| # Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. |
| # |
| # SPDX-License-Identifier: Apache-2.0 |
| # |
| # Licensed under the Apache License, Version 2.0 (the License); you may |
| # not use this file except in compliance with the License. |
| # You may obtain a copy of the License at |
| # |
| # www.apache.org/licenses/LICENSE-2.0 |
| # |
| # Unless required by applicable law or agreed to in writing, software |
| # distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| # WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| # See the License for the specific language governing permissions and |
| # limitations under the License. |
| # Description: |
| # Register level (low-level) command stream generation for Ethos-U55. Takes a list of NPU operations and generates |
| # all the register settings. Calculates dependencies between commands and inserts wait operations. And generates a bit |
| # stream suitable for interpretation by the Ethos-U55 processor. |
| from collections import defaultdict |
| from collections import namedtuple |
| from enum import Enum |
| from enum import IntEnum |
| from typing import List |
| from typing import Optional |
| |
| import numpy as np |
| |
| from . import numeric_util |
| from . import scaling |
| from .api import NpuActivation |
| from .api import NpuActivationOp |
| from .api import NpuAddressRange |
| from .api import NpuBlockOperation |
| from .api import NpuBlockTraversal |
| from .api import NpuConv2DOperation |
| from .api import NpuDataType |
| from .api import NpuDmaOperation |
| from .api import NpuElementWiseOp |
| from .api import NpuElementWiseOperation |
| from .api import NpuFeatureMap |
| from .api import NpuKernel |
| from .api import NpuLayout |
| from .api import NpuOperation |
| from .api import NpuOperationType |
| from .api import NpuPadding |
| from .api import NpuPoolingOp |
| from .api import NpuPoolingOperation |
| from .api import NpuQuantization |
| from .api import NpuResamplingMode |
| from .api import NpuRoundingMode |
| from .api import NpuShape3D |
| from .api import NpuTileBox |
| from .architecture_features import Accelerator |
| from .architecture_features import ArchitectureFeatures |
| from .architecture_features import Block |
| from .architecture_features import Rect |
| from .architecture_features import SharedBufferArea |
| from .architecture_features import SHRAMElements |
| from .debug_database import DebugDatabase |
| from .ethos_u55_regs.ethos_u55_regs import acc_format |
| from .ethos_u55_regs.ethos_u55_regs import activation |
| from .ethos_u55_regs.ethos_u55_regs import cmd0 |
| from .ethos_u55_regs.ethos_u55_regs import cmd1 |
| from .ethos_u55_regs.ethos_u55_regs import elementwise_mode |
| from .ethos_u55_regs.ethos_u55_regs import pooling_mode |
| from .ethos_u55_regs.ethos_u55_regs import resampling_mode |
| from .ethos_u55_regs.ethos_u55_regs import rounding |
| from .high_level_command_stream import CommandType |
| from .high_level_command_to_npu_op import convert_command_to_npu_op |
| from .high_level_command_to_npu_op import to_kernel |
| from .high_level_command_to_npu_op import unary_elementwise_ops |
| from .numeric_util import quantise_float32 |
| from .numeric_util import round_away_zero |
| from .numeric_util import round_up_to_int |
| from .operation import NpuBlockType |
| from .range_set import AccessDirection |
| from .range_set import MemoryAccessSet |
| from .range_set import MemoryRangeSet |
| from .shared_buffer_allocation import find_suitable_block_configs |
| from .shared_buffer_allocation import shared_buffer_allocation_for_npu_op |
| from .shared_buffer_allocation import SharedBufferAllocation |
| |
| |
| class RegisterMachine: |
| def __init__(self): |
| self.n_banks = 1 |
| self.registers = [defaultdict(lambda: None) for _ in range(self.n_banks)] |
| self.bank_idx = 0 |
| |
| def set_register(self, reg, value): |
| is_changed = self.registers[self.bank_idx][reg] != value |
| self.registers[self.bank_idx][reg] = value |
| # is_changed = True # force command |
| return is_changed |
| |
| def switch_bank(self): |
| self.bank_idx = (self.bank_idx + 1) % self.n_banks |
| |
| |
| class CmdMode(IntEnum): |
| NoPayload = 0x0000 |
| Payload32 = 0x4000 |
| Mask = 0xC000 |
| CmdOpMask = 0x03FF |
| |
| |
| class CommandStreamEmitter: |
| WORD_SIZE = 4 |
| |
| def __init__(self): |
| self.cmd_stream = [] |
| self.reg_machine = [RegisterMachine(), RegisterMachine()] |
| self.last_absolute_wait = defaultdict(int) |
| self.offset = 0 |
| |
| def get_reg_machine(self, cmd): |
| if "DMA" in cmd.name: |
| return self.reg_machine[1] |
| else: |
| return self.reg_machine[0] |
| |
| def size_in_bytes(self): |
| sz = 0 |
| for cmd in self.cmd_stream: |
| sz += len(cmd) * CommandStreamEmitter.WORD_SIZE |
| return sz |
| |
| def to_list(self) -> List[int]: |
| return [elem for cmd in self.cmd_stream for elem in cmd] |
| |
| def print_cmds(self): |
| print("Code: Command: Param: Payload:") |
| for words_for_one_command in self.cmd_stream: |
| code = words_for_one_command[0] & 0x0000FFFF # lower 16 bits |
| param = words_for_one_command[0] >> 16 # higher 16 bits |
| |
| payload_mode = CmdMode(code & CmdMode.Mask) |
| |
| # code and command |
| s = " 0x%04x " % code |
| if payload_mode == CmdMode.NoPayload: |
| s += str(cmd0(code & CmdMode.CmdOpMask)) |
| else: |
| s += str(cmd1(code & CmdMode.CmdOpMask)) |
| |
| s = s.ljust(40) |
| s += "%5d" % param |
| |
| # payload |
| if payload_mode == CmdMode.Payload32: |
| s += " 0x%08x (%d)" % (words_for_one_command[1], words_for_one_command[1]) |
| else: |
| s += " -" |
| |
| print(s) |
| |
| def cmd0_with_param(self, cmd: cmd0, param): |
| if isinstance(param, Enum): |
| param = int(param.value) |
| else: |
| param = int(param) |
| param = param & 0xFFFF |
| command = cmd.value | (param << 16) |
| if not self.get_reg_machine(cmd).set_register(cmd, (command, param)): |
| return |
| |
| # This is not a redundant command, actually write it |
| self.cmd_stream.append((command,)) |
| self.offset += CommandStreamEmitter.WORD_SIZE |
| |
| def cmd1_with_offset(self, cmd: cmd1, offset, param=0x0): |
| offset = int(offset) & 0xFFFFFFFFF |
| command = cmd.value | CmdMode.Payload32.value | (param << 16) |
| |
| if not self.get_reg_machine(cmd).set_register(cmd, (command, offset)): |
| return |
| |
| # This is not a redundant command, actually write it |
| self.cmd_stream.append((command, offset)) |
| self.offset += CommandStreamEmitter.WORD_SIZE * 2 |
| |
| def cmd_wait(self, cmd: cmd0, channel: int, outstanding_count: int): |
| param = (16 * channel) + outstanding_count |
| command = ((param & 0xFFFF) << 16) | cmd.value |
| self.cmd_stream.append((command,)) |
| self.offset += CommandStreamEmitter.WORD_SIZE |
| |
| def cmd_do_operation(self, cmd: cmd0, param=0): |
| param = int(param) |
