0 Replies Latest reply on May 1, 2013 8:14 AM by oscarbg2

    Bugs in 13.4 GL 4.3 exts:compute shader, SBOs and no_attachments



      I have been testing new OGL compute shader and storage buffer objects extension and found following bugs (13.4 on 7950):

      (please note all the samples I use for testing this work correctly on Nvidia OGL 4.3 cards)

      *using atomicMax and atomicMin on shared variables hang the GLSL compiler others like atomicOr are OK!

      groupshared uint ldsZMax;

      uint z;

      atomicMax( ldsZMax, z );

      *using a compute shader with following launch size and shared arrays usage:

      #define BLOCK_SIZE 32

      layout (local_size_x = BLOCK_SIZE, local_size_y = BLOCK_SIZE) in;

      shared double As[BLOCK_SIZE*BLOCK_SIZE];

      shared double Bs[BLOCK_SIZE*BLOCK_SIZE];

      crashes with:

      Compute shader(s) failed to link.

      Compute link error: HW_UNSUPPORTED.

      Compute shader not supported by hardware

      diminishing BLOCK_SIZE to less than 32 seems to work.. I have tested using

      layout (local_size_x = 32, local_size_y = 32) in;

      isn't a issue so 32 should work as for this conf each of this two shared arrays is size 8192 (sizeof(double)*32*32) so total shared mem usage is 2*8192 and is equal to HALF! of the reported max

      size (GL_MAX_COMPUTE_SHARED_MEMORY_SIZE: 32768).. I verify this issue is on shared mem size usage as using something like

      (with BLOCK_SIZE=32):

      shared double As[BLOCK_SIZE*BLOCK_SIZE-1];

      shared double Bs[BLOCK_SIZE*BLOCK_SIZE];

      seems to compile so please fix to be able to use not only up to 16383 bytes of shared mem but full 32768 bytes..

      *using sbo on non compute shaders (like fragment shaders seems no be not correct)

      *getting GL_MAX_COMPUTE_WORK_GROUP_COUNT and GL_MAX_COMPUTE_WORK_GROUP_SIZE I get using debug_output bug:

      glGetIntegerv parameter <pname> has an invalid enum '0x91be' (GL_INVALID_ENUM)

      other new like GL_MAX_COMPUTE_ATOMIC_COUNTERS seem to work..

      Related altough no_attachments extension is not adversited new entry points are present so I played with it using default and seems a simple test works on 79xx but not on 58xx






      a sample using this works on 7xxx series but not on 5xxx series..