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;
}
}
void *allocateMemory(unsigned size_in_bytes)
{
int size;
unsigned chunk;
int chunk_size;
if (!initialized)
{
initializeHeap();
initialized = true;
}
if (size_in_bytes == 0)
{
return nullptr;
}
size = (size_in_bytes + sizeof(int)-1) / sizeof(int);
chunk = findFreeChunk(size);
if (always_false)
{
displayChunks();
}
if (chunk == bogus_chunk)
{
printf("Ah NUTS, out of memory\n");
exit(-1); // forces main to return -1 !immediately!
}
splitChunkIfNecessary(chunk, size);
chunk_size = memory_pool[chunk];
assert(chunk_size >= size); // just a sanity check, pretty insane really
setBothSignatures(chunk, -chunk_size);
/* OK, all done... all we need to do now is return... oh,
* we're supposed to return a pointer, not an index.....
*/
return &memory_pool[chunk + 1];
}
