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 |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 28 | from .errors import UnsupportedFeatureError |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 29 | from .ethos_u55_regs.ethos_u55_regs import resampling_mode |
Louis Verhaard | 8912c53 | 2020-09-30 12:11:49 +0200 | [diff] [blame] | 30 | from .numeric_util import clamp_sigmoid |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 31 | from .numeric_util import full_shape |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 32 | from .numeric_util import round_away_zero |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 33 | from .operation import create_avgpool_nop |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 34 | from .operation import NpuBlockType |
| 35 | from .operation import Operation |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 36 | from .softmax import SoftMax |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 37 | from .tensor import create_const_tensor |
| 38 | from .tensor import create_reshape_tensor |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 39 | from .tensor import QuantizationParameters |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 40 | from .tensor import Tensor |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 41 | |
| 42 | passthrough_nodes = set(("Identity",)) |
| 43 | |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 44 | conv_op = set(("Conv2D", "QuantizedConv2D", "Conv2DBackpropInputSwitchedBias", "Conv2DBiasAct")) |
| 45 | fc_op = set( |
| 46 | ( |
| 47 | "MatMul", |
| 48 | "QuantizedMatMul", |
| 49 | "BlockLSTM", |
| 50 | "RnnAct", |
| 51 | "UnidirectionalSequenceRnnAct", |
| 52 | "BidirectionalSequenceRnnAct", |
| 53 | "LstmAct", |
| 54 | "UnidirectionalSequenceLstmAct", |
| 55 | "BidirectionalSequenceLstmAct", |
| 56 | "FullyConnectedAct", |
| 57 | ) |
| 58 | ) |
| 59 | depthwise_op = set(("DepthwiseConv2dNative", "DepthwiseConv2dBiasAct",)) |
| 60 | pool_op = set( |
| 61 | ("AvgPool", "MaxPool", "QuantizedAvgPool", "QuantizedMaxPool", "AvgPoolAct", "MaxPoolAct", "ResizeBilinear") |
| 62 | ) |
| 63 | reduce_sum_ops = set(("ReduceSum",)) |
| 64 | binary_elementwise_op = set(("AddAct", "MulAct", "SubAct", "Maximum", "Minimum")) |
| 65 | elementwise_op = set(("LeakyRelu", "Abs", "CLZ", "SHL", "SHR")) | binary_elementwise_op |
| 66 | relu_ops = set(("Relu", "Relu6", "ReluN1To1")) |
| 67 | activation_ops = set(("Sigmoid", "Tanh")) | relu_ops |
| 68 | memory_only_ops = set(("Reshape",)) |
| 69 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 70 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 71 | def remove_passthrough_tensor(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 72 | if len(tens.ops) == 1 and tens.ops[0].type in passthrough_nodes: |
| 73 | assert len(tens.ops[0].inputs) == 1 |
| 74 | tens = tens.ops[0].inputs[0] |
| 75 | return tens |
| 76 | |
| 77 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 78 | def rewrite_concat(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 79 | if len(tens.ops) == 1 and tens.ops[0].is_concat_op(): |
| 80 | concat_op = tens.ops[0] |
| 81 | if tens != concat_op.outputs[0]: |
| 82 | return tens # don't attempt to rewrite the min/max outputs of QuantizedConcat |
| 83 | |
| 84 | # Not supported so leave it and run on CPU |
| 85 | if not concat_op.run_on_npu: |
| 86 | return tens |
| 87 | |
| 88 | inputs, axis = concat_op.get_concat_inputs_axis() |
| 89 | |
| 90 | tens.ops = [] |
| 91 | offset = 0 |
| 92 | for idx, inp in enumerate(inputs): |
| 93 | new_op = Operation("ConcatSliceWrite", concat_op.name + str(idx)) |
| 94 | new_op.inputs = [inp] |
| 95 | new_op.outputs = [tens] |
| 96 | new_op.attrs["concat_axis"] = axis |
| 97 | new_op.attrs["concat_start"] = offset |
| 98 | offset += inp.shape[axis] |
| 99 | new_op.attrs["concat_end"] = offset |
| 100 | new_op.run_on_npu = True |
| 101 | tens.ops.append(new_op) |
| 102 | assert tens.shape[axis] == offset |
| 103 | |
Patrik Gustavsson | 29d568e | 2020-08-18 10:11:21 +0200 | [diff] [blame] | 104 | # If axis corresponds to C-dimension, NHCWB16 can only be used in the output if all the concat_start's are a |
| 105 | # multiple of 16. This as, it is only then the address offset for the ofm, for all operations, will be 16 byte |
| 106 | # 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] | 107 | # 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] | 108 | if axis == -1 or axis == (len(tens.shape) - 1): |
Patrik Gustavsson | 458a208 | 2020-08-13 13:41:05 +0200 | [diff] [blame] | 109 | for op in tens.ops: |
| 110 | if op.attrs["concat_start"] % 16 != 0: |
| 111 | tens.avoid_NHCWB16 = True |
| 112 | break |
| 113 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 114 | return tens |
| 115 | |
| 116 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 117 | def rewrite_split(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 118 | |
| 119 | if len(tens.ops) == 1 and tens.ops[0].is_split_op(): |
| 120 | split_op = tens.ops[0] |
| 121 | |
| 122 | # Not supported so leave it and run on CPU |
| 123 | if not split_op.run_on_npu: |
| 124 | return tens |
| 125 | |
| 126 | inp, outputs, axis, offset_start, offset_end = split_op.get_split_inputs_axis() |
| 127 | |
| 128 | tens.ops = [] |
| 129 | new_op = Operation("SplitSliceRead", split_op.name) |
| 130 | new_op.inputs = [inp] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 131 | |
| 132 | # For Split the offset cannot be extracted from the tensor so it has to |
| 133 | # be calculated from the index of the output tensor |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 134 | if axis is not None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 135 | # Get the start and end of the split |
| 136 | offset_start = [0] * len(tens.shape) |
| 137 | offset_end = [0] * len(tens.shape) |
| 138 | for out in outputs: |
| 139 | if out == tens: |
| 140 | break |
| 141 | offset_start[axis] += out.shape[axis] |
| 142 | |
Patrik Gustavsson | eebb1c2 | 2020-08-18 15:03:04 +0200 | [diff] [blame] | 143 | # If start offset is not a multiple of 16 in the C-dimension, NHCWB16 need to be avoided in the input |
| 144 | if (offset_start[-1] % 16) != 0: |
| 145 | inp.avoid_NHCWB16 = True |
| 146 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 147 | offset_end[axis] = offset_start[axis] + tens.shape[axis] |
| 148 | |
| 149 | new_op.attrs["split_start"] = offset_start |
| 150 | new_op.attrs["split_end"] = offset_end |
| 151 | new_op.run_on_npu = True |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 152 | new_op.set_output_tensor(tens) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 153 | |
| 154 | return tens |
| 155 | |
| 156 | |
| 157 | def needed_total_padding(input_size, stride, filter_size): |
| 158 | out_size = (input_size + stride - 1) // stride |
| 159 | needed_input = (out_size - 1) * stride + filter_size |
| 160 | total_padding = max(0, needed_input - input_size) |
| 161 | return total_padding |
| 162 | |
| 163 | |
| 164 | def calc_padding_and_skirt(padding_type, kernel_size, stride, input_dims): |
| 165 | ypad = needed_total_padding(int(input_dims[1]), int(stride[1]), int(kernel_size[0])) |
| 166 | xpad = needed_total_padding(int(input_dims[2]), int(stride[2]), int(kernel_size[1])) |
| 167 | if padding_type == b"SAME": |
| 168 | left_pad = (xpad + 0) // 2 |
| 169 | right_pad = (xpad + 1) // 2 |
| 170 | top_pad = (ypad + 0) // 2 |
| 171 | bottom_pad = (ypad + 1) // 2 |
| 172 | elif padding_type == b"VALID": |
| 173 | left_pad = 0 |
| 174 | right_pad = 0 |
| 175 | top_pad = 0 |
| 176 | bottom_pad = 0 |
