[FQ2I] Add abs to FQ2I (apache#10922)

* add abs to fq2i

* lint

* special case for fq2i

* np iinfo

python/tvm/relay/transform/fake_quantization_to_integer.py

@@ -109,6 +109,20 @@ def identity(expr, type_map):
 register_unary_identity("image.resize2d")
 
 
+@register_fake_quantization_to_integer("abs")
+def abs_(expr, type_map):
+    """Rewrite an abs op"""
+    assert len(expr.args) == 1
+    arg = expr.args[0]
+    t = type_map[arg]
+
+    min_value = relay.const(np.iinfo(t.dtype).min, t.dtype)
+    one = relay.const(1, t.dtype)
+    out = relay.op.where(relay.op.equal(min_value, arg), arg + one, arg)
+    out = relay.op.abs(out)
+    return [out, t]
+
+
 @register_fake_quantization_to_integer("nn.adaptive_avg_pool1d")
 def adaptive_avgpool1d(expr, type_map):
     """Rewrite an adaptive avgpool op"""

tests/python/relay/test_pass_fake_quantization_to_integer.py

@@ -374,6 +374,19 @@ def test_fake_quantize_image_resize_bilinear():
     compare_fq_to_int(op, [x_np], allow_rounding_error=True)
 
 
+def test_fake_quantize_abs():
+    x = relay.var("x", shape=[1, 3, 224, 224], dtype="int8")
+
+    zero = relay.const(0)
+    x = relay.qnn.op.dequantize(x, relay.const(2.0), zero)
+    op = relay.op.abs(x)
+    op = relay.qnn.op.quantize(op, relay.const(2.0), zero)
+
+    x_np = np.random.randint(-128, 127, size=[1, 3, 224, 224], dtype="int8")
+
+    compare_fq_to_int(op, [x_np])
+
+
 def test_fake_quantize_expand_dims():
     x = relay.var("x", shape=[1, 3, 224, 224], dtype="int8")