fix tflite rounding

CMakeLists.txt

@@ -95,6 +95,7 @@ if(MSVC)
   add_definitions(-D_ENABLE_EXTENDED_ALIGNED_STORAGE)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /EHsc")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /MP")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /bigobj")
   set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /bigobj")
   if(USE_MSVC_MT)
     foreach(flag_var

include/tvm/relay/qnn/attrs.h

@@ -32,6 +32,25 @@ namespace tvm {
 namespace relay {
 namespace qnn {
 
+/*! \brief Attribute for qnn add operator */
+struct QnnAddAttrs : public tvm::AttrsNode<QnnAddAttrs> {
+  std::string rounding;
+
+  TVM_DECLARE_ATTRS(QnnAddAttrs, "relay.attrs.QnnAddAttrs") {
+    TVM_ATTR_FIELD(rounding).set_default("UPWARD").describe(
+        "Defines the rounding direction when the value is midway between"
+        "two representable values. There are two 3 modes - UPWARD, TONEAREST"
+        "or TFLITE. UP/TONEAREST modes behave exactly same except at the"
+        "midpoints between the two representable values. At the midpoint,"
+        "UPWARD rounds towards positive infinity (for example -1.5 will be"
+        "rounded to -1). TONEAREST is the standard rounding where the"
+        "value is rounded away from zero at midpoints (for example, -1.5"
+        "rounds to -2). More context can be found at following glibc manual"
+        "https://www.gnu.org/software/libc/manual/html_node/Rounding.html."
+        "TFLITE mode is more complicated, referring to tflite implementation.");
+  }
+};
+
 /*! \brief Attribute for requantize operator */
 struct RequantizeAttrs : public tvm::AttrsNode<RequantizeAttrs> {
   int axis;
@@ -46,14 +65,15 @@ struct RequantizeAttrs : public tvm::AttrsNode<RequantizeAttrs> {
         .set_default(-1);
     TVM_ATTR_FIELD(rounding).set_default("UPWARD").describe(
         "Defines the rounding direction when the value is midway between"
-        "two representable values. There are two supported modes - UPWARD"
-        "or TONEAREST. Both modes behave exactly same except at the"
+        "two representable values. There are two 3 modes - UPWARD, TONEAREST"
+        "or TFLITE. UP/TONEAREST modes behave exactly same except at the"
         "midpoints between the two representable values. At the midpoint,"
         "UPWARD rounds towards positive infinity (for example -1.5 will be"
         "rounded to -1). TONEAREST is the standard rounding where the"
         "value is rounded away from zero at midpoints (for example, -1.5"
-        "rounds to -2). More context can be found at following gblic manual"
-        "https://www.gnu.org/software/libc/manual/html_node/Rounding.html.");
+        "rounds to -2). More context can be found at following glibc manual"
+        "https://www.gnu.org/software/libc/manual/html_node/Rounding.html."
+        "TFLITE mode is more complicated, referring to tflite implementation.");
     TVM_ATTR_FIELD(out_dtype)
         .set_default(NullValue<DataType>())
         .describe("Output data type, set to explicit type under mixed precision setting");

python/tvm/relay/frontend/tflite.py

@@ -45,7 +45,7 @@ def __init__(self, tensor_idx, tensor, buffer, qnn_params=None):
 
 class OperatorConverter(object):
     """Operator Converted for converting TFLite ops to Relay ops"""
-    def __init__(self, model, subgraph, exp_tab):
+    def __init__(self, model, subgraph, exp_tab, rounding):
 
         try:
             from tflite.BuiltinOperator import BuiltinOperator
@@ -60,6 +60,7 @@ def __init__(self, model, subgraph, exp_tab):
         self.builtin_op_code = build_str_map(BuiltinOperator())
         self.activation_fn_type = build_str_map(ActivationFunctionType())
         self.builtin_options = build_str_map(BuiltinOptions())
+        self.rounding = rounding
 
