From 9207170caaffbf597106d799f88e0c3c39bac8c0 Mon Sep 17 00:00:00 2001
From: Simon Bjurek <simbj106@student.liu.se>
Date: Wed, 12 Mar 2025 08:48:23 +0000
Subject: [PATCH] Recursive loops now seperate sfg property, fixed 0 input in
Division and Reciprocal etc
---
b_asic/core_operations.py | 4 +
b_asic/schedule.py | 46 ++++++++-
b_asic/scheduler_gui/scheduler_item.py | 13 ++-
b_asic/signal_flow_graph.py | 82 +++++++++------
examples/scheduling_pipelining_retiming.py | 111 +++++++++++++++++++++
test/unit/test_core_operations.py | 9 ++
test/unit/test_list_schedulers.py | 42 ++++++++
test/unit/test_sfg.py | 21 ++++
8 files changed, 297 insertions(+), 31 deletions(-)
create mode 100644 examples/scheduling_pipelining_retiming.py
diff --git a/b_asic/core_operations.py b/b_asic/core_operations.py
index 1928e81a..faf5d31d 100644
--- a/b_asic/core_operations.py
+++ b/b_asic/core_operations.py
@@ -523,6 +523,8 @@ class Division(AbstractOperation):
return TypeName("div")
def evaluate(self, a, b):
+ if b == 0:
+ return float("inf")
return a / b
@property
@@ -1372,6 +1374,8 @@ class Reciprocal(AbstractOperation):
return TypeName("rec")
def evaluate(self, a):
+ if a == 0:
+ return float("inf")
return 1 / a
diff --git a/b_asic/schedule.py b/b_asic/schedule.py
index 6dfe5809..7658fdb3 100644
--- a/b_asic/schedule.py
+++ b/b_asic/schedule.py
@@ -365,10 +365,10 @@ class Schedule:
available_time = (
cast(int, source.latency_offset)
+ self._start_times[source.operation.graph_id]
- - self._schedule_time * self._laps[signal.graph_id]
)
if available_time > self._schedule_time:
available_time -= self._schedule_time
+ available_time -= self._schedule_time * self._laps[signal.graph_id]
else:
available_time = (
cast(int, source.latency_offset)
@@ -458,6 +458,26 @@ class Schedule:
raise ValueError(
f"New schedule time ({time}) too short, minimum: {max_end_time}."
)
+
+ # if updating the scheduling time -> update laps due to operations
+ # reading and writing in different iterations (across the edge)
+ if self._schedule_time is not None:
+ for signal_id in self._laps.keys():
+ port = self._sfg.find_by_id(signal_id).destination
+
+ source_port = port.signals[0].source
+ source_op = source_port.operation
+ source_port_start_time = self._start_times[source_op.graph_id]
+ source_port_latency_offset = source_op.latency_offsets[
+ f"out{source_port.index}"
+ ]
+ if (
+ source_port_start_time + source_port_latency_offset
+ > self._schedule_time
+ and source_port_start_time + source_port_latency_offset <= time
+ ):
+ self._laps[signal_id] += 1
+
self._schedule_time = time
return self
@@ -855,6 +875,13 @@ class Schedule:
):
new_start = self._schedule_time
self._laps[op.input(0).signals[0].graph_id] -= 1
+ if (
+ new_start == 0
+ and isinstance(op, Input)
+ and self._laps[op.output(0).signals[0].graph_id] != 0
+ ):
+ new_start = 0
+ self._laps[op.output(0).signals[0].graph_id] -= 1
# Set new start time
self._start_times[graph_id] = new_start
return self
@@ -936,7 +963,24 @@ class Schedule:
destination_laps = []
for signal_id, lap in self._laps.items():
port = new_sfg.find_by_id(signal_id).destination
+
+ # if an operation lies across the scheduling time, place a delay after it
+ source_port = port.signals[0].source
+ source_op = source_port.operation
+ source_port_start_time = self._start_times[source_op.graph_id]
