[CK] Fix gptoss sink
## Motivation
This PR removes conditional logic for handling infinity values in the
sink mechanism across multiple FMHA pipeline implementations, defaulting
sink_size to 0 and adding a constraint in the kernel selection logic.
## Technical Details
Changes:
Removed __builtin_isinf_sign(sink_v) checks and conditional
initialization of LSE accumulators across 7 pipeline files
Added default initialization (= 0) for sink_size in 4 argument structs
Added F_sink == "f" constraint to kernel compatibility checking
## Test Plan
Local test
## Test Result
passed
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] CK Tile improvements and fixes for depthwise merged
convolutions forward (#4873)
## Motivation
Performance benchmarks showed that old CK's depthwise merged
convolutions are much faster than CK Tile's ones.
## Technical Details
After investigation it showed up that the requirement that A/CVectorload
is a multiple of gemm's rightmost dimension is too strict in case of
processing multiple groups, because if tensor is in NHWGC/NHWGK format,
then if C/K is equal to 1, we can use vectorloads on the G dimension,
which is added by this PR. Filter5x5 specialization was also added,
because some models are using it, it's similar to 3x3, the only
difference is the window size. This addition was needed, because of the
differences of tensor descriptor transformations betweeen CK and CK
Tile. In old CK the case of grouped depthwise 5x5 convs was supported
via Default specialization, but in CK Tile that case was not working
properly.
## Test Plan
Performance was tested by our internal test suite, which contains
several DL models.
## Test Result
Tests results showed significant performance uplift for depthwise(3x3,
5x5) cases
[CK][CK Tile] Fix batched gemm kernel 2 lds
## Motivation
Fix 2 lds batched gemm universal gemm call. Disable split k for not
valid atomic add instruction size.
## Technical Details
Fix 2 lds batched gemm universal gemm call. Disable split k for not
valid atomic add instruction size.
## Test Plan
CI overall
## Test Result
pending
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Add split-K support for ABQuantGrouped in
block_scale_gemm (#4816)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Changes
### Split-K support in `gemm_quant_kernel.hpp`
- **`SplitKBatchOffset`**: Added `aq_group_offset` and
`aq_k_split_offset` fields (mirroring the existing `bq_*` fields for B)
to track each split-K batch's position within the AQ scale tensor. For
`ABQuantGrouped`, both offsets are computed from `k_id * KRead` divided
by `AQuantGroupSize::kK`.
- **`MakeAQBlockWindow`**: Added an `aq_group_offset` parameter
(defaulting to 0 for non-split-K paths) so the AQ tensor view's K-group
dimension reflects only the remaining K-groups from the split-K offset,
consistent with how `MakeBQBlockWindow` handles the BQ tensor.
- **`RunGemm`**: Threads the `aq_k_split_offset` through to
`MakeAQBlockWindow` when in split-K mode.
### Constraints in `IsSupportedArgument()`
Four constraints gate split-K (`k_batch > 1`) for ABQuantGrouped:
1. **Mode check** — split-K is only allowed for `BQuantGrouped` (no
preshuffle) or `ABQuantGrouped` (no `APreshuffleQuant`). Any other quant
mode with `k_batch > 1` returns `false`.
2. **B quant group alignment** — `KRead` (per-batch K slice) must be
divisible by `BQuantGroupSize::kK`. Each batch must operate on complete
B quantization groups; a partial group would require splitting a scale
value across batches.
3. **A quant group alignment** (new, ABQuantGrouped only) — `KRead` must
also be divisible by `AQuantGroupSize::kK` for the same reason applied
to the AQ scale tensor.
4. **Minimum 2 K-tile iterations per batch** (new) — The
software-pipelined GEMM kernels (CompV3 family) prefetch one tile ahead,
so they require `per_batch_num_loop = KRead / KPerBlock >= 2`. When
`KRead == KPerBlock` (i.e. each batch is exactly one tile), the prefetch
reads into the next batch's memory region and produces incorrect
results. Configurations where `K == k_batch * KPerBlock` are therefore
rejected.