| command = ((param & 0xFFFF) << 16) | cmd.value |
| |
| self.cmd_stream.append((command,)) |
| self.offset += CommandStreamEmitter.WORD_SIZE |
| self.get_reg_machine(cmd).switch_bank() |
| |
| |
| # ------------------------------------------------------------------- |
| # REGISTER GENERATION |
| # ------------------------------------------------------------------- |
| |
| |
| class BasePointerIndex(IntEnum): |
| WeightTensor = 0 # base address index for the Weight tensor |
| ScratchTensor = 1 # base address index for the Scratch_tensor in the TensorArena |
| ScratchFastTensor = 2 # base address for the Scratch_fast_tensor |
| Mem2Mem = (1 << 8) | (3 << 0) # base address slot for memory 2 memory transfer |
| |
| |
| # TODO: Replace with definitions from ethos_u55_regs |
| class IFM2Broadcast(IntEnum): |
| BroadcastHdim = 1 << 0 |
| BroadcastWdim = 1 << 1 |
| BroadcastCdim = 1 << 2 |
| ReverseOperandOrder = 1 << 6 |
| UseIFM2Scalar = 1 << 7 |
| |
| |
| pooling_op_map = { |
| NpuPoolingOp.MAX: pooling_mode.MAX.value, |
| NpuPoolingOp.AVERAGE: pooling_mode.AVERAGE.value, |
| NpuPoolingOp.REDUCE_SUM: pooling_mode.REDUCE_SUM.value, |
| } |
| |
| elementwise_op_map = { |
| NpuElementWiseOp.MUL: elementwise_mode.MUL.value, |
| NpuElementWiseOp.ADD: elementwise_mode.ADD.value, |
| NpuElementWiseOp.SUB: elementwise_mode.SUB.value, |
| NpuElementWiseOp.MIN: elementwise_mode.MIN.value, |
| NpuElementWiseOp.MAX: elementwise_mode.MAX.value, |
| NpuElementWiseOp.LRELU: elementwise_mode.LRELU.value, |
| NpuElementWiseOp.ABS: elementwise_mode.ABS.value, |
| NpuElementWiseOp.CLZ: elementwise_mode.CLZ.value, |
| NpuElementWiseOp.SHR: elementwise_mode.SHR.value, |
| NpuElementWiseOp.SHL: elementwise_mode.SHL.value, |
| } |
| |
| activation_op_map = { |
| NpuActivationOp.NONE_OR_RELU: activation.NONE, |
| NpuActivationOp.TANH: activation.TANH, |
| NpuActivationOp.SIGMOID: activation.SIGMOID, |
| } |
| |
| # Maps an AccumulatorType enum to the corresponding acc_format value |
| acc_format_map = { |
| SHRAMElements.Acc16: acc_format.FP_S5_10.value, |
| SHRAMElements.Acc32: acc_format.INT_32BIT.value, |
| SHRAMElements.Acc40: acc_format.INT_40BIT.value, |
| } |
| |
| resampling_mode_map = { |
| NpuResamplingMode.NONE: resampling_mode.NONE, |
| NpuResamplingMode.NEAREST: resampling_mode.NEAREST, |
| NpuResamplingMode.TRANSPOSE: resampling_mode.TRANSPOSE, |
| } |
| |
| # Maps data type size in bits to activation precision |
| precision_map = {8: 0, 16: 1, 32: 2} |
| |
| # Maps rounding mode to the corresponding value |
| rounding_mode_map = { |
| NpuRoundingMode.TFL: rounding.TFL.value, |
| NpuRoundingMode.TRUNCATE: rounding.TRUNCATE.value, |
| NpuRoundingMode.NATURAL: rounding.NATURAL.value, |
| } |
| |
| |
| def quantise(value: float, quant: Optional[NpuQuantization]) -> int: |
| """Quantizes the given value""" |
| scale = 1 if quant is None or quant.scale_f32 is None else quant.scale_f32 |
| zp = 0 if quant is None else quant.zero_point |
| return quantise_float32(value, scale, zp) |
| |
| |
| def has_ifm2(npu_op: NpuBlockOperation) -> bool: |
| """Checks if op has non-scalar IFM2""" |
| return npu_op.ifm2 is not None and npu_op.ifm2_scalar is None |
| |
| |
| def is_dma_op(npu_op: NpuOperation) -> bool: |
| """Checks if op is a DMA operation""" |
| return npu_op.op_type == NpuOperationType.Dma |
| |
| |
| def generate_padding(emit: CommandStreamEmitter, padding: NpuPadding): |
| """Generates IFM_PAD registers""" |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_PAD_TOP, padding.top) |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_PAD_LEFT, padding.left) |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_PAD_BOTTOM, padding.bottom) |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_PAD_RIGHT, padding.right) |
| |
| |
| def generate_activation(emit: CommandStreamEmitter, activation: Optional[NpuActivation], ofm: NpuFeatureMap): |
| """Generates ACTIVATION registers""" |
| act = activation if activation is not None else NpuActivation(NpuActivationOp.NONE_OR_RELU) |
| |
| if act.min is None: |
| quantized_min = ofm.data_type.min_value() |
| else: |
| quantized_min = quantise(act.min, ofm.quantization) |
| if act.max is None: |
| quantized_max = ofm.data_type.max_value() |
| else: |
| quantized_max = quantise(act.max, ofm.quantization) |
| quantized_min = max(quantized_min, np.iinfo(np.int16).min, ofm.data_type.min_value()) |
| quantized_max = min(quantized_max, np.iinfo(np.int16).max, ofm.data_type.max_value()) |
| if act.op_type == NpuActivationOp.TABLE_LOOKUP: |
| assert 0 <= act.lookup_table_index < 8 |
| activation_value = 16 + act.lookup_table_index |
| if ofm.data_type == NpuDataType.INT32: |
| activation_value |= 3 << 12 # Force I8 range |
| quantized_min = max(-128, quantized_min) |
| quantized_max = min(127, quantized_max) |
| else: |
| activation_value = activation_op_map[act.op_type] |
| emit.cmd0_with_param(cmd0.NPU_SET_ACTIVATION, activation_value) |
| emit.cmd0_with_param(cmd0.NPU_SET_ACTIVATION_MIN, quantized_min) |
| emit.cmd0_with_param(cmd0.NPU_SET_ACTIVATION_MAX, quantized_max) |
| |
| |
| def generate_addresses(emit: CommandStreamEmitter, ptr_cmds: List[cmd1], addresses: List[int], layout: NpuLayout): |
| """Generates xFM_BASE registers""" |
| if layout == NpuLayout.NHCWB16: |
| # Check that all BasePointer addresses are aligned to 16 bytes |
| assert all((int(addr) % 16) == 0 for addr in addresses) |
| emit.cmd1_with_offset(ptr_cmds[0], addresses[0]) |
| emit.cmd1_with_offset(ptr_cmds[1], addresses[1]) |
| emit.cmd1_with_offset(ptr_cmds[2], addresses[2]) |
| emit.cmd1_with_offset(ptr_cmds[3], addresses[3]) |
| |
| |
| def generate_tiles(emit: CommandStreamEmitter, tile_cmds: List[cmd0], tiles: NpuTileBox): |
| """Generates xFM_HEIGHT0/HEIGHT1/WIDTH0 registers""" |
| emit.cmd0_with_param(tile_cmds[0], tiles.height_0 - 1) |
| emit.cmd0_with_param(tile_cmds[1], tiles.height_1 - 1) |
| emit.cmd0_with_param(tile_cmds[2], tiles.width_0 - 1) |
| |
| |
| def generate_strides( |
| emit: CommandStreamEmitter, fm: NpuFeatureMap, stride_c_cmd: cmd1, stride_y_cmd: cmd1, stride_x_cmd: cmd1 |
| ): |
| """Generates STRIDE_C/Y/X registers""" |
| strides = get_strides(fm) |
| emit.cmd1_with_offset(stride_c_cmd, strides.depth) # stride between 16-byte channel blocks (C) |
| emit.cmd1_with_offset(stride_y_cmd, strides.height) # stride between vertical values (H) |
| emit.cmd1_with_offset(stride_x_cmd, strides.width) # stride between horisontal values (W) |
| |
| |
| def generate_ifm_precision(emit: CommandStreamEmitter, fm: NpuFeatureMap, op_to_scale: int, precision_cmd: cmd0): |
| """Generates IFM/IFM2_PRECISION register""" |
| dtype = fm.data_type |
| prec = 1 if dtype.is_signed() else 0 |
| activation_precision = precision_map[dtype.size_in_bits()] |
| prec += activation_precision << 2 |
| |
| if fm.layout == NpuLayout.NHCWB16: |
| prec |= 1 << 6 |
