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 |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 20 | |
| 21 | import numpy as np |
| 22 | |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 23 | from . import fp_math |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 24 | from . import lut |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 25 | from . import rewrite_graph |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 26 | from . import scaling |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 27 | from .data_type import DataType |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 28 | from .debug_database import DebugDatabase |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 29 | from .errors import UnsupportedFeatureError |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 30 | from .ethos_u55_regs.ethos_u55_regs import resampling_mode |
Louis Verhaard | 8912c53 | 2020-09-30 12:11:49 +0200 | [diff] [blame] | 31 | from .numeric_util import clamp_sigmoid |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 32 | from .numeric_util import full_shape |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 33 | from .numeric_util import round_away_zero |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame^] | 34 | from .operation import create_activation_function |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 35 | from .operation import create_avgpool_nop |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 36 | from .operation import NpuBlockType |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 37 | from .operation import Op |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 38 | from .operation import Operation |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 39 | from .softmax import SoftMax |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 40 | from .tensor import check_quantized_tens_scaling_equal |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 41 | from .tensor import create_const_tensor |
| 42 | from .tensor import create_reshape_tensor |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 43 | from .tensor import QuantizationParameters |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 44 | from .tensor import Tensor |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 45 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 46 | passthrough_nodes = set((Op.Identity,)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 47 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 48 | memory_only_ops = set((Op.Reshape,)) |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 49 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 50 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 51 | def remove_passthrough_tensor(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 52 | if len(tens.ops) == 1 and tens.ops[0].type in passthrough_nodes: |
| 53 | assert len(tens.ops[0].inputs) == 1 |
| 54 | tens = tens.ops[0].inputs[0] |
| 55 | return tens |
| 56 | |
| 57 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 58 | def rewrite_concat(tens, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 59 | if len(tens.ops) == 1 and tens.ops[0].type.is_concat_op(): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 60 | concat_op = tens.ops[0] |
| 61 | if tens != concat_op.outputs[0]: |
| 62 | return tens # don't attempt to rewrite the min/max outputs of QuantizedConcat |
| 63 | |
| 64 | # Not supported so leave it and run on CPU |
| 65 | if not concat_op.run_on_npu: |
| 66 | return tens |
| 67 | |
| 68 | inputs, axis = concat_op.get_concat_inputs_axis() |
| 69 | |
| 70 | tens.ops = [] |
| 71 | offset = 0 |
| 72 | for idx, inp in enumerate(inputs): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 73 | new_op = Operation(Op.ConcatSliceWrite, concat_op.name + str(idx)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 74 | new_op.inputs = [inp] |
| 75 | new_op.outputs = [tens] |
| 76 | new_op.attrs["concat_axis"] = axis |
| 77 | new_op.attrs["concat_start"] = offset |
| 78 | offset += inp.shape[axis] |
| 79 | new_op.attrs["concat_end"] = offset |
| 80 | new_op.run_on_npu = True |
| 81 | tens.ops.append(new_op) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 82 | DebugDatabase.add_optimised(concat_op, new_op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 83 | assert tens.shape[axis] == offset |
| 84 | |
Patrik Gustavsson | 29d568e | 2020-08-18 10:11:21 +0200 | [diff] [blame] | 85 | # If axis corresponds to C-dimension, NHCWB16 can only be used in the output if all the concat_start's are a |
| 86 | # multiple of 16. This as, it is only then the address offset for the ofm, for all operations, will be 16 byte |
| 87 | # aligned. For other values of axis the address offsets will be 16 byte aligned, as they are all based on c = 0 |
Patrik Gustavsson | 458a208 | 2020-08-13 13:41:05 +0200 | [diff] [blame] | 88 | # and those addresses are always 16 byte aligned due to the NHCWB16 format. |
Patrik Gustavsson | 6c97e9a | 2020-09-23 11:02:18 +0200 | [diff] [blame] | 89 | if axis == -1 or axis == (len(tens.shape) - 1): |
Patrik Gustavsson | 458a208 | 2020-08-13 13:41:05 +0200 | [diff] [blame] | 90 | for op in tens.ops: |
| 91 | if op.attrs["concat_start"] % 16 != 0: |
| 92 | tens.avoid_NHCWB16 = True |
| 93 | break |
| 94 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 95 | return tens |
| 96 | |
| 97 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 98 | def rewrite_split(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 99 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 100 | if len(tens.ops) == 1 and tens.ops[0].type.is_split_op(): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 101 | split_op = tens.ops[0] |
| 102 | |
| 103 | # Not supported so leave it and run on CPU |
| 104 | if not split_op.run_on_npu: |
| 105 | return tens |
| 106 | |
| 107 | inp, outputs, axis, offset_start, offset_end = split_op.get_split_inputs_axis() |
| 108 | |
| 109 | tens.ops = [] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 110 | new_op = Operation(Op.SplitSliceRead, split_op.name) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 111 | new_op.inputs = [inp] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 112 | |
| 113 | # For Split the offset cannot be extracted from the tensor so it has to |
| 114 | # be calculated from the index of the output tensor |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 115 | if axis is not None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 116 | # Get the start and end of the split |
| 117 | offset_start = [0] * len(tens.shape) |
| 118 | offset_end = [0] * len(tens.shape) |
| 119 | for out in outputs: |
| 120 | if out == tens: |
| 121 | break |
| 122 | offset_start[axis] += out.shape[axis] |
| 123 | |
Patrik Gustavsson | eebb1c2 | 2020-08-18 15:03:04 +0200 | [diff] [blame] | 124 | # If start offset is not a multiple of 16 in the C-dimension, NHCWB16 need to be avoided in the input |
| 125 | if (offset_start[-1] % 16) != 0: |
| 126 | inp.avoid_NHCWB16 = True |
| 127 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 128 | offset_end[axis] = offset_start[axis] + tens.shape[axis] |
| 129 | |
| 130 | new_op.attrs["split_start"] = offset_start |
| 131 | new_op.attrs["split_end"] = offset_end |
| 132 | new_op.run_on_npu = True |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 133 | new_op.set_output_tensor(tens) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 134 | DebugDatabase.add_optimised(split_op, new_op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 135 | |
| 136 | return tens |
| 137 | |
| 138 | |
| 139 | def needed_total_padding(input_size, stride, filter_size): |
| 140 | out_size = (input_size + stride - 1) // stride |
| 141 | needed_input = (out_size - 1) * stride + filter_size |
| 142 | total_padding = max(0, needed_input - input_size) |
| 143 | return total_padding |
| 144 | |
| 145 | |
| 146 | def calc_padding_and_skirt(padding_type, kernel_size, stride, input_dims): |
| 147 | ypad = needed_total_padding(int(input_dims[1]), int(stride[1]), int(kernel_size[0])) |
| 148 | xpad = needed_total_padding(int(input_dims[2]), int(stride[2]), int(kernel_size[1])) |
| 149 | if padding_type == b"SAME": |
| 150 | left_pad = (xpad + 0) // 2 |
| 151 | right_pad = (xpad + 1) // 2 |
| 152 | top_pad = (ypad + 0) // 2 |
| 153 | bottom_pad = (ypad + 1) // 2 |
| 154 | elif padding_type == b"VALID": |
| 155 | left_pad = 0 |
| 156 | right_pad = 0 |
| 157 | top_pad = 0 |
| 158 | bottom_pad = 0 |