         # Add more operators
         self.convert_map = {
@@ -643,6 +644,9 @@ def _hard_swish(data):
 
         return out
 
+    # TODO in quantized mode, concat op implicitly invokes requantize in cpp
+    # implementation, TFLITE mode rounding needed to be selected in order
+    # to get bit-exact execution.
     def convert_concatenation(self, op):
         """Convert TFLite concatenation"""
         try:
@@ -854,14 +858,26 @@ def _convert_elemwise(self, relay_op, op):
         if lhs_tensor.qnn_params:
             assert rhs_tensor.qnn_params, "Both tensors should be quantized."
             assert output_tensor.qnn_params, "Output tensor should be quantized."
-            out = relay_op(lhs=lhs_expr,
-                           rhs=rhs_expr,
-                           lhs_scale=lhs_tensor.qnn_params['scale'],
-                           lhs_zero_point=lhs_tensor.qnn_params['zero_point'],
-                           rhs_scale=rhs_tensor.qnn_params['scale'],
-                           rhs_zero_point=rhs_tensor.qnn_params['zero_point'],
-                           output_scale=output_tensor.qnn_params['scale'],
-                           output_zero_point=output_tensor.qnn_params['zero_point'])
+            has_tflite_rounding_mode = [_qnn.op.add]
+            if relay_op in has_tflite_rounding_mode:
+                out = relay_op(lhs=lhs_expr,
+                               rhs=rhs_expr,
+                               lhs_scale=lhs_tensor.qnn_params['scale'],
+                               lhs_zero_point=lhs_tensor.qnn_params['zero_point'],
+                               rhs_scale=rhs_tensor.qnn_params['scale'],
+                               rhs_zero_point=rhs_tensor.qnn_params['zero_point'],
+                               output_scale=output_tensor.qnn_params['scale'],
+                               output_zero_point=output_tensor.qnn_params['zero_point'],
+                               rounding=self.rounding)
+            else:
+                out = relay_op(lhs=lhs_expr,
+                               rhs=rhs_expr,
+                               lhs_scale=lhs_tensor.qnn_params['scale'],
+                               lhs_zero_point=lhs_tensor.qnn_params['zero_point'],
+                               rhs_scale=rhs_tensor.qnn_params['scale'],
+                               rhs_zero_point=rhs_tensor.qnn_params['zero_point'],
+                               output_scale=output_tensor.qnn_params['scale'],
+                               output_zero_point=output_tensor.qnn_params['zero_point'])
         else:
             out = relay_op(lhs_expr, rhs_expr)
 
@@ -924,6 +940,9 @@ def convert_sub(self, op):
             return self._convert_elemwise(_qnn.op.subtract, op)
         return self._convert_elemwise(_op.subtract, op)
 
+    # TODO in quantized mode, mul op implicitly invokes requantize in cpp
+    # implementation, TFLITE mode rounding needed to be selected in order
+    # to get bit-exact execution.
     def convert_mul(self, op):
         """Convert TFLite MUL"""
         # Check if the input tensor is quantized, call QNN op
@@ -1327,6 +1346,7 @@ def _convert_reduce(self, relay_op, op):
                                      input_zero_point=input_tensor.qnn_params['zero_point'],
                                      output_scale=output_tensor.qnn_params['scale'],
                                      output_zero_point=output_tensor.qnn_params['zero_point'],
+                                     rounding=self.rounding,
                                      out_dtype=output_tensor_type_str)
 
         return out
@@ -1452,6 +1472,7 @@ def convert_fully_connected(self, op):
                                      input_zero_point=new_input_zero_point,
                                      output_scale=output_tensor.qnn_params['scale'],
                                      output_zero_point=output_tensor.qnn_params['zero_point'],
+                                     rounding=self.rounding,
                                      out_dtype=output_tensor_type_str)
 
             # Call activation function
@@ -1667,6 +1688,7 @@ def convert_conv(self, op, conv_type):
                                      input_zero_point=new_input_zero_point,
                                      output_scale=output_tensor.qnn_params['scale'],
                                      output_zero_point=output_tensor.qnn_params['zero_point'],
+                                     rounding=self.rounding,
                                      out_dtype=output_tensor_type_str)
 
             # Call activation function
@@ -2454,8 +2476,10 @@ def get_scalar_from_constant(expr):
     assert isinstance(expr, _expr.Constant) and not expr.data.shape, \
         "Expr is not a constant scalar."
     value = expr.data.asnumpy()
-    assert value.dtype == np.dtype(np.int32) or value.dtype == np.dtype(np.float32), \
-        "value must be float32/int32"
+    assert value.dtype == np.dtype(np.int32) or \
+           value.dtype == np.dtype(np.float32) or \
+           value.dtype == np.dtype(np.float64), \
+        "value must be float32/float64/int32"
     return np.asscalar(value)
 
 
@@ -2524,7 +2548,7 @@ def get_tensor_name(subgraph, tensor_idx):
     return subgraph.Tensors(tensor_idx).Name().decode("utf-8")
 
 
-def from_tflite(model, shape_dict, dtype_dict):
+def from_tflite(model, shape_dict, dtype_dict, rounding='TFLITE'):
     """Convert from tflite model into compatible relay Function.
 
