Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [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. |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 16 | # Description: |
| 17 | # Early optimisation of the network graph, using the rewrite_graph module to do the traversal of the graph. These are |
| 18 | # split into two parts optimise_graph_a and optimise_graph_b. |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 19 | import math |
Diqing Zhong | 016b827 | 2020-12-16 16:46:06 +0100 | [diff] [blame] | 20 | import uuid |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 21 | |
| 22 | import numpy as np |
| 23 | |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 24 | from . import fp_math |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 25 | from . import lut |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 26 | from . import rewrite_graph |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 27 | from . import scaling |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 28 | from .data_type import DataType |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 29 | from .debug_database import DebugDatabase |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 30 | from .errors import UnsupportedFeatureError |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 31 | from .errors import VelaError |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 32 | from .ethos_u55_regs.ethos_u55_regs import resampling_mode |
Louis Verhaard | 8912c53 | 2020-09-30 12:11:49 +0200 | [diff] [blame] | 33 | from .numeric_util import clamp_sigmoid |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 34 | from .numeric_util import full_shape |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 35 | from .numeric_util import round_away_zero |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 36 | from .operation import create_activation_function |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 37 | from .operation import NpuBlockType |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 38 | from .operation import Op |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 39 | from .operation import Operation |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 40 | from .operation import Padding |
Fredrik Svedberg | d9c2c42 | 2020-12-01 16:33:45 +0100 | [diff] [blame] | 41 | from .operation_util import create_avgpool_nop |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 42 | from .shape4d import Shape4D |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 43 | from .softmax import SoftMax |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 44 | from .tensor import check_quantized_tens_scaling_equal |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 45 | from .tensor import create_const_tensor |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 46 | from .tensor import QuantizationParameters |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 47 | from .tensor import Tensor |
Michael McGeagh | 7a6f843 | 2020-12-02 15:29:22 +0000 | [diff] [blame] | 48 | from .tflite_mapping import optype_to_builtintype |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 49 | |
Michael McGeagh | f3e3ad7 | 2020-12-02 12:39:03 +0000 | [diff] [blame] | 50 | passthrough_nodes = (Op.Identity,) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 51 | |
Michael McGeagh | f3e3ad7 | 2020-12-02 12:39:03 +0000 | [diff] [blame] | 52 | memory_only_ops = (Op.Reshape,) |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 53 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 54 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 55 | def remove_passthrough_tensor(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 56 | if len(tens.ops) == 1 and tens.ops[0].type in passthrough_nodes: |
| 57 | assert len(tens.ops[0].inputs) == 1 |
| 58 | tens = tens.ops[0].inputs[0] |
| 59 | return tens |
| 60 | |
| 61 | |
Patrik Gustavsson | 2c2522d | 2021-01-29 11:51:31 +0100 | [diff] [blame^] | 62 | def rewrite_concat_ops(op, arch): |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 63 | if not op.run_on_npu or not op.type.is_concat_op(): |
| 64 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 65 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 66 | axis_4D = 0 |
| 67 | ofm = op.ofm |
| 68 | ofm.ops = [] |
| 69 | offset = 0 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 70 | |
Patrik Gustavsson | 7bada40 | 2021-01-28 15:46:21 +0100 | [diff] [blame] | 71 | unfuse_activation_function(op) |
| 72 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 73 | if op.type == Op.Pack: |
| 74 | # Pack is also referred to as Stack |
| 75 | axis = int(op.attrs["axis"]) |
| 76 | desired_shape = op.inputs[0].shape[:axis] + [1] + op.inputs[0].shape[axis:] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 77 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 78 | if axis >= 0: |
| 79 | axis_4D = axis + (4 - len(desired_shape)) |
| 80 | else: |
| 81 | axis_4D = axis |
| 82 | |
| 83 | for idx, inp in enumerate(op.inputs): |
| 84 | op.ifm_shapes[idx] = Shape4D(desired_shape) |
| 85 | if Shape4D(inp.shape) != op.ifm_shapes[idx]: |
| 86 | inp.avoid_NHCWB16 = True |
| 87 | op.type = Op.PackReshaped |
| 88 | |
| 89 | inputs, axis = op.get_concat_inputs_axis() |
| 90 | |
| 91 | for idx, inp in enumerate(inputs): |
| 92 | if op.type != Op.PackReshaped: |
| 93 | op.ifm_shapes[idx] = Shape4D(inp.shape) |
Patrik Gustavsson | 3d73717 | 2020-12-22 10:40:51 +0100 | [diff] [blame] | 94 | if axis >= 0: |
| 95 | axis_4D = axis + (4 - len(inp.shape)) |
| 96 | else: |
| 97 | axis_4D = axis |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 98 | new_op = Operation(Op.ConcatSliceWrite, op.name + str(idx)) |
| 99 | new_op.inputs = [inp] |
| 100 | new_op.outputs = [ofm] |
| 101 | new_op.attrs["concat_axis"] = axis_4D |
| 102 | new_op.attrs["concat_start"] = offset |
| 103 | offset += op.ifm_shapes[idx].get_dim(axis_4D) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 104 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 105 | new_op.attrs["concat_end"] = offset |
| 106 | new_op.run_on_npu = True |
| 107 | ofm.ops.append(new_op) |
| 108 | DebugDatabase.add_optimised(op, new_op) |
| 109 | new_op.ifm_shapes.append(op.ifm_shapes[idx].clone()) |
| 110 | new_op.ofm_shapes.append(op.ofm_shapes[0].clone()) |
| 111 | assert ofm.shape[axis] == offset |
Patrik Gustavsson | 458a208 | 2020-08-13 13:41:05 +0200 | [diff] [blame] | 112 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 113 | # If axis corresponds to C-dimension, NHCWB16 can only be used in the output if all the concat_start's are a |
| 114 | # multiple of 16. This as, it is only then the address offset for the ofm, for all operations, will be 16 byte |
| 115 | # aligned. For other values of axis the address offsets will be 16 byte aligned, as they are all based on c = 0 |
| 116 | # and those addresses are always 16 byte aligned due to the NHCWB16 format. |
| 117 | if axis == -1 or axis == (len(ofm.shape) - 1): |
| 118 | for op in ofm.ops: |
| 119 | if op.attrs["concat_start"] % 16 != 0: |
| 120 | ofm.avoid_NHCWB16 = True |
| 121 | break |
| 122 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 123 | |
| 124 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 125 | def rewrite_split_ops(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 126 | |
Patrik Gustavsson | 224e99b | 2021-01-14 10:55:43 +0100 | [diff] [blame] | 127 | if len(tens.ops) == 1 and tens.ops[0].type.is_split_op() and tens.ops[0].type != Op.Unpack: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 128 | split_op = tens.ops[0] |
| 129 | |
| 130 | # Not supported so leave it and run on CPU |
| 131 | if not split_op.run_on_npu: |
| 132 | return tens |
| 133 | |
| 134 | inp, outputs, axis, offset_start, offset_end = split_op.get_split_inputs_axis() |
| 135 | |
| 136 | tens.ops = [] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 137 | new_op = Operation(Op.SplitSliceRead, split_op.name) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 138 | new_op.inputs = [inp] |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 139 | ofm_shape_idx = 0 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 140 | |
| 141 | # For Split the offset cannot be extracted from the tensor so it has to |
| 142 | # be calculated from the index of the output tensor |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 143 | if axis is not None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 144 | # Get the start and end of the split |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 145 | offset_start = [0] * 4 |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 146 | axis_4D_list = split_op.attrs.get("split_axis_4D", None) # Present for UnpackReshaped and some StridedSlice |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 147 | for idx, out in enumerate(outputs): |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 148 | if axis_4D_list is not None: |
| 149 | axis_4D = axis_4D_list[idx] |
Patrik Gustavsson | 3d73717 | 2020-12-22 10:40:51 +0100 | [diff] [blame] | 150 | else: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 151 | split_op.ofm_shapes[idx] = Shape4D(out.shape) |
| 152 | if axis >= 0: |
| 153 | axis_4D = axis + (4 - len(out.shape)) |
| 154 | else: |
| 155 | axis_4D = axis |
| 156 | |
| 157 | if out == tens: |
| 158 | ofm_shape_idx = idx |
| 159 | break |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 160 | |
| 161 | offset_start[axis_4D] += split_op.ofm_shapes[idx].get_dim(axis_4D) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 162 | |
Patrik Gustavsson | eebb1c2 | 2020-08-18 15:03:04 +0200 | [diff] [blame] | 163 | # If start offset is not a multiple of 16 in the C-dimension, NHCWB16 need to be avoided in the input |
| 164 | if (offset_start[-1] % 16) != 0: |
