| Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame^] | 1 | # Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. |
| 2 | # |
| 3 | # SPDX-License-Identifier: Apache-2.0 |
| 4 | # |
| 5 | # Licensed under the Apache License, Version 2.0 (the License); you may |
| 6 | # not use this file except in compliance with the License. |
| 7 | # You may obtain a copy of the License at |
| 8 | # |
| 9 | # www.apache.org/licenses/LICENSE-2.0 |
| 10 | # |
| 11 | # Unless required by applicable law or agreed to in writing, software |
| 12 | # distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| 13 | # WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | # See the License for the specific language governing permissions and |
| 15 | # limitations under the License. |
| 16 | # Description: |
| 17 | # Contains SoftMax |
| 18 | import numpy as np |
| 19 | |
| 20 | from . import scaling |
| 21 | from .data_type import DataType |
| 22 | from .operation import Operation |
| 23 | from .tensor import Tensor |
| 24 | from .tensor import TensorPurpose |
| 25 | |
| 26 | |
| 27 | class TensorUtil: |
| 28 | # TODO: Move these functions to Tensor/Operation classes |
| 29 | @staticmethod |
| 30 | def create_const_tensor( |
| 31 | name, shape, dtype, values, value_dtype=None, purpose=TensorPurpose.Unknown, quantization=None |
| 32 | ): |
| 33 | const_op = Operation("Const", name) |
| 34 | const_tensor = Tensor(shape, dtype, name + "_0") |
| 35 | const_tensor.purpose = purpose |
| 36 | const_tensor.quantization = quantization |
| 37 | const_tensor.values = np.array(values, dtype=value_dtype) |
| 38 | const_tensor.quant_values = np.frombuffer(const_tensor.values.tobytes(), dtype=np.uint8) |
| 39 | const_tensor.ops.append(const_op) |
| 40 | const_op.outputs.append(const_tensor) |
| 41 | return const_tensor |
| 42 | |
| 43 | @staticmethod |
| 44 | def add_ifm_tensor(op, tens): |
| 45 | op.inputs.append(tens) |
| 46 | tens.consumer_list.append(op) |
| 47 | |
| 48 | @staticmethod |
| 49 | def set_ofm_tensor(op, tens): |
| 50 | tens.ops = [op] |
| 51 | op.outputs = [tens] |
| 52 | |
| 53 | @staticmethod |
| 54 | def reshape(tens, shape, ifm_reshape=True): |
| 55 | if shape == tens.shape: |
| 56 | return tens |
| 57 | name = tens.name + "_reshape" |
| 58 | reshape_op = Operation("Reshape", name) |
| 59 | reshape_op.attrs["new_shape"] = shape |
| 60 | reshape_ifm = tens |
| 61 | reshape_ofm = tens.clone("_reshaped") |
| 62 | reshape_ofm.shape = reshape_ofm.storage_shape = reshape_ofm.bandwidth_shape = shape |
| 63 | if not ifm_reshape: |
| 64 | reshape_ifm, reshape_ofm = reshape_ofm, reshape_ifm |
| 65 | reshape_op.inputs = [reshape_ifm, TensorUtil.create_const_tensor(name + "_shape", [1], DataType.int32, shape)] |
| 66 | TensorUtil.set_ofm_tensor(reshape_op, reshape_ofm) |
| 67 | return reshape_ofm if ifm_reshape else reshape_ifm |
| 68 | |
| 69 | @staticmethod |
| 70 | def get_full_shape(shape): |
| 71 | d = len(shape) |
| 72 | if d in (1, 3): |
| 73 | return [1] * (4 - d) + shape |
| 74 | elif d == 2: |
| 75 | return [shape[0], 1, 1, shape[1]] |
| 76 | else: |
| 77 | return shape |
| 78 | |
| 79 | |
| 80 | class SoftMax: |
| 81 | # Turn off black formatting for the LUT tables to keep them compact |
| 82 | # fmt: off |
| 83 | EXP_LUT = [ |
| 84 | 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, |
