def test_full_integral_image_correctness():
''' Test generated full integral image correctness,
note that this relies on the corectness of interpreter and reference.py '''
block_size = (20, 20)
size = tuple(x*3 for x in block_size)
# generate random test image
test_image = [[float(random.randint(0, 255)) for i in xrange(size[0])] for j in xrange(size[1])]
# reference implementation
integral_ref = reference.gen_integral_image(test_image)
sq_integral_ref = reference.gen_integral_squared_image(test_image)
# pointer config
buffer_size = block_size[0]*block_size[1]
src_ptr = 0
integral_ptr = buffer_size
sq_integral_ptr = 2*buffer_size
# set up interpreter for integral image calculation
pe_dim = [s//b for s, b in zip(size, block_size)]
def code_gen(code, block_size, args):
return gen_code.gen_full_integral_image(code, src_ptr, integral_ptr, sq_integral_ptr, pe_dim, block_size)
code = Code()
code.set_generator(optimiser_wrapper(code_gen), block_size)
sim = Interpreter(code, test_image, block_size)
sim.run()
# get result of simulator with scaling, truncation turned off and float output
integral_test = sim.gen_output_image(1, False, False, True)
sq_integral_test = sim.gen_output_image(2, False, False, True)
# comparison of reference with blip sim
integral_err = compare_images(integral_ref, integral_test)
sq_integral_err = compare_images(sq_integral_ref, sq_integral_test)
err_eps = 0.001
if not ((integral_err < err_eps) and (sq_integral_err < err_eps)):
print 'integral comp:', integral_err
print 'squared integral comp:', sq_integral_err
print 'rendering instruction stream to file, can take a while'
try:
f = open('unoptimised_full_integral_image_trace.txt', 'w')
def tag_str(instr): return ', '.join(instr.tag) if hasattr(instr, 'tag') else ''
f.write('\n'.join(str(x).ljust(40) + ' tags: ' + tag_str(x) for x in code_gen(Code())))
f.close()
optim_gen = optimiser_wrapper(code_gen, block_size, {})
f = open('bad_full_integral_image_trace.txt', 'w')
def tag_str(instr): return ', '.join(instr.tag) if hasattr(instr, 'tag') else ''
f.write('\n'.join(str(x).ljust(40) + ' tags: ' + tag_str(x) for x in optim_gen(Code())))
f.close()
except Exception, e:
print 'could render instruction stream to file'
print 'err: ' + str(e)
assert False
def test_map_pixel_to_pixel():
def pixel_op(code, pixel_in, pixel_out, args, block_size):
th = args['th']
with scoped_alloc(code, 3) as (th_r, v, const_255):
yield Imm(th_r, th)
yield Cmp(pixel_in, th_r)
yield Imm(const_255, 255)
yield Mov(pixel_out, const_255, cond='GT')
yield Xor(pixel_out, pixel_out, pixel_out, cond='LE')
def run_test(image, th):
code = Code()
block_size = (16, 16)
in_ptr = 0
out_ptr = block_size[0]*block_size[1]
args = {'th': th}
def codegen(code, block_size, args):
return map_pixel_to_pixel(code, in_ptr, out_ptr, pixel_op, args, block_size)
code.set_generator(codegen, block_size, args)
sim = Interpreter(code, image, block_size)
sim.run()
return sim.gen_output_image(1)
def run_ref(image, th):
return [[255 if x > th else 0 for x in y] for y in image]
th = 100
image = [[random.randint(0, 255) for x in xrange(16)] for y in xrange(16)]
res_test = run_test(image, th)
res_ref = run_test(image, th)
assert compare_images(res_test, res_ref) < 0.01
def _test_image_function_single_pe(codegen, args, ref_implementation):
# setttings
width, height = 9, 9 # size only effects execution speed so keep it small
block_size = (width, height)
def run_test(image, args, block_size):
im_size = len(image[0]), len(image)
bwidth, bheight = block_size
assert(im_size == block_size) # only one pe
code = Code()
code.set_generator(codegen, block_size, args)
sim = Interpreter(code, image, block_size)
sim.run()
output = sim.gen_output_image(1, False)
return output
image = [[random.randint(0, 255) for x in xrange(width)] for y in xrange(height)]
ref_output = ref_implementation(image, args)
test_output = run_test(image, args, block_size)
if not (compare_images(ref_output, test_output) < 0.01):
print 'ref'
print '\n'.join(str(y) for y in ref_output)
print 'test'
