diff options
Diffstat (limited to 'ethosu/vela')
-rw-r--r-- | ethosu/vela/compiler_driver.py | 2 | ||||
-rw-r--r-- | ethosu/vela/debug_database.py | 79 | ||||
-rw-r--r-- | ethosu/vela/graph_optimiser_util.py | 6 | ||||
-rw-r--r-- | ethosu/vela/tflite_graph_optimiser.py | 19 |
4 files changed, 69 insertions, 37 deletions
diff --git a/ethosu/vela/compiler_driver.py b/ethosu/vela/compiler_driver.py index 014d12ae..439b954a 100644 --- a/ethosu/vela/compiler_driver.py +++ b/ethosu/vela/compiler_driver.py @@ -134,7 +134,7 @@ def next_sram_factor(alloc_results): def _record_operator(op, arch): - if op.type != Op.Const: + if op.type not in (Op.Const, Op.Placeholder): DebugDatabase.add_source(op) diff --git a/ethosu/vela/debug_database.py b/ethosu/vela/debug_database.py index bd45b468..c1389e51 100644 --- a/ethosu/vela/debug_database.py +++ b/ethosu/vela/debug_database.py @@ -1,4 +1,4 @@ -# SPDX-FileCopyrightText: Copyright 2020-2021 Arm Limited and/or its affiliates <open-source-office@arm.com> +# SPDX-FileCopyrightText: Copyright 2020-2022 Arm Limited and/or its affiliates <open-source-office@arm.com> # # SPDX-License-Identifier: Apache-2.0 # @@ -61,39 +61,58 @@ class DebugDatabase: [uid, str(op.type), op.kernel.width, op.kernel.height, ofm_shape[-2], ofm_shape[-3], ofm_shape[-1]] ) + # Ops are added when their type changes, and after optimisation. If an op was already + # added before optimisation was finished it will only be added again if it's entry + # has changed in any way from it's previous entry. @classmethod def add_optimised(cls, parent: Operation, op: Operation): assert isinstance(parent, Operation) and isinstance(op, Operation) - if op not in cls._optimisedUID: - if parent not in cls._sourceUID: - # The the parent wasn't in the source network try to look it - # up in the optimised network and use that op's source parent. - if parent in cls._optimisedUID: - src_uid = cls._optimisedUID[parent][1] - else: - if DebugDatabase.show_warnings: - print("Debug Database: Associated parent '{0}' not in network".format(parent.type)) - src_uid = DebugDatabase.NULLREF - else: - src_uid = cls._sourceUID[parent] - uid = len(cls._optimisedUID) - cls._optimisedUID[op] = (uid, src_uid) - if len(op.ofm_shapes) == 0: - ofm_shape = Shape4D(op.outputs[0].shape) + if parent not in cls._sourceUID: + # If the parent wasn't in the source network try to look it + # up in the optimised network and use that op's source parent. + if parent in cls._optimisedUID: + src_uid = cls._optimisedUID[parent][1] else: - ofm_shape = op.ofm_shapes[0] - cls._optimisedTable.append( - [ - uid, - src_uid, - str(op.type), - op.kernel.width, - op.kernel.height, - ofm_shape.width, - ofm_shape.height, - ofm_shape.depth, - ] - ) + if DebugDatabase.show_warnings: + print("Debug Database: Associated parent '{0}' not in network".format(parent.type)) + src_uid = DebugDatabase.NULLREF + else: + src_uid = cls._sourceUID[parent] + + # correction for missing shapes + if len(op.ofm_shapes) == 0: + ofm_shape = Shape4D(op.outputs[0].shape) + else: + ofm_shape = op.ofm_shapes[0] + + next_uid = len(cls._optimisedTable) # required because no longer 1:1 UID->table correspondence + opt_uid = cls._optimisedUID.get(op, (next_uid, 0))[0] # already seen or next uid (if not seen) + + opt_table_entry = [ + opt_uid, + src_uid, + str(op.type), + op.kernel.width, + op.kernel.height, + ofm_shape.width, + ofm_shape.height, + ofm_shape.depth, + ] + + if op not in cls._optimisedUID: + # optimised op does not exist + cls._optimisedUID[op] = (next_uid, src_uid) + cls._optimisedTable.append(opt_table_entry) + else: + # optimised op already exists + existing_entry = cls._optimisedTable[ + cls._optimisedUID[op][0] + ] # Existing entry is where the 'op' object was last inserted + if opt_table_entry != existing_entry: + # only add again if it's changed in any way + opt_table_entry[0] = next_uid # give it a new unique id (required) + cls._optimisedUID[op] = (next_uid, src_uid) + cls._optimisedTable.append(opt_table_entry) @classmethod def add_stream(cls, key): diff --git a/ethosu/vela/graph_optimiser_util.py b/ethosu/vela/graph_optimiser_util.py index fde01cfe..4e233c44 100644 --- a/ethosu/vela/graph_optimiser_util.py +++ b/ethosu/vela/graph_optimiser_util.py @@ -257,7 +257,6 @@ def move_splitsliceread_to_consumer(op, cons_op): cons_op.read_shapes[1] = op.read_shapes[0] cons_op.set_input_tensor(op.ifm, cons_op.type.info.indices.ifms[1]) cons_op.ifm_shapes[1] = op.ifm_shapes[0] - op.ofm.consumer_list.remove(cons_op) op.ofm.ops = [] op.ifm.consumer_list.remove(op) @@ -270,7 +269,7 @@ def check_memory_only_removed(op, arch): def record_optimised(op, arch): - if op.type != Op.Const: + if op.type not in (Op.Const, Op.Placeholder): DebugDatabase.add_optimised(op, op) @@ -392,12 +391,12 @@ def convert_depthwise_to_conv(op, arch, nng): weight_tensor.values = np.transpose(weight_tensor.values, (0, 1, 3, 2)) weight_tensor.set_all_shapes(list(weight_tensor.values.shape)) + DebugDatabase.add_optimised(op, op) else: raise UnsupportedFeatureError( f"Unsupported 'DEPTHWISE_CONV_2D' with depth_multiplier = {op.attrs['depth_multiplier']},", f" ifm channels = {ifm_shape.depth}, ofm channels = {ofm_shape.depth}", ) - DebugDatabase.add_optimised(op, op) return op @@ -426,4 +425,5 @@ def convert_to_lut(op, lut_values, lut_name): lut_tensor = lut.create_lut_tensor(op.name + "_values", lut_values, DataType.int8) op.set_activation_lut(lut_tensor) op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, op) return op diff --git a/ethosu/vela/tflite_graph_optimiser.py b/ethosu/vela/tflite_graph_optimiser.py index b858f648..36c1de5a 100644 --- a/ethosu/vela/tflite_graph_optimiser.py +++ b/ethosu/vela/tflite_graph_optimiser.py @@ -76,10 +76,10 @@ def create_avg_pool_for_concat(concat_op, name, ifm, ifm_shape: Shape4D, write_o avgpool_op.write_offset = write_offset avgpool_op.write_shape = ifm_shape ofm.ops.append(avgpool_op) - DebugDatabase.add_optimised(concat_op, avgpool_op) avgpool_op.ifm_shapes.append(ifm_shape) avgpool_op.ofm_shapes.append(concat_op.ofm_shapes[0]) avgpool_op.memory_function = Op.ConcatSliceWrite + DebugDatabase.add_optimised(concat_op, avgpool_op) return avgpool_op @@ -279,6 +279,7 @@ def fixup_conv2d_backprop(op, arch, nng): # Update strides op.attrs.update({"stride_w": 1, "stride_h": 1, "strides": (1, 1, 1, 1)}) + DebugDatabase.add_optimised(op, op) return op @@ -301,6 +302,7 @@ def convert_resize_1x1_to_add(op): op.inputs[1] = op.inputs[0] op.inputs[0] = tens op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, op) return op @@ -390,6 +392,7 @@ def convert_resizenn_ac_to_depthwise_conv(op, upscale_factor): # finally update the shape incase we've change the tensor shapes or connections op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, op) return op @@ -433,6 +436,7 @@ def convert_resize_to_upscale_and_average_pool(op): pre_op = scaled_op scaled_op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, scaled_op) # Last x2 upscaling if n > 1: @@ -463,6 +467,7 @@ def convert_resize_to_upscale_and_average_pool(op): scaled_op.outputs = outputs scaled_op.outputs[0].ops = [scaled_op] scaled_op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, scaled_op) return op @@ -531,6 +536,7 @@ def convert_resizebilinear_to_depthwise_convolutions(op, half_pixel_centers=True avgpool_op.add_input_tensor(ifm) avgpool_op.set_output_tensor(intermediate_tens) avgpool_op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, op) dw_conv = Operation(Op.DepthwiseConv2DBias, "depthwise_conv") dw_conv._original_type = Op.ResizeBilinear @@ -592,6 +598,8 @@ def convert_resizebilinear_to_depthwise_convolutions(op, half_pixel_centers=True fixup_bias_tensors(dw_conv, None, None, dtype=DataType.int32) dw_conv.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, dw_conv) + dw_conv = dw_conv.clone(f"_{index}") return op @@ -674,6 +682,7 @@ def unfuse_activation_function(op): act_op.add_input_tensor(intermediate_tens) op.set_output_tensor(intermediate_tens) act_op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, act_op) def rewrite_stridedslice_output(op, arch, nng): @@ -955,6 +964,7 @@ def convert_prelu(op, arch, nng): fm_id = ofm.clone(op.name + "_id", set_unique=True) mul_identity.set_output_tensor(fm_id) mul_identity.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, mul_identity) # Combine scaled and alpha multiplied values max_op = Operation(Op.Maximum, op.name + "_max") @@ -1389,7 +1399,7 @@ def fuse_activation_function_with_prev(op, arch, nng): prev_op.set_activation_lut(op.activation_lut) # Bypass op prev_op.set_output_tensor(ofm) - DebugDatabase.add_optimised(op, prev_op) + DebugDatabase.add_optimised(prev_op, prev_op) return op @@ -1482,6 +1492,8 @@ def replace_pad_by_hw_pad(op: Operation, arch, nng): op.attrs["padding"] = Padding.EXPLICIT op.attrs["explicit_padding"] = (top, left, bottom, right) op.set_ifm_ofm_shapes() + DebugDatabase.add_optimised(op, op) + return op @@ -1682,6 +1694,7 @@ def convert_mean_to_depthwise_conv_or_avgpool(op, arch, nng): # If the AvgPool version is used, we don't need to do anything else if op.type == Op.AvgPool: + DebugDatabase.add_optimised(op, op) return op # Make unit weight tensor quantization @@ -1712,6 +1725,7 @@ def convert_mean_to_depthwise_conv_or_avgpool(op, arch, nng): # Add bias tensor bias_shape = [shape[-1]] op.inputs.append(create_const_tensor("bias", bias_shape, DataType.int32, np.ones(bias_shape) * bias)) + DebugDatabase.add_optimised(op, op) return op @@ -1803,7 +1817,6 @@ def convert_shape_op_to_constant_tensor(op: Operation, arch, nng): # Convert this SHAPE op to const op.type = Op.Const - DebugDatabase.add_optimised(op, op) # Add size calculation to shape output tensors ofm.values = np.array(ifm.shape) |