|
4 | 4 | # LICENSE file in the root directory of this source tree. |
5 | 5 |
|
6 | 6 | from copy import copy |
7 | | -from typing import cast, Set, Type |
| 7 | +from typing import cast, Dict, Optional, Set, Tuple, Type |
8 | 8 |
|
9 | | -from executorch.backends.arm._passes.arm_pass_utils import create_node |
| 9 | +import torch |
| 10 | +from executorch.backends.arm._passes.arm_pass import ArmPass |
| 11 | +from executorch.backends.arm._passes.arm_pass_utils import create_node, set_node_arg |
| 12 | +from executorch.backends.arm._passes.fold_qdq_with_annotated_qparams_pass import ( |
| 13 | + get_output_qparams, |
| 14 | +) |
10 | 15 | from executorch.backends.arm._passes.quant_args import QuantArgs |
11 | 16 | from executorch.backends.arm.constants import DQ_OPS, Q_OPS |
12 | 17 | from executorch.exir.dialects._ops import ops as exir_ops |
@@ -65,3 +70,234 @@ def call(self, graph_module: GraphModule) -> PassResult: |
65 | 70 | graph_module = super().call(graph_module).graph_module |
66 | 71 | graph_module.recompile() |
67 | 72 | return PassResult(graph_module, modified) |
| 73 | + |
| 74 | + |
| 75 | +class InsertRescaleInt32Pass(ArmPass): |
| 76 | + """ |
| 77 | + Numerous TOSA ops require inputs and outputs to be 32-bit integers in their |
| 78 | + quantized implementations. This pass treats such operator nodes by |
| 79 | + inserting rescale ops before and after them if needed. Note that extra logic |
| 80 | + that handles the scales and zero points must be in place because the affected |
| 81 | + TOSA have naive implementations that do not account for the quantization |
| 82 | + parameters. |
| 83 | + """ |
| 84 | + |
| 85 | + _passes_required_after: Set[Type[ExportPass]] = set() |
| 86 | + |
| 87 | + included_targets = [ |
| 88 | + exir_ops.edge.aten.abs.default, |
| 89 | + exir_ops.edge.aten.eq.Tensor, |
| 90 | + exir_ops.edge.aten.ge.Tensor, |
| 91 | + exir_ops.edge.aten.gt.Tensor, |
| 92 | + exir_ops.edge.aten.le.Tensor, |
| 93 | + exir_ops.edge.aten.lt.Tensor, |
| 94 | + exir_ops.edge.aten.maximum.default, |
| 95 | + exir_ops.edge.aten.minimum.default, |
| 96 | + ] |
| 97 | + |
| 98 | + def _int32_qargs(self, s): |
| 99 | + """Helper creator function for INT32-based QuantArgs""" |
| 100 | + |
| 101 | + return QuantArgs( |
| 102 | + scale=s, |
| 103 | + zp=0, |
| 104 | + qmin=torch.iinfo(torch.int32).min, |
| 105 | + qmax=torch.iinfo(torch.int32).max, |
| 106 | + dtype=torch.int32, |
| 107 | + ) |
| 108 | + |
| 109 | + def _get_inputs_rescaled_qparams( |
| 110 | + self, target, input_qparams: Dict[int, QuantArgs] |
| 111 | + ) -> Dict[int, QuantArgs]: |
| 112 | + """Get the qparams for the INT32 operands to the op ``target`` |
| 113 | +
|
| 114 | + Inputs to the INT32-based operator must be rescaled from INT8 to INT32. |
| 115 | + This function computes the ``QuantArgs`` for each of the operands and returns |
| 116 | + it as a dict, mapping tensor index to ``QuantArgs``. |
| 117 | + """ |
| 118 | + |
| 119 | + if target in [ |
| 120 | + exir_ops.edge.aten.abs.default, |
| 121 | + exir_ops.edge.aten.eq.Tensor, |
| 122 | + exir_ops.edge.aten.ge.Tensor, |
| 123 | + exir_ops.edge.aten.gt.Tensor, |
| 124 | + exir_ops.edge.aten.le.Tensor, |
| 125 | + exir_ops.edge.aten.lt.Tensor, |
| 126 | + exir_ops.edge.aten.minimum.default, |
| 127 | + exir_ops.edge.aten.maximum.default, |
