Ville Pietilä
c9504f2c27
Update profiling doc.
2026-02-10 04:45:54 -05:00
Ville Pietilä
d32bdb1412
Merge branch 'vpietila/retina-net-fwd-convs' into vpietila/retina-net-training-perf
2026-02-09 06:54:16 -05:00
Ville Pietilä
b722492a30
Add BF16 example.
2026-02-09 06:36:20 -05:00
Ville Pietilä
60d1ec34a9
Improve benchmarking scripts.
2026-02-09 04:34:27 -05:00
Ville Pietilä
53201d2081
Disable some cases.
2026-02-04 10:47:12 -05:00
Ville Pietilä
7660fa0a2e
Fix gitignore.
2026-02-04 10:07:26 -05:00
Ville Pietilä
e2225e2baa
Git ignore rocprofv3 files.
2026-02-04 10:05:14 -05:00
Ville Pietilä
c97795b139
Remove .dat file.
2026-02-04 10:04:55 -05:00
Ville Pietilä
1c842f39a4
Git ignore profiler output.
2026-02-04 10:02:57 -05:00
Ville Pietilä
1c1ac4ef10
Small fixes to runner script.
2026-02-04 09:55:05 -05:00
Ville Pietilä
73b459c5a4
Runner script for benchmarking.
2026-02-04 09:38:16 -05:00
Ville Pietilä
3e6415a8ea
True baseline benchmarking results.
2026-02-04 07:56:15 -05:00
Ville Pietilä
20cf6df685
Best instances for benchmark shapes.
2026-02-04 07:27:34 -05:00
Ville Pietilä
2cfc4209bb
Profile optionally only a given instance.
2026-02-04 06:31:36 -05:00
Ville Pietilä
f6f381dbd4
Benchmarking shapes and baseline results.
2026-02-04 06:07:19 -05:00
Ville Pietilä
403f36ed26
Disable building all but fwd convs for CK profiler.
2026-02-04 06:07:06 -05:00
Ville Pietilä
d132df2bf5
Finalize conv specialization for filter 3x3, pad 1, stride 1, dilation 1 case.
2026-02-03 04:10:09 -05:00
Ville Pietilä
a814ba15fd
Add profiling documentation.
2026-02-02 11:15:46 -05:00
Ville Pietilä
bd81d645e2
Add one more specialization.
2026-02-02 03:03:13 -05:00
Bartłomiej Kocot
fbb073f276
Update device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v3.hpp
2026-01-31 20:46:58 +01:00
Jakub Piasecki
2086516deb
fixed building errors
2026-01-30 19:22:34 +00:00
Jakub Piasecki
ae2d2d9f2c
fixed conflicts
2026-01-30 18:47:19 +00:00
Bartlomiej Kocot
a7b57187cf
Grouped Convolution Backward Data Direct Load
...
Co-authored-by: Jakub Piasecki <jakpia21@gmail.com >
2026-01-30 18:45:23 +00:00
Graner, Johannes
c815e734c7
Add good instance
2026-01-30 10:50:06 -05:00
Ville Pietilä
4b7ec1bacb
Filter 3x3, pad1, stride1, dilation 1 - specialization.
2026-01-30 10:26:48 -05:00
Ville Pietilä
7ffa682bd3
Add missing applicability check to v3 fwd convs.
2026-01-30 05:05:38 -05:00
Ville Pietilä
966706bb21
Add new grouped conv instance to the gfx950 branch.
2026-01-30 04:13:36 -05:00
Ville Pietilä
fc7964462e
Conv specializations.
2026-01-29 11:50:50 -05:00
Graner, Johannes
5301efc8e4
Add NumGroupsToMerge to BwdWeight type string
2026-01-29 09:30:02 -05:00
Ville Pietilä
0fba67a7e7
Add fwd conv group merging to the v3 conv instances.
2026-01-29 08:16:11 -05:00
Ville Pietilä
d7c4775455
Improve logging.
2026-01-29 05:50:09 -05:00
Ville Pietilä
44960922a2
Merge remote-tracking branch 'origin/jograner/bwd-weight-splitk-autodeduce' into features/grouped-conv-perf-uplift
2026-01-28 10:57:40 -05:00
Ville Pietilä
c92b954537
Add new fwd conv fp16/bf16 instances optimized for unit group size.
2026-01-28 10:50:46 -05:00
Graner, Johannes
029efffeb5
Update test with new applicability
2026-01-28 09:19:41 -05:00
Graner, Johannes
0eee2d3392
Fix threshold calculation
2026-01-28 09:18:03 -05:00
Ville Pietilä
349fb5206e
Merge remote-tracking branch 'origin/develop' into vpietila/retina-net-training-perf
2026-01-28 02:38:04 -05:00
Graner, Johannes
55d8e9b4f0
Add missing logic to wmma multiple d kernel
2026-01-28 02:12:18 -05:00
damien-lejeune
91e32f305f
[CK Tile] multi reduce improvements ( #3607 )
...
