Function calcNrThreads

src/library/blas/gens/ger_lds.cpp:178–221

}

static void
calcNrThreads(
    size_t threads[2],
    const SubproblemDim *subdims,
    const PGranularity *pgran,
    const void *args,
    const void *_extra)
{
	const CLBlasKargs *kargs = (const CLBlasKargs *)args;
	const CLBLASKernExtra *extra = ( CLBLASKernExtra *)_extra;
	size_t BLOCKSIZE = pgran->wgSize[0] * pgran->wgSize[1]; // 1D Block
	size_t BH, BW;
    unsigned int VEC_LEN = extra->vecLenA;

	clblasOrder order = ( extra->flags & KEXTRA_COLUMN_MAJOR) ? clblasColumnMajor: clblasRowMajor;

	size_t nBlocksY;                //number of blocks in Y dir ( Although we say 1D block to opencl )
    size_t nBlocksX;                //number of blocks in X dir

    BH = subdims->y;
	BW = subdims->x;

    if ( order == clblasColumnMajor )
    {
		nBlocksY = ( kargs->M + BH*VEC_LEN - 1 ) / (BH*VEC_LEN);
        nBlocksX = ( kargs->N + BW - 1) / BW;
    }
    else
    {
		nBlocksY = ( kargs->M + BH - 1) / BH;
        nBlocksX = ( kargs->N + BW*VEC_LEN - 1) / (BW*VEC_LEN);
    }
	size_t blocks = nBlocksX * nBlocksY;
	threads[0] = blocks * BLOCKSIZE;
	threads[1] = 1;

	#ifdef DEBUG_GER
	printf("calcNrThreads called from GER_Reg.cpp.. wgSize[0]: %u\twgSize[1]: %u\n", pgran->wgSize[0], pgran->wgSize[1]);
	printf("subdim->y :%u\t subdim->x : %u\n", subdims->y, subdims->x);
	printf("kargs-> M : %d,  kargs-> N: %d,  BH: %d,  BW: %d\n", kargs->M, kargs->N, BH, BW);
	printf("blocks : %d\tglobalthreads[0]  : %u\t VecLen :%d\n", blocks, threads[0], VEC_LEN);
	#endif

}

//
// FIXME: Report correct return value - Needs change in KPRINTF