### Example update (`run_gemm_quant_example.inc`)
Updated the comment above the `IsSupportedArgument` call to document
that split-K is now supported for both `BQuantGrouped` (no preshuffle)
and `ABQuantGrouped` (no `APreshuffleQuant`).
## Unit Tests
Two new test files covering decode and prefill tile shapes across a
range of `k_batch` values (2–8), data types (FP8, BF8), and quantization
group sizes (1×1×128 and 1×128×128 for B):
- `test_gemm_quant_abquant_splitk_decode.cpp` — uses the decode tile
shape (M=16, N=64, K_tile=256)
- `test_gemm_quant_abquant_splitk_prefill.cpp` — uses the prefill tile
shape (M=128, N=128, K_tile=128)
Each test calls `run_test_with_validation` which runs the kernel and
checks correctness against a CPU reference. Configurations excluded from
tests are annotated with comments explaining which constraint they
violate (typically the `per_batch_num_loop >= 2` requirement).
## Prerequisites
This PR depends on #4429, which must be merged before this can be
merged.
Tile Engine support for gfx950
## Motivation
This PR adds support for the gfx950 GPU architecture to the Tile Engine
in Composable Kernel library, focusing on GEMM operations with FP8 and
BF8 data types.
## Technical Details
Added gfx950-specific MFMA warp GEMM implementations with conditional
compilation.
Updated default GEMM configuration parameters for tile sizes and warp
configurations.
Added Jenkins CI pipeline stage for testing TILE_ENGINE_GEMM on gfx950
hardware.
## Test Plan
Tile engine itself is a benchmarking utility, so if it passes the CI it
will be tested automatically.
## Test Result
Tile engine itself is a benchmarking utility, so if it passes the CI it
will be tested automatically.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Fix windows build issues
## Motivation
Full build on Windows is currently broken due to compiler errors, this
PR should help fix that. This is also holding up the following PR in the
TheRock: https://github.com/ROCm/TheRock/pull/3382
## Technical Details
1. I don't see a good reason to be nesting a windows include inside the
ck_tile namespace. It was causing compiler errors too: Windows.h comes
with min and max, which was conflicting with ck_tile::min and
ck_tile::max, so I moved it out. I also defined NOMINMAX to prevent this
inclusion in the future.
2. The TRUE/FALSE macros are already used by Windows.h, which causes an
error. So I've opted for True/False. You can see this pattern in other
rocm-libraries.
3. The M_PI macro isn't available, at least in the WIN32_LEAN_AND_MEAN
context, from \<cmath\> on Windows. We'll be able to use
std::numbers::v_pi\<float\> when we have C++20 support.
4. There was a missing \<chrono\> include.
## Test Plan
Test locally and make sure this doesn't impact existing CI.
## Test Result
Compiles locally and passes existing ci.
## Submission Checklist
- [ x ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK][CK TILE] Improve oob check
## Motivation
Improve OOB checks. Remove permutes which have been generated by thread
buffer zero clear. at now in assembly there is only condmask instead of
permute + condmask.
Change number of KPack for generated instances
## Technical Details
Remove permute instructions from assembly
## Test Plan
test_grouped_convnd_fwd_tile
## Test Result
passed
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Refactor `UniversalGemm::MakeA/B/C/DBlockViews` to
allow caller to pass desciptors directly (#4295)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Proposed changes
Currently `UniversalGemmKernel::MakeA/B/C/DBlockViews` directly create
tensor views from strides and sizes. This refactors the descriptor
creation out and add overloaded definitions, allowing descriptors to be
created separately by the caller instead of passing explicit strides,
with no functional changes.
This will enable further refactoring of `RunGemm` to do likewise,
enabling derived kernels like BatchedContractionKernel to avoid creating
separate versions (PR
[#3457](https://github.com/ROCm/composable_kernel/pull/3457)).
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [x] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [x] I have run `clang-format` on all changed files
- [ ] Any dependent changes have been merged
## Discussion
Since the logic within the MakeXBlockviews chains together operations on
tuples, and thus the descriptors are also passed as such, adding a
template parameter for the type of the input tuple was the simplest
option to enable the overload without too much verbiage. However, for
`MakeCBlockView` this adds a complications as the templated definitions
are prone to overlap. This for now is avoided by just moving the
arguments around for the descriptor version, which avoids the collision.
