In MKL manual this functionality is described as working with square matrices:
The mkl_?csrgemv routine performs a matrix-vector operation defined as
y := A*x
y := A'*x,
where: x and y are vectors, A is an m-by-m sparse square matrix in the CSR format (3-array variation), A' is the transpose of A.
So for rectangular matrices it works improperly in some cases.