print '\n'.join(str(y) for y in test_output)
return False
return True
def test_get2D_func():
''' Test if get2D codegen is correct. '''
import random
block_size = (5, 5)
out_ptr = block_size[0]*block_size[1]
im_size = tuple(x*3 for x in block_size) # only single block, this code doesn't account for interblock comm
src_image = [[random.randint(0, 255) for _ in xrange(im_size[0])] for _ in xrange(im_size[1])]
src = '''
@kernel
def main(in_ptr, out_ptr):
y = get_local_id(0)
x = get_local_id(1)
index = %(bwidth)i*y
out_ptr[index+x] = get2D(in_ptr, x-1, y-1, %(bwidth)i, %(bheight)i)
'''%{'bwidth':block_size[0], 'bheight':block_size[1]}
test_out = run_kernel(src, src_image, block_size, {'in_ptr':0, 'out_ptr':out_ptr}, 32)
ref_out = [[0 for x in y] for y in src_image]
for i, src_row in enumerate(src_image):
for j, src_px in enumerate(src_row):
x = j-1
y = i-1
if x >= 0 and x < im_size[0] and y >= 0 and y < im_size[1]:
ref_out[i][j] = src_image[y][x]
assert compare_images(test_out, ref_out) < 0.001
def gen_integral_image_correctness(): ''' test if generated integral image is correct, note that this relies on the corectness of interpreter and reference.py ''' # size = (120, 80) # block_size = (40, 40) size = (80, 80) block_size = size # generate random test image test_image = [[float(random.randint(0, 255)) for i in xrange(size[0])] for j in xrange(size[1])] # reference implementation integral_ref = reference.gen_integral_image(test_image) sq_integral_ref = reference.gen_integral_squared_image(test_image) # pointer config buffer_size = block_size[0]*block_size[1] src_ptr = 0 integral_ptr = buffer_size sq_integral_ptr = 2*buffer_size # set up interpreter for integral image calculation def code_gen(code, block_size, args): return gen_code.gen_integral_image(code, src_ptr, integral_ptr, sq_integral_ptr, block_size) code = Code() code.set_generator(optimiser_wrapper(code_gen), block_size) sim = Interpreter(code, test_image, block_size) sim.run() # get result of simulator with scaling, truncation turned off and float output integral_test = sim.gen_output_image(1, False, False, True) sq_integral_test = sim.gen_output_image(2, False, False, True) # comparison of reference with blip sim integral_err = compare_images(integral_ref, integral_test) sq_integral_err = compare_images(sq_integral_ref, sq_integral_test) err_eps = 0.001 if not ((integral_err < err_eps) and (sq_integral_err < err_eps)): print 'integral comp:', integral_err print 'squared integral comp:', sq_integral_err assert False
def test_full_integral_image_correctness(): ''' Test generated full integral image correctness, note that this relies on the correctness of interpreter and reference.py ''' block_size = (20, 20) size = tuple(x*3 for x in block_size) # generate random test image test_image = [[float(random.randint(0, 255)) for i in xrange(size[0])] for j in xrange(size[1])] # reference implementation integral_ref = reference.gen_integral_image(test_image) sq_integral_ref = reference.gen_integral_squared_image(test_image) # pointer config buffer_size = block_size[0]*block_size[1] src_ptr = 0 integral_ptr = buffer_size sq_integral_ptr = 2*buffer_size # set up interpreter for integral image calculation pe_dim = [s//b for s, b in zip(size, block_size)] code = Code() def code_gen(code, block_size, args): return gen_code.gen_full_integral_image(code, src_ptr, integral_ptr, sq_integral_ptr, pe_dim, block_size) code.set_generator(code_gen, block_size) sim = Interpreter(code, test_image, block_size) sim.run() # get result of simulator with scaling, truncation turned off and float output integral_test = sim.gen_output_image(1, False, False, True) sq_integral_test = sim.gen_output_image(2, False, False, True) # comparison of reference with blip sim integral_err = compare_images(integral_ref, integral_test) sq_integral_err = compare_images(sq_integral_ref, sq_integral_test) err_eps = 0.001 assert (integral_err < err_eps) and (sq_integral_err < err_eps)
def test_conv_3x3_multiple_block():
''' Simple 2 dimensional convolution test with multiple blocks.
