Well, some modules seemed actually to be slower on the GPU than on the CPU, even though the GPU was used when available… (esp. atrous).
So, you gain 10% on that particular image/history stack. What happens when you use other somewhat intensive modules (DorS, for instance, or capture sharpening in demosaic). There are situations where the tiling and/or memory transfers kill any gain from the parallel processing a GPU allows…
Plus - if you dedicate that shared/unified mem to OpenCL it won’t be used for dt system & CPU code reducing it’s performance 
You could measure OpenCL performance and calculate a “performance gain” and set that in the device specifing config though - but i completely doubt it’s worth that.
There is a subjective part in the appreciation as well, though: if a GPU (even a mediocre one) makes the editing feel more fluid (less of the “working” on the image), it may be worth it, even if the “objective” gain is limited.
But for me personally, 10% faster on a bulk export is not relevant, I usually can do something else in that time. Spending hours in trying to get OpenCL working in that situation just isn’t worth it for me.