def test_redundant_kernels():
dtype = "float32"
A = te.placeholder(shape=(1, ), name="A", dtype=dtype)
B = te.placeholder(shape=(1, ), name="B", dtype=dtype)
C = te.placeholder(shape=(1, ), name="C", dtype=dtype)
D = topi.less(A, C)
E = topi.less(B, C)
F = topi.logical_or(D, E)
G = topi.identity(F)
for target in ["opencl", "cuda"]:
if not tvm.testing.device_enabled(target):
continue
print("Running on target: %s" % target)
valid = [None]
with tvm.target.Target(target):
s = tvm.topi.testing.get_reduce_schedule(target)(G)
with tvm.transform.PassContext(config={
"tir.add_lower_pass": [(2,
get_verify_pass(valid, max_kernels=1))]
}):
tvm.build(s, [A, B, C, G], target)
assert valid[0]
def simulated_quantize_compute(attrs, inputs, out_type):
"""Compiler for simulated_quantize."""
assert len(inputs) == 4
assert attrs.sign
assert attrs.rounding == "round"
data, scale, clip_min, clip_max = inputs
if attrs.kind == QAnnotateKind.IDENTITY:
return [topi.identity(data)]
# simulate rounding error
scaled_data = topi.divide(data, scale)
clipped_data = topi.maximum(topi.minimum(scaled_data, clip_max), clip_min)
round_data = topi.round(clipped_data)
# recover data
rdata = topi.multiply(round_data, scale)
return [rdata]