// This does the exchange of the gauge field ghost zone and places it
// into the ghost array.
// This should be optimized so it is reused if called multiple times
void cpuGaugeField::exchangeGhost() const {
void **send = (void**)malloc(sizeof(void*)*QUDA_MAX_DIM);
for (int d=0; d<nDim; d++) {
send[d] = malloc(nFace * surface[d] * reconstruct * precision);
}
// get the links into a contiguous buffer
if (precision == QUDA_DOUBLE_PRECISION) {
packGhost((double**)gauge, (double**)send, nFace, x, volumeCB, surfaceCB);
} else {
packGhost((float**)gauge, (float**)send, nFace, x, volumeCB, surfaceCB);
}
// communicate between nodes
FaceBuffer faceBuf(x, nDim, reconstruct, nFace, precision);
faceBuf.exchangeCpuLink(ghost, send);
for (int i=0; i<4; i++) {
double sum = 0.0;
for (int j=0; j<nFace*surface[i]*reconstruct; j++) {
sum += ((double*)(ghost[i]))[j];
}
}
for (int d=0; d<nDim; d++) free(send[d]);
free(send);
}
void cpuColorSpinorField::exchangeGhost(QudaParity parity, int nFace, int dagger, const MemoryLocation *dummy1,
const MemoryLocation *dummy2, bool dummy3, bool dummy4) const
{
// allocate ghost buffer if not yet allocated
allocateGhostBuffer(nFace);
void **sendbuf = static_cast<void**>(safe_malloc(nDimComms * 2 * sizeof(void*)));
for (int i=0; i<nDimComms; i++) {
sendbuf[2*i + 0] = backGhostFaceSendBuffer[i];
sendbuf[2*i + 1] = fwdGhostFaceSendBuffer[i];
ghost_buf[2*i + 0] = backGhostFaceBuffer[i];
ghost_buf[2*i + 1] = fwdGhostFaceBuffer[i];
}
packGhost(sendbuf, parity, nFace, dagger);
exchange(ghost_buf, sendbuf, nFace);
host_free(sendbuf);
}