| 85 | 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, |
| 86 | 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, 0x00000002, |
| 87 | 0x00000002, 0x00000002, 0x00010002, 0x00000003, 0x00000003, 0x00000003, 0x00000003, 0x00000003, |
| 88 | 0x00000003, 0x00000003, 0x00000003, 0x00000003, 0x00000003, 0x00000003, 0x00000003, 0x00000003, |
| 89 | 0x00000003, 0x00000003, 0x00000003, 0x00010003, 0x00000004, 0x00000004, 0x00000004, 0x00000004, |
| 90 | 0x00000004, 0x00000004, 0x00000004, 0x00000004, 0x00000004, 0x00000004, 0x00000004, 0x00000004, |
| 91 | 0x00010004, 0x00000005, 0x00000005, 0x00000005, 0x00000005, 0x00000005, 0x00000005, 0x00000005, |
| 92 | 0x00000005, 0x00000005, 0x00010005, 0x00000006, 0x00000006, 0x00000006, 0x00000006, 0x00000006, |
| 93 | 0x00000006, 0x00000006, 0x00010006, 0x00000007, 0x00000007, 0x00000007, 0x00000007, 0x00000007, |
| 94 | 0x00000007, 0x00000007, 0x00010007, 0x00000008, 0x00000008, 0x00000008, 0x00000008, 0x00000008, |
| 95 | 0x00010008, 0x00000009, 0x00000009, 0x00000009, 0x00000009, 0x00000009, 0x00010009, 0x0000000a, |
| 96 | 0x0000000a, 0x0000000a, 0x0000000a, 0x0001000a, 0x0000000b, 0x0000000b, 0x0000000b, 0x0000000b, |
| 97 | 0x0001000b, 0x0000000c, 0x0000000c, 0x0000000c, 0x0001000c, 0x0000000d, 0x0000000d, 0x0000000d, |
| 98 | 0x0001000d, 0x0000000e, 0x0000000e, 0x0000000e, 0x0001000e, 0x0000000f, 0x0000000f, 0x0001000f, |
| 99 | 0x00000010, 0x00000010, 0x00010010, 0x00000011, 0x00000011, 0x00010011, 0x00000012, 0x00000012, |
| 100 | 0x00010012, 0x00000013, 0x00000013, 0x00010013, 0x00000014, 0x00010014, 0x00000015, 0x00000015, |
| 101 | 0x00010015, 0x00000016, 0x00010016, 0x00000017, 0x00010017, 0x00000018, 0x00010018, 0x00000019, |
| 102 | 0x00010019, 0x0000001a, 0x0001001a, 0x0000001b, 0x0001001b, 0x0000001c, 0x0001001c, 0x0000001d, |
| 103 | 0x0001001d, 0x0000001e, 0x0001001e, 0x0001001f, 0x00000020, 0x00010020, 0x00010021, 0x00000022, |
| 104 | 0x00010022, 0x00010023, 0x00000024, 0x00010024, 0x00000025, 0x00010025, 0x00010026, 0x00010027, |
| 105 | 0x00000028, 0x00020028, 0x0000002a, 0x0001002a, 0x0001002b, 0x0001002c, 0x0000002d, 0x0001002d, |
| 106 | 0x0001002e, 0x0001002f, 0x00010030, 0x00010031, 0x00010032, 0x00010033, 0x00010034, 0x00010035, |
| 107 | 0x00010036, 0x00010037, 0x00010038, 0x00020039, 0x0001003b, 0x0000003c, 0x0002003c, 0x0001003e, |
| 108 | 0x0002003f, 0x00000041, 0x00020041, 0x00010043, 0x00010044, 0x00020045, 0x00020047, 0x00010049, |
| 109 | 0x0001004a, 0x0002004b, 0x0001004d, 0x0002004e, 0x00010050, 0x00020051, 0x00020053, 0x00010055, |
| 110 | 0x00020056, 0x00020058, 0x0002005a, 0x0001005c, 0x0002005d, 0x0002005f, 0x00020061, 0x00020063, |
| 111 | 0x00020065, 0x00020067, 0x00020069, 0x0002006b, 0x0003006d, 0x00020070, 0x00020072, 0x00020074, |
| 112 | 0x00030076, 0x00020079, 0x0003007b, 0x0002007e, 0x00030080, 0x00020083, 0x00020085, 0x00040087, |
| 113 | 0x0002008b, 0x0003008d, 0x00030090, 0x00020093, 0x00030095, 0x00030098, 0x0003009b, 0x0004009e, |
| 114 | 0x000300a2, 0x000300a5, 0x000300a8, 0x000300ab, 0x000400ae, 0x000300b2, 0x000400b5, 0x000400b9, |
| 115 | 0x000300bd, 0x000400c0, 0x000400c4, 0x000400c8, 0x000400cc, 0x000400d0, 0x000500d4, 0x000400d9, |
| 116 | 0x000400dd, 0x000500e1, 0x000400e6, 0x000500ea, 0x000400ef, 0x000500f3, 0x000500f8, 0x000500fd, |
