ethosu/vela/cascade_builder.py

# Copyright (C) 2021 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:
# Groups Operators in a schedule together to form Cascades.
from collections import namedtuple

from .numeric_util import round_up
from .operation import NpuBlockType
from .operation import Op
from .operation import Padding
from .shape4d import Shape4D

non_cascadable_blocks = (
    NpuBlockType.Default,
    NpuBlockType.VectorProduct,
    NpuBlockType.ReduceSum,
)


class CascadeInfo:
    """Contains metadata about a cascade"""

    def __init__(self, start, end, buffers, mem_usage: int):
        self.start = start
        self.end = end
        self.buffers = buffers
        self.mem_usage = mem_usage


class BufferMap:
    """Caches the buffers seen"""

    def __init__(self):
        self.buffer_map = {}

    def get_buffer(self, producer, consumer, cost):
        assert producer or consumer
        key = (producer, consumer)
        if key not in self.buffer_map:
            # No cached buffer between these two SchedulerOperations
            if consumer is None:
                # There are either no consumers or multiple consumers - FeatureMap needs to be stored in full
                buffer_shape = producer.ofm.shape
                buffer_size = producer.ofm_size_in_bytes()
            elif producer is None:
                # First Op in subgraph or cascade - FeatureMap needs to be stored in full
                buffer_shape = consumer.ifm.shape
                buffer_size = consumer.ifm_size_in_bytes()
            elif producer.requires_full_ofm or consumer.requires_full_ifm:
                # FeatureMap needs to be stored in full
                buffer_shape = max(producer.ofm.shape, consumer.ifm.shape)
                buffer_size = max(producer.ofm_size_in_bytes(), consumer.ifm_size_in_bytes())
            else:
                # Use a rolling buffer
                buffer_shape = rolling_buffer_shape(cost[producer].stripe, cost[consumer].stripe_input)
                buffer_size = buffer_shape.elements() * producer.ofm.dtype.size_in_bytes()

            self.buffer_map[key] = (buffer_shape, buffer_size)

        return self.buffer_map[key]


def rolling_buffer_shape(producer_stripe: Shape4D, consumer_stripe_input: Shape4D) -> Shape4D:
    """Calculates the storage shape of the rolling buffer between two SchedulerOperations in a Cascade"""
    buffer_height = round_up(producer_stripe.height + consumer_stripe_input.height, consumer_stripe_input.height)
    # Rolling buffers have to conform to NHCWB16 format
    return consumer_stripe_input.with_height(buffer_height).with_depth(round_up(producer_stripe.depth, 16))


class CascadeBuilder:
    """Class for grouping SchedulerOperations into cascades"""

    def __init__(self, sched_ops, spilling, non_local_mem_usage=None):
        self.sched_ops = sched_ops
        self.no_cascade = 0
        self.non_local_mem_usage = non_local_mem_usage if non_local_mem_usage else {}
        self.spilling = spilling

    def _is_cascadable(self, sched_op, cost) -> bool:
        """Checks if 'sched_op' can be cascaded"""

        return (
            sched_op.op_type.npu_block_type not in non_cascadable_blocks
            and cost.stripe.height < sched_op.ofm.shape.height
            and sched_op.parent_op.read_offsets[0] is None
            and sched_op.parent_op.read_offsets[1] is None
            and self.element_wise_cascading_conformity(sched_op)
            and not sched_op.parent_op.type.is_resize_op()
            and sched_op.parent_op.attrs.get("padding", None) != Padding.TILE
        )

    def _is_mergeable(self, sched_op) -> bool:
        # Code based on merge_elementwise_op_ranges from live_range.py

        if not sched_op.op_type.is_elementwise_op():
            return False

        elem_op = sched_op.parent_op

        # Check if overwriting the inputs can be allowed
        OpShapeTens = namedtuple("OpShapeTens", ["op_shape", "tens"])
        outp = OpShapeTens(elem_op.ofm_shapes[0], elem_op.ofm)

        # check output tensor only has one producer
        if len(outp.tens.ops) != 1:
            return False

        inps = []
        if elem_op.ifm is not None:
            inps.append(OpShapeTens(elem_op.ifm_shapes[0], elem_op.ifm))
        if elem_op.ifm2 is not None:
            inps.append(OpShapeTens(elem_op.ifm_shapes[1], elem_op.ifm2))

