The VX function vxQueryGraph() does not seem to work well.

[Peng_X_]

When I use the VX function

vxQueryGraph(graph, VX_GRAPH_PERFORMANCE, &perf, sizeof(perf));

to get the performance of a graph, the result seems to be wrong. Is this because the Intel implementation doesn't use this function? Maybe  I should write my own function to get the performance?

[Peng_X_]

The problem has been solved. Unlike the implementation of AMD, Intel CV SDK disables the performance measurement by default (maybe to save a bit of time?). So we need to use 

vxDirective((vx_reference)context, VX_DIRECTIVE_ENABLE_PERFORMANCE);

to enable the measurement first. (If you want to measure the performance of a graph, you should use it before the execution of the graph. It you want  to measure the performance of a node, you'd better use it before you add the node to the graph.)

[Jeffrey_M_Intel1]

Glad you've found a solution.  By the way, VAD also has an increasing amount of features for measuring graph performance and trying performance experiments for optimization.