| 165 | inp.avoid_NHCWB16 = True |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 166 | else: |
| 167 | offset_start = full_shape(4, offset_start, 0) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 168 | |
| 169 | new_op.attrs["split_start"] = offset_start |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 170 | new_op.run_on_npu = True |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 171 | new_op.set_output_tensor(tens) |
Patrik Gustavsson | 224e99b | 2021-01-14 10:55:43 +0100 | [diff] [blame] | 172 | new_op.ifm_shapes.append(Shape4D(inp.shape)) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 173 | new_op.ofm_shapes.append(split_op.ofm_shapes[ofm_shape_idx].clone()) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 174 | DebugDatabase.add_optimised(split_op, new_op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 175 | |
| 176 | return tens |
| 177 | |
| 178 | |
| 179 | def needed_total_padding(input_size, stride, filter_size): |
| 180 | out_size = (input_size + stride - 1) // stride |
| 181 | needed_input = (out_size - 1) * stride + filter_size |
| 182 | total_padding = max(0, needed_input - input_size) |
| 183 | return total_padding |
| 184 | |
| 185 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 186 | def calc_padding_and_skirt(padding_type, kernel_size, stride, input_shape, explicit_padding): |
| 187 | ypad = needed_total_padding(int(input_shape.height), int(stride[1]), int(kernel_size[0])) |
| 188 | xpad = needed_total_padding(int(input_shape.width), int(stride[2]), int(kernel_size[1])) |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 189 | if padding_type == Padding.SAME: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 190 | left_pad = (xpad + 0) // 2 |
| 191 | right_pad = (xpad + 1) // 2 |
| 192 | top_pad = (ypad + 0) // 2 |
| 193 | bottom_pad = (ypad + 1) // 2 |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 194 | elif padding_type == Padding.VALID: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 195 | left_pad = 0 |
| 196 | right_pad = 0 |
| 197 | top_pad = 0 |
| 198 | bottom_pad = 0 |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 199 | elif padding_type == Padding.EXPLICIT: |
| 200 | # Padding is specified in a PAD operator which has been bypassed. |
| 201 | # The top and left padding are taken from the PAD; bottom and right are calculated. |
| 202 | top_pad, left_pad, _, _ = explicit_padding |
| 203 | bottom_pad = ypad - top_pad |
| 204 | right_pad = xpad - left_pad |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 205 | else: |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 206 | raise UnsupportedFeatureError(f"Unknown padding") |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 207 | padding = (top_pad, left_pad, bottom_pad, right_pad) |
| 208 | skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad) |
| 209 | return padding, skirt |
| 210 | |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 211 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 212 | def calc_upscaled_padding_and_skirt(padding_type, kernel_size, stride, input_shape, upscaling_factor): |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 213 | kernel_height, kernel_width = kernel_size[0], kernel_size[1] |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 214 | if padding_type == Padding.SAME: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 215 | ypad = needed_total_padding(int(input_shape.height) * upscaling_factor, int(stride[1]), int(kernel_height)) |
| 216 | xpad = needed_total_padding(int(input_shape.width) * upscaling_factor, int(stride[2]), int(kernel_width)) |
Jacob Bohlin | d47cc27 | 2020-08-24 11:42:14 +0200 | [diff] [blame] | 217 | right_pad = max(((xpad + 1) // upscaling_factor) - 1, 0) |
| 218 | bottom_pad = max(((ypad + 1) // upscaling_factor) - 1, 0) |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 219 | left_pad = max(kernel_width - 1 - right_pad, 0) |
| 220 | top_pad = max(kernel_height - 1 - bottom_pad, 0) |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 221 | elif padding_type == Padding.VALID: |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 222 | right_pad = max(kernel_width - 2, 0) |
| 223 | bottom_pad = max(kernel_height - 2, 0) |
| 224 | left_pad = kernel_width - 1 |
| 225 | top_pad = kernel_height - 1 |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 226 | else: |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 227 | raise UnsupportedFeatureError(f"Unknown padding") |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 228 | padding = (top_pad, left_pad, bottom_pad, right_pad) |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 229 | skirt = padding |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 230 | return padding, skirt |
| 231 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 232 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 233 | def fixup_conv2d_backprop(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 234 | if op.type == Op.Conv2DBackpropInput: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 235 | # flip the inputs |
| 236 | op.inputs[0], op.inputs[2] = op.inputs[2], op.inputs[0] |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 237 | op.set_ifm_ofm_shapes() |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 238 | op.type = Op.Conv2DBackpropInputSwitchedBias |
Louis Verhaard | 69b8480 | 2020-12-16 12:02:28 +0100 | [diff] [blame] | 239 | op.ifm.resampling_mode = resampling_mode.TRANSPOSE |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 240 | |
| 241 | # Update strides |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 242 | op.attrs.update({"stride_w": 1, "stride_h": 1, "strides": (1, 1, 1, 1)}) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 243 | |
| 244 | return op |
| 245 | |
| 246 | |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 247 | # Convert the op to an elementwise add |
| 248 | def convert_resizebilinear_1x1_to_add(op): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 249 | op.type = Op.Add |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 250 | op.name = op.name + "_add" |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 251 | op.attrs["resizebilinear"] = True |
| 252 | # Create an input tensor filled with zeros |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 253 | shape = op.ofm_shapes[0].as_list() |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 254 | tens = Tensor(shape, op.inputs[0].dtype, op.inputs[1].name + "_add") |
| 255 | tens.values = np.zeros(shape) |
| 256 | tens.quant_values = np.zeros(shape, np.uint8) |
| 257 | tens.quantization = QuantizationParameters(0.0, 255.0) |
| 258 | tens.quantization.scale_f32 = 1.0 |
| 259 | tens.quantization.zero_point = 0 |
| 260 | tens.consumer_list = [op] |
| 261 | tens_op = op.inputs[1].ops[0] |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 262 | tens_op.set_output_tensor(tens) |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 263 | # Set the add inputs |
| 264 | op.inputs[1] = op.inputs[0] |
| 265 | op.inputs[0] = tens |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 266 | op.set_ifm_ofm_shapes() |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 267 | |
| 268 | return op |
| 269 | |
| 270 | |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 271 | # Convert ResizeBilinear to a number of 2x2 pool ops |
| 272 | def convert_resizebilinear_to_2x2_pool(op): |
| 273 | count = 0 |
| 274 | pre_op = op |
| 275 | outputs = op.outputs |
| 276 | |
| 277 | op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)}) |
| 278 | if op.attrs["align_corners"]: |
| 279 | shape_modifier = 1 |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 280 | op.attrs["padding"] = Padding.VALID |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 281 | else: |
| 282 | shape_modifier = 0 |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 283 | op.attrs["padding"] = Padding.SAME |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 284 | op.inputs[0].resampling_mode = resampling_mode.NEAREST |
| 285 | |
Patrik Gustavsson | 2c2522d | 2021-01-29 11:51:31 +0100 | [diff] [blame^] | 286 | upscaled_shape = np.array(op.ifm_shapes[0].get_hw_as_list()) |
| 287 | out_shape = np.array(op.ofm_shapes[0].get_hw_as_list()) |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 288 | if (upscaled_shape == upscaled_shape * 2 - shape_modifier).all(): |
| 289 | return op |
| 290 | |
| 291 | while (upscaled_shape < out_shape).all(): |
| 292 | if count == 0: |
| 293 | scaled_op = pre_op |
| 294 | else: |
| 295 | scaled_op = op.clone("_{}".format(count)) |
| 296 | scaled_op.inputs[0] = pre_op.outputs[0] |
| 297 | |
| 298 | upscaled_shape = upscaled_shape * 2 - shape_modifier |
| 299 | |
| 300 | if (upscaled_shape == out_shape).all(): |
| 301 | scaled_op.outputs = outputs |
| 302 | scaled_op.outputs[0].ops = [scaled_op] |
| 303 | else: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 304 | shape = op.ofm_shapes[0].as_list() |
| 305 | shape[1:3] = upscaled_shape |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 306 | out_tens = Tensor(shape, DataType.int16, "{}_{}".format(op.outputs[0].name, count)) |
| 307 | out_tens.quantization = op.outputs[0].quantization.clone() |
| 308 | out_tens.quantization.quant_min = np.iinfo(np.int16).min |
| 309 | out_tens.quantization.quant_max = np.iinfo(np.int16).max |
| 310 | scaled_op.set_output_tensor(out_tens) |
| 311 | pre_op = scaled_op |
| 312 | count += 1 |
| 313 | |
| 314 | # Setup the scale value |
| 315 | if scaled_op.inputs[0].dtype.bits == 8 and scaled_op.outputs[0].dtype.bits == 16: |
Fredrik Svedberg | e82be7c | 2021-01-18 15:21:03 +0100 | [diff] [blame] | 316 | scaled_op.rescale = 128 |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 317 | elif scaled_op.inputs[0].dtype.bits == 16 and scaled_op.outputs[0].dtype.bits == 8: |
Fredrik Svedberg | e82be7c | 2021-01-18 15:21:03 +0100 | [diff] [blame] | 318 | scaled_op.rescale = 1 / 128 |
| 319 | else: |
| 320 | scaled_op.rescale = None |
Patrik Gustavsson | cc6915c | 2020-12-22 09:16:50 +0100 | [diff] [blame] | 321 | scaled_op.set_ifm_ofm_shapes() |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 322 | |
| 323 | return op |
| 324 | |
| 325 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 326 | def fixup_resizebilinear(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 327 | if op.type == Op.ResizeBilinear and op.run_on_npu: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 328 | if op.ifm_shapes[0] == op.ofm_shapes[0]: |
Charles Xu | 36ffaf3 | 2020-08-05 15:40:44 +0200 | [diff] [blame] | 329 | # Bypass nop resizebilinear |
| 330 | op.inputs = op.inputs[:1] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 331 | op.type = Op.Identity |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 332 | elif op.ifm_shapes[0].height == 1 and op.ifm_shapes[0].width == 1: |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 333 | convert_resizebilinear_1x1_to_add(op) |
| 334 | else: |
| 335 | convert_resizebilinear_to_2x2_pool(op) |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 336 | |
| 337 | return op |
| 338 | |
| 339 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 340 | def convert_nop_split_to_identity(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 341 | if op.type == Op.Split and op.attrs.get("num_splits") == 1: |
Dwight Lidman | c3862c2 | 2020-09-14 15:22:33 +0200 | [diff] [blame] | 342 | # the list comprehension should return a list with a single tensor |
| 343 | # if it shouldn't, remove_passthrough_tensor will fail appropriately |
| 344 | op.inputs = [i for i in op.inputs if i.shape == op.outputs[0].shape] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 345 | op.type = Op.Identity |
Dwight Lidman | c3862c2 | 2020-09-14 15:22:33 +0200 | [diff] [blame] | 346 | return op |
| 347 | |
| 348 | |
Patrik Gustavsson | 2c2522d | 2021-01-29 11:51:31 +0100 | [diff] [blame^] | 349 | def rewrite_fully_connected_input(op, arch, nng): |
| 350 | if op.type == Op.FullyConnected: |
| 351 | n_in_elems = op.weights.shape[-2] |
| 352 | elms = op.ifm.elements() |
| 353 | batch_size = elms // n_in_elems |
| 354 | assert batch_size * n_in_elems == elms |
| 355 | |
| 356 | if op.ifm.shape != [batch_size, n_in_elems]: |
| 357 | op.ifm.avoid_NHCWB16 = True |
| 358 | |
| 359 | op.ifm_shapes[0] = Shape4D([batch_size, 1, 1, n_in_elems]) |
| 360 | return op |
| 361 | |
| 362 | |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 363 | def convert_batched_fc_shape(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 364 | if op.type == Op.FullyConnected: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 365 | # Check if the first dimension indicates batching |
| 366 | if op.ifm_shapes[0].batch > 1: |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 367 | batching_split = {4: (2, 2), 8: (2, 4), 16: (4, 4)} |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 368 | n = op.ifm_shapes[0].batch |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 369 | h, w = batching_split.get(n, (1, n)) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 370 | op.ifm_shapes[0] = Shape4D([1, h, w, op.ifm_shapes[0].depth]) |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 371 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 372 | op.ifm.avoid_NHCWB16 = True |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 373 | |
| 374 | # Reshape Weights to be 4D. IO becomes HWIO |
| 375 | weight_tensor = op.inputs[1] |
| 376 | weight_tensor.quant_values = np.expand_dims(np.expand_dims(weight_tensor.quant_values, axis=0), axis=0) |
| 377 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
| 378 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 379 | n = op.ofm_shapes[0].batch |
| 380 | h, w = batching_split.get(n, (1, n)) |
| 381 | op.ofm_shapes[0] = Shape4D([1, h, w, op.ofm_shapes[0].depth]) |
| 382 | op.ofm.avoid_NHCWB16 = True |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 383 | return op |
| 384 | |
| 385 | |
Patrik Gustavsson | 7bada40 | 2021-01-28 15:46:21 +0100 | [diff] [blame] | 386 | def unfuse_activation_function(op): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 387 | if op.type == Op.ConcatTFLite and op.run_on_npu and op.activation is not None: |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 388 | act_op = Operation(op.activation.op_type, op.name + op.activation.op_type.name) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 389 | op.activation = None |
Fredrik Svedberg | 0f98b36 | 2020-09-29 10:00:39 +0200 | [diff] [blame] | 390 | out_tens = op.outputs[0] |
| 391 | intermediate_tens = out_tens.clone("_act_intermediate") |
| 392 | act_op.set_output_tensor(out_tens) |
| 393 | act_op.add_input_tensor(intermediate_tens) |
| 394 | op.set_output_tensor(intermediate_tens) |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 395 | act_op.set_ifm_ofm_shapes() |
Fredrik Svedberg | 0f98b36 | 2020-09-29 10:00:39 +0200 | [diff] [blame] | 396 | |
Louis Verhaard | 8912c53 | 2020-09-30 12:11:49 +0200 | [diff] [blame] | 397 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 398 | def rewrite_stridedslice_output(op, arch, nng): |
| 399 | if not op.run_on_npu or op.type != Op.StridedSlice: |
| 400 | return op |
| 401 | |
| 402 | new_axis_mask = op.attrs["new_axis_mask"] |
| 403 | shrink_axis_mask = op.attrs["shrink_axis_mask"] |
| 404 | |
| 405 | if shrink_axis_mask == 0 and new_axis_mask == 0: |
| 406 | return op |
| 407 | |
| 408 | axis_4D = [0] * len(op.outputs) |
| 409 | for idx, out_tens in enumerate(op.outputs): |
| 410 | output_shape = list(out_tens.shape) |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 411 | |
Dwight Lidman | 73320a4 | 2020-11-05 10:34:41 +0100 | [diff] [blame] | 412 | if shrink_axis_mask != 0: |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 413 | n = 0 |
| 414 | axis = 0 |
| 415 | while shrink_axis_mask: |
| 416 | prev_mask = shrink_axis_mask |
| 417 | n += 1 |
| 418 | shrink_axis_mask &= shrink_axis_mask - 1 |
| 419 | axis = int(math.log2(prev_mask - shrink_axis_mask)) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 420 | output_shape = output_shape[:axis] + [1] + output_shape[axis:] |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 421 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 422 | assert len(out_tens.shape) == (len(op.inputs[0].shape) - n) |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 423 | op.attrs["shrink_axis_mask"] = 0 |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 424 | if axis >= 0: |
| 425 | axis_4D[idx] = axis + (4 - len(output_shape)) |
| 426 | else: |
| 427 | axis_4D[idx] = axis |
| 428 | op.ofm_shapes[idx] = Shape4D(output_shape) |
| 429 | |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 430 | elif new_axis_mask != 0: |
| 431 | n = 0 |
| 432 | axis = 0 |
| 433 | while new_axis_mask: |
| 434 | prev_mask = new_axis_mask |
| 435 | n += 1 |
| 436 | new_axis_mask &= new_axis_mask - 1 |
| 437 | axis = int(math.log2(prev_mask - new_axis_mask)) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 438 | output_shape = output_shape[:axis] + output_shape[(axis + 1) :] |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 439 | new_axis_mask >>= 1 |
| 440 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 441 | assert len(out_tens.shape) == (len(op.inputs[0].shape) + n) |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 442 | op.attrs["new_axis_mask"] = 0 |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 443 | if axis >= 0: |
| 444 | axis_4D[idx] = axis + (4 - len(output_shape)) |
| 445 | else: |
| 446 | axis_4D[idx] = axis |
| 447 | op.ofm_shapes[idx] = Shape4D(output_shape) |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 448 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 449 | if op.ofm_shapes[idx] != Shape4D(out_tens.shape): |
| 450 | out_tens.avoid_NHCWB16 = True |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 451 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 452 | op.attrs["split_axis_4D"] = axis_4D |
| 453 | return op |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 454 | |
| 455 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 456 | def rewrite_unpack_output(op, arch, nng): |
| 457 | tens = op.outputs[0] |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 458 | if op.run_on_npu and op.type == Op.Unpack: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 459 | # Unpack is also referred to as Unstack |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 460 | axis = int(op.attrs["axis"]) |
| 461 | op.type = Op.UnpackReshaped |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 462 | desired_output_shape = tens.shape[:axis] + [1] + tens.shape[axis:] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 463 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 464 | if axis >= 0: |
| 465 | axis_4D = axis + (4 - len(desired_output_shape)) |
| 466 | else: |
| 467 | axis_4D = axis |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 468 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 469 | axis_4D_list = [0] * len(op.outputs) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 470 | for idx, out_tens in enumerate(op.outputs): |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 471 | op.ofm_shapes[idx] = Shape4D(desired_output_shape) |
| 472 | axis_4D_list[idx] = axis_4D |
| 473 | if op.ofm_shapes[idx] != Shape4D(out_tens.shape): |
| 474 | out_tens.avoid_NHCWB16 = True |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 475 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 476 | op.attrs["split_axis_4D"] = axis_4D_list |
| 477 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 478 | |