| 128 | + ]: |
| 129 | + # For these ops, use the smallest scale among the INT8 operands. |
| 130 | + min_scale = min( |
| 131 | + [qp.get_scale_per_tensor() for qp in input_qparams.values()] |
| 132 | + ) |
| 133 | + qparams = { |
| 134 | + i: self._int32_qargs(min_scale) for i in range(len(input_qparams)) |
| 135 | + } |
| 136 | + else: |
| 137 | + raise ValueError(f"Not a valid target: {target}") |
| 138 | + |
| 139 | + return qparams |
| 140 | + |
| 141 | + def _get_output_qparams( |
| 142 | + self, target, inputs_qparams: Dict[int, QuantArgs] |
| 143 | + ) -> Optional[QuantArgs]: |
| 144 | + """Given an op ``target`` and the ``QuantArgs`` for each of its inputs, compute |
| 145 | + the scale of the output based on how the operator itself affects it.""" |
| 146 | + |
| 147 | + if target in [ |
| 148 | + exir_ops.edge.aten.abs.default, |
| 149 | + exir_ops.edge.aten.maximum.default, |
| 150 | + exir_ops.edge.aten.minimum.default, |
| 151 | + ]: |
| 152 | + # The op has not altered the scale; the output scale is equal to |
| 153 | + # the operands' scales. |
| 154 | + return self._int32_qargs(inputs_qparams[0].get_scale_per_tensor()) |
| 155 | + elif target in [ |
| 156 | + exir_ops.edge.aten.eq.Tensor, |
| 157 | + exir_ops.edge.aten.ge.Tensor, |
| 158 | + exir_ops.edge.aten.gt.Tensor, |
| 159 | + exir_ops.edge.aten.le.Tensor, |
| 160 | + exir_ops.edge.aten.lt.Tensor, |
| 161 | + ]: |
| 162 | + # Output is bool for these ops and thus no qparams are present |
| 163 | + return None |
| 164 | + else: |
| 165 | + raise ValueError(f"Not a valid target: {target}") |
| 166 | + |
| 167 | + def _get_rescale_qparams( |
| 168 | + self, target, input_qparams: Dict[int, QuantArgs] |
| 169 | + ) -> Tuple[Dict[int, QuantArgs], Optional[QuantArgs]]: |
| 170 | + """ |
| 171 | + Get the quantization parameters of the INT32 inputs/outputs that will |
| 172 | + surround the node after the new RESCALE ops have been inserted. |
| 173 | + """ |
| 174 | + |
| 175 | + inputs_rescaled_qparams = self._get_inputs_rescaled_qparams( |
| 176 | + target, input_qparams |
| 177 | + ) |
| 178 | + output_qparams = self._get_output_qparams(target, inputs_rescaled_qparams) |
| 179 | + |
| 180 | + return (inputs_rescaled_qparams, output_qparams) |
| 181 | + |
| 182 | + def _rescale_inputs(self, graph, node, rescale_qargs: Dict[int, QuantArgs]) -> bool: |
| 183 | + qargs = node.meta["input_qparams"] |
| 184 | + |
| 185 | + args_copy = list(node.args) |
| 186 | + seen_args = set() |
| 187 | + modified = False |
| 188 | + for i in qargs: |
| 189 | + qp = qargs[i] |
| 190 | + if qp.dtype != torch.int8: |
| 191 | + continue |
| 192 | + |
| 193 | + arg_node = args_copy[i] |
| 194 | + if arg_node in seen_args: |
| 195 | + continue |
| 196 | + seen_args.add(arg_node) |
| 197 | + |
| 198 | + with graph.inserting_after(arg_node): |
| 199 | + rescale_node = create_node( |
| 200 | + graph, |
| 201 | + exir_ops.backend.tosa.RESCALE.default, |
| 202 | + ( |
| 203 | + arg_node, |
| 204 | + torch.int32, |
| 205 | + qp.get_scale_per_tensor() |
| 206 | + / rescale_qargs[ |
| 207 | + i |
| 208 | + ].get_scale_per_tensor(), # Old scale / new scale |
| 209 | + qp.get_zp_per_tensor(), # Old zero point |
| 210 | + rescale_qargs[i].get_zp_per_tensor(), # New zero point |
| 211 | + ), |
| 212 | + from_node=node, |
| 213 | + ) |
| 214 | + |
