gpu: intel: ocl: allow unlimited allocations #4354
Conversation
github-actions
bot
added
platform:gpu-intel
spalicki
force-pushed
the
spalicki/add_ocl_4gb_flag
branch
from
83b8209 to
bdf6bda
Compare
|
make test |
src/gpu/intel/ocl/usm_utils.cpp
Outdated
| bool large_buffer = size | ||
| > utils::downcast<const xpu::ocl::engine_impl_t *>(engine->impl()) | ||
| ->max_allocation_size(); | ||
| static cl_bitfield properties[] |
There was a problem hiding this comment.
Minor: I'm not sure static is good for properties object that's passed to CL calls especially in case it will need to be extended in the future. Also current code doesn't provide a scalable way of expanding it either, probably a TODO with that note could help to look into that direction when/if needed.
| #endif | ||
|
|
||
| int get_gpu_ram_sizes(size_t &ram_size, size_t &max_alloc_size) { | ||
| int get_gpu_ram_size(size_t &ram_size) { |
There was a problem hiding this comment.
Did you have a chance to verify the change works with all 4 supported memory kinds times correctness and fast performance mode where different approach used for memory object management?
As a part of this question also: should this call be updated?
There was a problem hiding this comment.
Well, it seems that CL_MEM_ALLOW_UNRESTRICTED_SIZE_INTEL is truly unrestricted, allowing to allocate any size even over the GPU VRAM. So, I am going to have to add some additional guard.
NVM it looks like driver bug.
The flag according to documentation only applies to clCreateBuffer, clCreateBufferWithProperties, clCreateBufferWithPropertiesINTEL, clSVMAlloc, clSharedMemAllocINTEL, clDeviceMemAllocINTEL and clHostMemAllocINTEL.
spalicki
force-pushed
the
spalicki/add_ocl_4gb_flag
branch
5 times, most recently
from
f3bf210 to
79caa93
Compare
spalicki
force-pushed
the
spalicki/add_ocl_4gb_flag
branch
from
79caa93 to
179f908
Compare
|
make test |
|
make test perf-gpu |
4 participants
Allow allocation of buffers bigger than
CL_DEVICE_MAX_MEM_ALLOC_SIZElimited byCL_DEVICE_GLOBAL_MEM_SIZE.