We present a new near-infrared imaging survey in the four CFHTLS deep fields: the WIRCam Deep Survey (WIRDS). WIRDS comprises extremely deep, high quality (FWHM ~0.6) J, H and K imaging covering a total effective area of 2.1 deg^2 and reaching AB 50%
completeness limits of ~24.5. We combine our images with the CFHTLS to create a unique eight-band ugrizJHK photometric catalogues in the CFHTLS deep fields; these four separate fields allow us to make a robust estimate of the effect of cosmic variance for all our measurements. We use these catalogues to estimate precise photometric redshifts, galaxy types and stellar masses for a unique sample of ~1.8 million galaxies. Our JHK number counts are consistent with previous studies. We apply the BzK selection to our gzK filter set and find that the star forming BzK selection successfully selects 76% of star-forming galaxies in the redshift range 1.4<z<2.5 in our photometric catalogue. The passive BzK selection returns 52% of the passive 1.4<z<2.5 population identified in the photometric catalogue. We present the galaxy stellar mass function as a function of redshift up to z=2 and present fits using double Schechter functions. A mass-dependent evolution of the mass function is seen with the numbers of galaxies with masses of log(M)<10.75 still evolving at z<1, but galaxies of higher mass reaching their present day numbers by z~0.8-1. This is consistent with the present picture of downsizing in galaxy evolution. We compare our results with the predictions of the GALFORM semi-analytical galaxy formation model and find that the simulations provide a relatively successful fit to the observed mass functions at intermediate masses (i.e. 10<log(M)<11). However, the GALFORM results under-predict the mass function at low masses, whilst the fit as a whole degrades beyond redshifts of z~1.2.
