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. |
| 16 | |
| 17 | |
| 18 | # Description: |
| 19 | # Serialises and packs an NPU subgraph into tensors. |
| 20 | |
| 21 | from .nn_graph import PassPlacement |
| 22 | from .tensor import MemArea, Tensor, TensorPurpose, TensorFormat |
| 23 | from .operation import Operation |
| 24 | from .data_type import DataType |
| 25 | import numpy as np |
| 26 | from . import driver_actions |
| 27 | import struct |
| 28 | |
| 29 | |
| 30 | def make_memory_tensor(name, mem_area, sz, want_values, arch): |
| 31 | tens = Tensor([sz], DataType.uint8, name) |
| 32 | tens.mem_area = mem_area |
| 33 | tens.purpose = TensorPurpose.FeatureMap |
| 34 | tens.set_format(TensorFormat.NHWC, arch) |
| 35 | if want_values: |
| 36 | tens.values = np.zeros(tens.shape, np.uint8) |
| 37 | return tens |
| 38 | |
| 39 | |
| 40 | def copy_compressed_values_to_memory_tensor(memory_tensor, src_tensor): |
| 41 | start_addr = src_tensor.address |
| 42 | for compressed_values in src_tensor.compressed_values: |
| 43 | end_addr = start_addr + len(compressed_values) |
| 44 | memory_tensor.values[start_addr:end_addr] = compressed_values |
| 45 | start_addr = end_addr |
| 46 | |
| 47 | |
| 48 | def serialise_npu_subgraph_into_tensors(nng, sg, arch, scratch_tens, flash_tens): |
| 49 | if sg.placement != PassPlacement.Npu: |
| 50 | return scratch_tens, flash_tens |
| 51 | |
| 52 | flash_area = arch.permanent_storage_mem_area |
| 53 | scratch_area = MemArea.Sram |
| 54 | |
| 55 | flash_size = sg.memory_used.get(flash_area, 0) |
| 56 | scratch_size = sg.memory_used.get(scratch_area, 0) |
| 57 | |
| 58 | # Prepare driver actions for this command tensor |
| 59 | da_list = [] |
| 60 | driver_actions.emit_fourcc(da_list, "COP1") |
| 61 | driver_actions.emit_config(da_list, 0, 1, arch) |
| 62 | driver_actions.emit_cmd_stream_header(da_list, len(sg.register_command_stream)) |
| 63 | |
| 64 | # Append command stream words |
| 65 | da_list.extend(sg.register_command_stream) |
| 66 | |
| 67 | # Convert to bytes |
| 68 | payload_bytes = struct.pack("<{0}I".format(len(da_list)), *da_list) |
| 69 | |
| 70 | command_stream_size_bytes = len(payload_bytes) |
| 71 | |
| 72 | # Adjust the bits per element calculation to exclude metadata generated by Vela |
| 73 | nng.total_size[flash_area] = nng.total_size.get(flash_area, 0) - flash_size - command_stream_size_bytes |
| 74 | nng.total_elements[flash_area] = nng.total_elements.get(flash_area, 0) - flash_size - command_stream_size_bytes |
| 75 | nng.total_size[scratch_area] = nng.total_size.get(scratch_area, 0) - scratch_size |
| 76 | nng.total_elements[scratch_area] = nng.total_elements.get(scratch_area, 0) - scratch_size |
| 77 | |
| 78 | if flash_tens == scratch_tens == None: |
| 79 | # First Npu subgraph, create scratch and flash tensors |
| 80 | sg.scratch_tensor = make_memory_tensor(sg.name + "_scratch", scratch_area, scratch_size, False, arch) |
| 81 | sg.scratch_tensor.purpose = TensorPurpose.Scratch |
| 82 | sg.flash_tensor = make_memory_tensor(sg.name + "_flash", flash_area, flash_size, True, arch) |
| 83 | else: |
| 84 | sg.scratch_tensor = scratch_tens |
| 85 | sg.scratch_tensor.shape[0] += scratch_size |
| 86 | sg.flash_tensor = flash_tens |
| 87 | sg.flash_tensor.shape[0] += flash_size |
| 88 | |
| 89 | for cps in sg.cascaded_passes: |
| 90 | for ps in cps.passes: |
| 91 | if ps.placement == PassPlacement.Npu and ps.weight_tensor != None: |
| 92 | # For DMA ops, ps.weight_tensor is referring to the SRAM weight tensor and therefore the address |
| 93 | # is pointing at the destination address of where the weights should be placed in SRAM. |
| 94 | # This ensures that the Flash weight tensor is used instead and thus gets the correct address. |
| 95 | if ps.weight_tensor.ops[0].type == "DMA": |
| 96 | copy_compressed_values_to_memory_tensor(sg.flash_tensor, ps.weight_tensor.ops[0].inputs[0]) |
| 97 | else: |
| 98 | copy_compressed_values_to_memory_tensor(sg.flash_tensor, ps.weight_tensor) |
| 99 | |
| 100 | copy_compressed_values_to_memory_tensor(sg.flash_tensor, ps.scale_tensor) |
| 101 | |
| 102 | sg.command_stream_tensor = make_memory_tensor( |
| 103 | sg.name + "_command_stream", flash_area, command_stream_size_bytes, True, arch |
| 104 | ) |
| 105 | sg.command_stream_tensor.values = np.frombuffer(payload_bytes, dtype=np.uint8) |
| 106 | |
| 107 | return sg.scratch_tensor, sg.flash_tensor |
| 108 | |
| 109 | |
| 110 | def add_const_tens_to_startup_cascaded_pass(startup_cps, tens): |
| 111 | op = Operation("Const", tens.name + "_const") |
| 112 | op.outputs = [tens] |
| 113 | tens.ops = [op] |
| 114 | startup_cps.passes[0].ops.insert(0, op) |
| 115 | startup_cps.passes[0].outputs.insert(0, tens) |
| 116 | startup_cps.outputs.insert(0, tens) |
| 117 | |
| 118 | |
| 119 | def rewrite_npu_call_ops(nng, sg, arch): |
| 120 | if sg.placement != PassPlacement.Cpu: |
| 121 | return |
| 122 | |
| 123 | startup_cps = sg.cascaded_passes[0] |
| 124 | |
| 125 | for idx, cps in enumerate(sg.cascaded_passes): |
| 126 | for ps in cps.passes: |
| 127 | for op in ps.ops: |
| 128 | if op.type == "NpuOp": |
| 129 | callee = op.attrs["subgraph"] |
| 130 | op.attrs["custom_options"] = {"type": op.type} |
| 131 | |
| 132 | sz = 0 |
| 133 | for tens in [callee.scratch_tensor, callee.flash_tensor, callee.command_stream_tensor]: |
| 134 | op.inputs.insert(0, tens) |
| 135 | ps.inputs.insert(0, tens) |
| 136 | cps.inputs.insert(0, tens) |
| 137 | if tens != callee.scratch_tensor: |
| 138 | add_const_tens_to_startup_cascaded_pass(startup_cps, tens) |
| 139 | sz += tens.storage_size() |
| 140 | |
| 141 | for prev_cps in sg.cascaded_passes[: idx + 1]: |
| 142 | prev_cps.sram_used += sz |
| 143 | |
| 144 | if callee.scratch_tensor is not None: |
| 145 | cps.sram_used += callee.scratch_tensor.storage_size() |