Function assertQuregFitsInGpuMem

quest/src/core/validation.cpp:1659–1699

}

void assertQuregFitsInGpuMem(int numQubits, int isDensMatr, int isDistrib, int isGpuAccel, QuESTEnv env, const char* caller) {

    // validate when GPU-acceleration is possible; EVEN if it's auto! Because the auto-deployer will never
    // fall-back to painfully slow CPU if the GPU memory is filled. So if GPU-accel is possible, it must fit qureg.
    if (isGpuAccel == 0 || env.isGpuAccelerated == 0)
        return;

    // we consult the current available local GPU memory (being more strict than is possible for RAM)
    size_t localCurrGpuMem = gpu_getCurrentAvailableMemoryInBytes();

    // check whether qureg (considering if distributed) fits between node GPU memory(s).
    // note this sets numQuregNodes=1 only when distribution is impossible/switched-off,
    // but not when it would later be automatically disabled. that's fine; the auto-deployer
    // will never disable distribution if local GPU memory can't store the qureg, so we don't 
    // need to validate the auto-deployed-to-non-distributed scenario. we only need to ensure 
    // that auto-deploying-to-distribution is permitted by GPU memory capacity.
    int numQuregNodes = (isDistrib == 0 || ! env.isDistributed)? 1 : env.numNodes;
    bool quregFitsInMem = mem_canQuregFitInMemory(numQubits, isDensMatr, numQuregNodes, localCurrGpuMem);

    tokenSubs vars = {
        {"${IS_DENS}",        isDensMatr},
        {"${NUM_QUBITS}",     numQubits},
        {"${QCOMP_BYTES}",    sizeof(qcomp)},
        {"${MIN_VRAM_AVAIL}", localCurrGpuMem}};

    // make error message specific to whether qureg is local or distributed
    if (numQuregNodes == 1) {
        vars["${EXP_BASE}"] = (isDensMatr)? 4 : 2;

        // require expensive node consensus in case of heterogeneous GPU hardware or loads
        assertAllNodesAgreeThat(quregFitsInMem, report::NEW_QUREG_CANNOT_FIT_INTO_NON_DISTRIB_CURRENT_GPU_MEM, vars, caller);

    // when distributed, comm buffers are considered (hence +1 below)
    } else {
        vars["${LOG2_NUM_AMPS}"] = 1 + mem_getEffectiveNumStateVecQubitsPerNode(numQubits, isDensMatr, numQuregNodes);
        vars["${NUM_GPUS}"] = numQuregNodes;

        // require expensive node consensus in case of heterogeneous GPU hardware or loads
        assertAllNodesAgreeThat(quregFitsInMem, report::NEW_QUREG_CANNOT_FIT_INTO_POTENTIALLY_DISTRIB_CURRENT_GPU_MEM, vars, caller);
    }
}

void validate_newQuregParams(int numQubits, int isDensMatr, int isDistrib, int isGpuAccel, int isMultithread, QuESTEnv env, const char* caller) {