+ source_port_latency_offset = source_op.latency_offsets[
+ f"out{source_port.index}"
+ ]
+ if (
+ source_port_start_time + source_port_latency_offset
+ > self._schedule_time
+ ):
+ lap += (
+ source_port_start_time + source_port_latency_offset
+ ) // self._schedule_time
+
destination_laps.append((port.operation.graph_id, port.index, lap))
+
for op, port, lap in destination_laps:
for delays in range(lap):
new_sfg = new_sfg.insert_operation_before(op, Delay(), port)
diff --git a/b_asic/scheduler_gui/scheduler_item.py b/b_asic/scheduler_gui/scheduler_item.py
index fbb84fcd..56a6e7fb 100644
--- a/b_asic/scheduler_gui/scheduler_item.py
+++ b/b_asic/scheduler_gui/scheduler_item.py
@@ -28,6 +28,7 @@ from b_asic.scheduler_gui.axes_item import AxesItem
from b_asic.scheduler_gui.operation_item import OperationItem
from b_asic.scheduler_gui.scheduler_event import SchedulerEvent
from b_asic.scheduler_gui.signal_item import SignalItem
+from b_asic.special_operations import Output
from b_asic.types import GraphID
@@ -121,7 +122,10 @@ class SchedulerItem(SchedulerEvent, QGraphicsItemGroup):
raise ValueError("No schedule installed.")
new_start_time = floor(pos) - floor(self._x_axis_indent)
slacks = self.schedule.slacks(item.graph_id)
- op_start_time = self.schedule.start_time_of_operation(item.graph_id)
+ op_start_time = (
+ self.schedule.start_time_of_operation(item.graph_id)
+ % self.schedule.schedule_time
+ )
if not -slacks[0] <= new_start_time - op_start_time <= slacks[1]:
# Cannot move due to dependencies
return False
@@ -245,6 +249,13 @@ class SchedulerItem(SchedulerEvent, QGraphicsItemGroup):
op_start_time = self.schedule.start_time_of_operation(item.graph_id)
new_start_time = floor(pos) - floor(self._x_axis_indent)
move_time = new_start_time - op_start_time
+ op = self._schedule.sfg.find_by_id(item.graph_id)
+ if (
+ isinstance(op, Output)
+ and op_start_time == self.schedule.schedule_time
+ and new_start_time < self.schedule.schedule_time
+ ):
+ move_time = new_start_time
if move_time:
self.schedule.move_operation(item.graph_id, move_time)
print(f"schedule.move_operation({item.graph_id!r}, {move_time})")
diff --git a/b_asic/signal_flow_graph.py b/b_asic/signal_flow_graph.py
index c0c71483..de1133ba 100644
--- a/b_asic/signal_flow_graph.py
+++ b/b_asic/signal_flow_graph.py
@@ -1767,13 +1767,52 @@ class SFG(AbstractOperation):
-------
The iteration period bound.
"""
+ loops = self.loops
+ if not loops:
+ return -1
+
+ op_and_latency = {}
+ for op in self.operations:
+ for loop in loops:
+ for element in loop:
+ if op.type_name() not in op_and_latency:
+ op_and_latency[op.type_name()] = op.latency
+ t_l_values = []
+
+ for loop in loops:
+ loop.pop()
+ time_of_loop = 0
+ number_of_t_in_loop = 0
+ for element in loop:
+ if ''.join([i for i in element if not i.isdigit()]) == 't':
+ number_of_t_in_loop += 1
+ for key, item in op_and_latency.items():
+ if key in element:
+ time_of_loop += item
+ if number_of_t_in_loop in (0, 1):
+ t_l_values.append(Fraction(time_of_loop, 1))
+ else:
+ t_l_values.append(Fraction(time_of_loop, number_of_t_in_loop))
+ return max(t_l_values)
+
+ @property
+ def loops(self) -> list[list[GraphID]]:
+ """
+ Return the recursive loops found in the SFG.
+
+ If -1, the SFG does not have any loops.
+
+ Returns
+ -------
+ A list of the recursive loops.
