25af574ff3
The logistic quantization error is introduced by accidently fused the requantize op to the logistic op. To fix the issue, an interface needs to be defined for this op, so this fixed output range property can be queried by the pass. In the followup cls, this interface will be used to replace the FixedResultScale trait. PiperOrigin-RevId: 314221651 Change-Id: Iece591195ca0146b93b5c1a1b9f65c0d205eed11
111 lines
4.4 KiB
TableGen
111 lines
4.4 KiB
TableGen
/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
|
|
|
|
Licensed 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.
|
|
==============================================================================*/
|
|
|
|
// This is the quantization definition file for TensorFlow.
|
|
|
|
#ifdef TF_Quantization
|
|
#else
|
|
#define TF_Quantization
|
|
|
|
include "mlir/IR/OpBase.td"
|
|
include "mlir/Dialect/Quant/QuantOpsBase.td"
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// QuantizedType definitions.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// The base class of a quantized type.
|
|
class QuantizedType<string n, list<int> params, bit signed>
|
|
: Type<And<[CPred<"$_self.isa<mlir::quant::QuantizedType>()">,
|
|
CPred<"$_self.cast<mlir::quant::QuantizedType>()" #
|
|
".getStorageTypeIntegralWidth() == " # !head(params)>]>,
|
|
"Q" # !if (signed, "I", "UI") # !head(params) # " type"> {
|
|
string name = n;
|
|
string asTraitArgsStr =
|
|
StrJoinInt<params>.result # !if(signed, ", true", ", false");
|
|
}
|
|
|
|
// Uniform quantized types. Two integers "smantissa" and "sexp" are used to
|
|
// express the Mantissa and Exponent components of the floating-point scale so
|
|
// the scale of the quantized type is "smantissa * 10 ^ sexp".
|
|
class UInt8UniformQuantizedType<int zero_pt, int smantissa, int sexp>
|
|
: QuantizedType<"Uniform",
|
|
[8, zero_pt, smantissa, sexp, 0, 255], 0>;
|
|
class Int8UniformQuantizedType<int zero_pt, int smantissa, int sexp>
|
|
: QuantizedType<"Uniform",
|
|
[8, zero_pt, smantissa, sexp, -128, 127], 1>;
|
|
|
|
// General uniform quantized types. The definitions can be used to specify
|
|
// operand's tensor types.
|
|
def QUI8 : QuantizedType<"Uniform", [8], 0>;
|
|
def QI8 : QuantizedType<"Uniform", [8], 1>;
|
|
def QUI16 : QuantizedType<"Uniform", [16], 0>;
|
|
def QI16 : QuantizedType<"Uniform", [16], 1>;
|
|
def QUI32 : QuantizedType<"Uniform", [32], 0>;
|
|
def QI32 : QuantizedType<"Uniform", [32], 1>;
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// TFL native op traits (for quantization).
|
|
//
|
|
// Ops in this link should have those traits specified:
|
|
// https://www.tensorflow.org/lite/performance/quantization_spec
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// TODO(b/157870442): replace all FixedResultScale trait
|
|
def FixedOutputRangeInterface : OpInterface<
|
|
"FixedOutputRangeInterface"> {
|
|
let description = [{
|
|
Interface for defining the fixed output range.
|
|
}];
|
|
|
|
let methods = [
|
|
InterfaceMethod<
|
|
[{Returns the fixed output range.}],
|
|
"UniformQuantizedType", "GetFixedOutputRange",
|
|
(ins "bool":$sign, "int":$bit_width)
|
|
>,
|
|
];
|
|
}
|
|
|
|
// Specify this trait if the op has a fixed output value range.
|
|
class FixedResultScale<QuantizedType qt> : NativeOpTrait<!strconcat(
|
|
"quant::FixedResult", qt.name, "Scale<", qt.asTraitArgsStr, ">::Impl")>;
|
|
|
|
// Specify this trait if the op requires same inputs and outputs quantization
|
|
// scales.
|
|
def SameOperandsAndResultsScale : NativeOpTrait<
|
|
"quant::SameOperandsAndResultsScale">;
|
|
|
|
// Specify this trait if the bias-th input of the op is a bias input, which
|
|
// needs a scale based on the scales of op1 and op2.
|
|
class AccumulatorUniformScale<int bias, int op1, int op2> : NativeOpTrait<
|
|
!strconcat("quant::AccumulatorUniformScale<",
|
|
StrJoinInt<[bias, op1, op2]>.result,
|
|
">::Impl")>;
|
|
|
|
// Specify the operand index of the coefficient operand for an affine op
|
|
// and also the quantization dimension if per-axis quantization is support.
|
|
// If the quantization dimension is -1, per-axis quantization isn't supported.
|
|
class AffineOpCoefficient<int dim, int index> : NativeOpTrait<
|
|
!strconcat("quant::AffineOpCoefficient<",
|
|
StrJoinInt<[dim, index]>.result,
|
|
">::Impl")>;
|
|
|
|
// Specify this trait if the op doesn't have quantizable output. We shouldn't
|
|
// apply quantization on this op.
|
|
def NoQuantizableResult : NativeOpTrait<"quant::NoQuantizableResult">;
|
|
|
|
#endif // TF_Quantization
|