Significant Performance Improvements For RadeonSI LLVM With New Patches

[Veerappan]

Well, the 3D performance *MAY* be significantly improved... I just ran a test of Unigine Heaven on my 7850, and the results are as linked:

http://openbenchmarking.org/result/1501029-SO-LLVMPERFT49

From 21 to 29 fps... nice improvement.

HOWEVER, the rendering was incorrect, so this may not be indicative of the final performance once the rendering is fixed.

[dungeon]

Well, the 3D performance *MAY* be significantly improved... I just ran a test of Unigine Heaven on my 7850, and the results are as linked:

http://openbenchmarking.org/result/1501029-SO-LLVMPERFT49

From 21 to 29 fps... nice improvement.

HOWEVER, the rendering was incorrect, so this may not be indicative of the final performance once the rendering is fixed.

If rendering is broken do not run benchmarks - that is first rule for proper benchmarking I don't think you will have that +28% gain once rendering is fixed, it will be most probably halfed, but OK that is just a guess

[Veerappan]

If rendering is broken do not run benchmarks - that is first rule for proper benchmarking I don't think you will have that +28% gain once rendering is fixed, it will be most probably halfed, but OK that is just a guess

Hence the giant pile of caveats attached to my post.

Maybe I'll have time over the weekend to try to do a full piglit run and see if there's any tests that can consistently isolate what regressed.

[bridgman]

I don't quite undestand the difference between various AMD drivers. AMD Catalyst drivers on Windows and Linux are also using LLVM. Do they have two different LLVM based compilers - one open-source and other closed source. Or is the compiler part shared between both drivers stacks?

Both stacks use LLVM, but at different points. There are normally two levels of compilation (input to IR to output) but the open source compute stack essentially combines them into a single step today.

** Catalyst compute:

EDG (Edison Design Group) front end parses OpenCL C99 and generates LLVM IR, LLVM + AMDIL back-end generates AMDIL
AMD proprietary shader compiler goes from AMDIL to HW ISA

For OpenCL 2.0, some of the HSA paths are used:

Clang (I think) parses OpenCL C99 and generates LLVM IR, LLVM + HSAIL back-end generates HSAIL
AMD proprietary shader compiler (running in HSA finalizer mode) goes from HSAIL to HW ISA

** Open source compute:

Clang parses OpenCL C99 and generates LLVM IR, LLVM + R600 backend generate HW ISA

** Catalyst graphics:

AMD proprietary compiler parses GLSL and generates AMDIL
AMD proprietary shader compiler goes from AMDIL to HW ISA

** Open source graphics:

Mesa compiler parses GLSL and generates GLSL IR, which is then converted to TGSI
Driver shader compiler parses TGSI, converts to LLVM IR, then LLVM + R600 backend generate HW ISA

[bridgman]

BTW we are looking into sharing more code between open and proprietary stacks in this area (as a consequence of HSA work) but there's a fair amount of work required to get there.

If you look at the previous post there is currently sharing between graphics & compute paths for each stack (Catalyst shares the AMD proprietary shader compiler between graphics & compute, open stack shares the LLVM r600 back end between graphics and compute) but no direct sharing between Catalyst & open stack. The LLVM r600 back end was originally derived from the proprietary AMDIL back end, but there's probably not a lot of the original code still there today (tstellar would know for sure).

[Szzz]

...

Thanks for the info.
When I got "LLVM ERROR: Cannot select" on Windows while trying to compile some OpenCL kernels I thought that it is the same R900 backend about which I usually read on Phoronix. Now I see that this is not the case. By the way, do you know, where should I submit the bugs in Catalyst?

[bridgman]

By the way, do you know, where should I submit the bugs in Catalyst?

I would use the "unofficial Bugzilla" at http://ati.cchtml.com/ - I think that's still the best way to get bugs recorded in a form that's useful to the developers.

Simple bugs/questions can go into the forums at developer.amd.com but that's not a great vehicle for complex issues - the forums are more useful for questions than for bug reports IMO.

[Kano]

Even reporting OpenGL bugs in the official amd steam forum is usally useless:

http://devgurus.amd.com/community/steam-linux

[commiethebeastie]

********************
Testing Time: 67.85s
********************
Failing Tests (5):
    LLVM :: CodeGen/R600/fceil64.ll
    LLVM :: CodeGen/R600/ffloor.ll
    LLVM :: CodeGen/R600/imm.ll
    LLVM :: CodeGen/R600/setcc-opt.ll
    LLVM :: CodeGen/R600/trunc-cmp-constant.ll

  Expected Passes    : 11879
  Expected Failures  : 74
  Unsupported Tests  : 180
  Unexpected Failures: 5
Makefile:98: recipe for target 'check-local' failed
make[3]: *** [check-local] Error 1
make[3]: Leaving directory '/home/user/sources/graphics/llvm/build-llvm/test'
/home/user/sources/graphics/llvm/Makefile.rules:1784: recipe for target 'check' failed
make[2]: *** [check] Error 2
make[2]: Leaving directory '/home/user/sources/graphics/llvm/build-llvm'

http://pastebin.com/w1pL4QN0

[mercutio]

when is 2d performance being fixed?

r9 290 and radeon 7750 are both amazingly slow for some things, like say dragging the window title in notion from one space to another, or like showing some text on urxvt. (alsamixer is one of the problematic programs - it takes about 4 seconds to display the volume settings showing the screen slowly draw)