| 177 | else: |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 178 | raise UnsupportedFeatureError("Unknown padding {}".format(str(padding_type))) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 179 | padding = (top_pad, left_pad, bottom_pad, right_pad) |
| 180 | skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad) |
| 181 | return padding, skirt |
| 182 | |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 183 | |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 184 | def calc_upscaled_padding_and_skirt(padding_type, kernel_size, stride, input_dims, upscaling_factor): |
| 185 | kernel_height, kernel_width = kernel_size[0], kernel_size[1] |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 186 | if padding_type == b"SAME": |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 187 | ypad = needed_total_padding(int(input_dims[1]) * upscaling_factor, int(stride[1]), int(kernel_height)) |
| 188 | xpad = needed_total_padding(int(input_dims[2]) * upscaling_factor, int(stride[2]), int(kernel_width)) |
| 189 | |
Jacob Bohlin | d47cc27 | 2020-08-24 11:42:14 +0200 | [diff] [blame] | 190 | right_pad = max(((xpad + 1) // upscaling_factor) - 1, 0) |
| 191 | bottom_pad = max(((ypad + 1) // upscaling_factor) - 1, 0) |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 192 | left_pad = max(kernel_width - 1 - right_pad, 0) |
| 193 | top_pad = max(kernel_height - 1 - bottom_pad, 0) |
| 194 | |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 195 | elif padding_type == b"VALID": |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 196 | right_pad = max(kernel_width - 2, 0) |
| 197 | bottom_pad = max(kernel_height - 2, 0) |
| 198 | left_pad = kernel_width - 1 |
| 199 | top_pad = kernel_height - 1 |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 200 | else: |
| 201 | assert 0, "Unknown padding" |
| 202 | |
| 203 | padding = (top_pad, left_pad, bottom_pad, right_pad) |
Jacob Bohlin | 9b64ba0 | 2020-07-07 17:15:22 +0200 | [diff] [blame] | 204 | skirt = padding |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 205 | return padding, skirt |
| 206 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 207 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 208 | def fixup_conv2d_backprop(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 209 | if op.type == "Conv2DBackpropInput": |
| 210 | # flip the inputs |
| 211 | op.inputs[0], op.inputs[2] = op.inputs[2], op.inputs[0] |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 212 | op.type = "Conv2DBackpropInputSwitchedBias" |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 213 | |
| 214 | # Update strides |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 215 | 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] | 216 | |
| 217 | return op |
| 218 | |
| 219 | |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 220 | # Convert the op to an elementwise add |
| 221 | def convert_resizebilinear_1x1_to_add(op): |
| 222 | op.type = "AddAct" |
| 223 | op.name = op.name + "_add" |
| 224 | op.attrs.update({"npu_block_type": NpuBlockType.ElementWise}) |
| 225 | op.attrs["resizebilinear"] = True |
| 226 | # Create an input tensor filled with zeros |
| 227 | shape = op.outputs[0].shape |
| 228 | tens = Tensor(shape, op.inputs[0].dtype, op.inputs[1].name + "_add") |
| 229 | tens.values = np.zeros(shape) |
| 230 | tens.quant_values = np.zeros(shape, np.uint8) |
| 231 | tens.quantization = QuantizationParameters(0.0, 255.0) |
| 232 | tens.quantization.scale_f32 = 1.0 |
| 233 | tens.quantization.zero_point = 0 |
| 234 | tens.consumer_list = [op] |
| 235 | tens_op = op.inputs[1].ops[0] |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 236 | tens_op.set_output_tensor(tens) |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 237 | # Set the add inputs |
| 238 | op.inputs[1] = op.inputs[0] |
| 239 | op.inputs[0] = tens |
| 240 | |
| 241 | return op |
| 242 | |
| 243 | |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 244 | # Convert ResizeBilinear to a number of 2x2 pool ops |
| 245 | def convert_resizebilinear_to_2x2_pool(op): |
| 246 | count = 0 |
| 247 | pre_op = op |
| 248 | outputs = op.outputs |
| 249 | |
| 250 | op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)}) |
| 251 | if op.attrs["align_corners"]: |
| 252 | shape_modifier = 1 |
| 253 | op.attrs["padding"] = b"VALID" |
| 254 | else: |
| 255 | shape_modifier = 0 |
| 256 | op.attrs["padding"] = b"SAME" |
| 257 | op.inputs[0].resampling_mode = resampling_mode.NEAREST |
| 258 | |
| 259 | upscaled_shape = np.array(op.inputs[0].shape[1:3]) |
| 260 | out_shape = np.array(op.outputs[0].shape[1:3]) |
| 261 | if (upscaled_shape == upscaled_shape * 2 - shape_modifier).all(): |
| 262 | return op |
| 263 | |
| 264 | while (upscaled_shape < out_shape).all(): |
| 265 | if count == 0: |
| 266 | scaled_op = pre_op |
| 267 | else: |
| 268 | scaled_op = op.clone("_{}".format(count)) |
| 269 | scaled_op.inputs[0] = pre_op.outputs[0] |
| 270 | |
| 271 | upscaled_shape = upscaled_shape * 2 - shape_modifier |
| 272 | |
| 273 | if (upscaled_shape == out_shape).all(): |
| 274 | scaled_op.outputs = outputs |
| 275 | scaled_op.outputs[0].ops = [scaled_op] |
| 276 | else: |
| 277 | shape = outputs[0].shape.copy() |
| 278 | shape[1:3] = upscaled_shape[0:2] |
| 279 | out_tens = Tensor(shape, DataType.int16, "{}_{}".format(op.outputs[0].name, count)) |
| 280 | out_tens.quantization = op.outputs[0].quantization.clone() |
| 281 | out_tens.quantization.quant_min = np.iinfo(np.int16).min |
| 282 | out_tens.quantization.quant_max = np.iinfo(np.int16).max |
| 283 | scaled_op.set_output_tensor(out_tens) |
| 284 | pre_op = scaled_op |
| 285 | count += 1 |
| 286 | |
| 287 | # Setup the scale value |
| 288 | if scaled_op.inputs[0].dtype.bits == 8 and scaled_op.outputs[0].dtype.bits == 16: |
| 289 | scaled_op.attrs["rescale"] = 128 |
| 290 | elif scaled_op.inputs[0].dtype.bits == 16 and scaled_op.outputs[0].dtype.bits == 8: |
| 291 | scaled_op.attrs["rescale"] = 1 / 128 |
| 292 | elif "rescale" in scaled_op.attrs: |
| 293 | del scaled_op.attrs["rescale"] |
| 294 | |
| 295 | return op |
| 296 | |
| 297 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 298 | def fixup_resizebilinear(op, arch, nng): |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 299 | if op.type == "ResizeBilinear" and op.run_on_npu: |
| 300 | if op.inputs[0].shape == op.outputs[0].shape: |
Charles Xu | 36ffaf3 | 2020-08-05 15:40:44 +0200 | [diff] [blame] | 301 | # Bypass nop resizebilinear |
| 302 | op.inputs = op.inputs[:1] |
| 303 | op.type = "Identity" |
Charles Xu | 87c1350 | 2020-08-06 12:17:26 +0200 | [diff] [blame] | 304 | elif op.inputs[0].shape[1] == 1 and op.inputs[0].shape[2] == 1: |
| 305 | convert_resizebilinear_1x1_to_add(op) |
| 306 | else: |
| 307 | convert_resizebilinear_to_2x2_pool(op) |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 308 | |
| 309 | return op |
| 310 | |
| 311 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 312 | def convert_nop_split_to_identity(op, arch, nng): |
Dwight Lidman | c3862c2 | 2020-09-14 15:22:33 +0200 | [diff] [blame] | 313 | if op.type == "Split" and op.attrs.get("num_splits") == 1: |
| 314 | # the list comprehension should return a list with a single tensor |
| 315 | # if it shouldn't, remove_passthrough_tensor will fail appropriately |
| 316 | op.inputs = [i for i in op.inputs if i.shape == op.outputs[0].shape] |
| 317 | op.type = "Identity" |
| 318 | return op |
| 319 | |
| 320 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 321 | def fixup_fully_connected_input(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 322 | if op.type == "FullyConnectedAct": |
| 323 | inp = op.inputs[0] |
| 324 | weights = op.inputs[1] |
| 325 | |
| 326 | n_in_elems = weights.shape[-2] |
| 327 | elms = inp.elements() |
