def test_basics3():
input = hl.ImageParam(hl.Float(32), 3, 'input')
r_sigma = hl.Param(hl.Float(32), 'r_sigma',
0.1) # Value needed if not generating an executable
s_sigma = 8 # This is passed during code generation in the C++ version
x = hl.Var('x')
y = hl.Var('y')
z = hl.Var('z')
c = hl.Var('c')
# Add a boundary condition
clamped = hl.Func('clamped')
clamped[x, y] = input[hl.clamp(x, 0,
input.width() - 1),
hl.clamp(y, 0,
input.height() - 1), 0]
# Construct the bilateral grid
r = hl.RDom([(0, s_sigma), (0, s_sigma)], 'r')
val = clamped[x * s_sigma + r.x - s_sigma // 2,
y * s_sigma + r.y - s_sigma // 2]
val = hl.clamp(val, 0.0, 1.0)
zi = hl.i32((val / r_sigma) + 0.5)
histogram = hl.Func('histogram')
histogram[x, y, z, c] = 0.0
ss = hl.select(c == 0, val, 1.0)
left = histogram[x, y, zi, c]
left += 5
left += ss
def test_basics2():
input = hl.ImageParam(hl.Float(32), 3, 'input')
r_sigma = hl.Param(hl.Float(32), 'r_sigma', 0.1) # Value needed if not generating an executable
s_sigma = 8 # This is passed during code generation in the C++ version
x = hl.Var('x')
y = hl.Var('y')
z = hl.Var('z')
c = hl.Var('c')
# Add a boundary condition
clamped = hl.Func('clamped')
clamped[x, y] = input[hl.clamp(x, 0, input.width()-1),
hl.clamp(y, 0, input.height()-1),0]
# Construct the bilateral grid
r = hl.RDom(0, s_sigma, 0, s_sigma, 'r')
val0 = clamped[x * s_sigma, y * s_sigma]
val00 = clamped[x * s_sigma * hl.cast(hl.Int(32), 1), y * s_sigma * hl.cast(hl.Int(32), 1)]
#val1 = clamped[x * s_sigma - s_sigma/2, y * s_sigma - s_sigma/2] # should fail
val22 = clamped[x * s_sigma - hl.cast(hl.Int(32), s_sigma//2),
y * s_sigma - hl.cast(hl.Int(32), s_sigma//2)]
val2 = clamped[x * s_sigma - s_sigma//2, y * s_sigma - s_sigma//2]
val3 = clamped[x * s_sigma + r.x - s_sigma//2, y * s_sigma + r.y - s_sigma//2]
return
def test_basics2():
input = hl.ImageParam(hl.Float(32), 3, 'input')
r_sigma = hl.Param(hl.Float(32), 'r_sigma', 0.1)
s_sigma = 8
x = hl.Var('x')
y = hl.Var('y')
z = hl.Var('z')
c = hl.Var('c')
# Add a boundary condition
clamped = hl.Func('clamped')
clamped[x, y] = input[hl.clamp(x, 0,
input.width() - 1),
hl.clamp(y, 0,
input.height() - 1), 0]
# Construct the bilateral grid
r = hl.RDom([(0, s_sigma), (0, s_sigma)], 'r')
val0 = clamped[x * s_sigma, y * s_sigma]
val00 = clamped[x * s_sigma * hl.i32(1), y * s_sigma * hl.i32(1)]
val22 = clamped[x * s_sigma - hl.i32(s_sigma // 2),
y * s_sigma - hl.i32(s_sigma // 2)]
val2 = clamped[x * s_sigma - s_sigma // 2, y * s_sigma - s_sigma // 2]
val3 = clamped[x * s_sigma + r.x - s_sigma // 2,
y * s_sigma + r.y - s_sigma // 2]
try:
val1 = clamped[x * s_sigma - s_sigma / 2, y * s_sigma - s_sigma / 2]
except RuntimeError as e:
assert 'Implicit cast from float32 to int' in str(e)
else:
assert False, 'Did not see expected exception!'
def test_basics():
input = hl.ImageParam(hl.UInt(16), 2, 'input')
x, y = hl.Var('x'), hl.Var('y')
blur_x = hl.Func('blur_x')
blur_xx = hl.Func('blur_xx')
blur_y = hl.Func('blur_y')
yy = hl.i32(1)
assert yy.type() == hl.Int(32)
z = x + 1
input[x, y]
input[0, 0]
input[z, y]
input[x + 1, y]
input[x, y] + input[x + 1, y]
if False:
aa = blur_x[x, y]
bb = blur_x[x, y + 1]
aa + bb
blur_x[x, y] + blur_x[x, y + 1]
(input[x, y] + input[x + 1, y]) / 2
blur_x[x, y]
blur_xx[x, y] = input[x, y]
blur_x[x, y] = (input[x, y] + input[x + 1, y] + input[x + 2, y]) / 3
blur_y[x, y] = (blur_x[x, y] + blur_x[x, y + 1] + blur_x[x, y + 2]) / 3
xi, yi = hl.Var('xi'), hl.Var('yi')
blur_y.tile(x, y, xi, yi, 8, 4).parallel(y).vectorize(xi, 8)
blur_x.compute_at(blur_y, x).vectorize(x, 8)
blur_y.compile_jit()
def main():
# define and compile the function
input = hl.ImageParam(hl.UInt(8), 3, "input")
erode = get_erode(input)
erode.compile_jit()
# preparing input and output memory buffers (numpy ndarrays)
input_data = get_input_data()
input_image = hl.Buffer(input_data)
input.set(input_image)
output_data = np.empty(input_data.shape, dtype=input_data.dtype, order="F")
output_image = hl.Buffer(output_data)
print("input_image", input_image)
print("output_image", output_image)
# do the actual computation
erode.realize(output_image)
# save results
input_path = "erode_input.png"
output_path = "erode_result.png"
imageio.imsave(input_path, input_data)
imageio.imsave(output_path, output_data)
print("\nerode realized on output image.", "Result saved at", output_path,
"( input data copy at", input_path, ")")
print("\nEnd of game. Have a nice day!")
