Method completeDecompositionSingle

src/library/tools/ktest/step.cpp:123–220

}

void
Step::completeDecompositionSingle()
{
    cl_int err;

    kextra_.dtype = kargs().dtype;
    kextra_.kernType = CLBLAS_COMPUTING_KERNEL;
    kextra_.flags = (KernelExtraFlags)(kextra_.flags |
            clblasArgsToKextraFlags(&step_.args, blasFunctionID()));
    if (deviceVendor(device()) == "Advanced Micro Devices, Inc.") {
        kextra_.flags = static_cast<KernelExtraFlags>
            (kextra_.flags | KEXTRA_VENDOR_AMD | KEXTRA_ENABLE_MAD);
    }

    step_.pgran.wfSize = deviceWavefront(device(), &err);

    step_.extraFlags = kextra_.flags;
    step_.patternID = selectPattern(&step_, 0);
    pattern_ = &clblasSolvers[step_.funcID].memPatterns[step_.patternID];

    if (0 == step_.subdims[0].bwidth
            && 0 == step_.subdims[0].bwidth
            && 0 == step_.subdims[0].bwidth) {
        getStepGranulation(&step_);

    }
    else if (pattern_->sops->checkCalcDecomp) {
        pattern_->sops->checkCalcDecomp(&step_.pgran, step_.subdims, 2,
                                        kextra_.dtype, PGRAN_CALC);
    }
    else {
        size_t wgX, wgY;
        size_t x0, y0;
        SolverFlags sflags;

        // Set up granulation for given dimensions

        wgY = step_.subdims[0].y/ step_.subdims[1].y;
        wgX = step_.subdims[0].x/ step_.subdims[1].x;

        x0 = step_.subdims[0].x;
        y0 = step_.subdims[0].y;

        if (funcBlasLevel(blasFunctionID()) == 2) {
            /* Level 2 decomposition size for vectors (dims[0].x) is 1.
             * We have to "restore" it to proceed.
             */
            size_t xBlocks;

            xBlocks = step_.subdims[0].bwidth / step_.subdims[1].bwidth;
            x0 = step_.subdims[1].x * xBlocks;
        }

        /*
         * adjust local size if a subproblem is not divisible
         * between all local threads
         */
        for (; (wgY > 1) && (y0 < wgY); wgY /= 2) { }