def doit(self, dag: Dag):
new_inputs = dag.inputs
self.allroots = list(dag.roots())
self.run(dag)
new_inputs.append(self.P)
new_inputs.append(self.R)
new_inputs.append(self.O)
# return dag that has taken precision as an input
return Dag(outputs=dag.outputs, inputs=new_inputs)
def matrix_multiply():
a00 = Input(name="a00")
a01 = Input(name="a01")
a10 = Input(name="a10")
a11 = Input(name="a11")
b00 = Input(name="b00")
b01 = Input(name="b01")
b10 = Input(name="b10")
b11 = Input(name="b11")
p0 = Mul(Add(a00, a11), Add(b00, b11), name="p0")
p1 = Mul(Add(a10, a11), b00, name="p1")
p2 = Mul(a00, Sub(b01, b11), name="p2")
p3 = Mul(a11, Sub(b10, b00), name="p3")
p4 = Mul(b11, Add(a00, a01), name="p4")
p5 = Mul(Add(b00, b01), Sub(a10, a00), name="p5")
p6 = Mul(Add(b10, b11), Sub(a01, a11), name="p6")
y00 = Add(Sub(Add(p0, p3), p4), p6, name="y00")
y01 = Add(p2, p4, name="y01")
y10 = Add(p1, p3, name="y10")
y11 = Add(Sub(Add(p0, p2), p1), p5, name="y11")
matrix_dag = Dag(outputs=[y00, y01, y10, y11], inputs=[
a00, a01, a10, a11, b00, b01, b10, b11])
return matrix_dag
def Matrix_Multiplication():
a = Input(name="a")
b = Input(name="b")
a00 = a[0][0]
a01 = a[0][1]
a10 = a[1][0]
a11 = a[1][1]
b00 = b[0][0]
b01 = b[0][1]
b10 = b[1][0]
b11 = b[1][1]
p0 = (a00 + a11) * (b00 + b11)
p1 = (a10 + a11) * b00
p2 = a00 * (b01 - b11)
p3 = a11 * (b10 - b00)
p4 = (a00 + a01) * b11
p5 = (a10 - a00) * (b00 + b01)
p6 = (a01 - a11) * (b10 + b11)
y00 = p0 + p3 - p4 + p6
y01 = p2 + p4
y10 = p1 + p3
y11 = p0 + p2 - p1 + p5
casestudy_dag = Dag(outputs=[y00, y01, y10, y11], inputs=[a, b])
return casestudy_dag
def generate_basic(a):
x = Input(name="x")
a = Constant(a, name="a")
output = Mul(a, x, name="res")
approx_dag = Dag(outputs=[output], inputs=[x])
return approx_dag
def gen_fig3():
#(a*b) + 4 - b
a = Input(name="a")
b = Input(name="b")
c = Constant(4, name="c")
d = Mul(a, b, name="d")
e = Add(d, c, name="e")
z = Sub(e, b, name="z")
fig3_dag = Dag(outputs=[z], inputs=[a, b])
return fig3_dag
def gen_dag1():
#(a*b) + 4 - b
x = Input(name="x")
y = Input(name="y")
z = Input(name="z")
i = Mul(x, y, name="i")
j = Sub(y, z, name="j")
k = Mul(i, j, name="k")
dag = Dag(outputs=[k], inputs=[x, y, z])
return dag
def gen_ex1():
#(a * b) + (b * c)
a = Input(name="a")
b = Input(name="b")
c = Input(name="c")
d = Mul(a, b, name="d")
e = Mul(b, c, name="e")
z = Add(e, d, name="z")
dag = Dag(outputs=[z], inputs=[a, b, c])
return dag
def generate_basic_lut():
x = Input(name="x")
amplitude = Input(name="a")
shift = Input(name="b")
output = Add(Mul(amplitude, LookupTable(
np.sin, x)), shift, name="res")
# output = Add(Mul(amplitude, x), shift, name="res")
sin_dag = Dag(outputs=[output], inputs=[x, amplitude, shift])
return sin_dag
def poly_approx():
a = Input(name="a")
c = Input(name="c")
first = a[0]
for i in range(1, 5): # for degree 4 polynomial
first = first * c + a[i]
casestudy_dag = Dag(outputs=[first], inputs=[a, c])
return casestudy_dag
def generate_square_wave():
x = Input(name="x")
a = LookupTable(np.sin, x)
b = Add(Mul(Constant(1/3.), LookupTable(np.sin, Mul(Constant(3.), x))), a)
c = Add(Mul(Constant(1/5.), LookupTable(np.sin, Mul(Constant(5.), x))), b)
d = Add(Mul(Constant(1/7.), LookupTable(np.sin, Mul(Constant(7.), x))), c)
e = Add(Mul(Constant(1/9.), LookupTable(np.sin,
Mul(Constant(9.), x))), d, name="res")
dag = Dag(outputs=[e], inputs=[x])
return dag
def gen_ex2():
# sqrt((a * a) + (b * b) + (c * c))
a = Input(name="a")
b = Input(name="b")
c = Input(name="c")
out = LookupTable(np.sqrt, Add(
Add(Mul(a, a), Mul(b, b)), Mul(c, c)), name="res", precision=8.)