It's not a great solution, so feel free to suggest a better one.
[CK_TILE] Extend support of mix precision microscaling BQuant
(#4267)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Proposed changes
Supported types combinations using BQuant=e8m0:
- A=bf16
- B=bf16,bf8,fp4
Summary:
- remove usage of `pk_fp4_raw_t`: consistent with other implementations
and avoid taking into account of the packed size explicitly. In general,
the raw type should not be used because CK Tile internally takes care of
the PackedSize, so using the raw type adds unnecessary complexity to the
implementation
- handle microscaling by checking for `e8m0` type for BQuant (previous
implementation was inconsistent)
- add support for scaling instructions in `DequantPack8`
- mx pipeline:
- extend existing pipeline to support different B types
- add support to scale and cast before writing to LDS or after reading
from LDS (this can be defined in the `Problem` by the user)
- block gemm:
- mx pipeline is now using block gemm BQuant
- block gemm BQuant can now load from LDS and apply scale and then call
block gemm universal operator. This adds new functionalities and remove
code duplication
- warp gemm:
- add case to support 128bit ds_read/write for both A and B when A=16bit
and B=8bit
- add examples and tests: note that some tests for bf16/fp4 already
existed but were removed during previous tests refactoring. I added them
again and other relevant tests for new types combinations
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [ ] I have run `clang-format` on all changed files
- [ ] Any dependent changes have been merged
## Discussion
If this is a relatively large or complex change, feel free to start a
discussion by explaining why you chose the solution you did and what
alternatives you considered
[CK TILE] Refactor sequence_reverse_inclusive_scan
## Proposed changes
Refactor ck tile `sequence_reverse_inclusive_scan` from recursive to
for-loop.
Tracking issue: #4229
This pull request introduces a new lightweight array type,
`static_array`, and refactors the sequence utilities to use it for
improved constexpr support and simplicity. The changes also include
updates to the build system to add container-related tests.
**Core Library Improvements:**
* Added a new header `static_array.hpp` that defines the `static_array`
type, a constexpr-friendly array with basic accessors and no custom
constructors.
* Updated includes in `core.hpp` and `sequence.hpp` to import
`static_array`.
[[1]](diffhunk://#diff-14b406eccf59794051a16c0c9c1a7e11234324bfdd107a5bbe0f173cd25bcddcR44)
[[2]](diffhunk://#diff-5042e5b47bb2ba78bbab2d284338cf0503bc8fb76a7d631cc2684ad6ca832a76R7)
**Refactoring to Use `static_array`:**
* Refactored sequence utilities in `sequence.hpp` to use `static_array`
instead of the previously forward-declared `array` type, including in
histogram and array generation logic.
[[1]](diffhunk://#diff-5042e5b47bb2ba78bbab2d284338cf0503bc8fb76a7d631cc2684ad6ca832a76L1108-R1133)
[[2]](diffhunk://#diff-5042e5b47bb2ba78bbab2d284338cf0503bc8fb76a7d631cc2684ad6ca832a76L1130-R1146)
* Rewrote the implementation of `sequence_reverse_inclusive_scan` to use
`static_array` for intermediate storage, improving constexpr evaluation
and clarity.
**Build System and Testing:**
* Added a new test subdirectory for container tests and a GoogleTest
executable for `unit_sequence.cpp` to the CMake build configuration.
[[1]](diffhunk://#diff-5d35ff7555d3f0b438d45cde06b661eb1332cdbec66287ac7ec3c478d688aae5R5)
[[2]](diffhunk://#diff-1f54f0d2b431b7fc74f7b4ffb66e80c381c904c3383b1d27987467e3482d6d7aR1-R7)
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
[CK_TILE] Update Stream-K Reduction Strategy Enum
## Motivation
Currently, Stream-K has 3 reduction options: 1) atomics, 2) The
reduction described in the Stream-K paper, and 3) a tree reduction. The
reduction strategy described in the original Stream-K paper has the
starting workgroup of each tile sequentially accumulating partial
results of other contributing workgroups in the tile, which requires a
linear number of steps. Hence, for clarity, this works updates the
naming of the `StreamKReductionStrategy` enum members to better describe
the existing reduction strategy options.