| 479 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 480 | def add_padding_fields(op, arch, nng): |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 481 | if op.run_on_npu: |
| 482 | if "padding" in op.attrs: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 483 | input_shape = op.ifm_shapes[0] |
| 484 | output_shape = op.ofm_shapes[0] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 485 | if op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op(): |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 486 | kernel_size = op.inputs[1].shape[:2] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 487 | elif op.type.is_pool_op() or op.type.npu_block_type == NpuBlockType.ReduceSum: |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 488 | kernel_size = op.attrs["ksize"][1:3] |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 489 | else: |
Michael McGeagh | 7a6f843 | 2020-12-02 15:29:22 +0000 | [diff] [blame] | 490 | raise UnsupportedFeatureError(f"Unknown operation that uses padding: {optype_to_builtintype(op.type)}") |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 491 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 492 | if op.type == Op.Conv2DBackpropInputSwitchedBias: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 493 | upscaling_factor = output_shape.height // input_shape.height |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 494 | padding, skirt = calc_upscaled_padding_and_skirt( |
| 495 | op.attrs["padding"], kernel_size, op.attrs["strides"], input_shape, upscaling_factor |
| 496 | ) |
| 497 | else: |
| 498 | dilation_h, dilation_w = op.get_dilation_h_w() |
| 499 | dilated_kernel_size = [dilation_h * (kernel_size[0] - 1) + 1, dilation_w * (kernel_size[1] - 1) + 1] |
| 500 | padding, skirt = calc_padding_and_skirt( |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 501 | op.attrs["padding"], |
| 502 | dilated_kernel_size, |
| 503 | op.attrs["strides"], |
| 504 | input_shape, |
| 505 | op.attrs.get("explicit_padding"), |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 506 | ) |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 507 | |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 508 | op.attrs["explicit_padding"] = padding |
| 509 | op.attrs["skirt"] = skirt |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 510 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 511 | return op |
| 512 | |
| 513 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 514 | # Check if the op can be reordered |
| 515 | def get_prepend_op(op): |
| 516 | inp = op.inputs[0] |
| 517 | # The op should be reordered between prev_op and prep_op |
| 518 | prev_op = inp.ops[-1] |
| 519 | prep_op = None |
| 520 | while prev_op.type in memory_only_ops and len(prev_op.outputs) == 1 and len(prev_op.outputs[0].consumers()) == 1: |
| 521 | prep_op = prev_op |
| 522 | inp = prev_op.inputs[0] |
| 523 | prev_op = inp.ops[-1] |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 524 | if prev_op is not None and len(prev_op.outputs) == 1 and len(prev_op.outputs[0].consumers()) == 1: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 525 | return prep_op |
| 526 | |
| 527 | return None |
| 528 | |
| 529 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 530 | def convert_depthwise_to_conv(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 531 | # Depthwise is equivalent to a single conv2d if the ifm depth is 1 and |
| 532 | # the ofm depth equals the depth multipler. |
| 533 | # If those conditions are true, then we can perform a simple |
| 534 | # switch of the operator type (and weight order) |
| 535 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 536 | if op.type == Op.DepthwiseConv2DBias and (op.attrs["depth_multiplier"] != 1): |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 537 | ifm_shape = op.ifm_shapes[0] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 538 | weight_tensor = op.inputs[1] |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 539 | ofm_shape = op.ofm_shapes[0] |
| 540 | if (ifm_shape.depth == 1) and (ofm_shape.depth == op.attrs["depth_multiplier"]): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 541 | # Change op type to Conv2d |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 542 | op.type = Op.Conv2DBias |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 543 | del op.attrs["channel_multiplier"] |
| 544 | del op.attrs["depth_multiplier"] |
| 545 | |
| 546 | weight_tensor.quant_values = np.transpose(weight_tensor.quant_values, (0, 1, 3, 2)) |
Michael McGeagh | 6a8d424 | 2020-07-28 12:17:59 +0100 | [diff] [blame] | 547 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 548 | else: |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 549 | raise UnsupportedFeatureError( |
Michael McGeagh | 7a6f843 | 2020-12-02 15:29:22 +0000 | [diff] [blame] | 550 | f"Unsupported 'DEPTHWISE_CONV_2D' with depth_multiplier = {op.attrs['depth_multiplier']},", |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 551 | f" ifm channels = {ifm_shape.depth}, ofm channels = {ofm_shape.depth}", |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 552 | ) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 553 | DebugDatabase.add_optimised(op, op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 554 | return op |
| 555 | |
| 556 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 557 | def reorder_depthwise_weights(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 558 | if op.type.is_depthwise_conv2d_op(): |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 559 | weight_tensor = op.inputs[1] |
| 560 | weight_tensor.quant_values = np.transpose(weight_tensor.quant_values, (0, 1, 3, 2)) |
Michael McGeagh | 6a8d424 | 2020-07-28 12:17:59 +0100 | [diff] [blame] | 561 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 562 | weight_tensor.weight_transpose_depthwise = True |
| 563 | |
| 564 | return op |
| 565 | |
| 566 | |
Diqing Zhong | 016b827 | 2020-12-16 16:46:06 +0100 | [diff] [blame] | 567 | def optimise_strided_conv(op, arch, nng): |
| 568 | stride_x, stride_y = op.get_kernel_stride() |
| 569 | ifm_tensor, _, weight_tensor, _ = op.get_ifm_ifm2_weights_ofm() |
| 570 | |
| 571 | if ( |
| 572 | op.type == Op.Conv2DBias |
| 573 | and op.op_index == 0 |
| 574 | and stride_x == 2 |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 575 | and op.ifm_shapes[0].depth <= 4 |
| 576 | and op.ifm_shapes[0].width % 2 == 0 |
Diqing Zhong | 016b827 | 2020-12-16 16:46:06 +0100 | [diff] [blame] | 577 | and weight_tensor is not None |
| 578 | and weight_tensor.shape[1] >= 2 |
| 579 | ): |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 580 | ifm_shape = op.ifm_shapes[0] |
Diqing Zhong | 016b827 | 2020-12-16 16:46:06 +0100 | [diff] [blame] | 581 | # IFM |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 582 | op.ifm_shapes[0] = Shape4D([ifm_shape.batch, ifm_shape.height, ifm_shape.width // 2, ifm_shape.depth * 2]) |
| 583 | op.ifm.avoid_NHCWB16 = True |
Diqing Zhong | 016b827 | 2020-12-16 16:46:06 +0100 | [diff] [blame] | 584 | |
| 585 | # Weights |
| 586 | weight_shape = weight_tensor.shape |
| 587 | if weight_shape[1] % 2 != 0: |
| 588 | weight_shape[1] = weight_shape[1] + 1 |
| 589 | padded_array = np.zeros(weight_shape) |
| 590 | for i in range(weight_shape[0]): |
| 591 | padded_array[i] = np.vstack( |
| 592 | [ |
| 593 | weight_tensor.quant_values[i], |
| 594 | np.full((1, weight_shape[2], weight_shape[3]), weight_tensor.quantization.zero_point), |
| 595 | ] |
| 596 | ) |
| 597 | weight_tensor.quant_values = padded_array |
| 598 | weight_shape[1] //= 2 |
| 599 | weight_shape[2] *= 2 |
| 600 | weight_tensor.quant_values = np.reshape(weight_tensor.quant_values, weight_shape) |
| 601 | weight_tensor.set_all_shapes(weight_shape) |
| 602 | # If multiple copies of the weights are used, we could avoid |
| 603 | # them having the same address by changing the value_id |
| 604 | weight_tensor.value_id = uuid.uuid4() |
| 605 | |
| 606 | # Strides |
| 607 | stride_x = 1 |
| 608 | op.attrs.update({"stride_w": stride_x, "stride_h": stride_y, "strides": (1, stride_y, stride_x, 1)}) |
| 609 | |
Diqing Zhong | 016b827 | 2020-12-16 16:46:06 +0100 | [diff] [blame] | 610 | return op |
| 611 | |
| 612 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 613 | def convert_conv_to_fc(op, arch, nng): |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 614 | # Conv 1x1 can be equivalent to Fully Connected. |
| 615 | # By representing certain convs as fully connected layers, Vela can better determine wether or not to use |
| 616 | # caching/double buffering for the weights. |
| 617 | # (Weights dont need to be reloaded for convs when IFM H and W are 1) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 618 | if op.type == Op.Conv2DBias: |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 619 | h = op.ifm_shapes[0].height |
| 620 | w = op.ifm_shapes[0].width |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 621 | kh, kw, _, _ = op.inputs[1].shape |
| 622 | if h == 1 and w == 1 and kh == 1 and kw == 1: |
| 623 | # Overwrite this op as a Fully Connected Op |
| 624 | op.name += "_fc" |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 625 | op.type = Op.FullyConnected |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 626 | op.attrs = { |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 627 | "weights_format": 0, |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 628 | } |
| 629 | # Reshape Weights to be 2D. HWIO becomes just IO (as H and W are 1, they can just be dropped) |
| 630 | weight_tensor = op.inputs[1] |
| 631 | weight_tensor.quant_values = weight_tensor.quant_values.squeeze(axis=(0, 1)) |
| 632 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 633 | |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 634 | DebugDatabase.add_optimised(op, op) |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 635 | return op |