| 117 | 0x00050102, 0x00050107, 0x0005010c, 0x00060111, 0x00050117, 0x0006011c, 0x00060122, 0x00060128, |
| 118 | 0x0006012e, 0x00060134, 0x0006013a, 0x00070140, 0x00060147, 0x0007014d, 0x00060154, 0x0007015a, |
| 119 | 0x00070161, 0x00060168, 0x0008016e, 0x00070176, 0x0008017d, 0x00080185, 0x0007018d, 0x00090194, |
| 120 | 0x0008019d, 0x000801a5, 0x000801ad, 0x000901b5, 0x000901be, 0x000901c7, 0x000901d0, 0x000901d9, |
| 121 | 0x000a01e2, 0x000901ec, 0x000a01f5, 0x000b01ff, 0x000a020a, 0x000b0214, 0x000a021f, 0x000b0229, |
| 122 | 0x000b0234, 0x000b023f, 0x000c024a, 0x000c0256, 0x000c0262, 0x000c026e, 0x000c027a, 0x000d0286, |
| 123 | 0x000d0293, 0x000d02a0, 0x000e02ad, 0x000e02bb, 0x000e02c9, 0x000e02d7, 0x000f02e5, 0x000f02f4, |
| 124 | 0x000f0303, 0x000f0312, 0x00100321, 0x00100331, 0x00110341, 0x00100352, 0x00120362, 0x00110374, |
| 125 | 0x00120385, 0x00120397, 0x001203a9, 0x001303bb, 0x001303ce, 0x001403e1, 0x001403f5, 0x00140409, |
| 126 | 0x0015041d, 0x00150432, 0x00160447, 0x0016045d, 0x00160473, 0x00170489, 0x001704a0, 0x001904b7, |
| 127 | 0x001804d0, 0x001904e8, 0x00190501, 0x001a051a, 0x001a0534, 0x001b054e, 0x001b0569, 0x001c0584, |
| 128 | 0x001c05a0, 0x001d05bc, 0x001e05d9, 0x001e05f7, 0x001e0615, 0x00200633, 0x00200653, 0x00200673, |
| 129 | 0x00210693, 0x002206b4, 0x002306d6, 0x002306f9, 0x0024071c, 0x00240740, 0x00260764, 0x0026078a, |
| 130 | 0x002607b0, 0x002807d6, 0x002907fe, 0x00290827, 0x002a0850, 0x002a087a, 0x002c08a4, 0x002c08d0, |
| 131 | 0x002e08fc, 0x002e092a, 0x002f0958, 0x00310987, 0x003109b8, 0x003209e9, 0x00330a1b, 0x00340a4e, |
| 132 | 0x00350a82, 0x00350ab7, 0x00380aec, 0x00380b24, 0x003a0b5c, 0x003a0b96, 0x003c0bd0, 0x003d0c0c, |
| 133 | 0x003e0c49, 0x003f0c87, 0x00400cc6, 0x00420d06, 0x00430d48, 0x00440d8b, 0x00460dcf, 0x00480e15, |
| 134 | 0x00480e5d, 0x00490ea5, 0x004c0eee, 0x004d0f3a, 0x004e0f87, 0x00500fd5, 0x00511025, 0x00531076, |
| 135 | 0x005610c9, 0x0056111f, 0x00581175, 0x005a11cd, 0x005c1227, 0x005e1283, 0x005e12e1, 0x0061133f, |
| 136 | 0x006413a0, 0x00651404, 0x00671469, 0x006914d0, 0x006c1539, 0x006c15a5, 0x00701611, 0x00721681, |
| 137 | 0x007416f3, 0x00761767, 0x007917dd, 0x007a1856, 0x007d18d0, 0x0080194d, 0x008319cd, 0x00841a50, |
| 138 | 0x00881ad4, 0x00891b5c, 0x008d1be5, 0x00911c72, 0x00911d03, 0x00961d94, 0x00981e2a, 0x009c1ec2, |
| 139 | 0x009e1f5e, 0x00a21ffc, 0x00a4209e, 0x00a92142, 0x00ab21eb, 0x00ae2296, 0x00b22344, 0x00b523f6, |
| 140 | 0x00b924ab, 0x00be2564, 0x00c02622, 0x00c526e2, 0x00c827a7, 0x00cc286f, 0x00d0293b, 0x00d52a0b, |
| 141 | 0x00d72ae0, 0x00dd2bb7, 0x00e12c94, 0x00e62d75, 0x00eb2e5b, 0x00ef2f46, 0x00f23035, 0x00f83127, |
| 142 | 0x00fe321f, 0x0101331d, 0x0108341e, 0x010c3526, 0x01123632, 0x01173744, 0x011c385b, 0x01233977, |
| 143 | 0x01273a9a, 0x012e3bc1, 0x01343cef, 0x013a3e23, 0x01403f5d, 0x0146409d, 0x014c41e3, 0x0154432f, |
| 144 | 0x01594483, 0x016145dc, 0x0168473d, 0x016f48a5, 0x01764a14, 0x017d4b8a, 0x01854d07, 0x018d4e8c, |
| 145 | 0x01945019, 0x019d51ad, 0x01a4534a, 0x01ad54ee, 0x01b5569b, 0x01be5850, 0x01c75a0e, 0x01d05bd5, |
| 146 | 0x01d85da5, 0x01e35f7d, 0x01eb6160, 0x01f6634b, 0x01ff6541, 0x02096740, 0x02146949, 0x021e6b5d, |
| 147 | 0x02296d7b, 0x02336fa4, 0x023f71d7, 0x024a7416, 0x02567660, 0x026278b6, 0x026d7b18, 0x027a7d85, |
| 148 | ] |
| 149 | |
| 150 | ONE_OVER_ONE_PLUS_X_LUT = [ |