return
def main():
input = hl.ImageParam(float_t, 3, "input")
levels = 10
interpolate = get_interpolate(input, levels)
# preparing input and output memory buffers (numpy ndarrays)
input_data = get_input_data()
assert input_data.shape[2] == 4
input_image = hl.Buffer(input_data)
input.set(input_image)
input_width, input_height = input_data.shape[:2]
t0 = datetime.now()
output_image = interpolate.realize(input_width, input_height, 3)
t1 = datetime.now()
print('Interpolated in %.5f secs' % (t1 - t0).total_seconds())
output_data = hl.buffer_to_ndarray(output_image)
# save results
input_path = "interpolate_input.png"
output_path = "interpolate_result.png"
imsave(input_path, input_data)
imsave(output_path, output_data)
print("\nblur realized on output image.", "Result saved at", output_path,
"( input data copy at", input_path, ")")
print("\nEnd of game. Have a nice day!")
def test_basics2():
input = hl.ImageParam(hl.Float(32), 3, 'input')
r_sigma = hl.Param(hl.Float(32), 'r_sigma',
0.1) # Value needed if not generating an executable
s_sigma = 8 # This is passed during code generation in the C++ version
x = hl.Var('x')
y = hl.Var('y')
z = hl.Var('z')
c = hl.Var('c')
# Add a boundary condition
clamped = hl.Func('clamped')
clamped[x, y] = input[hl.clamp(x, 0,
input.width() - 1),
hl.clamp(y, 0,
input.height() - 1), 0]
if True:
print("s_sigma", s_sigma)
print("s_sigma/2", s_sigma / 2)
print("s_sigma//2", s_sigma // 2)
print()
print("x * s_sigma", x * s_sigma)
print("x * 8", x * 8)
print("x * 8 + 4", x * 8 + 4)
print("x * 8 * 4", x * 8 * 4)
print()
print("x", x)
print("(x * s_sigma).type()", )
print("(x * 8).type()", (x * 8).type())
print("(x * 8 + 4).type()", (x * 8 + 4).type())
print("(x * 8 * 4).type()", (x * 8 * 4).type())
print("(x * 8 / 4).type()", (x * 8 / 4).type())
print("((x * 8) * 4).type()", ((x * 8) * 4).type())
print("(x * (8 * 4)).type()", (x * (8 * 4)).type())
assert (x * 8).type() == hl.Int(32)
assert (x * 8 * 4).type() == hl.Int(32) # yes this did fail at some point
assert ((x * 8) / 4).type() == hl.Int(32)
assert (x * (8 / 4)).type() == hl.Float(32) # under python3 division rules
assert (x * (8 // 4)).type() == hl.Int(32)
#assert (x * 8 // 4).type() == hl.Int(32) # not yet implemented
# Construct the bilateral grid
r = hl.RDom(0, s_sigma, 0, s_sigma, 'r')
val0 = clamped[x * s_sigma, y * s_sigma]
val00 = clamped[x * s_sigma * hl.cast(hl.Int(32), 1),
y * s_sigma * hl.cast(hl.Int(32), 1)]
#val1 = clamped[x * s_sigma - s_sigma/2, y * s_sigma - s_sigma/2] # should fail
val22 = clamped[x * s_sigma - hl.cast(hl.Int(32), s_sigma // 2),
y * s_sigma - hl.cast(hl.Int(32), s_sigma // 2)]
val2 = clamped[x * s_sigma - s_sigma // 2, y * s_sigma - s_sigma // 2]
val3 = clamped[x * s_sigma + r.x - s_sigma // 2,
y * s_sigma + r.y - s_sigma // 2]
return
def test_imageparam_bug():
"see https://github.com/rodrigob/Halide/issues/2"
x = hl.Var("x")
y = hl.Var("y")
fx = hl.Func("fx")
input = hl.ImageParam(hl.UInt(8), 1, "input")
fx[x, y] = input[y]
return
def main():
input_img = hl.ImageParam(hl.UInt(16), 3, 'input')
# number of intensity levels
levels = hl.Param(int_t, 'levels', 8)
# Parameters controlling the filter
alpha = hl.Param(float_t, 'alpha', 1.0 / 7.0)
beta = hl.Param(float_t, 'beta', 1.0)
local_laplacian = get_local_laplacian(input_img, levels, alpha, beta)
filter_test_image(local_laplacian, input_img)
def setup_inputs(self):
# input scalars
delo2 = hl.Param(hl.Float(64), "delo2")
delta = hl.Param(hl.Float(64), "delta")
rdelta = hl.Param(hl.Float(64), "rdelta")
# input vectors
expnt_in = hl.ImageParam(hl.Float(64), 1, "expnt_in")
rnorm_in = hl.ImageParam(hl.Float(64), 1, "rnorm_in")
x_in = hl.ImageParam(hl.Float(64), 1, "x_in")
y_in = hl.ImageParam(hl.Float(64), 1, "y_in")