+ """
inputs_used = []
for used_input in self._used_ids:
if 'in' in str(used_input):
used_input = used_input.replace("in", "")
inputs_used.append(int(used_input))
if inputs_used == []:
- raise ValueError("No inputs to sfg")
+ return []
for input in inputs_used:
input_op = self._input_operations[input]
queue: Deque[Operation] = deque([input_op])
@@ -1795,39 +1834,24 @@ class SFG(AbstractOperation):
visited.add(new_op)
else:
raise ValueError("Destination does not exist")
- if not dict_of_sfg:
- raise ValueError(
- "the SFG does not have any loops and therefore no iteration period bound."
- )
cycles = [
[node] + path
for node in dict_of_sfg
for path in self._dfs(dict_of_sfg, node, node)
]
- if not cycles:
- return -1
- op_and_latency = {}
- for op in self.operations:
- for lista in cycles:
- for element in lista:
- if op.type_name() not in op_and_latency:
- op_and_latency[op.type_name()] = op.latency
- t_l_values = []
- for loop in cycles:
- loop.pop()
- time_of_loop = 0
- number_of_t_in_loop = 0
- for element in loop:
- if ''.join([i for i in element if not i.isdigit()]) == 't':
- number_of_t_in_loop += 1
- for key, item in op_and_latency.items():
- if key in element:
- time_of_loop += item
- if number_of_t_in_loop in (0, 1):
- t_l_values.append(Fraction(time_of_loop, 1))
- else:
- t_l_values.append(Fraction(time_of_loop, number_of_t_in_loop))
- return max(t_l_values)
+
+ loops = self._get_non_redundant_cycles(cycles)
+ return loops
+
+ def _get_non_redundant_cycles(self, loops):
+ unique_lists = []
+ seen_cycles = set()
+ for loop in loops:
+ operation_set = frozenset(loop)
+ if operation_set not in seen_cycles:
+ unique_lists.append(loop)
+ seen_cycles.add(operation_set)
+ return unique_lists
def state_space_representation(self):
"""
diff --git a/examples/scheduling_pipelining_retiming.py b/examples/scheduling_pipelining_retiming.py
new file mode 100644
index 00000000..38a9601f
--- /dev/null
+++ b/examples/scheduling_pipelining_retiming.py
@@ -0,0 +1,111 @@
+"""
+=========================================
+Scheduling and Pipelining/Retiming
+=========================================
+
+When scheduling cyclically (modulo) there is implicit pipelining/retiming taking place.
+B-ASIC can easily be used to showcase this.
+"""
+
+import matplotlib.pyplot as plt
+import numpy as np
+from scipy import signal
+
+from b_asic.core_operations import Addition, ConstantMultiplication
+from b_asic.schedule import Schedule
+from b_asic.scheduler import ALAPScheduler
+from b_asic.sfg_generators import direct_form_1_iir
+from b_asic.signal_generator import Impulse
+from b_asic.simulation import Simulation
+
+# %%
+# Design a simple direct form IIR low-pass filter.
+N = 3
+Wc = 0.2
+b, a = signal.butter(N, Wc, btype="lowpass", output="ba")
+
+# %%
+# Generate the corresponding signal-flow-graph (SFG).
+sfg = direct_form_1_iir(b, a)
+sfg
+
+# %%
+# Set latencies and execution times of the operations.
+sfg.set_latency_of_type(Addition.type_name(), 1)
+sfg.set_latency_of_type(ConstantMultiplication.type_name(), 3)
+sfg.set_execution_time_of_type(Addition.type_name(), 1)
+sfg.set_execution_time_of_type(ConstantMultiplication.type_name(), 1)
+
+# %%
+# Print the critical path Tcp and the iteration period bound Tmin.
+T_cp = sfg.critical_path_time()
+print("Tcp:", T_cp)
+T_min = sfg.iteration_period_bound()
+print("Tmin:", T_min)
+
+# %%
+# Create an ALAP schedule
+schedule = Schedule(sfg, scheduler=ALAPScheduler(), cyclic=True)
+schedule.show()
+
+# %%
+# Move some operations "over the edge" in order to reach Tcp = Tmin.