| 151 | 0xffc17fff, 0xffc07fc0, 0xffc27f80, 0xffc07f42, 0xffc17f02, 0xffc17ec3, 0xffc27e84, 0xffc27e46, |
| 152 | 0xffc27e08, 0xffc37dca, 0xffc27d8d, 0xffc37d4f, 0xffc37d12, 0xffc37cd5, 0xffc37c98, 0xffc47c5b, |
| 153 | 0xffc47c1f, 0xffc47be3, 0xffc57ba7, 0xffc57b6c, 0xffc37b31, 0xffc67af4, 0xffc57aba, 0xffc67a7f, |
| 154 | 0xffc57a45, 0xffc67a0a, 0xffc779d0, 0xffc67997, 0xffc6795d, 0xffc77923, 0xffc778ea, 0xffc778b1, |
| 155 | 0xffc87878, 0xffc77840, 0xffc87807, 0xffc877cf, 0xffc97797, 0xffc87760, 0xffc97728, 0xffc976f1, |
| 156 | 0xffc976ba, 0xffc87683, 0xffca764b, 0xffca7615, 0xffca75df, 0xffca75a9, 0xffca7573, 0xffcb753d, |
| 157 | 0xffca7508, 0xffcb74d2, 0xffcb749d, 0xffca7468, 0xffcc7432, 0xffcc73fe, 0xffcb73ca, 0xffcc7395, |
| 158 | 0xffcd7361, 0xffcc732e, 0xffcc72fa, 0xffcd72c6, 0xffcd7293, 0xffcd7260, 0xffcc722d, 0xffce71f9, |
| 159 | 0xffcd71c7, 0xffce7194, 0xffce7162, 0xffce7130, 0xffcf70fe, 0xffce70cd, 0xffce709b, 0xffcf7069, |
| 160 | 0xffcf7038, 0xffcf7007, 0xffcf6fd6, 0xffcf6fa5, 0xffd06f74, 0xffd06f44, 0xffd06f14, 0xffd06ee4, |
| 161 | 0xffd06eb4, 0xffd06e84, 0xffd16e54, 0xffd16e25, 0xffd16df6, 0xffd16dc7, 0xffd06d98, 0xffd26d68, |
| 162 | 0xffd16d3a, 0xffd26d0b, 0xffd26cdd, 0xffd26caf, 0xffd26c81, 0xffd26c53, 0xffd36c25, 0xffd26bf8, |
| 163 | 0xffd36bca, 0xffd36b9d, 0xffd36b70, 0xffd26b43, 0xffd46b15, 0xffd36ae9, 0xffd46abc, 0xffd46a90, |
| 164 | 0xffd46a64, 0xffd46a38, 0xffd46a0c, 0xffd469e0, 0xffd469b4, 0xffd56988, 0xffd5695d, 0xffd56932, |
| 165 | 0xffd56907, 0xffd568dc, 0xffd568b1, 0xffd56886, 0xffd6685b, 0xffd56831, 0xffd66806, 0xffd667dc, |
| 166 | 0xffd667b2, 0xffd76788, 0xffd6675f, 0xffd76735, 0xffd6670c, 0xffd766e2, 0xffd666b9, 0xffd7668f, |
| 167 | 0xffd86666, 0xffd6663e, 0xffd86614, 0xffd765ec, 0xffd865c3, 0xffd8659b, 0xffd86573, 0xffd8654b, |
| 168 | 0xffd86523, 0xffd864fb, 0xffd964d3, 0xffd864ac, 0xffd96484, 0xffd8645d, 0xffd96435, 0xffd9640e, |
| 169 | 0xffd963e7, 0xffd963c0, 0xffd96399, 0xffda6372, 0xffd9634c, 0xffda6325, 0xffda62ff, 0xffda62d9, |
| 170 | 0xffda62b3, 0xffda628d, 0xffda6267, 0xffdb6241, 0xffda621c, 0xffdb61f6, 0xffda61d1, 0xffdc61ab, |
| 171 | 0xffd96187, 0xffdc6160, 0xffdb613c, 0xffdb6117, 0xffdb60f2, 0xffdc60cd, 0xffdc60a9, 0xffdb6085, |
| 172 | 0xffdc6060, 0xffdc603c, 0xffdc6018, 0xffdc5ff4, 0xffdc5fd0, 0xffdd5fac, 0xffdc5f89, 0xffdc5f65, |
| 173 | 0xffdd5f41, 0xffdd5f1e, 0xffdd5efb, 0xffdd5ed8, 0xffdd5eb5, 0xffdd5e92, 0xffdd5e6f, 0xffdd5e4c, |
| 174 | 0xffdd5e29, 0xffde5e06, 0xffde5de4, 0xffdd5dc2, 0xffde5d9f, 0xffde5d7d, 0xffde5d5b, 0xffde5d39, |
| 175 | 0xffdf5d17, 0xffde5cf6, 0xffde5cd4, 0xffdf5cb2, 0xffdf5c91, 0xffde5c70, 0xffdf5c4e, 0xffdf5c2d, |
| 176 | 0xffde5c0c, 0xffe05bea, 0xffdf5bca, 0xffdf5ba9, 0xffdf5b88, 0xffdf5b67, 0xffe05b46, 0xffe05b26, |
| 177 | 0xffdf5b06, 0xffe05ae5, 0xffe05ac5, 0xffe05aa5, 0xffe05a85, 0xffe05a65, 0xffe05a45, 0xffe15a25, |
| 178 | 0xffe05a06, 0xffe059e6, 0xffe159c6, 0xffe159a7, 0xffe05988, 0xffe15968, 0xffe15949, 0xffe1592a, |
| 179 | 0xffe1590b, 0xffe158ec, 0xffe258cd, 0xffe158af, 0xffe15890, 0xffe25871, 0xffe15853, 0xffe25834, |
| 180 | 0xffe25816, 0xffe257f8, 0xffe157da, 0xffe257bb, 0xffe3579d, 0xffe25780, 0xffe25762, 0xffe25744, |
| 181 | 0xffe35726, 0xffe25709, 0xffe256eb, 0xffe356cd, 0xffe356b0, 0xffe35693, 0xffe25676, 0xffe35658, |
| 182 | 0xffe3563b, 0xffe3561e, 0xffe35601, 0xffe355e4, 0xffe455c7, 0xffe355ab, 0xffe4558e, 0xffe35572, |