z_in = hl.ImageParam(hl.Float(64), 1, "z_in")
# input matrices
fm_in = hl.ImageParam(hl.Float(64), 2, "fm_in")
g_fock_in_in = hl.ImageParam(hl.Float(64), 2, "g_fock_in")
g_dens_in = hl.ImageParam(hl.Float(64), 2, "g_dens_in")
self.inputs.update({
x.name(): x
for x in [
delo2, delta, rdelta, expnt_in, rnorm_in, x_in, y_in, z_in,
fm_in, g_fock_in_in, g_dens_in
]
})
# clamp all inputs, to prevent out-of-bounds errors from odd tile sizes and such
expnt = hl.BoundaryConditions.constant_exterior(expnt_in, 0)
rnorm = hl.BoundaryConditions.constant_exterior(rnorm_in, 0)
x = hl.BoundaryConditions.constant_exterior(x_in, 0)
y = hl.BoundaryConditions.constant_exterior(y_in, 0)
z = hl.BoundaryConditions.constant_exterior(z_in, 0)
fm = hl.BoundaryConditions.constant_exterior(fm_in, 0)
g_fock_in = hl.BoundaryConditions.constant_exterior(g_fock_in_in, 0)
g_dens = hl.BoundaryConditions.constant_exterior(g_dens_in, 0)
self.clamps.update({
"expnt": expnt,
"rnorm": rnorm,
"x": x,
"y": y,
"z": z,
"fm": fm,
"g_fock_in_clamped": g_fock_in,
"g_dens": g_dens
})
# nbfn=number of basis functions. This is our problem size
self.nbfn = g_fock_in_in.height()
def test_basics():
input = hl.ImageParam(hl.UInt(16), 2, 'input')
x, y = hl.Var('x'), hl.Var('y')
blur_x = hl.Func('blur_x')
blur_xx = hl.Func('blur_xx')
blur_y = hl.Func('blur_y')
yy = hl.cast(hl.Int(32), 1)
assert yy.type() == hl.Int(32)
print("yy type:", yy.type())
z = x + 1
input[x,y]
input[0,0]
input[z,y]
input[x+1,y]
print("ping 0.2")
input[x,y]+input[x+1,y]
if False:
aa = blur_x[x,y]
bb = blur_x[x,y+1]
aa + bb
blur_x[x,y]+blur_x[x,y+1]
print("ping 0.3")
(input[x,y]+input[x+1,y]) / 2
print("ping 0.4")
blur_x[x,y]
print("ping 0.4.1")
blur_xx[x,y] = input[x,y]
print("ping 0.5")
blur_x[x,y] = (input[x,y]+input[x+1,y]+input[x+2,y])/3
print("ping 1")
blur_y[x,y] = (blur_x[x,y]+blur_x[x,y+1]+blur_x[x,y+2])/3
xi, yi = hl.Var('xi'), hl.Var('yi')
print("ping 2")
blur_y.tile(x, y, xi, yi, 8, 4).parallel(y).vectorize(xi, 8)
blur_x.compute_at(blur_y, x).vectorize(x, 8)
blur_y.compile_jit()
print("Compiled to jit")
return
def main():
# We'll define a simple one-stage pipeline:
brighter = hl.Func("brighter")
x, y = hl.Var("x"), hl.Var("y")
# The pipeline will depend on one scalar parameter.
offset = hl.Param(hl.UInt(8), name="offset")
# And take one grayscale 8-bit input buffer. The first
# constructor argument gives the type of a pixel, and the second
# specifies the number of dimensions (not the number of
# channels!). For a grayscale image this is two for a color
# image it's three. Currently, four dimensions is the maximum for
# inputs and outputs.
input = hl.ImageParam(hl.UInt(8), 2)
# If we were jit-compiling, these would just be an int and a
# hl.Buffer, but because we want to compile the pipeline once and
# have it work for any value of the parameter, we need to make a
# hl.Param object, which can be used like an hl.Expr, and an hl.ImageParam
# object, which can be used like a hl.Buffer.
# Define the hl.Func.
brighter[x, y] = input[x, y] + offset
# Schedule it.
brighter.vectorize(x, 16).parallel(y)
# This time, instead of calling brighter.realize(...), which
# would compile and run the pipeline immediately, we'll call a
# method that compiles the pipeline to an object file and header.
#
# For AOT-compiled code, we need to explicitly declare the
# arguments to the routine. This routine takes two. Arguments are
# usually Params or ImageParams.
fname = "lesson_10_halide"
brighter.compile_to(
{
hl.Output.object: "lesson_10_halide.o",
hl.Output.c_header: "lesson_10_halide.h",
hl.Output.python_extension: "lesson_10_halide.py.cpp"
}, [input, offset], "lesson_10_halide")
print("Halide pipeline compiled, but not yet run.")