| 636 | |
| 637 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 638 | def fixup_relus_with_differing_ifm_ofm_scaling(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 639 | if op.run_on_npu and op.type.is_relu_op(): |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 640 | ifm = op.inputs[0] |
| 641 | ofm = op.outputs[0] |
| 642 | # Relu with differing IFM and OFM scaling cannot be fused with another primary op |
| 643 | # and requires its own to be inserted |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 644 | if not check_quantized_tens_scaling_equal(ifm, ofm): |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 645 | # Override this op with its own primary op (avgpool) |
| 646 | relu_fused_op = create_avgpool_nop(op.name + "_avgpool") |
| 647 | # And fuse the original activation function to it |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 648 | relu_fused_op.activation = create_activation_function(op.type) |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 649 | # Tidy up and assign the ifm and ofm to the new op |
| 650 | ifm.consumer_list.remove(op) |
Andreas Nevalainen | f3d737e | 2020-09-25 14:12:43 +0200 | [diff] [blame] | 651 | |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 652 | relu_fused_op.add_input_tensor(ifm) |
| 653 | relu_fused_op.set_output_tensor(ofm) |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 654 | relu_fused_op.set_ifm_ofm_shapes() |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 655 | op = relu_fused_op |
| 656 | return op |
| 657 | |
| 658 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 659 | # TODO remove if mem only ops can all be removed |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 660 | # Reorder activation op if it's after the memory only operations |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 661 | def fixup_act_reorder(op, arch, nng): |
Michael McGeagh | f3e3ad7 | 2020-12-02 12:39:03 +0000 | [diff] [blame] | 662 | if op.type.is_relu_op() or op.type in (Op.Sigmoid, Op.Tanh): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 663 | prep_op = get_prepend_op(op) |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 664 | if prep_op is not None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 665 | act_op = op.clone("_reordered") |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 666 | act_op.ifm_shapes = list(op.ifm_shapes) |
| 667 | act_op.ofm_shapes = list(op.ofm_shapes) |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 668 | |
| 669 | # There is only one input tensor, overwrite it |
| 670 | act_op.set_input_tensor(prep_op.inputs[0], 0) |
| 671 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 672 | act_op_out = act_op.inputs[0].clone("_acted") |
| 673 | act_op_out.quantization = op.outputs[0].quantization.clone() |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 674 | act_op.set_output_tensor(act_op_out) |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 675 | act_op.ofm_shapes[0] = act_op.ifm_shapes[0].clone() |
| 676 | act_op.ifm_shapes[0] = prep_op.ifm_shapes[0].clone() |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 677 | |
| 678 | # Update the consumer list |
| 679 | act_op_out.consumer_list = op.outputs[0].consumer_list.copy() |
| 680 | act_op_out.consumer_list.append(prep_op) |
| 681 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 682 | prep_op.inputs[0] = act_op_out |
| 683 | prep_op.outputs[0].quantization = act_op_out.quantization.clone() |
| 684 | |
| 685 | # Mark the op so that it will be removed as passthrough later on |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 686 | op.type = Op.Identity |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 687 | |
| 688 | # Record optimisation in debug database |
| 689 | DebugDatabase.add_optimised(op, act_op) |
| 690 | DebugDatabase.add_optimised(op, op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 691 | return op |
| 692 | |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 693 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 694 | def fixup_elementwise_with_scalars(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 695 | if op.type.is_binary_elementwise_op(): |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 696 | ifm_tensor, ifm2_tensor, _, _ = op.get_ifm_ifm2_weights_ofm() |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 697 | if ifm2_tensor.shape != [] and ifm_tensor.shape != []: |
| 698 | diff = len(ifm_tensor.shape) - len(ifm2_tensor.shape) |
| 699 | if diff > 0: |
| 700 | ifm2_tensor.shape = full_shape(len(ifm_tensor.shape), ifm2_tensor.shape, 1) |
| 701 | elif diff < 0: |
| 702 | ifm_tensor.shape = full_shape(len(ifm2_tensor.shape), ifm_tensor.shape, 1) |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 703 | elif ifm_tensor.shape == [] and ifm_tensor.quant_values is None: |
| 704 | # IFM is marked as a scalar, but is a result of an operation; change it to a shape of size 1 |
| 705 | ifm_tensor.shape = len(ifm2_tensor.shape) * [1] |
| 706 | ifm_tensor.storage_shape = ifm_tensor.shape |
| 707 | elif ifm2_tensor.shape == [] and ifm2_tensor.quant_values is None: |
| 708 | # IFM2 is marked as a scalar, but is a result of an operation; change it to a shape of size 1 |
| 709 | ifm2_tensor.shape = len(ifm_tensor.shape) * [1] |
| 710 | ifm2_tensor.storage_shape = ifm2_tensor.shape |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 711 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 712 | |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 713 | |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 714 | # Set input/output tensor equivalence to the same id for memory operations |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 715 | def set_tensor_equivalence(op, arch, nng): |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 716 | if op.type in memory_only_ops: |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 717 | eid = op.outputs[0].equivalence_id |
| 718 | for inp in op.inputs: |
| 719 | inp.equivalence_id = eid |
| 720 | return op |
| 721 | |
| 722 | |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 723 | def set_ifm_ofm_op_shapes(op, arch, nng): |
| 724 | if op.run_on_npu and op.type.needs_shapes(): |
| 725 | if op.ifm_shapes or op.ofm_shapes: |
| 726 | # Shapes already set |
| 727 | return op |
| 728 | op.set_ifm_ofm_shapes() |
| 729 | return op |
| 730 | |
| 731 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 732 | def convert_softmax(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 733 | if op.type == Op.Softmax and op.run_on_npu: |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 734 | softmax = SoftMax(op) |
| 735 | op = softmax.get_graph() |
| 736 | return op |
| 737 | |
| 738 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 739 | def convert_mul_max_to_abs_or_lrelu(op, arch, nng): |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 740 | r"""Whenever there is a subgraph with this topology: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 741 | |
| 742 | Input X For X = -1 or X > 0 |
| 743 | | \ / This subgraph can be replaced with either |
| 744 | | Mul an Abs (if X = -1) or a LeakyReLU (if X > 0) |
| 745 | | / |
| 746 | Max |
| 747 | """ |
| 748 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 749 | if op.type == Op.Maximum: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 750 | # finds the Mul input(s) to the Max |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 751 | muls = [i for i in op.inputs if i.ops[0].type == Op.Mul] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 752 | if len(muls) == 1: |
| 753 | mul = muls[0].ops[0] |
| 754 | elif len(muls) == 2: |
| 755 | # In the case both inputs are Muls, find the one with the same input as the Max |
| 756 | mul = [m for m in muls if len(set(op.inputs + m.ops[0].inputs)) == 1][0].ops[0] |
| 757 | else: |
| 758 | # No Mul inputs |
| 759 | return op |
| 760 | |
| 761 | # make sure the Mul doesn't have any other consumers |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 762 | mul_ofm = mul.outputs[0] |
| 763 | if len(mul_ofm.consumers()) != 1: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 764 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 765 | # make sure the Mul doesn't have a fused activation function |
| 766 | if mul.activation: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 767 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 768 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 769 | if ifm is None or ofm is None: |
| 770 | return op |
| 771 | |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 772 | if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype: |
| 773 | return op |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 774 | if not check_quantized_tens_scaling_equal(ifm, ofm) or not check_quantized_tens_scaling_equal(ifm, mul_ofm): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 775 | # rewrite to LeakyRelu currently only makes sense if the quantization is identical |
| 776 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 777 | |
| 778 | # finds the branched input that goes to both the Max and the Mul |
| 779 | shared = set(op.inputs) & set(mul.inputs) |
| 780 | if len(shared) == 1: |
| 781 | shared_in = shared.pop() |
| 782 | # find the constant scalar input to the Mul |
| 783 | const_tens = (set(mul.inputs) - {shared_in}).pop() |
| 784 | # check that it is a scalar |
| 785 | if const_tens.shape != []: |
| 786 | return op |
| 787 | const = const_tens.ops[0] |
| 788 | # check that it is a constant |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 789 | if const.type != Op.Const: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 790 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 791 | # Remove the Mul from the shared input's consumers |