| 328 | batch_size = elms // n_in_elems |
| 329 | assert batch_size * n_in_elems == elms |
| 330 | |
| 331 | desired_shape = [batch_size, n_in_elems] |
| 332 | if inp.shape != desired_shape: |
| 333 | # mismatch, insert a reshape to fix this. |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 334 | op.set_input_tensor(create_reshape_tensor(inp, desired_shape), 0) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 335 | |
| 336 | return op |
| 337 | |
| 338 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 339 | def convert_batched_fc_to_conv(op, arch, nng): |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 340 | if op.type == "FullyConnectedAct": |
| 341 | ifm = op.inputs[0] |
| 342 | ofm = op.outputs[0] |
| 343 | # Check if the FC is 2D and first dimension indicates batching |
| 344 | if len(ifm.shape) == len(ofm.shape) == 2 and ifm.shape[0] != 1: |
| 345 | n = ifm.shape[0] |
| 346 | batching_split = {4: (2, 2), 8: (2, 4), 16: (4, 4)} |
| 347 | h, w = batching_split.get(n, (1, n)) |
| 348 | |
| 349 | # Convert to convolution |
| 350 | op.name += "_conv" |
| 351 | op.type = "Conv2DBiasAct" |
| 352 | faf = op.attrs.get("fused_activation_function", None) |
| 353 | op.attrs = { |
| 354 | "dilation": (1, 1, 1, 1), |
| 355 | "dilation_h_factor": 1, |
| 356 | "dilation_w_factor": 1, |
| 357 | "fused_activation_function": faf, |
| 358 | "npu_block_type": NpuBlockType.ConvolutionMxN, |
| 359 | "padding": b"SAME", |
| 360 | "stride_h": 1, |
| 361 | "stride_w": 1, |
| 362 | "strides": (1, 1, 1, 1), |
| 363 | } |
| 364 | |
| 365 | prev_op = ifm.ops[0] |
| 366 | desired_shape = [1, h, w, ifm.shape[-1]] |
| 367 | if len(ifm.consumer_list) == 1 and prev_op is not None and prev_op.type == "Reshape": |
| 368 | # There is a preceding Reshape |
| 369 | # Compare input of prev_op and input of op, to see if prev_op can be removed |
| 370 | ifm_prev_op = prev_op.inputs[0] |
| 371 | if ifm_prev_op.shape == ifm.shape and ifm_prev_op.quantization.is_scaling_equal(ifm.quantization): |
| 372 | # prev_op can be removed |
| 373 | op.set_input_tensor(ifm_prev_op, 0) |
| 374 | else: |
| 375 | op.inputs[0].set_all_shapes(desired_shape) |
| 376 | prev_op.set_input_tensor( |
| 377 | create_const_tensor(prev_op.inputs[1].name, [1], DataType.int32, desired_shape), 1 |
| 378 | ) |
| 379 | prev_op.attrs["new_shape"] = desired_shape |
| 380 | else: |
| 381 | # Add reshape op to the input if there is no preceding reshape |
| 382 | ifm.consumer_list.remove(op) |
| 383 | op.set_input_tensor(create_reshape_tensor(ifm, desired_shape), 0) |
| 384 | |
| 385 | # Reshape Weights to be 4D. IO becomes HWIO |
| 386 | weight_tensor = op.inputs[1] |
| 387 | weight_tensor.quant_values = np.expand_dims(np.expand_dims(weight_tensor.quant_values, axis=0), axis=0) |
| 388 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
| 389 | |
| 390 | desired_shape = [1, h, w, ofm.shape[-1]] |
| 391 | if ( |
| 392 | len(ofm.consumer_list) == 1 |
| 393 | and ofm.consumer_list[0] is not None |
| 394 | and ofm.consumer_list[0].type == "Reshape" |
| 395 | ): |
| 396 | # There is a subsequent Reshape |
| 397 | # Compare desired shape and output of consumer op, to see if consumer op can be removed |
| 398 | ofm_cons_op = ofm.consumer_list[0].outputs[0] |
| 399 | if desired_shape == ofm_cons_op.shape and ofm.quantization.is_scaling_equal(ofm_cons_op.quantization): |
| 400 | op.outputs[0] = ofm_cons_op |
| 401 | op.outputs[0].ops = [op] |
| 402 | else: |
| 403 | op.outputs[0].set_all_shapes(desired_shape) |
| 404 | else: |
| 405 | # Add rehape op to the output |
| 406 | op.set_output_tensor(create_reshape_tensor(ofm, desired_shape, False)) |
| 407 | return op |
| 408 | |
| 409 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 410 | def fixup_pack_input(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 411 | if op.type == "Pack": |
| 412 | # Pack is also referred to as Stack |
| 413 | # Requires the rewrite_concat function to be called on the op afterwards |
| 414 | axis = int(op.attrs["axis"]) |
| 415 | desired_shape = op.inputs[0].shape[:axis] + [1] + op.inputs[0].shape[axis:] |
| 416 | |
| 417 | # Construct 1 shape tensor to be used by all inserted reshape ops |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 418 | 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] | 419 | |
| 420 | for idx, inp in enumerate(op.inputs): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 421 | reshape_out = inp.clone("_reshaped") |
Michael McGeagh | 6a8d424 | 2020-07-28 12:17:59 +0100 | [diff] [blame] | 422 | reshape_out.set_all_shapes(desired_shape) |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 423 | |
| 424 | reshape_op = Operation("Reshape", "{}{}_reshape".format(op.name, idx)) |
| 425 | reshape_op.attrs["new_shape"] = desired_shape |
| 426 | reshape_op.inputs = [inp, new_shape_tens] |
| 427 | reshape_op.set_output_tensor(reshape_out) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 428 | |
| 429 | op.inputs[idx] = reshape_out |
| 430 | |
| 431 | op.type = "PackReshaped" |
| 432 | |
| 433 | return op |
| 434 | |
| 435 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 436 | def unfuse_activation_function(op, arch, nng): |
Fredrik Svedberg | 0f98b36 | 2020-09-29 10:00:39 +0200 | [diff] [blame] | 437 | unfuse_ops = ("ConcatTFLite",) |
| 438 | if op.type in unfuse_ops and op.run_on_npu and op.attrs.get("fused_activation_function", None) is not None: |
| 439 | act = op.attrs["fused_activation_function"] |
| 440 | del op.attrs["fused_activation_function"] |
| 441 | act_op = Operation(act, op.name + act) |
| 442 | out_tens = op.outputs[0] |
| 443 | intermediate_tens = out_tens.clone("_act_intermediate") |
| 444 | act_op.set_output_tensor(out_tens) |
| 445 | act_op.add_input_tensor(intermediate_tens) |
| 446 | op.set_output_tensor(intermediate_tens) |
| 447 | |
| 448 | return op |
| 449 | |
Louis Verhaard | 8912c53 | 2020-09-30 12:11:49 +0200 | [diff] [blame] | 450 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 451 | def fixup_unpack_output(tens, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 452 | op = tens.ops[0] |
| 453 | if op.type in set(("Unpack", "StridedSlice")): |
| 454 | # Unpack is also referred to as Unstack |
| 455 | # Requires the rewrite_split function to be called on the op afterwards |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 456 | |
| 457 | reshape_input_shape = tens.shape |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 458 | if op.type == "StridedSlice": |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 459 | new_axis_mask = op.attrs["new_axis_mask"] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 460 | shrink_axis_mask = op.attrs["shrink_axis_mask"] |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 461 | ellipsis_mask = op.attrs["ellipsis_mask"] |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 462 | |
| 463 | if (new_axis_mask != 0 and shrink_axis_mask != 0) or ellipsis_mask != 0: |
| 464 | # Not supported, will be put on CPU |
| 465 | return tens |
| 466 | if shrink_axis_mask == 0 and new_axis_mask == 0: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 467 | # Equal Rank StridedSlice, no need to insert reshape |
| 468 | return tens |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 469 | elif shrink_axis_mask != 0: |
| 470 | n = 0 |
| 471 | axis = 0 |
| 472 | while shrink_axis_mask: |
| 473 | prev_mask = shrink_axis_mask |
| 474 | n += 1 |
| 475 | shrink_axis_mask &= shrink_axis_mask - 1 |
| 476 | axis = int(math.log2(prev_mask - shrink_axis_mask)) |
| 477 | reshape_input_shape = reshape_input_shape[:axis] + [1] + reshape_input_shape[axis:] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 478 | |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 479 | assert len(tens.shape) == (len(op.inputs[0].shape) - n) |