# To continue this lesson, look in the file lesson_10_aot_compilation_run.cpp
return 0
def test_multipass_constraints():
input = hl.ImageParam(hl.Float(32), 2, "input")
f = hl.Func("f")
x = hl.Var("x")
y = hl.Var("y")
f[x, y] = input[x + 1, y + 1] + input[x - 1, y - 1]
f[x, y] += 3.0
f.update().vectorize(x, 4)
o = f.output_buffer()
# Now make some hard-to-resolve constraints
input.dim(0).set_bounds(min=input.dim(1).min() - 5,
extent=input.dim(1).extent() + o.dim(0).extent())
o.dim(0).set_bounds(min=0,
extent=hl.select(
o.dim(0).extent() < 22,
o.dim(0).extent() + 1,
o.dim(0).extent()))
# Make a bounds query buffer
query_buf = hl.Buffer.make_bounds_query(type=hl.Float(32), sizes=[7, 8])
query_buf.set_min([2, 2])
f.infer_input_bounds(query_buf)
if input.get().dim(0).min() != -4 or \
input.get().dim(0).extent() != 34 or \
input.get().dim(1).min() != 1 or \
input.get().dim(1).extent() != 10 or \
query_buf.dim(0).min() != 0 or \
query_buf.dim(0).extent() != 24 or \
query_buf.dim(1).min() != 2 or \
query_buf.dim(1).extent() != 8:
print("Constraints not correctly satisfied:\n", "in:",
input.get().dim(0).min(),
input.get().dim(0).extent(),
input.get().dim(1).min(),
input.get().dim(1).extent(), "out:",
query_buf.dim(0).min(),
query_buf.dim(0).extent(),
query_buf.dim(1).min(),
query_buf.dim(1).extent())
assert False
def test_scalar_funcs():
input = hl.ImageParam(hl.UInt(16), 0, 'input')
f = hl.Func('f')
g = hl.Func('g')
input[()]
(input[()] + input[()]) / 2
f[()]
g[()]
f[()] = (input[()] + input[()] + input[()]) / 3
g[()] = (f[()] + f[()] + f[()]) / 3
g.compile_jit()
def main():
input = hl.ImageParam(hl.Float(32), 2, 'input')
r_sigma = hl.Param(hl.Float(32), 'r_sigma', 0.1) # Value needed if not generating an executable
s_sigma = 8 # This is passed during code generation in the C++ version
bilateral_grid = get_bilateral_grid(input, r_sigma, s_sigma)
# Set `generate` to False to run the jit immediately and get instant gratification.
#generate = True
generate = False
if generate:
generate_compiled_file(bilateral_grid)
else:
filter_test_image(bilateral_grid, input)
print("\nEnd of game. Have a nice day!")
return
def main():
input = hl.ImageParam(hl.UInt(16), 3, 'input')
# number of intensity levels
levels = hl.Param(int_t, 'levels', 8)
#Parameters controlling the filter
alpha = hl.Param(float_t, 'alpha', 1.0 / 7.0)
beta = hl.Param(float_t, 'beta', 1.0)
local_laplacian = get_local_laplacian(input, levels, alpha, beta)
generate = False # Set to False to run the jit immediately and get instant gratification.
if generate:
generate_compiled_file(local_laplacian)
else:
filter_test_image(local_laplacian, input)
return
def main():
input = hl.ImageParam(float_t, 3, "input")
levels = 10
interpolate = get_interpolate(input, levels)
# preparing input and output memory buffers (numpy ndarrays)
input_data = get_input_data()
assert input_data.shape[2] == 4
input_image = hl.Buffer(input_data)
input.set(input_image)
input_width, input_height = input_data.shape[:2]
t0 = datetime.now()
output_image = interpolate.realize(input_width, input_height, 3)
t1 = datetime.now()
elapsed = (t1 - t0).total_seconds()
print('Interpolated in {:.5f} secs'.format(elapsed))
output_data = np.asanyarray(output_image)
# convert output
input_data = (input_data * 255).astype(np.uint8)
output_data = (output_data * 255).astype(np.uint8)
# save results
input_path = "interpolate_input.png"
output_path = "interpolate_result.png"
imageio.imsave(input_path, input_data)
imageio.imsave(output_path, output_data)
print()
print(
'blur realized on output image. Result saved at {} (input data copy at {})'
.format(output_path, input_path))
print()
print("End of game. Have a nice day!")
import halide as hl
import imageio
import numpy as np
# Constructing Halide functions statically.
input = hl.ImageParam(hl.Float(32), 3)
f = hl.Func('f')
x, y, c = hl.Var('x'), hl.Var('y'), hl.Var('c')
# Double the values and clamp them by 1.
f[x, y, c] = hl.min(2 * input[x, y, c], 1.0)
# Actually compiling/executing the Halide functions.
#
# Setup the input by loading an image (Halide assumes Fortran ordering).
input_buffer = hl.Buffer(
np.asfortranarray(
imageio.imread('images/rgb.png').astype(np.float32) / 255.0))
input.set(input_buffer)
# Process the input by calling f.realize
output = f.realize(input_buffer.width(), input_buffer.height(),
input_buffer.channels())
# Save the image to a file by converting to a numpy array.
output = np.array(output)
imageio.imsave('output.png', (output * 255.0).astype(np.uint8))
def focus_stack_pipeline():
outputs = []
start_w, start_h = 3000, 2000
number_of_layers = 5
layer_sizes = [[start_w, start_h]]
for i in range(0, number_of_layers):
# Grab from prev layer
w,h = layer_sizes[-1]
layer_sizes.append([int(math.ceil(w/2.0)),int(math.ceil(h/2.0))])
