def top(input, filter, bias, ):
input_extent_3_required_s = (((((final_extent_2 + 31)//32) * final_extent_3) + -1)//hcl.select(((final_extent_2 + 31)//32) > 1, ((final_extent_2 + 31)//32), 1))
final_total_extent_1 = (hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_1) * hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_0))
final_total_extent_2 = (final_total_extent_1 * hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_2))
final_total_extent_3 = (final_total_extent_2 * hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_3))
f_conv_n_extent_realized_s = hcl.select(hcl.select((((final_extent_2 * final_extent_3) + -1)//hcl.select(final_extent_2 > 1, final_extent_2, 1)) > (final_extent_3 + -1), (((final_extent_2 * final_extent_3) + -1)//hcl.select(final_extent_2 > 1, final_extent_2, 1)), (final_extent_3 + -1)) > (((((final_extent_2 + 31)//32) * final_extent_3) + -1)//(hcl.select(((final_extent_2 + -1)//32) > 0, ((final_extent_2 + -1)//32), 0) + 1)), hcl.select((((final_extent_2 * final_extent_3) + -1)//hcl.select(final_extent_2 > 1, final_extent_2, 1)) > (final_extent_3 + -1), (((final_extent_2 * final_extent_3) + -1)//hcl.select(final_extent_2 > 1, final_extent_2, 1)), (final_extent_3 + -1)), (((((final_extent_2 + 31)//32) * final_extent_3) + -1)//(hcl.select(((final_extent_2 + -1)//32) > 0, ((final_extent_2 + -1)//32), 0) + 1)))
f_conv_z_extent_realized = hcl.select(((hcl.select(((final_extent_2 + -1)//32) > 0, ((final_extent_2 + -1)//32), 0) * 32) + 32) > final_extent_2, ((hcl.select(((final_extent_2 + -1)//32) > 0, ((final_extent_2 + -1)//32), 0) * 32) + 32), final_extent_2)
f_conv = hcl.compute((final_extent_0, ((((final_extent_1 + -1)//32) * 32) + 32), f_conv_z_extent_realized, (f_conv_n_extent_realized_s + 1)), lambda x, y, z, w: 0, name = "f_conv", dtype = hcl.Float(bits = 32))
with hcl.Stage("f_conv"):
with hcl.for_(0, (final_extent_2 * final_extent_3), name = "f_conv_s0_z_par") as f_conv_s0_z_par:
with hcl.for_(final_min_1, final_extent_1, name = "f_conv_s0_y") as f_conv_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "f_conv_s0_x") as f_conv_s0_x:
f_conv[f_conv_s0_x, f_conv_s0_y, ((f_conv_s0_z_par % hcl.select(final_extent_2 > 1, final_extent_2, 1)) + final_min_2), ((f_conv_s0_z_par//hcl.select(final_extent_2 > 1, final_extent_2, 1)) + final_min_3)] = bias[((f_conv_s0_z_par % hcl.select(final_extent_2 > 1, final_extent_2, 1)) + final_min_2)]
with hcl.for_(0, (((final_extent_2 + 31)//32) * final_extent_3), name = "f_conv_s1_z_z_par") as f_conv_s1_z_z_par:
f_conv_s1_z_z_t_base_s = (f_conv_s1_z_z_par % hcl.select(((final_extent_2 + 31)//32) > 1, ((final_extent_2 + 31)//32), 1))
with hcl.for_(0, 32, name = "f_conv_s1_r__z") as f_conv_s1_r__z:
with hcl.for_(0, ((final_extent_1 + 31)//32), name = "f_conv_s1_y_y") as f_conv_s1_y_y:
with hcl.for_(0, 32, name = "f_conv_s1_z_z_t") as f_conv_s1_z_z_t:
with hcl.for_(0, 32, name = "f_conv_s1_y_y_t") as f_conv_s1_y_y_t:
with hcl.for_(final_min_0, final_extent_0, name = "f_conv_s1_x") as f_conv_s1_x:
with hcl.for_(0, 3, name = "f_conv_s1_r__y_r21") as f_conv_s1_r__y_r21:
with hcl.for_(0, 3, name = "f_conv_s1_r__x_r20") as f_conv_s1_r__x_r20:
t51_s = (f_conv_s1_z_z_par//hcl.select(((final_extent_2 + 31)//32) > 1, ((final_extent_2 + 31)//32), 1))
f_conv[f_conv_s1_x, (((f_conv_s1_y_y * 32) + final_min_1) + f_conv_s1_y_y_t), (((f_conv_s1_z_z_t_base_s * 32) + final_min_2) + f_conv_s1_z_z_t), ((f_conv_s1_z_z_par//hcl.select(((final_extent_2 + 31)//32) > 1, ((final_extent_2 + 31)//32), 1)) + final_min_3)] = (f_conv[f_conv_s1_x, (((f_conv_s1_y_y * 32) + final_min_1) + f_conv_s1_y_y_t), (((f_conv_s1_z_z_t_base_s * 32) + final_min_2) + f_conv_s1_z_z_t), (final_min_3 + t51_s)] + (filter[f_conv_s1_r__x_r20, f_conv_s1_r__y_r21, f_conv_s1_r__z, (((f_conv_s1_z_z_t_base_s * 32) + final_min_2) + f_conv_s1_z_z_t)] * input[(f_conv_s1_r__x_r20 + f_conv_s1_x), ((((f_conv_s1_y_y * 32) + final_min_1) + f_conv_s1_y_y_t) + f_conv_s1_r__y_r21), f_conv_s1_r__z, (final_min_3 + t51_s)]))