| 480 | op.attrs["shrink_axis_mask"] = 0 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 481 | |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 482 | elif new_axis_mask != 0: |
| 483 | n = 0 |
| 484 | axis = 0 |
| 485 | while new_axis_mask: |
| 486 | prev_mask = new_axis_mask |
| 487 | n += 1 |
| 488 | new_axis_mask &= new_axis_mask - 1 |
| 489 | axis = int(math.log2(prev_mask - new_axis_mask)) |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 490 | reshape_input_shape = reshape_input_shape[:axis] + reshape_input_shape[(axis + 1) :] |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 491 | new_axis_mask >>= 1 |
| 492 | |
| 493 | assert len(tens.shape) == (len(op.inputs[0].shape) + n) |
| 494 | op.attrs["new_axis_mask"] = 0 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 495 | else: |
| 496 | axis = int(op.attrs["axis"]) |
| 497 | op.type = "UnpackReshaped" |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 498 | reshape_input_shape = tens.shape[:axis] + [1] + tens.shape[axis:] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 499 | |
| 500 | # Construct 1 shape tensor to be used by all inserted reshape ops |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 501 | 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] | 502 | |
| 503 | for idx, out_tens in enumerate(op.outputs): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 504 | reshape_in = out_tens.clone("_reshaped") |
Michael McGeagh | 6a8d424 | 2020-07-28 12:17:59 +0100 | [diff] [blame] | 505 | reshape_in.set_all_shapes(reshape_input_shape) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 506 | reshape_in.ops = [op] |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 507 | |
| 508 | reshape_op = Operation("Reshape", "{}{}_reshape".format(op.name, idx)) |
| 509 | reshape_op.attrs["new_shape"] = reshape_input_shape |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 510 | reshape_op.inputs = [reshape_in, new_shape_tens] |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 511 | reshape_op.set_output_tensor(out_tens) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 512 | |
| 513 | op.outputs[idx] = reshape_in |
| 514 | |
| 515 | return tens |
| 516 | |
| 517 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 518 | def add_padding_fields(op, arch, nng): |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 519 | if op.run_on_npu: |
| 520 | if "padding" in op.attrs: |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 521 | if op.type in conv_op | depthwise_op: |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 522 | kernel_size = op.inputs[1].shape[:2] |
| 523 | input_shape = op.inputs[0].shape |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 524 | elif op.type in pool_op | reduce_sum_ops: |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 525 | kernel_size = op.attrs["ksize"][1:3] |
| 526 | input_shape = op.inputs[0].shape |
| 527 | elif op.type == "ExtractImagePatches": |
| 528 | kernel_size = op.attrs["ksizes"][1:3] |
| 529 | input_shape = op.inputs[0].shape |
| 530 | else: |
| 531 | raise UnsupportedFeatureError("Unknown operation that uses padding: {}".format(op.type)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 532 | |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 533 | if op.type == "Conv2DBackpropInputSwitchedBias": |
| 534 | upscaling_factor = op.outputs[0].shape[1] // input_shape[1] |
| 535 | padding, skirt = calc_upscaled_padding_and_skirt( |
| 536 | op.attrs["padding"], kernel_size, op.attrs["strides"], input_shape, upscaling_factor |
| 537 | ) |
| 538 | else: |
| 539 | dilation_h, dilation_w = op.get_dilation_h_w() |
| 540 | dilated_kernel_size = [dilation_h * (kernel_size[0] - 1) + 1, dilation_w * (kernel_size[1] - 1) + 1] |
| 541 | padding, skirt = calc_padding_and_skirt( |
| 542 | op.attrs["padding"], dilated_kernel_size, op.attrs["strides"], input_shape |
| 543 | ) |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 544 | |
Jacob Bohlin | 90033f3 | 2020-08-28 15:45:44 +0200 | [diff] [blame] | 545 | op.attrs["explicit_padding"] = padding |
| 546 | op.attrs["skirt"] = skirt |
Jacob Bohlin | cf7da10 | 2020-05-20 09:03:40 +0200 | [diff] [blame] | 547 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 548 | return op |
| 549 | |
| 550 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 551 | # Check if the op can be reordered |
| 552 | def get_prepend_op(op): |
| 553 | inp = op.inputs[0] |
| 554 | # The op should be reordered between prev_op and prep_op |
| 555 | prev_op = inp.ops[-1] |
| 556 | prep_op = None |
| 557 | while prev_op.type in memory_only_ops and len(prev_op.outputs) == 1 and len(prev_op.outputs[0].consumers()) == 1: |
| 558 | prep_op = prev_op |
| 559 | inp = prev_op.inputs[0] |
| 560 | prev_op = inp.ops[-1] |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 561 | 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] | 562 | return prep_op |
| 563 | |
| 564 | return None |
| 565 | |
| 566 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 567 | def mark_npu_block_type(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 568 | npu_block_type = NpuBlockType.Default |
| 569 | if op.type in conv_op: |
| 570 | npu_block_type = NpuBlockType.ConvolutionMxN |
| 571 | elif op.type in fc_op: |
| 572 | npu_block_type = NpuBlockType.VectorProduct |
| 573 | elif op.type in depthwise_op: |
| 574 | npu_block_type = NpuBlockType.ConvolutionDepthWise |
| 575 | elif op.type in pool_op: |
| 576 | npu_block_type = NpuBlockType.Pooling |
| 577 | elif op.type in elementwise_op: |
| 578 | npu_block_type = NpuBlockType.ElementWise |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 579 | elif op.type in reduce_sum_ops: |
| 580 | npu_block_type = NpuBlockType.ReduceSum |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 581 | |
| 582 | op.attrs["npu_block_type"] = npu_block_type |
| 583 | return op |
| 584 | |
| 585 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 586 | def convert_depthwise_to_conv(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 587 | # Depthwise is equivalent to a single conv2d if the ifm depth is 1 and |
| 588 | # the ofm depth equals the depth multipler. |
| 589 | # If those conditions are true, then we can perform a simple |
| 590 | # switch of the operator type (and weight order) |
| 591 | |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 592 | if (op.type in depthwise_op) and (op.attrs["depth_multiplier"] != 1): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 593 | ifm_tensor = op.inputs[0] |
| 594 | weight_tensor = op.inputs[1] |
| 595 | ofm_tensor = op.outputs[0] |
| 596 | if (ifm_tensor.shape[3] == 1) and (ofm_tensor.shape[3] == op.attrs["depth_multiplier"]): |
| 597 | # Change op type to Conv2d |
| 598 | op.type = op.type.replace("DepthwiseConv2d", "Conv2D") |
| 599 | del op.attrs["channel_multiplier"] |
| 600 | del op.attrs["depth_multiplier"] |
| 601 | |
| 602 | 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] | 603 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 604 | else: |
Louis Verhaard | 7db7896 | 2020-05-25 15:05:26 +0200 | [diff] [blame] | 605 | raise UnsupportedFeatureError( |
| 606 | "Unsupported DepthwiseConv2d with depth_multiplier = {}, ifm channels = {}, ofm channels = {}".format( |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 607 | op.attrs["depth_multiplier"], ifm_tensor.shape[3], ofm_tensor.shape[3] |
| 608 | ) |
| 609 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 610 | return op |
| 611 | |
