Граф коммитов

280 Коммитов

Автор SHA1 Сообщение Дата
Ben Sander 0409bf639c Add HIP_FAIL_SOC.
Fail sub-optimal-copies rather than perform them slowly.
SOC occur on async copy of unpinned memory, or P2P copy between GPUs
that are not peers.
2017-01-25 21:53:17 -06:00
Ben Sander 1635b8f43f Read HCC_OPT_FLUSH and optimize dispatch accordingly.
If HCC is in this mode, we can use less aggressive flushes in some
cases.
2017-01-25 21:50:52 -06:00
Ben Sander 0dabdeb01f Move core env var processing to env.cpp 2017-01-23 22:34:41 -06:00
Ben Sander 96eac67929 Add debug tips to docs 2017-01-23 22:34:41 -06:00
Ben Sander 927ac3d81c Add HIP_SYNC_HOST_ALLOC, HipReadEnv 2017-01-19 23:55:24 -06:00
Ben Sander 1f5d16afe7 Doc update - describe debug techniques
Also tweak sample to remove unneeded HIP_KERNEL_NAME.
Comment update
2017-01-19 12:40:45 -06:00
Aditya Atluri 73fcce26f9 changed copyright year from 2016 to 2017 in src directory
Change-Id: Idb97db509b2b4b1656b2df7a14a62ade38c9d574
2017-01-11 18:05:41 -06:00
Ben Sander a3e0012567 Add HIP_MAX_QUEUES feature.
Includes some tricky manipulation of the locks for contexts and streams.
issue is that stealing a stream requires we lock the context to
walk the streams to find a victim.  To avoid deadlock, we can't
have a stream locked when we lock the context.  This implementation
releases the stream lock, then acquires the context and selects the
victim.
A more stable implemenation might be to copy the stream list
from a context so that a lock is not required to walk all streams.
Smart shared_ptr could be used to prevent the streams from being
deallocated during the walk.
2017-01-09 21:02:56 -06:00
Ben Sander 93fbc9cf7b First pass at virtualized queue support.
Also updated stream debug messages to consistently use trace_helper.
2017-01-09 21:02:53 -06:00
Ben Sander 3a42a7642a tolerate spaces in hip args 2017-01-09 20:57:13 -06:00
Rahul Garg 5fb09879c7 Added state for hipDevice.
Change-Id: Idbc3c04cd054a01b634856a1e0a23ff172e991aa
2017-01-09 23:54:01 +05:30
Ben Sander c325c988b1 Support size_t in memset kernel.
Add disable for HSA_AMD_AGENT_INFO_MAX_WAVES_PER_CU
Remove one copy of completion_future in memset.
2016-12-22 12:25:09 -06:00
Ben Sander 37d8cafb12 Increment API sequence number.
Change name to tls_tidInfo
2016-12-21 15:30:36 -06:00
Rahul Garg fbf7ed63a8 Fix for HCSWAP-67
Change-Id: I0b2ce5ab933237947fb41d89769db3da16e5be6a

Conflicts:
	src/hip_hcc.cpp
2016-12-19 16:19:51 +05:30
Ben Sander 6ed7e1c1c1 Remove USE_DISPATCH_HSA_KERNEL=0 path. 2016-12-17 07:22:56 -06:00
pensun 778c6626fd HIP resource leaks fix from Jack
Change-Id: I93f3ad7cb94ff1cba1577bd8acc90e826693d12e
2016-12-05 20:21:33 -06:00
Maneesh Gupta 46ffc69557 Revert "Enable USE_DISPATCH_HSA_KERNEL."
This reverts commit 097e4eb9d8.
2016-12-05 16:55:26 +05:30
Ben Sander 097e4eb9d8 Enable USE_DISPATCH_HSA_KERNEL.
Optimize hipLaunchModule dispatch latency.
2016-12-04 00:13:19 -06:00
Ben Sander ff2f54c1bf Add additional controls for forcing serialization and blocking.
Move HIP_COHERENT_HOST_ALLOC so it is read once at init time.
Add HIP_LAUNCH_BLOCKING_KERNELS, HIP_API_BLOCKING.
Update docs on debug and chicken bits.