| 183 | 0xffe45555, 0xffe35539, 0xffe4551c, 0xffe45500, 0xffe454e4, 0xffe454c8, 0xffe454ac, 0xffe45490, |
| 184 | 0xffe45474, 0xffe55458, 0xffe4543d, 0xffe45421, 0xffe55405, 0xffe553ea, 0xffe453cf, 0xffe553b3, |
| 185 | 0xffe45398, 0xffe5537c, 0xffe55361, 0xffe55346, 0xffe5532b, 0xffe55310, 0xffe552f5, 0xffe552da, |
| 186 | 0xffe652bf, 0xffe552a5, 0xffe5528a, 0xffe6526f, 0xffe55255, 0xffe6523a, 0xffe65220, 0xffe55206, |
| 187 | 0xffe651eb, 0xffe651d1, 0xffe651b7, 0xffe6519d, 0xffe65183, 0xffe65169, 0xffe7514f, 0xffe65136, |
| 188 | 0xffe6511c, 0xffe75102, 0xffe650e9, 0xffe750cf, 0xffe650b6, 0xffe7509c, 0xffe75083, 0xffe6506a, |
| 189 | 0xffe75050, 0xffe75037, 0xffe7501e, 0xffe75005, 0xffe74fec, 0xffe74fd3, 0xffe74fba, 0xffe74fa1, |
| 190 | 0xffe84f88, 0xffe74f70, 0xffe84f57, 0xffe74f3f, 0xffe84f26, 0xffe74f0e, 0xffe84ef5, 0xffe84edd, |
| 191 | 0xffe84ec5, 0xffe84ead, 0xffe74e95, 0xffe84e7c, 0xffe84e64, 0xffe94e4c, 0xffe84e35, 0xffe84e1d, |
| 192 | 0xffe84e05, 0xffe94ded, 0xffe84dd6, 0xffe84dbe, 0xffe94da6, 0xffe94d8f, 0xffe84d78, 0xffe84d60, |
| 193 | 0xffea4d48, 0xffe84d32, 0xffe94d1a, 0xffe94d03, 0xffe84cec, 0xffe94cd4, 0xffe94cbd, 0xffea4ca6, |
| 194 | 0xffe94c90, 0xffe84c79, 0xffea4c61, 0xffe94c4b, 0xffe94c34, 0xffea4c1d, 0xffe94c07, 0xffea4bf0, |
| 195 | 0xffe94bda, 0xffea4bc3, 0xffea4bad, 0xffe94b97, 0xffea4b80, 0xffea4b6a, 0xffea4b54, 0xffea4b3e, |
| 196 | 0xffea4b28, 0xffea4b12, 0xffea4afc, 0xffea4ae6, 0xffea4ad0, 0xffeb4aba, 0xffea4aa5, 0xffea4a8f, |
| 197 | 0xffeb4a79, 0xffea4a64, 0xffea4a4e, 0xffeb4a38, 0xffeb4a23, 0xffea4a0e, 0xffeb49f8, 0xffea49e3, |
| 198 | 0xffeb49cd, 0xffeb49b8, 0xffeb49a3, 0xffeb498e, 0xffea4979, 0xffeb4963, 0xffeb494e, 0xffec4939, |
| 199 | 0xffeb4925, 0xffea4910, 0xffec48fa, 0xffeb48e6, 0xffeb48d1, 0xffec48bc, 0xffeb48a8, 0xffec4893, |
| 200 | 0xffeb487f, 0xffec486a, 0xffeb4856, 0xffec4841, 0xffec482d, 0xffeb4819, 0xffec4804, 0xffec47f0, |
| 201 | 0xffec47dc, 0xffec47c8, 0xffec47b4, 0xffec47a0, 0xffec478c, 0xffec4778, 0xffec4764, 0xffec4750, |
| 202 | 0xffec473c, 0xffed4728, 0xffec4715, 0xffec4701, 0xffed46ed, 0xffec46da, 0xffed46c6, 0xffec46b3, |
| 203 | 0xffec469f, 0xffed468b, 0xffed4678, 0xffec4665, 0xffed4651, 0xffed463e, 0xffed462b, 0xffec4618, |
| 204 | 0xffed4604, 0xffed45f1, 0xffed45de, 0xffed45cb, 0xffed45b8, 0xffed45a5, 0xffed4592, 0xffed457f, |
| 205 | 0xffee456c, 0xffed455a, 0xffed4547, 0xffed4534, 0xffee4521, 0xffed450f, 0xffed44fc, 0xffee44e9, |
| 206 | 0xffed44d7, 0xffee44c4, 0xffee44b2, 0xffed44a0, 0xffee448d, 0xffee447b, 0xffed4469, 0xffee4456, |
| 207 | 0xffee4444, 0xffee4432, 0xffee4420, 0xffee440e, 0xffee43fc, 0xffee43ea, 0xffee43d8, 0xffee43c6, |
| 208 | 0xffee43b4, 0xffee43a2, 0xffee4390, 0xffef437e, 0xffee436d, 0xffee435b, 0xffef4349, 0xffee4338, |
| 209 | 0xffee4326, 0xffef4314, 0xffee4303, 0xffef42f1, 0xffee42e0, 0xffef42ce, 0xffee42bd, 0xffef42ab, |
| 210 | 0xffef429a, 0xffee4289, 0xfff04277, 0xffee4267, 0xffef4255, 0xffef4244, 0xffef4233, 0xffef4222, |
| 211 | 0xffee4211, 0xffef41ff, 0xfff041ee, 0xffef41de, 0xffef41cd, 0xffee41bc, 0xfff041aa, 0xffef419a, |
| 212 | 0xffef4189, 0xffef4178, 0xfff04167, 0xffef4157, 0xffef4146, 0xfff04135, 0xffef4125, 0xfff04114, |
| 213 | 0xffef4104, 0xfff040f3, 0xffef40e3, 0xfff040d2, 0xfff040c2, 0xffef40b2, 0xfff040a1, 0xfff04091, |
| 214 | 0xfff04081, 0xffef4071, 0xfff04060, 0xfff04050, 0xfff04040, 0xfff04030, 0xfff04020, 0xfff04010 |