# Add last size in once more to get the 2nd top lap layer (gaussian) for
# the energy/deviation split.
layer_sizes.append(layer_sizes[-1])
input = hl.ImageParam(hl.UInt(8), 3)
input.dim(0).set_estimate(0, start_w)
input.dim(1).set_estimate(0, start_h)
input.dim(2).set_estimate(0, 3)
lap_inputs = []
max_energy_inputs = []
for i in range(0,number_of_layers+1):
lap_layer = hl.ImageParam(hl.Float(32), 3, "lap{}".format(i))
lap_inputs.append(lap_layer)
w,h = layer_sizes[i]
lap_layer.dim(0).set_estimate(0, w)
lap_layer.dim(1).set_estimate(0, h)
lap_layer.dim(2).set_estimate(0, 3)
if i == number_of_layers:
# last (top - small) layer
# Add the last laplacian (really direct from gaussian) layer
# in twice. We output one maxed on entropies and one maxed on
# deviations.
lap_layer = hl.ImageParam(hl.Float(32), 3, "lap{}".format(i+1))
lap_inputs.append(lap_layer)
lap_layer.dim(0).set_estimate(0, w)
lap_layer.dim(1).set_estimate(0, h)
lap_layer.dim(2).set_estimate(0, 3)
entropy_layer = hl.ImageParam(hl.Float(32), 2, "entroy{}".format(i))
max_energy_inputs.append(entropy_layer)
entropy_layer.dim(0).set_estimate(0, w)
entropy_layer.dim(1).set_estimate(0, h)
deviation_layer = hl.ImageParam(hl.Float(32), 2, "deviation{}".format(i))
max_energy_inputs.append(deviation_layer)
deviation_layer.dim(0).set_estimate(0, w)
deviation_layer.dim(1).set_estimate(0, h)
else:
max_energy_layer = hl.ImageParam(hl.Float(32), 2, "max_energy{}".format(i))
max_energy_inputs.append(max_energy_layer)
max_energy_layer.dim(0).set_estimate(0, w)
max_energy_layer.dim(1).set_estimate(0, h)
x, y, c = hl.Var("x"), hl.Var("y"), hl.Var("c")
hist_index = hl.Var('hist_index')
clamped = f32(x, y, c, mirror(input, 3000, 2000))
f = hl.Func("input32")
f[x, y, c] = clamped[x, y, c]
energy_outputs = []
gaussian_layers = [f]
laplacian_layers = []
merged_laps = []
for layer_num in range(0, number_of_layers):
# Add the layer size in also
w,h = layer_sizes[layer_num]
start_layer = gaussian_layers[-1]
# Blur the image
gaussian_layer = gaussian(x, y, c, start_layer)
# Grab next layer size
# w,h = layer_sizes[layer_num+1]
# Reduce the layer size and add it into the list
next_layer = reduce_layer(x, y, c, gaussian_layer)
gaussian_layers.append(next_layer)
# Expand back up
expanded = expand_layer(x, y, c, next_layer)
# Generate the laplacian from the
# original - blurred/reduced/expanded version
laplacian_layer = laplacian(x, y, c, start_layer, expanded)
laplacian_layers.append(laplacian_layer)
# Calculate energies for the gaussian layer
prev_energies = mirror(max_energy_inputs[layer_num], w, h)
next_energies = region_energy(x, y, c, laplacian_layer)
prev_laplacian = mirror(lap_inputs[layer_num], w, h)
merged_energies = energy_maxes(x, y, c, prev_energies, next_energies)
merged_lap = merge_laplacian(x, y, c, merged_energies, next_energies, prev_laplacian, laplacian_layer)
energy_outputs.append([[w,h,True],merged_energies])
merged_laps.append(merged_lap)
# Add estimates
next_layer.set_estimate(x, 0, w)
next_layer.set_estimate(y, 0, h)
next_layer.set_estimate(c, 0, 3)
# Handle last layer differently
w,h = layer_sizes[-1]
# The next_lap is really just the last gaussian layer
next_lap = gaussian_layers[-1]
prev_entropy_laplacian = mirror(lap_inputs[-2], w, h)
prev_entropy = mirror(max_energy_inputs[-2], w, h)
next_entropy = entropy(x, y, c, next_lap, w, h, hist_index)
merged_entropy = energy_maxes(x, y, c, prev_entropy, next_entropy)
merged_lap_on_entropy = merge_laplacian(x, y, c, merged_entropy, next_entropy, prev_entropy_laplacian, next_lap)
merged_laps.append(merged_lap_on_entropy)
prev_deviation_laplacian = mirror(lap_inputs[-1], w, h)
prev_deviation = mirror(max_energy_inputs[-1], w, h)
next_deviation = deviation(x, y, c, next_lap)
merged_deviation = energy_maxes(x, y, c, prev_deviation, next_deviation)
merged_lap_on_deviation = merge_laplacian(x, y, c, merged_deviation, next_deviation, prev_deviation_laplacian, next_lap)
merged_laps.append(merged_lap_on_deviation)
energy_outputs.append([[w,h,True],merged_entropy])
energy_outputs.append([[w,h,True],merged_deviation])
print("NUM LAYERS: ", len(gaussian_layers), len(laplacian_layers), layer_sizes)
# Add all of the laplacian layers to the output first
i = 0
for merged_lap in merged_laps:
w,h = layer_sizes[i]
mid = (i < (len(merged_laps) - 2))
outputs.append([[w,h,False,mid], merged_lap])
i += 1
# Then energies
for energy_output in energy_outputs:
outputs.append(energy_output)
new_outputs = []
for size, output in outputs:
w = size[0]
h = size[1]
gray = len(size) > 2 and size[2]
mid = len(size) > 3 and size[3]
if mid:
uint8_output = output
else:
uint8_output = output
uint8_output.set_estimate(x, 0, w)
uint8_output.set_estimate(y, 0, h)
if not gray:
uint8_output.set_estimate(c, 0, 3)
new_outputs.append([size, uint8_output])
outputs = new_outputs
print("OUTPUT LAYERS: ")
pprint(outputs)
output_funcs = [output for _, output in outputs]
pipeline = hl.Pipeline(output_funcs)
return {
'pipeline': pipeline,
'inputs': [input] + lap_inputs + max_energy_inputs
}
def test_extern():
"""
Shows an example of Halide calling a C library loaded
in the Python process via ctypes
"""
# Requires Makefile support to build the external function in linkable form
print("TODO: test_extern not yet implemented in Python; skipping...")
