Skip to content

[Fix] Set identity block scales for CDNA4 MFMA_Scale in fp8 row-scale… #744

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
amd-songpiao wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[Fix] Set identity block scales for CDNA4 MFMA_Scale in fp8 row-scale… #744

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions kernels/fp8_gemm_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -207,6 +207,12 @@ def wait_barrier(count):
class Mfma16x16x128:
def __init__(self, n_tiles_a, n_tiles_b):
self.atom = fx.make_mma_atom(fx.rocdl.cdna4.MFMA_Scale(16, 16, 128, fx.Float8E4M3FN))
# The CDNA4 scaled-MFMA atom defaults its E8M0 block scales to 0
# (== 2**-127 ~= 0), which annihilates every product. This kernel does
# the row/col dequant in the epilogue (StoreC), so the MMA itself wants
# the identity scale: each of the 4 per-32-block E8M0 bytes set to 127
# (2**0 == 1.0), i.e. 0x7F7F7F7F for both A and B.
self.atom = self.atom.set_value({"scale_a": fx.Int32(0x7F7F7F7F), "scale_b": fx.Int32(0x7F7F7F7F)})
self.accum_type = Vec.make_type(4, fx.Float32)
self.zero_value = Vec.filled(4, 0.0, fx.Float32)
self.n_tiles_a = n_tiles_a
Expand Down
Loading