This patch implements "managed" memory for AMD GCN GPUs in OpenMP. It
builds on the support added to the NVPTX libgomp for CUDA Managed Memory, a
week or two ago.
These features were first posted here a few years ago, as part of a larger
Unified Shared Memory patch series, and then in a slightly changed version just
over a year ago. Hopefully this time the controversial bits have been removed.
Since we do not use HIP we cannot use hipMallocManaged, so this patch attempts
to replicate the same effect by setting the appropriate attributes. This works
on more devices than support proper USM, but still I cannot be sure that the
settings are correct for every device out there (I have tested on gfx900,
gfx906, gfx908, gfx90a, and gfx1100).
The HSA header file update uses the most recent files relicensed for us by AMD,
at the time of the first patch posting. Those files have certainly moved on in
the upstream sources, but I did not ask to get those relicensed.
include/ChangeLog:
* hsa.h: Import newer version.
* hsa_ext_amd.h: Likewise.
* hsa_ext_image.h: Likewise.
libgomp/ChangeLog:
* Makefile.in: Regenerate.
* libgomp-plugin.h (gomp_simple_alloc_init_context): New prototype.
(gomp_simple_alloc_register_memory): New prototype.
(gomp_simple_alloc): New prototype.
(gomp_simple_free): New prototype.
(gomp_simple_realloc): New prototype.
* libgomp.h (gomp_simple_alloc_init_context): Move to libgomp-plugin.h.
(gomp_simple_alloc_register_memory): Likewise.
(gomp_simple_alloc): Likewise.
(gomp_simple_free): Likewise.
(gomp_simple_realloc): Likewise.
* libgomp.texi: Update AMD managed memory description.
* plugin/Makefrag.am (libgomp_plugin_gcn_la_SOURCES): Add
simple-allocator.c and plugin/mutex.c.
* plugin/plugin-gcn.c: Include sys/mman.h and unistd.h.
(struct hsa_runtime_fn_info): Add hsa_amd_svm_attributes_set_fn.
(dump_hsa_system_info): Add HSA_AMD_SYSTEM_INFO_SVM_SUPPORTED and
HSA_AMD_SYSTEM_INFO_SVM_ACCESSIBLE_BY_DEFAULT to the GCN_DEBUG output.
(init_hsa_runtime_functions): Add hsa_amd_svm_attributes_set.
(isa_matches_agent): Add a new error message for the case where the
ISA doesn't match but the name does.
(managed_ctx): New variable.
(managed_heap_create): New function.
(GOMP_OFFLOAD_managed_alloc): Likewise.
(GOMP_OFFLOAD_managed_free): Likewise.
* simple-allocator.c (gomp_fatal): New macro.
* testsuite/lib/libgomp.exp (check_effective_target_omp_managedmem):
Add amdgcn support checker.
(check_effective_target_offload_target_amdgcn_with_xnack): New.
* testsuite/libgomp.c-c++-common/requires-4.c: Ignore xnack warning.
* testsuite/libgomp.c-c++-common/requires-4a.c: Ignore xnack warning.
* testsuite/libgomp.c-c++-common/requires-5.c: Ignore xnack warning.
* testsuite/libgomp.c++/alloc-managed-1.C: Add -mxnack=on, if needed.
* testsuite/libgomp.c/alloc-managed-1.c: Likewise.
* testsuite/libgomp.c/alloc-managed-2.c: Likewise.
* testsuite/libgomp.c/alloc-managed-3.c: Likewise.
* testsuite/libgomp.c/alloc-managed-4.c: Likewise.
* testsuite/libgomp.fortran/alloc-managed-1.f90: Likewise.
* plugin/mutex.c: New file.
This patch introduces a new custom memory allocator for use with pinned
memory (in the case where the Cuda allocator isn't available). In future,
this allocator will also be used for Managed Memory. Both memories are
incompatible with the system malloc because allocated memory cannot share a
page with memory allocated for other purposes.
This means that small allocations will no longer consume an entire page of
pinned memory. Unfortunately, it also means that pinned memory pages will
never be unmapped (although they may be reused). This isn't a technical
limitation; the "free" algorithm could be extended in future, if needed.
The implementation is not perfect; there are various corner cases (especially
related to extending onto new pages) where allocations and reallocations may
be sub-optimal, but it should still be a step forward in support for small
allocations.
I have considered using libmemkind's "fixed" memory but rejected it for three
reasons: 1) libmemkind may not always be present at runtime, 2) there's no
currently documented means to extend a "fixed" kind one page at a time
(although the code appears to have an undocumented function that may do the
job, and/or extending libmemkind to support the MAP_LOCKED mmap flag with its
regular kinds would be straight-forward), 3) Managed Memory benefits from
having the metadata located in different memory and using an external
implementation makes it hard to guarantee this.
libgomp/ChangeLog:
* Makefile.am (libgomp_la_SOURCES): Add simple-allocator.c.
* Makefile.in: Regenerate.
* basic-allocator.c: Mention simple-allocator in the comment.
* config/linux/allocator.c: Include unistd.h.
(pin_ctx): New variable.
(ctxlock): New variable.
(linux_init_pin_ctx): New function.
(linux_memspace_alloc): Use simple-allocator for pinned memory.
(linux_memspace_free): Likewise.
(linux_memspace_realloc): Likewise.
* libgomp.h (gomp_simple_alloc_init_context): New prototype.
(gomp_simple_alloc_register_memory): New prototype.
(gomp_simple_alloc): New prototype.
(gomp_simple_free): New prototype.
(gomp_simple_realloc): New prototype.
* libgomp.texi: Update pinned memory trait documentation.
* testsuite/libgomp.c/alloc-pinned-8.c: New test.
* simple-allocator.c: New file.