| 612 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 613 | def reorder_depthwise_weights(op, arch, nng): |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 614 | if op.type in depthwise_op: |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 615 | weight_tensor = op.inputs[1] |
| 616 | 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] | 617 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 618 | weight_tensor.weight_transpose_depthwise = True |
| 619 | |
| 620 | return op |
| 621 | |
| 622 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 623 | def convert_conv_to_fc(op, arch, nng): |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 624 | # Conv 1x1 can be equivalent to Fully Connected. |
| 625 | # By representing certain convs as fully connected layers, Vela can better determine wether or not to use |
| 626 | # caching/double buffering for the weights. |
| 627 | # (Weights dont need to be reloaded for convs when IFM H and W are 1) |
| 628 | if op.type == "Conv2DBiasAct": |
| 629 | _, h, w, _ = op.inputs[0].shape |
| 630 | kh, kw, _, _ = op.inputs[1].shape |
| 631 | if h == 1 and w == 1 and kh == 1 and kw == 1: |
| 632 | # Overwrite this op as a Fully Connected Op |
| 633 | op.name += "_fc" |
| 634 | op.type = "FullyConnectedAct" |
| 635 | faf = op.attrs.get("fused_activation_function", None) |
| 636 | op.attrs = { |
| 637 | "fused_activation_function": faf, |
| 638 | "weights_format": 0, |
| 639 | "npu_block_type": NpuBlockType.VectorProduct, |
| 640 | } |
| 641 | # Reshape Weights to be 2D. HWIO becomes just IO (as H and W are 1, they can just be dropped) |
| 642 | weight_tensor = op.inputs[1] |
| 643 | weight_tensor.quant_values = weight_tensor.quant_values.squeeze(axis=(0, 1)) |
| 644 | weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape)) |
| 645 | # The output from a fully connected is expected to be 2D so we need to add a reshape layer to convert it |
| 646 | # back to 4D afterwards as the next layer is expecting that shape |
| 647 | orig_ofm_tensor = op.outputs[0] |
| 648 | # 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}) |
| 649 | fc_ofm_tensor = orig_ofm_tensor.clone("_fc") |
| 650 | fc_ofm_tensor.set_all_shapes([1, fc_ofm_tensor.shape[-1]]) |
| 651 | fc_ofm_tensor.ops = [op] |
| 652 | # 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] | 653 | reshape_name = op.name + "_reshape" |
| 654 | new_shape_tens = create_const_tensor(reshape_name + "_shape", [1], DataType.int32, orig_ofm_tensor.shape) |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 655 | reshape_op = Operation("Reshape", reshape_name) |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 656 | reshape_op.attrs["new_shape"] = orig_ofm_tensor.shape |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 657 | reshape_op.inputs = [fc_ofm_tensor, new_shape_tens] |
| 658 | reshape_op.set_output_tensor(orig_ofm_tensor) |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 659 | # Replace this ops OFM to point to the 2D tensor |
| 660 | op.outputs[0] = fc_ofm_tensor |
| 661 | return op |
| 662 | |
| 663 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 664 | def fixup_relus_with_differing_ifm_ofm_scaling(op, arch, nng): |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 665 | if op.run_on_npu and op.type in relu_ops: |
| 666 | ifm = op.inputs[0] |
| 667 | ofm = op.outputs[0] |
| 668 | # Relu with differing IFM and OFM scaling cannot be fused with another primary op |
| 669 | # and requires its own to be inserted |
| 670 | if not ifm.is_scaling_equal(ofm): |
| 671 | # Override this op with its own primary op (avgpool) |
| 672 | relu_fused_op = create_avgpool_nop(op.name + "_avgpool") |
| 673 | # And fuse the original activation function to it |
| 674 | relu_fused_op.attrs["fused_activation_function"] = op.type |
| 675 | # Tidy up and assign the ifm and ofm to the new op |
| 676 | ifm.consumer_list.remove(op) |
Andreas Nevalainen | f3d737e | 2020-09-25 14:12:43 +0200 | [diff] [blame] | 677 | |
| 678 | # if not 4d, reshape ifm/ofm |
| 679 | if len(ifm.shape) < 4: |
| 680 | ifm_shaped = create_reshape_tensor(ifm, full_shape(4, ifm.shape, 1)) |
| 681 | ifm = ifm_shaped |
| 682 | if len(ofm.shape) < 4: |
| 683 | ofm_shaped = create_reshape_tensor(ofm, full_shape(4, ofm.shape, 1), False) |
| 684 | ofm = ofm_shaped |
| 685 | |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 686 | relu_fused_op.add_input_tensor(ifm) |
| 687 | relu_fused_op.set_output_tensor(ofm) |
| 688 | op = relu_fused_op |
| 689 | return op |
| 690 | |
| 691 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 692 | # Reorder activation op if it's after the memory only operations |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 693 | def fixup_act_reorder(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 694 | if op.type in activation_ops: |
| 695 | prep_op = get_prepend_op(op) |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 696 | if prep_op is not None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 697 | act_op = op.clone("_reordered") |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 698 | |
| 699 | # There is only one input tensor, overwrite it |
| 700 | act_op.set_input_tensor(prep_op.inputs[0], 0) |
| 701 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 702 | act_op_out = act_op.inputs[0].clone("_acted") |
| 703 | act_op_out.quantization = op.outputs[0].quantization.clone() |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 704 | act_op.set_output_tensor(act_op_out) |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 705 | |
| 706 | # Update the consumer list |
| 707 | act_op_out.consumer_list = op.outputs[0].consumer_list.copy() |
| 708 | act_op_out.consumer_list.append(prep_op) |
| 709 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 710 | prep_op.inputs[0] = act_op_out |
| 711 | prep_op.outputs[0].quantization = act_op_out.quantization.clone() |
| 712 | |
| 713 | # Mark the op so that it will be removed as passthrough later on |
| 714 | op.type = "Identity" |
| 715 | return op |
| 716 | |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 717 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 718 | def fixup_elementwise_with_scalars(op, arch, nng): |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 719 | if op.type in binary_elementwise_op: |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 720 | ifm_tensor, ifm2_tensor, _, _ = op.get_ifm_ifm2_weights_ofm() |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 721 | if ifm2_tensor.shape != [] and ifm_tensor.shape != []: |
| 722 | diff = len(ifm_tensor.shape) - len(ifm2_tensor.shape) |
| 723 | if diff > 0: |
| 724 | ifm2_tensor.shape = full_shape(len(ifm_tensor.shape), ifm2_tensor.shape, 1) |
| 725 | elif diff < 0: |
| 726 | 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] | 727 | elif ifm_tensor.shape == [] and ifm_tensor.quant_values is None: |
| 728 | # IFM is marked as a scalar, but is a result of an operation; change it to a shape of size 1 |
| 729 | ifm_tensor.shape = len(ifm2_tensor.shape) * [1] |
| 730 | ifm_tensor.storage_shape = ifm_tensor.shape |
| 731 | elif ifm2_tensor.shape == [] and ifm2_tensor.quant_values is None: |
| 732 | # IFM2 is marked as a scalar, but is a result of an operation; change it to a shape of size 1 |
| 733 | ifm2_tensor.shape = len(ifm_tensor.shape) * [1] |
| 734 | ifm2_tensor.storage_shape = ifm2_tensor.shape |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 735 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 736 | |
Louis Verhaard | e0ef273 | 2020-06-03 08:56:44 +0200 | [diff] [blame] | 737 | |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 738 | # Set input/output tensor equivalence to the same id for memory operations |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 739 | def set_tensor_equivalence(op, arch, nng): |
