for i,_ in enumerate(jByJT):
jByJT[i,i] += lmbda*jByJT[i,i]
jByDiff= -np.dot(fxj['jacob'],fxj['fx'])
delta=np.linalg.solve(jByJT,jByDiff)
return delta
"""
ir = ir0 - s/4 + d/4 # ir - convert to ps from steps
if = if0 + s/4 - d/4
or = or0 - s/4 - d/4 # ir - convert to ps from steps
of = of0 + s/4 + d/4
(s-d)/2=if-ir
(s+d)/2=of-or
s=if-ir+of-or
d=ir-if+of-or
"""
\ No newline at end of file
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment