From 75c839bfb80e1632aa8a5aaaad8bedff69d5319a Mon Sep 17 00:00:00 2001
From: Simon Bjurek <simbj106@student.liu.se>
Date: Wed, 19 Feb 2025 15:38:53 +0000
Subject: [PATCH] Add IO times for list schedulers
---
b_asic/core_schedulers.py | 13 +-
b_asic/resources.py | 2 +-
b_asic/schedule.py | 4 +-
b_asic/scheduler.py | 147 +++-
b_asic/sfg_generators.py | 4 +-
.../auto_scheduling_with_custom_io_times.py | 85 +++
examples/ldlt_matrix_inverse.py | 28 +-
examples/memory_constrained_scheduling.py | 134 ++++
test/test_core_schedulers.py | 642 +++++++++++++++++-
9 files changed, 1026 insertions(+), 33 deletions(-)
create mode 100644 examples/auto_scheduling_with_custom_io_times.py
create mode 100644 examples/memory_constrained_scheduling.py
diff --git a/b_asic/core_schedulers.py b/b_asic/core_schedulers.py
index 1b5b4f18..32cb23db 100644
--- a/b_asic/core_schedulers.py
+++ b/b_asic/core_schedulers.py
@@ -124,7 +124,8 @@ class EarliestDeadlineScheduler(ListScheduler):
deadlines = {}
for op_id, start_time in schedule_copy.start_times.items():
- deadlines[op_id] = start_time + schedule.sfg.find_by_id(op_id).latency
+ if not op_id.startswith("in"):
+ deadlines[op_id] = start_time + schedule.sfg.find_by_id(op_id).latency
return sorted(deadlines, key=deadlines.get)
@@ -137,7 +138,10 @@ class LeastSlackTimeScheduler(ListScheduler):
schedule_copy = copy.copy(schedule)
ALAPScheduler().apply_scheduling(schedule_copy)
- return sorted(schedule_copy.start_times, key=schedule_copy.start_times.get)
+ sorted_ops = sorted(
+ schedule_copy.start_times, key=schedule_copy.start_times.get
+ )
+ return [op for op in sorted_ops if not op.startswith("in")]
class MaxFanOutScheduler(ListScheduler):
@@ -152,7 +156,8 @@ class MaxFanOutScheduler(ListScheduler):
for op_id, start_time in schedule_copy.start_times.items():
fan_outs[op_id] = len(schedule.sfg.find_by_id(op_id).output_signals)
- return sorted(fan_outs, key=fan_outs.get, reverse=True)
+ sorted_ops = sorted(fan_outs, key=fan_outs.get, reverse=True)
+ return [op for op in sorted_ops if not op.startswith("in")]
class HybridScheduler(ListScheduler):
@@ -199,4 +204,4 @@ class HybridScheduler(ListScheduler):
sorted_op_list = [pair[0] for pair in fan_out_sorted_items]
- return sorted_op_list
+ return [op for op in sorted_op_list if not op.startswith("in")]
diff --git a/b_asic/resources.py b/b_asic/resources.py
index 97dffb68..bc850f1f 100644
--- a/b_asic/resources.py
+++ b/b_asic/resources.py
@@ -907,7 +907,7 @@ class ProcessCollection:
def split_on_ports(
self,
- heuristic: str = "left_edge",
+ heuristic: str = "graph_color",
read_ports: Optional[int] = None,
write_ports: Optional[int] = None,
total_ports: Optional[int] = None,
diff --git a/b_asic/schedule.py b/b_asic/schedule.py
index 6cfabb17..6e231c2c 100644
--- a/b_asic/schedule.py
+++ b/b_asic/schedule.py
@@ -119,9 +119,9 @@ class Schedule:
self._remove_delays_no_laps()
max_end_time = self.get_max_end_time()
- if schedule_time is None:
+ if not self._schedule_time:
self._schedule_time = max_end_time
- elif schedule_time < max_end_time:
+ elif self._schedule_time < max_end_time:
raise ValueError(f"Too short schedule time. Minimum is {max_end_time}.")
def __str__(self) -> str:
diff --git a/b_asic/scheduler.py b/b_asic/scheduler.py
index 4f1a58f2..c4d5de35 100644
--- a/b_asic/scheduler.py
+++ b/b_asic/scheduler.py
@@ -1,3 +1,4 @@
+import sys
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING, Optional, cast
@@ -9,6 +10,7 @@ from b_asic.types import TypeName
if TYPE_CHECKING:
from b_asic.operation import Operation
from b_asic.schedule import Schedule
+ from b_asic.signal_flow_graph import SFG
from b_asic.types import GraphID
@@ -44,9 +46,21 @@ class Scheduler(ABC):
] + cast(int, source_port.latency_offset)
+# TODO: Rename max_concurrent_reads/writes to max_concurrent_read_ports or something to signify difference
+
+
class ListScheduler(Scheduler, ABC):
- def __init__(self, max_resources: Optional[dict[TypeName, int]] = None) -> None:
- if max_resources:
+ def __init__(
+ self,
+ max_resources: Optional[dict[TypeName, int]] = None,
+ max_concurrent_reads: Optional[int] = None,
+ max_concurrent_writes: Optional[int] = None,
+ input_times: Optional[dict["GraphID", int]] = None,
+ output_delta_times: Optional[dict["GraphID", int]] = None,
+ cyclic: Optional[bool] = False,
+ ) -> None:
+ super()
+ if max_resources is not None:
if not isinstance(max_resources, dict):
raise ValueError("max_resources must be a dictionary.")
for key, value in max_resources.items():
@@ -54,12 +68,20 @@ class ListScheduler(Scheduler, ABC):
raise ValueError("max_resources key must be a valid type_name.")
if not isinstance(value, int):
raise ValueError("max_resources value must be an integer.")
