Adobe has recently enhanced Photoshop CC to accelerate certain filters via OpenCL. AMD contacted NitroWare with this information and claims of 11-fold performance increases with "Smart Sharpen" on Kaveri, specifically. The computer hardware site decided to test these claims on a Radeon HD 7850 using the test metrics that AMD provided them.
Sure enough, he noticed a 16-fold gain in performance. Without OpenCL, the filter's loading bar was on screen for over ten seconds; with it enabled, there was no bar.
Dominic from NitroWare is careful to note that an HD 7850 is significantly higher performance than an APU (barring some weird scenario involving memory transfers or something). This might mark the beginning of Adobe's road to sensible heterogeneous computing outside of video transcoding. Of course, this will also be exciting for AMD. While they cannot keep up with Intel, thread per thread, they are still a heavyweight in terms of total performance. With Photoshop, people might actually notice it.
Cool, I will test it with an
Cool, I will test it with an nvidia based GPU, and see if the performance results are similar.
edit: It performs about the same, it does not use much GPU usage, so likely after a certain point, the GPU simply stops adding performance.(GTX 760 only clocks up to about 1000MHz and uses about 70-78% GPU usage, applying a smart filter to noise on a 30000×30000 image with a total processing time of 30.3 seconds. At lower resolutions, the GPU usage barely spikes as it all takes place in less than a second)
The CPU usage also never maxes out, so I wonder what the bottleneck is that is stopping the GPU using its full power
Using the same settings from the video, I got a processing time on the smart sharped, of 0.8 seconds
System specs:
CPU: Phenom II x6 1075t
RAM 12GB DDR3 1600 9-9-9-24
GPU: GTX 760 (GPU clock 1115MHz, boost 2.0 clock 1305MHz)(though in photoshop, it only went up to 1000 while processing)
GPU usage is not the best way
GPU usage is not the best way to measure these things. To know what's going on, you really need to profile execution time at various points. That's not to say it won't give you some idea, but it will not give you a good one.
I guess I was just hoping to
I guess I was just hoping to see more CPU and GPU usage when it was taking time to apply the effect. Overall, it is extremely fast on pretty much any normal resolution that someone would use.