| 159 | else: |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 160 | raise UnsupportedFeatureError("Unknown padding {}".format(str(padding_type))) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 161 | padding = (top_pad, left_pad, bottom_pad, right_pad) |
| 162 | skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad) |
| 163 | return padding, skirt |
| 164 | |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 165 | |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 166 | def calc_upscaled_padding_and_skirt(padding_type, kernel_size, stride, input_dims, upscaling_factor): |
| 167 | kernel_height, kernel_width = kernel_size[0], kernel_size[1] |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 168 | if padding_type == b"SAME": |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 169 | ypad = needed_total_padding(int(input_dims[1]) * upscaling_factor, int(stride[1]), int(kernel_height)) |
| 170 | xpad = needed_total_padding(int(input_dims[2]) * upscaling_factor, int(stride[2]), int(kernel_width)) |
| 171 | |
Jacob Bohlin | d47cc27 | 2020-08-24 11:42:14 +0200 | [diff] [blame] | 172 | right_pad = max(((xpad + 1) // upscaling_factor) - 1, 0) |
| 173 | bottom_pad = max(((ypad + 1) // upscaling_factor) - 1, 0) |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 174 | left_pad = max(kernel_width - 1 - right_pad, 0) |
| 175 | top_pad = max(kernel_height - 1 - bottom_pad, 0) |
| 176 | |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 177 | elif padding_type == b"VALID": |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 178 | right_pad = max(kernel_width - 2, 0) |
| 179 | bottom_pad = max(kernel_height - 2, 0) |
| 180 | left_pad = kernel_width - 1 |
| 181 | top_pad = kernel_height - 1 |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 182 | else: |
| 183 | assert 0, "Unknown padding" |
| 184 | |
| 185 | padding = (top_pad, left_pad, bottom_pad, right_pad) |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 186 | skirt = padding |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 187 | return padding, skirt |
| 188 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 189 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 190 | def fixup_conv2d_backprop(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 191 | if op.type == Op.Conv2DBackpropInput: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 192 | # flip the inputs |
| 193 | op.inputs[0], op.inputs[2] = op.inputs[2], op.inputs[0] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 194 | op.type = Op.Conv2DBackpropInputSwitchedBias |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 195 | |
| 196 | # Update strides |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 197 | 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] | 198 | |
| 199 | return op |
| 200 | |
| 201 | |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 202 | # Convert the op to an elementwise add |
| 203 | def convert_resizebilinear_1x1_to_add(op): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 204 | op.type = Op.Add |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 205 | op.name = op.name + "_add" |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 206 | op.attrs["resizebilinear"] = True |
| 207 | # Create an input tensor filled with zeros |
| 208 | shape = op.outputs[0].shape |
| 209 | tens = Tensor(shape, op.inputs[0].dtype, op.inputs[1].name + "_add") |
| 210 | tens.values = np.zeros(shape) |
| 211 | tens.quant_values = np.zeros(shape, np.uint8) |
| 212 | tens.quantization = QuantizationParameters(0.0, 255.0) |
| 213 | tens.quantization.scale_f32 = 1.0 |
| 214 | tens.quantization.zero_point = 0 |
| 215 | tens.consumer_list = [op] |
| 216 | tens_op = op.inputs[1].ops[0] |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 217 | tens_op.set_output_tensor(tens) |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 218 | # Set the add inputs |
| 219 | op.inputs[1] = op.inputs[0] |
| 220 | op.inputs[0] = tens |
| 221 | |
| 222 | return op |
| 223 | |
| 224 | |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 225 | # Convert ResizeBilinear to a number of 2x2 pool ops |
| 226 | def convert_resizebilinear_to_2x2_pool(op): |
| 227 | count = 0 |
| 228 | pre_op = op |
| 229 | outputs = op.outputs |
| 230 | |
| 231 | op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)}) |
| 232 | if op.attrs["align_corners"]: |
| 233 | shape_modifier = 1 |
| 234 | op.attrs["padding"] = b"VALID" |
| 235 | else: |
| 236 | shape_modifier = 0 |
| 237 | op.attrs["padding"] = b"SAME" |
| 238 | op.inputs[0].resampling_mode = resampling_mode.NEAREST |
| 239 | |
| 240 | upscaled_shape = np.array(op.inputs[0].shape[1:3]) |
| 241 | out_shape = np.array(op.outputs[0].shape[1:3]) |
| 242 | if (upscaled_shape == upscaled_shape * 2 - shape_modifier).all(): |
| 243 | return op |
| 244 | |
| 245 | while (upscaled_shape < out_shape).all(): |
| 246 | if count == 0: |
| 247 | scaled_op = pre_op |
| 248 | else: |
| 249 | scaled_op = op.clone("_{}".format(count)) |
| 250 | scaled_op.inputs[0] = pre_op.outputs[0] |
| 251 | |
| 252 | upscaled_shape = upscaled_shape * 2 - shape_modifier |
| 253 | |
| 254 | if (upscaled_shape == out_shape).all(): |
| 255 | scaled_op.outputs = outputs |
| 256 | scaled_op.outputs[0].ops = [scaled_op] |
| 257 | else: |
| 258 | shape = outputs[0].shape.copy() |
| 259 | shape[1:3] = upscaled_shape[0:2] |
| 260 | out_tens = Tensor(shape, DataType.int16, "{}_{}".format(op.outputs[0].name, count)) |
| 261 | out_tens.quantization = op.outputs[0].quantization.clone() |
| 262 | out_tens.quantization.quant_min = np.iinfo(np.int16).min |
| 263 | out_tens.quantization.quant_max = np.iinfo(np.int16).max |
| 264 | scaled_op.set_output_tensor(out_tens) |
| 265 | pre_op = scaled_op |
| 266 | count += 1 |
| 267 | |
| 268 | # Setup the scale value |
| 269 | if scaled_op.inputs[0].dtype.bits == 8 and scaled_op.outputs[0].dtype.bits == 16: |
| 270 | scaled_op.attrs["rescale"] = 128 |
| 271 | elif scaled_op.inputs[0].dtype.bits == 16 and scaled_op.outputs[0].dtype.bits == 8: |
| 272 | scaled_op.attrs["rescale"] = 1 / 128 |
| 273 | elif "rescale" in scaled_op.attrs: |
| 274 | del scaled_op.attrs["rescale"] |
| 275 | |
| 276 | return op |
| 277 | |
| 278 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 279 | def fixup_resizebilinear(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 280 | if op.type == Op.ResizeBilinear and op.run_on_npu: |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 281 | if op.inputs[0].shape == op.outputs[0].shape: |
Charles Xu | 36ffaf3 | 2020-08-05 15:40:44 +0200 | [diff] [blame] | 282 | # Bypass nop resizebilinear |
| 283 | op.inputs = op.inputs[:1] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 284 | op.type = Op.Identity |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 285 | elif op.inputs[0].shape[1] == 1 and op.inputs[0].shape[2] == 1: |
| 286 | convert_resizebilinear_1x1_to_add(op) |
| 287 | else: |
| 288 | convert_resizebilinear_to_2x2_pool(op) |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 289 | |
| 290 | return op |
| 291 | |
| 292 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 293 | def convert_nop_split_to_identity(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 294 | if op.type == Op.Split and op.attrs.get("num_splits") == 1: |
Dwight Lidman | c3862c2 | 2020-09-14 15:22:33 +0200 | [diff] [blame] | 295 | # the list comprehension should return a list with a single tensor |
| 296 | # if it shouldn't, remove_passthrough_tensor will fail appropriately |
| 297 | 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] | 298 | op.type = Op.Identity |
Dwight Lidman | c3862c2 | 2020-09-14 15:22:33 +0200 | [diff] [blame] | 299 | return op |
| 300 | |
| 301 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 302 | def fixup_fully_connected_input(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 303 | if op.type == Op.FullyConnected: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 304 | inp = op.inputs[0] |
| 305 | weights = op.inputs[1] |
| 306 | |
| 307 | n_in_elems = weights.shape[-2] |
| 308 | elms = inp.elements() |
| 309 | batch_size = elms // n_in_elems |
| 310 | assert batch_size * n_in_elems == elms |
| 311 | |
| 312 | desired_shape = [batch_size, n_in_elems] |
| 313 | if inp.shape != desired_shape: |
| 314 | # mismatch, insert a reshape to fix this. |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 315 | op.set_input_tensor(create_reshape_tensor(inp, desired_shape), 0) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 316 | |