-
- if max_resources:
self._max_resources = max_resources
else:
self._max_resources = {}
+ self._max_concurrent_reads = (
+ max_concurrent_reads if max_concurrent_reads else sys.maxsize
+ )
+ self._max_concurrent_writes = (
+ max_concurrent_writes if max_concurrent_writes else sys.maxsize
+ )
+
+ self._input_times = input_times if input_times else {}
+ self._output_delta_times = output_delta_times if output_delta_times else {}
+
def apply_scheduling(self, schedule: "Schedule") -> None:
"""Applies the scheduling algorithm on the given Schedule.
@@ -69,40 +91,63 @@ class ListScheduler(Scheduler, ABC):
Schedule to apply the scheduling algorithm on.
"""
sfg = schedule.sfg
- start_times = schedule.start_times
used_resources_ready_times = {}
remaining_resources = self._max_resources.copy()
sorted_operations = self._get_sorted_operations(schedule)
- # place all inputs at time 0
+ schedule.start_times = {}
+
+ remaining_reads = self._max_concurrent_reads
+
+ # initial input placement
+ if self._input_times:
+ for input_id in self._input_times:
+ schedule.start_times[input_id] = self._input_times[input_id]
+
for input_op in sfg.find_by_type_name(Input.type_name()):
- start_times[input_op.graph_id] = 0
+ if input_op.graph_id not in self._input_times:
+ schedule.start_times[input_op.graph_id] = 0
current_time = 0
+ timeout_counter = 0
while sorted_operations:
# generate the best schedulable candidate
candidate = sfg.find_by_id(sorted_operations[0])
counter = 0
while not self._candidate_is_schedulable(
- start_times,
+ schedule.start_times,
+ sfg,
candidate,
current_time,
remaining_resources,
+ remaining_reads,
+ self._max_concurrent_writes,
sorted_operations,
):
if counter == len(sorted_operations):
counter = 0
current_time += 1
+ timeout_counter += 1
+
+ if timeout_counter > 10:
+ msg = "Algorithm did not schedule any operation for 10 time steps, try relaxing constraints."
+ raise TimeoutError(msg)
+
+ remaining_reads = self._max_concurrent_reads
+
# update available operators
for operation, ready_time in used_resources_ready_times.items():
if ready_time == current_time:
remaining_resources[operation.type_name()] += 1
+
else:
candidate = sfg.find_by_id(sorted_operations[counter])
counter += 1
+ timeout_counter = 0
+
# if the resource is constrained, update remaining resources
if candidate.type_name() in remaining_resources:
remaining_resources[candidate.type_name()] -= 1
@@ -115,19 +160,24 @@ class ListScheduler(Scheduler, ABC):
current_time + candidate.latency
)
+ remaining_reads -= candidate.input_count
+
# schedule the best candidate to the current time
sorted_operations.remove(candidate.graph_id)
- start_times[candidate.graph_id] = current_time
-
- schedule.set_schedule_time(current_time)
+ schedule.start_times[candidate.graph_id] = current_time
self._handle_outputs(schedule)
+
+ if not schedule.cyclic:
+ max_start_time = max(schedule.start_times.values())
+ if current_time < max_start_time:
+ current_time = max_start_time
+ current_time = max(current_time, schedule.get_max_end_time())
+ schedule.set_schedule_time(current_time)
+
schedule.remove_delays()
- # move all inputs ALAP now that operations have moved
- for input_op in schedule.sfg.find_by_type_name(Input.type_name()):
- input_op = cast(Input, input_op)
- schedule.move_operation_alap(input_op.graph_id)
+ self._handle_inputs(schedule)
# move all dont cares ALAP
for dc_op in schedule.sfg.find_by_type_name(DontCare.type_name()):
@@ -137,9 +187,12 @@ class ListScheduler(Scheduler, ABC):
@staticmethod
def _candidate_is_schedulable(
start_times: dict["GraphID"],
+ sfg: "SFG",
operation: "Operation",
current_time: int,
remaining_resources: dict["GraphID", int],
+ remaining_reads: int,
+ max_concurrent_writes: int,
remaining_ops: list["GraphID"],
) -> bool:
if (
@@ -148,23 +201,79 @@ class ListScheduler(Scheduler, ABC):
):
return False
+ op_finish_time = current_time + operation.latency
+ future_ops = [
+ sfg.find_by_id(item[0])
+ for item in start_times.items()
+ if item[1] + sfg.find_by_id(item[0]).latency == op_finish_time
+ ]
+
+ future_ops_writes = sum([op.input_count for op in future_ops])
+
+ if (
+ not operation.graph_id.startswith("out")
+ and future_ops_writes >= max_concurrent_writes
+ ):
+ return False
+
+ read_counter = 0
earliest_start_time = 0
for op_input in operation.inputs:
source_op = op_input.signals[0].source.operation
+ if isinstance(source_op, Delay):
+ continue
+
source_op_graph_id = source_op.graph_id
if source_op_graph_id in remaining_ops:
return False
+ if start_times[source_op_graph_id] != current_time - 1:
+ # not a direct connection -> memory read required
+ read_counter += 1
+
+ if read_counter > remaining_reads:
+ return False
+
proceeding_op_start_time = start_times.get(source_op_graph_id)
