Code

[STF] stackable stf resources#2674

Code

Draft

caugonnet wants to merge 86 commits intoNVIDIA:mainNVIDIA/cccl:mainfrom caugonnet:stackable_ctx_datacaugonnet/cccl:stackable_ctx_dataCopy head branch name to clipboard

+2,890-50

Commits

Commits on Nov 7, 2024

• Introduce "stackable" resources for to improve how we nest contexts

  caugonnet

  committed
  56b5e9f

  Copy full SHA for 56b5e9f

• Add missing constructor

  caugonnet

  committed
  7b2bd58

  Copy full SHA for 7b2bd58

• Add two more tests with stacked resources

  caugonnet

  committed
  95869b0

  Copy full SHA for 95869b0

• fix copyright notice

  caugonnet

  committed
  771bcb0

  Copy full SHA for 771bcb0

• fix formatting issue in the doxygen comment

  caugonnet

  committed
  438bfa3

  Copy full SHA for 438bfa3

• remove dead code and debug printfs

  caugonnet

  committed
  22ba9d2

  Copy full SHA for 22ba9d2

• replace assert by _CCCL_ASSERT

  caugonnet

  committed
  0f9425b

  Copy full SHA for 0f9425b

• Replace at by a [] operator

  caugonnet

  committed
  c68c5fb

  Copy full SHA for c68c5fb

• clang-format

  caugonnet

  committed
  dd8b008

  Copy full SHA for dd8b008

• avoid copies

  caugonnet

  committed
  f78594f

  Copy full SHA for f78594f

• Cleanup the stackable resource implementation

  caugonnet

  committed
  fb3de98

  Copy full SHA for fb3de98

Commits on Jan 14, 2025

• Merge branch 'main' into stackable_ctx_data

  caugonnet

  committed
  3d1605d

  Copy full SHA for 3d1605d

• Automatically call pop() on (stackable) logical data when the pop() method of the context is called (if logical data's pop() was not called explicitly).

  caugonnet

  committed
  cd63feb

  Copy full SHA for cd63feb

• change test to automatically pop data

  caugonnet

  committed
  dc7e04a

  Copy full SHA for dc7e04a

Commits on Jan 15, 2025

• stackable_ctx::impl is not copyable but movable

  caugonnet

  committed
  bdddfe9

  Copy full SHA for bdddfe9

• clang-format

  caugonnet

  committed
  8902d6f

  Copy full SHA for 8902d6f

• stackable_logical_data::impl is movable but not copyable

  caugonnet

  committed
  7c72f4b

  Copy full SHA for 7c72f4b

• unittests are broken as they compare a fullpath with a relative path in the wrong directory : so we record the full path so that it works at runtime

  caugonnet

  committed
  8dce985

  Copy full SHA for 8dce985

• Save WIP : add unittests in the stackable_ctx.cuh header

  caugonnet

  committed
  de5317b

  Copy full SHA for de5317b

• Add a missing ::std

  caugonnet

  committed
  38f7b2b

  Copy full SHA for 38f7b2b

Commits on Jan 16, 2025

• WIP: create stackable_task_dep which inherit from task_dep and parse that

  caugonnet

  committed
  a68b769

  Copy full SHA for a68b769

Commits on Jan 17, 2025

• Save WIP : try to create stackable logical data at the root ctx so that we can export them to lower contexts

  caugonnet

  committed
  90e9975

  Copy full SHA for 90e9975

Commits on Jan 26, 2025

• Use a stream_adapter for the allocator of the inner graph in an algorithm

  caugonnet

  committed
  8e5a436

  Copy full SHA for 8e5a436

Commits on Jan 27, 2025

• Merge branch 'main' into stackable_ctx_data

  caugonnet

  authored
  f36be8c

  Copy full SHA for f36be8c

• Merge branch 'NVIDIA:main' into stf_algorithm_alloc_adapter

  caugonnet

  authored
  0656dde

  Copy full SHA for 0656dde

Commits on Jan 28, 2025

• Merge branch 'main' into stf_algorithm_alloc_adapter

  caugonnet

  authored
  6675af2

  Copy full SHA for 6675af2

• Entirely rework the stream_adapter implementation so that it is movable and to check clear() was called, factorize code to setup allocators in algorithms

