1 Reply Latest reply on May 22, 2009 9:07 AM by kamokr

    Any kernel writes only zeroes

      output resource created with calResCreate2D and CAL_FORMAT_UBYTE_1

      My goal is basically to mix two RGB images into one grayscale. In order to speed up memory transfers (realtime video processing) I use memory pinning for input and output buffers. It's modified user_memory sample from CAL SDK 1.4.0 beta:

      // first initializing CAL and compiling kernel

      // then

      CALuint isize=sizeof(CALuint)*width*height;

      CALuint osize=sizeof(CALubyte)*width*height;


      idata0 = (CALuint*)_aligned_malloc(isize, 4096);

      idata1 = (CALuint*)_aligned_malloc(isize, 4096);

      odata0 = (CALubyte*)_aligned_malloc(osize, 4096);


      calExtGetProc((CALextproc*)&calResCreate2D, CAL_EXT_RES_CREATE, "calResCreate2D");


      calResCreate2D(&inputRes0, device, (CALvoid*)idata0, width, height, CAL_FORMAT_UINT_1, isize, 0);

      calResCreate2D(&inputRes1, device, (CALvoid*)idata1, width, height, CAL_FORMAT_UINT_1, isize, 0);

      calResCreate2D(&outputRes0, device, (CALvoid*)odata0, width, height, CAL_FORMAT_UBYTE_1, osize, 0);

      // then mapping, binding and initializing input buffers with random data

      // and finally calling calCtxRunProgram


      I have following problems:

      1. The required pitch alignment in data elements is 64. When width is 640, calResCreate2D for odata0 fails, but works for 256, 512, 768 etc. Is UBYTE_1 one data element, or 1/4 of data element?

      2. Any kernel fills output buffer with zeroes, even this one:

      kernel void SomeKernel(unsigned int i0<>, unsigned int i1<>, out unsigned char o0<>


        o0 = 0xff;



      All of this works well for CAL_FORMAT_UINT_1 output buffer.