+ proceeding_op_finish_time = proceeding_op_start_time + source_op.latency
+
+ # if not proceeding_op_finish_time == current_time:
+ # # not direct connection -> memory required, check if okay
+ # satisfying_remaining_reads = remaining_reads >= operation.input_count
+ # satisfying_remaining_writes = remaining_writes >= operation.output_count
+ # if not (satisfying_remaining_reads and satisfying_remaining_writes):
+ # return False
- if not isinstance(source_op, Delay):
- earliest_start_time = max(
- earliest_start_time, proceeding_op_start_time + source_op.latency
- )
+ earliest_start_time = max(earliest_start_time, proceeding_op_finish_time)
return earliest_start_time <= current_time
@abstractmethod
def _get_sorted_operations(schedule: "Schedule") -> list["GraphID"]:
raise NotImplementedError
+
+ def _handle_inputs(self, schedule: "Schedule") -> None:
+ for input_op in schedule.sfg.find_by_type_name(Input.type_name()):
+ input_op = cast(Input, input_op)
+ if input_op.graph_id not in self._input_times:
+ schedule.move_operation_alap(input_op.graph_id)
+
+ def _handle_outputs(
+ self, schedule: "Schedule", non_schedulable_ops: Optional[list["GraphID"]] = []
+ ) -> None:
+ super()._handle_outputs(schedule, non_schedulable_ops)
+
+ schedule.set_schedule_time(schedule.get_max_end_time())
+
+ for output in schedule.sfg.find_by_type_name(Output.type_name()):
+ output = cast(Output, output)
+ if output.graph_id in self._output_delta_times:
+ delta_time = self._output_delta_times[output.graph_id]
+ if schedule.cyclic:
+ schedule.start_times[output.graph_id] = schedule.schedule_time
+ schedule.move_operation(output.graph_id, delta_time)
+ else:
+ schedule.start_times[output.graph_id] = (
+ schedule.schedule_time + delta_time
+ )
diff --git a/b_asic/sfg_generators.py b/b_asic/sfg_generators.py
index 3e76f159..2d68215c 100644
--- a/b_asic/sfg_generators.py
+++ b/b_asic/sfg_generators.py
@@ -415,7 +415,7 @@ def radix_2_dif_fft(points: int) -> SFG:
inputs = []
for i in range(points):
- inputs.append(Input(name=f"Input: {i}"))
+ inputs.append(Input())
ports = inputs
number_of_stages = int(np.log2(points))
@@ -430,7 +430,7 @@ def radix_2_dif_fft(points: int) -> SFG:
ports = _get_bit_reversed_ports(ports)
outputs = []
for i, port in enumerate(ports):
- outputs.append(Output(port, name=f"Output: {i}"))
+ outputs.append(Output(port))
return SFG(inputs=inputs, outputs=outputs)
diff --git a/examples/auto_scheduling_with_custom_io_times.py b/examples/auto_scheduling_with_custom_io_times.py
new file mode 100644
index 00000000..8913bfd8
--- /dev/null
+++ b/examples/auto_scheduling_with_custom_io_times.py
@@ -0,0 +1,85 @@
+"""
+=========================================
+Auto Scheduling With Custom IO times
+=========================================
+
+"""
+
+from b_asic.core_operations import Butterfly, ConstantMultiplication
+from b_asic.core_schedulers import ASAPScheduler, HybridScheduler
+from b_asic.schedule import Schedule
+from b_asic.sfg_generators import radix_2_dif_fft
+
+sfg = radix_2_dif_fft(points=8)
+
+# %%
+# The SFG is
+sfg
+
+# %%
+# Set latencies and execution times.
+sfg.set_latency_of_type(Butterfly.type_name(), 3)
+sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+sfg.set_execution_time_of_type(Butterfly.type_name(), 1)
+sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+
+# %%
+# Generate an ASAP schedule for reference
+schedule = Schedule(sfg, scheduler=ASAPScheduler())
+schedule.show()
+
+# %%
+# Generate a non-cyclic Schedule from HybridScheduler with custom IO times.
+resources = {Butterfly.type_name(): 1, ConstantMultiplication.type_name(): 1}
+input_times = {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ "in3": 3,
+ "in4": 4,
+ "in5": 5,
+ "in6": 6,
+ "in7": 7,
+}
+output_delta_times = {
+ "out0": -2,
+ "out1": -1,
+ "out2": 0,
+ "out3": 1,
+ "out4": 2,
+ "out5": 3,
+ "out6": 4,
+ "out7": 5,
+}
+schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ resources,
+ input_times=input_times,
+ output_delta_times=output_delta_times,
+ ),
+)
+schedule.show()
+
+# %%
+# Generate a new Schedule with cyclic scheduling enabled
+output_delta_times = {
+ "out0": 0,
+ "out1": 1,
+ "out2": 2,
+ "out3": 3,
+ "out4": 4,
+ "out5": 5,
+ "out6": 6,
+ "out7": 7,
+}
+schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ resources,
+ input_times=input_times,
+ output_delta_times=output_delta_times,
+ ),
+ cyclic=True,
+)
+schedule.show()
diff --git a/examples/ldlt_matrix_inverse.py b/examples/ldlt_matrix_inverse.py
index 845bd14b..cf5961aa 100644
--- a/examples/ldlt_matrix_inverse.py
+++ b/examples/ldlt_matrix_inverse.py
@@ -64,8 +64,31 @@ print("Scheduling time:", schedule.schedule_time)
schedule.show()
# %%
-# Create a HybridScheduler schedule that satisfies the resource constraints.
-schedule = Schedule(sfg, scheduler=HybridScheduler(resources))
+# Create a HybridScheduler schedule that satisfies the resource constraints with custom IO times.
+# This is the schedule we will synthesize an architecture for.
+input_times = {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ "in3": 3,
+ "in4": 4,
+ "in5": 5,
+}
+output_delta_times = {
+ "out0": 0,
+ "out1": 1,
+ "out2": 2,
+ "out3": 3,
+ "out4": 4,
+ "out5": 5,
+}
+schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ resources, input_times=input_times, output_delta_times=output_delta_times
+ ),
+ cyclic=True,
+)
print("Scheduling time:", schedule.schedule_time)
schedule.show()
@@ -116,4 +139,3 @@ arch = Architecture(
# %%
arch
-# schedule.edit()
diff --git a/examples/memory_constrained_scheduling.py b/examples/memory_constrained_scheduling.py
new file mode 100644
index 00000000..6c0ea9a8
--- /dev/null
+++ b/examples/memory_constrained_scheduling.py
@@ -0,0 +1,134 @@
+"""
+=========================================
+Memory Constrained Scheduling
+=========================================
+
+"""
+
+from b_asic.architecture import Architecture, Memory, ProcessingElement
+from b_asic.core_operations import Butterfly, ConstantMultiplication
+from b_asic.core_schedulers import ASAPScheduler, HybridScheduler
+from b_asic.schedule import Schedule
+from b_asic.sfg_generators import radix_2_dif_fft
+from b_asic.special_operations import Input, Output
+
+sfg = radix_2_dif_fft(points=16)
+
+# %%
+# The SFG is
+sfg
+
+# %%
+# Set latencies and execution times.
+sfg.set_latency_of_type(Butterfly.type_name(), 3)
+sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+sfg.set_execution_time_of_type(Butterfly.type_name(), 1)
+sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+
+# # %%
+# Generate an ASAP schedule for reference
+schedule = Schedule(sfg, scheduler=ASAPScheduler())
+schedule.show()
+
+# %%
+# Generate a PE constrained HybridSchedule
+resources = {Butterfly.type_name(): 1, ConstantMultiplication.type_name(): 1}
+schedule = Schedule(sfg, scheduler=HybridScheduler(resources))
+schedule.show()
+
+# %%
+operations = schedule.get_operations()
+bfs = operations.get_by_type_name(Butterfly.type_name())
+bfs.show(title="Butterfly executions")
+const_muls = operations.get_by_type_name(ConstantMultiplication.type_name())
+const_muls.show(title="ConstMul executions")
+inputs = operations.get_by_type_name(Input.type_name())
+inputs.show(title="Input executions")
+outputs = operations.get_by_type_name(Output.type_name())
+outputs.show(title="Output executions")
+
+bf_pe = ProcessingElement(bfs, entity_name="bf")
+mul_pe = ProcessingElement(const_muls, entity_name="mul")
+
+pe_in = ProcessingElement(inputs, entity_name='input')
+pe_out = ProcessingElement(outputs, entity_name='output')
+
+mem_vars = schedule.get_memory_variables()
+mem_vars.show(title="All memory variables")
+direct, mem_vars = mem_vars.split_on_length()
+mem_vars.show(title="Non-zero time memory variables")
+mem_vars_set = mem_vars.split_on_ports(read_ports=1, write_ports=1, total_ports=2)
+
+# %%
+memories = []
+for i, mem in enumerate(mem_vars_set):
+ memory = Memory(mem, memory_type="RAM", entity_name=f"memory{i}")
+ memories.append(memory)
+ mem.show(title=f"{memory.entity_name}")
+ memory.assign("left_edge")
+ memory.show_content(title=f"Assigned {memory.entity_name}")
+
+direct.show(title="Direct interconnects")
+
+# %%
+arch = Architecture(
+ {bf_pe, mul_pe, pe_in, pe_out},
+ memories,
+ direct_interconnects=direct,
+)
+arch
+
+# %%
+# Generate another HybridSchedule but this time constrain the amount of reads and writes to reduce the amount of memories
+resources = {Butterfly.type_name(): 1, ConstantMultiplication.type_name(): 1}
+schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ resources, max_concurrent_reads=2, max_concurrent_writes=2
+ ),
+)
+schedule.show()
+
+# %% Print the max number of read and write port accesses to non-direct memories
+direct, mem_vars = schedule.get_memory_variables().split_on_length()
+print("Max read ports:", mem_vars.read_ports_bound())
+print("Max write ports:", mem_vars.write_ports_bound())
+
+# %% Proceed to construct PEs and plot executions and non-direct memory variables
+operations = schedule.get_operations()
+bfs = operations.get_by_type_name(Butterfly.type_name())
+bfs.show(title="Butterfly executions")
+const_muls = operations.get_by_type_name(ConstantMultiplication.type_name())
+const_muls.show(title="ConstMul executions")
+inputs = operations.get_by_type_name(Input.type_name())
+inputs.show(title="Input executions")
+outputs = operations.get_by_type_name(Output.type_name())
+outputs.show(title="Output executions")
+
+bf_pe = ProcessingElement(bfs, entity_name="bf")
+mul_pe = ProcessingElement(const_muls, entity_name="mul")
+
+pe_in = ProcessingElement(inputs, entity_name='input')
+pe_out = ProcessingElement(outputs, entity_name='output')
+
+mem_vars.show(title="Non-zero time memory variables")
+mem_vars_set = mem_vars.split_on_ports(read_ports=1, write_ports=1, total_ports=2)
+
+# %% Allocate memories by graph-coloring
+memories = []
+for i, mem in enumerate(mem_vars_set):
+ memory = Memory(mem, memory_type="RAM", entity_name=f"memory{i}")
+ memories.append(memory)
+ mem.show(title=f"{memory.entity_name}")
+ memory.assign("left_edge")
+ memory.show_content(title=f"Assigned {memory.entity_name}")
+
+direct.show(title="Direct interconnects")
+
+# %% Synthesize the new architecture, now only using two memories but with data rate
+arch = Architecture(
+ {bf_pe, mul_pe, pe_in, pe_out},
+ memories,
+ direct_interconnects=direct,
+)
+arch
diff --git a/test/test_core_schedulers.py b/test/test_core_schedulers.py
index f934f157..8404db9c 100644
--- a/test/test_core_schedulers.py
+++ b/test/test_core_schedulers.py
@@ -1,13 +1,26 @@
import pytest
-from b_asic.core_operations import Addition, Butterfly, ConstantMultiplication
+from b_asic.core_operations import (
+ MADS,
+ Addition,
+ Butterfly,
+ ConstantMultiplication,
+ Reciprocal,
+)
from b_asic.core_schedulers import (
ALAPScheduler,
ASAPScheduler,
EarliestDeadlineScheduler,
+ HybridScheduler,
+ LeastSlackTimeScheduler,
+ MaxFanOutScheduler,
)
from b_asic.schedule import Schedule
-from b_asic.sfg_generators import direct_form_1_iir, radix_2_dif_fft
+from b_asic.sfg_generators import (
+ direct_form_1_iir,
+ ldlt_matrix_inverse,
+ radix_2_dif_fft,
+)
class TestASAPScheduler:
@@ -484,3 +497,628 @@ class TestEarliestDeadlineScheduler:
"out3": 7,
}
assert schedule.schedule_time == 7
+
+
+class TestLeastSlackTimeScheduler:
+ def test_empty_sfg(self, sfg_empty):
+ with pytest.raises(
+ ValueError, match="Empty signal flow graph cannot be scheduled."
+ ):
+ Schedule(sfg_empty, scheduler=LeastSlackTimeScheduler())
+
+ def test_direct_form_1_iir(self):
+ sfg = direct_form_1_iir([1, 2, 3], [1, 2, 3])
+
+ sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+ sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+ sfg.set_latency_of_type(Addition.type_name(), 3)
+ sfg.set_execution_time_of_type(Addition.type_name(), 1)
+
+ resources = {Addition.type_name(): 1, ConstantMultiplication.type_name(): 1}
+ schedule = Schedule(
+ sfg, scheduler=LeastSlackTimeScheduler(max_resources=resources)
+ )
+
+ assert schedule.start_times == {
+ "cmul4": 0,
+ "cmul3": 1,
+ "in0": 2,
+ "cmul0": 2,
+ "add1": 3,
+ "cmul1": 3,
+ "cmul2": 4,
+ "add3": 6,
+ "add0": 7,
+ "add2": 10,
+ "out0": 13,
+ }
+ assert schedule.schedule_time == 13
+
+ def test_direct_form_2_iir_inf_resources_no_exec_time(
+ self, sfg_direct_form_iir_lp_filter
+ ):
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(Addition.type_name(), 5)
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(
+ ConstantMultiplication.type_name(), 4
+ )
+
+ schedule = Schedule(
+ sfg_direct_form_iir_lp_filter, scheduler=LeastSlackTimeScheduler()
+ )
+
+ # should be the same as for ASAP due to infinite resources, except for input
+ assert schedule.start_times == {
+ "in0": 9,
+ "cmul1": 0,
+ "cmul4": 0,
+ "cmul2": 0,
+ "cmul3": 0,
+ "add3": 4,
+ "add1": 4,
+ "add0": 9,
+ "cmul0": 14,
+ "add2": 18,
+ "out0": 23,
+ }
+ assert schedule.schedule_time == 23
+
+ def test_direct_form_2_iir_1_add_1_mul_no_exec_time(
+ self, sfg_direct_form_iir_lp_filter
+ ):
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(Addition.type_name(), 5)
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(
+ ConstantMultiplication.type_name(), 4
+ )
+
+ max_resources = {ConstantMultiplication.type_name(): 1, Addition.type_name(): 1}
+
+ schedule = Schedule(
+ sfg_direct_form_iir_lp_filter,
+ scheduler=LeastSlackTimeScheduler(max_resources),
+ )
+ assert schedule.start_times == {
+ "cmul4": 0,
+ "cmul3": 4,
+ "cmul1": 8,
+ "add1": 8,
+ "cmul2": 12,
+ "in0": 13,
+ "add0": 13,
+ "add3": 18,
+ "cmul0": 18,
+ "add2": 23,
+ "out0": 28,
+ }
+
+ assert schedule.schedule_time == 28
+
+ def test_direct_form_2_iir_1_add_1_mul_exec_time_1(
+ self, sfg_direct_form_iir_lp_filter
+ ):
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(
+ ConstantMultiplication.type_name(), 3
+ )
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(Addition.type_name(), 2)
+ sfg_direct_form_iir_lp_filter.set_execution_time_of_type(
+ ConstantMultiplication.type_name(), 1
+ )
+ sfg_direct_form_iir_lp_filter.set_execution_time_of_type(
+ Addition.type_name(), 1
+ )
+
+ max_resources = {ConstantMultiplication.type_name(): 1, Addition.type_name(): 1}
+
+ schedule = Schedule(
+ sfg_direct_form_iir_lp_filter,
+ scheduler=LeastSlackTimeScheduler(max_resources),
+ )
+ assert schedule.start_times == {
+ "cmul4": 0,
+ "cmul3": 1,
+ "cmul1": 2,
+ "cmul2": 3,
+ "add1": 4,
+ "in0": 6,
+ "add0": 6,
+ "add3": 7,
+ "cmul0": 8,
+ "add2": 11,
+ "out0": 13,
+ }
+
+ assert schedule.schedule_time == 13
+
+ def test_direct_form_2_iir_2_add_3_mul_exec_time_1(
+ self, sfg_direct_form_iir_lp_filter
+ ):
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(
+ ConstantMultiplication.type_name(), 3
+ )
+ sfg_direct_form_iir_lp_filter.set_latency_of_type(Addition.type_name(), 2)
+ sfg_direct_form_iir_lp_filter.set_execution_time_of_type(
+ ConstantMultiplication.type_name(), 1
+ )
+ sfg_direct_form_iir_lp_filter.set_execution_time_of_type(
+ Addition.type_name(), 1
+ )
+
+ max_resources = {ConstantMultiplication.type_name(): 3, Addition.type_name(): 2}
+
+ schedule = Schedule(
+ sfg_direct_form_iir_lp_filter,
+ scheduler=LeastSlackTimeScheduler(max_resources),
+ )
+ assert schedule.start_times == {
+ "cmul1": 0,
+ "cmul4": 0,
+ "cmul3": 0,
+ "cmul2": 1,
+ "add1": 3,
+ "add3": 4,
+ "in0": 5,
+ "add0": 5,
+ "cmul0": 7,
+ "add2": 10,
+ "out0": 12,
+ }
+
+ assert schedule.schedule_time == 12
+
+ def test_radix_2_fft_8_points(self):
+ sfg = radix_2_dif_fft(points=8)
+
+ sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+ sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+ sfg.set_latency_of_type(Butterfly.type_name(), 1)
+ sfg.set_execution_time_of_type(Butterfly.type_name(), 1)
+
+ resources = {Butterfly.type_name(): 2, ConstantMultiplication.type_name(): 2}
+ schedule = Schedule(
+ sfg, scheduler=LeastSlackTimeScheduler(max_resources=resources)
+ )
+
+ assert schedule.start_times == {
+ "in1": 0,
+ "in3": 0,
+ "in5": 0,
+ "in7": 0,
+ "bfly6": 0,
+ "bfly8": 0,
+ "in2": 1,
+ "in6": 1,
+ "cmul2": 1,
+ "cmul3": 1,
+ "bfly11": 1,
+ "bfly7": 1,
+ "in0": 2,
+ "in4": 2,
+ "cmul0": 2,
+ "bfly0": 2,
+ "cmul4": 2,
+ "bfly5": 3,
+ "bfly1": 3,
+ "cmul1": 4,
+ "bfly2": 4,
+ "bfly9": 4,
+ "bfly10": 5,
+ "bfly3": 5,
+ "out0": 5,
+ "out4": 5,
+ "bfly4": 6,
+ "out1": 6,
+ "out2": 6,
+ "out5": 6,
+ "out6": 6,
+ "out7": 7,
+ "out3": 7,
+ }
+ assert schedule.schedule_time == 7
+
+
+class TestMaxFanOutScheduler:
+ def test_empty_sfg(self, sfg_empty):
+ with pytest.raises(
+ ValueError, match="Empty signal flow graph cannot be scheduled."
+ ):
+ Schedule(sfg_empty, scheduler=MaxFanOutScheduler())
+
+ def test_direct_form_1_iir(self):
+ sfg = direct_form_1_iir([1, 2, 3], [1, 2, 3])
+
+ sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+ sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+ sfg.set_latency_of_type(Addition.type_name(), 3)
+ sfg.set_execution_time_of_type(Addition.type_name(), 1)
+
+ resources = {Addition.type_name(): 1, ConstantMultiplication.type_name(): 1}
+ schedule = Schedule(sfg, scheduler=MaxFanOutScheduler(max_resources=resources))
+
+ assert schedule.start_times == {
+ "in0": 0,
+ "cmul0": 0,
+ "cmul1": 1,
+ "cmul2": 2,
+ "cmul4": 3,
+ "cmul3": 4,
+ "add3": 4,
+ "add1": 6,
+ "add0": 9,
+ "add2": 12,
+ "out0": 15,
+ }
+ assert schedule.schedule_time == 15
+
+
+class TestHybridScheduler:
+ def test_empty_sfg(self, sfg_empty):
+ with pytest.raises(
+ ValueError, match="Empty signal flow graph cannot be scheduled."