+schedule.move_operation('out0', 2)
+schedule.move_operation('add2', 2)
+schedule.move_operation('add0', 2)
+schedule.move_operation('add3', 2)
+schedule.set_schedule_time(5)
+schedule.show()
+
+# %%
+# Print the new critical path Tcp that is now equal to Tmin.
+T_cp = schedule.sfg.critical_path_time()
+print("Tcp:", T_cp)
+T_min = schedule.sfg.iteration_period_bound()
+print("Tmin:", T_min)
+
+# %%
+# Show the reconstructed SFG that is now pipelined/retimed compared to the original.
+schedule.sfg
+
+# %%
+# Simulate the impulse response of the original and reconstructed SFGs.
+# Plot the frequency responses of the original filter, the original SFG and the reconstructed SFG to verify
+# that the schedule is valid.
+sim1 = Simulation(sfg, [Impulse()])
+sim1.run_for(1000)
+
+sim2 = Simulation(schedule.sfg, [Impulse()])
+sim2.run_for(1000)
+
+w, h = signal.freqz(b, a)
+
+# Plot 1: Original filter
+spectrum_0 = 20 * np.log10(np.abs(h))
+plt.figure()
+plt.plot(w / np.pi, spectrum_0)
+plt.title("Original filter")
+plt.xlabel("Normalized frequency (x pi rad/sample)")
+plt.ylabel("Magnitude (dB)")
+plt.grid(True)
+plt.show()
+
+# Plot 2: Simulated SFG
+spectrum_1 = 20 * np.log10(np.abs(signal.freqz(sim1.results['0'])[1]))
+plt.figure()
+plt.plot(w / np.pi, spectrum_1)
+plt.title("Simulated SFG")
+plt.xlabel("Normalized frequency (x pi rad/sample)")
+plt.ylabel("Magnitude (dB)")
+plt.grid(True)
+plt.show()
+
+# Plot 3: Recreated SFG
+spectrum_2 = 20 * np.log10(np.abs(signal.freqz(sim2.results['0'])[1]))
+plt.figure()
+plt.plot(w / np.pi, spectrum_2)
+plt.title("Pipelined/retimed SFG")
+plt.xlabel("Normalized frequency (x pi rad/sample)")
+plt.ylabel("Magnitude (dB)")
+plt.grid(True)
+plt.show()
diff --git a/test/unit/test_core_operations.py b/test/unit/test_core_operations.py
index e8241e63..3e2230cc 100644
--- a/test/unit/test_core_operations.py
+++ b/test/unit/test_core_operations.py
@@ -185,6 +185,11 @@ class TestDivision:
test_operation = Division(Addition(Input(), Constant(3)), Constant(3))
assert test_operation.is_linear
+ def test_zero_input(self):
+ test_operation = Division()
+ assert test_operation.evaluate_output(0, [0, 1]) == 0
+ assert test_operation.evaluate_output(0, [1, 0]) == float("inf")
+
class TestSquareRoot:
"""Tests for SquareRoot class."""