Michael McGeagh | 11b0bdb | 2020-09-08 11:07:35 +0100 | [diff] [blame] | 740 | if op.type in memory_only_ops: |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 741 | eid = op.outputs[0].equivalence_id |
| 742 | for inp in op.inputs: |
| 743 | inp.equivalence_id = eid |
| 744 | return op |
| 745 | |
| 746 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 747 | def convert_softmax(op, arch, nng): |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 748 | if op.type == "Softmax" and op.run_on_npu: |
| 749 | softmax = SoftMax(op) |
| 750 | op = softmax.get_graph() |
| 751 | return op |
| 752 | |
| 753 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 754 | def convert_mul_max_to_abs_or_lrelu(op, arch, nng): |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 755 | r"""Whenever there is a subgraph with this topology: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 756 | |
| 757 | Input X For X = -1 or X > 0 |
| 758 | | \ / This subgraph can be replaced with either |
| 759 | | Mul an Abs (if X = -1) or a LeakyReLU (if X > 0) |
| 760 | | / |
| 761 | Max |
| 762 | """ |
| 763 | |
| 764 | if op.type == "Maximum": |
| 765 | # finds the Mul input(s) to the Max |
| 766 | muls = [i for i in op.inputs if i.ops[0].type == "MulAct"] |
| 767 | if len(muls) == 1: |
| 768 | mul = muls[0].ops[0] |
| 769 | elif len(muls) == 2: |
| 770 | # In the case both inputs are Muls, find the one with the same input as the Max |
| 771 | mul = [m for m in muls if len(set(op.inputs + m.ops[0].inputs)) == 1][0].ops[0] |
| 772 | else: |
| 773 | # No Mul inputs |
| 774 | return op |
| 775 | |
| 776 | # make sure the Mul doesn't have any other consumers |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 777 | mul_ofm = mul.outputs[0] |
| 778 | if len(mul_ofm.consumers()) != 1: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 779 | return op |
| 780 | # make sure the Mul doesn't have a faf |
| 781 | if mul.attrs["fused_activation_function"]: |
| 782 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 783 | ifm, _, _, ofm = op.get_ifm_weights_biases_ofm() |
| 784 | if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype: |
| 785 | return op |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 786 | if not ifm.is_scaling_equal(ofm) or not ifm.is_scaling_equal(mul_ofm): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 787 | # rewrite to LeakyRelu currently only makes sense if the quantization is identical |
| 788 | return op |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 789 | |
| 790 | # finds the branched input that goes to both the Max and the Mul |
| 791 | shared = set(op.inputs) & set(mul.inputs) |
| 792 | if len(shared) == 1: |
| 793 | shared_in = shared.pop() |
| 794 | # find the constant scalar input to the Mul |
| 795 | const_tens = (set(mul.inputs) - {shared_in}).pop() |
| 796 | # check that it is a scalar |
| 797 | if const_tens.shape != []: |
| 798 | return op |
| 799 | const = const_tens.ops[0] |
| 800 | # check that it is a constant |
| 801 | if const.type != "Const": |
| 802 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 803 | # Remove the Mul from the shared input's consumers |
| 804 | shared_in.consumer_list.remove(mul) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 805 | else: |
| 806 | return op |
| 807 | |
| 808 | val = const.outputs[0].values |
| 809 | if val >= 0: |
| 810 | new_op = "LeakyRelu" |
| 811 | op.attrs["alpha"] = val |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 812 | # to produce bit exact results, the alpha is not enough; |
| 813 | # save additional scaling info in attr "alpha_scale", to be used as input |
| 814 | # to the LUT construction |
| 815 | alpha_scalar = const_tens.quant_values - const_tens.quantization.zero_point |
| 816 | mul_ifm_scale = np.double(ifm.quantization.scale_f32) |
| 817 | mul_ifm2_scale = np.double(const_tens.quantization.scale_f32) |
| 818 | mul_ofm_scale = np.double(mul_ofm.quantization.scale_f32) |
| 819 | alpha_scale, alpha_shift = scaling.elementwise_mul_scale(mul_ifm_scale, mul_ifm2_scale, mul_ofm_scale) |
| 820 | op.attrs["alpha_scaling"] = (alpha_scalar, alpha_scale, alpha_shift) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 821 | elif val == -1: |
| 822 | new_op = "Abs" |
| 823 | else: |
| 824 | return op |
| 825 | |
| 826 | op.type = op.type.replace("Maximum", new_op) |
| 827 | op.name = op.name.replace("Maximum", new_op) |
| 828 | op.outputs[0].name = op.outputs[0].name.replace("Maximum", new_op) |
| 829 | op.inputs = [shared_in] |
| 830 | return op |
| 831 | |
| 832 | |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 833 | def convert_lrelu_to_mul_max(op, arch): |
| 834 | # Converts LeakyRelu to Max(alpha * IFM, identity * IFM) |
| 835 | # (the opposite of convert_mul_max_to_abs_or_lrelu) |
| 836 | ifm, _, _, ofm = op.get_ifm_weights_biases_ofm() |
| 837 | |
| 838 | # Add multiplication with alpha |
| 839 | mul_alpha = Operation("MulAct", op.name + "_mul_alpha") |
| 840 | mul_alpha.add_input_tensor(ifm) |
| 841 | # Create const tensor containing alpha as scalar |
| 842 | alpha = op.attrs["alpha"] |
| 843 | quantization = ifm.quantization.clone() |
| 844 | quantization.min = 0 |
| 845 | quantization.max = alpha * (quantization.quant_max - quantization.quant_min) |
| 846 | quantization.scale_f32 = alpha |
| 847 | quantization.zero_point = 0 |
| 848 | alpha_tens = create_const_tensor(op.name + "_alpha_scalar", [], ifm.dtype, [1], np.int8, quantization=quantization) |
| 849 | mul_alpha.add_input_tensor(alpha_tens) |
| 850 | fm_alpha = ofm.clone(op.name + "_alpha") |
| 851 | mul_alpha.set_output_tensor(fm_alpha) |
| 852 | |
| 853 | if ifm.is_scaling_equal(ofm): |
| 854 | # No identity multiplication is needed |
| 855 | fm_id = ifm |
| 856 | else: |
| 857 | # Add multiplication with identity |
| 858 | mul_identity = Operation("MulAct", op.name + "_mul_identity") |
| 859 | mul_identity.add_input_tensor(ifm) |
| 860 | # Create const tensor containing identity as scalar |
| 861 | quantization = ifm.quantization.clone() |
| 862 | quantization.min = 0 |
| 863 | quantization.max = quantization.quant_max - quantization.quant_min |
| 864 | quantization.scale_f32 = 1 |
| 865 | quantization.zero_point = 0 |
| 866 | identity_tens = create_const_tensor( |
| 867 | op.name + "_id_scalar", [], ifm.dtype, [1], np.uint8, quantization=quantization |
| 868 | ) |
| 869 | mul_identity.add_input_tensor(identity_tens) |
| 870 | fm_id = ofm.clone(op.name + "_id") |
| 871 | mul_identity.set_output_tensor(fm_id) |
| 872 | |
| 873 | # Convert LeakyRelu to Max, add the results of the multiplication(s) as inputs |
| 874 | op.type = "Maximum" |
| 875 | op.name = op.name.replace("LeakyRelu", "Maximum") |
| 876 | op.inputs = [] |
| 877 | ifm.consumer_list.remove(op) |
| 878 | op.add_input_tensor(fm_alpha) |
| 879 | op.add_input_tensor(fm_id) |
| 880 | return op |
| 881 | |
| 882 | |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 883 | def convert_to_lut(op, lut_values): |
| 884 | # Rewrite the operation by Add with scalar 0 + LUT activation |
| 885 | ifm = op.inputs[0] |
Louis Verhaard | 58520b9 | 2020-08-24 16:45:38 +0200 | [diff] [blame] | 886 | assert ifm.dtype.size_in_bytes() == 1 |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 887 | op.type = "AddAct" |
| 888 | op.name = op.name + "_add" |
| 889 | op.attrs.update({"npu_block_type": NpuBlockType.ElementWise}) |
| 890 | # Mark as no-op to enable potential fusing optimizations |
| 891 | op.attrs["is_nop"] = True |
| 892 | # Create an input tensor containing scalar zero |
