[CK_TILE] Add FAv3 fwd pipeline (#2731)

* Add FAv3 fwd pipeline

* Unpack v_pk_mul to hide v_mov

* Avoid compiler moving l compute across phase

* Sync sched_group_barrier() setting for masking cases
This commit is contained in:
Po Yen Chen
2025-09-01 09:16:45 +08:00
committed by GitHub
parent 705804d9bf
commit d876e87fe4
16 changed files with 3270 additions and 0 deletions

View File

@@ -0,0 +1,31 @@
#!/bin/sh
# TODO: run this script from CK root or build directory
EXE="$(find . -name tile_example_fmha_fwd_v3 -type f | head -n 1)"
VALID=0
for causal in 0 1 ; do
for prec in "fp16" "bf16" ; do
for hdim in 128 ; do
for perm in 0 ; do
if [ $causal -eq 0 ]; then
mask=0
else
mask=b:-1,0
fi
$EXE -prec=$prec -b=32 -h=16 -s=512 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=16 -h=16 -s=1024 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=8 -h=16 -s=2048 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=4 -h=16 -s=4096 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=2 -h=16 -s=8192 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=16 -s=16384 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=64 -s=16384 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=16 -h_k=1 -s=65536 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=40 -s=37200 -d=$hdim -mask=$mask -iperm=$perm -operm=$perm -v=$VALID
done
done
done
done