| 317 | return op |
| 318 | |
| 319 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 320 | def convert_batched_fc_to_conv(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 321 | if op.type == Op.FullyConnected: |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 322 | ifm = op.inputs[0] |
| 323 | ofm = op.outputs[0] |
| 324 | # Check if the FC is 2D and first dimension indicates batching |
| 325 | if len(ifm.shape) == len(ofm.shape) == 2 and ifm.shape[0] != 1: |
| 326 | n = ifm.shape[0] |
| 327 | batching_split = {4: (2, 2), 8: (2, 4), 16: (4, 4)} |
| 328 | h, w = batching_split.get(n, (1, n)) |
| 329 | |
| 330 | # Convert to convolution |
| 331 | op.name += "_conv" |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 332 | op.type = Op.Conv2DBias |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 333 | op.attrs = { |
| 334 | "dilation": (1, 1, 1, 1), |
| 335 | "dilation_h_factor": 1, |
| 336 | "dilation_w_factor": 1, |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 337 | "padding": b"SAME", |
| 338 | "stride_h": 1, |
| 339 | "stride_w": 1, |
| 340 | "strides": (1, 1, 1, 1), |
| 341 | } |
| 342 | |
| 343 | prev_op = ifm.ops[0] |
| 344 | desired_shape = [1, h, w, ifm.shape[-1]] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 345 | if len(ifm.consumer_list) == 1 and prev_op is not None and prev_op.type == Op.Reshape: |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 346 | # There is a preceding Reshape |
| 347 | # Compare input of prev_op and input of op, to see if prev_op can be removed |
| 348 | ifm_prev_op = prev_op.inputs[0] |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 349 | if ifm_prev_op.shape == ifm.shape and check_quantized_tens_scaling_equal(ifm_prev_op, ifm.quantization): |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 350 | # prev_op can be removed |
| 351 | op.set_input_tensor(ifm_prev_op, 0) |
| 352 | else: |
| 353 | op.inputs[0].set_all_shapes(desired_shape) |
| 354 | prev_op.set_input_tensor( |
| 355 | create_const_tensor(prev_op.inputs[1].name, [1], DataType.int32, desired_shape), 1 |
| 356 | ) |
| 357 | prev_op.attrs["new_shape"] = desired_shape |
| 358 | else: |
| 359 | # Add reshape op to the input if there is no preceding reshape |
| 360 | ifm.consumer_list.remove(op) |
| 361 | op.set_input_tensor(create_reshape_tensor(ifm, desired_shape), 0) |
| 362 | |
| 363 | # Reshape Weights to be 4D. IO becomes HWIO |
| 364 | weight_tensor = op.inputs[1] |
| 365 | weight_tensor.quant_values = np.expand_dims(np.expand_dims(weight_tensor.quant_values, axis=0), axis=0) |
| 366 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
| 367 | |
| 368 | desired_shape = [1, h, w, ofm.shape[-1]] |
| 369 | if ( |
| 370 | len(ofm.consumer_list) == 1 |
| 371 | and ofm.consumer_list[0] is not None |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 372 | and ofm.consumer_list[0].type == Op.Reshape |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 373 | ): |
| 374 | # There is a subsequent Reshape |
| 375 | # Compare desired shape and output of consumer op, to see if consumer op can be removed |
| 376 | ofm_cons_op = ofm.consumer_list[0].outputs[0] |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 377 | if desired_shape == ofm_cons_op.shape and check_quantized_tens_scaling_equal(ofm, ofm_cons_op): |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 378 | op.outputs[0] = ofm_cons_op |
| 379 | op.outputs[0].ops = [op] |
| 380 | else: |
| 381 | op.outputs[0].set_all_shapes(desired_shape) |
| 382 | else: |
| 383 | # Add rehape op to the output |
| 384 | op.set_output_tensor(create_reshape_tensor(ofm, desired_shape, False)) |
| 385 | return op |
| 386 | |
| 387 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 388 | def fixup_pack_input(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 389 | if op.type == Op.Pack: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 390 | # Pack is also referred to as Stack |
| 391 | # Requires the rewrite_concat function to be called on the op afterwards |
| 392 | axis = int(op.attrs["axis"]) |
| 393 | desired_shape = op.inputs[0].shape[:axis] + [1] + op.inputs[0].shape[axis:] |
| 394 | |
| 395 | # Construct 1 shape tensor to be used by all inserted reshape ops |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 396 | new_shape_tens = create_const_tensor(op.name + "_reshape_shape", [1], DataType.int32, desired_shape) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 397 | |
| 398 | for idx, inp in enumerate(op.inputs): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 399 | reshape_out = inp.clone("_reshaped") |
Michael McGeagh | 6a8d424 | 2020-07-28 12:17:59 +0100 | [diff] [blame] | 400 | reshape_out.set_all_shapes(desired_shape) |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 401 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 402 | reshape_op = Operation(Op.Reshape, "{}{}_reshape".format(op.name, idx)) |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 403 | reshape_op.attrs["new_shape"] = desired_shape |
| 404 | reshape_op.inputs = [inp, new_shape_tens] |
| 405 | reshape_op.set_output_tensor(reshape_out) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 406 | DebugDatabase.add_optimised(op, reshape_op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 407 | |
| 408 | op.inputs[idx] = reshape_out |
| 409 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 410 | op.type = Op.PackReshaped |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 411 | |
| 412 | return op |
| 413 | |
| 414 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 415 | def unfuse_activation_function(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 416 | 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^] | 417 | 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] | 418 | op.activation = None |
Fredrik Svedberg | 0f98b36 | 2020-09-29 10:00:39 +0200 | [diff] [blame] | 419 | out_tens = op.outputs[0] |
| 420 | intermediate_tens = out_tens.clone("_act_intermediate") |
| 421 | act_op.set_output_tensor(out_tens) |
| 422 | act_op.add_input_tensor(intermediate_tens) |
| 423 | op.set_output_tensor(intermediate_tens) |
| 424 | |
| 425 | return op |
| 426 | |
Louis Verhaard | 8912c53 | 2020-09-30 12:11:49 +0200 | [diff] [blame] | 427 | |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 428 | def fixup_stridedslice_output(tens, arch, nng): |
| 429 | op = tens.ops[0] |
Dwight Lidman | 73320a4 | 2020-11-05 10:34:41 +0100 | [diff] [blame] | 430 | if op.run_on_npu and op.type == Op.StridedSlice: |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 431 | reshape_input_shape = tens.shape |
| 432 | new_axis_mask = op.attrs["new_axis_mask"] |
| 433 | shrink_axis_mask = op.attrs["shrink_axis_mask"] |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 434 | |
Dwight Lidman | 73320a4 | 2020-11-05 10:34:41 +0100 | [diff] [blame] | 435 | if shrink_axis_mask != 0: |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 436 | n = 0 |
| 437 | axis = 0 |
| 438 | while shrink_axis_mask: |
| 439 | prev_mask = shrink_axis_mask |
| 440 | n += 1 |
| 441 | shrink_axis_mask &= shrink_axis_mask - 1 |
| 442 | axis = int(math.log2(prev_mask - shrink_axis_mask)) |
| 443 | reshape_input_shape = reshape_input_shape[:axis] + [1] + reshape_input_shape[axis:] |
| 444 | |
| 445 | assert len(tens.shape) == (len(op.inputs[0].shape) - n) |
| 446 | op.attrs["shrink_axis_mask"] = 0 |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 447 | elif new_axis_mask != 0: |
| 448 | n = 0 |
| 449 | axis = 0 |
| 450 | while new_axis_mask: |
| 451 | prev_mask = new_axis_mask |
| 452 | n += 1 |
| 453 | new_axis_mask &= new_axis_mask - 1 |
| 454 | axis = int(math.log2(prev_mask - new_axis_mask)) |
| 455 | reshape_input_shape = reshape_input_shape[:axis] + reshape_input_shape[(axis + 1) :] |
| 456 | new_axis_mask >>= 1 |
| 457 | |
| 458 | assert len(tens.shape) == (len(op.inputs[0].shape) + n) |
| 459 | op.attrs["new_axis_mask"] = 0 |
Dwight Lidman | dda21af | 2020-11-11 15:44:57 +0100 | [diff] [blame] | 460 | else: |
| 461 | # Equal Rank StridedSlice, no need to insert reshape |
| 462 | return tens |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 463 | |
| 464 | # Construct 1 shape tensor to be used by all inserted reshape ops |
| 465 | new_shape_tens = create_const_tensor(op.name + "_reshape_shape", [1], DataType.int32, tens.shape) |
| 466 | |
| 467 | for idx, out_tens in enumerate(op.outputs): |
| 468 | reshape_in = out_tens.clone("_reshaped") |
| 469 | reshape_in.set_all_shapes(reshape_input_shape) |
| 470 | reshape_in.ops = [op] |
| 471 | |
| 472 | reshape_op = Operation(Op.Reshape, "{}{}_reshape".format(op.name, idx)) |
| 473 | reshape_op.attrs["new_shape"] = reshape_input_shape |
| 474 | reshape_op.inputs = [reshape_in, new_shape_tens] |
| 475 | reshape_op.set_output_tensor(out_tens) |
| 476 | |
| 477 | op.outputs[idx] = reshape_in |
| 478 | |
| 479 | return tens |
| 480 | |
| 481 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 482 | def fixup_unpack_output(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 483 | op = tens.ops[0] |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 484 | if op.run_on_npu and op.type == Op.Unpack: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 485 | # Unpack is also referred to as Unstack |
| 486 | # Requires the rewrite_split function to be called on the op afterwards |
Diqing Zhong | c7c0b1b | 2020-10-26 11:45:25 +0100 | [diff] [blame] | 487 | axis = int(op.attrs["axis"]) |
| 488 | op.type = Op.UnpackReshaped |
| 489 | reshape_input_shape = tens.shape[:axis] + [1] + tens.shape[axis:] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 490 | |
| 491 | # Construct 1 shape tensor to be used by all inserted reshape ops |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 492 | new_shape_tens = create_const_tensor(op.name + "_reshape_shape", [1], DataType.int32, tens.shape) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 493 | |
| 494 | for idx, out_tens in enumerate(op.outputs): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 495 | reshape_in = out_tens.clone("_reshaped") |
Michael McGeagh | 6a8d424 | 2020-07-28 12:17:59 +0100 | [diff] [blame] | 496 | reshape_in.set_all_shapes(reshape_input_shape) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 497 | reshape_in.ops = [op] |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 498 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 499 | reshape_op = Operation(Op.Reshape, "{}{}_reshape".format(op.name, idx)) |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 500 | reshape_op.attrs["new_shape"] = reshape_input_shape |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 501 | reshape_op.inputs = [reshape_in, new_shape_tens] |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 502 | reshape_op.set_output_tensor(out_tens) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 503 | DebugDatabase.add_optimised(op, reshape_op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 504 | |
| 505 | op.outputs[idx] = reshape_in |
| 506 | |
| 507 | return tens |
| 508 | |
| 509 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 510 | def add_padding_fields(op, arch, nng): |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 511 | if op.run_on_npu: |
| 512 | if "padding" in op.attrs: |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 513 | 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] | 514 | kernel_size = op.inputs[1].shape[:2] |
| 515 | input_shape = op.inputs[0].shape |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 516 | 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] | 517 | kernel_size = op.attrs["ksize"][1:3] |
| 518 | input_shape = op.inputs[0].shape |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 519 | else: |
| 520 | raise UnsupportedFeatureError("Unknown operation that uses padding: {}".format(op.type)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 521 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 522 | if op.type == Op.Conv2DBackpropInputSwitchedBias: |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 523 | upscaling_factor = op.outputs[0].shape[1] // input_shape[1] |
| 524 | padding, skirt = calc_upscaled_padding_and_skirt( |
| 525 | op.attrs["padding"], kernel_size, op.attrs["strides"], input_shape, upscaling_factor |
| 526 | ) |
| 527 | else: |
| 528 | dilation_h, dilation_w = op.get_dilation_h_w() |
| 529 | dilated_kernel_size = [dilation_h * (kernel_size[0] - 1) + 1, dilation_w * (kernel_size[1] - 1) + 1] |
| 530 | padding, skirt = calc_padding_and_skirt( |
| 531 | op.attrs["padding"], dilated_kernel_size, op.attrs["strides"], input_shape |
| 532 | ) |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 533 | |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 534 | op.attrs["explicit_padding"] = padding |
| 535 | op.attrs["skirt"] = skirt |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 536 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 537 | return op |
| 538 | |
| 539 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 540 | # Check if the op can be reordered |
| 541 | def get_prepend_op(op): |
| 542 | inp = op.inputs[0] |
| 543 | # The op should be reordered between prev_op and prep_op |
| 544 | prev_op = inp.ops[-1] |
| 545 | prep_op = None |
| 546 | while prev_op.type in memory_only_ops and len(prev_op.outputs) == 1 and len(prev_op.outputs[0].consumers()) == 1: |
| 547 | prep_op = prev_op |
| 548 | inp = prev_op.inputs[0] |
| 549 | prev_op = inp.ops[-1] |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 550 | 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] | 551 | return prep_op |
| 552 | |
| 553 | return None |
| 554 | |
| 555 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 556 | def convert_depthwise_to_conv(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 557 | # Depthwise is equivalent to a single conv2d if the ifm depth is 1 and |
| 558 | # the ofm depth equals the depth multipler. |
| 559 | # If those conditions are true, then we can perform a simple |
| 560 | # switch of the operator type (and weight order) |
| 561 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 562 | if op.type == Op.DepthwiseConv2DBias and (op.attrs["depth_multiplier"] != 1): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 563 | ifm_tensor = op.inputs[0] |
| 564 | weight_tensor = op.inputs[1] |
| 565 | ofm_tensor = op.outputs[0] |
| 566 | if (ifm_tensor.shape[3] == 1) and (ofm_tensor.shape[3] == op.attrs["depth_multiplier"]): |
| 567 | # Change op type to Conv2d |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 568 | op.type = Op.Conv2DBias |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 569 | del op.attrs["channel_multiplier"] |
| 570 | del op.attrs["depth_multiplier"] |
| 571 | |
| 572 | 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] | 573 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 574 | else: |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 575 | raise UnsupportedFeatureError( |
| 576 | "Unsupported DepthwiseConv2d with depth_multiplier = {}, ifm channels = {}, ofm channels = {}".format( |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 577 | op.attrs["depth_multiplier"], ifm_tensor.shape[3], ofm_tensor.shape[3] |
| 578 | ) |
| 579 | ) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 580 | DebugDatabase.add_optimised(op, op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 581 | return op |
| 582 | |
| 583 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 584 | def reorder_depthwise_weights(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 585 | if op.type.is_depthwise_conv2d_op(): |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 586 | weight_tensor = op.inputs[1] |
| 587 | 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] | 588 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 589 | weight_tensor.weight_transpose_depthwise = True |
| 590 | |
| 591 | return op |
| 592 | |
| 593 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 594 | def convert_conv_to_fc(op, arch, nng): |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 595 | # Conv 1x1 can be equivalent to Fully Connected. |
| 596 | # By representing certain convs as fully connected layers, Vela can better determine wether or not to use |
| 597 | # caching/double buffering for the weights. |
| 598 | # (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] | 599 | if op.type == Op.Conv2DBias: |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 600 | _, h, w, _ = op.inputs[0].shape |
| 601 | kh, kw, _, _ = op.inputs[1].shape |
| 602 | if h == 1 and w == 1 and kh == 1 and kw == 1: |
| 603 | # Overwrite this op as a Fully Connected Op |
| 604 | op.name += "_fc" |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 605 | op.type = Op.FullyConnected |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 606 | op.attrs = { |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 607 | "weights_format": 0, |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 608 | } |