| 893 | quantization = QuantizationParameters(0.0, 255.0) |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 894 | quantization.scale_f32 = ifm.quantization.scale_f32 |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 895 | quantization.zero_point = 0 |
| 896 | tens = create_const_tensor(op.inputs[0].name + "_add", [], ifm.dtype, [0], np.uint8, quantization=quantization) |
| 897 | op.add_input_tensor(tens) |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 898 | # The LUT must be applied without any preceding rescaling (the LUT itself performs the rescale), |
| 899 | # so even if the OFM has a different scale than the IFM, the generated OFM scale instructions |
| 900 | # should be the same as the IFM |
| 901 | op.attrs["forced_output_quantization"] = ifm.quantization |
| 902 | lut_tensor = lut.create_lut_tensor(op.name + "_lut", lut_values, DataType.int8) |
| 903 | op.set_activation_lut(lut_tensor) |
| 904 | return op |
| 905 | |
| 906 | |
| 907 | def convert_to_lut8(op, fn): |
| 908 | # Converts op to a no-op + int8/uint8 LUT which is generated with the given function. |
| 909 | # fn is a function(real) -> real |
| 910 | ifm, _, _, ofm = op.get_ifm_weights_biases_ofm() |
| 911 | if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype: |
| 912 | return op |
| 913 | # Generate the LUT |
| 914 | ifm_scale = np.double(ifm.quantization.scale_f32) |
| 915 | ofm_scale = np.double(ofm.quantization.scale_f32) |
| 916 | zp_in = ifm.quantization.zero_point |
| 917 | zp_out = ofm.quantization.zero_point |
| 918 | values = [] |
| 919 | ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128) |
| 920 | quantized_min = min(ix) |
| 921 | quantized_max = max(ix) |
| 922 | for x in ix: |
| 923 | x_real = ifm_scale * (x - zp_in) |
| 924 | y_real = fn(x_real) |
| 925 | lut_result = round_away_zero(zp_out + y_real / ofm_scale) |
| 926 | lut_result = min(quantized_max, max(quantized_min, lut_result)) |
| 927 | values.append(lut_result) |
| 928 | return convert_to_lut(op, values) |
| 929 | |
| 930 | |
| 931 | def convert_lrelu_to_lut(op, arch): |
| 932 | ifm, _, _, ofm = op.get_ifm_weights_biases_ofm() |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 933 | # Generate the LUT |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 934 | alpha = op.attrs["alpha"] |
| 935 | ifm_scale = np.double(ifm.quantization.scale_f32) |
| 936 | ofm_scale = np.double(ofm.quantization.scale_f32) |
| 937 | zp_in = ifm.quantization.zero_point |
| 938 | zp_out = ofm.quantization.zero_point |
| 939 | identity_scale, identity_shift = scaling.elementwise_mul_scale(ifm_scale, 1, ofm_scale) |
| 940 | alpha_scalar = 1 |
| 941 | alpha_scale, alpha_shift = scaling.elementwise_mul_scale(ifm_scale, alpha, ofm_scale) |
| 942 | if "alpha_scaling" in op.attrs: |
| 943 | # The LeakyRelu was the result from convert_mul_max_to_abs_or_lrelu |
| 944 | alpha_scalar, alpha_scale, alpha_shift = op.attrs["alpha_scaling"] |
| 945 | values = [] |
Louis Verhaard | 58520b9 | 2020-08-24 16:45:38 +0200 | [diff] [blame] | 946 | ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128) |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 947 | quantized_min = min(ix) |
| 948 | quantized_max = max(ix) |
| 949 | for x in ix: |
| 950 | if x < zp_in: |
| 951 | lut_result = zp_out + fp_math.multiply_by_quantized_multiplier( |
| 952 | alpha_scalar * (x - zp_in), alpha_scale, alpha_shift |
| 953 | ) |
| 954 | else: |
| 955 | lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(x - zp_in, identity_scale, identity_shift) |
| 956 | lut_result = min(quantized_max, max(quantized_min, lut_result)) |
| 957 | values.append(lut_result) |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 958 | return convert_to_lut(op, values) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 959 | |
| 960 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 961 | def convert_lrelu(op, arch, nng): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 962 | # Converts LeakyRelu to a LUT based solution if possible, otherwise a mul + max |
| 963 | if op.type != "LeakyRelu": |
| 964 | return op |
| 965 | ifm, _, _, ofm = op.get_ifm_weights_biases_ofm() |
Louis Verhaard | d7911c4 | 2020-08-25 13:36:41 +0200 | [diff] [blame] | 966 | if ifm.dtype in (DataType.uint8, DataType.int8) and ifm.dtype == ofm.dtype: |
| 967 | # use LUT for int8/uint8 |
| 968 | return convert_lrelu_to_lut(op, arch) |
| 969 | if ifm.is_scaling_equal(ofm) and ifm.dtype == ofm.dtype and ifm.dtype == DataType.int16: |
| 970 | # use LeakyRelu unmodified for int16 with equal input/output scaling |
| 971 | return op |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 972 | return convert_lrelu_to_mul_max(op, arch) |
| 973 | |
| 974 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 975 | def convert_tanh_sigmoid_to_lut(op, arch, nng): |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 976 | # Converts int8/uint8 Sigmoid and Tanh to a LUT based solution |
| 977 | if op.type == "Sigmoid": |
Louis Verhaard | 8912c53 | 2020-09-30 12:11:49 +0200 | [diff] [blame] | 978 | return convert_to_lut8(op, clamp_sigmoid) |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 979 | elif op.type == "Tanh": |
| 980 | return convert_to_lut8(op, math.tanh) |
| 981 | return op |
| 982 | |
| 983 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 984 | def remove_unwanted_reshapes(op, arch, nng): |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 985 | # Try to remove reshapes enclosing ElementWise operator with only one non-constant input |
| 986 | if not op.run_on_npu or op.attrs["npu_block_type"] != NpuBlockType.ElementWise: |
| 987 | return op |
| 988 | |
| 989 | # Check if the ElementWise operator only have one non-constant input |
| 990 | non_const_tens = [x for x in op.inputs if x.ops[0].type != "Const"] |
| 991 | if len(non_const_tens) != 1: |
| 992 | return op |
| 993 | ifm = non_const_tens[0] |
| 994 | |
| 995 | # Check if operation is enclosed by Reshapes that can be removed |
| 996 | ofm = op.outputs[0] |
| 997 | prev_op = ifm.ops[0] |
| 998 | if ( |
| 999 | len(ifm.consumer_list) == 1 |
| 1000 | and prev_op.type == "Reshape" |
| 1001 | and len(ofm.consumer_list) == 1 |
| 1002 | and ofm.consumer_list[0].type == "Reshape" |
| 1003 | ): |
| 1004 | # Operation is enclosed by reshapes, check if they can be removed |
| 1005 | prev_op_ifm, _, _, prev_op_ofm = prev_op.get_ifm_weights_biases_ofm() |
| 1006 | cons_op = ofm.consumer_list[0] |
| 1007 | cons_op_ifm = ofm |
| 1008 | cons_op_ofm = cons_op.outputs[0] |
| 1009 | if len(prev_op_ifm.shape) == len(cons_op_ofm.shape): |
| 1010 | # Check if quantization is the same in the input and output for the reshape ops |
| 1011 | if prev_op_ifm.quantization.is_scaling_equal( |
| 1012 | prev_op_ofm.quantization |
| 1013 | ) and cons_op_ifm.quantization.is_scaling_equal(cons_op_ofm.quantization): |
Patrik Gustavsson | 7ad862a | 2020-09-29 14:09:43 +0200 | [diff] [blame] | 1014 | op.set_input_tensor(prev_op_ifm, 0) |
| 1015 | op.set_output_tensor(cons_op_ofm) |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 1016 | return op |
| 1017 | |
| 1018 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 1019 | def fuse_activation_function_with_prev(op, arch, nng): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1020 | # if op is a no-op: attempts to move the activation function to the preceding op |
| 1021 | if not op.attrs.get("is_nop", False) or op.attrs.get("fused_activation_function", None) is None: |
| 1022 | return op |
| 1023 | ifm, _, _, ofm = op.get_ifm_weights_biases_ofm() |
| 1024 | # finds the input(s) to the operation |
| 1025 | prev_op = ifm.ops[0] |
| 1026 | # Note: the below checks on prev_op require that a first optimize pass on the full graph has been performed |
| 1027 | fuse = ( |