## Technical Details
Prior to this change, the enum is as follows:
```cpp
enum StreamKReductionStrategy : uint32_t
{
Atomic = 0u,
Reduction = 1u,
TreeReduction = 2u
};
```
But, the distinction between `Reduction` and `TreeReduction` is not very
clear and has some redundancy.
Hence, the updated enum is as follows:
```cpp
enum StreamKReductionStrategy : uint32_t
{
Atomic = 0u,
Linear = 1u,
Tree = 2u
};
```
All references to `StreamKReductionStrategy` were updated to reflect
this change.
## Test Plan
No new functionality was added, so no new tests were added; I just
validated existing tests and examples.
## Test Result
All tests passed locally.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Fix FP8 MXGEMM numerical error in async load path
(#4704)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Summary
Fixes FP8 MXGEMM producing half the expected result (e.g., 128 instead
of 256 with all 1s input).
**Bug introduced in:** `b7de1e14cea70681a23cd1a136df42910c776e4a` -
"[CK_TILE] Add blockscale GEMM support for EightWarps on gfx950 (#4280)"
## Root Cause
In the `static_move_ys=true` code path in `tile_window.hpp`, the IMM
optimization computes `lds_ys_offset` using a default-constructed tensor
descriptor:
```cpp
make_tensor_coordinate(decltype(tensor_descriptor){}, idx_ys_offset)
```
This default-constructed descriptor has different strides than the actual DRAM tensor descriptor used for dram_ys_offset. When these offsets are mixed in the address calculation:
```cpp
imm_valid = lds_ys_offset % IMM_RANGE; // From wrong descriptor
wave_offset = dram_ys_offset - imm_valid; // From correct descriptor
```
The final address wave_offset + imm_valid ≠ dram_ys_offset, causing incorrect memory accesses.
Fix
```cpp
Set imm_valid = 0 to bypass the IMM optimization and ensure the full
offset is passed through wave_offset:
constexpr auto imm_valid = 0; // Avoids inconsistency between
lds_ys_offset and dram_ys_offset
```
This disables the 12-bit immediate field optimization in the buffer_load_lds instruction but guarantees correctness. A proper fix would require making the DRAM tensor descriptor constexpr, which is not feasible since tensor strides depend on runtime parameters (LDA, LDB).
[CK_TILE][FMHA] Support gfx11
## Motivation
Add support of gfx11 architectures (RDNA3) to FMHA.
## Technical Details
Distributions (matrix elements to lane registers mapping) of gfx11 WMMA
are completely different from distributions of gfx9 MFMA and gfx12 WMMA.
There are two cases in FMHA where this difference matters:
* usage of results (matrix C) of one GEMM as input (matrix A) of another
GEMM.
* random number generation for dropout (implementation for gfx9 MFMA,
gfx12 WMMA and host validation produce the same results).
Both cases are solved by a special remapping implemented using
`__builtin_amdgcn_permlanex16` and `__builtin_amdgcn_perm`.
Additional changes:
* FMHA tests are now build and run only for those types for which
instances exist (gfx11 supports only fp16 and bf16).
* Two fixes for uninitialized values (`mask.sink` and
`do_fp8_static_quant`): they may contain garbage resulting in incorrect
dispatching logic, sometimes tests report that there are no instance
available for current parameters.
* Small fix to remove expcnt(0) from s_waitcnt instruction on gfx11 when
they are not requested (i.e. every time), likely has no effect on
performance but makes disassembly a bit clearer.
## Test Plan
```
ninja test_ck_tile_fmha
bin/test_ck_tile_fmha_fwd_fp16
bin/test_ck_tile_fmha_fwd_bf16
bin/test_ck_tile_fmha_bwd_fp16
bin/test_ck_tile_fmha_bwd_bf16
```
## Test Result
All tests must pass (some tests may be skipped).
