Block GPUTransientMemoryAllocator::allocate(GPUProgressTracking & progress, GPUTransientChunk & chunk,
size_t size, size_t alignment,
SeqNum cmdList)
{
size = roundUpToMultiple(size, alignment);
auto &free = chunk.m_free;
auto b = free.fitAtBegin(size, alignment);
#if 0
XOR_GPU_TRANSIENT_VERBOSE("Trying to allocate %zu for list %lld in existing chunk (%lld, %lld).\n",
size, cmdList,
free.begin, free.end);
#endif
// If the allocation fits in the previous active chunk, just use that.
if (b)
{
free.begin = b.end;
#if 0
XOR_GPU_TRANSIENT_VERBOSE(" Allocation successful. Chunk is now (%lld, %lld).\n",
free.begin, free.end);
#endif
return b;
}
// If not, get a new chunk.
else
{
XOR_GPU_TRANSIENT_VERBOSE(" Existing chunk cannot hold allocation, getting new chunk for list %lld.\n",
cmdList);
XOR_CHECK(size <= static_cast<size_t>(m_chunkSize),
"Allocation does not fit in one chunk");
ChunkNumber newChunk = findFreeChunk(progress);
XOR_CHECK(newChunk >= 0, "There are no free or waitable chunks.");
m_usedChunks.emplace_back(cmdList, newChunk);
int64_t begin = newChunk * m_chunkSize;
free = Block(begin, begin + m_chunkSize);
auto b = free.fitAtBegin(size, alignment);
XOR_ASSERT(b.valid(), "Allocation failed with an empty chunk");
free.begin = b.end;
return b;
}
}
static size_t
collDefBufferSizes()
{
int *streamIndexList, vlistID, nvars, varID, iorank;
int modelID;
size_t sumGetBufferSizes = 0;
int rankGlob = commInqRankGlob ();
int nProcsModel = commInqNProcsModel ();
int root = commInqRootGlob ();
xassert(rxWin != NULL);
unsigned nstreams = reshCountType ( &streamOps );
streamIndexList = xmalloc((size_t)nstreams * sizeof (streamIndexList[0]));
reshGetResHListOfType ( nstreams, streamIndexList, &streamOps );
for (unsigned streamNo = 0; streamNo < nstreams; streamNo++)
{
// space required for data
vlistID = streamInqVlist ( streamIndexList[streamNo] );
nvars = vlistNvars ( vlistID );
for ( varID = 0; varID < nvars; varID++ )
{
iorank = vlistInqVarIOrank ( vlistID, varID );
xassert ( iorank != CDI_UNDEFID );
if ( iorank == rankGlob )
{
for ( modelID = 0; modelID < nProcsModel; modelID++ )
{
int decoChunk;
{
int varSize = vlistInqVarSize(vlistID, varID);
int nProcsModel = commInqNProcsModel();
decoChunk =
(int)ceilf(cdiPIOpartInflate_
* (float)(varSize + nProcsModel - 1)
/ (float)nProcsModel);
}
xassert ( decoChunk > 0 );
rxWin[modelID].size += (size_t)decoChunk * sizeof (double)
/* re-align chunks to multiple of double size */
+ sizeof (double) - 1
/* one header for data record, one for
* corresponding part descriptor*/
+ 2 * sizeof (struct winHeaderEntry)
/* FIXME: heuristic for size of packed Xt_idxlist */
+ sizeof (Xt_int) * (size_t)decoChunk * 3;
rxWin[modelID].dictSize += 2;
}
}
}
// space required for the 3 function calls streamOpen, streamDefVlist, streamClose
// once per stream and timestep for all collprocs only on the modelproc root
rxWin[root].size += numRPCFuncs * sizeof (struct winHeaderEntry)
/* serialized filename */
+ MAXDATAFILENAME
/* data part of streamDefTimestep */
+ (2 * CDI_MAX_NAME + sizeof (taxis_t));
rxWin[root].dictSize += numRPCFuncs;
}
free ( streamIndexList );
for ( modelID = 0; modelID < nProcsModel; modelID++ )
{
/* account for size header */
rxWin[modelID].dictSize += 1;
rxWin[modelID].size += sizeof (struct winHeaderEntry);
rxWin[modelID].size = roundUpToMultiple(rxWin[modelID].size,
PIO_WIN_ALIGN);
sumGetBufferSizes += (size_t)rxWin[modelID].size;
}
xassert ( sumGetBufferSizes <= MAXWINBUFFERSIZE );
return sumGetBufferSizes;
}
