2 Replies Latest reply on Mar 12, 2015 9:48 PM by abcdabcd

    Questions about GPU Perfstudio




      I have a few questions about GPU Perfstudio.


      1) Why I cannot know values of compute-shader-realated counters when I use OpenGL instead of DirectX? I do not know why the current GPU Perfstudio version only supports CS counters with DirectX.


      2) What does TexUnitBusy exactly mean? According to the GPU Perf API document, TexUnitBusy means a percentage of GPUTime the texture unit is active, and TexUnitBusy is measured with all extra fetches and any cache or memory effects taken into account. I understand that the texture unit (maybe the L2 texture cache?) deals with all memory requests including fetches of textures, vertices, frame buffers, etc. In fact, even though textures were not used in a scene, the TexUnitBusy value was higher than zero (maybe) due to vertex fetches. Do I understand correctly?


      Additionally, what does (VS/DS/GS/PS)TexBusy mean? I found that the sum of (VS/DS/GS/PS)TexBusy was less than the value of TexUnitBusy, and the ratio of (VS/DS/GS/PS)TexBusy to (VS/DS/GS/PS)ALUBusy was similar when screen resolutions (and TexMissRate) were changed. It means that (VS/DS/GS/PS)TexBusy may be affected by not total texture mapping costs including memory accesses but only texture instruction processing. Do I understand correctly? If so, can I interpret that TexUnitBusy is related to both texture processing of each shader and memory accesses? If a texture cache hit rate is 100%, will the value of TexUnitBusy be equal to the sum of (VS/DS/GS/PS)TexBusy? Otherwise, will TexUnitBusy be zero due to no off-chip memory accesses in this case?


      To sum up, I am curious the relationship between TexUnitBusy and (VS/DS/GS./PS)TexBusy. Currently, I understand as follows:


      * TexUnitBusy = GPU time for texture mapping/filtering and all off-chip memory accesses. The total percentage of (VS/DS/GS/PS)Texbusy should be included in the percentage of TexUnitBusy.

      * (VS/DS/GS/PS)TexBusy = GPU time for texture instruction processing in each shader. Memory fetch costs are not included.


      Is it correct?


      3) When can I download the updated version with the fix of the incorrect average calculation bug in the frame profiler which I asked one month ago?



        • Re: Questions about GPU Perfstudio



          Question 1: We think you might be using pre-GCN hardware, because the ComputeShader counters are available for GL for all GCN-based hardware. OpenGL didn’t have support for Compute Shaders until v4.3 (2012), and we didn’t add support for Compute Shader counters until sometime after that (late 2013).  At that point, we only added them for SI/CI. 


          Question 2: No answer yet.


          Question 3: The fix has not yet been implemented





            • Re: Questions about GPU Perfstudio

              Thank you for your answer.


              1) As you expected, I have a pre-GCN GPU (7550M). Do you have no plan to support SI/CI counters on pre-GCN GPU with OpenGL? Both the current OpenGL version and pre-GCN driver support compute shaders, and I need to analyze a compute-shader-based application.


              2) Can you answer to my question 2? If my question is too wordy to understand, I can simply ask as follows.

              Does TexUnitBusy include GPU time for both texture mapping and all off-chip memory accesses? Also, does (VS/DS/GS/PS)TexBusy include GPU time for memory accesses? I expect that the answer to the first question is yes and to the second question is no.