     Parameters
@@ -2538,6 +2562,9 @@ def from_tflite(model, shape_dict, dtype_dict):
     dtype_dict : dict of str to str
         Input types of the model.
 
+    rounding : str
+        Rounding mode for tflite model
+
     Returns
     -------
     mod : tvm.IRModule
@@ -2576,7 +2603,7 @@ def from_tflite(model, shape_dict, dtype_dict):
         exp_tab.set_expr(model_input_name, _expr.var(model_input_name, shape=shape, dtype=dtype))
 
     # op code in model
-    op_converter = OperatorConverter(model, subgraph, exp_tab)
+    op_converter = OperatorConverter(model, subgraph, exp_tab, rounding)
     op_converter.check_unsupported_ops()
     op_converter.convert_op_to_relay()
 

python/tvm/relay/qnn/op/qnn.py

@@ -300,7 +300,8 @@ def add(lhs,
         rhs_scale,
         rhs_zero_point,
         output_scale,
-        output_zero_point):
+        output_zero_point,
+        rounding="UPWARD"):
     """Quantized addition with numpy-style broadcasting.
 
     Parameters
@@ -329,6 +330,9 @@ def add(lhs,
     output_zero_point: relay.Expr
        The zero point of output quantized expr.
 
+    rounding: str, optional
+       rounding mode of qnn add
+
     Returns
     -------
     result : relay.Expr
@@ -338,7 +342,8 @@ def add(lhs,
     return _make.add(lhs, rhs,
                      lhs_scale, lhs_zero_point,
                      rhs_scale, rhs_zero_point,
-                     output_scale, output_zero_point)
+                     output_scale, output_zero_point,
+                     rounding)
 
 
 def dense(data,

python/tvm/relay/quantize/quantize.py

@@ -172,7 +172,7 @@ def qconfig(**kwargs):
         is None, which means will try to call all operartors' annotate rewrite
         function.
 
-    rounding: "UPWARD" or "TONEAREST"
+    rounding: "UPWARD" or "TONEAREST" or "TFLITE"
         Rounding direction for fixed point multiplications.
 
     Returns

src/relay/qnn/op/add.cc

@@ -30,9 +30,11 @@ namespace tvm {
 namespace relay {
 namespace qnn {
 
+TVM_REGISTER_NODE_TYPE(QnnAddAttrs);
+
 /*
  * \brief Canonicalizes the QNN add op.
- * \param attrs The empty attribute.
+ * \param attrs The QNN add attrs.
  * \param new_args The new mutated args to the call node.
  * \param arg_types The types of input and output.
  * \return The sequence of Relay ops for add op.
@@ -42,9 +44,55 @@ Expr QnnAddCanonicalize(const Attrs& attrs, const Array<Expr>& new_args,
   // Get the args.
   QnnBinaryOpArguments args(new_args);
 
+  // Get the attrs.
+  const QnnAddAttrs* add_attrs = attrs.as<QnnAddAttrs>();
+  CHECK(add_attrs != nullptr);
+  auto& rounding = add_attrs->rounding;
+
   // Get the input dtype and shape.
   QnnBinaryOpTensorType input_type(arg_types, 0);
 
