libgomp: parallel reverse offload

Extend OpenMP reverse offload support to allow running the host kernels
on multiple threads.  The device plugin API for reverse offload is now made
non-blocking, meaning that running the host kernel in the wrong device
context is no longer a problem.  The NVPTX message passing interface now
uses a ring buffer aproximately matching GCN.

libgomp/ChangeLog:

	* config/gcn/target.c (GOMP_target_ext): Add "signal" field.
	Fix atomics race condition.
	* config/nvptx/libgomp-nvptx.h (REV_OFFLOAD_QUEUE_SIZE): New define.
	(struct rev_offload): Implement ring buffer.
	* config/nvptx/target.c (GOMP_target_ext): Likewise.
	* env.c (initialize_env): Read GOMP_REVERSE_OFFLOAD_THREADS.
	* libgomp-plugin.c (GOMP_PLUGIN_target_rev): Replace "aq" parameter
	with "signal" and "use_aq".
	* libgomp-plugin.h (GOMP_PLUGIN_target_rev): Likewise.
	* libgomp.h (gomp_target_rev): Likewise.
	* plugin/plugin-gcn.c (process_reverse_offload): Add "signal".
	(console_output): Pass signal value through.
	* plugin/plugin-nvptx.c (GOMP_OFFLOAD_openacc_async_construct):
	Attach new threads to the numbered device.
	Change the flag to CU_STREAM_NON_BLOCKING.
	(GOMP_OFFLOAD_run): Implement ring-buffer and remove signalling.
	* target.c (gomp_target_rev): Rename to ...
	(gomp_target_rev_internal): ... this, and change "dev_num" to
	"devicep".
	(gomp_target_rev_worker_thread): New function.
	(gomp_target_rev): New function (old name).
	* libgomp.texi: Document GOMP_REVERSE_OFFLOAD_THREADS.
	* testsuite/libgomp.c/reverse-offload-threads-1.c: New test.
	* testsuite/libgomp.c/reverse-offload-threads-2.c: New test.

0 files changed, 0 insertions, 0 deletions