#37 Build error

[Anton]

Compilation stops at src/cudainfo.cu with heavy memory usage.
CUDA 10.1
GCC 8.3.0
Qt 5.12.3
Arch Linux

[Andriy Golovnya]

Hi!
This is a bug in CUDA Toolset sinse version 8.0 or so.
It seams they refuse to fix it till now.
I happens because of size of sible linear test blocks like this:

for(i = 0; i < CZ_CALC_BLOCK_LOOPS; i++) {
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
}

If you reduce them to 2 lines it works (or at least it worked for CUDA Toolset 8.0)

for(i = 0; i < CZ_CALC_BLOCK_LOOPS; i++) {
        CZ_CALC_FMAD_256(val1, val2);
        CZ_CALC_FMAD_256(val1, val2);
}

Remember to adjust the constant accodingly:

#define CZ_CALC_BLOCK_NUM   8           /*!< Number of instruction blocks in loop. */

You may also increase lopps count to compensete for the calculation loss:

#define CZ_CALC_BLOCK_LOOPS 32          /*!< Number of loops to run calculation loop. */

Let me know if it still helps to avoid the memeory usage problem!

WBR, Ag

[Anton]

Hi, Andriy!
Unfortunately, the workaround you suggested didn't help, and the problem remains.
AK