QNN quantize and dequantize operators.

include/tvm/relay/qnn/attrs.h

@@ -65,6 +65,38 @@ struct RequantizeAttrs : public tvm::AttrsNode<RequantizeAttrs> {
   }
 };
 
+/*! \brief Attribute for quantize operator */
+struct QuantizeAttrs : public tvm::AttrsNode<QuantizeAttrs> {
+  int32_t output_zero_point;
+  double output_scale;
+  DataType out_dtype;
+
+  TVM_DECLARE_ATTRS(QuantizeAttrs, "relay.attrs.QuantizeAttrs") {
+    TVM_ATTR_FIELD(out_dtype)
+      .describe("Output data type, can be one of [int8 or uint8].");
+
+    TVM_ATTR_FIELD(output_zero_point)
+      .describe("The zero_point for the activation of this op.");
+
+    TVM_ATTR_FIELD(output_scale)
+      .describe("The scale for the activation of this op.");
+  }
+};
+
+/*! \brief Attribute for dequantize operator */
+struct DequantizeAttrs : public tvm::AttrsNode<DequantizeAttrs> {
+  int32_t input_zero_point;
+  double input_scale;
+
+  TVM_DECLARE_ATTRS(QuantizeAttrs, "relay.attrs.QuantizeAttrs") {
+    TVM_ATTR_FIELD(input_zero_point)
+      .describe("The zero_point for the input tensor of this op.");
+
+    TVM_ATTR_FIELD(input_scale)
+      .describe("The scale for the input tensor of this op.");
+  }
+};
+
 }  // namespace qnn
 }  // namespace relay
 }  // namespace tvm

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

@@ -72,3 +72,50 @@ def requantize(data,
                             output_zero_point,
                             rounding,
                             out_dtype)
+
+
+def quantize(input_data, output_zero_point, output_scale, out_dtype='int8'):
+    r""" Quantize op
+     This operator takes float32 as input and produces quantized int8 or unit8 as output.
+     The input tensor can be of any shape. The output shape is the same as input shape.
+     ..math::
+            \mbox{out}[x] =
+                \mbox{clamp(round(input_tensor/output_scale) + output_zero_point);
+                 out_dtype::min, out_dtype::max}
+     Parameters
+    ----------
+    input_data : tvm.relay.Expr
+        The input tensor to be quantized. Can be of type float32.
+    output_zero_point :
+        The output zero_point.
+    output_scale:
+        The output scale.
+    input_dtype:
+        The data type of the input tensor. Can be [int8, uint8]
+    Returns
+    -------
+    result : tvm.relay.Expr
+        The computed result.
+    """
+    return _make.quantize(input_data, output_zero_point, output_scale, out_dtype)
+
+
+def dequantize(input_data, input_zero_point, input_scale):
+    r""" Dequantize op
+     This operator takes quantized int8 and unit8 as input and produces
+    dequantized float32 as output. The output shape is the same as input shape. The input
+    tensor can be of any shape.
+     Parameters
+    ----------
+    input_data : tvm.relay.Expr
+        The input tensor to be dequantized. Can be of type [int8, uint8].
+    input_zero_point :
+        The output zero_point.
+    input_scale:
+        The output scale.
+    Returns
+    -------
+    result : tvm.relay.Expr
+        The computed result.
+    """
+    return _make.dequantize(input_data, input_zero_point, input_scale)

