[FEA] LTO IR Support (3) - Replace JITIFY usage with LIBRTCX by lamarrr · Pull Request #22654 · NVIDIA/cudf

lamarrr · 2026-05-24T05:55:55Z

Description

This PR:

Replaces all uses of JITIFY with LIBRTCX
Minor refactoring of LibRTCX
Adds new features:
- JIT Cache preloading
- NVTX ranges on all compilation/linking steps
- CUDA cache management (LIBCUDF_JIT_DISABLE_CUDA_CACHE)
- More benchmarking and profiling functionalities (LIBCUDF_JIT_DUMP_TRACE, LIBCUDF_JIT_DUMP_TIME_PROFILE, LIBCUDF_KERNEL_CACHE_DISABLED)
- Removes the 64-bit source-code workaround _FILE_OFFSET_BITS=64 (replaced by ASM embedding)

Immediate Impact

Reduces libcudf.so from 516 MB to 494 MB, a 4.26% binary size decrease due to LIBRTCX's compressed CUDA source code embedding, unlike JITIFY's per-kernel preprocessed CUDA source strings

Checklist

I am familiar with the Contributing Guidelines.
New or existing tests cover these changes.
The documentation is up to date with these changes.

- Updated CMakeLists.txt to remove unnecessary comments and added src/jit/jit.cpp to the build. - Enhanced JitEmbed.cmake for better dependency management and code clarity. - Renamed structures in column_view.cuh for consistency and clarity. - Updated thunk.cuh to reflect new structure names and ensure proper mapping. - Modified types.cuh to align with new structure names and improve documentation. - Cleaned up transform.hpp by removing unnecessary whitespace. - Introduced jit.cpp and jit.hpp for improved JIT compilation handling. - Added README.md for documentation on JIT LTO functionality. - Updated cache.hpp and rtc.cpp to include necessary checks and improve error handling. - Refactored context.cpp and context.hpp to integrate new JIT bundle structure. - Adjusted transform2.cu to ensure proper namespace usage. - Updated rtc.cpp test cases to reflect changes in structure names and improve test clarity.

- Introduced new header files in `librtcx/system_includes` for various standard library components, including algorithm, array, atomic, barrier, bit, cassert, complex, and more, all utilizing the CUDA standard library. - Updated the `context` class to replace the old RTC cache implementation with the new `rtcx` namespace, ensuring proper initialization and usage of the cache. - Refactored includes and namespaces to maintain consistency and improve code organization.

… and remove deprecated nullable_element functions

…ndle initialization

…orm2.cu, and update export.hpp copyright year

… and add new compiler flags

…y_uncached function

… and improve column handling - Updated `transform_input` to remove `mutable_column_view` and streamline input handling. - Introduced `transform_output` struct to specify output types and nullability policies. - Modified `transform_extended2` to return a `table` instead of a vector of columns, enhancing output management. - Refactored `column_accessor` to utilize `detail::column_device_view_base` for better abstraction. - Removed deprecated reflection functions for input accessors and adjusted related helper functions. - Enhanced `transform_udf` to work with the new column view structure, ensuring compatibility with the updated input/output specifications. - Updated kernel launch and argument handling to accommodate changes in input and output types. - Improved nullability checks and handling in the transform execution flow. - Added support for pre-allocated string offsets in output columns to optimize memory usage.

…orm-support

…olumn_device_view_core

lamarrr · 2026-05-28T08:56:33Z

/ok to test 2d0e7c2

lamarrr · 2026-05-28T09:00:23Z

Can we remove https://github.com/rapidsai/cudf/blob/2d4cc67d7bced44c48ee7efa8d6096372daf9032/cpp/include/cudf/detail/operation-udf.hpp?

No, we can't, I accidentally omitted it for some of the kernels while re-applying my local patch files.
It is required not just for the explainability of the flow but also to instantiate the kernels, as NVCC/NVRTC can't have defines with complex CUDA strings in them.

…cudf into lto-ir-3--rtcx-adoption

lamarrr · 2026-05-28T09:15:52Z

/ok to test 2d0e7c2

copy-pr-bot · 2026-05-28T09:15:57Z

/ok to test 2d0e7c2

@lamarrr, there was an error processing your request: E2

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/2/

lamarrr · 2026-05-28T09:16:07Z

/ok to test da973b9

vyasr

I'm comfortable moving forward with this so we can start on the proposed next steps of the split.

lamarrr · 2026-06-02T01:12:08Z

awaiting re-review from @bdice

bdice

Nice work. I have a few minor comments. Please merge when you're ready.