dag = Dag(outputs=[out], inputs=[a, b, c])
return dag
def generate_poly_approx(a, b, c, d):
x = Input(name="x")
a = Constant(a, name="a")
b = Constant(b, name="b")
c = Constant(c, name="c")
d = Constant(d, name="d")
output = Mul(
Add(Mul(Add(Mul(Add(Mul(a, x), b), x), c), x), d), x, name="res")
approx_dag = Dag(outputs=[output], inputs=[x])
return approx_dag
def RGB_to_YCbCr():
a = Input(name="a")
col_1 = Constant(.299) * a[0] + Constant(.587) * \
a[1] + Constant(.114) * a[2]
col_2 = Constant(-.16875) * \
a[0] + Constant(-.33126) * a[1] + Constant(.5) * a[2]
col_3 = Constant(.5) * a[0] + Constant(-.41869) * \
a[1] + Constant(-.08131) * a[2]
casestudy_dag = Dag(outputs=[col_1, col_2, col_3], inputs=[a])
return casestudy_dag
def RGB_to_YCbCr():
r = Input(name="r")
g = Input(name="g")
b = Input(name="b")
col_1 = Add(Add(Mul(Constant(.299, name="C1"), r), Mul(Constant(.587, name="C2"), g)),
Mul(Constant(.114, name="C3"), b), name="col_1")
col_2 = Add(Add(Mul(Constant(-.16875, name="C4"), r), Mul(Constant(-.33126, name="C5"), g)),
Mul(Constant(.5, name="C6"), b), name="col_2")
col_3 = Add(Add(Mul(Constant(.5, name="C7"), r), Mul(Constant(-.41869, name="C8"), g)),
Mul(Constant(-.08131, name="C9"), b), name="col_3")
casestudy_dag = Dag(outputs=[col_1, col_2, col_3], inputs=[r, g, b])
return casestudy_dag
def gen_ex3():
a = Input(name="a")
b = Input(name="b")
c = Input(name="c")
d = Input(name="d")
out = Add(Mul(a, Mul(b, Mul(c, d))), Mul(a, b), name="res")
dag = Dag(outputs=[out], inputs=[a, b, c, d])
return dag
# def test_fig3():
# dag = gen_fig3()
# bf = BitFlow(dag, {"z": 8.}, {'a': (-3., 2.),
# 'b': (4., 8.)}, lr=1e-2, range_lr=1e-2, train_range=True, training_size=50000, testing_size=10000, distribution=2, incorporate_ulp_loss=False, batch_size=16, test_optimizer=True)
# bf.train(epochs=10, decay=0.8)
# rounded_dag = bf.rounded_dag
# original_dag = bf.original_dag
# verilog_gen = BitFlow2Verilog(
# "fig3", bf.P, bf.R, bf.filtered_vars, original_dag, {"z": 8.})
# verilog_gen.evaluate()
# # check saving object works
# BitFlow.save("./models/fig3", bf)
# new_bf = BitFlow.load("./models/fig3")
# new_bf.train(epochs=5)
# assert new_bf.range_lr == bf.range_lr
return
def doit(self, dag: Dag):
self.run(dag)
return Dag(outputs=dag.outputs, inputs=dag.inputs)