| |
| prec |= op_to_scale << 8 |
| emit.cmd0_with_param(precision_cmd, prec) |
| |
| |
| def generate_ofm_precision(emit: CommandStreamEmitter, npu_op: NpuBlockOperation, use_global_scale: bool): |
| """Generates OFM_PRECISION register""" |
| dtype = npu_op.ofm.data_type |
| prec = 1 if dtype.is_signed() else 0 |
| activation_precision = precision_map[dtype.size_in_bits()] |
| prec += activation_precision << 1 |
| |
| if use_global_scale: |
| # Set global scale bit, as opposed to using per channel scale |
| prec |= 1 << 8 |
| if npu_op.ofm.layout == NpuLayout.NHCWB16: |
| prec |= 1 << 6 |
| prec |= rounding_mode_map[npu_op.rounding_mode] << 14 |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_PRECISION, prec) |
| |
| |
| def generate_ifm2_broadcast(emit: CommandStreamEmitter, npu_op: NpuElementWiseOperation): |
| """Generates IFM2_BROADCAST register for binary elementwise operations""" |
| ifm2_broadcast = 0 |
| ifm = npu_op.ifm |
| ifm2 = npu_op.ifm2 |
| if npu_op.reversed_operands: |
| ifm2_broadcast |= IFM2Broadcast.ReverseOperandOrder |
| if npu_op.ifm2_scalar is not None: |
| # IFM2 is a constant, set UseIFM2Scalar bit to IFM2_BROADCAST |
| ifm2_broadcast |= IFM2Broadcast.UseIFM2Scalar |
| else: |
| if ifm.shape.height != ifm2.shape.height: |
| # Broadcast in 'H' dimension |
| assert ifm2.shape.height == 1 |
| ifm2_broadcast |= IFM2Broadcast.BroadcastHdim |
| |
| if ifm.shape.width != ifm2.shape.width: |
| # Broadcast in 'W' dimension |
| assert ifm2.shape.width == 1 |
| ifm2_broadcast |= IFM2Broadcast.BroadcastWdim |
| |
| if ifm.shape.depth != ifm2.shape.depth: |
| # Broadcast in 'C' dimension |
| assert ifm2.shape.depth == 1 |
| ifm2_broadcast |= IFM2Broadcast.BroadcastCdim |
| |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM2_BROADCAST, ifm2_broadcast) |
| |
| |
| def generate_ifm(emit: CommandStreamEmitter, ifm: NpuFeatureMap): |
| """Generates general IFM registers""" |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_REGION, ifm.region) |
| generate_addresses( |
| emit, |
| [cmd1.NPU_SET_IFM_BASE0, cmd1.NPU_SET_IFM_BASE1, cmd1.NPU_SET_IFM_BASE2, cmd1.NPU_SET_IFM_BASE3], |
| ifm.tiles.addresses, |
| ifm.layout, |
| ) |
| generate_tiles( |
| emit, [cmd0.NPU_SET_IFM_HEIGHT0_M1, cmd0.NPU_SET_IFM_HEIGHT1_M1, cmd0.NPU_SET_IFM_WIDTH0_M1], ifm.tiles |
| ) |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_DEPTH_M1, ifm.shape.depth - 1) |
| generate_strides(emit, ifm, cmd1.NPU_SET_IFM_STRIDE_C, cmd1.NPU_SET_IFM_STRIDE_Y, cmd1.NPU_SET_IFM_STRIDE_X) |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_ZERO_POINT, int(ifm.quantization.zero_point)) |
| |
| |
| def generate_ifm2(emit: CommandStreamEmitter, ifm2: NpuFeatureMap, has_scalar: bool): |
| """Generates general IFM2 registers""" |
| if not has_scalar: |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM2_REGION, ifm2.region) |
| generate_addresses( |
| emit, |
| [cmd1.NPU_SET_IFM2_BASE0, cmd1.NPU_SET_IFM2_BASE1, cmd1.NPU_SET_IFM2_BASE2, cmd1.NPU_SET_IFM2_BASE3], |
| ifm2.tiles.addresses, |
| ifm2.layout, |
| ) |
| generate_tiles( |
| emit, [cmd0.NPU_SET_IFM2_HEIGHT0_M1, cmd0.NPU_SET_IFM2_HEIGHT1_M1, cmd0.NPU_SET_IFM2_WIDTH0_M1], ifm2.tiles |
| ) |
| generate_strides(emit, ifm2, cmd1.NPU_SET_IFM2_STRIDE_C, cmd1.NPU_SET_IFM2_STRIDE_Y, cmd1.NPU_SET_IFM2_STRIDE_X) |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM2_ZERO_POINT, int(ifm2.quantization.zero_point)) |
| |
| |
| def generate_ofm(emit: CommandStreamEmitter, ofm: NpuFeatureMap): |
| """Generates general OFM registers""" |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_REGION, ofm.region) |
| generate_addresses( |
| emit, |
| [cmd1.NPU_SET_OFM_BASE0, cmd1.NPU_SET_OFM_BASE1, cmd1.NPU_SET_OFM_BASE2, cmd1.NPU_SET_OFM_BASE3], |
| ofm.tiles.addresses, |
| ofm.layout, |
| ) |
| generate_tiles( |
| emit, [cmd0.NPU_SET_OFM_HEIGHT0_M1, cmd0.NPU_SET_OFM_HEIGHT1_M1, cmd0.NPU_SET_OFM_WIDTH0_M1], ofm.tiles |
| ) |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_HEIGHT_M1, ofm.shape.height - 1) |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_WIDTH_M1, ofm.shape.width - 1) |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_DEPTH_M1, ofm.shape.depth - 1) |
| generate_strides(emit, ofm, cmd1.NPU_SET_OFM_STRIDE_C, cmd1.NPU_SET_OFM_STRIDE_Y, cmd1.NPU_SET_OFM_STRIDE_X) |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_ZERO_POINT, int(ofm.quantization.zero_point)) |
| |
| |
| def generate_kernel(emit: CommandStreamEmitter, kernel: NpuKernel, block_traversal: NpuBlockTraversal): |
| """Generates KERNEL related registers""" |
| emit.cmd0_with_param(cmd0.NPU_SET_KERNEL_HEIGHT_M1, kernel.dilation_y * (kernel.height - 1)) |
| emit.cmd0_with_param(cmd0.NPU_SET_KERNEL_WIDTH_M1, kernel.dilation_x * (kernel.width - 1)) |
| # set kernel x stride low bit |
| stride = (kernel.stride_x - 1) & 1 |
| # set kernel y stride low bit |
| stride |= (kernel.stride_y - 1 & 1) << 1 |
| # set kernel x stride extension bits |
| stride |= (kernel.stride_x - 1 >> 1) << 6 |
| # set kernel y stride extension bits |
| stride |= (kernel.stride_y - 1 >> 1) << 9 |
| stride |= (kernel.dilation_x - 1) << 3 |
| stride |= (kernel.dilation_y - 1) << 4 |
| if block_traversal == NpuBlockTraversal.PART_KERNEL_FIRST: |
| stride |= 1 << 2 |
| emit.cmd0_with_param(cmd0.NPU_SET_KERNEL_STRIDE, stride) |
| |
| |
| def generate_weights(emit: CommandStreamEmitter, weights: List[NpuAddressRange], arch: ArchitectureFeatures): |
| """Generates WEIGHT registers""" |
| if len(weights) == 0: |
| return |
| emit.cmd0_with_param(cmd0.NPU_SET_WEIGHT_REGION, weights[0].region) |
| # Set weights sources for active and present cores |
| for core, (addr, length) in enumerate( |
| [ |
| (cmd1.NPU_SET_WEIGHT_BASE, cmd1.NPU_SET_WEIGHT_LENGTH), |
| (cmd1.NPU_SET_WEIGHT1_BASE, cmd1.NPU_SET_WEIGHT1_LENGTH), |
| ] |
| ): |
| if core < len(weights): |
| emit.cmd1_with_offset(addr, weights[core].address) |
| emit.cmd1_with_offset(length, weights[core].length) |
| elif core < arch.ncores: |
| emit.cmd1_with_offset(addr, weights[0].address) |
| emit.cmd1_with_offset(length, 0) |
| |
| |
| def generate_biases(emit: CommandStreamEmitter, biases: List[NpuAddressRange], arch: ArchitectureFeatures): |
| """Generates SCALE registers""" |
| if len(biases) == 0: |
| return |
| emit.cmd0_with_param(cmd0.NPU_SET_SCALE_REGION, biases[0].region) |
| # Set weights sources for active and present cores |
| for core, (addr, length) in enumerate( |
| [(cmd1.NPU_SET_SCALE_BASE, cmd1.NPU_SET_SCALE_LENGTH), (cmd1.NPU_SET_SCALE1_BASE, cmd1.NPU_SET_SCALE1_LENGTH)] |
| ): |
| if core < len(biases): |
| emit.cmd1_with_offset(addr, biases[core].address) |
| emit.cmd1_with_offset(length, biases[core].length) |
| elif core < arch.ncores: |
| emit.cmd1_with_offset(addr, biases[0].address) |
| emit.cmd1_with_offset(length, 0) |
| |
| |
| def generate_block_config( |