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
fix: correct ULP calculation in get_absolute_threshold for
BF16 tolerance (#4556)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
BF16 grouped GEMM tests were failing on gfx1201 with errors like:
```
Error: Incorrect results! out[5457621] != ref[5457621]: -66 != -65.5
max err: 0.5, number of errors: 1
```
The calculated absolute tolerance (atol ~0.26) was too small to account
for legitimate hardware vs software BF16 conversion differences (0.5
ULP).
## Changes
1. **Discrete exponent calculation**: Changed from continuous `log2()`
to `floor(log2())` to match actual IEEE 754 floating-point exponent
levels
2. **Full ULP for output_error**: Changed from 0.5 to 1.0 ULP to account
for hardware `__bf16` vs software `float_to_bf16()` conversion
differences
## Calculation Example
For the failing case with value ~66:
**Before (incorrect):**
```
expo = log2(66) = 6.044...
atol = 2^(6.044 - 7) * 0.5 = 2^(-0.956) * 0.5 ≈ 0.26
Error 0.5 > 0.26 → Test fails ❌
```
**After (correct):**
```
discrete_expo = floor(log2(66)) = 6
atol = 2^(6 - 7) * 1.0 = 2^(-1) * 1.0 = 0.5
Error 0.5 ≤ 0.5 → Test passes ✓
```
The ULP for values in [64, 128) is 2^(-1) = 0.5, and the error of 0.5 is
exactly 1 ULP, which is the maximum expected difference between hardware
and software BF16 conversions at tie cases.
## Rationale
Hardware and software BF16 conversions can differ by up to 1 ULP at tie
cases due to different rounding strategies (hardware vs IEEE 754
round-to-nearest-even). The discrete exponent ensures ULP is calculated
correctly for all values within an exponent range.
**Modified file**:
`projects/composablekernel/include/ck_tile/host/check_err.hpp`
Fix the Composable Kernel CI and versions incompatibility
(#4640)
## Motivation
This PR has 4 patches:
1. Fix the CI error of grouped gemm.
2. Fix the incompatibility of old linux version.
3. Fix the potential errors of flatmm.
4. Address the previous comments of abquant eight warps pipeline
solution.
[CK][CK TILE] Add has hot loop check for pipeline v1
## Motivation
Add has hot loop check for pipeline v1 (v1 basic and v1 basic async).
Enable more tests which have been fixed by this change.
## Technical Details
Hot loop has been executed without num loop check.
## Test Plan
test_grouped_convnd_fwd_tile
## Test Result
Passed
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-651
AICK-663
[CK TILE] fix numerical errors of preshuffle_b
This pull request introduces several improvements and fixes related to
quantized grouped GEMM (General Matrix Multiply) pipelines and their
supporting utilities.
# The numerical issue
## Steps to reproduce
```bash
Run
./bin/tile_example_gemm_weight_preshuffle -prec=fp8
./bin/tile_example_gemm_weight_preshuffle -prec=int4
```
# Solution
The main changes address type correctness, improve data layout and
shuffling logic, and expand test coverage to better validate different
GEMM configurations.
**Key changes include:**
### Data layout and shuffling logic
* Refactored the logic in `shuffle_b_permuteN` to use `constexpr`
variables for `KLane` and `ItemsPerAccess`, simplifying tile view
construction and correcting the permutation order for improved
efficiency and correctness (`tensor_shuffle_utils.hpp`).
* Fixed the calculation of `KLaneBytes` in weight preshuffle pipeline
policies to account for internal data type conversion (e.g., from
`pk_int4_t` to `fp8`), ensuring accurate memory access and alignment in
quantized GEMM policies (`wp_pipeline_agmem_bgmem_creg_base_policy.hpp`,
`gemm_wp_abquant_pipeline_ag_bg_cr_base_policy.hpp`).