src/relay/qnn/op/dequantize.cc

@@ -0,0 +1,105 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2019 by Contributors
+ * \file src/relay/qnn/op/dequantize.cc
+ * \brief QNN dequantize operator. Dequantize operator converts from quantized
+ * domain to unquantized domain.
+ */
+
+#include <tvm/relay/analysis.h>
+//#include <tvm/relay/op.h>
+#include <tvm/relay/op_attr_types.h>
+#include <tvm/relay/qnn/attrs.h>
+#include "../../pass/pattern_util.h"
+#include "../util.h"
+
+namespace tvm {
+namespace relay {
+namespace qnn {
+
+TVM_REGISTER_NODE_TYPE(DequantizeAttrs);
+
+bool DequantizeRel(const Array<Type>& types,
+                   int num_inputs,
+                   const Attrs& attrs,
+                   const TypeReporter& reporter) {
+  CHECK_EQ(types.size(), 2);
+  const auto* data = types[0].as<TensorTypeNode>();
+  const auto input_dtype = data->dtype;
+  CHECK(input_dtype == Int(8) || input_dtype == UInt(8))
+    << "Input type should be one of the quantized types [unit8, int8] but was " <<  input_dtype;
+  const Array<tvm::Expr> oshape = data->shape;
+  // assign output type, output will always be float 32.
+  reporter->Assign(types[1], TensorTypeNode::make(oshape, Float(32)));
+  return true;
+}
+
+Expr MakeDequantize(Expr data,
+                    int32_t input_zero_point,
+                    double input_scale) {
+  auto attrs = make_node<DequantizeAttrs>();
+  attrs->input_scale = input_scale;
+  attrs->input_zero_point = input_zero_point;
+  static const Op& op = Op::Get("qnn.dequantize");
+  return CallNode::make(op, {data}, Attrs(attrs), {});
+}
+
+Expr DequantizeLower(const Expr& input_tensor, const DequantizeAttrs* param) {
+  const auto input_zero_point = MakeConstantScalar(Int(32), param->input_zero_point);
+  const auto input_scale = MakeConstantScalar(Float(32), param->input_scale);
+  auto shift = Subtract(Cast(input_tensor, Int(32)), input_zero_point);
+  auto scaled_output = Multiply(Cast(shift, Float(32)), input_scale);
+  return scaled_output;
+}
+
+Expr DequantizeLegalize(const Attrs& attrs, const Array<Expr>& new_args,
+                          const Array<tvm::relay::Type>& arg_types) {
+  CHECK_EQ(new_args.size(), 1);
+  auto& data = new_args[0];
+  const auto* param = attrs.as<DequantizeAttrs>();
+  CHECK(param != nullptr);
+
+  CHECK_EQ(arg_types.size(), 1);
+  auto input_dtype = arg_types[0];
+  auto input_tensor_type = input_dtype.as<TensorTypeNode>();
+  CHECK(input_tensor_type != nullptr) << "Type information missing."
+                                      << " Please run infer_type pass.";
+  return DequantizeLower(data, param);
+}
+
+RELAY_REGISTER_OP("qnn.dequantize")
+  .describe(R"code(Dequantizes the input and produces float32 output.
+The input is always quantized (int8, uint8) and will be converted to float32 given input scale and zero_point.
+- **data**: Quantized tensor of any shape to dequantize. The input data can be of floating point
+)code" TVM_ADD_FILELINE)
+.set_attrs_type_key("relay.attrs.DequantizeAttrs")
+.set_num_inputs(1)
+.add_argument("data", "Tensor", "The tensor to dequantize.")
+.set_support_level(11)
+.add_type_rel("Dequantize", DequantizeRel)
+.set_attr<FTVMLegalize>("FTVMLegalize", DequantizeLegalize);
+
+TVM_REGISTER_API("relay.qnn.op._make.dequantize")
+.set_body_typed(MakeDequantize);
+
+}  // namespace qnn
+}  // namespace relay
+}  // namespace tvm