+ ):
+ Schedule(sfg_empty, scheduler=HybridScheduler())
+
+ def test_direct_form_1_iir(self):
+ sfg = direct_form_1_iir([1, 2, 3], [1, 2, 3])
+
+ sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+ sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+ sfg.set_latency_of_type(Addition.type_name(), 3)
+ sfg.set_execution_time_of_type(Addition.type_name(), 1)
+
+ resources = {Addition.type_name(): 1, ConstantMultiplication.type_name(): 1}
+ schedule = Schedule(sfg, scheduler=HybridScheduler(max_resources=resources))
+
+ assert schedule.start_times == {
+ "cmul4": 0,
+ "cmul3": 1,
+ "in0": 2,
+ "cmul0": 2,
+ "add1": 3,
+ "cmul1": 3,
+ "cmul2": 4,
+ "add3": 6,
+ "add0": 7,
+ "add2": 10,
+ "out0": 13,
+ }
+ assert schedule.schedule_time == 13
+
+ def test_radix_2_fft_8_points(self):
+ sfg = radix_2_dif_fft(points=8)
+
+ sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+ sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+ sfg.set_latency_of_type(Butterfly.type_name(), 1)
+ sfg.set_execution_time_of_type(Butterfly.type_name(), 1)
+
+ resources = {Butterfly.type_name(): 2, ConstantMultiplication.type_name(): 2}
+ schedule = Schedule(sfg, scheduler=HybridScheduler(max_resources=resources))
+
+ assert schedule.start_times == {
+ "in1": 0,
+ "in3": 0,
+ "in5": 0,
+ "in7": 0,
+ "bfly6": 0,
+ "bfly8": 0,
+ "in2": 1,
+ "in6": 1,
+ "cmul2": 1,
+ "cmul3": 1,
+ "bfly11": 1,
+ "bfly7": 1,
+ "in0": 2,
+ "in4": 2,
+ "cmul0": 2,
+ "bfly0": 2,
+ "cmul4": 2,
+ "bfly5": 3,
+ "bfly1": 3,
+ "cmul1": 4,
+ "bfly2": 4,
+ "bfly9": 4,
+ "bfly10": 5,
+ "bfly3": 5,
+ "out0": 5,
+ "out4": 5,
+ "bfly4": 6,
+ "out1": 6,
+ "out2": 6,
+ "out5": 6,
+ "out6": 6,
+ "out7": 7,
+ "out3": 7,
+ }
+ assert schedule.schedule_time == 7
+
+ def test_radix_2_fft_8_points_specified_IO_times_cyclic(self):
+ sfg = radix_2_dif_fft(points=8)
+
+ sfg.set_latency_of_type(Butterfly.type_name(), 3)
+ sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+ sfg.set_execution_time_of_type(Butterfly.type_name(), 1)
+ sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+
+ resources = {Butterfly.type_name(): 1, ConstantMultiplication.type_name(): 1}
+ input_times = {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ "in3": 3,
+ "in4": 4,
+ "in5": 5,
+ "in6": 6,
+ "in7": 7,
+ }
+ output_times = {
+ "out0": -2,
+ "out1": -1,
+ "out2": 0,
+ "out3": 1,
+ "out4": 2,
+ "out5": 3,
+ "out6": 4,
+ "out7": 5,
+ }
+ schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ resources, input_times=input_times, output_delta_times=output_times
+ ),
+ cyclic=True,
+ )
+
+ assert schedule.start_times == {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ "in3": 3,
+ "in4": 4,
+ "in5": 5,
+ "in6": 6,
+ "in7": 7,
+ "bfly0": 4,
+ "bfly8": 5,
+ "bfly11": 6,
+ "bfly6": 7,
+ "cmul2": 8,
+ "cmul0": 9,
+ "bfly1": 9,
+ "cmul3": 10,
+ "bfly7": 10,
+ "bfly2": 11,
+ "bfly5": 12,
+ "cmul4": 13,
+ "bfly9": 13,
+ "bfly10": 15,
+ "cmul1": 15,
+ "bfly3": 16,
+ "bfly4": 17,
+ "out0": 18,
+ "out1": 19,
+ "out2": 20,
+ "out3": 1,
+ "out4": 2,
+ "out5": 3,
+ "out6": 4,
+ "out7": 5,
+ }
+ assert schedule.schedule_time == 20
+
+ def test_radix_2_fft_8_points_specified_IO_times_non_cyclic(self):
+ sfg = radix_2_dif_fft(points=8)
+
+ sfg.set_latency_of_type(Butterfly.type_name(), 3)
+ sfg.set_latency_of_type(ConstantMultiplication.type_name(), 2)
+ sfg.set_execution_time_of_type(Butterfly.type_name(), 1)
+ sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+
+ resources = {Butterfly.type_name(): 1, ConstantMultiplication.type_name(): 1}
+ input_times = {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ "in3": 3,
+ "in4": 4,
+ "in5": 5,
+ "in6": 6,
+ "in7": 7,
+ }
+ output_times = {
+ "out0": -2,
+ "out1": -1,
+ "out2": 0,
+ "out3": 1,
+ "out4": 2,
+ "out5": 3,
+ "out6": 4,
+ "out7": 5,
+ }
+ schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ resources, input_times=input_times, output_delta_times=output_times
+ ),
+ cyclic=False,
+ )
+
+ assert schedule.start_times == {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ "in3": 3,
+ "in4": 4,
+ "in5": 5,
+ "in6": 6,
+ "in7": 7,
+ "bfly0": 4,
+ "bfly8": 5,
+ "bfly11": 6,
+ "bfly6": 7,
+ "cmul2": 8,
+ "cmul0": 9,
+ "bfly1": 9,
+ "cmul3": 10,
