[LLVMGPU][ROCm] Add MFMA_F32_16x16x4_F32 instruction #17847
Conversation
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
2 times, most recently
from
ddacba7
to
9412a5e
Compare
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
3 times, most recently
from
8227633
to
a34e030
Compare
pashu123
changed the title
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
2 times, most recently
from
73f0863
to
a2ddc34
Compare
pashu123
changed the title
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
from
a2ddc34
to
49e0bea
Compare
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
2 times, most recently
from
73bcdaa
to
4ca5488
Compare
| // not supoorted by amgpu.mfma op. | ||
| Value zeroIdx = builder.create<arith::ConstantIndexOp>(loc, 0); | ||
| lhs = builder.create<vector::ExtractElementOp>(loc, lhs, zeroIdx); | ||
| rhs = builder.create<vector::ExtractElementOp>(loc, rhs, zeroIdx); |
There was a problem hiding this comment.
Use vector.extract with an attribute index instead
There was a problem hiding this comment.
Updated! Any reasons why?
There was a problem hiding this comment.
extract element allows for dynamism over the index and, in this application, requires you to materialize the index attribute as a constant. we don't need either of the two
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
3 times, most recently
from
f9953e0
to
435b878
Compare
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
2 times, most recently
from
8988ba7
to
24f1f84
Compare
| lhs = builder.create<vector::ExtractOp>(loc, lhs, SmallVector<int64_t>{0}); | ||
| rhs = builder.create<vector::ExtractOp>(loc, rhs, SmallVector<int64_t>{0}); |
There was a problem hiding this comment.
I think ArrayRef{int64_t{0}} will also work
pashu123
force-pushed
the
wmma_ab_f32_c_f32
branch
from
24f1f84
to
da1cd74
Compare
|
@pashu123 DCO is failing, you need to sign all your commits and force-push |
|
I think this broke SDXL compilation with Sorry, CI results have been very noisy, particularly this morning. I also filed nod-ai/SHARK-TestSuite#286 for the poor failure mode in that workflow a few days ago (was hoping that would be noticed here during review and resolved, but I should have commented too). |
|
BTW please add PR descriptions in the future. I'm not sure from the logs and code if this was expected to change lowering paths and affect existing models, for example. |
This reverts commit d65c6d4.
|
Oooh, compilation succeeded but the |
Reverts #17847 This broke SDXL rocm pipeline tests on mi300, see #17847 (comment). The tests aren't showing error messages (`root:benchmark_sdxl_rocm.py:31 Command failed with error: b''`) so I can't easily tell what the issue is, nod-ai/SHARK-TestSuite#286 is filed to improve the situation there.
Signed-off-by: Lubo Litchev <lubol@google.com>
…rg#17894) Reverts iree-org#17847 This broke SDXL rocm pipeline tests on mi300, see iree-org#17847 (comment). The tests aren't showing error messages (`root:benchmark_sdxl_rocm.py:31 Command failed with error: b''`) so I can't easily tell what the issue is, nod-ai/SHARK-TestSuite#286 is filed to improve the situation there. Signed-off-by: Lubo Litchev <lubol@google.com>
No description provided.