return 0
x = hl.Var("x")
data = np.random.random(10).astype(np.float64)
expected_result = np.sort(data)
output_data = np.empty(10, dtype=np.float64)
sort_func = hl.Func("extern_sort_func")
# gsl_sort,
# see http://www.gnu.org/software/gsl/manual/html_node/Sorting-vectors.html#Sorting-vectors
input = hl.ImageParam(hl.Float(64), 1, "input_data")
extern_name = "the_sort_func"
params = [hl.ExternFuncArgument(input)]
output_types = [hl.Int(32)]
dimensionality = 1
sort_func.define_extern(extern_name, params, output_types, dimensionality)
try:
sort_func.compile_jit()
except RuntimeError:
pass
else:
raise Exception(
"compile_jit should have raised a 'Symbol not found' RuntimeError")
import ctypes
sort_lib = ctypes.CDLL("the_sort_function.so")
print(sort_lib.the_sort_func)
try:
sort_func.compile_jit()
except RuntimeError:
print("ctypes CDLL did not work out")
else:
print("ctypes CDLL worked !")
lib_path = "the_sort_function.so"
#lib_path = "/home/rodrigob/code/references/" \
# "Halide_master/python_bindings/tests/the_sort_function.nohere.so"
load_error = load_library_into_llvm(lib_path)
assert load_error == False
sort_func.compile_jit()
# now that things are loaded, we try to call them
input.set(data)
sort_func.realize(output_data)
assert np.isclose(expected_result, output_data)
return
def main():
# We'll define the simple one-stage pipeline that we used in lesson 10.
brighter = hl.Func("brighter")
x, y = hl.Var("x"), hl.Var("y")
# Declare the arguments.
offset = hl.Param(hl.UInt(8))
input = hl.ImageParam(hl.UInt(8), 2)
args = [input, offset]
# Define the hl.Func.
brighter[x, y] = input[x, y] + offset
# Schedule it.
brighter.vectorize(x, 16).parallel(y)
# The following line is what we did in lesson 10. It compiles an
# object file suitable for the system that you're running this
# program on. For example, if you compile and run this file on
# 64-bit linux on an x86 cpu with sse4.1, then the generated code
# will be suitable for 64-bit linux on x86 with sse4.1.
brighter.compile_to_file("lesson_11_host", args, "lesson_11_host")
# We can also compile object files suitable for other cpus and
# operating systems. You do this with an optional third argument
# to compile_to_file which specifies the target to compile for.
create_android = True
create_windows = True
create_ios = True
if create_android:
# Let's use this to compile a 32-bit arm android version of this code:
target = hl.Target()
target.os = hl.TargetOS.Android # The operating system
target.arch = hl.TargetArch.ARM # The CPU architecture
target.bits = 32 # The bit-width of the architecture
arm_features = [] # A list of features to set
target.set_features(arm_features)
# Pass the target as the last argument.
brighter.compile_to_file("lesson_11_arm_32_android", args,
"lesson_11_arm_32_android", target)
if create_windows:
# And now a Windows object file for 64-bit x86 with AVX and SSE 4.1:
target = hl.Target()
target.os = hl.TargetOS.Windows
target.arch = hl.TargetArch.X86
target.bits = 64
target.set_features([hl.TargetFeature.AVX, hl.TargetFeature.SSE41])
brighter.compile_to_file("lesson_11_x86_64_windows", args,
"lesson_11_x86_64_windows", target)
if create_ios:
# And finally an iOS mach-o object file for one of Apple's 32-bit
# ARM processors - the A6. It's used in the iPhone 5. The A6 uses
# a slightly modified ARM architecture called ARMv7s. We specify
# this using the target features field. Support for Apple's
# 64-bit ARM processors is very new in llvm, and still somewhat
# flaky.
target = hl.Target()
target.os = hl.TargetOS.IOS
target.arch = hl.TargetArch.ARM
target.bits = 32
target.set_features([hl.TargetFeature.ARMv7s])
brighter.compile_to_file("lesson_11_arm_32_ios", args,
"lesson_11_arm_32_ios", target)
# Now let's check these files are what they claim, by examining
# their first few bytes.
if create_android:
# 32-arm android object files start with the magic bytes:
# uint8_t []
arm_32_android_magic = [
0x7f,
ord('E'),
ord('L'),
ord('F'), # ELF format
1, # 32-bit
1, # 2's complement little-endian
1
] # Current version of elf
length = len(arm_32_android_magic)
f = open("lesson_11_arm_32_android.o", "rb")
try:
header_bytes = f.read(length)
except:
print("Android object file not generated")
return -1
f.close()
header = list(unpack("B" * length, header_bytes))
if header != arm_32_android_magic:
print([x == y for x, y in zip(header, arm_32_android_magic)])
raise Exception(
"Unexpected header bytes in 32-bit arm object file.")