| emit: CommandStreamEmitter, |
| npu_op: NpuBlockOperation, |
| arch: ArchitectureFeatures, |
| shared_buffer: SharedBufferAllocation, |
| ) -> NpuShape3D: |
| """Selects a suitable block config if none has been set, and generates OFM_BLK_HEIGHT/WIDTH/DEPTH registers""" |
| block_config = npu_op.block_config |
| if block_config is None or block_config.height < 0: |
| # Note: this code only used if the public API to generate command streams is used; |
| # in the "normal" flow, the block config selected by the scheduler is used |
| if npu_op.weights: |
| assert block_config is not None, "block_config.depth must be provided for ops with weights" |
| # Block config has not been provided: find one |
| blocks = find_suitable_block_configs(arch, shared_buffer) |
| # Return the block with biggest volume |
| # TODO: use a better algorithm to find the best block |
| best_block = None |
| best_value = 0 |
| for block in blocks: |
| if block_config is not None and block[3] != block_config.depth: |
| continue |
| value = block[0] * block[1] * block[3] |
| if value > best_value: |
| best_value = value |
| best_block = block |
| assert best_block is not None, f"No suitable block config was found, {npu_op.op_type}" |
| block_config = NpuShape3D(height=best_block[0], width=best_block[1], depth=best_block[3]) |
| alloc = shared_buffer.try_block(Block(block_config.width, block_config.height, block_config.depth)) |
| assert alloc is not None, f"Block config {block_config} does not fit, op: {npu_op.op_type}" |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_BLK_HEIGHT_M1, block_config.height - 1) |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_BLK_WIDTH_M1, block_config.width - 1) |
| emit.cmd0_with_param(cmd0.NPU_SET_OFM_BLK_DEPTH_M1, block_config.depth - 1) |
| return block_config |
| |
| |
| def generate_shram_registers_elementwise( |
| emit: CommandStreamEmitter, |
| npu_op: NpuElementWiseOperation, |
| arch: ArchitectureFeatures, |
| shared_buffer: SharedBufferAllocation, |
| ): |
| """Generates IB_END/IB_START/AB_START registers for elementwise operations""" |
| # For elementwise set the required SHRAM to be equal to the total size of available SHRAM |
| uses_lut = npu_op.activation is not None and npu_op.activation.op_type == NpuActivationOp.TABLE_LOOKUP |
| shram_required = arch.available_shram_banks(uses_lut) |
| |
| # Acc buffers not needed so set AB_START to size of SHRAM |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_IB_END, shram_required) |
| emit.cmd0_with_param(cmd0.NPU_SET_AB_START, shram_required) |
| if has_ifm2(npu_op): |
| # Set IFM2_IB_START to the latter half of the IB space |
| ifm_ib_start = shared_buffer.bank_locations[SharedBufferArea.IFM] |
| emit.cmd0_with_param( |
| cmd0.NPU_SET_IFM2_IB_START, (shram_required - ifm_ib_start) // shared_buffer.ifm_count + ifm_ib_start, |
| ) |
| emit.cmd0_with_param(cmd0.NPU_SET_ACC_FORMAT, acc_format_map[shared_buffer.use_accumulator_element]) |
| |
| |
| def generate_shram_registers_non_elementwise(emit: CommandStreamEmitter, shared_buffer: SharedBufferAllocation): |
| """Generates IB_END/IB_START/AB_START registers for non-elementwise operations""" |
| emit.cmd0_with_param( |
| cmd0.NPU_SET_IFM_IB_END, |
| shared_buffer.bank_locations[SharedBufferArea.IFM] + shared_buffer.banks_required[SharedBufferArea.IFM], |
| ) |
| emit.cmd0_with_param(cmd0.NPU_SET_AB_START, shared_buffer.bank_locations[SharedBufferArea.Accumulators]) |
| emit.cmd0_with_param(cmd0.NPU_SET_ACC_FORMAT, acc_format_map[shared_buffer.use_accumulator_element]) |
| |
| |
| def generate_common( |
| emit: CommandStreamEmitter, |
| npu_op: NpuBlockOperation, |
| block_traversal: NpuBlockTraversal, |
| arch: ArchitectureFeatures, |
| use_global_scale: bool = False, |
| op_to_scale: int = 0, |
| ): |
| """Generate registers that are common to most operations""" |
| assert npu_op.ifm is not None and npu_op.ofm is not None |
| generate_ifm(emit, npu_op.ifm) |
| generate_ifm_precision(emit, npu_op.ifm, op_to_scale, cmd0.NPU_SET_IFM_PRECISION) |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM_UPSCALE, resampling_mode_map[npu_op.ifm_upscale]) |
| if npu_op.padding is not None: |
| generate_padding(emit, npu_op.padding) |
| generate_ofm(emit, npu_op.ofm) |
| generate_ofm_precision(emit, npu_op, use_global_scale) |
| if npu_op.op_type != NpuOperationType.ElementWise: |
| assert npu_op.kernel is not None |
| generate_kernel(emit, npu_op.kernel, block_traversal) |
| generate_weights(emit, npu_op.weights, arch) |
| generate_biases(emit, npu_op.biases, arch) |
| generate_activation(emit, npu_op.activation, npu_op.ofm) |
| |
| |
| # ------------------------------------------------------------------- |
| # SCALING |
| # ------------------------------------------------------------------- |
| |
| |
| def generate_ofm_scaling_for_pooling(emit: CommandStreamEmitter, pool_op: NpuPoolingOperation): |
| """Generates OFM_SCALE register for pooling operations""" |
| # For valid padding vela has to output scaling values |
| kernel = pool_op.kernel |
| ifm_quant = pool_op.ifm.quantization |
| ofm_quant = pool_op.ofm.quantization |
| if pool_op.activation is not None and pool_op.activation.op_type in (NpuActivationOp.SIGMOID, NpuActivationOp.TANH): |
| assert ifm_quant.scale_f32 is not None |
| rescale = 0x3000 * ifm_quant.scale_f32 |
| if pool_op.ifm.data_type == NpuDataType.INT16: |
| # Calculate scale and shift for the output scale of 1/(3*4096) |
| shift = 0 |
| max_rescale = np.iinfo(np.int16).max / 2 |
| while rescale <= max_rescale and shift <= 30: |
| shift += 1 |
| rescale *= 2 |
| scale = int(rescale) |
| else: |
| rescale_bits = len(bin(round_up_to_int(rescale))) - 2 + 1 |
| scale, shift = scaling.quantise_pooling_scale(kernel.height * kernel.width, rescale_bits) |
| scale = int(round_away_zero(scale * rescale)) |
| elif pool_op.fused_quantize: |
| # Quantize op requires different scaling |
| ifm_scale_f64 = np.double(ifm_quant.scale_f32) |
| ofm_scale_f64 = np.double(ofm_quant.scale_f32) |
| scale, shift = scaling.quantise_scale(ifm_scale_f64 / ofm_scale_f64) |
| elif pool_op.rescale is not None: |
| # for ResizeBilinear operations with "rescale" in primary_op.attrs |
| rescale = pool_op.rescale |
| rescale_bits = len(bin(round_up_to_int(rescale))) - 2 + 1 |
| scale, shift = scaling.quantise_pooling_scale(kernel.height * kernel.width, rescale_bits) |
| scale = int(round_away_zero(scale * rescale)) |
| else: |
| # In case avg pool fused with concat or other memory operation, rescaling might be needed. |
| # kernel height == kernel width == 1 is always true in this case |
| # Normally the scale is maximised, to get maximum precision, which means that |
| # if rescale != 1, scale need to consider the number of bits needed for rescaling |
| if ofm_quant.scale_f32 is not None and ifm_quant.scale_f32 is not None: |
