Method execute

Source/Falcor/Utils/Algorithm/PrefixSum.cpp:77–169

}

void PrefixSum::execute(
    RenderContext* pRenderContext,
    ref<Buffer> pData,
    uint32_t elementCount,
    uint32_t* pTotalSum,
    ref<Buffer> pTotalSumBuffer,
    uint64_t pTotalSumOffset
)
{
    FALCOR_PROFILE(pRenderContext, "PrefixSum::execute");

    FALCOR_ASSERT(pRenderContext);
    FALCOR_ASSERT(elementCount > 0);
    FALCOR_ASSERT(pData && pData->getSize() >= elementCount * sizeof(uint32_t));

    // Clear total sum to zero.
    pRenderContext->clearUAV(mpTotalSum->getUAV().get(), uint4(0));

    uint32_t maxElementCountPerIteration = kGroupSize * kGroupSize * 2;
    uint32_t totalElementCount = elementCount;
    uint32_t iterationsCount = div_round_up(totalElementCount, maxElementCountPerIteration);

    for (uint32_t iter = 0; iter < iterationsCount; iter++)
    {
        // Compute number of thread groups in the first pass. Each thread operates on two elements.
        uint32_t numPrefixGroups = std::max(1u, div_round_up(std::min(elementCount, maxElementCountPerIteration), kGroupSize * 2));
        FALCOR_ASSERT(numPrefixGroups > 0 && numPrefixGroups <= kGroupSize);

        // Copy previus iterations total sum to read buffer.
        pRenderContext->copyResource(mpPrevTotalSum.get(), mpTotalSum.get());

        // Pass 1: compute per-thread group prefix sums.
        {
            // Clear group sums to zero.
            pRenderContext->clearUAV(mpPrefixGroupSums->getUAV().get(), uint4(0));

            // Set constants and data.
            auto var = mpPrefixSumGroupVars->getRootVar();
            var["CB"]["gNumGroups"] = numPrefixGroups;
            var["CB"]["gTotalNumElems"] = totalElementCount;
            var["CB"]["gIter"] = iter;
            var["gData"] = pData;

            mpComputeState->setProgram(mpPrefixSumGroupProgram);
            pRenderContext->dispatch(mpComputeState.get(), mpPrefixSumGroupVars.get(), {numPrefixGroups, 1, 1});
        }

        // Add UAV barriers for our buffers to make sure writes from the previous pass finish before the next pass.
        // This is necessary since the buffers are bound as UAVs in both passes and there are no resource transitions.
        pRenderContext->uavBarrier(pData.get());
        pRenderContext->uavBarrier(mpPrefixGroupSums.get());

        // Pass 2: finalize prefix sum by adding the sums to the left to each group.
        // This is only necessary if we have more than one group.
        if (numPrefixGroups > 1)
        {
            // Compute number of thread groups. Each thread operates on one element.
            // Note that we're skipping the first group of 2N elements, as no add is needed (their group sum is zero).