final = hcl.compute((64, 64, 32, 4), lambda x, y, z, w: 0, name = "final", dtype = hcl.Float(bits = 32))
with hcl.Stage("final"):
with hcl.for_(final_min_3, final_extent_3, name = "final_s0_n") as final_s0_n:
with hcl.for_(final_min_2, final_extent_2, name = "final_s0_z") as final_s0_z:
with hcl.for_(final_min_1, final_extent_1, name = "final_s0_y") as final_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "final_s0_x") as final_s0_x:
final[final_s0_x, final_s0_y, final_s0_z, final_s0_n] = hcl.select(f_conv[final_s0_x, final_s0_y, final_s0_z, final_s0_n] > hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.000000), f_conv[final_s0_x, final_s0_y, final_s0_z, final_s0_n], hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.000000))
return final
def kernel(A, B):
C = hcl.compute((10, 32), lambda *args: A[args] + B[args], "C")
with hcl.Stage("Super") as m:
hcl.update(C, lambda *args: C[args] + 1, "update")
with hcl.Stage("Plus") as stage:
with hcl.for_(0, 10) as j:
C[j, 0] = 10
return C
def top(input, ):
final_total_extent_1 = (
hcl.cast(dtype=hcl.Int(bits=64), expr=final_extent_1) *
hcl.cast(dtype=hcl.Int(bits=64), expr=final_extent_0))
max_local = hcl.compute((final_extent_0, final_extent_1),
lambda x, y: 0,
name="max_local",
dtype=hcl.UInt(bits=16))
with hcl.Stage("max_local"):
with hcl.for_(final_min_1, final_extent_1,
name="max_local_s0_y") as max_local_s0_y:
with hcl.for_(final_min_0, final_extent_0,
name="max_local_s0_x") as max_local_s0_x:
maximum = hcl.compute((1, 1),
lambda x, y: 0,
name="maximum",
dtype=hcl.UInt(bits=16))
with hcl.Stage("maximum"):
maximum[max_local_s0_x,
max_local_s0_y] = hcl.cast(dtype=hcl.UInt(bits=16),
expr=0)
with hcl.for_(
0, 3,
name="maximum_s1_box__y") as maximum_s1_box__y:
with hcl.for_(
0, 3,
name="maximum_s1_box__x") as maximum_s1_box__x:
maximum[max_local_s0_x,
max_local_s0_y] = hcl.select(
maximum[max_local_s0_x, max_local_s0_y]
> input[(max_local_s0_x +
maximum_s1_box__x),
(max_local_s0_y +
maximum_s1_box__y)],
maximum[max_local_s0_x,
max_local_s0_y],
input[(max_local_s0_x +
maximum_s1_box__x),
(max_local_s0_y +
maximum_s1_box__y)])
max_local[max_local_s0_x,
max_local_s0_y] = maximum[max_local_s0_x,
max_local_s0_y]
final = hcl.compute((640, 480),
lambda x, y: 0,
name="final",
dtype=hcl.UInt(bits=16))
with hcl.Stage("final"):
with hcl.for_(final_min_1, final_extent_1,
name="final_s0_y") as final_s0_y:
with hcl.for_(final_min_0, final_extent_0,
name="final_s0_x") as final_s0_x:
final[final_s0_x, final_s0_y] = max_local[final_s0_x,
final_s0_y]
return final
def kernel(trainData, testData, itemMem, idMem, rdv1, rdv2):
def train_encoding(m, preTrainData):
train_temp = hcl.compute((trainData.shape[1], dim), lambda x, y: itemMem[trainData[m][x]][y] ^ idMem[x][y], name = "train_temp")
k1 = hcl.reduce_axis(0, trainData.shape[1], 'k1')
train_result = hcl.compute((dim,), lambda x: hcl.sum(train_temp[k1, x], axis = k1, dtype=hcl.Int()), name = "train_result")
with hcl.for_(0, dim) as n:
preTrainData[m][n] = train_result[n]
with hcl.if_((m + 1) % 1000 == 0):
hcl.print((m+1), "Finish encoding %d training data\n")
def test_encoding(m, preTestData):
test_temp = hcl.compute((testData.shape[1], dim), lambda x, y: itemMem[testData[m][x]][y]^idMem[x][y], name = "test_temp")
k2 = hcl.reduce_axis(0, testData.shape[1], 'k2')
test_result = hcl.compute((dim,), lambda x: hcl.sum(test_temp[k2, x], axis = k2, dtype=hcl.Int()), name = "test_result")
with hcl.for_(0, dim) as n:
preTestData[m][n] = test_result[n]
with hcl.if_((m+1)%100 == 0):
hcl.print((m+1), "Finish encoding %d testing data\n")
#Encoding
hcl.print((), "Encoding the training data into HDVs.\n")
preTrainData = hcl.compute((trainData.shape[0], dim), lambda x, y: 0, "preTrainData")
hcl.mutate((trainData.shape[0], ), lambda x: train_encoding(x, preTrainData))