@@ -577,6 +582,10 @@ class TestReciprocal:
test_operation = Reciprocal()
assert test_operation.evaluate_output(0, [1 + 1j]) == 0.5 - 0.5j
+ def test_zero_input(self):
+ test_operation = Reciprocal()
+ assert test_operation.evaluate_output(0, [0]) == float("inf")
+
class TestDepends:
def test_depends_addition(self):
diff --git a/test/unit/test_list_schedulers.py b/test/unit/test_list_schedulers.py
index 34daa4d8..d788abd2 100644
--- a/test/unit/test_list_schedulers.py
+++ b/test/unit/test_list_schedulers.py
@@ -468,6 +468,7 @@ class TestMaxFanOutScheduler:
"out0": 36,
}
assert schedule.schedule_time == 36
+ _validate_recreated_sfg_ldlt_matrix_inverse(schedule, 3)
class TestHybridScheduler:
@@ -829,6 +830,7 @@ class TestHybridScheduler:
"out0": 16,
}
assert schedule.schedule_time == 16
+ _validate_recreated_sfg_ldlt_matrix_inverse(schedule, 2)
def test_ldlt_inverse_2x2_specified_IO_times_cyclic(self):
sfg = ldlt_matrix_inverse(N=2)
@@ -876,6 +878,21 @@ class TestHybridScheduler:
}
assert schedule.schedule_time == 16
+ # validate regenerated sfg with random 2x2 real s.p.d. matrix
+ A = np.random.rand(2, 2)
+ A = np.dot(A, A.T)
+ A_inv = np.linalg.inv(A)
+ input_signals = []
+ for i in range(2):
+ for j in range(i, 2):
+ input_signals.append(Constant(A[i, j]))
+
+ sim = Simulation(schedule.sfg, input_signals)
+ sim.run_for(2)
+ assert np.allclose(sim.results["0"], [A_inv[0, 0], A_inv[0, 0]])
+ assert np.allclose(sim.results["1"], [0, A_inv[0, 1]])
+ assert np.allclose(sim.results["2"], [0, A_inv[1, 1]])
+
def test_invalid_max_resources(self):
sfg = ldlt_matrix_inverse(N=2)
@@ -1157,6 +1174,7 @@ class TestHybridScheduler:
direct, mem_vars = schedule.get_memory_variables().split_on_length()
assert mem_vars.read_ports_bound() == 3
assert mem_vars.write_ports_bound() == 1
+ _validate_recreated_sfg_ldlt_matrix_inverse(schedule, 3)
def test_32_point_fft_custom_io_times(self):
POINTS = 32
@@ -1672,6 +1690,7 @@ class TestHybridScheduler:
}
assert all([val == 0 for val in schedule.laps.values()])
+ _validate_recreated_sfg_ldlt_matrix_inverse(schedule, 3)
def test_latency_offsets_cyclic(self):
sfg = ldlt_matrix_inverse(
@@ -1953,3 +1972,26 @@ def _validate_recreated_sfg_fft(schedule: Schedule, points: int) -> None:
a = res[str(i)]
b = exp_res[i]
assert np.isclose(a, b)
+
+
+def _validate_recreated_sfg_ldlt_matrix_inverse(schedule: Schedule, N: int) -> None:
+ # random real s.p.d matrix
+ A = np.random.rand(N, N)
+ A = np.dot(A, A.T)
+
+ # iterate through the upper diagonal and construct the input to the SFG
+ input_signals = []
+ for i in range(N):
+ for j in range(i, N):
+ input_signals.append(Constant(A[i, j]))
+
+ A_inv = np.linalg.inv(A)
+ sim = Simulation(schedule.sfg, input_signals)
+ sim.run_for(1)
+
+ # iterate through the upper diagonal and check
+ count = 0
+ for i in range(N):
+ for j in range(i, N):
+ assert np.isclose(sim.results[str(count)], A_inv[i, j])
+ count += 1
diff --git a/test/unit/test_sfg.py b/test/unit/test_sfg.py
index 471a8bde..090a11dd 100644
--- a/test/unit/test_sfg.py
+++ b/test/unit/test_sfg.py
@@ -1913,6 +1913,27 @@ class TestIterationPeriodBound:
assert sfg_two_inputs_two_outputs.iteration_period_bound() == -1
+class TestLoops:
+ def test_accumulator(self, sfg_simple_accumulator):
+ loops = sfg_simple_accumulator.loops
+ assert loops == [["add0", "t0", "add0"]]
+
+ def test_simple_filter(self, sfg_simple_filter):
+ loops = sfg_simple_filter.loops
+ assert loops == [["add0", "t0", "cmul0", "add0"]]
+
+ def test_direct_form_iir_filter(self, sfg_direct_form_iir_lp_filter):
+ loops = sfg_direct_form_iir_lp_filter.loops
+ assert loops == [
+ ["add0", "t0", "cmul4", "add1", "add0"],
+ ["add0", "t0", "t1", "cmul3", "add1", "add0"],
+ ]
+
+ def test_empty_sfg(self):
+ loops = SFG([], []).loops
+ assert loops == []
+
+
class TestStateSpace:
def test_accumulator(self, sfg_simple_accumulator):
ss = sfg_simple_accumulator.state_space_representation()
--
GitLab