| 792 | shared_in.consumer_list.remove(mul) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 793 | else: |
| 794 | return op |
| 795 | |
| 796 | val = const.outputs[0].values |
| 797 | if val >= 0: |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 798 | new_op = Op.LeakyRelu |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 799 | op.attrs["alpha"] = val |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 800 | # to produce bit exact results, the alpha is not enough; |
| 801 | # save additional scaling info in attr "alpha_scale", to be used as input |
| 802 | # to the LUT construction |
| 803 | alpha_scalar = const_tens.quant_values - const_tens.quantization.zero_point |
| 804 | mul_ifm_scale = np.double(ifm.quantization.scale_f32) |
| 805 | mul_ifm2_scale = np.double(const_tens.quantization.scale_f32) |
| 806 | mul_ofm_scale = np.double(mul_ofm.quantization.scale_f32) |
| 807 | alpha_scale, alpha_shift = scaling.elementwise_mul_scale(mul_ifm_scale, mul_ifm2_scale, mul_ofm_scale) |
| 808 | op.attrs["alpha_scaling"] = (alpha_scalar, alpha_scale, alpha_shift) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 809 | elif val == -1: |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 810 | new_op = Op.Abs |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 811 | else: |
| 812 | return op |
| 813 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 814 | op.type = new_op |
| 815 | op.name = op.name.replace("Maximum", new_op.name) |
| 816 | op.outputs[0].name = op.outputs[0].name.replace("Maximum", new_op.name) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 817 | op.inputs = [shared_in] |
Patrik Gustavsson | c509d33 | 2020-12-22 13:53:52 +0100 | [diff] [blame] | 818 | op.set_ifm_ofm_shapes() |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 819 | |
| 820 | # Record optimisation in debug database |
| 821 | DebugDatabase.add_optimised(op, op) |
| 822 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 823 | return op |
| 824 | |
| 825 | |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 826 | def convert_lrelu_to_mul_max(op, arch): |
| 827 | # Converts LeakyRelu to Max(alpha * IFM, identity * IFM) |
| 828 | # (the opposite of convert_mul_max_to_abs_or_lrelu) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 829 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 830 | if ifm is None or ofm is None: |
| 831 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 832 | |
| 833 | # Add multiplication with alpha |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 834 | mul_alpha = Operation(Op.Mul, op.name + "_mul_alpha") |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 835 | mul_alpha.add_input_tensor(ifm) |
| 836 | # Create const tensor containing alpha as scalar |
| 837 | alpha = op.attrs["alpha"] |
| 838 | quantization = ifm.quantization.clone() |
| 839 | quantization.min = 0 |
| 840 | quantization.max = alpha * (quantization.quant_max - quantization.quant_min) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 841 | quantization.zero_point = 0 |
Louis Verhaard | ece4e65 | 2021-01-07 13:35:47 +0100 | [diff] [blame] | 842 | if np.isinf(1 / np.float32(alpha)): |
| 843 | # Handling of alpha near zero |
| 844 | quantization.scale_f32 = 1 |
| 845 | scalar = 0 |
| 846 | else: |
| 847 | quantization.scale_f32 = alpha |
| 848 | scalar = 1 |
| 849 | alpha_tens = create_const_tensor( |
| 850 | op.name + "_alpha_scalar", [], ifm.dtype, [scalar], np.int8, quantization=quantization |
| 851 | ) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 852 | mul_alpha.add_input_tensor(alpha_tens) |
| 853 | fm_alpha = ofm.clone(op.name + "_alpha") |
| 854 | mul_alpha.set_output_tensor(fm_alpha) |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 855 | mul_alpha.set_ifm_ofm_shapes() |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 856 | DebugDatabase.add_optimised(op, mul_alpha) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 857 | |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 858 | if check_quantized_tens_scaling_equal(ifm, ofm): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 859 | # No identity multiplication is needed |
| 860 | fm_id = ifm |
| 861 | else: |
| 862 | # Add multiplication with identity |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 863 | mul_identity = Operation(Op.Mul, op.name + "_mul_identity") |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 864 | mul_identity.add_input_tensor(ifm) |
| 865 | # Create const tensor containing identity as scalar |
| 866 | quantization = ifm.quantization.clone() |
| 867 | quantization.min = 0 |
| 868 | quantization.max = quantization.quant_max - quantization.quant_min |
| 869 | quantization.scale_f32 = 1 |
| 870 | quantization.zero_point = 0 |
| 871 | identity_tens = create_const_tensor( |
| 872 | op.name + "_id_scalar", [], ifm.dtype, [1], np.uint8, quantization=quantization |
| 873 | ) |
| 874 | mul_identity.add_input_tensor(identity_tens) |
Louis Verhaard | ece4e65 | 2021-01-07 13:35:47 +0100 | [diff] [blame] | 875 | # Make sure that fm_id is allocated to a different address than fm_alpha |
| 876 | fm_id = ofm.clone(op.name + "_id", set_unique=True) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 877 | mul_identity.set_output_tensor(fm_id) |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 878 | mul_identity.set_ifm_ofm_shapes() |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 879 | DebugDatabase.add_optimised(op, mul_identity) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 880 | |
| 881 | # Convert LeakyRelu to Max, add the results of the multiplication(s) as inputs |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 882 | op.type = Op.Maximum |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 883 | op.name = op.name.replace("LeakyRelu", "Maximum") |
| 884 | op.inputs = [] |
| 885 | ifm.consumer_list.remove(op) |
| 886 | op.add_input_tensor(fm_alpha) |
| 887 | op.add_input_tensor(fm_id) |
Patrik Gustavsson | c509d33 | 2020-12-22 13:53:52 +0100 | [diff] [blame] | 888 | op.set_ifm_ofm_shapes() |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 889 | |
| 890 | DebugDatabase.add_optimised(op, op) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 891 | return op |
| 892 | |
| 893 | |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 894 | def convert_to_lut(op, lut_values, lut_name): |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 895 | # Rewrite the operation by Add with scalar 0 + LUT activation |
| 896 | ifm = op.inputs[0] |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 897 | if ifm is None: |
| 898 | return op |
Louis Verhaard | 58520b9 | 2020-08-24 16:45:38 +0200 | [diff] [blame] | 899 | assert ifm.dtype.size_in_bytes() == 1 |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 900 | op.type = Op.Add |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 901 | op.name = op.name + "_lut_" + lut_name |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 902 | # Mark as no-op to enable potential fusing optimizations |
| 903 | op.attrs["is_nop"] = True |
| 904 | # Create an input tensor containing scalar zero |
| 905 | quantization = QuantizationParameters(0.0, 255.0) |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 906 | quantization.scale_f32 = ifm.quantization.scale_f32 |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 907 | quantization.zero_point = 0 |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 908 | tens = create_const_tensor(op.inputs[0].name + "_scalar0", [], ifm.dtype, [0], np.uint8, quantization=quantization) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 909 | op.add_input_tensor(tens) |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 910 | op.ifm_shapes.append(Shape4D(tens.shape)) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 911 | |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 912 | # The LUT must be applied without any preceding rescaling (the LUT itself performs the rescale), |
| 913 | # so even if the OFM has a different scale than the IFM, the generated OFM scale instructions |
| 914 | # should be the same as the IFM |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 915 | op.forced_output_quantization = ifm.quantization |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 916 | lut_tensor = lut.create_lut_tensor(op.name + "_values", lut_values, DataType.int8) |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 917 | op.set_activation_lut(lut_tensor) |
Patrik Gustavsson | c509d33 | 2020-12-22 13:53:52 +0100 | [diff] [blame] | 918 | op.set_ifm_ofm_shapes() |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 919 | return op |
| 920 | |
| 921 | |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 922 | def convert_to_lut8(op, fn, fn_name): |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 923 | # Converts op to a no-op + int8/uint8 LUT which is generated with the given function. |
| 924 | # fn is a function(real) -> real |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 925 | ifm, ofm = op.get_ifm_ofm() |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 926 | if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype: |
| 927 | return op |
| 928 | # Generate the LUT |
| 929 | ifm_scale = np.double(ifm.quantization.scale_f32) |
| 930 | ofm_scale = np.double(ofm.quantization.scale_f32) |
| 931 | zp_in = ifm.quantization.zero_point |
| 932 | zp_out = ofm.quantization.zero_point |
| 933 | values = [] |
| 934 | ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128) |
| 935 | quantized_min = min(ix) |
| 936 | quantized_max = max(ix) |
| 937 | for x in ix: |
| 938 | x_real = ifm_scale * (x - zp_in) |
| 939 | y_real = fn(x_real) |
| 940 | lut_result = round_away_zero(zp_out + y_real / ofm_scale) |
| 941 | lut_result = min(quantized_max, max(quantized_min, lut_result)) |