[[1]](diffhunk://#diff-93f16cd76e6e24404777e682a5ac8e039913ddd6a438c7efd61fdda42276e4efL274-R275)
[[2]](diffhunk://#diff-9c3d0fc3c014feed435bfd93ba1f8f9fb3e054dcc322deada3addf70bee5a58cL100-R105)
### Test infrastructure enhancements
* Unit tests did not catch this issue since there were no tests for fp8.
Added new configuration structs (`config_mn_16x16`, `config_mn_32x32`)
to support additional GEMM tile shapes and updated tests to run with
these configurations for broader coverage
(`test_gemm_pipeline_util.hpp`).
[[1]](diffhunk://#diff-5a5962b2c4aa7f6a87d1d6201ad383135e30df13b42654e997d870d57420d5b8R86-R103)
[[2]](diffhunk://#diff-5a5962b2c4aa7f6a87d1d6201ad383135e30df13b42654e997d870d57420d5b8L255-R269)
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Add padding to cshuffle epilogue to avoid bank conflict
(#4274)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Proposed changes
Added the padding to CShuffle Epilogue to avoid the bank conflicts of
64. Synced up and learned from the internal repo.
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [ ] I have run `clang-format` on all changed files
- [ ] Any dependent changes have been merged
## Discussion
If this is a relatively large or complex change, feel free to start a
discussion by explaining why you chose the solution you did and what
alternatives you considered
[CK_TILE]: PreshuffleB + PreshuffleBQuant for ABQuant
pipeline (#4268)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Proposed changes
Implement BQuantPreshuffle option for the ABQuant PreshuffleB pipeline.
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [X] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [X] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [X] I have added inline documentation which enables the maintainers
with understanding the motivation
- [X] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [X] I have run `clang-format` on all changed files
- [X] Any dependent changes have been merged
[CK_TILE] Blockscale Gemm Fix Multi-Arch Compilation
## Motivation
This PR updates CK_TILE blockscale GEMM-quant kernels and launch helpers
to compile across multiple GPU architectures by introducing compile-time
availability gating and a new attribute tag mechanism for kernel
symbol/attribute specialization.
## Technical Details
- Add an architecture-guarded `kIsAvailable` flag to the gfx950 pipeline
and propagate availability handling into `QuantGemmKernel`.
- Extend `make_kernel`/`kentry` to accept an `Attr` tag enabling
per-kernel compile-time attributes (e.g., `no-packed-fp32-ops`) and
unique symbols.
- Update the blockscale GEMM quant example to pass kernel attributes and
adjust gfx950 gating.
## Test Plan
- CI
- Local test: `cmake .. --preset dev -DGPU_TARGETS='gfx942;gfx950'
-GNinja && ninja tile_example_gemm_quant`
- Local test with ROCm/aiter#1954
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] CK Tile grouped convolution direct load
## Motivation
CK Tile grouped convolution forward direct load support.
## Technical Details
Basic pipeline for direct load and new instances for forward for v1 and
v4 pipelines.
## Test Plan
test_grouped_convnd_fwd_tile
## Test Result
CI pending
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-130
[CK_TILE] Add blockscale GEMM support for EightWarps on
gfx950 (#4280)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Proposed changes
gemm blockscale eightwarps support
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [x] I have run `clang-format` on all changed files
- [x] Any dependent changes have been merged
## Discussion
If this is a relatively large or complex change, feel free to start a
discussion by explaining why you chose the solution you did and what
alternatives you considered
[CK_TILE] Add support and tests for V6 pipeline in conv fwd
(#4357)
Added support for conv v6 pipeline in ck tile's convolution forward
kernel. CK Tile v6 pipeline is the equivalent to old ck's V5 pipeline
and should be faster than other pipelines for some cases. This PR also
adds tests inside profiler that's currently inside experimental
directory, so now we should be able to detect regressions easier.
[CK_TILE] Fix MMA concepts compiler error
## Motivation
CK Tile is required to support certain older OSs; on these OSs, cpp 20
is not fully supported. For ROCm 7.2, compiler errors occur on one of
these older OSs. An example of this error is as follows:
```bash
/composable_kernel/include/ck_tile/core/arch/mma/amdgcn_mma.hpp:34:28: error: expected concept name with optional arguments
34 | { MmaOp::kAMBlock } -> std::convertible_to<unsigned int>;
|
```
The goal of this PR is to resolve these compiler errors.
