[ETHOSN] Remove requantize dependency on resize (#14422)

A following requantize operation can instead be offloaded as a
standalone operation, leaving to the support library to optimize.

Change-Id: I573d8d67bcdf54f23d4c73ea70dbbe3f0c9be9cb

Author: lhutton1

python/tvm/relay/op/contrib/ethosn.py

@@ -206,11 +206,7 @@ def qnn_requantize_pattern():
         return pattern
 
     def qnn_resize_pattern():
-        pattern = is_op("image.resize2d")(wildcard()).has_attr({"method": "nearest_neighbor"})
-        pattern = is_op("qnn.requantize")(
-            pattern, is_constant(), is_constant(), is_constant(), is_constant()
-        )
-        return pattern
+        return is_op("image.resize2d")(wildcard()).has_attr({"method": "nearest_neighbor"})
 
     def qnn_mul_pattern():
         """

src/relay/backend/contrib/ethosn/ethosn_api.cc

@@ -838,40 +838,27 @@ EthosnError EthosnAPI::ReinterpretQuantize(const Expr& expr,
 }
 
 EthosnError EthosnAPI::Resize(const Expr& expr, ResizeParams* params) {
-  Call requantize = Downcast<Call>(expr);
-  Call resize = Downcast<Call>(requantize->args[0]);
+  Call resize = Downcast<Call>(expr);
+  const auto* input_ttype = resize->args[0]->checked_type().as<TensorTypeNode>();
+
+  const auto* attrs = resize->attrs.as<Resize2DAttrs>();
+  uint32_t height, width;
+  EthosnError err = Tvm2Npu(attrs->size, &height, &width);
+  params->resize_info = sl::ResizeInfo{sl::ResizeAlgorithm::NEAREST_NEIGHBOUR, height, width,
+                                       params->input_info.m_QuantizationInfo};
 
-  const auto* input_dtype = resize->args[0]->checked_type().as<TensorTypeNode>();
   sl::TensorShape input_tensor_shape = {1, 1, 1, 1};
-  EthosnError err = Tvm2Npu(input_dtype->shape, &input_tensor_shape);
   sl::DataType input_tensor_dtype;
-  err += Tvm2Npu(input_dtype->dtype, &input_tensor_dtype);
-  float input_sc;
-  int input_zp;
-  err += AsConstant(requantize->args[2], &input_zp);
-  err += AsConstant(requantize->args[1], &input_sc);
-  sl::QuantizationInfo input_q_info;
-  err += Tvm2Npu(input_zp, input_sc, &input_q_info);
+  err = Tvm2Npu(input_ttype->shape, &input_tensor_shape);
+  err += Tvm2Npu(input_ttype->dtype, &input_tensor_dtype);
   params->input_info =
-      sl::TensorInfo(input_tensor_shape, input_tensor_dtype, sl::DataFormat::NHWC, input_q_info);
+      sl::TensorInfo(input_tensor_shape, input_tensor_dtype, params->input_info.m_DataFormat,
+                     params->input_info.m_QuantizationInfo);
 
-  float output_sc;
-  int output_zp;
-  err += AsConstant(requantize->args[3], &output_sc);
-  err += AsConstant(requantize->args[4], &output_zp);
-  sl::QuantizationInfo resize_q_info;
-  err += Tvm2Npu(output_zp, output_sc, &resize_q_info);
-  const auto* attrs = resize->attrs.as<Resize2DAttrs>();
-  uint32_t height, width;
-  err += Tvm2Npu(attrs->size, &height, &width);
-  params->resize_info =
-      sl::ResizeInfo{sl::ResizeAlgorithm::NEAREST_NEIGHBOUR, height, width, resize_q_info};
-
-  sl::TensorInfo output_info = params->input_info;
-  output_info.m_Dimensions[1] = params->resize_info.m_NewHeight;
-  output_info.m_Dimensions[2] = params->resize_info.m_NewWidth;
-  output_info.m_QuantizationInfo = params->resize_info.m_OutputQuantizationInfo;
-  params->output_info = output_info;
+  sl::TensorInfo output_tensor_info;
+  err += Tvm2Npu(resize->checked_type(), &output_tensor_info);
+  output_tensor_info.m_QuantizationInfo = params->input_info.m_QuantizationInfo;
+  params->output_info = output_tensor_info;
 
   return err;
 }

tests/python/contrib/test_ethosn/test_resize.py

@@ -29,31 +29,18 @@ def _get_model(
     shape,
     dtype,
     size,
-    input_zp,
-    input_sc,
-    output_zp,
-    output_sc,
     coordinate_transformation_mode,
     rounding_method,
 ):
     x = relay.var("x", shape=shape, dtype=dtype)
-    resize = relay.image.resize2d(
+    return relay.image.resize2d(
         data=x,
         size=size,
         layout="NHWC",
         method="nearest_neighbor",
         coordinate_transformation_mode=coordinate_transformation_mode,
         rounding_method=rounding_method,
     )
-    model = relay.qnn.op.requantize(
-        resize,
-        input_scale=relay.const(input_sc, "float32"),
-        input_zero_point=relay.const(input_zp, "int32"),
-        output_scale=relay.const(output_sc, "float32"),
-        output_zero_point=relay.const(output_zp, "int32"),
-        out_dtype=dtype,
-    )
-    return model
 
 
 @requires_ethosn
@@ -82,10 +69,6 @@ def test_resize(dtype, shape, size, coordinate_transformation_mode, rounding_met
             shape=shape,
             dtype=dtype,
             size=size,
-            input_zp=zp_min + 128,
-            input_sc=0.0784314,
-            output_zp=zp_min + 128,
-            output_sc=0.0784314,
             coordinate_transformation_mode=coordinate_transformation_mode,
             rounding_method=rounding_method,
         )
@@ -113,16 +96,11 @@ def test_resize(dtype, shape, size, coordinate_transformation_mode, rounding_met
 def test_resize_failure(size, err_msg):
     """Check Resize error messages."""
     dtype = "int8"
-    zp_min = np.iinfo(dtype).min
 
     model = _get_model(
         shape=(1, 10, 10, 1),
         dtype=dtype,
         size=size,
-        input_zp=zp_min + 128,
-        input_sc=0.0784314,
-        output_zp=zp_min + 128,
-        output_sc=0.0784314,
         coordinate_transformation_mode="half_pixel",
         rounding_method="round_prefer_ceil",
     )