src/relay/qnn/op/quantize_op.cc

@@ -0,0 +1,123 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2019 by Contributors
+ * \file src/relay/qnn/op/quantize_op.cc
+ * \brief QNN dequantize operator. Dequantize operator converts from quantized
+ * domain to unquantized domain.
+ */
+
+#include <tvm/relay/analysis.h>
+//#include <tvm/relay/op.h>
+#include <tvm/relay/op_attr_types.h>
+#include <tvm/relay/qnn/attrs.h>
+#include "../../pass/pattern_util.h"
+#include "../util.h"
+
+namespace tvm {
+namespace relay {
+namespace qnn {
+
+TVM_REGISTER_NODE_TYPE(QuantizeAttrs);
+
+bool QuantizeRel(const Array<Type>& types,
+                 int num_inputs,
+                 const Attrs& attrs,
+                 const TypeReporter& reporter) {
+  CHECK_EQ(types.size(), 2);
+  const auto* data = types[0].as<TensorTypeNode>();
+  const auto input_dtype = data->dtype;
+  CHECK(input_dtype == Float(32))
+    << "Input type should be one of float32 but was " <<  input_dtype;
+  const auto* param = attrs.as<QuantizeAttrs>();
+  const Array<tvm::Expr> oshape = data->shape;
+  const DataType out_dtype = param->out_dtype;
+  CHECK(out_dtype == Int(8) || out_dtype == UInt(8))
+    << "Output type should be one of [int8, unit8 ] but was " << out_dtype;
+  // assign output type
+  reporter->Assign(types[1], TensorTypeNode::make(oshape, out_dtype));
+  return true;
+}
+
+Expr MakeQuantize(Expr data,
+                  int32_t output_zero_point,
+                  double output_scale,
+                  DataType out_dtype) {
+  auto attrs = make_node<QuantizeAttrs>();
+  attrs->output_scale = output_scale;
+  attrs->output_zero_point = output_zero_point;
+  attrs->out_dtype = std::move(out_dtype);
+  static const Op& op = Op::Get("qnn.quantize");
+  return CallNode::make(op, {data}, Attrs(attrs), {});
+}
+
+Expr QuantizeLower(const Expr& input_tensor, const QuantizeAttrs* param) {
+  const auto out_dtype = param->out_dtype;
+  const auto output_zero_point = MakeConstantScalar(Int(32), param->output_zero_point);
+  const auto scale = MakeConstantScalar(Float(32), param->output_scale);
+  const int32_t min_val = GetQmin(out_dtype);
+  const int32_t max_val = GetQmax(out_dtype);
+  auto scale_data = Cast(Round(Divide(input_tensor, scale)), Int(32));
+  // we are trying to do - std::min(std::max(unclamped, min_val), max_val);
+  auto add_zero_point = Add(scale_data, output_zero_point);
+  auto clamped_output = Clip(add_zero_point, min_val, max_val);
+  auto clamp_out_dtype = Cast(clamped_output, out_dtype);
+  return clamp_out_dtype;
+}
+
+Expr QuantizeLegalize(const Attrs& attrs, const Array<Expr>& new_args,
+                        const Array<tvm::relay::Type>& arg_types) {
+  CHECK_EQ(new_args.size(), 1);
+  auto& data = new_args[0];
+  const auto* param = attrs.as<QuantizeAttrs>();
+  CHECK(param != nullptr);
+
+  CHECK_EQ(arg_types.size(), 1);
+  auto input_dtype = arg_types[0];
+  auto input_tensor_type = input_dtype.as<TensorTypeNode>();
+  CHECK(input_tensor_type != nullptr) << "Type information missing."
+                                      << " Please run infer_type pass.";
+  return QuantizeLower(data, param);
+}
+
+RELAY_REGISTER_OP("qnn.quantize")
+.describe(R"code(Quantizes the input and produces quantized output.
+The input can be either float or quantized(int8, unit8). If the input is float,
+this op takes scale and zero point and quantize the float value to
+quantized output, in int8 or uint8 format. If the input is quantized value,
+the op requantize the input (of a certain type, with a given scale and zero
+point) to the output of the same or different type with a same or different
+scale and zero point.
+- **data**: Tensor of any shape to quantize. The input data can be of floating point
+          or quantized.
+)code" TVM_ADD_FILELINE)
+.set_attrs_type_key("relay.attrs.QuantizeAttrs")
+.set_num_inputs(1)
+.add_argument("data", "Tensor", "The tensor to quantize.")
+.set_support_level(11)
+.add_type_rel("Quantize", QuantizeRel)
+.set_attr<FTVMLegalize>("FTVMLegalize", QuantizeLegalize);
+
+TVM_REGISTER_API("relay.qnn.op._make.quantize")
+.set_body_typed(MakeQuantize);
+
+}  // namespace qnn
+}  // namespace relay
+}  // namespace tvm

src/relay/qnn/op/requantize.cc

@@ -19,7 +19,7 @@
 
 /*!
  *  Copyright (c) 2019 by Contributors
- * \file requantize.cc
+ * \file src/relay/qnn/op/requantize.cc
  * \brief QNN requantize operator.
  */
 
@@ -228,14 +228,14 @@ bool RequantizeRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
   const auto* data = types[0].as<TensorTypeNode>();
   const auto in_dtype = data->dtype;
   CHECK(in_dtype == Int(8) || in_dtype == UInt(8) || in_dtype == Int(32))
-      << "Input type should be an integer but was " << in_dtype;
+      << "Input type should be one of [int8, uint8, int32] but was " << in_dtype;
 
   const Array<tvm::Expr> oshape = data->shape;
   // assign output type
   const RequantizeAttrs* param = attrs.as<RequantizeAttrs>();
   auto out_dtype = param->out_dtype;
   CHECK(out_dtype == Int(8) || out_dtype == UInt(8) || out_dtype == Int(32))
-      << "Output type should be an integer but was " << out_dtype;
+      << "Output type should be one of [int8, uint8, int32] integer but was " << out_dtype;
   reporter->Assign(types[1], TensorTypeNode::make(oshape, out_dtype));
   return true;
 }