+  if (rounding == "TFLITE") {
+    double lhs_scale_val = GetScalarFromConstant<float>(args.lhs_scale);
+    double rhs_scale_val = GetScalarFromConstant<float>(args.rhs_scale);
+    double out_scale_val = GetScalarFromConstant<float>(args.output_scale);
+    double twice_max_input_scale = 2 * std::max(lhs_scale_val, rhs_scale_val);
+    double real_lhs_scale_val = lhs_scale_val / twice_max_input_scale;
+    double real_rhs_scale_val = rhs_scale_val / twice_max_input_scale;
+    double real_out_scale_val = twice_max_input_scale / ((1 << 20) * out_scale_val);
+
+    auto real_lhs_scale = MakeConstantScalar<double>(DataType::Float(64), real_lhs_scale_val);
+    auto real_rhs_scale = MakeConstantScalar<double>(DataType::Float(64), real_rhs_scale_val);
+    auto real_out_scale = MakeConstantScalar<double>(DataType::Float(64), real_out_scale_val);
+    auto one_scalar = MakeConstantScalar<double>(DataType::Float(64), 1);
+    auto zero_scalar = MakeConstantScalar<int>(DataType::Int(32), 0);
+    auto left_shift_scalar = MakeConstantScalar<int>(DataType::Int(32), 1 << 20);
+
+    Expr adapted_lhs = Cast(args.lhs, DataType::Int(32));
+    if (!IsEqualScalar(args.lhs_zero_point, zero_scalar)) {
+      adapted_lhs = Subtract(adapted_lhs, Cast(args.lhs_zero_point, DataType::Int(32)));
+    }
+    adapted_lhs = Multiply(adapted_lhs, left_shift_scalar);
+
+    Expr adapted_rhs = Cast(args.rhs, DataType::Int(32));
+    if (!IsEqualScalar(args.rhs_zero_point, zero_scalar)) {
+      adapted_rhs = Subtract(adapted_rhs, Cast(args.rhs_zero_point, DataType::Int(32)));
+    }
+    adapted_rhs = Multiply(adapted_rhs, left_shift_scalar);
+
+    auto requantized_lhs = Requantize(adapted_lhs, input_type.shape, real_lhs_scale, zero_scalar,
+                                      one_scalar, zero_scalar, DataType::Int(32), rounding);
+
+    auto requantized_rhs = Requantize(adapted_rhs, input_type.shape, real_rhs_scale, zero_scalar,
+                                      one_scalar, zero_scalar, DataType::Int(32), rounding);
+
+    auto output = Add(requantized_lhs, requantized_rhs);
+    output = Requantize(output, input_type.shape, real_out_scale, zero_scalar, one_scalar,
+                        args.output_zero_point, DataType::Int(32), rounding);
+    // Go back to lower precision.
+    return ConvertDtype(output, input_type.dtype);
+  }
+
   // FIXME (anijain2305) - The lowering can be further optimized. Instead of inserting requantize in
   // the start, we can insert requantize at the end if both input tensors have same qnn params. In
   // that case, we can first add the tensors, subtract the zero point, and requantize at the end.
@@ -86,9 +134,23 @@ Expr QnnAddCanonicalize(const Attrs& attrs, const Array<Expr>& new_args,
   return ConvertDtype(output, input_type.dtype);
 }
 
+Expr MakeQnnAdd(Expr lhs, Expr rhs, Expr lhs_scale, Expr lhs_zero_point, Expr rhs_scale,
+                Expr rhs_zero_point, Expr output_scale, Expr output_zero_point,
+                std::string rounding) {
+  auto attrs = make_object<QnnAddAttrs>();
+  attrs->rounding = std::move(rounding);
+
+  static const Op& op = Op::Get("qnn.add");
+  return Call(op,
+              {lhs, rhs, lhs_scale, lhs_zero_point, rhs_scale, rhs_zero_point, output_scale,
+               output_zero_point},
+              Attrs(attrs), {});
+}
+
 // QNN Addition operator.
-QNN_REGISTER_BINARY_OP("add")
+QNN_REGISTER_BINARY_OP_WITH_BODY("add", MakeQnnAdd)
     .describe("Elementwise add with with broadcasting for quantized tensors.")
+    .set_attrs_type<QnnAddAttrs>()
     .set_support_level(11)
     .set_attr<FTVMLegalize>("FTVMQnnCanonicalize", QnnAddCanonicalize);
 

src/relay/qnn/op/concatenate.cc

@@ -48,7 +48,10 @@ bool QnnConcatenateRel(const Array<Type>& types, int num_inputs, const Attrs& at
                 << PrettyPrint(types[1]));
   }
   for (const auto& input_scale : input_scales_tuple->fields) {
-    CHECK(IsScalarType(input_scale, DataType::Float(32)));  // input_scales[idx]
+    const auto* input_scale_type = input_scale.as<TensorTypeNode>();
+    CHECK(input_scale_type && IsScalarType(input_scale, input_scale_type->dtype) &&
+          (input_scale_type->dtype == DataType::Float(32) ||
+           input_scale_type->dtype == DataType::Float(64)));  // input_scale[idx]
   }
 
   const auto* input_zero_points_tuple = types[2].as<TupleTypeNode>();
@@ -61,8 +64,12 @@ bool QnnConcatenateRel(const Array<Type>& types, int num_inputs, const Attrs& at
     CHECK(IsScalarType(input_zero_point, DataType::Int(32)));  // input_zero_points[idx]
   }
 