hdTrainData = hcl.compute((trainData.shape[0], dim), lambda x, y: 0, "hdTrainData", dtype=hcl.UInt(1))
with hcl.Stage("S1"):
with hcl.if_(trainData.shape[1] % 2 == 0):
hcl.print((), "Use the random vector\n")
hcl.update(hdTrainData, lambda x, y: hcl.select(preTrainData[x][y] + rdv1[x][y] - trainData.shape[1]/2 > 0, 1, 0))
with hcl.else_():
hcl.update(hdTrainData, lambda x, y: hcl.select(preTrainData[x][y] - trainData.shape[1]/2 > 0, 1, 0))
hcl.print((),"Encoding the testing data into HDVs.\n")
preTestData = hcl.compute((testData.shape[0], dim), lambda x, y: 0, "preTestData")
hcl.mutate((testData.shape[0], ), lambda x: test_encoding(x, preTestData))
hdTestData = hcl.compute((testData.shape[0], dim), lambda x, y: 0, "hdTestData", dtype=hcl.UInt(1))
with hcl.Stage("S2"):
with hcl.if_(testData.shape[1] % 2 == 0):
hcl.print((), "Use the random vector\n")
hcl.update(hdTestData, lambda x, y: hcl.select(preTestData[x][y] + rdv2[x][y] - testData.shape[1]/2 > 0, 1, 0))
with hcl.else_():
hcl.update(hdTestData, lambda x, y: hcl.select(preTestData[x][y] - testData.shape[1]/2 > 0, 1, 0))
###data_packing
pack_train = hcl.pack(hdTrainData, axis=1, dtype=hcl.UInt(bw), name="pack_train")
pack_test = hcl.pack(hdTestData, axis=1, dtype=hcl.UInt(bw), name="pack_test")
return pack_train, pack_test
def fft(X_real, X_imag, IndexTable, F_real, F_imag):
L = X_real.shape[0]
if np.log2(L) % 1 > 0:
raise ValueError("Length of input vector (1d tensor) must be power of 2")
num_stages = int(np.log2(L))
# bit reverse permutation
hcl.update(F_real, lambda i: X_real[IndexTable[i]], name='F_real_update')
hcl.update(F_imag, lambda i: X_imag[IndexTable[i]], name='F_imag_update')
with hcl.Stage("Out"):
one = hcl.scalar(1, dtype="int32")
with hcl.for_(0, num_stages) as stage:
DFTpts = one[0] << (stage + 1)
numBF = DFTpts / 2
e = -2 * np.pi / DFTpts
a = hcl.scalar(0)
with hcl.for_(0, numBF) as j:
c = hcl.scalar(hcl.cos(a[0]))
s = hcl.scalar(hcl.sin(a[0]))
a[0] = a[0] + e
with hcl.for_(j, L + DFTpts - 1, DFTpts) as i:
i_lower = i + numBF
temp_r = hcl.scalar(F_real[i_lower] * c - F_imag[i_lower] * s)
temp_i = hcl.scalar(F_imag[i_lower] * c + F_real[i_lower] * s)
F_real[i_lower] = F_real[i] - temp_r[0]
F_imag[i_lower] = F_imag[i] - temp_i[0]
F_real[i] = F_real[i] + temp_r[0]
F_imag[i] = F_imag[i] + temp_i[0]
def kernel(A):
with hcl.Stage():
with hcl.for_(0, 10) as i:
with hcl.for_(0, 10) as j:
with hcl.if_(j >= i):
hcl.break_()
A[i] += j
def kernel(A, B, C, O):
dtype_xyz = hcl.Struct({
"x": hcl.Int(),
"y": hcl.Int(),
"z": hcl.Int()
})
dtype_out = hcl.Struct({
"v0": hcl.Int(),
"v1": hcl.Int(),
"v2": hcl.Int(),
"v3": hcl.Int(),
"v4": hcl.Int(),
"v5": hcl.Int()
})
D = hcl.compute(A.shape, lambda x: (A[x], B[x], C[x]), dtype=dtype_xyz)
E = hcl.compute(A.shape,
lambda x:
(D[x].x * D[x].x, D[x].y * D[x].y, D[x].z * D[x].z, D[
x].x * D[x].y, D[x].y * D[x].z, D[x].x * D[x].z),
dtype=dtype_out)
with hcl.Stage():
with hcl.for_(0, 100) as i:
for j in range(0, 6):
O[i][j] = E[i].__getattr__("v" + str(j))
def func(data):
out = hcl.compute((4, 4), lambda x, y: 0, "out", dtype)
with hcl.Stage("S"):
with hcl.for_(0, 4, name="i") as i:
with hcl.for_(0, 4, name="j") as j:
out[i, j] = data[i, j] + 1
return out
def kernel(A, B, C):
with hcl.Stage("S"):
with hcl.for_(0, 10) as i:
# set the LSB of B to be the same as A
B[i][0] = A[i][0]
# set the lower 4-bit of C
C[i][4:0] = A[i]
def algorithm(A, B):
@hcl.def_([A.shape, B.shape, ()])
def update_B(A, B, x):
B[x] = A[x] + 1
with hcl.Stage():
with hcl.for_(0, 10) as i:
update_B(A, B, i)
def kernel(A):
with hcl.Stage():
with hcl.if_(A[0] > 5):
A[0] = 5
with hcl.elif_(A[0] > 3):
A[0] = 3
with hcl.else_():
A[0] = 0
def kernel(A):
with hcl.Stage():
i = hcl.scalar(0)
with hcl.while_(True):
with hcl.if_(i[0] > 5):
hcl.break_()
A[i[0]] = i[0]
i[0] += 1
def kernel(A, B):
C = hcl.compute(A.shape, lambda *args : 0, "C")
with hcl.Stage("stage"):
with hcl.for_(0, 10, name="i") as i:
with hcl.for_(0, 32, name="j") as j:
B[i, j] = A[i, j] + B[i, j]
C[i, j] = 2 * B[i, j]
return C
def kernel(A, B):
C = hcl.compute((10, 32), lambda *args : 0, "C")
D = hcl.compute(C.shape, lambda *args: 0, "D")
with hcl.Stage("Super") as m:
with hcl.for_(0, 10, name="j") as j:
hcl.update(D, lambda *args: j*A[args] + B[args], name="update.D")
hcl.update(C, lambda *args: A[args] + j*D[args], name="update.C")
return C
def top(input, ):
final_total_extent_1 = (hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_1) * hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_0))
blur_x = hcl.compute((final_extent_0, (final_extent_1 + 2)), lambda x, y: 0, name = "blur_x", dtype = hcl.UInt(bits = 16))
with hcl.Stage("blur_x"):
with hcl.for_(final_min_1, (final_extent_1 + 2), name = "blur_x_s0_y") as blur_x_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "blur_x_s0_x") as blur_x_s0_x:
blur_x[blur_x_s0_x, blur_x_s0_y] = ((input[(blur_x_s0_x + 2), blur_x_s0_y] + (input[blur_x_s0_x, blur_x_s0_y] + input[(blur_x_s0_x + 1), blur_x_s0_y]))//hcl.cast(dtype = hcl.UInt(bits = 16), expr = 3))
blur_y = hcl.compute((final_extent_0, final_extent_1), lambda x, y: 0, name = "blur_y", dtype = hcl.UInt(bits = 16))
with hcl.Stage("blur_y"):
with hcl.for_(final_min_1, final_extent_1, name = "blur_y_s0_y") as blur_y_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "blur_y_s0_x") as blur_y_s0_x:
blur_y[blur_y_s0_x, blur_y_s0_y] = ((blur_x[blur_y_s0_x, (blur_y_s0_y + 2)] + (blur_x[blur_y_s0_x, blur_y_s0_y] + blur_x[blur_y_s0_x, (blur_y_s0_y + 1)]))//hcl.cast(dtype = hcl.UInt(bits = 16), expr = 3))
final = hcl.compute((640, 480), lambda x, y: 0, name = "final", dtype = hcl.UInt(bits = 16))
with hcl.Stage("final"):
with hcl.for_(final_min_1, final_extent_1, name = "final_s0_y") as final_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "final_s0_x") as final_s0_x:
final[final_s0_x, final_s0_y] = blur_y[final_s0_x, final_s0_y]
return final
def top(input, ):
final_total_extent_1 = (
hcl.cast(dtype=hcl.Int(bits=64), expr=final_extent_1) *
hcl.cast(dtype=hcl.Int(bits=64), expr=final_extent_0))
mean_local = hcl.compute((final_extent_0, final_extent_1),
lambda x, y: 0,
name="mean_local",
dtype=hcl.UInt(bits=16))
with hcl.Stage("mean_local"):
with hcl.for_(final_min_1, final_extent_1,
name="mean_local_s0_y") as mean_local_s0_y:
with hcl.for_(final_min_0, final_extent_0,
name="mean_local_s0_x") as mean_local_s0_x:
mean_local[mean_local_s0_x,
mean_local_s0_y] = hcl.cast(dtype=hcl.UInt(bits=16),
expr=0)
with hcl.for_(final_min_1, final_extent_1,
name="mean_local_s1_y") as mean_local_s1_y:
with hcl.for_(final_min_0, final_extent_0,
name="mean_local_s1_x") as mean_local_s1_x:
with hcl.for_(
0, 3,
name="mean_local_s1_box__y") as mean_local_s1_box__y:
with hcl.for_(0, 3, name="mean_local_s1_box__x"
) as mean_local_s1_box__x:
mean_local[mean_local_s1_x, mean_local_s1_y] = (
mean_local[mean_local_s1_x, mean_local_s1_y] +
(input[(mean_local_s1_box__x + mean_local_s1_x),
(mean_local_s1_box__y + mean_local_s1_y)] //
hcl.cast(dtype=hcl.UInt(bits=16), expr=9)))
final = hcl.compute((6418, 4818),
lambda x, y: 0,
name="final",
dtype=hcl.UInt(bits=16))
with hcl.Stage("final"):
with hcl.for_(final_min_1, final_extent_1,
name="final_s0_y") as final_s0_y:
with hcl.for_(final_min_0, final_extent_0,
name="final_s0_x") as final_s0_x:
final[final_s0_x, final_s0_y] = mean_local[final_s0_x,
final_s0_y]
return final
def test_schedule_intra_stage():
hcl.init()
def popcount(A, B): # each element in A is a 32-bit integer
with hcl.for_(0, A.shape[0], name="x") as x:
with hcl.for_(0, A.shape[1], name="y") as y:
B[x, y] = 0
with hcl.for_(0, 32) as i:
B[x, y] += A[x, y][i]
A = hcl.placeholder((10, 20))
B = hcl.placeholder(A.shape)
with hcl.Stage() as C:
popcount(A, B)
def test_unroll():
s = hcl.create_schedule([A, B])
