Aling 'linalg-to-xegpu' pass with patched XeGPU dialect #201
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
dchigarev
force-pushed
the
fix-linalg-to-xe
branch
from
616c9e1 to
716af02
Compare
dchigarev
commented
Jul 31, 2024
Comment on lines
745
to
762
| loc, vecLoadType, tile, vnniAxisAttr, transpose, | ||
| loc, vecLoadType, tile, packedAttr, transpose, transpose_bit, |
Contributor
Author
There was a problem hiding this comment.
vnniAxis->packedAttr: instead of a vnni axis (0, 1) specify "packed" attribute that's equivalent ofvnni_axis=0transpose_bit: allows to transpose data while loading. Isn't used by this lowering pass
dchigarev
commented
Jul 31, 2024
|
|
||
| // Load A sub-tiles. | ||
| SmallVector<Value> loadVecA = | ||
| loadNdDescTiles(rewriter, loc, tilesA, readCacheHint, vnniConfA); |
Contributor
Author
There was a problem hiding this comment.
vnniConfA can't be used during loading since vnniAxis=1 is now longer supported. However we still need this config to compute proper tiles for xegpu.dpas later in the code.
dchigarev
force-pushed
the
fix-linalg-to-xe
branch
from
aeada62 to
435b520
Compare
Signed-off-by: dchigarev <dmitry.chigarev@intel.com>
Signed-off-by: dchigarev <dmitry.chigarev@intel.com>
dchigarev
force-pushed
the
fix-linalg-to-xe
branch
from
435b520 to
2778459
Compare
kurapov-peter
approved these changes
Aug 1, 2024
dchigarev
force-pushed
the
fix-linalg-to-xe
branch
from
f78f6d2 to
829b9d4
Compare
dchigarev
commented
Aug 1, 2024
dchigarev
commented
Aug 1, 2024
| // Create output initial value load tiles. | ||
| // CHECK: %[[rootC:.+]] = xegpu.create_nd_tdesc %[[C]] | ||
| // CHECK: %[[tC:.+]] = xegpu.update_nd_offset %[[rootC]], [0, 0] | ||
| // CHECK: %[[tC:.+]] = xegpu.update_nd_offset %[[rootC]], [%c0, %c0] |
Contributor
Author
There was a problem hiding this comment.
imex doesn't support constant offsets (see intel/mlir-extensions#815)
dchigarev
commented
Aug 1, 2024
| // Extract DPAS-sized chunks from larger loaded tile A. | ||
| // Tile B is already in the correct shape. | ||
| // CHECK: %[[vA_flat:.+]] = vector.shape_cast %[[vA]] : vector<32x8x2xf16> to vector<512xf16> | ||
| // CHECK: %[[vA_flat:.+]] = vector.shape_cast %[[vA]] : vector<32x16xf16> to vector<512xf16> |
Contributor
Author
There was a problem hiding this comment.
we do not load the A matrix via vnni_axis=1 anymore (see packed_attr)
|
The IMEX changes are merged in Menooker:dev. |
Signed-off-by: dchigarev <dmitry.chigarev@intel.com>
dchigarev
commented
Aug 2, 2024
| # required functionality is merged. | ||
| gc_fetch_content(imex 496b240093b5e132b60c5ee69878300fe69be300 https://github.com/Menooker/mlir-extensions | ||
| SET IMEX_CHECK_LLVM_VERSION=ON IMEX_ENABLE_L0_RUNTIME=0 | ||
| gc_fetch_content(imex d5bbd635dee500b8cff138686833bacfac5ade78 https://github.com/Menooker/mlir-extensions |
Contributor
Author
There was a problem hiding this comment.
updated to the latest commit in dev branch
Signed-off-by: dchigarev <dmitry.chigarev@intel.com>
Signed-off-by: dchigarev <dmitry.chigarev@intel.com>
Signed-off-by: dchigarev <dmitry.chigarev@intel.com>
dchigarev
commented
Aug 5, 2024
Comment on lines
-132
to
-136
| cl_platform_id platform; // OpenCL platform | ||
| cl_device_id device; // device ID | ||
| CL_SAFE_CALL(clGetPlatformIDs(1, &platform, NULL)); | ||
| CL_SAFE_CALL(clGetDeviceIDs(platform, *devtype, 1, &device, NULL)); | ||
| return device; |
Contributor
Author
There was a problem hiding this comment.
The old logic searched for a device of the requested type only in one platform (and couldn't find any GPU on my machine). Rewritten the logic to iterate over all available platforms and return a first suitable device
5 tasks
Menooker
approved these changes
Aug 7, 2024
Menooker
approved these changes
Aug 7, 2024
LongshengDu
approved these changes
Aug 7, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Closes #192
This PR updates
linalg-to-xegpupass to make it compatible withxegpu-to-vc-funcpass from IMEX.The PR also adds a simple e2e test for
linalg->xegpu->gpu exepipeline.