| 609 | # Reshape Weights to be 2D. HWIO becomes just IO (as H and W are 1, they can just be dropped) |
| 610 | weight_tensor = op.inputs[1] |
| 611 | weight_tensor.quant_values = weight_tensor.quant_values.squeeze(axis=(0, 1)) |
| 612 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
| 613 | # The output from a fully connected is expected to be 2D so we need to add a reshape layer to convert it |
| 614 | # back to 4D afterwards as the next layer is expecting that shape |
| 615 | orig_ofm_tensor = op.outputs[0] |
| 616 | # Reshape this ops output to be 2D: {(N*H*W), C} (We know N H and W are all 1 so this becomes {1, C}) |
| 617 | fc_ofm_tensor = orig_ofm_tensor.clone("_fc") |
| 618 | fc_ofm_tensor.set_all_shapes([1, fc_ofm_tensor.shape[-1]]) |
| 619 | fc_ofm_tensor.ops = [op] |
| 620 | # Add a reshape after the new OFM to convert it back to the original 4D shape |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 621 | reshape_name = op.name + "_reshape" |
| 622 | new_shape_tens = create_const_tensor(reshape_name + "_shape", [1], DataType.int32, orig_ofm_tensor.shape) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 623 | reshape_op = Operation(Op.Reshape, reshape_name) |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 624 | reshape_op.attrs["new_shape"] = orig_ofm_tensor.shape |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 625 | reshape_op.inputs = [fc_ofm_tensor, new_shape_tens] |
| 626 | reshape_op.set_output_tensor(orig_ofm_tensor) |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 627 | # Replace this ops OFM to point to the 2D tensor |
| 628 | op.outputs[0] = fc_ofm_tensor |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 629 | # Record optimisation in debug database |
| 630 | DebugDatabase.add_optimised(op, reshape_op) |
| 631 | DebugDatabase.add_optimised(op, op) |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 632 | return op |
| 633 | |
| 634 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 635 | def fixup_relus_with_differing_ifm_ofm_scaling(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 636 | if op.run_on_npu and op.type.is_relu_op(): |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 637 | ifm = op.inputs[0] |
| 638 | ofm = op.outputs[0] |
| 639 | # Relu with differing IFM and OFM scaling cannot be fused with another primary op |
| 640 | # and requires its own to be inserted |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 641 | if not check_quantized_tens_scaling_equal(ifm, ofm): |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 642 | # Override this op with its own primary op (avgpool) |
| 643 | relu_fused_op = create_avgpool_nop(op.name + "_avgpool") |
| 644 | # And fuse the original activation function to it |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame^] | 645 | relu_fused_op.activation = create_activation_function(op.type) |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 646 | # Tidy up and assign the ifm and ofm to the new op |
| 647 | ifm.consumer_list.remove(op) |
Andreas Nevalainen | f3d737e | 2020-09-25 14:12:43 +0200 | [diff] [blame] | 648 | |
| 649 | # if not 4d, reshape ifm/ofm |
| 650 | if len(ifm.shape) < 4: |
| 651 | ifm_shaped = create_reshape_tensor(ifm, full_shape(4, ifm.shape, 1)) |
| 652 | ifm = ifm_shaped |
| 653 | if len(ofm.shape) < 4: |
| 654 | ofm_shaped = create_reshape_tensor(ofm, full_shape(4, ofm.shape, 1), False) |
| 655 | ofm = ofm_shaped |
| 656 | |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 657 | relu_fused_op.add_input_tensor(ifm) |
| 658 | relu_fused_op.set_output_tensor(ofm) |
| 659 | op = relu_fused_op |
| 660 | return op |
| 661 | |
| 662 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 663 | # Reorder activation op if it's after the memory only operations |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 664 | def fixup_act_reorder(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 665 | if op.type.is_relu_op() or op in set((Op.Sigmoid, Op.Tanh)): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 666 | prep_op = get_prepend_op(op) |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 667 | if prep_op is not None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 668 | act_op = op.clone("_reordered") |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 669 | |
| 670 | # There is only one input tensor, overwrite it |
| 671 | act_op.set_input_tensor(prep_op.inputs[0], 0) |
| 672 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 673 | act_op_out = act_op.inputs[0].clone("_acted") |
| 674 | act_op_out.quantization = op.outputs[0].quantization.clone() |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 675 | act_op.set_output_tensor(act_op_out) |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 676 | |
| 677 | # Update the consumer list |
| 678 | act_op_out.consumer_list = op.outputs[0].consumer_list.copy() |
| 679 | act_op_out.consumer_list.append(prep_op) |
| 680 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 681 | prep_op.inputs[0] = act_op_out |
| 682 | prep_op.outputs[0].quantization = act_op_out.quantization.clone() |
| 683 | |
| 684 | # 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] | 685 | op.type = Op.Identity |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 686 | |
| 687 | # Record optimisation in debug database |
| 688 | DebugDatabase.add_optimised(op, act_op) |
| 689 | DebugDatabase.add_optimised(op, op) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 690 | return op |
| 691 | |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 692 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 693 | def fixup_elementwise_with_scalars(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 694 | if op.type.is_binary_elementwise_op(): |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 695 | ifm_tensor, ifm2_tensor, _, _ = op.get_ifm_ifm2_weights_ofm() |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 696 | if ifm2_tensor.shape != [] and ifm_tensor.shape != []: |
| 697 | diff = len(ifm_tensor.shape) - len(ifm2_tensor.shape) |
| 698 | if diff > 0: |
| 699 | ifm2_tensor.shape = full_shape(len(ifm_tensor.shape), ifm2_tensor.shape, 1) |
| 700 | elif diff < 0: |
| 701 | 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] | 702 | elif ifm_tensor.shape == [] and ifm_tensor.quant_values is None: |
| 703 | # IFM is marked as a scalar, but is a result of an operation; change it to a shape of size 1 |
| 704 | ifm_tensor.shape = len(ifm2_tensor.shape) * [1] |
| 705 | ifm_tensor.storage_shape = ifm_tensor.shape |
| 706 | elif ifm2_tensor.shape == [] and ifm2_tensor.quant_values is None: |
| 707 | # IFM2 is marked as a scalar, but is a result of an operation; change it to a shape of size 1 |
| 708 | ifm2_tensor.shape = len(ifm_tensor.shape) * [1] |
| 709 | ifm2_tensor.storage_shape = ifm2_tensor.shape |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 710 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 711 | |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 712 | |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 713 | # Set input/output tensor equivalence to the same id for memory operations |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 714 | def set_tensor_equivalence(op, arch, nng): |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 715 | if op.type in memory_only_ops: |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 716 | eid = op.outputs[0].equivalence_id |
| 717 | for inp in op.inputs: |
| 718 | inp.equivalence_id = eid |
| 719 | return op |
| 720 | |
| 721 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 722 | def convert_softmax(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 723 | if op.type == Op.Softmax and op.run_on_npu: |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 724 | softmax = SoftMax(op) |
| 725 | op = softmax.get_graph() |
| 726 | return op |
| 727 | |
| 728 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 729 | def convert_mul_max_to_abs_or_lrelu(op, arch, nng): |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 730 | r"""Whenever there is a subgraph with this topology: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 731 | |
| 732 | Input X For X = -1 or X > 0 |
| 733 | | \ / This subgraph can be replaced with either |
| 734 | | Mul an Abs (if X = -1) or a LeakyReLU (if X > 0) |
| 735 | | / |
| 736 | Max |
| 737 | """ |
| 738 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 739 | if op.type == Op.Maximum: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 740 | # finds the Mul input(s) to the Max |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 741 | 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] | 742 | if len(muls) == 1: |
| 743 | mul = muls[0].ops[0] |
| 744 | elif len(muls) == 2: |
| 745 | # In the case both inputs are Muls, find the one with the same input as the Max |
| 746 | mul = [m for m in muls if len(set(op.inputs + m.ops[0].inputs)) == 1][0].ops[0] |
| 747 | else: |
| 748 | # No Mul inputs |
| 749 | return op |
| 750 | |
| 751 | # make sure the Mul doesn't have any other consumers |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 752 | mul_ofm = mul.outputs[0] |
| 753 | if len(mul_ofm.consumers()) != 1: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 754 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 755 | # make sure the Mul doesn't have a fused activation function |
| 756 | if mul.activation: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 757 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 758 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 759 | if ifm is None or ofm is None: |
| 760 | return op |
| 761 | |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 762 | if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype: |
| 763 | return op |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 764 | 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] | 765 | # rewrite to LeakyRelu currently only makes sense if the quantization is identical |
| 766 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 767 | |
| 768 | # finds the branched input that goes to both the Max and the Mul |
| 769 | shared = set(op.inputs) & set(mul.inputs) |
| 770 | if len(shared) == 1: |
| 771 | shared_in = shared.pop() |
| 772 | # find the constant scalar input to the Mul |
| 773 | const_tens = (set(mul.inputs) - {shared_in}).pop() |
| 774 | # check that it is a scalar |
| 775 | if const_tens.shape != []: |
| 776 | return op |
| 777 | const = const_tens.ops[0] |
| 778 | # check that it is a constant |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 779 | if const.type != Op.Const: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 780 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 781 | # Remove the Mul from the shared input's consumers |
| 782 | shared_in.consumer_list.remove(mul) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 783 | else: |
| 784 | return op |
| 785 | |
| 786 | val = const.outputs[0].values |
| 787 | if val >= 0: |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 788 | new_op = Op.LeakyRelu |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 789 | op.attrs["alpha"] = val |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 790 | # to produce bit exact results, the alpha is not enough; |
| 791 | # save additional scaling info in attr "alpha_scale", to be used as input |
| 792 | # to the LUT construction |
| 793 | alpha_scalar = const_tens.quant_values - const_tens.quantization.zero_point |
| 794 | mul_ifm_scale = np.double(ifm.quantization.scale_f32) |
| 795 | mul_ifm2_scale = np.double(const_tens.quantization.scale_f32) |
| 796 | mul_ofm_scale = np.double(mul_ofm.quantization.scale_f32) |
| 797 | alpha_scale, alpha_shift = scaling.elementwise_mul_scale(mul_ifm_scale, mul_ifm2_scale, mul_ofm_scale) |
| 798 | op.attrs["alpha_scaling"] = (alpha_scalar, alpha_scale, alpha_shift) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 799 | elif val == -1: |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 800 | new_op = Op.Abs |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 801 | else: |
| 802 | return op |
| 803 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 804 | op.type = new_op |
| 805 | op.name = op.name.replace("Maximum", new_op.name) |
| 806 | 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] | 807 | op.inputs = [shared_in] |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 808 | |
| 809 | # Record optimisation in debug database |
| 810 | DebugDatabase.add_optimised(op, op) |
| 811 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 812 | return op |
| 813 | |
| 814 | |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 815 | def convert_lrelu_to_mul_max(op, arch): |
| 816 | # Converts LeakyRelu to Max(alpha * IFM, identity * IFM) |
| 817 | # (the opposite of convert_mul_max_to_abs_or_lrelu) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 818 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 819 | if ifm is None or ofm is None: |
| 820 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 821 | |
| 822 | # Add multiplication with alpha |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 823 | mul_alpha = Operation(Op.Mul, op.name + "_mul_alpha") |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 824 | mul_alpha.add_input_tensor(ifm) |
| 825 | # Create const tensor containing alpha as scalar |
| 826 | alpha = op.attrs["alpha"] |
| 827 | quantization = ifm.quantization.clone() |
| 828 | quantization.min = 0 |
| 829 | quantization.max = alpha * (quantization.quant_max - quantization.quant_min) |
| 830 | quantization.scale_f32 = alpha |
| 831 | quantization.zero_point = 0 |
| 832 | alpha_tens = create_const_tensor(op.name + "_alpha_scalar", [], ifm.dtype, [1], np.int8, quantization=quantization) |
| 833 | mul_alpha.add_input_tensor(alpha_tens) |
| 834 | fm_alpha = ofm.clone(op.name + "_alpha") |
| 835 | mul_alpha.set_output_tensor(fm_alpha) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 836 | DebugDatabase.add_optimised(op, mul_alpha) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 837 | |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 838 | if check_quantized_tens_scaling_equal(ifm, ofm): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 839 | # No identity multiplication is needed |
| 840 | fm_id = ifm |
| 841 | else: |
| 842 | # Add multiplication with identity |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 843 | mul_identity = Operation(Op.Mul, op.name + "_mul_identity") |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 844 | mul_identity.add_input_tensor(ifm) |
| 845 | # Create const tensor containing identity as scalar |
| 846 | quantization = ifm.quantization.clone() |
| 847 | quantization.min = 0 |
| 848 | quantization.max = quantization.quant_max - quantization.quant_min |
| 849 | quantization.scale_f32 = 1 |
| 850 | quantization.zero_point = 0 |
| 851 | identity_tens = create_const_tensor( |
| 852 | op.name + "_id_scalar", [], ifm.dtype, [1], np.uint8, quantization=quantization |
| 853 | ) |
| 854 | mul_identity.add_input_tensor(identity_tens) |
| 855 | fm_id = ofm.clone(op.name + "_id") |
| 856 | mul_identity.set_output_tensor(fm_id) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 857 | DebugDatabase.add_optimised(op, mul_alpha) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 858 | |
| 859 | # 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] | 860 | op.type = Op.Maximum |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 861 | op.name = op.name.replace("LeakyRelu", "Maximum") |
| 862 | op.inputs = [] |
| 863 | ifm.consumer_list.remove(op) |
| 864 | op.add_input_tensor(fm_alpha) |
| 865 | op.add_input_tensor(fm_id) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 866 | |
| 867 | DebugDatabase.add_optimised(op, op) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 868 | return op |
| 869 | |
| 870 | |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 871 | def convert_to_lut(op, lut_values, lut_name): |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 872 | # Rewrite the operation by Add with scalar 0 + LUT activation |
| 873 | ifm = op.inputs[0] |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 874 | if ifm is None: |
| 875 | return op |
Louis Verhaard | 58520b9 | 2020-08-24 16:45:38 +0200 | [diff] [blame] | 876 | assert ifm.dtype.size_in_bytes() == 1 |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 877 | op.type = Op.Add |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 878 | op.name = op.name + "_lut_" + lut_name |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 879 | # Mark as no-op to enable potential fusing optimizations |
| 880 | op.attrs["is_nop"] = True |
| 881 | # Create an input tensor containing scalar zero |
| 882 | quantization = QuantizationParameters(0.0, 255.0) |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 883 | quantization.scale_f32 = ifm.quantization.scale_f32 |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 884 | quantization.zero_point = 0 |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 885 | 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] | 886 | op.add_input_tensor(tens) |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 887 | # The LUT must be applied without any preceding rescaling (the LUT itself performs the rescale), |