* WIP: refactoring
* Swap operation/data nested loops order
* Improve memory coalescing
* Add comments
* Enforce same identity element for the reduce operations
* Re-add compile time constant
* Comment + re-add __builtin_amdgcn_readfirstlane(0) to the loop init
---------
Co-authored-by: Damien Lejeune <damien.lejeune@amd.com >
2026-01-27 12:56:09 -08:00
linqunAMD
23cefda140
[ck] add gridwise base class for in all xdl kernel ( #186 ) ( #3544 )
...
1. Add base class GridwiseGemm_xdl_cshuffle_base for all gridwise_gemm_xdl classes.
- to select correct LDS layout and epilogue behavior , three additional parameters is added.
- ForceNaiveLdsLayout: disable XOR based LDS layout when it is true
- DirectLoad: pipeline only use directload, we need force naive layout and ignore any padding on gfx9
- IsMxGemm: epilogue has two addtional dimensions
2. Move all LDS descriptor layout related fucntion to base class, including
- GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1
- GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1
- GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
3. Move several LDS related helper funtions to base class, including
- GetSharedMemoryNumberOfByte
- GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1
- GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1
- GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
4. Move all c epilogue related code to base class, and 4 kind of implementation are provided
- RunEpilogueNoShuffle
- RunEpilogue
- RunMultiDEpilogue
- RunMoeEpilogue
2026-01-27 12:49:47 -08:00
Michał Kulikowski
b737f1dee5
[CK]Refactoring threadwise_tensor_slice_transfer_v3r1.hpp ( #3263 )
...
Signed-off-by: Michal Kulikowski <Michal.Kulikowski@amd.com >
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com >
2026-01-27 10:48:16 -08:00
Illia Silin
b26cb596b0
fix some syntax errors ( #3658 )
2026-01-27 09:59:39 -08:00
Ville Pietilä
42fde4860a
Custom global mem write.
2026-01-27 11:24:29 -05:00
Ville Pietilä
970869661c
Global mem write debugging.
2026-01-27 11:23:37 -05:00
spolifroni-amd
0cc83cb8e8
CK: removed the api reference ( #3571 )
...
* removed the api reference
* updating to the latest rocm-docs-core min version
* fixed a formatting issue with buffer views
* removed reference links from code snippets
* removed reference links from code snippets
---------
Co-authored-by: John Afaganis <john.afaganis@amd.com >
2026-01-27 07:36:47 -08:00
Graner, Johannes
74eb200c73
Fix error threshold calculations
2026-01-27 09:23:12 -05:00
Graner, Johannes
ad3954f119
Enable bwd weight splitk autodeduction with cap
2026-01-27 08:46:53 -05:00
Max Podkorytov
b66597ed96
Add build time optimization documentation ( #3608 )
...
This document describes techniques for reducing C++ template instantiation
overhead in the Composable Kernel codebase, including:
- Replacing recursive templates with pack expansion (O(N) → O(1) depth)
- Using named functors instead of lambdas to share instantiations
- Replacing template recursion with constexpr loops
- Using fold expressions for accumulation operations
These techniques can significantly reduce build times for template-heavy code.
2026-01-27 06:07:27 -07:00
Bartłomiej Kocot
3d67e6c492
[CK TILE] Enable CK TILE Conv Fwd tests in CI and fix check_err ( #3624 )
...
* [CK TILE] Enable CK TILE Conv Fwd tests in CI and fix check_err
* Update test_grouped_convnd_fwd_tile.cpp
* Update test_grouped_convnd_fwd_tile.cpp
* Update conv_tuning_params.hpp
* clang format fix
* Update CMakeLists.txt
2026-01-27 11:04:11 +02:00
Johannes Graner
c190d8d61f
[CK tests] Extend conv GPU reference ( #3539 )
...
* test_convnd_fwd
* test_convnd_bwd_data
* test_conv_bwd_data_scale
* test_grouped_convnd_fwd_clamp
* test_grouped_convnd_fwd_scale
* multiple A/B tensors and D tensor for fwd GPU ref
* test_grouped_convnd_fwd_scaleadd_ab
* test_grouped_convnd_fwd_bias_clamp
* test_grouped_convnd_fwd_bilinear
* test_grouped_convnd_fwd_gk_bias_clamp
* Extend GPU reference to enable batchnorm epilogue
* test_grouped_convnd_fwd{,_gk}_bias_bnorm_clamp
* test_grouped_conv_bwd_data_bilinear
* test_grouped_convnd_bwd_weight_bilinear
* Add missing template instantiation
* Perform operations in float in reference
* Slightly increase tolerance for batchnorm profiler
* Revert "Slightly increase tolerance for batchnorm profiler"
This reverts commit a3b24752296da4576060e2f75fa10e
2026-01-27 09:49:42 +01:00
Ville Pietilä
6361810fb5
Fix merge conflict.
2026-01-27 03:08:22 -05:00