lamarrr · 2026-06-05T16:37:29Z

thanks for the feedback @bdice & @vyasr

lamarrr · 2026-06-05T16:38:06Z

/merge

lamarrr added 30 commits February 11, 2026 08:46

changed symbol visibility

613eb37

Rename struct jit_statistics to jit_compilation_stats for clarity

1d53fc8

cmake dependency fix

112ae32

refactoring

32383c0

refactoring

21a2610

checkpoint

badacea

started making operators AST-based

6f829ac

update

ec95217

update

5e0f2e7

Merge remote-tracking branch 'upstream/main' into lto-ir-expr-ast-based

93b57a5

update

5ffddda

Merge remote-tracking branch 'upstream/main' into lto-ir-expr-ast-based

4414b4e

Merge remote-tracking branch 'upstream/main' into drop-jitify

7e5fb74

remove faked headers

ca50cc9

Update conda environment files to require python-zstd version 1.5.7.2…

db2e2dd

… and remove deprecated nullable_element functions

Refactor init_flags and initialization logic to remove RTC and JIT bu…

75916e4

…ndle initialization

Clean up JIT-related code: remove RTC test file, delete unused transf…

8bb341d

…orm2.cu, and update export.hpp copyright year

Refactor JIT compilation options: streamline precompiled header usage…

4bb406a

… and add new compiler flags

Remove TODO comment regarding CMake tracking in jit_bundle_t constructor

e53bf52

Update log message to specify compiled CUDA library in compile_librar…

bdc6597

…y_uncached function

[FEA] Support Multi-Output JIT Transforms

6c9be74

update

e8250e2

Merge remote-tracking branch 'upstream/main' into drop-jitify

e55ba0f

Merge remote-tracking branch 'upstream/main' into multi-output-transf…

aa2fd4d

…orm-support

initial diff

f708ff9

Add [[nodiscard]] attribute to head() and data() methods in mutable_c…

69ff5e1

…olumn_device_view_core

Refactoring + add support for zero inputs

2ab3bc5

Merge branch 'main' into lto-ir-3--rtcx-adoption

2d0e7c2

lamarrr added 2 commits May 28, 2026 09:15

fix error

3846c9e

Merge branch 'lto-ir-3--rtcx-adoption' of https://github.com/lamarrr/…

da973b9

…cudf into lto-ir-3--rtcx-adoption

lamarrr requested review from bdice and vyasr May 28, 2026 15:32

vyasr mentioned this pull request May 28, 2026

Support precompiled LTO-IR in librtcx #22704

Closed

3 tasks

vyasr approved these changes Jun 2, 2026

View reviewed changes

vyasr mentioned this pull request Jun 3, 2026

refactor(rtcx): decouple librtcx from cudf and add standalone CMake build #22744

Merged

3 tasks

bdice approved these changes Jun 3, 2026

View reviewed changes

Comment thread cpp/src/jit/cache.cpp

Comment thread cpp/src/runtime/context.cpp

Comment thread cpp/src/transform/jit/kernel.cu

rapids-bot Bot merged commit 363920c into NVIDIA:main Jun 5, 2026
217 of 219 checks passed

jihoonson mentioned this pull request Jun 9, 2026

RTCX failure loading nvJitLink/nvrtc in AST CompiledExpression tests across multiple Spark shims NVIDIA/cudf-spark#14996

Closed

This was referenced Jun 9, 2026

Add filter_join_indices_output_size #22694

Merged

[FEA] LTO IR Support (4) - Implement LTO Transform Kernels #22680

Merged

coderabbitai Bot mentioned this pull request Jun 17, 2026

[FEA] RTCX: Replace Software SHA256 with SIMD-accelerated xxHash128 for JIT cache #22920

Closed

3 tasks

thirtiseven mentioned this pull request Jun 23, 2026

Docker build wrapper should provide a writable libcudf RTCX cache NVIDIA/cudf-spark-jni#4735

Closed

coderabbitai Bot mentioned this pull request Jun 30, 2026

[FEA] Support input of dictionary-encoded types in Transforms #21552

Merged

3 tasks

lamarrr mentioned this pull request Jul 1, 2026

[BUG] Repeated calls to cudaGetDeviceProperties causes launch overhead in JIT kernels #23074

Closed

This was referenced Jul 7, 2026

Use rtcx from the librtcx repository #23137

Merged

[FEA] Add Java bindings for AST JIT operations #23117

Merged

coderabbitai Bot mentioned this pull request Jul 22, 2026

[FEA] Multi-String Output & LTO Transform Examples #23173

Open

3 tasks

coderabbitai Bot mentioned this pull request Aug 1, 2026

Fix JIT LTO linking on builds targeting only feature-set variants #23506

Open

3 tasks

coderabbitai Bot mentioned this pull request Aug 14, 2026

Update rtcx #23626

Merged

3 tasks

Navigation Menu

Uh oh!

[FEA] LTO IR Support (3) - Replace JITIFY usage with LIBRTCX - #22654

[FEA] LTO IR Support (3) - Replace JITIFY usage with LIBRTCX#22654
rapids-bot[bot] merged 307 commits into
NVIDIA:mainfrom
lamarrr:lto-ir-3--rtcx-adoption

lamarrr commented May 24, 2026 •

edited

Loading

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

copy-pr-bot Bot commented May 28, 2026

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

vyasr left a comment

Uh oh!

lamarrr commented Jun 2, 2026

Uh oh!

bdice left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lamarrr commented Jun 5, 2026 •

edited

Loading

Uh oh!

lamarrr commented Jun 5, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

lamarrr commented May 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Immediate Impact

Checklist

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

copy-pr-bot Bot commented May 28, 2026

Uh oh!

lamarrr commented May 28, 2026

Uh oh!

vyasr left a comment

Choose a reason for hiding this comment

Uh oh!

lamarrr commented Jun 2, 2026

Uh oh!

bdice left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lamarrr commented Jun 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lamarrr commented Jun 5, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

lamarrr commented May 24, 2026 •

edited

Loading

lamarrr commented Jun 5, 2026 •

edited

Loading