s[C].unroll(C.x, factor=3)
ir = hcl.lower(s)
assert "unrolled \"factor\"=3" in str(ir)
def test_reorder():
s = hcl.create_schedule([A, B])
s[C].reorder(C.y, C.x)
ir = hcl.lower(s)
assert str(ir.body.body.body.body).startswith("for (y, 0, 20)")
assert str(ir.body.body.body.body.body).startswith("for (x, 0, 10)")
def test_fuse():
s = hcl.create_schedule([A, B])
s[C].fuse(C.x, C.y)
ir = hcl.lower(s)
assert str(
ir.body.body.body.body).startswith("for (x.y.fused, 0, 200)")
def test_split():
s = hcl.create_schedule([A, B])
s[C].split(C.x, factor=3)
ir = hcl.lower(s)
assert str(ir.body.body.body.body).startswith("for (x.outer, 0, 4)")
assert str(
ir.body.body.body.body.body).startswith("for (x.inner, 0, 3)")
assert str(
ir.body.body.body.body.body.body).startswith("for (y, 0, 20)")
assert str(ir.body.body.body.body.body.body.body).startswith(
"if ((x.inner < (10 - (x.outer*3))))")
test_unroll()
test_reorder()
test_fuse()
test_split()
def vadd_vhls_ip(op1, op2, size, name=None):
if name is None: name = "vadd"
with hcl.Stage("ExternModule.vadd") as Module:
register_tensors([op1, op2])
Module.ext_ip_name = name
Module.inputs = [op1, op2, size]
# include cpp/hpp files
deps = os.path.dirname(os.path.abspath(__file__))
source = ["vadd.cpp"]
Module.source = include_dependency(source)
create_extern_module(Module, ip_type="HLS")
def insertion_sort(A):
# Introduce a stage.
with hcl.Stage("S"):
# for i in range(1, A.shape[0])
# We can name the axis
with hcl.for_(1, A.shape[0], name="i") as i:
key = hcl.local(A[i], "key")
j = hcl.local(i - 1, "j")
# while(j >= 0 && key < A[j])
with hcl.while_(hcl.and_(j >= 0, key < A[j])):
A[j + 1] = A[j]
j[0] -= 1
A[j + 1] = key[0]
def systolic_array(A, B):
# define modules with loop
@hcl.def_([(1,), (1,), ()])
def pe(a, b, x):
with hcl.if_(x == 0):
result = a * b
hcl.return_(a)
with hcl.elif_(x == 1):
hcl.return_(b)
with hcl.else_():
hcl.return_(result)
# PE = {f'pe_{i}' : partial(pe) for i in range(w*h)}
PE = {}
for i in range(w * h):
with hcl.Stage("pe_{}".format(i)):
PE['pe_{}'.format(i)] = partial(pe)
# each k calls of update function calculate one block of result matrix
# b_row: block row index
# b_col: block col index
def update(b_row, b_col, k, O):
# fetch input
localA = []
localB = []
for input_a in range(h):
localA.append(hcl.compute((1,), lambda x : A[input_a + h * b_row, k], "localA_{}".format(input_a)))
for input_b in range(w):
localB.append(hcl.compute((1,), lambda x : B[k, input_b + w * b_col], "localB_{}".format(input_b)))
# systolic connection
net = [[None] * h] * w
for i in range(h + w - 1):
for row in range(i + 1):
col = i - row
if col < 0 or col > w-1 or row > h-1: continue
## instantiate a PE and record partial results
input_a = localA[row] if col == 0 else hcl.compute((1,), lambda x : net[row][col-1][0], "input_a{}{}".format(row, col))
input_b = localB[col] if row == 0 else hcl.compute((1,), lambda x : net[row-1][col][1], "input_b{}{}".format(row, col))
out = hcl.compute((3,), lambda x : PE['pe_%d' % (row * w + col)](
input_a, input_b, x), "out_{}{}".format(row, col))
O[row + h * b_row, col + w * b_col] += out[2]
net[row][col] = out
block_rows = int(m / h)
block_cols = int(n / w)
O = hcl.compute((m, n), lambda *args : 0, name="Output")
hcl.mutate((block_rows, block_cols, k), lambda b_row, b_col, k: update(b_row, b_col, k, O), name="update")
return O
def top(
A,
B,
):
final_total_extent_1 = (
hcl.cast(dtype=hcl.Int(bits=64), expr=final_extent_1) *
hcl.cast(dtype=hcl.Int(bits=64), expr=final_extent_0))
prod = hcl.compute((final_extent_0, final_extent_1),
lambda x, y: 0,
name="prod",
dtype=hcl.Float(bits=32))
with hcl.Stage("prod"):
with hcl.for_(final_min_1, final_extent_1,
name="prod_s0_y") as prod_s0_y:
with hcl.for_(final_min_0, final_extent_0,
name="prod_s0_x") as prod_s0_x:
prod[prod_s0_x, prod_s0_y] = hcl.cast(dtype=hcl.Float(bits=32),
expr=0.000000)
with hcl.for_(final_min_1, final_extent_1,