| 215 | ] |
| 216 | # fmt: on |
| 217 | |
| 218 | def __init__(self, op): |
| 219 | self.op = op |
| 220 | |
| 221 | def get_graph(self): |
| 222 | ifm = self.op.inputs[0] |
| 223 | ofm = self.op.outputs[0] |
| 224 | |
| 225 | if ifm.dtype == DataType.int16 and ofm.dtype == DataType.int16: |
| 226 | return self.get_graph_int16(ifm, ofm) |
| 227 | else: |
| 228 | self.op.run_on_npu = False |
| 229 | return self.op |
| 230 | |
| 231 | def get_graph_int16(self, ifm, ofm): |
| 232 | ifm = TensorUtil.reshape(ifm, TensorUtil.get_full_shape(ifm.shape)) |
| 233 | ofm = TensorUtil.reshape(ofm, TensorUtil.get_full_shape(ofm.shape), False) |
| 234 | no_scale_quant = ifm.quantization.clone() |
| 235 | no_scale_quant.scale_f32 = None |
| 236 | |
| 237 | # PASS 0 - Depthwise Maxpool |
| 238 | maxpool_op = self.op.clone("_maxpool0") |
| 239 | maxpool_op.type = "MaxPool" |
| 240 | maxpool_h = ifm.shape[1] * ifm.shape[2] |
| 241 | maxpool_w = ifm.shape[3] |
| 242 | maxpool_ifm_shape = [1, maxpool_h, maxpool_w, 1] |
| 243 | maxpool_op.attrs["padding"] = b"VALID" |
| 244 | maxpool_op.attrs["stride_w"] = 1 |
| 245 | maxpool_op.attrs["stride_h"] = 1 |
| 246 | maxpool_op.attrs["filter_width"] = maxpool_w |
| 247 | maxpool_op.attrs["filter_height"] = 1 |
| 248 | maxpool_op.attrs["strides"] = [1, maxpool_op.attrs["stride_h"], maxpool_op.attrs["stride_w"], 1] |
| 249 | maxpool_op.attrs["ksize"] = [1, maxpool_op.attrs["filter_height"], maxpool_op.attrs["filter_width"], 1] |
| 250 | maxpool_op.inputs = [TensorUtil.reshape(ifm, maxpool_ifm_shape)] |
| 251 | maxpool_ofm = Tensor([1, maxpool_h, 1, 1], DataType.int16, maxpool_op.name + "_0") |
| 252 | maxpool_ofm.quantization = no_scale_quant |
| 253 | TensorUtil.set_ofm_tensor(maxpool_op, maxpool_ofm) |
| 254 | |
| 255 | # PASS 1 - Sub |
| 256 | sub1_op = Operation("SubAct", self.op.name + "_sub1") |
| 257 | TensorUtil.add_ifm_tensor(sub1_op, ifm) |
| 258 | TensorUtil.add_ifm_tensor(sub1_op, TensorUtil.reshape(maxpool_ofm, [1, ifm.shape[1], ifm.shape[2], 1])) |
| 259 | sub1_ofm = Tensor(ifm.shape, DataType.int32, sub1_op.name + "_0") |
| 260 | sub1_ofm.quantization = ifm.quantization.clone() |
| 261 | TensorUtil.set_ofm_tensor(sub1_op, sub1_ofm) |
| 262 | |
| 263 | # PASS 2 - Mul |
| 264 | beta = self.op.attrs.get("beta", 1.0) |
| 265 | mul2_out_range = 10.0 / 65535.0 |
| 266 | mul2_scale, _ = scaling.elementwise_mul_scale(sub1_ofm.quantization.scale_f32, beta, mul2_out_range) |
| 267 | mul2_quant = ifm.quantization.clone() |
| 268 | mul2_quant.scale_f32 = beta |
| 269 | mul2_op = Operation("MulAct", self.op.name + "_mul2") |
| 270 | TensorUtil.add_ifm_tensor(mul2_op, sub1_ofm) |
| 271 | TensorUtil.add_ifm_tensor( |
| 272 | mul2_op, |
| 273 | TensorUtil.create_const_tensor( |
| 274 | mul2_op.name + "_const", [1, 1, 1, 1], DataType.int32, [mul2_scale], np.uint32, quantization=mul2_quant |
| 275 | ), |
| 276 | ) |
| 277 | mul2_ofm = Tensor(ifm.shape, DataType.int32, mul2_op.name + "_0") |
| 278 | mul2_ofm.quantization = ofm.quantization.clone() |
| 279 | mul2_ofm.quantization.scale_f32 = mul2_out_range |
| 280 | TensorUtil.set_ofm_tensor(mul2_op, mul2_ofm) |
| 281 | |
| 282 | # PASS 3 - Add+LUT(exp) |
| 283 | add_op = Operation("AddAct", self.op.name + "_add3") |
| 284 | TensorUtil.add_ifm_tensor(add_op, mul2_ofm) |
| 285 | TensorUtil.add_ifm_tensor( |
