Code

support for extended memops#59

Code

Merged

e-ago merged 99 commits intomastergpudirect/libgdsync:masterfrom new_memopsgpudirect/libgdsync:new_memopsCopy head branch name to clipboardMay 30, 2018

+3,482-858

Commits

Commits on May 10, 2018

• fixes for experimental CUDA MemOps

  drossetti

  committed
  69fe4e9

• fix bug such that GDS_DISABLE_WEAK_CONSISTENCY was not effective

  drossetti

  committed
  e8461b6

• bump version to 2.2

  drossetti

  committed
  4a4a2be

• add support for WRITE_MEMORY

  drossetti

  committed
  2e4426c

• introduce GDSCHECK

  drossetti

  committed
  1b95132

• use gds_prepare_write_memory to init descriptor

  drossetti

  committed
  638dc49

• use gpu.h tracing funs

  drossetti

  committed
  220ea6a

• convert most of prints into gpu_dbg()

  drossetti

  committed
  9fac000

• translate CUDA_ERROR_NOT_SUPPORTED, propagate error from gds_stream_batch_ops in gds_stream_post_descriptors

  drossetti

  committed
  9a36de6

• remove most of tools APIs, easily replaced by using gds descriptors

  drossetti

  committed
  de395b0

• rename GDS_DISABLE_INLINECOPY to GDS_DISABLE_WRITEMEMORY

  drossetti

  committed
  64c6104

• replacing gds_stream_post_poll_dword with gds_prepare_wait_value32+gds_stream_post_descriptors in gds_poll_lat

  drossetti

  committed
  9a2d701

• introduce -w (use_wrmem) to enable use of WRITE_MEMORY. add GDSCHECK to API calls.

  drossetti

  committed
  dfeb15d

• support global id attribute. should solve issue #10

  drossetti

  committed
  51aa395

• use gpu debug/error/tracing functions. refactor code to support non-peersync mode (still not working)

  drossetti

  committed
  f95f717

• get rid of autogen.sh, use autoreconf instead

  drossetti

  committed
  2c0539f

• fix CU_STREAM_MEMORY_BARRIER flags

  drossetti

  committed
  235b206

• improve error reporting

  drossetti

  committed
  c5bbf41

• refactor passing extra configure params

  drossetti

  committed
  2388e32

• cleanups, passing on K40m

  drossetti

  committed
  9e1731a

• increase CHUNK_SIZE to 16 in gds_sanity

  drossetti

  committed
  be2e8cd

• use CUDATK if set

  drossetti

  committed
  b5efbca

• fix merge breakages

  drossetti

  committed
  186b2b4

• don't use extended memops by default

  drossetti

  committed
  bbaf2e2

Commits on May 11, 2018

• move CUDA <= 9.2 memop check in general configure section. add checking for new memops in --enable-extended-memops section.

  drossetti

  committed
  5f56b54

• introduce OFED and GDRCOPY env vars. pass --enable-extended-memops by default.

  drossetti

  committed
  81d3196

• introduce GDS_MEMBAR_MLX5 membar flag as an optimization. use it in gds_fill_membar, assuming mlx5

  Show description for f49eb80
  drossetti

  committed
  f49eb80

• consistently deploy HAS_DECL macros. refactor weak consistency enabling code and make it per-device. abstract per-device max_batch_size.

  drossetti

  committed
  6afebc4

• warm gpu_stream_client as well

  drossetti

  committed
  8d1f7bb

• !peersync is not supported so message accordingly

  drossetti

  committed
  edce254

Commits on May 15, 2018

• add support for NVTX style profiling annotations

  drossetti

  committed
  0c2285e

Commits on May 16, 2018

• fix bug such that nvtx was always enabled

  drossetti

  committed
  86ae4a8

• introduce gds_post_descriptors

  drossetti

  committed
  e7e4589

• fix decl of gds_post_descriptors

  drossetti

  committed
  08f0ab1

• implement missing cases in gds_post_descriptors and gds_post_ops_on_cpu. add atomic helpers

  drossetti

  committed
  9b93e8c

• fix typo in comment

  drossetti

  committed
  a75b939

• add gpu_warnc & gpu_warn_once

  drossetti

  committed
  85ee32c

• make gpu_launch_calc_kernel_on_stream calculate grid

  drossetti

  committed
  0577d24

• warn once if size==0

  drossetti

  committed
  6c27f5a

• route dbg macro to gpu_dbg. introduce --skip-kernel-launch option, while --size=0 launches a void kernel. Implement !peersync case in terms of CUDA stream callbacks, relies on CPU posting descripto…

  Show description for 5eb44b0
  drossetti

  committed
  5eb44b0

• fix help test for -P option

  drossetti

  committed
  33bd1b9

Commits on May 17, 2018

• add support for WRITE_MEMORY

  drossetti

  committed
  733d772

• introduce GDSCHECK

  drossetti

  committed
  debd246

• use gds_prepare_write_memory to init descriptor

  drossetti

  committed
  8c62fb4

• use gpu.h tracing funs

  drossetti

  committed
  90ad6f7

• convert most of prints into gpu_dbg()

  drossetti

  committed
  ba15238

• translate CUDA_ERROR_NOT_SUPPORTED, propagate error from gds_stream_batch_ops in gds_stream_post_descriptors

  drossetti

  committed
  2ca857d

• remove most of tools APIs, easily replaced by using gds descriptors

  drossetti

  committed
  0e407cc

• rename GDS_DISABLE_INLINECOPY to GDS_DISABLE_WRITEMEMORY

  drossetti

  committed
  2f23a59

• replacing gds_stream_post_poll_dword with gds_prepare_wait_value32+gds_stream_post_descriptors in gds_poll_lat

  drossetti

  committed
  eadd319

• introduce -w (use_wrmem) to enable use of WRITE_MEMORY. add GDSCHECK to API calls.

