Minimal viable support for f16 dot product with f16 accumulator on the block pointer path #1211
Conversation
This file contains 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
8 tasks
jopperm
force-pushed
the
jopperm/blockptr_f16accu
branch
2 times, most recently
from
37835e5
to
df9b126
Compare
jopperm
commented
May 31, 2024
|
https://github.com/triton-lang/triton/blob/v2.1.0/python/tutorials/08-experimental-block-pointer.py#L161,#L176 |
jopperm
commented
May 31, 2024
Comment on lines
+227
to
+231
| if res_dtype in [torch.float16]: | ||
| # We observed high relative errors on small numbers when only using 16 bit for accumulation; | ||
| # hence, use a more restricted input set here. | ||
| a = torch.randint(low=-8, high=8, size=(512, 512), device='xpu', dtype=dtype) / 16 | ||
| b = torch.randint(low=-8, high=8, size=(512, 512), device='xpu', dtype=dtype) / 16 |
There was a problem hiding this comment.
As an alternative to raising the tolerances, we could use restricted inputs (here: [-0.5, 0.5] in 1/16th increments). WDYT?
Signed-off-by: Julian Oppermann <julian.oppermann@codeplay.com>
jopperm
force-pushed
the
jopperm/blockptr_f16accu
branch
from
55a34ee
to
01da0fc
Compare
| if res_dtype in [torch.float16]: | ||
| # We observed high relative errors on small numbers when only using 16 bit for accumulation; | ||
| # hence, use a more restricted input set here. | ||
| a = torch.randint(low=-8, high=8, size=(512, 512), device='xpu', dtype=dtype) / 16 |
There was a problem hiding this comment.
Why are we generating random integer values for float16?
There was a problem hiding this comment.
While the results overall look reasonable to me, there are outliers which would require the relative tolerance to be >> 10, so I looked for alternatives to restrict the inputs. I'll bring this up later in the call.
8 tasks
|
Superseded by #1258. |
whitneywhtsang
pushed a commit
that referenced
this pull request
Jun 20, 2024
This PR adds support for using the `bf16 += bf16 x bf16` variant of the DPAS instruction in the `10-experimental-block-pointer.py` tutorial. I extended the `triton_gen.dpas` op's verifier to support non-f32 accumulator types, and added corresponding testcases for bf16 (as well as for f16 accumulation which I missed in #1211). I had to raise the absolute tolerance again (compared to f16 accumulation). `1e-2` for `bf16` and `1e-3` for `f16` matches what `torch.finfo` returns as "resolution" for the data types, though that might be just a coincidence. --------- Signed-off-by: Julian Oppermann <julian.oppermann@codeplay.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Adds support for accumulating the dot result with 16-bit precision.