  caugonnet

  committed
  b1b9c62

  Copy full SHA for b1b9c62

• Add a new tests and some comments

  caugonnet

  committed
  ed1f244

  Copy full SHA for ed1f244

• Merge branch 'stf_algorithm_alloc_adapter' into stackable_ctx_data

  caugonnet

  committed
  995fea4

  Copy full SHA for 995fea4

• Save work in progress with stackable resources

  caugonnet

  committed
  885307e

  Copy full SHA for 885307e

Commits on Jan 29, 2025

• Add a test to see if we can access stackable logical data produced in a nested context

  caugonnet

  committed
  cf3694e

  Copy full SHA for cf3694e

• clang-format

  caugonnet

  committed
  5fa37d5

  Copy full SHA for 5fa37d5

• rename l to ld for better clarity

  caugonnet

  committed
  7b106f1

  Copy full SHA for 7b106f1

• Export "out" to a context below the one it was created in

  caugonnet

  committed
  228e033

  Copy full SHA for 228e033

• Reenable the stream_adapter allocator and add some information for DOT

  caugonnet

  committed
  f62b7f5

  Copy full SHA for f62b7f5

• save WIP

  caugonnet

  committed
  8a4c1c4

  Copy full SHA for 8a4c1c4

• Change how we track data which should be popped automatically when a stackable_ctx is popped

  caugonnet

  committed
  4e6f1c1

  Copy full SHA for 4e6f1c1

• add sanity check

  caugonnet

  committed
  cdd9a88

  Copy full SHA for cdd9a88

Commits on Jan 30, 2025

• Implement a mechanism to retain stackable_logical_data created in a nested

  Show description for 95cfd1d
  caugonnet

  committed
  95cfd1d

  Copy full SHA for 95cfd1d

• Do not use unique_ptr for the sake of simplicity in the fhe example

  caugonnet

  committed
  cac2f9a

  Copy full SHA for cac2f9a

• Avoid ambiguous inheritance with reserved::to_task_dep which converts a (pack of) stackable_task_dep to its underlying task_dep

  caugonnet

  committed
  f268800

  Copy full SHA for f268800

• stackable_task_dep does not inherit on task_dep anymore and add a missing to_task_dep overload

  caugonnet

  committed
  c3c1412

  Copy full SHA for c3c1412

• Merge branch 'main' into stackable_ctx_data

  caugonnet

  authored
  53021f9

  Copy full SHA for 53021f9

Commits on Jan 31, 2025

• Add missing overloads

  caugonnet

  committed
  cda57e2

  Copy full SHA for cda57e2

• Merge branch 'main' into stackable_ctx_data

  caugonnet

  committed
  05051b0

  Copy full SHA for 05051b0

• Add more missing overloads, add some annotations in tests for push/pop

  caugonnet

  committed
  f969e52

  Copy full SHA for f969e52

• Add a test which moves stackable data

  caugonnet

  committed
  c82484a

  Copy full SHA for c82484a

• clang-format

  caugonnet

  committed
  cc37f78

  Copy full SHA for cc37f78

• stackable_ctx::logical_token and another stackable_ctx::logical_data overload

  caugonnet

  committed
  9619ca6

  Copy full SHA for 9619ca6

• clang-format

  caugonnet

  committed
  8132468

  Copy full SHA for 8132468

• Test if we can freeze tokens, or use them as stackable data (WIP)

  caugonnet

  committed
  bbdfb98

  Copy full SHA for bbdfb98

• Adapt get_frozen to work with tokens

  caugonnet

  committed
  5b10614

  Copy full SHA for 5b10614

• Implement PDGEMM with stacked contexts

  caugonnet

  committed
  6945e6b

  Copy full SHA for 6945e6b

• Add a test to see if write-back mechanisms are enforced properly

  caugonnet

  committed
  9629e5f

  Copy full SHA for 9629e5f

• Erase the logical data before unfreezing the instance

  caugonnet

  committed
  e658104

  Copy full SHA for e658104

• Add a missing inline

  caugonnet

  committed
  2ac38e1

  Copy full SHA for 2ac38e1

• Add a more complex export test

  caugonnet

  committed
  4339e3d

  Copy full SHA for 4339e3d

Commits on Feb 1, 2025

• Avoid intermediate variable

  caugonnet

  committed
  9784959

  Copy full SHA for 9784959

• improve synchronization test after ctx.pop()