| 1028 | prev_op.run_on_npu |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 1029 | and "npu_block_type" in prev_op.attrs |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1030 | and prev_op.attrs["npu_block_type"] != NpuBlockType.Default |
| 1031 | and len(ifm.ops) == 1 |
| 1032 | and len(prev_op.outputs[0].consumers()) == 1 |
| 1033 | and prev_op.attrs.get("fused_activation_function", None) is None |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1034 | ) |
| 1035 | if op.activation_lut is not None and arch.shram_reserved_unused_banks == 0: |
| 1036 | # TODO: if SHRAM LUT space is shared with SHRAM ACC (32, 64 MAC), |
| 1037 | # LUT currently only works correctly for elementwise ops |
| 1038 | fuse = False |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1039 | if not fuse: |
| 1040 | return op |
| 1041 | # Move the fused activation function + corresponding info to prev_op |
Louis Verhaard | 98a3499 | 2020-09-01 10:39:04 +0200 | [diff] [blame] | 1042 | for attr in ("fused_activation_function", "forced_output_quantization"): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1043 | if attr in op.attrs: |
| 1044 | prev_op.attrs[attr] = op.attrs[attr] |
| 1045 | if op.activation_lut is not None: |
| 1046 | prev_op.set_activation_lut(op.activation_lut) |
| 1047 | # Bypass op |
Louis Verhaard | 98a3499 | 2020-09-01 10:39:04 +0200 | [diff] [blame] | 1048 | prev_op.set_output_tensor(ofm) |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1049 | return op |
| 1050 | |
| 1051 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 1052 | def add_attrs_to_resizebilinear(op, arch, nng): |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 1053 | if op.type == "ResizeBilinear" and op.run_on_npu: |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1054 | input_tensor = op.inputs[0] |
| 1055 | upscaled_shape = [input_tensor.shape[1] * 2, input_tensor.shape[2] * 2] |
| 1056 | out_shape = op.outputs[0].shape[1:3] |
| 1057 | if not op.attrs["align_corners"] and out_shape == upscaled_shape: |
| 1058 | # this means the output is supposed to be a x2 upscale, |
| 1059 | # so we need to do SAME padding |
| 1060 | op.attrs["padding"] = b"SAME" |
| 1061 | elif op.attrs["align_corners"] and out_shape == [upscaled_shape[0] - 1, upscaled_shape[1] - 1]: |
| 1062 | # here we can just run the avg pool without padding and |
| 1063 | # produce a (M * 2 - 1, N * 2 - 1) sized output |
| 1064 | op.attrs["padding"] = b"VALID" |
| 1065 | else: |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 1066 | return op |
Dwight Lidman | 42fed94 | 2020-05-29 09:37:03 +0200 | [diff] [blame] | 1067 | input_tensor.resampling_mode = resampling_mode.NEAREST |
Tim Hall | c30f495 | 2020-06-15 20:47:35 +0100 | [diff] [blame] | 1068 | 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] | 1069 | return op |
| 1070 | |
| 1071 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 1072 | def fixup_bias_tensors(op, arch, nng): |
Jacob Bohlin | a41cd4d | 2020-08-26 18:21:28 +0200 | [diff] [blame] | 1073 | if op.needs_bias() and not op.inputs[-1]: |
| 1074 | # Op has no bias, add bias tensor filled with zeros |
| 1075 | nr_biases = op.inputs[1].shape[-1] |
| 1076 | bias_values = [0] * nr_biases |
| 1077 | bias_tensor = create_const_tensor(op.name + "_bias", [nr_biases], DataType.int32, bias_values) |
| 1078 | bias_tensor.quant_values = bias_tensor.values |
| 1079 | op.set_input_tensor(bias_tensor, -1) |
Jacob Bohlin | 67e0d8f | 2020-08-20 10:53:02 +0200 | [diff] [blame] | 1080 | |
| 1081 | return op |
| 1082 | |
| 1083 | |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 1084 | def supported_operator_check(op, arch, nng): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1085 | op.run_on_npu = arch.supported_operators.is_operator_supported(op) |
| 1086 | return op |
| 1087 | |
| 1088 | |
| 1089 | def optimise_graph_a(nng, arch, verbose_graph=False): |
| 1090 | if verbose_graph: |
| 1091 | nng.print_graph() |
| 1092 | |
| 1093 | op_rewrite_list = [ |
| 1094 | # mark block type and check if the operations are supported |
| 1095 | mark_npu_block_type, |
Tim Hall | 4e12776 | 2020-05-15 16:05:49 +0100 | [diff] [blame] | 1096 | set_tensor_equivalence, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1097 | supported_operator_check, |
| 1098 | # then do any rewrites of supported operators |
| 1099 | convert_depthwise_to_conv, |
Michael McGeagh | 8d939c0 | 2020-07-29 13:11:43 +0100 | [diff] [blame] | 1100 | convert_conv_to_fc, |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 1101 | convert_softmax, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1102 | fixup_fully_connected_input, |
Patrik Gustavsson | cb33704 | 2020-09-16 14:55:40 +0200 | [diff] [blame] | 1103 | convert_batched_fc_to_conv, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1104 | fixup_pack_input, |
Fredrik Svedberg | 0f98b36 | 2020-09-29 10:00:39 +0200 | [diff] [blame] | 1105 | unfuse_activation_function, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1106 | fixup_conv2d_backprop, |
Michael McGeagh | 8dbf8cf | 2020-09-08 11:09:48 +0100 | [diff] [blame] | 1107 | fixup_relus_with_differing_ifm_ofm_scaling, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1108 | fixup_act_reorder, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1109 | mark_npu_block_type, |
Charles Xu | 7879222 | 2020-05-13 10:15:26 +0200 | [diff] [blame] | 1110 | fixup_elementwise_with_scalars, |
Jacob Bohlin | e843d33 | 2020-06-23 12:12:56 +0200 | [diff] [blame] | 1111 | reorder_depthwise_weights, |
Charles Xu | 9a03fdf | 2020-07-02 15:12:40 +0200 | [diff] [blame] | 1112 | fixup_resizebilinear, |
Jacob Bohlin | a41cd4d | 2020-08-26 18:21:28 +0200 | [diff] [blame] | 1113 | fixup_bias_tensors, |
Dwight Lidman | c3862c2 | 2020-09-14 15:22:33 +0200 | [diff] [blame] | 1114 | convert_nop_split_to_identity, |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1115 | convert_mul_max_to_abs_or_lrelu, |
Patrik Gustavsson | fa4cb29 | 2020-09-10 08:19:36 +0200 | [diff] [blame] | 1116 | remove_unwanted_reshapes, |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1117 | convert_lrelu, |
Louis Verhaard | f03bad3 | 2020-09-25 08:30:44 +0200 | [diff] [blame] | 1118 | convert_tanh_sigmoid_to_lut, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1119 | ] |
| 1120 | |
| 1121 | for idx, sg in enumerate(nng.subgraphs): |
| 1122 | # rewrite graph pass |
| 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, [fixup_unpack_output], op_rewrite_list, rewrite_unsupported=False |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1125 | ) |
| 1126 | |
| 1127 | for idx, sg in enumerate(nng.subgraphs): |
Louis Verhaard | b9fc33c | 2020-08-13 11:47:36 +0200 | [diff] [blame] | 1128 | # remove passthrough tensors and attempt further optimizations |
| 1129 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 1130 | 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] | 1131 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1132 | |
| 1133 | if verbose_graph: |
| 1134 | nng.print_graph() |
| 1135 | return nng |
| 1136 | |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 1137 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1138 | def optimise_graph_b(nng, arch, verbose_graph=False): |
| 1139 | if verbose_graph: |
| 1140 | nng.print_graph() |
| 1141 | |
| 1142 | for idx, sg in enumerate(nng.subgraphs): |
| 1143 | # combined rewrite graph pass |
Patrik Gustavsson | 3010d9b | 2020-10-01 08:22:10 +0200 | [diff] [blame^] | 1144 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(nng, sg, arch, [rewrite_concat, rewrite_split], []) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1145 | |
| 1146 | if verbose_graph: |
| 1147 | nng.print_graph() |
| 1148 | return nng |