Conflicts:
	src/hip_hcc.cpp
2016-12-02 18:03:59 -06:00
pensun 0dfcd3e664 Change to use produce device name by default
Change-Id: Ie2cee2a2e94a08b5874a2f5abee5d1ab6c9fdf47
2016-11-29 11:34:06 -06:00
Ben Sander ce92a53f25 Add more debug info 2016-11-26 08:56:02 -06:00
Ben Sander dec59d9909 Improve docs in some places
Change-Id: If31e84fbf0c8595ca72edb842dce7ce47783579b
2016-11-23 08:16:18 -06:00
Ben Sander b6ae6b08fb Improve debug capabilities.
Print TID mapping at init when HIP_TRACE_API=1.
Print base host/dev info from tracker during copy.

Change-Id: I84e26d7b801567e5a91baad36126fb590920ec87
2016-11-23 08:16:18 -06:00
Rahul Garg 2dcf20ac6f Removed hsaKmtReleaseSystemProperties call
Change-Id: I7cb992cccf587c333f0ca0cb518409f3944bdb06
2016-11-22 06:15:35 +05:30
Maneesh Gupta c0419cc749 Refactor for building HIP as dynamic library
Change-Id: I65a3d9d589c4fdbbdcf1611e5427224253be8260
2016-11-18 14:33:20 +05:30
Aditya Atluri 12dd9df88f Added i8 packed math intrinsics
1. Added add, sub, mul packed math i8 intrinsics
2. Removed c++ packed data structures included from HCC

Change-Id: I1d109c5ce10c48b7cd3ea059478b88fc1de78499
TODO: Add better packed data structures support
2016-11-17 01:09:12 -06:00
Maneesh Gupta 888a3528d2 Enable USE_COPY_EXT_V2 by default
Change-Id: I2c0dc80f85a0ccb5744715b5418a604e38b249ed
2016-11-15 10:42:27 +05:30
Aditya Atluri abf6872b2b fixed multi-dim module kernel launch
Change-Id: Id1d81f2375d058979ab526433f905cf0ea3d23d6
2016-11-11 12:25:23 -06:00
Ben Sander 1e5515ee9f Add option to deny peer access.
Also fix test.

Change-Id: I1b247f6c4271442b008e560669bca4daf8eb94c7
2016-11-10 23:12:48 -06:00
Ben Sander 65584e48de Use forceUnpinnedCopy to resolve P2p corner cases.
Change-Id: I2aebb419881246cebb696bec87798635bc71acc2
2016-11-10 23:12:48 -06:00
Ben Sander d3d6feb4de Enable async copy again.
Also add HIP_FORCE_SYNC_COPY chicken bit.

Change-Id: I76a385410494b99bf27305d3c08f55dd81987565
2016-11-10 23:12:48 -06:00
Ben Sander ced9d72d94 Refactor copy and P2P logic.
Prefer use of source-engine for DMA copies, even if user submits copy
in a stream attached to a different device.
The stream is now used only for synchronization, and HIP
makes the most optimal decision for which engine to perform the
copy - typically the source copy engine.

HIP now makes decision on which engine should perform the copy
and passes this to HCC using new apis.
HIP has additional information about peer
visibility and will make a decision which agent should perform
the copy .

Change-Id: I0cf4cfebeae256e6ca795f08a7ed7130f4857d1f
2016-11-10 23:12:48 -06:00
Ben Sander d728819d17 Improve Peer support and testing.
Change-Id: Icadc65988aaf145a265587ab0357c5bf4d26f3eb
2016-11-06 03:22:36 -06:00
Ben Sander 092b3dacda Set forceHostCopyEngine for other copy dirs. Support HIP_FORCE_P2P_HOST
Also: more debug for copy and P2p.