| 942 | values.append(lut_result) |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 943 | return convert_to_lut(op, values, fn_name) |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 944 | |
| 945 | |
| 946 | def convert_lrelu_to_lut(op, arch): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 947 | ifm, ofm = op.get_ifm_ofm() |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 948 | # Generate the LUT |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 949 | alpha = op.attrs["alpha"] |
| 950 | ifm_scale = np.double(ifm.quantization.scale_f32) |
| 951 | ofm_scale = np.double(ofm.quantization.scale_f32) |
| 952 | zp_in = ifm.quantization.zero_point |
| 953 | zp_out = ofm.quantization.zero_point |
| 954 | identity_scale, identity_shift = scaling.elementwise_mul_scale(ifm_scale, 1, ofm_scale) |
| 955 | alpha_scalar = 1 |
| 956 | alpha_scale, alpha_shift = scaling.elementwise_mul_scale(ifm_scale, alpha, ofm_scale) |
| 957 | if "alpha_scaling" in op.attrs: |
| 958 | # The LeakyRelu was the result from convert_mul_max_to_abs_or_lrelu |
| 959 | alpha_scalar, alpha_scale, alpha_shift = op.attrs["alpha_scaling"] |
| 960 | values = [] |
Louis Verhaard | 58520b9 | 2020-08-24 16:45:38 +0200 | [diff] [blame] | 961 | ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128) |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 962 | quantized_min = min(ix) |
| 963 | quantized_max = max(ix) |
| 964 | for x in ix: |
| 965 | if x < zp_in: |
| 966 | lut_result = zp_out + fp_math.multiply_by_quantized_multiplier( |
| 967 | alpha_scalar * (x - zp_in), alpha_scale, alpha_shift |
| 968 | ) |
| 969 | else: |
| 970 | lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(x - zp_in, identity_scale, identity_shift) |
| 971 | lut_result = min(quantized_max, max(quantized_min, lut_result)) |
| 972 | values.append(lut_result) |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 973 | return convert_to_lut(op, values, "lrelu") |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 974 | |
| 975 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 976 | def convert_lrelu(op, arch, nng): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 977 | # Converts LeakyRelu to a LUT based solution if possible, otherwise a mul + max |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 978 | if op.type != Op.LeakyRelu: |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 979 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 980 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 981 | if ifm is None or ofm is None: |
| 982 | return op |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 983 | if ifm.dtype in (DataType.uint8, DataType.int8) and ifm.dtype == ofm.dtype: |
| 984 | # use LUT for int8/uint8 |
| 985 | return convert_lrelu_to_lut(op, arch) |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 986 | if check_quantized_tens_scaling_equal(ifm, ofm) and ifm.dtype == ofm.dtype == DataType.int16: |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 987 | # use LeakyRelu unmodified for int16 with equal input/output scaling |
| 988 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 989 | return convert_lrelu_to_mul_max(op, arch) |
| 990 | |
| 991 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 992 | def convert_tanh_sigmoid_to_lut(op, arch, nng): |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 993 | # Converts int8/uint8 Sigmoid and Tanh to a LUT based solution |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 994 | if op.type == Op.Sigmoid: |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 995 | return convert_to_lut8(op, clamp_sigmoid, "sigmoid") |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 996 | elif op.type == Op.Tanh: |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 997 | return convert_to_lut8(op, math.tanh, "tanh") |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 998 | return op |
| 999 | |
| 1000 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1001 | def remove_reshapes(op, arch): |
| 1002 | if op.run_on_npu and op.type == Op.Reshape: |
| 1003 | ofm = op.ofm |
| 1004 | ifm = op.ifm |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 1005 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1006 | # Check if quantization is the same in the input and output for the reshape ops |
| 1007 | if not check_quantized_tens_scaling_equal(ifm, ofm): |
| 1008 | # TODO Both tensors are needed, since quantisation properties currently are linked to Tensors. |
| 1009 | # In order to remove this reshape either quantization properties need to be moved to Operator, |
| 1010 | # or the reshape need to be replace with a NOP. |
| 1011 | return |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 1012 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1013 | # Check if ifm is a sg input |
| 1014 | if ifm.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const): |
| 1015 | # put the reshape on CPU |
| 1016 | op.run_on_npu = False |
| 1017 | return |
| 1018 | |
| 1019 | # Check if Reshape ifm/ofm are network ifm/ofm |
| 1020 | ifm_is_sg_ofm = any(ifm_cons is None for ifm_cons in ifm.consumer_list) |
| 1021 | ofm_is_sg_ofm = any(ofm_cons is None for ofm_cons in ofm.consumer_list) |
| 1022 | |
| 1023 | if ifm_is_sg_ofm and ofm_is_sg_ofm: |
| 1024 | # Both ifm and ofm are sg outputs,add reshape to the ifm and put it on CPU |
| 1025 | ifm_cons_list_copy = ifm.consumer_list.copy() |
| 1026 | ifm_ops_copy = ifm.ops.copy() |
| 1027 | for ifm_cons in ifm_cons_list_copy: |
| 1028 | if ifm_cons is None: |
| 1029 | # Create a reshape op with ifm as output |
| 1030 | name = ifm.name + "_cpu_reshape" |
| 1031 | reshape_ifm = ifm.clone() |
| 1032 | reshape_op = Operation(Op.Reshape, name) |
| 1033 | reshape_op.attrs["new_shape"] = ifm.shape |
| 1034 | reshape_op.add_input_tensor(reshape_ifm) |
| 1035 | reshape_op.add_input_tensor(create_const_tensor(name + "_shape", [1], DataType.int32, ifm.shape)) |
| 1036 | reshape_op.set_output_tensor(ifm) |
| 1037 | reshape_op.set_ifm_ofm_shapes() |
| 1038 | reshape_op.run_on_npu = False |
| 1039 | reshape_op.ofm.ops = [reshape_op] |
| 1040 | reshape_op.ofm.consumer_list = [None] |
| 1041 | |
| 1042 | # Set reshape_ifm producers |
| 1043 | for prev_op in ifm_ops_copy: |
| 1044 | prev_op.outputs = [reshape_ifm] |
| 1045 | reshape_ifm.ops.append(prev_op) |
| 1046 | |
| 1047 | # Set reshape_ifm consumers |
| 1048 | for ifm_cons in ifm_cons_list_copy: |
| 1049 | if ifm_cons is not None: |
| 1050 | for ifm_idx, cons_ifm in enumerate(ifm_cons.inputs): |
| 1051 | if cons_ifm == ifm: |
| 1052 | ifm_cons.set_input_tensor(reshape_ifm, ifm_idx) |
| 1053 | |
| 1054 | ifm = reshape_ifm |
| 1055 | break |
| 1056 | ifm_is_sg_ofm = False |
| 1057 | |
| 1058 | if ofm_is_sg_ofm: |
| 1059 | # Bypassed by replacing ifm with ofm |
| 1060 | ofm.ops = [] |
| 1061 | for prev_op in ifm.ops: |
| 1062 | prev_op.outputs = [ofm] |
| 1063 | ofm.ops.append(prev_op) |
| 1064 | |
| 1065 | # All ifm consumers need to use ofm as input |
| 1066 | for ifm_cons in ifm.consumer_list: |
| 1067 | for ifm_idx, cons_ifm in enumerate(ifm_cons.inputs): |
| 1068 | if cons_ifm == ifm: |
| 1069 | ifm_cons.set_input_tensor(ofm, ifm_idx) |
| 1070 | if op.ifm_shapes[0] != op.ofm_shapes[0]: |
| 1071 | ofm.avoid_NHCWB16 = True |
| 1072 | else: |
| 1073 | # Bypassed Reshape by replacing ofm with ifm |
| 1074 | for cons in ofm.consumer_list: |
| 1075 | for ifm_idx, cons_ifm in enumerate(cons.inputs): |
| 1076 | if cons_ifm == ofm: |
| 1077 | cons.set_input_tensor(ifm, ifm_idx) |
| 1078 | if op.ifm_shapes[0] != op.ofm_shapes[0]: |
| 1079 | ifm.avoid_NHCWB16 = True |
| 1080 | |
| 1081 | |
| 1082 | def check_reshapes(op, arch): |
| 1083 | if op.run_on_npu and op.type == Op.Reshape: |
| 1084 | ofm = op.ofm |
| 1085 | |
| 1086 | if check_quantized_tens_scaling_equal(op.ifm, ofm): |
| 1087 | # Reshape should have been removed |
| 1088 | raise VelaError(f"Reshape op {op} expected to have been removed, still remains") |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 1089 | |
| 1090 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1091 | def fuse_activation_function_with_prev(op, arch, nng): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1092 | # if op is a no-op: attempts to move the activation function to the preceding op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1093 | if not op.attrs.get("is_nop", False) or op.activation is None: |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1094 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1095 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 1096 | if ifm is None or ofm is None: |
| 1097 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1098 | # finds the input(s) to the operation |
| 1099 | prev_op = ifm.ops[0] |
| 1100 | # Note: the below checks on prev_op require that a first optimize pass on the full graph has been performed |
| 1101 | fuse = ( |
| 1102 | prev_op.run_on_npu |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1103 | and prev_op.type.npu_block_type != NpuBlockType.Default |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1104 | and len(ifm.ops) == 1 |
| 1105 | and len(prev_op.outputs[0].consumers()) == 1 |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1106 | and prev_op.activation is None |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1107 | ) |
| 1108 | if op.activation_lut is not None and arch.shram_reserved_unused_banks == 0: |
| 1109 | # TODO: if SHRAM LUT space is shared with SHRAM ACC (32, 64 MAC), |
| 1110 | # LUT currently only works correctly for elementwise ops |
| 1111 | fuse = False |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1112 | if not fuse: |
| 1113 | return op |
| 1114 | # Move the fused activation function + corresponding info to prev_op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1115 | prev_op.activation = op.activation |
