Announcement

Collapse
No announcement yet.

Valve Funds Glassy Mesa Development For Better Driver Performance

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • #46
    It's neat how LLVM seems to be leaking into everything providing interesting ways of speeding stuff up. Emscriptem is pretty awesome as well.

    Comment


    • #47
      Originally posted by log0 View Post
      It would have been great if he had more numbers (Intel has a shader db, they use for benchmarking).

      But he is already getting up to 20% performance improvements, after just 10 weeks hacking on it. Given how much time Intel devs spend on shader optimization, they should be quite interested imho.

      It should be quite interesting for the gallium fraction too. They are already using llvm, would be only natural to switch to it completely.
      It seems there are a few people working on SSA for the existing IR's it will be interesting to see if this provides the same type of performance increase once optimisations have been applied without using llvm and adding yet another IR into the mix.

      Comment


      • #48
        Originally posted by tarceri View Post
        It seems there are a few people working on SSA for the existing IR's it will be interesting to see if this provides the same type of performance increase once optimisations have been applied without using llvm and adding yet another IR into the mix.
        I understand the desire to go with a custom solution, most certainly resulting in leaner code. But someone will have to implement the optimizations for the new IR first. And then there is OpenCL and SPIR(LLVM IR) which will be using LLVM anyway I think.

        Comment


        • #49
          Originally posted by log0 View Post
          I understand the desire to go with a custom solution, most certainly resulting in leaner code. But someone will have to implement the optimizations for the new IR first. And then there is OpenCL and SPIR(LLVM IR) which will be using LLVM anyway I think.
          Yes there are pros and cons to each approach. But as far as "someone will have to implement the optimizations" you can also say the same thing about Glassy Mesa, I don't know all the details but as far as I understand it still requires quite a bit of work to get it up to par with the current glsl front end. I'd imagine it will be a big job just on the testing side to not cause a large amoung of regressions. One other interesting thing is that it completely replaces the current glsl parser also albeit with the official reference front end.

          Comment


          • #50
            Originally posted by tarceri View Post
            albeit with the official reference front end.
            There is NO official reference front end for GLSL. Sadly.

            (Though if Mesa succeed with creating one, then maybe, just maybe AMD will use it for fglrx. Which would be really nice.)

            Comment


            • #51
              Originally posted by przemoli View Post
              There is NO official reference front end for GLSL. Sadly.

              (Though if Mesa succeed with creating one, then maybe, just maybe AMD will use it for fglrx. Which would be really nice.)
              Look here: http://www.opengl.org/sdk/tools/glslang/

              Comment


              • #52
                Originally posted by przemoli View Post
                There is NO official reference front end for GLSL. Sadly.

                (Though if Mesa succeed with creating one, then maybe, just maybe AMD will use it for fglrx. Which would be really nice.)
                Glslang is the official reference compiler front end for the OpenGL ES and OpenGL shading languages.
                http://www.khronos.org/opengles/sdk/...ence-Compiler/

                Comment


                • #53
                  Originally posted by przemoli View Post
                  There is NO official reference front end for GLSL. Sadly.

                  (Though if Mesa succeed with creating one, then maybe, just maybe AMD will use it for fglrx. Which would be really nice.)
                  Valve got Khronos to endorse glslang as the official one. Mesa was considered, but wasn't far enough along for Valve which wanted to use some GL4 code immediately.

                  Comment

                  Working...
                  X