name="prod_s1_y") as prod_s1_y:
with hcl.for_(final_min_0, final_extent_0,
name="prod_s1_x") as prod_s1_x:
with hcl.for_(0, 1024, name="prod_s1_r__x") as prod_s1_r__x:
prod[prod_s1_x, prod_s1_y] = (prod[prod_s1_x, prod_s1_y] +
(A[prod_s1_x, prod_s1_r__x] *
B[prod_s1_r__x, prod_s1_y]))
final = hcl.compute((1024, 1024),
lambda x, y: 0,
name="final",
dtype=hcl.Float(bits=32))
with hcl.Stage("final"):
with hcl.for_(final_min_1, final_extent_1,
name="final_s0_y") as final_s0_y:
with hcl.for_(final_min_0, final_extent_0,
name="final_s0_x") as final_s0_x:
final[final_s0_x, final_s0_y] = prod[final_s0_x, final_s0_y]
return final
def kernel(A):
B = hcl.compute((2, 2),lambda x, y: 0, "B", dtype) # syntax sugar
with hcl.Stage("S"):
LB = hcl.compute((2, 4),lambda x, y: 0, "LB", dtype)
with hcl.for_(0, 2, name="x") as x:
with hcl.for_(0, 2, name="y") as y:
with hcl.for_(0, 2, name="LB_i") as LB_i:
with hcl.for_(0, 4, name="LB_j") as LB_j:
LB[LB_i, LB_j] = A[x * 2 + LB_i, LB_j]
val = hcl.scalar(0,"val")
with hcl.for_(0, 2, name="r") as r:
with hcl.for_(0, 2, name="c") as c:
val.v += LB[r, y * 2 + c]
B[x, y] = val / 4
return B
def pool(data):
out = hcl.compute((2, 2), lambda x, y: 0, "out", dtype)
with hcl.Stage("S"):
LB = hcl.compute((2, 4), lambda x, y: 0, "LB", dtype)
with hcl.for_(0, 2, name="x") as x:
with hcl.for_(0, 2, name="y") as y:
with hcl.for_(0, 2, name="LB_i") as LB_i:
with hcl.for_(0, 4, name="LB_j") as LB_j:
LB[LB_i, LB_j] = data[x * 2 + LB_i, LB_j]
val = hcl.scalar(0, "val")
with hcl.for_(0, 2, name="r") as r:
with hcl.for_(0, 2, name="c") as c:
val.v += LB[r, y * 2 + c]
out[x, y] = val / 4
return out
def top(input, ):
final_total_extent_1 = (hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_1) * hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_0))
final_total_extent_2 = (final_total_extent_1 * hcl.cast(dtype = hcl.Int(bits = 64), expr = final_extent_2))
linear = hcl.compute(((final_extent_0 + 2), (final_extent_1 + 2), final_extent_2), lambda x, y, z: 0, name = "linear", dtype = hcl.Float(bits = 32))
with hcl.Stage("linear"):
with hcl.for_(final_min_2, final_extent_2, name = "linear_s0_c") as linear_s0_c:
with hcl.for_(final_min_1, (final_extent_1 + 2), name = "linear_s0_y") as linear_s0_y:
with hcl.for_(final_min_0, (final_extent_0 + 2), name = "linear_s0_x") as linear_s0_x:
t4 = input[linear_s0_x, linear_s0_y, linear_s0_c]
linear[linear_s0_x, linear_s0_y, linear_s0_c] = hcl.select((hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.040450) < t4), hcl.power(((t4 * hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.947867)) + hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.052133)), hcl.cast(dtype = hcl.Float(bits = 32), expr = 2.400000)), (t4 * hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.077399)))
blur_x = hcl.compute((final_extent_0, (final_extent_1 + 2), final_extent_2), lambda x, y, z: 0, name = "blur_x", dtype = hcl.Float(bits = 32))
with hcl.Stage("blur_x"):
with hcl.for_(final_min_2, final_extent_2, name = "blur_x_s0_c") as blur_x_s0_c:
with hcl.for_(final_min_1, (final_extent_1 + 2), name = "blur_x_s0_y") as blur_x_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "blur_x_s0_x") as blur_x_s0_x:
blur_x[blur_x_s0_x, blur_x_s0_y, blur_x_s0_c] = ((linear[(blur_x_s0_x + 2), blur_x_s0_y, blur_x_s0_c] + (linear[blur_x_s0_x, blur_x_s0_y, blur_x_s0_c] + linear[(blur_x_s0_x + 1), blur_x_s0_y, blur_x_s0_c])) * hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.333333))
blur_y = hcl.compute((final_extent_0, final_extent_1, final_extent_2), lambda x, y, z: 0, name = "blur_y", dtype = hcl.Float(bits = 32))
with hcl.Stage("blur_y"):
with hcl.for_(final_min_2, final_extent_2, name = "blur_y_s0_c") as blur_y_s0_c:
with hcl.for_(final_min_1, final_extent_1, name = "blur_y_s0_y") as blur_y_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "blur_y_s0_x") as blur_y_s0_x:
