SMEM-P: fix scoping - define tTMEM_LOADcS_frg unconditionally
This commit is contained in:
@@ -316,9 +316,9 @@ class FmhaKernel:
|
||||
frg_cnt = 4
|
||||
frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt
|
||||
tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile))
|
||||
# Coordinate fragments for SMEM-P mapping
|
||||
# Coordinate fragments for SMEM-P mapping (needed unconditionally for scoping)
|
||||
tTMEM_LOADcS_frg = cute.logical_divide(tTMEM_LOADcS, cute.make_layout(frg_tile))
|
||||
if self.use_smem_p:
|
||||
tTMEM_LOADcS_frg = cute.logical_divide(tTMEM_LOADcS, cute.make_layout(frg_tile))
|
||||
print(f"[SMEM-P CUTLASS] Created tTMEM_LOADcS_frg shape: {cute.shape(tTMEM_LOADcS_frg)}")
|
||||
|
||||
for j in range(frg_cnt):
|
||||
|
||||
Reference in New Issue
Block a user