return -1
if create_windows:
# 64-bit windows object files start with the magic 16-bit value 0x8664
# (presumably referring to x86-64)
# uint8_t []
win_64_magic = [0x64, 0x86]
f = open("lesson_11_x86_64_windows.obj", "rb")
try:
header_bytes = f.read(2)
except:
print("Windows object file not generated")
return -1
f.close()
header = list(unpack("B" * 2, header_bytes))
if header != win_64_magic:
raise Exception(
"Unexpected header bytes in 64-bit windows object file.")
return -1
if create_ios:
# 32-bit arm iOS mach-o files start with the following magic bytes:
# uint32_t []
arm_32_ios_magic = [
0xfeedface, # Mach-o magic bytes
#0xfe, 0xed, 0xfa, 0xce, # Mach-o magic bytes
12, # CPU type is ARM
11, # CPU subtype is ARMv7s
1
] # It's a relocatable object file.
f = open("lesson_11_arm_32_ios.o", "rb")
try:
header_bytes = f.read(4 * 4)
except:
print("ios object file not generated")
return -1
f.close()
header = list(unpack("I" * 4, header_bytes))
if header != arm_32_ios_magic:
raise Exception(
"Unexpected header bytes in 32-bit arm ios object file.")
return -1
# It looks like the object files we produced are plausible for
# those targets. We'll count that as a success for the purposes
# of this tutorial. For a real application you'd then need to
# figure out how to integrate Halide into your cross-compilation
# toolchain. There are several small examples of this in the
# Halide repository under the apps folder. See HelloAndroid and
# HelloiOS here:
# https:#github.com/halide/Halide/tree/master/apps/
print("Success!")
return 0
def test_complexstub():
constant_image = _make_constant_image()
input = hl.ImageParam(hl.UInt(8), 3, 'input')
input.set(constant_image)
x, y, c = hl.Var(), hl.Var(), hl.Var()
target = hl.get_jit_target_from_environment()
float_arg = 1.25
int_arg = 33
r = complexstub(target,
typed_buffer_input=constant_image,
untyped_buffer_input=constant_image,
simple_input=input,
array_input=[input, input],
float_arg=float_arg,
int_arg=[int_arg, int_arg],
untyped_buffer_output_type="uint8",
vectorize=True)
# return value is a tuple; unpack separately to avoid
# making the callsite above unreadable
(simple_output, tuple_output, array_output, typed_buffer_output,
untyped_buffer_output, static_compiled_buffer_output) = r
b = simple_output.realize(32, 32, 3, target)
assert b.type() == hl.Float(32)
for x in range(32):
for y in range(32):
for c in range(3):
expected = constant_image[x, y, c]
actual = b[x, y, c]
assert expected == actual, "Expected %s Actual %s" % (expected,
actual)
b = tuple_output.realize(32, 32, 3, target)
assert b[0].type() == hl.Float(32)
assert b[1].type() == hl.Float(32)
assert len(b) == 2
for x in range(32):
for y in range(32):
for c in range(3):
expected1 = constant_image[x, y, c] * float_arg
expected2 = expected1 + int_arg
actual1, actual2 = b[0][x, y, c], b[1][x, y, c]
assert expected1 == actual1, "Expected1 %s Actual1 %s" % (
expected1, actual1)
assert expected2 == actual2, "Expected2 %s Actual1 %s" % (
expected2, actual2)
assert len(array_output) == 2
for a in array_output:
b = a.realize(32, 32, target)
assert b.type() == hl.Int(16)
for x in range(32):
for y in range(32):
expected = constant_image[x, y, 0] + int_arg
actual = b[x, y]
assert expected == actual, "Expected %s Actual %s" % (expected,
actual)