| rescale = ifm_quant.scale_f32 / ofm_quant.scale_f32 |
| rescale_bits = 0 |
| if kernel.height == kernel.width == 1: |
| if rescale > 1: |
| rescale_bits = len(bin(round_up_to_int(rescale))) - 2 + 1 |
| elif rescale < 1: |
| rescale_bits = -(len(bin(round_up_to_int(1 / rescale))) - 2 - 1) |
| scale, shift = scaling.quantise_pooling_scale(kernel.height * kernel.width, rescale_bits) |
| scale = int(round_away_zero(scale * rescale)) |
| else: |
| scale = 1 |
| shift = 0 |
| |
| emit.cmd1_with_offset(cmd1.NPU_SET_OFM_SCALE, scale, shift) |
| |
| |
| def generate_scaling_for_elementwise(emit: CommandStreamEmitter, npu_op: NpuElementWiseOperation) -> int: |
| """ |
| Generates OFM/OPA/OPB_SCALE registers for elementwise operators. |
| Returns the operator to scale |
| """ |
| op_to_scale = 0 |
| if npu_op.sub_op_type in (NpuElementWiseOp.ADD, NpuElementWiseOp.MUL, NpuElementWiseOp.SUB): |
| input_scale = npu_op.ifm.quantization.scale_f32 if npu_op.ifm.quantization else None |
| input2_scale = npu_op.ifm2.quantization.scale_f32 if npu_op.ifm2.quantization else None |
| output_scale = npu_op.ofm.quantization.scale_f32 if npu_op.ofm.quantization else None |
| |
| if npu_op.activation is not None and npu_op.activation.op_type in ( |
| NpuActivationOp.SIGMOID, |
| NpuActivationOp.TANH, |
| ): |
| output_scale = 1 / 0x3000 |
| |
| if npu_op.sub_op_type == NpuElementWiseOp.MUL: |
| if None in (input_scale, input2_scale, output_scale): |
| ofm_scale = 1 |
| shift = 0 |
| else: |
| ofm_scale, shift = scaling.elementwise_mul_scale(input_scale, input2_scale, output_scale) |
| emit.cmd1_with_offset(cmd1.NPU_SET_OFM_SCALE, ofm_scale, shift) |
| else: # Add/Sub |
| if None in (input_scale, input2_scale, output_scale): |
| opa_scale = opb_scale = ofm_scale = 1 |
| opa_shift = shift = 0 |
| if npu_op.rescale is not None: |
| ofm_scale, shift = npu_op.rescale |
| elif input_scale == input2_scale: |
| opa_scale, opb_scale, ofm_scale, shift = scaling.simplified_elementwise_add_sub_scale( |
| input_scale, input2_scale, output_scale |
| ) |
| opa_shift = 0 # Unused for this case |
| else: |
| # Use advanced implementation only when input scales differ |
| bitdepth = npu_op.ifm.data_type.size_in_bits() |
| (opa_scale, opa_shift, ofm_scale, shift, op_to_scale,) = scaling.advanced_elementwise_add_sub_scale( |
| input_scale, input2_scale, output_scale, bitdepth |
| ) |
| opb_scale = 0 # Unused for this case |
| if npu_op.reversed_operands: |
| # If the operand order is reversed we also have to swap which operand is scaled |
| if op_to_scale == scaling.OperandToScale.OPa: |
| op_to_scale = scaling.OperandToScale.OPb |
| else: |
| op_to_scale = scaling.OperandToScale.OPa |
| emit.cmd1_with_offset(cmd1.NPU_SET_OPA_SCALE, opa_scale, opa_shift) |
| emit.cmd1_with_offset(cmd1.NPU_SET_OPB_SCALE, opb_scale) |
| emit.cmd1_with_offset(cmd1.NPU_SET_OFM_SCALE, ofm_scale, shift) |
| elif npu_op.sub_op_type in (NpuElementWiseOp.LRELU, NpuElementWiseOp.ABS): |
| output_scale = npu_op.ofm.quantization.scale_f32 |
| ofm_scale, shift = scaling.quantise_scale(output_scale) |
| emit.cmd1_with_offset(cmd1.NPU_SET_OFM_SCALE, ofm_scale, shift) |
| else: |
| emit.cmd1_with_offset(cmd1.NPU_SET_OFM_SCALE, 1, 0) |
| return op_to_scale |
| |
| |
| # ------------------------------------------------------------------- |
| # ADDRESSING/STRIDES (helper functions) |
| # ------------------------------------------------------------------- |
| |
| |
| def ranges_overlap(range1: NpuAddressRange, range2: NpuAddressRange) -> bool: |
| """Checks if the ranges overlap""" |
| return range1.region == range2.region and numeric_util.overlaps( |
| range1.address, range1.address + range1.length, range2.address, range2.address + range2.length |
| ) |
| |
| |
| def get_strides(fm: NpuFeatureMap) -> NpuShape3D: |
| """Calculates STRIDE_C/Y/X""" |
| if fm.strides is not None: |
| return fm.strides |
| elem_size = fm.data_type.size_in_bytes() |
| if fm.layout == NpuLayout.NHWC: |
| stride_c = elem_size |
| stride_x = fm.shape.depth * stride_c |
| stride_y = fm.shape.width * stride_x |
| else: |
| stride_x = 16 * elem_size |
| stride_c = stride_x * fm.shape.width |
| stride_y = elem_size * fm.shape.width * numeric_util.round_up(fm.shape.depth, 16) |
| return NpuShape3D(depth=stride_c, height=stride_y, width=stride_x) |
| |
| |
| def get_address(fm: NpuFeatureMap, strides: NpuShape3D, y: int, x: int, c: int) -> int: |
| """Returns address of given coordinate""" |
| t = 0 |
| BRICK = 16 |
| stride_c = BRICK * fm.data_type.size_in_bytes() if fm.layout == NpuLayout.NHWC else strides.depth |
| stride_x = BRICK * fm.data_type.size_in_bytes() if fm.layout == NpuLayout.NHCWB16 else strides.width |
| if x >= fm.tiles.width_0: |
| x -= fm.tiles.width_0 |
| t = 1 |
| if y >= fm.tiles.height_1: |
| y -= fm.tiles.height_1 |
| t += 2 |
| elif y >= fm.tiles.height_0: |
| y -= fm.tiles.height_0 |
| t += 2 |
| elem_size = fm.data_type.size_in_bytes() |
| return ( |
| fm.tiles.addresses[t] + y * strides.height + x * stride_x + (c // BRICK) * stride_c + int(c % BRICK) * elem_size |
| ) |
| |
| |
| def get_address_range( |
| fm: NpuFeatureMap, strides: NpuShape3D, y0: int, x0: int, c0: int, y1: int, x1: int, c1: int |
| ) -> NpuAddressRange: |
| """Gets address range for (y0, x0, c0) - (y1, x1, c1)""" |
| addr0 = get_address(fm, strides, y0, x0, c0) |
| addr1 = get_address(fm, strides, y1, x1, c1) |
| return NpuAddressRange(region=fm.region, address=addr0, length=addr1 - addr0 + fm.data_type.size_in_bytes()) |
| |
| |
| def get_address_ranges(fm: NpuFeatureMap) -> List[Optional[NpuAddressRange]]: |
| """Returns 4 adddress ranges, one for every tile, None if the tile is not in use""" |
| strides = get_strides(fm) |
| height, width, depth = fm.shape.height, fm.shape.width, fm.shape.depth |
| height_0, height_1, width_0 = fm.tiles.height_0, fm.tiles.height_1, fm.tiles.width_0 |
| t0 = get_address_range(fm, strides, 0, 0, 0, min(height, height_0) - 1, min(width, width_0) - 1, depth - 1,) |
| if width > width_0: |
| t1 = get_address_range(fm, strides, 0, width_0, 0, min(height, height_1) - 1, width - 1, depth - 1) |
| else: |
| t1 = None |
| if height > height_0: |
| t2 = get_address_range(fm, strides, height_0, 0, 0, height - 1, min(width, width_0) - 1, depth - 1) |
| else: |
| t2 = None |
| if t1 is not None and t2 is not None: |
| t3 = get_address_range(fm, strides, height_0, width_0, 0, height - 1, width - 1, depth - 1) |
| else: |
| t3 = None |
| return [t0, t1, t2, t3] |
| |
| |
| # ------------------------------------------------------------------- |
| # DMA_WAIT/KERNEL_WAIT |
| # ------------------------------------------------------------------- |
| |
| |
| Watermark = namedtuple("Watermark", ["npu", "dma"]) |
| |
| |
| def memory_range_set(range: NpuAddressRange) -> MemoryRangeSet: |
| return MemoryRangeSet(range.region, range.address, range.address + range.length) |