+ "bfly7": 10,
+ "bfly2": 11,
+ "bfly5": 12,
+ "cmul4": 13,
+ "bfly9": 13,
+ "bfly10": 15,
+ "cmul1": 15,
+ "bfly3": 16,
+ "bfly4": 17,
+ "out0": 18,
+ "out1": 19,
+ "out2": 20,
+ "out3": 21,
+ "out4": 22,
+ "out5": 23,
+ "out6": 24,
+ "out7": 25,
+ }
+ assert schedule.schedule_time == 25
+
+ def test_ldlt_inverse_2x2(self):
+ sfg = ldlt_matrix_inverse(N=2)
+
+ sfg.set_latency_of_type(MADS.type_name(), 3)
+ sfg.set_latency_of_type(Reciprocal.type_name(), 2)
+ sfg.set_execution_time_of_type(MADS.type_name(), 1)
+ sfg.set_execution_time_of_type(Reciprocal.type_name(), 1)
+
+ resources = {MADS.type_name(): 1, Reciprocal.type_name(): 1}
+ schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(resources),
+ )
+
+ assert schedule.start_times == {
+ "in0": 0,
+ "rec0": 0,
+ "in1": 2,
+ "dontcare1": 2,
+ "mads0": 2,
+ "in2": 5,
+ "mads3": 5,
+ "rec1": 8,
+ "dontcare0": 10,
+ "mads2": 10,
+ "mads1": 13,
+ "out2": 10,
+ "out1": 13,
+ "out0": 16,
+ }
+ assert schedule.schedule_time == 16
+
+ def test_ldlt_inverse_2x2_specified_IO_times_cyclic(self):
+ sfg = ldlt_matrix_inverse(N=2)
+
+ sfg.set_latency_of_type(MADS.type_name(), 3)
+ sfg.set_latency_of_type(Reciprocal.type_name(), 2)
+ sfg.set_execution_time_of_type(MADS.type_name(), 1)
+ sfg.set_execution_time_of_type(Reciprocal.type_name(), 1)
+
+ resources = {MADS.type_name(): 1, Reciprocal.type_name(): 1}
+ input_times = {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ }
+ output_times = {
+ "out0": 0,
+ "out1": 1,
+ "out2": 2,
+ }
+ schedule = Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ resources, input_times=input_times, output_delta_times=output_times
+ ),
+ cyclic=True,
+ )
+
+ assert schedule.start_times == {
+ "in0": 0,
+ "in1": 1,
+ "in2": 2,
+ "rec0": 0,
+ "dontcare1": 2,
+ "mads0": 2,
+ "mads3": 5,
+ "rec1": 8,
+ "dontcare0": 10,
+ "mads2": 10,
+ "mads1": 13,
+ "out0": 16,
+ "out1": 1,
+ "out2": 2,
+ }
+ assert schedule.schedule_time == 16
+
+ def test_max_invalid_resources(self):
+ sfg = ldlt_matrix_inverse(N=2)
+
+ sfg.set_latency_of_type(MADS.type_name(), 3)
+ sfg.set_latency_of_type(Reciprocal.type_name(), 2)
+ sfg.set_execution_time_of_type(MADS.type_name(), 1)
+ sfg.set_execution_time_of_type(Reciprocal.type_name(), 1)
+
+ resources = 2
+ with pytest.raises(ValueError, match="max_resources must be a dictionary."):
+ Schedule(sfg, scheduler=HybridScheduler(resources))
+
+ resources = "test"
+ with pytest.raises(ValueError, match="max_resources must be a dictionary."):
+ Schedule(sfg, scheduler=HybridScheduler(resources))
+
+ resources = []
+ with pytest.raises(ValueError, match="max_resources must be a dictionary."):
+ Schedule(sfg, scheduler=HybridScheduler(resources))
+
+ resources = {1: 1}
+ with pytest.raises(
+ ValueError, match="max_resources key must be a valid type_name."
+ ):
+ Schedule(sfg, scheduler=HybridScheduler(resources))
+
+ resources = {MADS.type_name(): "test"}
+ with pytest.raises(ValueError, match="max_resources value must be an integer."):
+ Schedule(sfg, scheduler=HybridScheduler(resources))
+
+ # def test_ldlt_inverse_2x2_read_constrained(self):
+ # sfg = ldlt_matrix_inverse(N=2)
+
+ # sfg.set_latency_of_type(MADS.type_name(), 3)
+ # sfg.set_latency_of_type(Reciprocal.type_name(), 2)
+ # sfg.set_execution_time_of_type(MADS.type_name(), 1)
+ # sfg.set_execution_time_of_type(Reciprocal.type_name(), 1)
+
+ # resources = {MADS.type_name(): 1, Reciprocal.type_name(): 1}
+ # schedule = Schedule(
+ # sfg,
+ # scheduler=HybridScheduler(
+ # max_resources = resources,
+ # max_concurrent_reads = 3,
+ # ),
+ # )
+
+ def test_ldlt_inverse_2x2_read_constrained_too_low(self):
+ sfg = ldlt_matrix_inverse(N=2)
+
+ sfg.set_latency_of_type(MADS.type_name(), 3)
+ sfg.set_latency_of_type(Reciprocal.type_name(), 2)
+ sfg.set_execution_time_of_type(MADS.type_name(), 1)
+ sfg.set_execution_time_of_type(Reciprocal.type_name(), 1)
+
+ resources = {MADS.type_name(): 1, Reciprocal.type_name(): 1}
+ with pytest.raises(
+ TimeoutError,
+ match="Algorithm did not schedule any operation for 10 time steps, try relaxing constraints.",
+ ):
+ Schedule(
+ sfg,
+ scheduler=HybridScheduler(
+ max_resources=resources,
+ max_concurrent_reads=2,
+ ),
+ )
--
GitLab