## Technical Details
The existing guards around the mma concepts only check if the concepts
language feature is supported, as follows:
```cpp
#if defined(__cpp_concepts) && __cpp_concepts >= 201907L
// ...
template <typename CtrlFlags>
concept CtrlFlagsGfx9I = requires(CtrlFlags ctrlFlags) {
// Flag members for Gfx9 MFMA instructions
{ CtrlFlags::Cbsz } -> std::convertible_to<int>;
{ CtrlFlags::Abid } -> std::convertible_to<int>;
{ CtrlFlags::Blgp } -> std::convertible_to<int>;
};
#endif // defined(__cpp_concepts) && __cpp_concepts >= 201907L
```
That said, in cases where functionality from the `<concepts>` header is
used (e.g., `std::convertible_to`), this guard fails to check whether
the `<concepts>` header is available.
This change adds an additional check to the concepts that make use of
functionality from the `<concepts>` header to ensure the header is
available.
## Test Plan
I tested the changes on the relevant docker for gfx90a, gfx950, and
gfx942 and the compiler issue is not present.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Add FP8 KV_BLOCKSCALE support for batch prefill
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Implement per-page K/V quantization for paged attention:
- Add KV_BLOCKSCALE enum to BlockAttentionQuantScaleEnum
- Use exp2 shift trick to eliminate explicit P scaling overhead
- Prefetch physical pages offset for KV cache, overlaps with
computations
## Proposed changes
Please describe the motivation behind the pull request, whether it
enables a new feature or fixes a bug. If there are associated pull
requests or issues, please link them to the pull request.
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [ ] I have run `clang-format` on all changed files
- [ ] Any dependent changes have been merged
## Discussion
If this is a relatively large or complex change, feel free to start a
discussion by explaining why you chose the solution you did and what
alternatives you considered
* WIP: add splitk to bquant
* feat: add support for bf8i4 and fp8i4 by calculating correct stride for packed data types
* chore: remove temporary test script
* fix: incorrect tile window length for splitted bq tensor window
* chore: improve comments
* test: add unit tests to cover bquant splitk functionality
* fix: conflict resolution by renaming variables
* [Compiler] Addressing new compiler warnings
Clang enables new lifetime warnings in production and we see build
errors due to this with the staging compiler.
The attributes added in this PR are suggested by the compiler. However,
I'm not very familiar with the code base, so the changes may be
incorrect.
* Update some more instances
* Adds file-level ignores via clang diagnostic pragma
The number of instances was large, so I decided to use file-level scope
to disable the warning via pragma clang diagnostic ignored.
It also showed this warning coming from the gtest dependency. For that,
I did add the respective command line flag to the CMake variables. I
don't know if this is acceptable or not.
* This adds the remaining instances
For a build on gfx90a.