| |
| |
| def get_dma_memory_accesses(dma_op: NpuDmaOperation) -> MemoryAccessSet: |
| """Returns the address that are read and written by the given DMA operation""" |
| res = MemoryAccessSet() |
| res.add(memory_range_set(dma_op.src), AccessDirection.Read) |
| res.add(memory_range_set(dma_op.dest), AccessDirection.Write) |
| return res |
| |
| |
| def get_op_memory_accesses(npu_op: NpuBlockOperation, arch: ArchitectureFeatures) -> MemoryAccessSet: |
| """Returns the addresses that are read and written by the given operation""" |
| assert npu_op.ifm is not None and npu_op.ofm is not None |
| # Read addresses |
| read_ranges = get_address_ranges(npu_op.ifm) |
| if has_ifm2(npu_op): |
| assert npu_op.ifm2 is not None |
| read_ranges.extend(get_address_ranges(npu_op.ifm2)) |
| read_ranges.extend(npu_op.weights) |
| read_ranges.extend(npu_op.biases) |
| if npu_op.activation is not None and npu_op.activation.op_type == NpuActivationOp.TABLE_LOOKUP: |
| address = arch.available_shram_banks(True) * arch.shram_bank_size |
| read_ranges.append(NpuAddressRange(region=BasePointerIndex.Mem2Mem, address=address, length=2048)) |
| # Written addresses |
| write_ranges = get_address_ranges(npu_op.ofm) |
| # Add write access to SHRAM, needed when LUTs can overwrite accumulator banks |
| uses_lut = npu_op.activation is not None and npu_op.activation.op_type == NpuActivationOp.TABLE_LOOKUP |
| written_shram_size = arch.available_shram_banks(uses_lut) * arch.shram_bank_size |
| write_ranges.append(NpuAddressRange(region=BasePointerIndex.Mem2Mem, address=0, length=written_shram_size)) |
| |
| res = MemoryAccessSet() |
| for read_range in read_ranges: |
| if read_range is not None: |
| res.add(memory_range_set(read_range), AccessDirection.Read) |
| for write_range in write_ranges: |
| if write_range is not None: |
| res.add(memory_range_set(write_range), AccessDirection.Write) |
| return res |
| |
| |
| def get_wait_dependency( |
| arch: ArchitectureFeatures, npu_op_list: List[NpuOperation], memory_accesses, op_index: int, watermark: Watermark |
| ): |
| """Used to calculate whether DMA wait or kernel wait operations are needed""" |
| npu_op = npu_op_list[op_index] |
| op_access = memory_accesses[npu_op] |
| index = op_index - 1 |
| |
| # NPU dependency tracking |
| npu_outstanding = -1 |
| npu_ops = 0 |
| npu_index = watermark.npu |
| |
| # DMA dependency tracking |
| dma_outstanding = -1 |
| dma_ops = 0 |
| dma_index = watermark.dma |
| |
| # Seek back in the command stream looking for NPU or DMA dependencies |
| # but only as far as the first dependency or the watermarks (dependencies |
| # before this point have been satisfied already). |
| # The watermark moves to after the latest element we must wait for, not |
| # the command that issues the wait. |
| # NPU->NPU dependency is handled via blockdep. |
| while (index >= npu_index) or (index >= dma_index): |
| prev_op = npu_op_list[index] |
| prev_access = memory_accesses[prev_op] |
| |
| # Check NPU consuming DMA output |
| if is_dma_op(prev_op): |
| if index >= dma_index: |
| if not is_dma_op(npu_op): |
| if (dma_outstanding == -1) and prev_access.conflicts(op_access): |
| dma_outstanding = dma_ops |
| dma_ops += 1 # Count DMA ops in the pipeline |
| if dma_ops >= arch.max_outstanding_dma: |
| dma_index = max(index + 1, dma_index) |
| # Check DMA consuming NPU output |
| else: |
| if index >= npu_index: |
| if is_dma_op(npu_op) and npu_outstanding == -1 and prev_access.conflicts(op_access): |
| npu_outstanding = npu_ops |
| npu_ops += 1 # Count NPU ops in the pipeline |
| if npu_ops >= arch.max_outstanding_kernels: |
| npu_index = max(index + 1, npu_index) |
| |
| index -= 1 |
| |
| # Update DMA watermark if we didn't see any and the NPU pipeline is full |
| if (dma_ops == 0) and (npu_ops >= arch.max_outstanding_kernels): |
| dma_index = op_index |
| |
| # Bring the search watermark forwards as we complete for those dependencies |
| watermark = Watermark(npu_index, dma_index) |
| outstanding = Watermark(npu_outstanding, dma_outstanding) |
| |
| return watermark, outstanding |
| |
| |
| def generate_cmd_waits(emit: CommandStreamEmitter, cmd_waits: Watermark): |
| if cmd_waits.npu >= 0: |
| emit.cmd_wait(cmd0.NPU_OP_KERNEL_WAIT, 0, cmd_waits.npu) |
| |
| if cmd_waits.dma >= 0: |
| emit.cmd_wait(cmd0.NPU_OP_DMA_WAIT, 0, cmd_waits.dma) |
| |
| |
| # ------------------------------------------------------------------- |
| # BLOCKDEP |
| # ------------------------------------------------------------------- |
| |
| |
| def is_dependent_on_prev_op(prev_op: NpuBlockOperation, npu_op: NpuBlockOperation) -> bool: |
| """Checks if npu_op's input is dependent on prev_op's output""" |
| assert npu_op.ifm is not None |
| assert prev_op.ofm is not None |
| curr_input_ranges = get_address_ranges(npu_op.ifm) |
| |
| if has_ifm2(npu_op): |
| assert npu_op.ifm2 is not None |
| curr_input_ranges.extend(get_address_ranges(npu_op.ifm2)) |
| for prev_range in get_address_ranges(prev_op.ofm): |
| if prev_range is None: |
| continue |
| for curr_range in curr_input_ranges: |
| if curr_range is not None and ranges_overlap(prev_range, curr_range): |
| return True |
| return False |
| |
| |
| def shape3d_to_rect(shape: NpuShape3D) -> Rect: |
| return Rect(0, 0, 0, shape.width - 1, shape.height - 1, shape.depth - 1) |
| |
| |
| def get_ifm_ofm_block_depth(arch: ArchitectureFeatures, npu_op: NpuBlockOperation) -> int: |
| # Note: NOT equivalent to the normal ifm block depth calculation since |
| # it takes into account 'depthless' block operations by returning full |
| # depth |
| if npu_op.op_type == NpuOperationType.Conv2D: |
| res = arch.calc_ifm_block_depth(npu_op.ifm.shape.depth, npu_op.ifm.data_type.size_in_bits()) |
| return res |
| return npu_op.ofm.shape.depth |
| |
| |
| def calc_blockdep( |
| arch: ArchitectureFeatures, |
| prev_op: Optional[NpuBlockOperation], |
| prev_block_config: Optional[NpuShape3D], |
| npu_op: NpuBlockOperation, |
| block_config: NpuShape3D, |
| ) -> int: |
| """Calculates the value of the BLOCKDEP register""" |
| if prev_op is None: |
| return 0 |
| if not is_dependent_on_prev_op(prev_op, npu_op): |
| return ArchitectureFeatures.MAX_BLOCKDEP |
| if prev_op.ofm.shape != npu_op.ifm.shape: |
| return 0 |
| prev_ifm_block_depth = get_ifm_ofm_block_depth(arch, prev_op) |
| prev_ofm_block = Block(prev_block_config.width, prev_block_config.height, prev_block_config.depth) |
| prev_ofm_rect = shape3d_to_rect(prev_op.ofm.shape) |
| prev_ifm_rect = shape3d_to_rect(prev_op.ifm.shape) |
| cur_ifm_block_depth = get_ifm_ofm_block_depth(arch, npu_op) |
| cur_ofm_block = Block(block_config.width, block_config.height, block_config.depth) |
| cur_ofm_rect = shape3d_to_rect(npu_op.ofm.shape) |
| cur_ifm_rect = shape3d_to_rect(npu_op.ifm.shape) |
| cur_padLT = (0, 0) if npu_op.padding is None else (npu_op.padding.left, npu_op.padding.top) |
