[microNPU][ETHOSU] Fix rounding mode in requantize operation (apache#…

…15929)

Before the fix, TFL rounding was used in the requantize operation and this led to a discrepancy in the results in some cases

python/tvm/relay/backend/contrib/ethosu/legalize.py

@@ -1216,6 +1216,7 @@ def callback(
             ifm_zero_point=int(params.ifm.q_params.zero_point),
             ofm_scale=float(params.ofm.q_params.scale_f32),
             ofm_zero_point=int(params.ofm.q_params.zero_point),
+            rounding_mode="NATURAL",
         )
 
 

python/tvm/relay/backend/contrib/ethosu/op/identity.py

@@ -36,8 +36,16 @@ def create_ethosu_identity_compute(attrs, args, out_type):
     ofm_scale = attrs.ofm_scale
     ofm_zero_point = attrs.ofm_zero_point
     activation = attrs.activation
+    rounding_mode = attrs.rounding_mode
     op = identity_compute(
-        ifm, lut, ifm_scale, ifm_zero_point, ofm_scale, ofm_zero_point, activation
+        ifm,
+        lut,
+        ifm_scale,
+        ifm_zero_point,
+        ofm_scale,
+        ofm_zero_point,
+        activation,
+        rounding_mode,
     )
     return [op]
 
@@ -61,6 +69,7 @@ def ethosu_identity(
     ofm_scale: float = 1,
     ofm_zero_point: int = 0,
     activation: str = "NONE",
+    rounding_mode: str = "TFL",
 ) -> tvm.relay.Call:
     """The Identity operator that runs on the NPU.
 
@@ -87,12 +96,17 @@ def ethosu_identity(
             "TANH" - tanh activation function.
             "SIGMOID" - sigmoid activation function.
             "LUT" - use a look-up table to perform the activation function.
+    rounding_mode : str, optional
+        The rounding mode to apply to the Output Feature Map tensor.
+            "TFL" - Tensorflow Lite rounding scheme.
+            "TRUNCATE" - Truncate towards zero.
+            "NATURAL" - Round to nearest value, with x.5 rounded up towards +infinity.
 
     Returns
     -------
     out : tvm.relay.Call
         A call to the ethosu_identity op.
     """
     return _make.ethosu_identity(
-        ifm, lut, ifm_scale, ifm_zero_point, ofm_scale, ofm_zero_point, activation
+        ifm, lut, ifm_scale, ifm_zero_point, ofm_scale, ofm_zero_point, activation, rounding_mode
     )

python/tvm/relay/backend/contrib/ethosu/te/identity.py

@@ -31,6 +31,7 @@ def identity_compute(
     ofm_scale: float,
     ofm_zero_point: int,
     activation: str,
+    rounding_mode: str,
 ) -> te.Tensor:
     """A compute operator for the NPU identity operator.
 
@@ -54,14 +55,19 @@ def identity_compute(
             "TANH" - tanh activation function.
             "SIGMOID" - sigmoid activation function.
             "LUT" - use a look-up table to perform the activation function.
+    rounding_mode : str
+        The rounding mode to apply to the Output Feature Map tensor.
+            "TFL" - Tensorflow Lite rounding scheme.
+            "TRUNCATE" - Truncate towards zero.
+            "NATURAL" - Round to nearest value, with x.5 rounded up towards +infinity.
 
     Returns
     -------
     te.Tensor
         The Output Feature Map tensor.
     """
     dmaed_ifm = read_compute(ifm, ifm_zero_point, ifm_scale)
-    id_attrs = {"op": "ethosu_identity", "activation": activation}
+    id_attrs = {"op": "ethosu_identity", "activation": activation, "rounding_mode": rounding_mode}
 
     has_lut = activation in ("TANH", "LUT", "SIGMOID")
 

python/tvm/relay/backend/contrib/ethosu/tir/identity.py

@@ -166,7 +166,7 @@ def get_identity_params(
             padding=SerialPadding(0, 0, 0, 0),
             activation=serial_activation,
             upscale="NONE",
-            rounding_mode="TFL",
+            rounding_mode=attrs["rounding_mode"],
             block_config=SerialBlockConfig(0, 0, 0),
         ),
         output_pointer,

src/relay/op/contrib/ethosu/identity.cc

@@ -63,13 +63,14 @@ bool EthosuIdentityRel(const Array<Type>& types, int num_inputs, const Attrs& at
 }
 
 Expr MakeEthosuIdentity(Expr ifm, Expr lut, double ifm_scale, int ifm_zero_point, double ofm_scale,
-                        int ofm_zero_point, String activation) {
+                        int ofm_zero_point, String activation, String rounding_mode) {
   auto attrs = make_object<EthosuIdentityAttrs>();
   attrs->ifm_scale = ifm_scale;
   attrs->ifm_zero_point = ifm_zero_point;
   attrs->ofm_scale = ofm_scale;
   attrs->ofm_zero_point = ofm_zero_point;
   attrs->activation = std::move(activation);
+  attrs->rounding_mode = std::move(rounding_mode);
   static const Op& op = Op::Get("contrib.ethosu.identity");
   return Call(op, {ifm, lut}, Attrs(attrs), {});
 }

src/relay/op/contrib/ethosu/op_attrs.h

@@ -319,6 +319,7 @@ struct EthosuIdentityAttrs : public tvm::AttrsNode<EthosuIdentityAttrs> {
   double ofm_scale;
   int ofm_zero_point;
   String activation;
+  String rounding_mode;
 
   TVM_DECLARE_ATTRS(EthosuIdentityAttrs, "relay.attrs.EthosuIdentityAttrs") {
     TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
@@ -335,6 +336,13 @@ struct EthosuIdentityAttrs : public tvm::AttrsNode<EthosuIdentityAttrs> {
             "'SIGMOID' - sigmoid activation function. "
             "'LUT' - use a look-up table to perform the activation function.")
         .set_default("NONE");
+    TVM_ATTR_FIELD(rounding_mode)
+        .describe(
+            "The rounding mode to apply to the Output Feature Map tensor. "
+            "'TFL' - Tensorflow Lite rounding scheme. "
+            "'TRUNCATE' - Truncate towards zero."
+            "'NATURAL' - Round to nearest value, with x.5 rounded up towards +infinity.")
+        .set_default("TFL");
   }
 };
 

tests/python/contrib/test_ethosu/cascader/test_ethosu_identity_matcher.py

@@ -39,6 +39,7 @@ def test_ethosu_identity_matcher():
         ofm_scale=1,
         ofm_zero_point=0,
         activation="NONE",
+        rounding_mode="TFL",
     )
 
     length = len(ifm.shape)

tests/python/contrib/test_ethosu/test_codegen.py

@@ -1233,6 +1233,7 @@ def split_func(x):
     [
         [(1, 8, 8, 3), 1.0, 0, 1.0, 0],
         [(1, 20, 30, 3), 1.345, 34, 0.32, -23],
+        [(1, 1, 4, 8), 0.0078125, 0, 0.00997, -30],
     ],
 )
 def test_ethosu_requantize(accel_type, ifm_shape, ifm_scale, ifm_zp, ofm_scale, ofm_zp):