blob: 3638b3b356579a7abb60121661ccaaaccf2ff0b4 [file] [log] [blame]
Eric Kunzee5e26762020-10-13 16:11:07 -07001
2// Copyright (c) 2020, ARM Limited.
3//
4// Licensed under the Apache License, Version 2.0 (the "License");
5// you may not use this file except in compliance with the License.
6// You may obtain a copy of the License at
7//
8// http://www.apache.org/licenses/LICENSE-2.0
9//
10// Unless required by applicable law or agreed to in writing, software
11// distributed under the License is distributed on an "AS IS" BASIS,
12// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13// See the License for the specific language governing permissions and
14// limitations under the License.
15
16#ifndef TOSA_REFERENCE_QUANT_UTIL_H
17#define TOSA_REFERENCE_QUANT_UTIL_H
18
19#include "arith_util.h"
20#include "func_debug.h"
21#include "ops/template_types.h"
22#include "tosa_generated.h"
23
24using namespace tosa;
25
26namespace TosaReference
27{
28
29template <DType AccDType>
30class QuantUtil
31{
32public:
33 using T = typename GetEigenType<AccDType>::type;
34
35 static void reciprocal_scale(int32_t value,
36 // Output
37 int32_t& multiplier,
38 int32_t& shift)
39 {
40 ASSERT_MSG(value > 0, "AvgPool2d reciprocal_scale() error: # of elements should be > 1 but is %d", value);
41 uint32_t value_u32 = (uint32_t)value;
42 int32_t k = 32 - LEADING_ZEROS_32(value_u32 - 1); // (1<<k)/2 < value <= (1<<k)
43 int64_t numerator = ((1L << 30) + 1) << k;
44 multiplier = numerator / value; // (1<<30) <= multiplier < (1<<31)
45 shift = 30 + k;
46 }
47
48 static int32_t apply_scale(T value, int32_t multiplier, int32_t shift, bool enabled_adjusted_rounding = true)
49 {
50 if (AccDType == DType_FLOAT)
51 {
52 return value;
53 }
54 ASSERT_MSG(multiplier >= 0, "apply_scale() error: multiplier should >= 0 but is %d", multiplier);
55 int64_t round = (shift > 0) ? (1L << (shift - 1)) : 0;
56 if (enabled_adjusted_rounding)
57 {
58 if (AccDType != DType_INT48)
59 {
60 if (shift > 31 && value >= 0)
61 round += (1L << 30);
62 if (shift > 31 && value < 0)
63 round -= (1L << 30);
64 }
65 else
66 { // input data could be int16, which leads to 48 bits accumulator
67 ASSERT_MSG(multiplier < (1 << 15), "apply_scale() error: multiplier should <= %d in 48 bit mode",
68 (1 << 15));
69 }
70 }
71 int64_t result = (int64_t)value * multiplier + round;
72 result = result >> shift;
73 ASSERT_MSG(result >= -(1L << 31) && result < (1L << 31),
74 "apply_scale() error: scaled result exceed int32 numeric range");
75 return static_cast<int32_t>(result);
76 }
77};
78
79class TypeChecker
80{
81public:
82 static bool is_integer(DType dtype)
83 {
84 if (dtype == DType_INT4 || dtype == DType_INT8 || dtype == DType_AINT8 || dtype == DType_UINT8 ||
85 dtype == DType_INT16 || dtype == DType_INT32 || dtype == DType_INT48)
86 {
87 return true;
88 }
89 return false;
90 }
91 static bool is_symmetric(DType dtype)
92 {
93 if (dtype == DType_INT4 || dtype == DType_INT8 || dtype == DType_INT16 || dtype == DType_INT32 ||
94 dtype == DType_INT48)
95 {
96 return true;
97 }
98 return false;
99 }
100};
101}; // namespace TosaReference
102
103#endif