        # find an input tensor that can be overwritten by the output
        for inp in inps:
            if (
                # check op input and output shapes allow overlapping
                inp.op_shape == outp.op_shape
                # check input tensor is valid
                and inp.tens is not None
                and inp.tens.shape != []
                # check input and output tensors are compatible
                and inp.tens.format == outp.tens.format
                and inp.tens.dtype == outp.tens.dtype
                # check input tensor only has one consumer
                and len(inp.tens.consumer_list) == 1
            ):
                return True

        return False

    def _estimate_sram_usage(self, sched_op, cost) -> int:
        """Estimate the SRAM required for the Op if all FeatureMaps are in SRAM"""
        ifm2_size = sched_op.ifm2_size_in_bytes()
        if sched_op.requires_full_ifm:
            ifm_size = sched_op.ifm_size_in_bytes()
        else:
            ifm_size = (
                cost.stripe_input.with_depth(round_up(cost.stripe_input.depth, 16)).elements()
                * sched_op.ifm.dtype.size_in_bytes()
            )
        if sched_op.requires_full_ofm:
            ofm_size = sched_op.ofm_size_in_bytes()
        else:
            ofm_size = (
                cost.stripe.with_depth(round_up(cost.stripe.depth, 16)).elements() * sched_op.ofm.dtype.size_in_bytes()
            )

        if self._is_mergeable(sched_op):
            # ofm will use the ifm buffer to reduce SRAM usage, hence ofm_size = 0
            ofm_size = 0

        return ifm_size + ifm2_size + ofm_size + self.non_local_mem_usage.get(sched_op, 0)

    @staticmethod
    def element_wise_cascading_conformity(sched_op):
        """Check the inputs of the op to see if it's a candidate for cascading."""
        # Cascading sub-operators of Softmax results in a crash when handling Sub and RescaleAdd ops

        ifm = sched_op.parent_op.ifm
        ifm2 = sched_op.parent_op.ifm2

        if sched_op.op_type in [Op.RescaleAdd]:
            return False

        if sched_op.parent_op.type.is_binary_elementwise_op() and ifm and ifm2:
            # We cannot rule out cascadability if at least one IFM is constant
            return Op.Const in (ifm.ops[0], ifm2.ops[0])
        else:
            # Either one IFM is not variable or it is not a binary elementwise op - we cannot rule out cascadability
            return True

    def build_cascades(self, ref_schedule, fallback_schedule, guiding_mem_limit):
        ref_cost = ref_schedule.cost_map
        fallback_cost = fallback_schedule.cost_map
        cost = {}
        cascade_map = {}
        buffers = BufferMap()

        # Peak memory usage so far - updated continously, unless dedicated SRAM where this is a hard limit
        peak_sram_usage = guiding_mem_limit

        idx = 0
        while idx < len(self.sched_ops):
            op = self.sched_ops[idx]
            if op in cost:
                # Already processed this Op
                idx += 1
                continue

            if not self._is_cascadable(op, ref_cost[op]):
                # Op is not a candidate for cascading - assign fallback cost
                cost[op] = fallback_cost[op]
                if not self.spilling:
                    peak_sram_usage = max(self._estimate_sram_usage(op, fallback_cost[op]), peak_sram_usage)
                idx += 1
                continue

            # Propose a cascade starting with this Op
            cascade_start = op.index
            # Keep track of which Ops are in the proposed cascade as well as the best cascade so far
            ops_in_cascade = [op]
            ops_in_best_cascade = [op]
            # Get the size of the weight buffer(s)
            weight_buffer = sum(tens.storage_size() for tens in ref_cost[op].buffered_weight_tensors)

            # The first IFM needs to be stored in full
            cascade_ifm_size = op.ifm_size_in_bytes() if not self.spilling else 0

            # Add non-local memory usage
            cascade_ifm_size += self.non_local_mem_usage.get(op, 0)

            # Sum of all intermediate cascade buffers (including weight buffers)
            cascade_buffers = weight_buffer
            # Best cascade size - Initially it's the fallback cost of the first Op in the cascade
            best_cascade_size = self._estimate_sram_usage(op, fallback_cost[op])

            # Op is the producer of the OFM consumed by the next Op to consider
            producer = op
            while True:
                dependants = producer.get_dependants()
                if len(dependants) != 1:
                    # producer is either the last Op in the schedule or the start of a branch
                    break

                current_op = dependants[0]
                if (
                    current_op in cost
                    or current_op not in ref_cost
                    or not self._is_cascadable(current_op, ref_cost[current_op])
                    or producer.ofm.shape != current_op.ifm.shape
                    or current_op.requires_full_ifm
                    or producer.requires_full_ofm
                ):
                    # Current op has already been processed or cannot be cascaded
                    break

                if producer.index + 1 != current_op.index:
                    # Cascading is possible, but requires reordering of operations in the schedule,
                    # this is currently not supported
                    break

