Commit Graph

4 Commits

Author SHA1 Message Date
John Shumway
9ac51aa0f4 Add describe() method to device ops for runtime introspection (#3375)
Introduces a polymorphic describe() method to BaseOperator that enables runtime introspection of kernel configurations through a unified interface.

Key changes:

* Add virtual describe() method to BaseOperator returning Description objects
* Implement describe() in 6 device operation classes (conv fwd/bwd variants)
* Create conv_describe.hpp with factory function for ConvDescription
* Extract type definitions to conv_types.hpp to resolve circular dependencies
* Add InstanceStringDescription for kernels without full ConvDescription support

Other Improvements:

* Update tests to use describe() instead of GetInstanceString()
* Remove circular dependency include from conv_traits.hpp
* Add ODD_C to ConvFwdSpecialization enum and fix OddC mapping
* Replace silent fallback in conv_layout() with compile-time error

This provides a foundation for runtime kernel introspection and better tooling support for analyzing and debugging kernel configurations.
2025-12-14 12:49:12 -08:00
Robin Voetter
6219b12730 [CK_BUILDER] convolution testing (#3267)
* Add README.md for testing

* Add tensor_memory_manager.

* ck-builder: tensor memory manager rebase fixes

This fixes some issues caused by the API being changed recently.
Also, this streamlines the ckt namespace to always be ck_tile::builder::test,
as this is already being used by other tests

Really, this commit should be squashed into the previous,
but I'm keeping it separate for brevity.

* ck-builder: test arguments initial prototype

* ck-builder: test system initial prototype

* ck-builder: fix non-standardized copyright comments

* ck-builder: new prototype

* ck-builder: group testing inputs/outputs into a separate structure

This is basically the return of the tensor memory manager after all,
except that the design is more closely tied to the actual operation.
Using a struct allows us to add additional input/output tensors
without breaking code (by defaulting those new parameters). Note
that the tensors are split into a separate inputs/outputs because we
usually want to allocate the output _twice_: once for the real
computation and once for the reference computation.

* ck-builder: simplify prototype naming; start docs

* ck-builder: update testing readme

* ck-builder: testing documentation

* ck-builder: HipStatusMatcher

This matcher can be used to check HIP status codes and provide
nice and readable error messages.

* ck-builder: tensor_buffer.hpp tests

* ck-builder: conv_fwd.hpp tests

* ck-builder: add example end-to-end test in conv fwd 2d fp16

* ck-builder: simplify extent usage

* ck-builder: update testing doc

* ck-builder: skip end to end test on non-gfx9

* fix check_copyright_year interpreter

/bin/bash is not guaranteed to exist on Linux. Signed,
a NixOS user

* ck-builder: fix copyrights

* ck-builder: reduce conv fwd testing size

This test allocated 24GB of memory, too much for 16GB cards.

---------

Co-authored-by: John Shumway <jshumway@amd.com>
2025-12-13 15:33:41 +01:00
Ville Pietilä
d66e5f667c [CK_BUILDER] Improve CK Builder and CK Builder tests (#3382)
* Remove stale documentation.

* Add placeholder for conv algorithm design description. Add link to conv factory description.

* Improve testing transfer parameters.

* Python script to check the block tilings.

* Improve tests and conv types serialization.

* Change representation of boolean values from 1/0 to true/false in instance strings.

* Change representation of boolean values from 1/0 to true/false in conv algorithm types.

* Test code improvements.

* Improve covn descriptions tests.

* Improve conv signature definition in conv fwd builder tests.

* clang-format.

* Remove obsolete script.

* Revert StaticAssertTypeEq changes in conv layout tests.

* Remove obsolete using declaration.

---------

Co-authored-by: Ville Pietilä <>
2025-12-11 09:50:00 +02:00
Bartłomiej Kocot
04612c30ce [CK_BUILDER] Ck Tile Grouped convolution factory (#3352)
* [BUILDER] Ck Tile Grouped convolution factory

* Part 2

* Fixes after rebase

* Remove leftovers
2025-12-08 10:32:56 +01:00