Update memory allocation guide in using pool apis

This is to allow allocations in system memory that exceed sizes
reported by a CPU device

Change-Id: I3d10d192aafcefbe4107f69b7c5e30bf7f836619


[ROCm/ROCR-Runtime commit: 3201f68f72]
This commit is contained in:
Ramesh Errabolu
2019-06-05 11:49:44 -05:00
parent 2b9e13a56c
commit 61b9d4e8b2
7 changed files with 98 additions and 26 deletions
@@ -175,6 +175,9 @@ class MemoryRegion : public core::MemoryRegion {
size_t max_single_alloc_size_;
// Used to collect total system memory
static size_t max_sysmem_alloc_size_;
HSAuint64 virtual_size_;
mutable KernelMutex access_lock_;
@@ -52,6 +52,10 @@
#include "core/inc/exceptions.h"
namespace amd {
// Tracks aggregate size of system memory available on platform
size_t MemoryRegion::max_sysmem_alloc_size_ = 0;
void* MemoryRegion::AllocateKfdMemory(const HsaMemFlags& flag,
HSAuint32 node_id, size_t size) {
void* ret = NULL;
@@ -119,7 +123,7 @@ MemoryRegion::MemoryRegion(bool fine_grain, bool full_profile, core::Agent* owne
virtual_size_ = kGpuVmSize;
} else if (IsSystem()) {
mem_flag_.ui32.PageSize = HSA_PAGE_SIZE_4KB;
mem_flag_.ui32.NoSubstitute = 1;
mem_flag_.ui32.NoSubstitute = 0;
mem_flag_.ui32.HostAccess = 1;
mem_flag_.ui32.CachePolicy = HSA_CACHING_CACHED;
@@ -127,9 +131,20 @@ MemoryRegion::MemoryRegion(bool fine_grain, bool full_profile, core::Agent* owne
(full_profile) ? os::GetUserModeVirtualMemorySize() : kGpuVmSize;
}
// Bind if memory region is coarse or fine grain
mem_flag_.ui32.CoarseGrain = (fine_grain) ? 0 : 1;
// Adjust allocatable size per page align
max_single_alloc_size_ = AlignDown(static_cast<size_t>(GetPhysicalSize()), kPageSize_);
mem_flag_.ui32.CoarseGrain = (fine_grain) ? 0 : 1;
// Keep track of total system memory available
// @note: System memory is surfaced as both coarse
// and fine grain memory regions. To track total system
// memory only fine grain is considered as it avoids
// double counting
if (IsSystem() && (fine_grain)) {
max_sysmem_alloc_size_ += max_single_alloc_size_;
}
assert(GetVirtualSize() != 0);
assert(GetPhysicalSize() <= GetVirtualSize());
@@ -147,7 +162,10 @@ hsa_status_t MemoryRegion::Allocate(size_t& size, AllocateFlags alloc_flags, voi
return HSA_STATUS_ERROR_INVALID_ALLOCATION;
}
if (size > max_single_alloc_size_) {
// Alocation requests for system memory considers aggregate
// memory available on all CPU devices
if (size > ((IsSystem() ?
max_sysmem_alloc_size_ : max_single_alloc_size_))) {
return HSA_STATUS_ERROR_INVALID_ALLOCATION;
}
@@ -285,9 +303,11 @@ hsa_status_t MemoryRegion::GetInfo(hsa_region_info_t attribute,
break;
case HSA_REGION_INFO_ALLOC_MAX_SIZE:
switch (mem_props_.HeapType) {
case HSA_HEAPTYPE_SYSTEM:
*((size_t*)value) = max_sysmem_alloc_size_;
break;
case HSA_HEAPTYPE_FRAME_BUFFER_PRIVATE:
case HSA_HEAPTYPE_FRAME_BUFFER_PUBLIC:
case HSA_HEAPTYPE_SYSTEM:
case HSA_HEAPTYPE_GPU_SCRATCH:
*((size_t*)value) = max_single_alloc_size_;
break;
@@ -365,10 +385,23 @@ hsa_status_t MemoryRegion::GetPoolInfo(hsa_amd_memory_pool_info_t attribute,
case HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_GRANULE:
case HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALIGNMENT:
return GetInfo(static_cast<hsa_region_info_t>(attribute), value);
break;
case HSA_AMD_MEMORY_POOL_INFO_ACCESSIBLE_BY_ALL:
*((bool*)value) = IsSystem() ? true : false;
break;
case HSA_AMD_MEMORY_POOL_INFO_ALLOC_MAX_SIZE:
switch (mem_props_.HeapType) {
case HSA_HEAPTYPE_FRAME_BUFFER_PRIVATE:
case HSA_HEAPTYPE_FRAME_BUFFER_PUBLIC:
case HSA_HEAPTYPE_GPU_SCRATCH:
return GetInfo(HSA_REGION_INFO_ALLOC_MAX_SIZE, value);
case HSA_HEAPTYPE_SYSTEM:
// Aggregate size available for allocation
*((size_t*)value) = max_sysmem_alloc_size_;
break;
default:
*((size_t*)value) = 0;
}
break;
default:
return HSA_STATUS_ERROR_INVALID_ARGUMENT;
}
@@ -648,8 +648,26 @@ typedef enum {
} hsa_amd_segment_t;
/**
* @brief A memory pool represents physical storage on an agent.
*/
* @brief A memory pool encapsulates physical storage on an agent
* along with a memory access model.
*
* @details A memory pool encapsulates a physical partition of an agent's
* memory system along with a memory access model. Division of a single
* memory system into separate pools allows querying each partition's access
* path properties (see ::hsa_amd_agent_memory_pool_get_info). Allocations
* from a pool are preferentially bound to that pool's physical partition.
* Binding to the pool's preferential physical partition may not be
* possible or persistent depending on the system's memory policy
* and/or state which is beyond the scope of HSA APIs.
*
* For example, a multi-node NUMA memory system may be represented by multiple
* pool's with each pool providing size and access path information for the
* partition it represents. Allocations from a pool are preferentially bound
* to the pool's partition (which in this example is a NUMA node) while
* following its memory access model. The actual placement may vary or migrate
* due to the system's NUMA policy and state, which is beyond the scope of
* HSA APIs.
*/
typedef struct hsa_amd_memory_pool_s {
/**
* Opaque handle.
@@ -729,6 +747,11 @@ typedef enum {
* attribute is bool.
*/
HSA_AMD_MEMORY_POOL_INFO_ACCESSIBLE_BY_ALL = 15,
/**
* Maximum aggregate allocation size in bytes. The type of this attribute
* is size_t.
*/
HSA_AMD_MEMORY_POOL_INFO_ALLOC_MAX_SIZE = 16,
} hsa_amd_memory_pool_info_t;
/**
@@ -817,8 +840,8 @@ hsa_status_t HSA_API hsa_amd_agent_iterate_memory_pools(
* @retval ::HSA_STATUS_ERROR_INVALID_MEMORY_POOL The memory pool is invalid.
*
* @retval ::HSA_STATUS_ERROR_INVALID_ALLOCATION The host is not allowed to
* allocate memory in @p memory_pool, or @p size is greater than the value of
* HSA_AMD_MEMORY_POOL_INFO_ALLOC_MAX_SIZE in @p memory_pool.
* allocate memory in @p memory_pool, or @p size is greater than
* the value of HSA_AMD_MEMORY_POOL_INFO_ALLOC_MAX_SIZE in @p memory_pool.
*
* @retval ::HSA_STATUS_ERROR_INVALID_ARGUMENT @p ptr is NULL, or @p size is 0,
* or flags is not 0.