| 888 | # so even if the OFM has a different scale than the IFM, the generated OFM scale instructions |
| 889 | # should be the same as the IFM |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 890 | op.forced_output_quantization = ifm.quantization |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 891 | 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] | 892 | op.set_activation_lut(lut_tensor) |
| 893 | return op |
| 894 | |
| 895 | |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 896 | def convert_to_lut8(op, fn, fn_name): |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 897 | # Converts op to a no-op + int8/uint8 LUT which is generated with the given function. |
| 898 | # fn is a function(real) -> real |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 899 | ifm, ofm = op.get_ifm_ofm() |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 900 | if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype: |
| 901 | return op |
| 902 | # Generate the LUT |
| 903 | ifm_scale = np.double(ifm.quantization.scale_f32) |
| 904 | ofm_scale = np.double(ofm.quantization.scale_f32) |
| 905 | zp_in = ifm.quantization.zero_point |
| 906 | zp_out = ofm.quantization.zero_point |
| 907 | values = [] |
| 908 | ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128) |
| 909 | quantized_min = min(ix) |
| 910 | quantized_max = max(ix) |
| 911 | for x in ix: |
| 912 | x_real = ifm_scale * (x - zp_in) |
| 913 | y_real = fn(x_real) |
| 914 | lut_result = round_away_zero(zp_out + y_real / ofm_scale) |
| 915 | lut_result = min(quantized_max, max(quantized_min, lut_result)) |
| 916 | values.append(lut_result) |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 917 | return convert_to_lut(op, values, fn_name) |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 918 | |
| 919 | |
| 920 | def convert_lrelu_to_lut(op, arch): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 921 | ifm, ofm = op.get_ifm_ofm() |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 922 | # Generate the LUT |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 923 | alpha = op.attrs["alpha"] |
| 924 | ifm_scale = np.double(ifm.quantization.scale_f32) |
| 925 | ofm_scale = np.double(ofm.quantization.scale_f32) |
| 926 | zp_in = ifm.quantization.zero_point |
| 927 | zp_out = ofm.quantization.zero_point |
| 928 | identity_scale, identity_shift = scaling.elementwise_mul_scale(ifm_scale, 1, ofm_scale) |
| 929 | alpha_scalar = 1 |
| 930 | alpha_scale, alpha_shift = scaling.elementwise_mul_scale(ifm_scale, alpha, ofm_scale) |
| 931 | if "alpha_scaling" in op.attrs: |
| 932 | # The LeakyRelu was the result from convert_mul_max_to_abs_or_lrelu |
| 933 | alpha_scalar, alpha_scale, alpha_shift = op.attrs["alpha_scaling"] |
| 934 | values = [] |
Louis Verhaard | 58520b9 | 2020-08-24 16:45:38 +0200 | [diff] [blame] | 935 | ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128) |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 936 | quantized_min = min(ix) |
| 937 | quantized_max = max(ix) |
| 938 | for x in ix: |
| 939 | if x < zp_in: |
| 940 | lut_result = zp_out + fp_math.multiply_by_quantized_multiplier( |
| 941 | alpha_scalar * (x - zp_in), alpha_scale, alpha_shift |
| 942 | ) |
| 943 | else: |
| 944 | lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(x - zp_in, identity_scale, identity_shift) |
| 945 | lut_result = min(quantized_max, max(quantized_min, lut_result)) |
| 946 | values.append(lut_result) |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 947 | return convert_to_lut(op, values, "lrelu") |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 948 | |
| 949 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 950 | def convert_lrelu(op, arch, nng): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 951 | # 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] | 952 | if op.type != Op.LeakyRelu: |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 953 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 954 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 955 | if ifm is None or ofm is None: |
| 956 | return op |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 957 | if ifm.dtype in (DataType.uint8, DataType.int8) and ifm.dtype == ofm.dtype: |
| 958 | # use LUT for int8/uint8 |
| 959 | return convert_lrelu_to_lut(op, arch) |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 960 | 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] | 961 | # use LeakyRelu unmodified for int16 with equal input/output scaling |
| 962 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 963 | return convert_lrelu_to_mul_max(op, arch) |
| 964 | |
| 965 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 966 | def convert_tanh_sigmoid_to_lut(op, arch, nng): |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 967 | # Converts int8/uint8 Sigmoid and Tanh to a LUT based solution |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 968 | if op.type == Op.Sigmoid: |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 969 | return convert_to_lut8(op, clamp_sigmoid, "sigmoid") |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 970 | elif op.type == Op.Tanh: |
Louis Verhaard | 2e186c7 | 2020-10-09 10:47:04 +0200 | [diff] [blame] | 971 | return convert_to_lut8(op, math.tanh, "tanh") |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 972 | return op |
| 973 | |
| 974 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 975 | def remove_unwanted_reshapes(op, arch, nng): |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 976 | # Try to remove reshapes enclosing ElementWise operator with only one non-constant input |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 977 | if not op.run_on_npu or not op.type.is_elementwise_op(): |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 978 | return op |
| 979 | |
| 980 | # Check if the ElementWise operator only have one non-constant input |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 981 | non_const_tens = [x for x in op.inputs if x.ops[0].type != Op.Const] |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 982 | if len(non_const_tens) != 1: |
| 983 | return op |
| 984 | ifm = non_const_tens[0] |
| 985 | |
| 986 | # Check if operation is enclosed by Reshapes that can be removed |
| 987 | ofm = op.outputs[0] |
| 988 | prev_op = ifm.ops[0] |
| 989 | if ( |
| 990 | len(ifm.consumer_list) == 1 |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 991 | and prev_op.type == Op.Reshape |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 992 | and len(ofm.consumer_list) == 1 |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 993 | and ofm.consumer_list[0].type == Op.Reshape |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 994 | ): |
| 995 | # Operation is enclosed by reshapes, check if they can be removed |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 996 | prev_op_ifm, prev_op_ofm = prev_op.get_ifm_ofm() |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 997 | cons_op = ofm.consumer_list[0] |
| 998 | cons_op_ifm = ofm |
| 999 | cons_op_ofm = cons_op.outputs[0] |
| 1000 | if len(prev_op_ifm.shape) == len(cons_op_ofm.shape): |
| 1001 | # Check if quantization is the same in the input and output for the reshape ops |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 1002 | if check_quantized_tens_scaling_equal(prev_op_ifm, prev_op_ofm) and check_quantized_tens_scaling_equal( |
| 1003 | cons_op_ifm, cons_op_ofm |
| 1004 | ): |
Patrik Gustavsson | 7ad862a | 2020-09-29 14:09:43 +0200 | [diff] [blame] | 1005 | op.set_input_tensor(prev_op_ifm, 0) |
| 1006 | op.set_output_tensor(cons_op_ofm) |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 1007 | return op |
| 1008 | |
| 1009 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1010 | def fuse_activation_function_with_prev(op, arch, nng): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1011 | # 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] | 1012 | 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] | 1013 | return op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1014 | ifm, ofm = op.get_ifm_ofm() |
Tim Hall | 9358296 | 2020-09-09 21:58:15 +0100 | [diff] [blame] | 1015 | if ifm is None or ofm is None: |