| 215 | + node.replace_input_with(arg_node, rescale_node) |
| 216 | + modified = True |
| 217 | + |
| 218 | + return modified |
| 219 | + |
| 220 | + def _rescale_outputs(self, graph, node, rescale_qargs: Optional[QuantArgs]) -> bool: |
| 221 | + if "output_qparams" not in node.meta or len(node.meta["output_qparams"]) == 0: |
| 222 | + return False |
| 223 | + |
| 224 | + qargs = get_output_qparams(node) |
| 225 | + assert len(qargs) == 1 |
| 226 | + assert rescale_qargs is not None |
| 227 | + |
| 228 | + qarg = qargs[0] |
| 229 | + if qarg.dtype != torch.int8: |
| 230 | + return False |
| 231 | + |
| 232 | + users_copy = list(node.users) |
| 233 | + |
| 234 | + with graph.inserting_after(node): |
| 235 | + rescale_node = create_node( |
| 236 | + graph, |
| 237 | + exir_ops.backend.tosa.RESCALE.default, |
| 238 | + ( |
| 239 | + node, |
| 240 | + torch.int8, |
| 241 | + rescale_qargs.get_scale_per_tensor() |
| 242 | + / qarg.get_scale_per_tensor(), # Old scale / new scale |
| 243 | + rescale_qargs.get_zp_per_tensor(), # Old zero point |
| 244 | + qarg.get_zp_per_tensor(), # New zero point |
| 245 | + ), |
| 246 | + from_node=node, |
| 247 | + ) |
| 248 | + |
| 249 | + for user in users_copy: |
| 250 | + user.replace_input_with(node, rescale_node) |
| 251 | + |
| 252 | + return True |
| 253 | + |
| 254 | + def call(self, graph_module: GraphModule) -> PassResult: |
| 255 | + graph = graph_module.graph |
| 256 | + |
| 257 | + modified = False |
| 258 | + for node in list(graph.nodes): |
| 259 | + node = cast(Node, node) |
| 260 | + |
| 261 | + if node.op != "call_function" or node.target not in self.included_targets: |
| 262 | + continue |
| 263 | + |
| 264 | + if "input_qparams" not in node.meta or len(node.meta["input_qparams"]) == 0: |
| 265 | + continue |
| 266 | + input_qparams = node.meta["input_qparams"] |
| 267 | + |
| 268 | + inputs_rescale_qargs, output_rescale_qargs = self._get_rescale_qparams( |
| 269 | + node.target, input_qparams |
| 270 | + ) |
| 271 | + |
| 272 | + inputs_was_rescaled = self._rescale_inputs( |
| 273 | + graph, node, inputs_rescale_qargs |
| 274 | + ) |
| 275 | + outputs_was_rescaled = False |
| 276 | + if inputs_was_rescaled: |
| 277 | + outputs_was_rescaled = self._rescale_outputs( |
| 278 | + graph, node, output_rescale_qargs |
| 279 | + ) |
| 280 | + modified = True |
| 281 | + |
| 282 | + # Update node metadata |
| 283 | + |
| 284 | + if inputs_was_rescaled: |
| 285 | + assert len(inputs_rescale_qargs) == len(node.meta["input_qparams"]) |
| 286 | + node.meta["input_qparams"] = inputs_rescale_qargs |
| 287 | + |
| 288 | + if outputs_was_rescaled: |
| 289 | + assert len(node.meta["output_qparams"]) == 1 |
| 290 | + node.meta["output_qparams"] = {0: output_rescale_qargs} |
| 291 | + |
| 292 | + # If the output type is specified in the node, change it such |
| 293 | + # that it matches the subsequent rescale node(s) that this node |
| 294 | + # now has output edges to. |
| 295 | + if "dtype" in node.kwargs: |
| 296 | + set_node_arg(node, "dtype", torch.int32) |
| 297 | + |
| 298 | + if modified: |
| 299 | + # Retrace the graph to update the fake tensor types |
| 300 | + graph_module = super().call(graph_module).graph_module |
| 301 | + graph_module.recompile() |
| 302 | + |
| 303 | + return PassResult(graph_module, modified) |
0 commit comments