| 1116 | prev_op.forced_output_quantization = op.forced_output_quantization |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1117 | if op.activation_lut is not None: |
| 1118 | prev_op.set_activation_lut(op.activation_lut) |
| 1119 | # Bypass op |
Louis Verhaard | 98a3499 | 2020-09-01 10:39:04 +0200 | [diff] [blame] | 1120 | prev_op.set_output_tensor(ofm) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 1121 | DebugDatabase.add_optimised(op, prev_op) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1122 | return op |
| 1123 | |
| 1124 | |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 1125 | def optimise_pad(op, arch, nng): |
| 1126 | """ |
| 1127 | Converts tens1 -> PAD -> tens2 -> CONV to tens1 -> CONV |
| 1128 | if both operations can be run on the NPU. |
| 1129 | """ |
| 1130 | if ( |
| 1131 | (op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op()) |
| 1132 | and op.run_on_npu |
| 1133 | and op.attrs["padding"] == Padding.VALID |
| 1134 | ): |
| 1135 | pad_op = op.ifm.ops[0] |
| 1136 | if pad_op.type != Op.Pad or not pad_op.run_on_npu: |
| 1137 | return op |
| 1138 | # Bypass the PAD operator |
| 1139 | op.set_input_tensor(pad_op.ifm, 0) |
| 1140 | # Adjust the padding attributes of the convolution operator |
| 1141 | op.attrs["padding"] = Padding.EXPLICIT |
| 1142 | padding = pad_op.inputs[1].values # 4x2 tensor, first dimension is N, H, W, C |
| 1143 | top, left, bottom, right = (padding[1][0], padding[2][0], padding[1][1], padding[2][1]) |
| 1144 | op.attrs["explicit_padding"] = (top, left, bottom, right) |
| 1145 | op.set_ifm_ofm_shapes() |
| 1146 | return op |
| 1147 | |
| 1148 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1149 | def add_attrs_to_resizebilinear(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1150 | if op.type == Op.ResizeBilinear and op.run_on_npu: |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1151 | input_tensor = op.inputs[0] |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1152 | input_shape = op.ifm_shapes[0] |
| 1153 | upscaled_height = input_shape.height * 2 |
| 1154 | upscaled_width = input_shape.width * 2 |
| 1155 | out_shape = op.ofm_shapes[0] |
| 1156 | if not op.attrs["align_corners"] and out_shape.height == upscaled_height and out_shape.width == upscaled_width: |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1157 | # this means the output is supposed to be a x2 upscale, |
| 1158 | # so we need to do SAME padding |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 1159 | op.attrs["padding"] = Padding.SAME |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1160 | elif ( |
| 1161 | op.attrs["align_corners"] |
| 1162 | and out_shape.height == (upscaled_height - 1) |
| 1163 | and out_shape.width == (upscaled_width - 1) |
| 1164 | ): |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1165 | # here we can just run the avg pool without padding and |
| 1166 | # produce a (M * 2 - 1, N * 2 - 1) sized output |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 1167 | op.attrs["padding"] = Padding.VALID |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1168 | else: |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 1169 | return op |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1170 | input_tensor.resampling_mode = resampling_mode.NEAREST |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 1171 | op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)}) |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1172 | return op |
| 1173 | |
| 1174 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1175 | def fixup_bias_tensors(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1176 | if op.type.needs_bias() and op.bias is None: |
Jacob Bohlin | a41cd4d | 2020-08-26 18:21:28 +0200 | [diff] [blame] | 1177 | # Op has no bias, add bias tensor filled with zeros |
| 1178 | nr_biases = op.inputs[1].shape[-1] |
| 1179 | bias_values = [0] * nr_biases |
| 1180 | bias_tensor = create_const_tensor(op.name + "_bias", [nr_biases], DataType.int32, bias_values) |
| 1181 | bias_tensor.quant_values = bias_tensor.values |
| 1182 | op.set_input_tensor(bias_tensor, -1) |
Jacob Bohlin | 67e0d8f | 2020-08-20 10:53:02 +0200 | [diff] [blame] | 1183 | |
| 1184 | return op |
| 1185 | |
| 1186 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1187 | def supported_operator_check(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1188 | op.run_on_npu = arch.supported_operators.is_operator_supported(op) |
| 1189 | return op |
| 1190 | |
| 1191 | |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 1192 | def _record_optimised(op, arch): |
| 1193 | if op.type != Op.Const: |
| 1194 | DebugDatabase.add_optimised(op, op) |
| 1195 | |
| 1196 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1197 | def optimise_graph_a(nng, arch, verbose_graph=False): |
| 1198 | if verbose_graph: |
| 1199 | nng.print_graph() |
| 1200 | |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 1201 | pre_process_list = [ |
| 1202 | supported_operator_check, |
| 1203 | set_ifm_ofm_op_shapes, |
| 1204 | # TODO: memory-only Op removal |
| 1205 | ] |
| 1206 | |
| 1207 | for idx, sg in enumerate(nng.subgraphs): |
| 1208 | # rewrite graph pass |
| 1209 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 1210 | nng, sg, arch, [], pre_process_list, rewrite_unsupported=False, |
| 1211 | ) |
| 1212 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1213 | # Handle Concat Ops |
| 1214 | for idx, sg in enumerate(nng.subgraphs): |
| 1215 | # rewrite graph pass |
Patrik Gustavsson | 2c2522d | 2021-01-29 11:51:31 +0100 | [diff] [blame^] | 1216 | rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [rewrite_concat_ops]) |
| 1217 | sg.refresh_after_modification() |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1218 | |
| 1219 | # Handle Split Ops |
| 1220 | for idx, sg in enumerate(nng.subgraphs): |
| 1221 | # rewrite graph pass |
| 1222 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 1223 | nng, |
| 1224 | sg, |
| 1225 | arch, |
| 1226 | [], |
| 1227 | [rewrite_unpack_output, rewrite_stridedslice_output, convert_nop_split_to_identity], |
| 1228 | rewrite_unsupported=False, |
| 1229 | ) |
| 1230 | |
| 1231 | for idx, sg in enumerate(nng.subgraphs): |
| 1232 | # rewrite graph pass |
| 1233 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 1234 | nng, sg, arch, [rewrite_split_ops], [], rewrite_unsupported=False, |
| 1235 | ) |
| 1236 | |
| 1237 | # Removal of reshapes |
| 1238 | for sg in nng.subgraphs: |
| 1239 | rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [remove_reshapes]) |
| 1240 | sg.refresh_after_modification() |
| 1241 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1242 | op_rewrite_list = [ |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 1243 | set_tensor_equivalence, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1244 | convert_depthwise_to_conv, |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 1245 | convert_conv_to_fc, |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 1246 | convert_softmax, |
Diqing Zhong | 016b827 | 2020-12-16 16:46:06 +0100 | [diff] [blame] | 1247 | optimise_strided_conv, |
Patrik Gustavsson | 2c2522d | 2021-01-29 11:51:31 +0100 | [diff] [blame^] | 1248 | rewrite_fully_connected_input, |
Diqing Zhong | 94457b1 | 2020-12-09 15:22:40 +0100 | [diff] [blame] | 1249 | convert_batched_fc_shape, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1250 | fixup_conv2d_backprop, |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 1251 | fixup_relus_with_differing_ifm_ofm_scaling, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1252 | fixup_act_reorder, |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1253 | fixup_elementwise_with_scalars, # TODO Move to early stage? |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 1254 | reorder_depthwise_weights, |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 1255 | fixup_resizebilinear, |
Jacob Bohlin | a41cd4d | 2020-08-26 18:21:28 +0200 | [diff] [blame] | 1256 | fixup_bias_tensors, |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1257 | convert_mul_max_to_abs_or_lrelu, |
| 1258 | convert_lrelu, |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 1259 | convert_tanh_sigmoid_to_lut, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1260 | ] |
| 1261 | |
| 1262 | for idx, sg in enumerate(nng.subgraphs): |
| 1263 | # rewrite graph pass |
| 1264 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Dwight Lidman | 73320a4 | 2020-11-05 10:34:41 +0100 | [diff] [blame] | 1265 | nng, sg, arch, [], op_rewrite_list, rewrite_unsupported=False, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1266 | ) |
| 1267 | |
| 1268 | for idx, sg in enumerate(nng.subgraphs): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1269 | # remove passthrough tensors and attempt further optimizations |
| 1270 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 1271 | nng, |
| 1272 | sg, |
| 1273 | arch, |
| 1274 | [remove_passthrough_tensor], |
| 1275 | [fuse_activation_function_with_prev, optimise_pad, add_padding_fields], |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1276 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1277 | |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1278 | # Post-optimisation operator debug tracing, and checking that no undesired reshapes are left in the graph |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 1279 | for sg in nng.subgraphs: |
Patrik Gustavsson | 3a26920 | 2021-01-21 08:28:55 +0100 | [diff] [blame] | 1280 | rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [check_reshapes, _record_optimised]) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1281 | |
| 1282 | if verbose_graph: |
| 1283 | nng.print_graph() |
| 1284 | return nng |