| 1016 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1017 | # finds the input(s) to the operation |
| 1018 | prev_op = ifm.ops[0] |
| 1019 | # Note: the below checks on prev_op require that a first optimize pass on the full graph has been performed |
| 1020 | fuse = ( |
| 1021 | prev_op.run_on_npu |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1022 | and prev_op.type.npu_block_type != NpuBlockType.Default |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1023 | and len(ifm.ops) == 1 |
| 1024 | and len(prev_op.outputs[0].consumers()) == 1 |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1025 | and prev_op.activation is None |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1026 | ) |
| 1027 | if op.activation_lut is not None and arch.shram_reserved_unused_banks == 0: |
| 1028 | # TODO: if SHRAM LUT space is shared with SHRAM ACC (32, 64 MAC), |
| 1029 | # LUT currently only works correctly for elementwise ops |
| 1030 | fuse = False |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1031 | if not fuse: |
| 1032 | return op |
| 1033 | # Move the fused activation function + corresponding info to prev_op |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1034 | prev_op.activation = op.activation |
| 1035 | prev_op.forced_output_quantization = op.forced_output_quantization |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1036 | if op.activation_lut is not None: |
| 1037 | prev_op.set_activation_lut(op.activation_lut) |
| 1038 | # Bypass op |
Louis Verhaard | 98a3499 | 2020-09-01 10:39:04 +0200 | [diff] [blame] | 1039 | prev_op.set_output_tensor(ofm) |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 1040 | DebugDatabase.add_optimised(op, prev_op) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1041 | return op |
| 1042 | |
| 1043 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1044 | def add_attrs_to_resizebilinear(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1045 | if op.type == Op.ResizeBilinear and op.run_on_npu: |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1046 | input_tensor = op.inputs[0] |
| 1047 | upscaled_shape = [input_tensor.shape[1] * 2, input_tensor.shape[2] * 2] |
| 1048 | out_shape = op.outputs[0].shape[1:3] |
| 1049 | if not op.attrs["align_corners"] and out_shape == upscaled_shape: |
| 1050 | # this means the output is supposed to be a x2 upscale, |
| 1051 | # so we need to do SAME padding |
| 1052 | op.attrs["padding"] = b"SAME" |
| 1053 | elif op.attrs["align_corners"] and out_shape == [upscaled_shape[0] - 1, upscaled_shape[1] - 1]: |
| 1054 | # here we can just run the avg pool without padding and |
| 1055 | # produce a (M * 2 - 1, N * 2 - 1) sized output |
| 1056 | op.attrs["padding"] = b"VALID" |
| 1057 | else: |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 1058 | return op |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1059 | input_tensor.resampling_mode = resampling_mode.NEAREST |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 1060 | 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] | 1061 | return op |
| 1062 | |
| 1063 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1064 | def fixup_bias_tensors(op, arch, nng): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 1065 | if op.type.needs_bias() and op.bias is None: |
Jacob Bohlin | a41cd4d | 2020-08-26 18:21:28 +0200 | [diff] [blame] | 1066 | # Op has no bias, add bias tensor filled with zeros |
| 1067 | nr_biases = op.inputs[1].shape[-1] |
| 1068 | bias_values = [0] * nr_biases |
| 1069 | bias_tensor = create_const_tensor(op.name + "_bias", [nr_biases], DataType.int32, bias_values) |
| 1070 | bias_tensor.quant_values = bias_tensor.values |
| 1071 | op.set_input_tensor(bias_tensor, -1) |
Jacob Bohlin | 67e0d8f | 2020-08-20 10:53:02 +0200 | [diff] [blame] | 1072 | |
| 1073 | return op |
| 1074 | |
| 1075 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1076 | def supported_operator_check(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1077 | op.run_on_npu = arch.supported_operators.is_operator_supported(op) |
| 1078 | return op |
| 1079 | |
| 1080 | |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 1081 | def _record_optimised(op, arch): |
| 1082 | if op.type != Op.Const: |
| 1083 | DebugDatabase.add_optimised(op, op) |
| 1084 | |
| 1085 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1086 | def optimise_graph_a(nng, arch, verbose_graph=False): |
| 1087 | if verbose_graph: |
| 1088 | nng.print_graph() |
| 1089 | |
| 1090 | op_rewrite_list = [ |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 1091 | set_tensor_equivalence, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1092 | supported_operator_check, |
| 1093 | # then do any rewrites of supported operators |
| 1094 | convert_depthwise_to_conv, |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 1095 | convert_conv_to_fc, |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 1096 | convert_softmax, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1097 | fixup_fully_connected_input, |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 1098 | convert_batched_fc_to_conv, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1099 | fixup_pack_input, |
Fredrik Svedberg | 0f98b36 | 2020-09-29 10:00:39 +0200 | [diff] [blame] | 1100 | unfuse_activation_function, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1101 | fixup_conv2d_backprop, |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 1102 | fixup_relus_with_differing_ifm_ofm_scaling, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1103 | fixup_act_reorder, |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 1104 | fixup_elementwise_with_scalars, |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 1105 | reorder_depthwise_weights, |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 1106 | fixup_resizebilinear, |
Jacob Bohlin | a41cd4d | 2020-08-26 18:21:28 +0200 | [diff] [blame] | 1107 | fixup_bias_tensors, |
Dwight Lidman | c3862c2 | 2020-09-14 15:22:33 +0200 | [diff] [blame] | 1108 | convert_nop_split_to_identity, |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1109 | convert_mul_max_to_abs_or_lrelu, |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 1110 | remove_unwanted_reshapes, |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1111 | convert_lrelu, |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 1112 | convert_tanh_sigmoid_to_lut, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1113 | ] |
| 1114 | |
| 1115 | for idx, sg in enumerate(nng.subgraphs): |
| 1116 | # rewrite graph pass |
| 1117 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Dwight Lidman | 73320a4 | 2020-11-05 10:34:41 +0100 | [diff] [blame] | 1118 | nng, sg, arch, [], op_rewrite_list, rewrite_unsupported=False, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1119 | ) |
| 1120 | |
| 1121 | for idx, sg in enumerate(nng.subgraphs): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1122 | # remove passthrough tensors and attempt further optimizations |
| 1123 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame] | 1124 | nng, sg, arch, [remove_passthrough_tensor], [fuse_activation_function_with_prev, add_padding_fields] |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1125 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1126 | |
Tim Hall | e6ccd87 | 2020-11-09 16:46:37 +0000 | [diff] [blame] | 1127 | # Post-optimisation operator debug tracing |
| 1128 | for sg in nng.subgraphs: |
| 1129 | rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [_record_optimised]) |
| 1130 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1131 | if verbose_graph: |
| 1132 | nng.print_graph() |
| 1133 | return nng |
| 1134 | |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 1135 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1136 | def optimise_graph_b(nng, arch, verbose_graph=False): |
| 1137 | if verbose_graph: |
| 1138 | nng.print_graph() |
| 1139 | |
| 1140 | for idx, sg in enumerate(nng.subgraphs): |
| 1141 | # combined rewrite graph pass |
Dwight Lidman | c6ac194 | 2020-10-02 14:55:45 +0200 | [diff] [blame] | 1142 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Dwight Lidman | 73320a4 | 2020-11-05 10:34:41 +0100 | [diff] [blame] | 1143 | nng, sg, arch, [fixup_unpack_output, fixup_stridedslice_output, rewrite_concat, rewrite_split], [] |
Dwight Lidman | c6ac194 | 2020-10-02 14:55:45 +0200 | [diff] [blame] | 1144 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1145 | |
| 1146 | if verbose_graph: |
| 1147 | nng.print_graph() |
| 1148 | return nng |