'''
import random
block_size = (8, 8)
out_ptr = block_size[0]*block_size[1]
im_size = tuple(x*2 for x in block_size) # only single block, this code doesn't account for interblock comm
src_image = [[random.randint(0, 255) for _ in xrange(im_size[0])] for _ in xrange(im_size[1])]
src = '''
@kernel
def main(in_ptr, out_ptr):
coeff = [[-1, 0, 1], [-1, 0, 1], [-1, 0, 1]]
y = get_local_id(0)
x = get_local_id(1)
index = %(bwidth)i*y+x
acc = 0
current_y = y - 1
for i in range(3):
current_x = x - 1
for j in xrange(3):
current_coeff = coeff[i][j]
in_v = get2D(in_ptr, current_x, current_y, %(bwidth)i, %(bheight)i)
v = current_coeff * in_v
acc += v
current_x += 1
current_y += 1
out_ptr[index] = acc
'''%{'bwidth':block_size[0], 'bheight':block_size[1]}
test_out = run_kernel(src, src_image, block_size, {'in_ptr':0, 'out_ptr':out_ptr}, 32)
ref_out = [[0 for x in y] for y in src_image]
coeffs = [[-1, 0, 1]]*3
for i, src_row in enumerate(src_image):
for j, src_px in enumerate(src_row):
acc = 0
for k, coeff_row in enumerate(coeffs):
y = i + k - 1
for l, coeff in enumerate(coeff_row):
x = j + l -1
if x >= 0 and x < len(src_row) and y >= 0 and y < len(src_image):
acc += coeff * src_image[y][x]
ref_out[i][j] = acc
assert compare_images(test_out, ref_out) < 0.001
def test_map_image_to_pixel():
def pixel_op(code, pos, in_ptr, out_ptr, args, block_size):
''' Simple image shift implementation. '''
offset = args['offset']
x, y = pos
width, height = block_size
c_in_ptr = in_ptr + width*y + (x + offset)
c_out_ptr = out_ptr + width*y + x
with scoped_alloc(code, 1) as v:
for instr in load_mem_value(code, c_in_ptr, pos, v, block_size):
yield instr
yield MemWImm(c_out_ptr, v)
def run_test(image, offset):
code = Code()
block_size = (16, 16)
in_ptr = 0
out_ptr = block_size[0]*block_size[1]
args = {'offset' : offset}
def codegen(code, block_size, args):
return map_image_to_pixel(code, in_ptr, out_ptr, pixel_op, args, block_size)
code.set_generator(codegen, block_size, args)
sim = Interpreter(code, image, block_size)
sim.run()
return sim.gen_output_image(1)
def run_ref(image, offset):
iwidth, iheight = len(image[0]), len(image)
res = [[0 for x in xrange(iwidth)] for y in xrange(iheight)]
def in_image(j, i):
return i >= 0 and i < iheight and j >= 0 and j < iwidth
for i, row in enumerate(image):
for j, _ in enumerate(row):
x, y = j + offset, i
res = image[y][x] if in_image(x, y) else 0
return res
offset = 2
image = [[random.randint(0, 255) for x in xrange(32)] for y in xrange(32)]
res_test = run_test(image, offset)
res_ref = run_test(image, offset)
assert compare_images(res_test, res_ref) < 0.00001
def test_gray():
''' Basic test, generate a gray image. '''
block_size = (8, 8)
im_size = tuple(x*2 for x in block_size)
src_image = [[0 for _ in xrange(im_size[0])] for _ in xrange(im_size[1])]
src = '''
@kernel
def main():
y = get_local_id(0)
x = get_local_id(1)
index = y*%(bwidth)i + x
out_ptr = %(bwidth)i * %(bheight)i
out_ptr[index] = 128
'''%{'bwidth':block_size[0], 'bheight':block_size[1]}
test_out = run_kernel(src, src_image, block_size, {}, 16)
ref_out = [[128. for _ in xrange(im_size[0])] for _ in xrange(im_size[1])]