* fix clang format
* Adding couple more instances from gfx1200 build
* Fixed another few instances
---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
* Change call to the intrinsics
* fix clang format
* Undo changes under include/ck/utility
* Use named variable as vector size
---------
Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
* chore: split block scale example instances in more separate files to speed up compile times
* wip: fp4 scaffolding for abquant
* feat: add fp4 decoding-while-loading to abquant pipeline
* feat: add support for fp4 CPU verification in abquant
* chore: add time tracking to reference calculation
* feat: add a4w4 test for blockscale gemm
* feat: optimize reference calculation by preconverting values to AccType
* feat: add fp4 to fp8 look-up table
* fix: reference to wrong ComputeDataType field in QuantProblem
* feat: type utilities for determining MFMA compute types
* feat: packed fp4 for abquant weight preshuffle
* feat: add separate tests for a4w4 base case, padding and preshuffleB
* fix: fp4 conversion on gfx950 attempting to use non-supported method
* fix: test case was using quant group sizes which don't work on gfx950 due to larger mfma tile size
* chore: add fp4 preshuffleb mode to block scale example
* chore: sanity check for packed types being 1 byte
* chore: clarify tensor dimension indices with constants
* chore: replace traits check with specialized check for packed types
* style: some minor refactoring and cleanup
* fix: correct conversion table for FNUZ fp8
* chore: add fp4 instances to main abquant instances again
* chore: use same initialization branch for int4 and fp4
* chore: add missing initialization for fp4 in block scale gemm example
---------
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
* initial commit
* preshuffleQuant support for ABQuant
* fix mxfp4 to use correct QuantGroupSize
* addressing review comments and seperated Preshufflequant for A and B
* updated grouped gemm example for updated traits definition
* fix for CI failure
* updated grouped_gemm_abquant test for updated traits definition
* updated grouped_gemm_abquant test for updated traits definition
- Add multi-dimensional page index support (YsGatherDims) in tile_scatter_gather
- Add is_gather_dim() and get_gather_index() for multi-dim page lookup
- Override MakeVDramTileDistribution() for VECTORIZED_LAYOUT to match
GEMM's BWarpDstrEncoding (K decomposition: {K2, K0, K1})
- Add GetGemmKDecomposition() to retrieve kABKLane and kKPerThread
- Add static_assert for RowMajor VLayout requirement in batch prefill
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* WIP: host level interwave pipeline compiles
* WIP: interwave implementation computes correct GEMM result when no aquant
* WIP: quantization works for subset of problem shapes
* WIP: quantization works for subset of problem shapes
* WIP: interwave memory pipeline passes local test
* feat: Add interwave pipeline implementation for memory pipline in aquant
* test: add unit test for aquant memory pipeline
* WIP: host level interwave pipeline compiles
* WIP: interwave implementation computes correct GEMM result when no aquant
* WIP: quantization works for subset of problem shapes
* WIP: quantization works for subset of problem shapes
* WIP: interwave memory pipeline passes local test
* feat: Add interwave pipeline implementation for memory pipline in aquant
* fix: compilation error on gfx950
* chore: remove debug statements from the code
* test: resolve merge conflict
* test: remove non rcr unit tests from test suite
* Fix alignment issue in Stream-K workspace buffer
In CK Tile Stream-K, the workspace buffer is used to hold flags and
partials, where the first i bytes holds the flags and the remaining
bytes hold partials. This change adds padding to the flags prefix of the
workspace buffer to ensure the number of bytes is 128B-aligned. Without
this alignment, since workgroups do not skip cache when reading from
partials, they may read stale partials data in cache, leading to
incorrect results. The added padding avoids the stale data reading.
This change also re-enables the test_ck_tile_streamk_reduction tests.
* Compute reference GEMM on GPU for test verification to decrease testing time
* add block scale parameters to kernel
* add block scale to kernel
* add smoke test
* format
* Revert "format"
This reverts commit 356c3c9706.
* only format my code
* format py
* fix auto not allowd in function prototype
* change instance tttt to ttff
* fix structured binding issue
* change s_acc elementwise op
* async pipeline add block scale
* add quantation P using shift exp2
* precompute (m - shift) once per row
* change blk scale seqstrt ptr name
* fix some name
* fix for deduction guide
* fix some comments
* add P scale to qr_ksvs_pipeline
* add comment to idx_identity
* change the method of calculating descale block index
* unify naming style: use block_scale_ as name prefix
* unify naming style
* update the CHANGELOG.md
* Add FP8 block scale quantization support for FMHA forward kernel
---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Add signal-based synchronization for persistent GEMM kernels where
input data becomes available incrementally. Uses modulo wraparound
(like PyTorch's AsyncMM) for chunk index calculation:
chunk_idx = ((tile_idx + tile_idx_pivot) / tiles_per_chunk) % num_chunks
Key components:
- PersistentAsyncInputScheduler struct with tiles_per_chunk_m,
chunk_signals, tile_idx_pivot_m, and num_chunks fields
- wait_eq_wave method using __builtin_amdgcn_s_sleep for power efficiency
- IsSupportedArgument validation for scheduler parameters
- Example demonstrating async input scheduling with simulated producer
- GTest unit tests covering all layout combinations