                # Get the size of the FeatureMap buffers between current and neighbouring Ops
                op_full_ifm = current_op.ifm_size_in_bytes()
                op_full_ofm = current_op.ofm_size_in_bytes()
                _, op_ifm_buffer = buffers.get_buffer(producer, current_op, ref_cost)

                # Get the size of the weight buffer(s)
                op_weight_buffer = sum(tens.storage_size() for tens in ref_cost[current_op].buffered_weight_tensors)

                # Calculate the uncascaded memory requirement for current Op
                uncascaded_sram_usage = op_full_ifm + op_full_ofm + self.non_local_mem_usage.get(current_op, 0)

                # Add current Op to cascade
                ops_in_cascade.append(current_op)

                # Increase the accumulated intermediate buffers in the cascade
                cascade_buffers += op_ifm_buffer + op_weight_buffer

                if self.spilling:
                    # For Dedicated SRAM only the intermediate buffers are in SRAM
                    if uncascaded_sram_usage < peak_sram_usage or cascade_buffers > peak_sram_usage:
                        # Cascade until an Op fits in its entirety or the accumulated buffers no longer fit
                        break
                    else:
                        # Any addition to the cascade that fits is the new best cascade for Dedicated SRAM
                        ops_in_best_cascade = [op for op in ops_in_cascade]
                        best_cascade_size = cascade_buffers

                else:
                    # Calculate the total size of the current cascade
                    cascade_size = cascade_ifm_size + cascade_buffers + op_full_ofm

                    # Determine if cascading search should stop
                    if (
                        uncascaded_sram_usage < peak_sram_usage
                        and best_cascade_size < peak_sram_usage
                        or (cascade_ifm_size + cascade_buffers) > best_cascade_size
                    ):
                        # Both the existing cascade and current Op fits
                        break

                    """
                    One of two conditions will update the best cascade:

                    - cascade_size < best_cascade_size or
                    - cascade_size < uncascaded_sram_usage

                    The last condition is illustrated below, showing an example where it is
                    better to choose a larger cascade_size (with more OPs) because it will
                    use less total SRAM usage.

                    For simplicity, all featuremaps have same size.

                    Cascade OP1-OP2, OP3 is standalone

                                ->  |OP1| -> roll buffer -> |OP2| -> FM -> |OP3| -> FM
                               /
                    |OP0| -> FM
                               \
                                ->  ....


                    best_cascade_size    : FM + roll buffer + FM
                    uncascaded_sram_usage: FM + FM + FM

                    compared with:

                    Cascade OP1-OP3

                                ->  |OP1| -> roll buffer -> |OP2| -> roll buffer -> |OP3| -> FM
                               /
                    |OP0| -> FM
                               \
                                ->  ....


                    cascade_size         : FM + roll buffer + roll buffer + FM


                    So, for this use case the comparison will be

                    (FM + roll buffer + roll buffer + FM) <  (FM + roll buffer + FM) or
                    (FM + roll buffer + roll buffer + FM) <  (FM + FM + FM)

                    hence, better to choose Cascade OP1-OP3 in this case.
                    """
                    if cascade_size < best_cascade_size or cascade_size < uncascaded_sram_usage:
                        best_cascade_size = cascade_ifm_size + cascade_buffers + op_full_ofm
                        ops_in_best_cascade = [op for op in ops_in_cascade]

                producer = current_op

            if len(ops_in_best_cascade) > 1:
                # A cascade was created - assign cascade and ref_cost to all of the Ops
                cascade_end = cascade_start + (len(ops_in_best_cascade) - 1)
                buffers_in_cascade = {}
                prev_op = None
                for cascaded_op in ops_in_best_cascade:
                    assert cascade_start <= cascaded_op.index <= cascade_end
                    cost[cascaded_op] = ref_cost[cascaded_op]
                    cost[cascaded_op].cascade = cascade_end
                    if prev_op:
                        rolling_buffer_shape, _ = buffers.get_buffer(prev_op, cascaded_op, ref_cost)
                        buffers_in_cascade[cascaded_op] = rolling_buffer_shape

                    prev_op = cascaded_op

                # Create a CascadeInfo for the cascade
                cascade_map[cascade_end] = CascadeInfo(
                    cascade_start, cascade_end, buffers_in_cascade, best_cascade_size
                )
                if not self.spilling:
                    # Update peak memory usage
                    peak_sram_usage = max(best_cascade_size, peak_sram_usage)
            else:
                # Assign fallback cost to the initial Op
                cost[op] = fallback_cost[op]
                if not self.spilling:
                    peak_sram_usage = max(self._estimate_sram_usage(op, fallback_cost[op]), peak_sram_usage)

        # Update costing and cascade information for the ref_schedule
        ref_schedule.cost_map = cost
        ref_schedule.cascades = cascade_map
        return ref_schedule