| 286 | add_op, |
| 287 | TensorUtil.create_const_tensor( |
| 288 | add_op.name + "_const", [1, 1, 1, 1], DataType.int32, [32767], np.uint32, quantization=no_scale_quant |
| 289 | ), |
| 290 | ) |
| 291 | add_op.set_activation_lut( |
| 292 | TensorUtil.create_const_tensor( |
| 293 | add_op.name + "_lut", [1, 1, 1, 512], DataType.int32, self.EXP_LUT, np.uint32, TensorPurpose.LUT |
| 294 | ) |
| 295 | ) |
| 296 | exp_ofm = Tensor(mul2_ofm.shape, DataType.int16, add_op.name + "_0") |
| 297 | exp_ofm.quantization = mul2_ofm.quantization.clone() |
| 298 | TensorUtil.set_ofm_tensor(add_op, exp_ofm) |
| 299 | |
| 300 | # PASS 4 - Reduce sum |
| 301 | reduce_sum_op = Operation("ReduceSum", self.op.name + "_reduce_sum4") |
| 302 | reduce_sum_op.attrs["padding"] = b"VALID" |
| 303 | reduce_sum_op.attrs["stride_w"] = 1 |
| 304 | reduce_sum_op.attrs["stride_h"] = 1 |
| 305 | reduce_sum_op.attrs["filter_width"] = 1 |
| 306 | reduce_sum_op.attrs["filter_height"] = 1 |
| 307 | reduce_sum_op.attrs["strides"] = [1, reduce_sum_op.attrs["stride_h"], reduce_sum_op.attrs["stride_w"], 1] |
| 308 | reduce_sum_op.attrs["ksize"] = [1, reduce_sum_op.attrs["filter_height"], reduce_sum_op.attrs["filter_width"], 1] |
| 309 | TensorUtil.add_ifm_tensor(reduce_sum_op, exp_ofm) |
| 310 | |
| 311 | reduce_sum_shape = [1, exp_ofm.shape[1], exp_ofm.shape[2], 1] |
| 312 | sum_of_exp = Tensor(reduce_sum_shape, DataType.int32, reduce_sum_op.name + "_0") |
| 313 | sum_of_exp.quantization = no_scale_quant |
| 314 | TensorUtil.set_ofm_tensor(reduce_sum_op, sum_of_exp) |
| 315 | |
| 316 | # PASS 5 - CLZ |
| 317 | clz_op = Operation("CLZ", self.op.name + "_clz5") |
| 318 | TensorUtil.add_ifm_tensor(clz_op, sum_of_exp) |
| 319 | headroom_plus_one = Tensor(reduce_sum_shape, DataType.int32, clz_op.name + "_0") |
| 320 | headroom_plus_one.quantization = no_scale_quant |
| 321 | TensorUtil.set_ofm_tensor(clz_op, headroom_plus_one) |
| 322 | |
| 323 | # PASS 6 - Sub |
| 324 | sub6_op = Operation("SubAct", self.op.name + "_sub6") |
| 325 | TensorUtil.add_ifm_tensor(sub6_op, headroom_plus_one) |
| 326 | TensorUtil.add_ifm_tensor( |
| 327 | sub6_op, |
| 328 | TensorUtil.create_const_tensor( |
| 329 | sub6_op.name + "_const", [1, 1, 1, 1], DataType.int32, [31], np.uint32, quantization=no_scale_quant |
| 330 | ), |
| 331 | ) |
| 332 | # TODO: Adding this attribute to reverse the operand order is not ideal |
| 333 | # it should be handled automatically by register_command_stream_generator |
| 334 | # or added as an internal operator. |
| 335 | sub6_op.attrs["reverse_op_order"] = True |
| 336 | reciprocal_right_shift = Tensor(reduce_sum_shape, DataType.int32, sub6_op.name + "_0") |
| 337 | reciprocal_right_shift.quantization = no_scale_quant |
| 338 | TensorUtil.set_ofm_tensor(sub6_op, reciprocal_right_shift) |
| 339 | |
| 340 | # PASS 7 - SHL |
| 341 | shl7_op = Operation("SHL", self.op.name + "_shl7") |
| 342 | TensorUtil.add_ifm_tensor(shl7_op, reciprocal_right_shift) |
| 343 | TensorUtil.add_ifm_tensor( |
| 344 | shl7_op, |
| 345 | TensorUtil.create_const_tensor( |
| 346 | shl7_op.name + "_const", [1, 1, 1, 1], DataType.int32, [1], np.uint32, quantization=no_scale_quant |
| 347 | ), |
| 348 | ) |
| 349 | # TODO: See above |
| 350 | shl7_op.attrs["reverse_op_order"] = True |
| 351 | constant_one = Tensor(reduce_sum_shape, DataType.int32, shl7_op.name + "0") |
