This commit is contained in:
2025-03-21 17:46:45 +08:00
parent dc1469a188
commit af94addb3a
2 changed files with 7 additions and 3 deletions
+1 -1
View File
@@ -133,7 +133,7 @@ if __name__ == "__main__":
out_strides = k.sts[0].real_strides()
if len(out_strides) == 3:
if full_shape[1] < 128:
if full_shape[2] <= 32: k.apply_opt(Opt(OptOps.UNROLL, 0, 0))
if full_shape[2] <= 16: k.apply_opt(Opt(OptOps.UNROLL, 0, 0))
else: k.apply_opt(Opt(OptOps.UNROLL, 0, 8))
k.apply_opt(Opt(OptOps.UPCAST, 1, full_shape[1]))
if out_strides[0] < 128:
+6 -2
View File
@@ -20,8 +20,12 @@ def multi_mul(a0, a1, b0, b1, c0, c1, d0, d1, acc=None):
swizzle.append(64+i)
swizzle.append(96+i)
swizzle = tuple(swizzle)
if a0.op is not Ops.CAST: return None
if a1.op is not Ops.CAST: return None
if a0.op is not Ops.CAST:
#print("rejected on a0")
return None
if a1.op is not Ops.CAST:
#print("rejected on a1")
return None
assert a0.op is Ops.CAST
assert b0.op is Ops.CAST
assert c0.op is Ops.CAST