blur_y[blur_y_s0_x, blur_y_s0_y, blur_y_s0_c] = ((blur_x[blur_y_s0_x, (blur_y_s0_y + 2), blur_y_s0_c] + (blur_x[blur_y_s0_x, blur_y_s0_y, blur_y_s0_c] + blur_x[blur_y_s0_x, (blur_y_s0_y + 1), blur_y_s0_c])) * hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.333333))
srgb = hcl.compute((final_extent_0, final_extent_1, final_extent_2), lambda x, y, z: 0, name = "srgb", dtype = hcl.Float(bits = 32))
with hcl.Stage("srgb"):
with hcl.for_(final_min_2, final_extent_2, name = "srgb_s0_c") as srgb_s0_c:
with hcl.for_(final_min_1, final_extent_1, name = "srgb_s0_y") as srgb_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "srgb_s0_x") as srgb_s0_x:
t5 = blur_y[srgb_s0_x, srgb_s0_y, srgb_s0_c]
srgb[srgb_s0_x, srgb_s0_y, srgb_s0_c] = hcl.select((hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.003131) < t5), ((hcl.power(t5, hcl.cast(dtype = hcl.Float(bits = 32), expr = 0.416667)) * hcl.cast(dtype = hcl.Float(bits = 32), expr = 1.055000)) + hcl.cast(dtype = hcl.Float(bits = 32), expr = -0.055000)), (t5 * hcl.cast(dtype = hcl.Float(bits = 32), expr = 12.920000)))
final = hcl.compute((766, 1278, 3), lambda x, y, z: 0, name = "final", dtype = hcl.Float(bits = 32))
with hcl.Stage("final"):
with hcl.for_(final_min_2, final_extent_2, name = "final_s0_c") as final_s0_c:
with hcl.for_(final_min_1, final_extent_1, name = "final_s0_y") as final_s0_y:
with hcl.for_(final_min_0, final_extent_0, name = "final_s0_x") as final_s0_x:
final[final_s0_x, final_s0_y, final_s0_c] = srgb[final_s0_x, final_s0_y, final_s0_c]
return final
def zculling(size_pixels,size,fragment,z_buffer,pixels):
pixel_cntr = hcl.scalar(0,dtype=hcl.Int())
with hcl.Stage("S2"):
with hcl.for_(0,size) as n:
x = hcl.scalar(fragment[n][0],dtype=hcl.Int())
y = hcl.scalar(fragment[n][1],dtype=hcl.Int())
z = hcl.scalar(fragment[n][2])
color = hcl.scalar(fragment[n][3])
with hcl.if_( z < z_buffer[y][x] ):
pixels[pixel_cntr][0] = x.v
pixels[pixel_cntr][1] = y.v
pixels[pixel_cntr][2] = color.v
pixel_cntr.v += 1
z_buffer[y][x] = z.v
size_pixels[0] = pixel_cntr.v
def test_if():
hcl.init()
def absolute(A, B):
with hcl.for_(0, A.shape[0], name="x") as x:
with hcl.for_(0, A.shape[1], name="y") as y:
with hcl.if_(A[x, y] >= 0):
B[x, y] = A[x, y]
with hcl.else_():
B[x, y] = -A[x, y]
A = hcl.placeholder((10, 20), name="A", dtype="float32")
B = hcl.placeholder(A.shape, name="B", dtype="float32")
with hcl.Stage() as C:
absolute(A, B)
s = hcl.create_schedule([A, B])
o, i = s[C].split(C.x, factor=3)
s[C].reorder(i, o)
# test lower
ir = hcl.lower(s)
assert str(ir.body.body.body.body).startswith("for (x.inner, 0, 3)")
assert str(ir.body.body.body.body.body).startswith("for (x.outer, 0, 4)")
assert str(ir.body.body.body.body.body.body).startswith("for (y, 0, 20)")
assert str(ir.body.body.body.body.body.body.body.condition).startswith(
"(x.inner < (10 - (x.outer*3)))")
assert str(
ir.body.body.body.body.body.body.body.then_case.condition).startswith(
"(0.000000f <= A[(y + ((x.inner + (x.outer*3))*20))])")
assert str(
ir.body.body.body.body.body.body.body.then_case.then_case
).startswith(
"B[(y + ((x.inner + (x.outer*3))*20))] = A[(y + ((x.inner + (x.outer*3))*20))]"
)
assert str(
ir.body.body.body.body.body.body.body.then_case.else_case
).startswith(
"B[(y + ((x.inner + (x.outer*3))*20))] = (A[(y + ((x.inner + (x.outer*3))*20))]*-1.000000f)"
)
# test build
f = hcl.build(s)
a_np = np.random.random((A.shape))
a_hcl = hcl.asarray(a_np, dtype="float32")
b_hcl = hcl.asarray(np.zeros(B.shape), dtype="float32")
f(a_hcl, b_hcl)
b_np = np.abs(a_np)
np.testing.assert_allclose(b_np, b_hcl.asnumpy())
def byte_swap_rtl(input_vec, ret=None, name=None):
if name is None: name = "my_byteswap"
Len = input_vec.shape[0]
return_tensors = False
if ret is None:
return_tensors = True
ret = hcl.compute(input_vec.shape, lambda *args: 0, "vec")
# functional behavior
with hcl.Stage("ExternModule") as Module:
hcl.update(ret,
lambda *args: input_vec[args] << 16 | input_vec[args] >> 16,
"swap")
dicts = {}
dicts["name"] = name
tensors = [input_vec]
dicts["args"] = [(_.name, _.dtype) for _ in tensors]
# declare headers and typedef
dicts["header"] = "unsigned int my_byteswap(unsigned int x);"
dicts["func"] = """
for (int k = 0; k < {}; k++) {{
vec[k] = my_byteswap({}[k]);
}}
""".format(Len, input_vec.name)
# add dependency files or folders
# the dependencies are copied to project folder
deps = os.path.dirname(os.path.abspath(__file__))
dicts["deps"] = deps + "/lib1"
# custom compilation command (root path: project)
# commands executed before impl or emulation
dicts["cmds"] = "cd lib1; " + \
"aocl library hdl-comp-pkg opencl_lib.xml -o opencl_lib.aoco;" + \
"aocl library create -name opencl_lib opencl_lib.aoco;"
# custom compiler flgas (load custom libs)
dicts["flags"] = "-I lib1 -L lib1 -l opencl_lib.aoclib"
create_extern_module(Module, dicts, ip_type="rtl")
if return_tensors: return ret
def toynn_vhls_ip(input_1, output_1, name=None):
if name is None: name = "myproject"
# Function behavior definition
with hcl.Stage("ExternModule.toyNN") as Module:
register_tensors([input_1, output_1])
Module.ext_ip_name = name
Module.inputs = [input_1, output_1]
# Include cpp/hpp files
if not os.path.exists("firmware"):
urllib.request.urlretrieve(
"https://raw.githubusercontent.com/Hecmay/debug.trace/main/toynn.tar.gz",
filename="toynn.tar.gz")
os.system("tar -zxvf toynn.tar.gz")
source = [
"firmware/myproject.cpp", "firmware/nnet_utils/", "firmware/weights/"
]
Module.source = include_dependency(source)
create_extern_module(Module, ip_type="HLS")
def vadd_rtl(A, B, length, ret=None, name=None):
if name is None: name = "vadd_rtl"
Len = A.shape[0]
assert A.shape == B.shape, "shape not match"
assert Len == length, "shape not match"
return_tensors = False
if ret is None:
return_tensors = True
ret = hcl.compute(A.shape, lambda *args: 0, "ret")
# functional behavior
with hcl.Stage("ExternModule") as Module:
hcl.update(ret, lambda *args:
A[args] + B[args], "vadd")
dicts = {}
dicts["name"] = name
tensors = [A, B]
dicts["args"] = [(_.name, _.dtype) for _ in tensors]
# RTL IP is wrapped as a separate OpenCL kernel in Vitis
# add dependency files or folders
# the dependencies are copied to project folder
deps = os.path.dirname(os.path.abspath(__file__))
dicts["deps"] = deps + "/scripts"
# custom compilation command (root path: project)
# commands executed before impl or emulation
dicts["cmds"] = "vivado -mode batch -source " + \
"scripts/gen_xo.tcl -tclargs vadd.xo vadd hw_emu {} {}"
# custom compiler flgas (load custom libs)
dicts["flags"] = "vadd.xo"
create_extern_module(Module, dicts, ip_type="rtl")
if return_tensors: return ret
def HJ_PDE_solver(V_new, V_init, thetas):
# Calculate spatial derivative based on index and dimension number
def spatial_derivative(i, j, k, dim):
left = i * j - k
right = i * j + k
return left, right
# Calculate Hamiltonian for every grid point in V_init
with hcl.Stage("Hamiltonian"):
with hcl.for_(1, V_init.shape[0], name="i") as i:
with hcl.for_(1, V_init.shape[1], name="j") as j:
with hcl.for_(1, V_init.shape[2], name="k") as k:
# Calculate dV_dx
dV_dx_L, dV_dx_R = spatial_derivative(i, j, k, 0)
dV_dy_L, dV_dy_R = spatial_derivative(i, j, k, 1)
dV_dtheta_L, dV_dtheta_R = spatial_derivative(i, j, k, 2)
# Calculate average gradient
dV_dx_C = (dV_dx_L + dV_dx_R) / 2
dV_dy_C = (dV_dy_L + dV_dy_R) / 2
dV_dtheta_C = (dV_dtheta_L + dV_dtheta_R) / 2
# Get optimal control
uOpt = 1
# Velocity
v = 1
# Assume that mode is min
with hcl.if_(dV_dtheta_C > 0):
uOpt = -uOpt
# Calculate dynamics function
#V_new[i,j,k] = 1 * cos(thetas[k]) * dV_dx_C +1 * sin(thetas[k]) * dV_dy_C +uOpt * dV_theta_C
#angle = hcl.scalar(thetas[k], "angle")
V_new[i, j,
k] = v * hcl.cos(thetas[k]) * dV_dx_C + v * hcl.sin(
thetas[k]) * dV_dy_C + dV_dtheta_C * uOpt