  drossetti

  committed
  6000cbe

• support global id attribute. should solve issue #10

  drossetti

  committed
  da10534

• use gpu debug/error/tracing functions. refactor code to support non-peersync mode (still not working)

  drossetti

  committed
  e94c87a

• get rid of autogen.sh, use autoreconf instead

  drossetti

  committed
  59caac4

• fix CU_STREAM_MEMORY_BARRIER flags

  drossetti

  committed
  f2102b4

• improve error reporting

  drossetti

  committed
  05bb8ed

• refactor passing extra configure params

  drossetti

  committed
  bad134f

• cleanups, passing on K40m

  drossetti

  committed
  2c99cc8

• increase CHUNK_SIZE to 16 in gds_sanity

  drossetti

  committed
  98933af

• use CUDATK if set

  drossetti

  committed
  ac20a78

• don't use extended memops by default

  drossetti

  committed
  50fb20a

• move CUDA <= 9.2 memop check in general configure section. add checking for new memops in --enable-extended-memops section.

  drossetti

  committed
  e232f25

• introduce OFED and GDRCOPY env vars. pass --enable-extended-memops by default.

  drossetti

  committed
  b4d5e7f

• introduce GDS_MEMBAR_MLX5 membar flag as an optimization. use it in gds_fill_membar, assuming mlx5

  Show description for 09055e3
  drossetti

  committed
  09055e3

• consistently deploy HAS_DECL macros. refactor weak consistency enabling code and make it per-device. abstract per-device max_batch_size.

  drossetti

  committed
  19753e4

• warm gpu_stream_client as well

  drossetti

  committed
  d02e5cc

• !peersync is not supported so message accordingly

  drossetti

  committed
  778dead

• Merge pull request #63 from gpudirect/streamcb

  Show description for 63f5955
  e-ago

  authored
  63f5955

• Merge branch 'new_memops' into nvtx

  e-ago

  authored
  e9726b7

• Merge pull request #62 from gpudirect/nvtx

  Show description for c1470fc
  e-ago

  authored
  c1470fc

Commits on May 18, 2018

• fix free on error in stream callback. force use_desc_apis when !peersync.

  drossetti

  committed
  6227ee4

• handle stream callback errors

  drossetti

  committed
  95a9a30

• handle stream callback errors. introduce --gpu-mem (replacing gpu_id==-1) and --skip-kernel-launch (-K) options, turn -K proto option into -k (lower case). comment out useless code for RC protocol.…

  Show description for 7be8f7b
  drossetti

  committed
  7be8f7b

• Merge pull request #65 from gpudirect/fixtests

  Show description for 80df4c7
  e-ago

  authored
  80df4c7

Commits on May 22, 2018

• NVTX annotate pp_post_work()

  drossetti

  committed
  fa155dd

• introduce --hide-cpu-launch-latency/-L option to ease benchmarking.

  Show description for babd92c
  drossetti

  committed
  babd92c

• cleanup: remove unused options

  drossetti

  committed
  7f6866c

• refactor gpu registration to accomodate on-demand registration happening before QP/CQ creation

  Show description for d671068
  drossetti

  committed
  d671068

• improve help. actually implement -m option. fix detection of WRITE_MEMORY.

  drossetti

  committed
  2f8d3d7

• remove stale text

  drossetti

  committed
  27ac382

• remove cudaDeviceSynchronize from NVTX macros. always flush stderr in gpu_msg().

  drossetti

  committed
  834f026

• add a comment clarifying that gds_ordinal_from_device() needs a better solution

  drossetti

  committed
  78e2c84

• add/fix options help text

  drossetti

  committed
  d83b81c

• add more tracing. fix PROF usage for !peersync case and make decrese resolution.

  drossetti

  committed
  f771a63

Commits on May 23, 2018

• add GDS_ASSERT utility function

  drossetti

  committed
  36a8f39

• fortify gds_ordinal_from_device implementation

  drossetti

  committed
  904bc6e

Commits on May 24, 2018

• fix implementation of GDS_ASSERT()

  drossetti

  committed
  9240b4d

• cleanup warnings of Ubuntu

  drossetti

  committed
  475813b

• abort if relaxed ordering is not consistently detected

  drossetti

  committed
  7c77f2e

• fix bug when calling gds_prepare_write_memory when CU_STREAM_MEM_OP_WRITE_MEMORY is not available

  drossetti

  committed
  c0965b9

• add dbg tracing for POLL_NOR

  drossetti

  committed
  e15534b

Commits on May 25, 2018

• document --gpu-mem option in help

  drossetti

  committed
  d767910

• improve messaging in case of errors in support_weak_consistency

  drossetti

  committed
  3a8faef

• fix more warnings on Ubuntu

  drossetti

  committed
  23da4fa

• remove stale code. introduce GDS_DISABLE_WAIT_NOR off by default. fix error prints.

  drossetti

  committed
  929516e

• disabling WAIT NOR as a work-around for #68

  drossetti

  committed
  2fb9295

Commits on May 29, 2018

• error prints MPI rank

  drossetti

  committed
  c11aee7

• Merge branch 'new_memops' of https://github.com/gpudirect/libgdsync into new_memops

  drossetti

  committed
  1489810

Commits on May 30, 2018

• Merge branch 'master' into new_memops

  e-ago

  authored
  4e430de