| 352 | constant_one.quantization = no_scale_quant |
| 353 | TensorUtil.set_ofm_tensor(shl7_op, constant_one) |
| 354 | |
| 355 | # PASS 8 - Sub |
| 356 | sub8_op = Operation("SubAct", self.op.name + "_sub8") |
| 357 | TensorUtil.add_ifm_tensor(sub8_op, sum_of_exp) |
| 358 | TensorUtil.add_ifm_tensor(sub8_op, constant_one) |
| 359 | sum_of_exps_minus_one = Tensor(reduce_sum_shape, DataType.int32, sub8_op.name + "_0") |
| 360 | sum_of_exps_minus_one.quantization = no_scale_quant |
| 361 | TensorUtil.set_ofm_tensor(sub8_op, sum_of_exps_minus_one) |
| 362 | |
| 363 | # PASS 9 - SHL |
| 364 | shl9_op = Operation("SHL", self.op.name + "_shl9") |
| 365 | TensorUtil.add_ifm_tensor(shl9_op, sum_of_exps_minus_one) |
| 366 | TensorUtil.add_ifm_tensor(shl9_op, headroom_plus_one) |
| 367 | shifted_sum_minus_one = Tensor(reduce_sum_shape, DataType.int32, shl9_op.name + "_0") |
| 368 | shifted_sum_minus_one.quantization = no_scale_quant |
| 369 | TensorUtil.set_ofm_tensor(shl9_op, shifted_sum_minus_one) |
| 370 | |
| 371 | # PASS 10 - SHR |
| 372 | shr10_op = Operation("SHR", self.op.name + "_shr10") |
| 373 | TensorUtil.add_ifm_tensor(shr10_op, shifted_sum_minus_one) |
| 374 | TensorUtil.add_ifm_tensor( |
| 375 | shr10_op, |
| 376 | TensorUtil.create_const_tensor( |
| 377 | shr10_op.name + "_const", [1, 1, 1, 1], DataType.int32, [15], np.uint32, quantization=no_scale_quant |
| 378 | ), |
| 379 | ) |
| 380 | shifted_sum_minus_one_16 = Tensor(reduce_sum_shape, DataType.int32, shr10_op.name + "_0") |
| 381 | shifted_sum_minus_one_16.quantization = shifted_sum_minus_one.quantization.clone() |
| 382 | TensorUtil.set_ofm_tensor(shr10_op, shifted_sum_minus_one_16) |
| 383 | |
| 384 | # PASS 11 - Sub+LUT(one over one plus x) |
| 385 | sub11_op = Operation("SubAct", self.op.name + "_sub11") |
| 386 | TensorUtil.add_ifm_tensor(sub11_op, shifted_sum_minus_one_16) |
| 387 | TensorUtil.add_ifm_tensor( |
| 388 | sub11_op, |
| 389 | TensorUtil.create_const_tensor( |
| 390 | sub11_op.name + "_const", [1, 1, 1, 1], DataType.int32, [32768], np.uint32, quantization=no_scale_quant |
| 391 | ), |
| 392 | ) |
| 393 | sub11_op.set_activation_lut( |
| 394 | TensorUtil.create_const_tensor( |
| 395 | sub11_op.name + "_lut", |
| 396 | [1, 1, 1, 512], |
| 397 | DataType.int32, |
| 398 | self.ONE_OVER_ONE_PLUS_X_LUT, |
| 399 | np.uint32, |
| 400 | TensorPurpose.LUT, |
| 401 | ) |
| 402 | ) |
| 403 | reciprocal_scale = Tensor(reduce_sum_shape, DataType.int16, sub11_op.name + "_0") |
| 404 | reciprocal_scale.quantization = no_scale_quant |
| 405 | TensorUtil.set_ofm_tensor(sub11_op, reciprocal_scale) |
| 406 | |
| 407 | # PASS 12 - Multiply |
| 408 | mul_op = Operation("MulAct", self.op.name + "_mul12") |
| 409 | TensorUtil.add_ifm_tensor(mul_op, exp_ofm) |
| 410 | TensorUtil.add_ifm_tensor(mul_op, reciprocal_scale) |
| 411 | mul_ofm = Tensor(exp_ofm.shape, DataType.int32, mul_op.name + "_0") |
| 412 | mul_ofm.quantization = no_scale_quant |
| 413 | TensorUtil.set_ofm_tensor(mul_op, mul_ofm) |
| 414 | |
| 415 | # PASS 13 - SHR |
| 416 | shr13_op = Operation("SHR", self.op.name + "_shr13") |
| 417 | TensorUtil.add_ifm_tensor(shr13_op, mul_ofm) |
| 418 | TensorUtil.add_ifm_tensor(shr13_op, reciprocal_right_shift) |
| 419 | TensorUtil.set_ofm_tensor(shr13_op, ofm) |
| 420 | |
| 421 | return shr13_op |