Change-Id: I87030c525410e041b2a00baaf6c68e6c0977ff42
2016-11-04 19:53:23 -05:00
Ben Sander 5d79384832 Refactor resolve-mem step1
Change-Id: I7b8b2bbb56d7b31a97b48ebd42002641cd07a460
2016-11-04 09:37:56 -05:00
Ben Sander 3f0a2b8dc1 Add debug for Peer APIs. Enable PeerMemcpy APIs by default.
Change-Id: I46e39a9e7b07686a78484c1f3b5495b08e052fbb
2016-11-04 08:51:16 -05:00
Aditya Atluri f48c53534e added inter thread data movement intrinsics
Change-Id: I2a8a8ed49429cb7f96439bd28c4b83b5142737df
2016-11-01 16:37:33 -05:00
Ben Sander 024d9ab090 Print short hipLaunchKernel correctly.
Change-Id: I6ca03d7c707cd03d6982199830213953d5855f17
2016-10-27 23:09:32 -05:00
Ben Sander bb58f4f6fc Add initial hipProfileStart/Stop
And modify sample to show how to use.
Still needs some work to understand interaction with CXL.

Change-Id: I2579824d2dd7863ea23874d34f0dabb3cb305d3e
2016-10-27 23:09:32 -05:00
Ben Sander ef8eac9b66 Add two levels of HIP_PROFILE_API (1=short,2=long)
Change-Id: I7ef98589f8731fb879db109fd573c62b489f2b61
2016-10-27 23:09:31 -05:00
Ben Sander ab1836544a Fix scoped marker so begin/end ATP timestamps correct
Change-Id: Ic944d3fc00d7bc31b756c0e6c327b99eb489537e
2016-10-27 23:09:31 -05:00
Ben Sander e9056798f6 Rename HIP_ATP_MARKER and profiling vars
HIP_PROFILE_API
HIP_DB_START_API
HIP_DB_STOP_API

Change-Id: I6c4da67212ff8217e6356a2622d4c6278a188c34
2016-10-27 23:09:31 -05:00
Ben Sander f5e8090f2f Allow HIP_DB to be number or string flags (ie HIP_DB=api+mem+sync)
Add callbacks for processing env vars.

Change-Id: I4ddf50e2da56b1dae43f50657bc693b07b23c03d
2016-10-27 23:09:31 -05:00
Ben Sander 710be682ca Add HIP_PROFILE_START_API, HIP_PROFILE_STOP_API
Refactor HIP_INIT_API to call recordApiTrace.

Change-Id: Ieff4b5018236f59e49e1b9841474440a34f821df
2016-10-27 23:09:31 -05:00
Ben Sander 739bc37503 Add per-thread API seqnum to debug
Change-Id: Ib13733a3e84cd56bae13a32bae40f936c20b7543
2016-10-27 23:09:31 -05:00
Ben Sander 346c519ace Improve HIP TID printing in debug mode.
Map long thread-id to a short one that is printed with each message.
Remove clunky stirng creation code for tid_tr.
Print TID on every message.

Change-Id: I780a91d8ce789cb4957789036b478bf5cde8c4e4
2016-10-27 23:09:31 -05:00
Aditya Atluri e1c1b4c009 reverted change for cache size query
Change-Id: I44a1f43818cd287a2a3b6265f43d183f9bd5b71c
2016-10-25 11:03:35 -05:00
Aditya Atluri 820a914b98 correct cachesize to output correct value
Change-Id: I5db031591eb718b0c12e78a35e4b19349de9526d
2016-10-25 09:33:45 -05:00
Ben Sander 203348ac9b Fix P2P for async
Also improve HIP debug message: Add more DB_COPY1 messages. memcpyStr,
expand HIP_DB bitmask.
2016-10-20 10:02:23 -05:00
Ben Sander 000d75de95 Add HIP_WAIT_MODE env var.
Also weaken cases where hipSetDeviceFlags returns hipErrorInvalidValue.

Change-Id: I7f113338be6fe498eaf1ab40fd0fd6b23849bb5e
2016-10-18 22:27:16 -05:00