  caugonnet

  committed
  23d11e2

  Copy full SHA for 23d11e2

• Add access_mode_is_compatible to check if one access mode is compatible with a previous one

  caugonnet

  committed
  1e9cc86

  Copy full SHA for 1e9cc86

• Add a method to update the logical data within a task_dep_untyped object

  caugonnet

  committed
  6143948

  Copy full SHA for 6143948

• Implement automatic stackable_logical_data push

  caugonnet

  committed
  85b6492

  Copy full SHA for 85b6492

• Remove debug messages

  caugonnet

  committed
  0b370d9

  Copy full SHA for 0b370d9

Commits on Feb 2, 2025

• WIP: rework how we destroy stackable_logical_data and how we defer the unfreeze operation after the finalize stage

  caugonnet

  committed
  8c4d3c3

  Copy full SHA for 8c4d3c3

• avoid calling the default ctor of stackable_ctx by error

  caugonnet

  committed
  1ad199e

  Copy full SHA for 1ad199e

• If we have a retained data, then the logical data was already destroyed

  caugonnet

  committed
  08f1298

  Copy full SHA for 08f1298

• remove a debug printf

  caugonnet

  committed
  4799218

  Copy full SHA for 4799218

• clang-format

  caugonnet

  committed
  7a4493b

  Copy full SHA for 7a4493b

Commits on Feb 3, 2025

• Remove the unit test to promote access modes in stackable_logical_data

  caugonnet

  committed
  2ba4c55

  Copy full SHA for 2ba4c55

• Merge branch 'main' into stackable_ctx_data

  caugonnet

  authored
  0183f71

  Copy full SHA for 0183f71

• Implement frozen_logical_data::get_access_mode to query the access mode used to

  Show description for 96c7896
  caugonnet

  committed
  96c7896

  Copy full SHA for 96c7896

• Merge branch 'stf_frozen_get_access_mode' into stackable_ctx_data

  caugonnet

  committed
  0ca15f3

  Copy full SHA for 0ca15f3

• Use the .get_access_mode method of the frozen_logical_data object to avoid storing modes separately in the frozen_modes vector

  caugonnet

  committed
  1cc523d

  Copy full SHA for 1cc523d

• Merge branch 'main' into stackable_ctx_data

  caugonnet

  authored
  d3461b1

  Copy full SHA for d3461b1

• Implement logical_data_no_export on shapes

  caugonnet

  committed
  aa73deb

  Copy full SHA for aa73deb

• retain data state only at a specific depth

  caugonnet

  committed
  bd16b6b

  Copy full SHA for bd16b6b

• More overloads for logical_data_no_export and fix a move error

  caugonnet

  committed
  db0baca

  Copy full SHA for db0baca

Commits on Feb 4, 2025

• retain at the appropriate level

  caugonnet

  committed
  7f997f0

  Copy full SHA for 7f997f0

Commits on Feb 5, 2025

• remove the need_untrack argument which is always false now

  caugonnet

  committed
  5163fdc

  Copy full SHA for 5163fdc

• Add comments to explain the implementation

  caugonnet

  committed
  981eb49

  Copy full SHA for 981eb49

• We don't use untrack_pushed_data anymore

  caugonnet

  committed
  2ecd810

  Copy full SHA for 2ecd810

• clang-format

  caugonnet

  committed
  1a4687c

  Copy full SHA for 1a4687c

• Add a test to ensure we detect improper accesses to non exportable data

  caugonnet

  committed
  64a8bba

  Copy full SHA for 64a8bba

• clang-format

  caugonnet

  committed
  baa2613

  Copy full SHA for baa2613

Commits on Feb 6, 2025

• Reimplement chain of CUDA kernels in the CUDA graph backend to avoid child graphs

  caugonnet

  committed
  72dc649

  Copy full SHA for 72dc649

• Merge branch 'stf_kernel_chain_graph' into stackable_ctx_data

  caugonnet

  committed
  392283f

  Copy full SHA for 392283f