CUDNN Backpropagation

I write a code to learn something with Cudnn, but the backpropagation returns wired numbers(first 5 and last 5 in each pointer):

Forward:
Conv, MaxPool, Conv, MaxPool, Fully, Activ, Lrn, Fully, Activ, Sfw,

Data Output 69.00000 69.00000 69.00000 69.00000 69.00000 ||91.00000 91.00000 91.00000 91.00000 91.00000

Conv Output 85.73157 76.21236 61.58315 68.90920 61.47983 ||-8.27115 -6.75207 -7.97190 -8.39558 -7.92991

Pool Output 85.73157 68.90920 61.47983 54.82409 49.50154 ||15.73470 22.46315 26.53646 25.11255 26.56086

Conv Output -109.32211 -100.58031 -105.15018 -91.54110 -93.69791 ||25.38896 18.39327 9.84402 3.03636 -3.78350

Pool Output -100.58031 -91.54110 -91.15739 -88.66731 -92.70054 ||9.56151 5.09123 0.82925 -10.66174 -15.90880

Fully Output -2.84714 -5.92057 -2.43803 -0.17245 3.85035 ||-2.42728 4.99564 3.96398 -4.13377 -0.18600

Acti Output 0.05483 0.00268 0.08032 0.45699 0.97917 ||0.08112 0.99328 0.98137 0.01577 0.45363

LRN Output 0.05483 0.00268 0.08032 0.45698 0.97914 ||0.08111 0.99323 0.98133 0.01577 0.45363

Fully Output -0.11432 -0.00142 0.20275 -0.14892 0.12301 ||0.23554 0.19915 -0.16288 -0.05834 0.11955

Acti Output 0.47145 0.49965 0.55051 0.46284 0.53071 ||0.55861 0.54962 0.45937 0.48542 0.52985

SfwMax Output 0.09617 0.09892 0.10409 0.09535 0.10204 ||0.10493 0.10399 0.09502 0.09753 0.10196

Predict 5.00000 , 1.00000

Backward:


Sfw, Activ, Fully, Lrn, Activ, Fully, Pool, Conv, Pool, Conv,

Lbl Grad 0.09617 -0.90108 0.10409 0.09535 0.10204 ||0.10493 0.10399 0.09502 0.09753 0.10196

SfwMax Grad 0.09617 -0.90108 0.10409 0.09535 0.10204 ||0.10493 0.10399 0.09502 0.09753 0.10196

Acti Grad 0.02396 -0.22527 0.02576 0.02371 0.02541 ||0.02587 0.02574 0.02360 0.02436 0.02540

Fully Grad -0.00536 0.00395 -0.00198 0.00224 -0.00172 ||0.00310 -0.00525 0.00264 -0.00210 -0.00412

LRN Grad -0.00536 0.00395 -0.00198 0.00224 -0.00172 ||0.00310 -0.00525 0.00264 -0.00210 -0.00412

Acti Grad -0.00028 0.00001 -0.00015 0.00056 -0.00004 ||0.00023 -0.00004 0.00005 -0.00003 -0.00102

Fully Grad -0.00005 0.00003 -0.00004 0.00006 0.00005 ||0.00000 0.00000 0.00000 0.00000 0.00000

Pool Grad 0.00000 -0.00005 0.00000 0.00003 0.00000 ||0.00000 0.00000 0.00000 0.00000 0.00000

Conv Grad 0.00000 0.00000 -0.00000 -0.00001 0.00000 ||0.00000 0.00000 0.00000 0.00000 0.00000

Pool Grad 0.00000 0.00000 0.00000 0.00000 -0.00000 ||0.00000 0.00000 0.00000 0.00000 0.00000

Does anybody have an idea about the mistake?
If the Code helps:

Mostly revelant files are Network.cpp, Propagation.cu and Layer.cpp.