-  CHECK(IsScalarType(types[3], DataType::Float(32)));  // output_scale
-  CHECK(IsScalarType(types[4], DataType::Int(32)));    // output_zero_point
+  const auto* output_scale_type = types[3].as<TensorTypeNode>();
+  CHECK(output_scale_type && IsScalarType(types[3], output_scale_type->dtype) &&
+        (output_scale_type->dtype == DataType::Float(32) ||
+         output_scale_type->dtype == DataType::Float(64)));  // output_scale
+
+  CHECK(IsScalarType(types[4], DataType::Int(32)));  // output_zero_point
 
   // Collect the input tensor and output tensor devoid of scale and zero points to reuse Relay
   // Concatenate infer type function.

src/relay/qnn/op/convolution.cc

@@ -57,13 +57,18 @@ bool QnnConv2DRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
   CHECK(param->out_dtype.bits() > 0) << "Output dtype bits should be greater than 0.";
 
   // Check the types of scale and zero points.
-  CHECK(IsScalarType(types[2], DataType::Int(32)));    // input_zero_point
-  CHECK(IsScalarType(types[3], DataType::Int(32)));    // kernel_zero_point
-  CHECK(IsScalarType(types[4], DataType::Float(32)));  // input_scale
+  CHECK(IsScalarType(types[2], DataType::Int(32)));  // input_zero_point
+  CHECK(IsScalarType(types[3], DataType::Int(32)));  // kernel_zero_point
+  const auto* input_scale_type = types[4].as<TensorTypeNode>();
+  CHECK(input_scale_type && IsScalarType(types[4], input_scale_type->dtype) &&
+        (input_scale_type->dtype == DataType::Float(32) ||
+         input_scale_type->dtype == DataType::Float(64)));  // input_scale
+
   // Kernel scale can be a vector of length output_channels or a scalar.
   size_t axis = param->kernel_layout.find('O');
   CHECK(axis != std::string::npos) << "Kernel layout attribute is not defined";
-  AssignType(types[5], DataType::Float(32), weight->shape[axis], reporter);  // kernel scale
+  const auto* kernel_scale_type = types[5].as<TensorTypeNode>();
+  AssignType(types[5], kernel_scale_type->dtype, weight->shape[axis], reporter);  // kernel scale
 
   // Collect the input tensor and output tensor devoid of scale and zero points to reuse Relay
   // Conv2D infer type function.

src/relay/qnn/op/dequantize.cc

@@ -46,8 +46,10 @@ bool DequantizeRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
       << input_dtype;
 
   // Check the types of scale and zero points.
-  CHECK(IsScalarType(types[1], DataType::Float(32)));  // input_scale
-  CHECK(IsScalarType(types[2], DataType::Int(32)));    // input_zero_point
+  const auto* input_scale_type = types[1].as<TensorTypeNode>();
+  CHECK(input_scale_type && (input_scale_type->dtype == DataType::Float(32) ||
+                             input_scale_type->dtype == DataType::Float(64)));
+  CHECK(IsScalarType(types[2], DataType::Int(32)));  // input_zero_point
 
   const Array<tvm::PrimExpr> oshape = data->shape;
   // assign output type, output will always be float 32.
@@ -66,7 +68,8 @@ Expr MakeDequantize(Expr data, Expr input_scale, Expr input_zero_point) {
 Expr DequantizeLower(const Expr& input_tensor, const Expr& input_scale,
                      const Expr& input_zero_point) {
   auto shift = Subtract(Cast(input_tensor, DataType::Int(32)), input_zero_point);
-  auto scaled_output = Multiply(Cast(shift, DataType::Float(32)), input_scale);
+  auto scaled_output =
+      Multiply(Cast(shift, DataType::Float(32)), Cast(input_scale, DataType::Float(32)));
   return scaled_output;
 }