# TODO: Output<Buffer<>> has additional behaviors useful when a Stub
# is used within another Generator; this isn't yet implemented since there
# isn't yet Python bindings for Generator authoring. This section
# of the test may need revision at that point.
b = typed_buffer_output.realize(32, 32, 3, target)
assert b.type() == hl.Float(32)
for x in range(32):
for y in range(32):
for c in range(3):
expected = constant_image[x, y, c]
actual = b[x, y, c]
assert expected == actual, "Expected %s Actual %s" % (expected,
actual)
b = untyped_buffer_output.realize(32, 32, 3, target)
assert b.type() == hl.UInt(8)
for x in range(32):
for y in range(32):
for c in range(3):
expected = constant_image[x, y, c]
actual = b[x, y, c]
assert expected == actual, "Expected %s Actual %s" % (expected,
actual)
b = static_compiled_buffer_output.realize(4, 4, 1, target)
assert b.type() == hl.UInt(8)
for x in range(4):
for y in range(4):
for c in range(1):
expected = constant_image[x, y, c] + 42
actual = b[x, y, c]
assert expected == actual, "Expected %s Actual %s" % (expected,
actual)
def gen_g(self):
''' define g() function '''
# vars
i, j, k, l = [self.vars[c] for c in "ijkl"]
# clamped inputs
x, y, z, expnt, fm, rnorm = [
self.clamps[c] for c in ["x", "y", "z", "expnt", "fm", "rnorm"]
]
# unclamped input (for sizing)
fm_in = self.inputs["fm_in"]
# scalar inputs
delo2, delta, rdelta = [
self.inputs[c] for c in ["delo2", "delta", "rdelta"]
]
dx = hl.Func("dx")
dy = hl.Func("dy")
dz = hl.Func("dz")
r2 = hl.Func("g_r2")
expnt2 = hl.Func("expnt2")
expnt_inv = hl.Func("expnt_inv")
self.add_funcs_by_name([dx, dy, dz, r2, expnt2, expnt_inv])
dx[i, j] = x[i] - x[j]
dy[i, j] = y[i] - y[j]
dz[i, j] = z[i] - z[j]
r2[i,
j] = dx[i, j] * dx[i, j] + dy[i, j] * dy[i, j] + dz[i, j] * dz[i, j]
expnt2[i, j] = expnt[i] + expnt[j]
expnt_inv[i, j] = hl.f64(1.0) / expnt2[i, j]
fac2 = hl.Func("fac2")
ex_arg = hl.Func("ex_arg")
ex = hl.Func("ex")
denom = hl.Func("denom")
fac4d = hl.Func("fac4d")
self.add_funcs_by_name([fac2, ex_arg, ex, denom, fac4d])
fac2[i, j] = expnt[i] * expnt[j] * expnt_inv[i, j]
ex_arg[i, j, k, l] = -fac2[i, j] * r2[i, j] - fac2[k, l] * r2[k, l]
ex[i, j, k, l] = hl.select(ex_arg[i, j, k, l] < hl.f64(-37.0),
hl.f64(0.0), hl.exp(ex_arg[i, j, k, l]))
denom[i, j, k,
l] = expnt2[i, j] * expnt2[k, l] * hl.sqrt(expnt2[i, j] +
expnt2[k, l])
fac4d[i, j, k,
l] = expnt2[i, j] * expnt2[k, l] / (expnt2[i, j] + expnt2[k, l])
x2 = hl.Func("g_x2")
y2 = hl.Func("g_y2")
z2 = hl.Func("g_z2")
rpq2 = hl.Func("rpq2")
self.add_funcs_by_name([x2, y2, z2, rpq2])
x2[i, j] = (x[i] * expnt[i] + x[j] * expnt[j]) * expnt_inv[i, j]
y2[i, j] = (y[i] * expnt[i] + y[j] * expnt[j]) * expnt_inv[i, j]
z2[i, j] = (z[i] * expnt[i] + z[j] * expnt[j]) * expnt_inv[i, j]
rpq2[i, j, k, l] = ((x2[i, j] - x2[k, l]) * (x2[i, j] - x2[k, l]) +
(y2[i, j] - y2[k, l]) * (y2[i, j] - y2[k, l]) +
(z2[i, j] - z2[k, l]) * (z2[i, j] - z2[k, l]))
f0t = hl.Func("f0t")
f0n = hl.Func("f0n")
f0x = hl.Func("f0x")
f0val = hl.Func("f0val")
self.add_funcs_by_name([f0t, f0n, f0x, f0val])
f0t[i, j, k, l] = fac4d[i, j, k, l] * rpq2[i, j, k, l]
f0n[i, j, k, l] = hl.clamp(hl.i32((f0t[i, j, k, l] + delo2) * rdelta),
fm_in.dim(0).min(),
fm_in.dim(0).max())
f0x[i, j, k, l] = delta * f0n[i, j, k, l] - f0t[i, j, k, l]
f0val[i, j, k, l] = hl.select(
f0t[i, j, k, l] >= hl.f64(28.0),
hl.f64(0.88622692545276) / hl.sqrt(f0t[i, j, k, l]),
fm[f0n[i, j, k, l], 0] + f0x[i, j, k, l] *
(fm[f0n[i, j, k, l], 1] + f0x[i, j, k, l] * hl.f64(0.5) *
(fm[f0n[i, j, k, l], 2] + f0x[i, j, k, l] * hl.f64(1. / 3.) *
(fm[f0n[i, j, k, l], 3] +
f0x[i, j, k, l] * hl.f64(0.25) * fm[f0n[i, j, k, l], 4]))))
g = hl.Func("g")
self.add_funcs_by_name([g])
if self.tracing and self.tracing_g:
g_trace_in = hl.ImageParam(hl.Float(64), 4, "g_trace_in")
g_trace = hl.BoundaryConditions.constant_exterior(g_trace_in, 0)
self.inputs["g_trace_in"] = g_trace_in
self.clamps["g_trace"] = g_trace
g_trace.compute_root()
g[i, j, k,
l] = (hl.f64(2.00) * hl.f64(pow(pi, 2.50)) / denom[i, j, k, l]
) * ex[i, j, k, l] * f0val[i, j, k, l] * rnorm[i] * rnorm[
j] * rnorm[k] * rnorm[l] + g_trace[i, j, k, l]
else:
g_trace = None
g[i, j, k,
l] = (hl.f64(2.00) * hl.f64(pow(pi, 2.50)) /
denom[i, j, k, l]) * ex[i, j, k, l] * f0val[
i, j, k, l] * rnorm[i] * rnorm[j] * rnorm[k] * rnorm[l]