T* memAlloc(const size_t &elements)
{
managerInit();
T* ptr = NULL;
size_t alloc_bytes = divup(sizeof(T) * elements, 1024) * 1024;
if (elements > 0) {
// FIXME: Add better checks for garbage collection
// Perhaps look at total memory available as a metric
if (memory_map.size() > MAX_BUFFERS || used_bytes >= MAX_BYTES) {
garbageCollect();
}
for(mem_iter iter = memory_map.begin(); iter != memory_map.end(); iter++) {
mem_info info = iter->second;
if (info.is_free && info.bytes == alloc_bytes) {
iter->second.is_free = false;
used_bytes += alloc_bytes;
return (T *)iter->first;
}
}
// Perform garbage collection if memory can not be allocated
ptr = (T *)malloc(alloc_bytes);
mem_info info = {false, alloc_bytes};
memory_map[ptr] = info;
used_bytes += alloc_bytes;
}
return ptr;
}
T* memAlloc(const size_t &elements)
{
managerInit();
T* ptr = NULL;
size_t alloc_bytes = divup(sizeof(T) * elements, memory_resolution) * memory_resolution;
if (elements > 0) {
std::lock_guard<std::mutex> lock(memory_map_mutex);
// FIXME: Add better checks for garbage collection
// Perhaps look at total memory available as a metric
if (memory_map.size() > MAX_BUFFERS ||
used_bytes >= MAX_BYTES) {
garbageCollect();
}
for(mem_iter iter = memory_map.begin();
iter != memory_map.end(); ++iter) {
mem_info info = iter->second;
if ( info.is_free &&
!info.is_unlinked &&
info.bytes == alloc_bytes) {
iter->second.is_free = false;
used_bytes += alloc_bytes;
used_buffers++;
return (T *)iter->first;
}
}
// Perform garbage collection if memory can not be allocated
ptr = (T *)malloc(alloc_bytes);
if (ptr == NULL) {
AF_ERROR("Can not allocate memory", AF_ERR_NO_MEM);
}
mem_info info = {false, false, alloc_bytes};
memory_map[ptr] = info;
used_bytes += alloc_bytes;
used_buffers++;
total_bytes += alloc_bytes;
}
return ptr;
}
T* memAlloc(const size_t &elements)
{
managerInit();
int n = getActiveDeviceId();
T* ptr = NULL;
size_t alloc_bytes = divup(sizeof(T) * elements, memory_resolution) * memory_resolution;
if (elements > 0) {
// FIXME: Add better checks for garbage collection
// Perhaps look at total memory available as a metric
if (memory_maps[n].size() >= MAX_BUFFERS || used_bytes[n] >= MAX_BYTES) {
garbageCollect();
}
for(mem_iter iter = memory_maps[n].begin();
iter != memory_maps[n].end(); ++iter) {
mem_info info = iter->second;
if ( info.is_free &&
!info.is_unlinked &&
info.bytes == alloc_bytes) {
iter->second.is_free = false;
used_bytes[n] += alloc_bytes;
used_buffers[n]++;
return (T *)iter->first;
}
}
// Perform garbage collection if memory can not be allocated
if (cudaMalloc((void **)&ptr, alloc_bytes) != cudaSuccess) {
garbageCollect();
CUDA_CHECK(cudaMalloc((void **)(&ptr), alloc_bytes));
}
mem_info info = {false, false, alloc_bytes};
memory_maps[n][ptr] = info;
used_bytes[n] += alloc_bytes;
used_buffers[n]++;
total_bytes[n] += alloc_bytes;
}
return ptr;
}
T* pinnedAlloc(const size_t &elements)
{
managerInit();
T* ptr = NULL;
// Allocate the higher megabyte. Overhead of creating pinned memory is
// more so we want more resuable memory.
size_t alloc_bytes = divup(sizeof(T) * elements, 1048576) * 1048576;
if (elements > 0) {
// FIXME: Add better checks for garbage collection
// Perhaps look at total memory available as a metric
if (pinned_maps.size() >= MAX_BUFFERS || pinned_used_bytes >= MAX_BYTES) {
pinnedGarbageCollect();
}
for(mem_iter iter = pinned_maps.begin();
iter != pinned_maps.end(); ++iter) {
mem_info info = iter->second;
if (info.is_free && info.bytes == alloc_bytes) {
iter->second.is_free = false;
pinned_used_bytes += alloc_bytes;
return (T *)iter->first;
}
}
// Perform garbage collection if memory can not be allocated
if (cudaMallocHost((void **)&ptr, alloc_bytes) != cudaSuccess) {
pinnedGarbageCollect();
CUDA_CHECK(cudaMallocHost((void **)(&ptr), alloc_bytes));
}
mem_info info = {false, false, alloc_bytes};
pinned_maps[ptr] = info;
pinned_used_bytes += alloc_bytes;
}
return (T*)ptr;
}
T* pinnedAlloc(const size_t &elements)
{
managerInit();
return (T *)pinnedBufferAlloc(elements * sizeof(T));
}
T *memAlloc(const size_t &elements)
{
managerInit();
return (T *)bufferAlloc(elements * sizeof(T));
}
/**
Make a new timer.
Note - the timer index selected will be used for all subsequent timers created.
@param timer The hardware timer to use - valid options are 0, 1 and 2. 0 is the default.
*/
Timer::Timer(int timer)
{
if(!manager.timer_count++)
managerInit(timer);
}