assert compare_images(test_out, ref_out) < 0.001
def test_map_neighborhood_to_pixel():
def pixel_op(code, mask_val, image_val, acc, args, block_size):
''' Simple convolution implementation. '''
with scoped_alloc(code, 2) as (v, mask_val_r):
yield Imm(mask_val_r, mask_val)
yield Mul(v, mask_val_r, image_val)
yield Add(acc, acc, v)
def run_test(image, coeff):
code = Code()
block_size = (16, 16)
in_ptr = 0
out_ptr = block_size[0]*block_size[1]
args = {}
def codegen(code, block_size, args):
return map_neighborhood_to_pixel(code, in_ptr, out_ptr, coeff, pixel_op, args, block_size)
code.set_generator(codegen, block_size, args)
sim = Interpreter(code, image, block_size)
sim.run()
return sim.gen_output_image(1)
def run_ref(image, coeff):
iwidth, iheight = len(image[0]), len(image)
res = [[0 for x in xrange(iwidth)] for y in xrange(iheight)]
for i, row in enumerate(image):
for j, v in enumerate(row):
acc = 0
for ii, c_row in enumerate(coeff):
if ii >= 0 and ii < iheight:
for jj, c in enumerate(c_row):
if jj >= 0 and jj < iwidth:
acc += c*image[ii][jj]
res[j][i] = acc
return res
coeff = [[-1, 0, 1]]*3
image = [[random.randint(0, 255) for x in xrange(32)] for y in xrange(32)]
res_test = run_test(image, coeff)
res_ref = run_test(image, coeff)
assert compare_images(res_test, res_ref) < 0.00001
def test_conv_3x1():
''' Simple 1 dimensional convolution test. '''
import random
block_size = (8, 8)
out_ptr = block_size[0]*block_size[1]
im_size = tuple(x*1 for x in block_size) # only single block, this code doesn't account for interblock comm
src_image = [[random.randint(0, 255) for _ in xrange(im_size[0])] for _ in xrange(im_size[1])]
src = '''
@kernel
def main(in_ptr, out_ptr):
coeff = [-1, 0, 1]
y = get_local_id(0)
x = get_local_id(1)
index = %(bwidth)i*y
acc = 0
current_x = x - 1
for i in range(3):
in_v = in_ptr[index+current_x]
acc += (coeff[i] * in_v if current_x < %(bwidth)i else 0) if current_x >= 0 else 0
current_x += 1
out_ptr[index+x] = acc
'''%{'bwidth':block_size[0]}
test_out = run_kernel(src, src_image, block_size, {'in_ptr':0, 'out_ptr':out_ptr}, 16)
ref_out = [[0 for x in y] for y in src_image]
coeffs = [-1, 0, 1]
for i, src_row in enumerate(src_image):
for j, src_px in enumerate(src_row):
acc = 0
for k, coeff in enumerate(coeffs):
x = j + k -1
if x >= 0 and x < len(src_row):
acc += coeff * src_row[x]
ref_out[i][j] = acc
assert compare_images(test_out, ref_out) < 0.001
def test_treshold():
import random
block_size = (8, 8)
th = 100
out_ptr = block_size[0]*block_size[1]
im_size = tuple(x*2 for x in block_size)
src_image = [[random.randint(0, 255) for _ in xrange(im_size[0])] for _ in xrange(im_size[1])]
src = '''
@kernel
def main(in_ptr, out_ptr, th):
y = get_local_id(0)
x = get_local_id(1)
index = %(bwidth)i*y + x
in_v = in_ptr[index]
out_v = 255 if in_v > th else 0
out_ptr[index] = out_v
'''%{'bwidth':block_size[0]}
test_out = run_kernel(src, src_image, block_size, {'in_ptr':0, 'out_ptr':out_ptr, 'th':th}, 16)
ref_out = [[255 if x > th else 0 for x in y] for y in src_image]
assert compare_images(test_out, ref_out) < 0.001