| blockdep = arch.calc_block_dep( |
| prev_ifm_rect, |
| prev_ofm_rect, |
| prev_ifm_block_depth, |
| prev_ofm_block, |
| to_kernel(prev_op.kernel), |
| cur_ifm_rect, |
| cur_ofm_rect, |
| cur_ifm_block_depth, |
| cur_ofm_block, |
| to_kernel(npu_op.kernel), |
| cur_padLT, |
| ) |
| return blockdep |
| |
| |
| # ------------------------------------------------------------------- |
| # PRINT |
| # ------------------------------------------------------------------- |
| |
| |
| def print_feature_map(fm: NpuFeatureMap, name: str): |
| if fm is not None: |
| q = ( |
| "no quantization" |
| if fm.quantization is None |
| else f"scale: {fm.quantization.scale_f32}, zero: {fm.quantization.zero_point}" |
| ) |
| h, w, c = fm.shape |
| sz = h * w * c * fm.data_type.size_in_bytes() |
| print(f" {name}: h={h},w={w},c={c}, region={fm.region}, {fm.layout}, {fm.data_type}, size={sz}, {q}") |
| strides = get_strides(fm) |
| stride_str = f"Stride y/x/c: {strides.height}/{strides.width}/{strides.depth}" |
| t = fm.tiles |
| addresses = [hex(addr) for addr in t.addresses] |
| print(f" {stride_str}, tiles: w0={t.width_0}, h0={t.height_0}, h1={t.height_1}, base={addresses}") |
| |
| |
| def print_operation(npu_op: NpuOperation, index: int = 0): |
| pass_info = f", {npu_op.cmd}" if hasattr(npu_op, "cmd") else "" |
| if is_dma_op(npu_op): |
| print(f"{index} DMA_START src={npu_op.src}, dest={npu_op.dest}{pass_info}") |
| return |
| k = None if npu_op.kernel is None else to_kernel(npu_op.kernel) |
| if npu_op.op_type in (NpuOperationType.Pooling, NpuOperationType.ElementWise): |
| print(f"{index} {npu_op.sub_op_type.name} {npu_op.op_type.name}:{pass_info}") |
| else: |
| if ( |
| npu_op.op_type == NpuOperationType.Conv2D |
| and k.elements_wh() * k.stride.x * k.stride.y * k.dilation.x * k.dilation.y == 1 |
| ): |
| fc = "FullyConnected " |
| else: |
| fc = "" |
| print(f"{index} {fc}{npu_op.op_type.name}{pass_info}") |
| print_feature_map(npu_op.ifm, "IFM") |
| if npu_op.ifm2_scalar is not None: |
| quant_val = quantise(npu_op.ifm2_scalar, npu_op.ifm2.quantization) |
| print(f" IFM2: Scalar={npu_op.ifm2_scalar} (quantized: {quant_val}), {npu_op.ifm2.quantization}") |
| else: |
| print_feature_map(npu_op.ifm2, "IFM2") |
| print_feature_map(npu_op.ofm, "OFM") |
| if k is not None and npu_op.op_type != NpuOperationType.ElementWise: |
| print(f" Kernel: {k}") |
| if npu_op.padding is not None: |
| print(f" {npu_op.padding}") |
| for weights in npu_op.weights: |
| print(f" Weights: {weights}") |
| for bias in npu_op.biases: |
| print(f" Scales: {bias}") |
| if npu_op.activation is not None: |
| act = npu_op.activation |
| if act.op_type != NpuActivationOp.NONE_OR_RELU or act.min is not None or act.max is not None: |
| lut = f", lut index={act.lookup_table_index}" if act.op_type == NpuActivationOp.TABLE_LOOKUP else "" |
| print(f" Activation: {act.op_type.name}, min={act.min}, max={act.max}{lut}") |
| if npu_op.op_type == NpuOperationType.Conv2D: |
| print(f" {npu_op.block_traversal}") |
| bh, bw, bc = npu_op.block_config |
| rescale = f", rescale={npu_op.rescale}" if hasattr(npu_op, "rescale") else "" |
| print(f" Block config: h={bh},w={bw},c={bc}, {npu_op.ifm_upscale}, {npu_op.rounding_mode}{rescale}") |
| |
| |
| def print_operations(npu_op_list: List[NpuOperation]): |
| for index, npu_op in enumerate(npu_op_list): |
| print_operation(npu_op, index) |
| |
| |
| # ------------------------------------------------------------------- |
| # OPERATIONS |
| # ------------------------------------------------------------------- |
| |
| |
| def generate_operation_code(emit: CommandStreamEmitter, npu_op: NpuOperation): |
| """Generates NPU_OP_* command""" |
| op_type = npu_op.op_type |
| if op_type == NpuOperationType.Dma: |
| emit.cmd_do_operation(cmd0.NPU_OP_DMA_START, npu_op.channel * 16 + npu_op.mode) |
| elif op_type == NpuOperationType.Conv2D: |
| emit.cmd_do_operation(cmd0.NPU_OP_CONV) |
| elif op_type == NpuOperationType.ConvDepthWise: |
| emit.cmd_do_operation(cmd0.NPU_OP_DEPTHWISE) |
| elif op_type == NpuOperationType.Pooling: |
| emit.cmd_do_operation(cmd0.NPU_OP_POOL, param=pooling_op_map[npu_op.sub_op_type]) |
| elif op_type == NpuOperationType.ElementWise: |
| emit.cmd_do_operation(cmd0.NPU_OP_ELEMENTWISE, param=elementwise_op_map[npu_op.sub_op_type]) |
| else: |
| assert 0, "Unsupported operation" |
| |
| |
| def generate_conv2d_op( |
| emit: CommandStreamEmitter, npu_op: NpuConv2DOperation, arch: ArchitectureFeatures |
| ) -> NpuShape3D: |
| """Generates register commands for Conv2D operations""" |
| generate_common(emit, npu_op, npu_op.block_traversal, arch) |
| ifm_resampling_mode = resampling_mode_map[npu_op.ifm_upscale] |
| shared_buffer = shared_buffer_allocation_for_npu_op(arch, npu_op, NpuBlockType.ConvolutionMxN, ifm_resampling_mode) |
| block_config = generate_block_config(emit, npu_op, arch, shared_buffer) |
| generate_shram_registers_non_elementwise(emit, shared_buffer) |
| return block_config |
| |
| |
| def generate_conv_depthwise_op(emit: CommandStreamEmitter, npu_op: NpuPoolingOperation, arch: ArchitectureFeatures): |
| """Generates register commands for depthwise convolution operations""" |
| generate_common(emit, npu_op, NpuBlockTraversal.DEPTH_FIRST, arch) |
| ifm_resampling_mode = resampling_mode_map[npu_op.ifm_upscale] |
| shared_buffer = shared_buffer_allocation_for_npu_op( |
| arch, npu_op, NpuBlockType.ConvolutionDepthWise, ifm_resampling_mode |
| ) |
| block_config = generate_block_config(emit, npu_op, arch, shared_buffer) |
| generate_shram_registers_non_elementwise(emit, shared_buffer) |
| return block_config |
| |
| |
| def generate_pooling_op(emit: CommandStreamEmitter, npu_op: NpuPoolingOperation, arch: ArchitectureFeatures): |
| """Generates register commands for pooling operations""" |
| use_global_scale = ( |
| npu_op.sub_op_type in (NpuPoolingOp.AVERAGE, NpuPoolingOp.REDUCE_SUM) and sum(npu_op.padding) == 0 |
| ) |
| generate_common(emit, npu_op, NpuBlockTraversal.DEPTH_FIRST, arch, use_global_scale=use_global_scale) |
| # Pooling op specific |
| if use_global_scale: |
| generate_ofm_scaling_for_pooling(emit, npu_op) |
| ifm_resampling_mode = resampling_mode_map[npu_op.ifm_upscale] |
| npu_block_type = NpuBlockType.ReduceSum if npu_op.sub_op_type == NpuPoolingOp.REDUCE_SUM else NpuBlockType.Pooling |
| shared_buffer = shared_buffer_allocation_for_npu_op(arch, npu_op, npu_block_type, ifm_resampling_mode) |
| block_config = generate_block_config(emit, npu_op, arch, shared_buffer) |
| generate_shram_registers_non_elementwise(emit, shared_buffer) |
| return block_config |
| |
| |
| def generate_elementwise_op(emit: CommandStreamEmitter, npu_op: NpuElementWiseOperation, arch: ArchitectureFeatures): |
| """Generates register commands for elementwise operations""" |
| use_global_scale = npu_op.sub_op_type in ( |
| NpuElementWiseOp.ADD, |
| NpuElementWiseOp.SUB, |
| NpuElementWiseOp.MUL, |
| NpuElementWiseOp.LRELU, |
| NpuElementWiseOp.ABS, |
| ) |
| op_to_scale = generate_scaling_for_elementwise(emit, npu_op) |
| generate_common( |
| emit, npu_op, NpuBlockTraversal.DEPTH_FIRST, arch, use_global_scale=use_global_scale, op_to_scale=op_to_scale |
| ) |
| # Elementwise op specific |
| if npu_op.sub_op_type not in unary_elementwise_ops: |
| # Binary operation; generate IFM2 registers |
| assert npu_op.ifm2 is not None |
| has_scalar = npu_op.ifm2_scalar is not None |
| generate_ifm2(emit, npu_op.ifm2, has_scalar) |
| generate_ifm_precision(emit, npu_op.ifm2, 0, cmd0.NPU_SET_IFM2_PRECISION) |
| generate_ifm2_broadcast(emit, npu_op) |
| if has_scalar: |
| quantized_scalar = quantise(npu_op.ifm2_scalar, npu_op.ifm2.quantization) |
| assert npu_op.ifm2.data_type.min_value() <= quantized_scalar <= npu_op.ifm2.data_type.max_value() |
| emit.cmd0_with_param(cmd0.NPU_SET_IFM2_SCALAR, quantized_scalar) |
| ifm_resampling_mode = resampling_mode_map[npu_op.ifm_upscale] |
| shared_buffer = shared_buffer_allocation_for_npu_op(arch, npu_op, NpuBlockType.ElementWise, ifm_resampling_mode) |
| block_config = generate_block_config(emit, npu_op, arch, shared_buffer) |
| generate_shram_registers_elementwise(emit, npu_op, arch, shared_buffer) |
| return block_config |
| |
| |
| def generate_dma_op(emit: CommandStreamEmitter, dma_op: NpuDmaOperation): |
| """Generates register commands for DMA operations""" |
| emit.cmd0_with_param(cmd0.NPU_SET_DMA0_SRC_REGION, dma_op.src.region) |
| emit.cmd1_with_offset(cmd1.NPU_SET_DMA0_SRC, dma_op.src.address) |
| emit.cmd0_with_param(cmd0.NPU_SET_DMA0_DST_REGION, dma_op.dest.region) |
| |
| emit.cmd1_with_offset(cmd1.NPU_SET_DMA0_DST, dma_op.dest.address) |
| emit.cmd1_with_offset(cmd1.NPU_SET_DMA0_LEN, dma_op.src.length) |
| |
| |
| def generate_registers_for_op( |
| emit: CommandStreamEmitter, npu_op: NpuOperation, arch: ArchitectureFeatures |
| ) -> Optional[NpuShape3D]: |
| """ |
| Generates register commands for the given operation, but not the final NPU_OP_... command. |
| Returns the selected block config |
| """ |
| op_type = npu_op.op_type |
| block_config = None |
| if op_type == NpuOperationType.Conv2D: |
| block_config = generate_conv2d_op(emit, npu_op, arch) |
| elif op_type == NpuOperationType.ConvDepthWise: |
| block_config = generate_conv_depthwise_op(emit, npu_op, arch) |
| elif op_type == NpuOperationType.Pooling: |
| block_config = generate_pooling_op(emit, npu_op, arch) |
| elif op_type == NpuOperationType.ElementWise: |
| block_config = generate_elementwise_op(emit, npu_op, arch) |
| elif op_type == NpuOperationType.Dma: |
| generate_dma_op(emit, npu_op) |
| else: |
| assert 0, "Unsupported operation" |
| return block_config |
| |
| |
| def generate_command_stream( |
| emit: CommandStreamEmitter, npu_op_list: List[NpuOperation], arch: ArchitectureFeatures, add_to_debug_db=None |
| ): |
| """Generates register commands for the given list of NPU operations""" |
| # Calculate memory accesses for every operation |
| memory_accesses = {} |
| for npu_op in npu_op_list: |
| if is_dma_op(npu_op): |
| memory_accesses[npu_op] = get_dma_memory_accesses(npu_op) |
| else: |
| memory_accesses[npu_op] = get_op_memory_accesses(npu_op, arch) |
| if arch.is_yoda_system: |
| emit.cmd0_with_param(cmd0.NPU_SET_PARALLEL_MODE, arch.ncores - 1) |
| dep_watermark = Watermark(0, 0) |
| prev_op = None |
| prev_block_config = None |
| # Generate register commands for all operations |
| for op_index, npu_op in enumerate(npu_op_list): |
| dep_watermark, cmd_waits = get_wait_dependency(arch, npu_op_list, memory_accesses, op_index, dep_watermark) |
| block_config = generate_registers_for_op(emit, npu_op, arch) |
| if not is_dma_op(npu_op): |
| # Generate BLOCKDEP |
| assert block_config is not None |
| blockdep = calc_blockdep(arch, prev_op, prev_block_config, npu_op, block_config) |
| blockdep = min(blockdep, arch.max_blockdep) |
| emit.cmd0_with_param(cmd0.NPU_SET_BLOCKDEP, blockdep) |
| prev_op = npu_op |
| prev_block_config = block_config |
| |
| generate_cmd_waits(emit, cmd_waits) |
| # Generate the actual NPU_OP command |
| generate_operation_code(emit, npu_op) |
| if add_to_debug_db is not None: |
| add_to_debug_db(npu_op, emit.offset) |
| # Fill in final part of command stream: |
| emit.cmd_do_operation(cmd0.NPU_OP_STOP, param=0xFFFF) |
| |
| |
| def generate_register_command_stream_for_sg(nng, sg, arch, verbose=False): |
| """Generates command stream for the subgraph, adds it to sg.register_command_stream""" |
| # Convert high level command stream to list of NpuOperation |
| npu_op_list = [] |
| npu_op_to_cmd = dict() # map from npu op to high level command |
| for cmd in sg.high_level_command_stream: |
| if cmd.cmdtype == CommandType.NpuStripe and cmd.ps.npu_block_type == NpuBlockType.Default: |
| print("Warning: Skipping register command stream generation for", cmd.ps) |
| else: |
| npu_op = convert_command_to_npu_op(cmd, arch) |
| npu_op_list.append(npu_op) |
| npu_op_to_cmd[npu_op] = cmd |
| if verbose: |
| print_operations(npu_op_list) |
| # Generate register commands |
| stream_id = DebugDatabase.add_stream(sg) |
| DebugDatabase.set_stream_offset(sg, 0) # Default to zero, can only set during file writing |
| emit = CommandStreamEmitter() |
| |
| def add_to_debug_db(npu_op: NpuOperation, offset: int): |
| """Adds info to the debug database""" |
| if not is_dma_op(npu_op): |
| cmd = npu_op_to_cmd[npu_op] |
| DebugDatabase.add_command(stream_id, offset, cmd.ps.primary_op) |
| |
| generate_command_stream(emit, npu_op_list, arch, add_to_debug_db) |
| sg.register_command_stream = emit.to_list() |
| if verbose: |
| emit.print_cmds() |
| print("number of commands", len(emit.cmd_stream)) |
| print("command stream length in words", len(sg.register_command_stream)) |
| |
| |
| def generate_register_command_stream(npu_op_list: List[NpuOperation], accelerator: Accelerator) -> List[int]: |
| """ |
| Public facing API for generating an ethosu register command stream. |
| Calculates dependencies between commands and inserts wait operations if needed. |
| |
| :param npu_op_list: List[NpuOperation] list of high level NPU operations |
| :param accelerator: architecture_features.Accelerator enum to pick the correct ethosu accelerator |
| :return ethosu instructions, as a list of 32-bit integers |
| """ |
| emit = CommandStreamEmitter() |
| arch = ArchitectureFeatures( |
| vela_config=None, |
| system_config=None, |
| accelerator_config=accelerator.value, |
| override_block_config=None, |
| block_config_limit=None, |
| global_memory_clock_scale=1.0, |
| max_blockdep=ArchitectureFeatures.MAX_BLOCKDEP, |
| weight_estimation_scaling=1.0, |
| ) |
| generate_command_stream(emit, npu_op_list, arch) |
| return emit.to_list() |