def tune_variance_zero_mean():
with open(get_kernel_path() + 'wienerfilter.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
size = np.int32(height * width)
tune_params = OrderedDict()
tune_params["block_size_x"] = [2**i for i in range(5, 11)]
tune_params["num_blocks"] = [2**i for i in range(5, 11)]
max_blocks = max(tune_params["num_blocks"])
output = np.zeros(max_blocks, dtype=np.float32)
args = [size, output, image]
problem_size = ("num_blocks", 1)
tune_kernel("computeVarianceZeroMean",
kernel_string,
problem_size,
args,
tune_params,
grid_div_x=[],
verbose=True)
def test_wiener():
with open(get_kernel_path() + 'wienerfilter.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
problem_size = (width, height)
output = np.zeros(problem_size, dtype=np.float32)
args = [height, width, output, image]
params = OrderedDict()
params["block_size_x"] = 32
params["block_size_y"] = 8
params["reuse_computation"] = 1
answer = run_kernel("computeVarianceEstimates",
kernel_string,
problem_size,
args,
params,
grid_div_y=["block_size_y"])
reference = run_kernel("computeVarianceEstimates_naive",
kernel_string,
problem_size,
args,
params,
grid_div_y=["block_size_y"])
assert np.allclose(answer[2], reference[2], atol=1e-6)
def tune_wiener():
with open(get_kernel_path() + 'wienerfilter.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
problem_size = (width, height)
output = np.zeros(problem_size, dtype=np.float32)
args = [height, width, output, image]
tune_params = OrderedDict()
tune_params["block_size_x"] = [32 * i for i in range(1, 33)]
tune_params["block_size_y"] = [2**i for i in range(6)]
#first the naive kernel
#tune_kernel("computeVarianceEstimates_naive", kernel_string, problem_size, args, tune_params, grid_div_y=["block_size_y"])
#more sophisticated kernel
tune_params["reuse_computation"] = [0, 1]
tune_kernel("computeVarianceEstimates",
kernel_string,
problem_size,
args,
tune_params,
grid_div_y=["block_size_y"])
def tune_fastnoise():
with open(get_kernel_path() + 'fastnoisefilter.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
problem_size = (width, height)
output = np.zeros(problem_size, dtype=np.float32)
args = [height, width, output, image]
tune_params = OrderedDict()
tune_params["block_size_x"] = [32 * i for i in range(1, 33)]
tune_params["block_size_y"] = [2**i for i in range(6)]
kernels = [
"normalized_gradient", "gradient", "convolveHorizontally",
"convolveVertically", "normalize"
]
for k in kernels:
tune_kernel(k, kernel_string, problem_size, args, tune_params)
def test_find_peak():
with open(get_kernel_path() + 'peaktocorrelationenergy.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test_small.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
problem_size = (width, height)
#generate some bogus crosscorr data
crosscorr = np.random.randn(height, width, 2).astype(np.float32)
#compute reference in Python
peak_index = np.argmax(np.absolute(crosscorr[:, :, 0]))
peak_value = np.absolute(crosscorr[:, :, 0].flatten()[peak_index])
params = {"block_size_x": 512, "num_blocks": 64}
problem_size = ("num_blocks", 1)
num_blocks = np.int32(params["num_blocks"])
peakval = np.zeros((1), dtype=np.float32)
peakvals = np.zeros((num_blocks), dtype=np.float32)
peakindx = np.zeros((num_blocks), dtype=np.int32)
loc = np.zeros((1), dtype=np.int32)
val = np.zeros((1), dtype=np.float32)
args = [height, width, peakval, peakvals, peakindx, crosscorr]
output1 = run_kernel("findPeak",
kernel_string,
problem_size,
args,
params,
grid_div_x=[])
peakvals = output1[3]
peakindx = output1[4]
args = [loc, val, peakindx, peakvals, num_blocks]
output2 = run_kernel("maxlocFloats",
kernel_string, (1, 1),
args,
params,
grid_div_x=[])
loc = output2[0][0]
val = output2[1][0]
print("answer")
print("loc=", loc, "val=", val)
print("reference")
print("loc=", peak_index, "val=", peak_value)
assert loc == peak_index
assert np.isclose(val, peak_value, atol=1e-6)
def tune_zeromean():
with open(get_kernel_path() + 'zeromeantotalfilter.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
tune_vertical(kernel_string, image, height, width)
tune_horizontal(kernel_string, image, height, width)
tune_transpose(kernel_string, image, height, width)
def tune_pce():
with open(get_kernel_path()+'peaktocorrelationenergy.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "Pentax_OptioA40_0_30731.JPG", mode="F")
image = fastnoise(image)
image2 = imread(get_testdata_path() + "Pentax_OptioA40_0_30757.JPG", mode="F")
image2 = fastnoise(image2)
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
image_freq, image2_freq = tune_complex_and_flip(kernel_string, height, width, image, image2)
crosscorr = tune_crosscorr(kernel_string, height, width, image_freq, image2_freq)
loc, val = tune_find_peak(kernel_string, height, width, crosscorr)
energy = tune_energy(kernel_string, height, width, crosscorr, loc)
pce_score = (val[0] * val[0]) / energy
print("Finished tuning PCE, pce_score=", pce_score)
def test_fastnoise():
with open(get_kernel_path()+'fastnoisefilter.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
problem_size = (width, height)
output1 = np.zeros_like(image)
output2 = np.zeros_like(image)
output3 = np.zeros_like(image)
args = [height, width, output1, output2, image]
params = OrderedDict()
params["block_size_x"] = 32
params["block_size_y"] = 16
d = np.gradient(image)
norm = np.sqrt( (d[0]*d[0]) + (d[1]*d[1]) )
scale = 1.0 / (1.0 + norm)
dys = d[0] * scale
dxs = d[1] * scale
answer = run_kernel("normalized_gradient",
kernel_string, problem_size, args, params)
assert np.allclose(answer[2], dxs, atol=1e-6)
assert np.allclose(answer[3], dys, atol=1e-6)
args = [height, width, output3, dxs, dys]
answer = run_kernel("gradient",
kernel_string, problem_size, args, params)
reference = np.gradient(dys, axis=0) + np.gradient(dxs, axis=1)
assert np.allclose(answer[2], reference, atol=1e-6)
def test_complex_and_flip2():
with open(get_kernel_path() + 'peaktocorrelationenergy.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test_small.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
problem_size = (width, height)
output = np.zeros((height, width, 2), dtype=np.float32)
args = [height, width, output, output, image, image]
params = OrderedDict()
params["block_size_x"] = 32
params["block_size_y"] = 16
answer = run_kernel("toComplexAndFlip2",
kernel_string,
problem_size,
args,
params,
grid_div_y=["block_size_y"],
grid_div_x=["block_size_x"])
output1 = answer[2].reshape(height, width, 2)
output1 = output1[:, :, 0] + 1j * output[:, :, 1]
reference1 = image + 1j * np.zeros((height, width), dtype=np.float32)
assert np.allclose(output1, reference1, atol=1e-6)
reference2 = image.flatten()[::-1].reshape(height, width)
reference2 = reference2
output2 = answer[3].reshape(height, width, 2)
assert np.allclose(output2[:, :, 0], reference2, atol=1e-6)
assert np.allclose(output2[:, :, 1],
np.zeros((height, width), dtype=np.float32),
atol=1e-6)
def test_variance_zero_mean():
with open(get_kernel_path() + 'wienerfilter.cu', 'r') as f:
kernel_string = f.read()
image = imread(get_testdata_path() + "test.jpg", mode="F")
height = np.int32(image.shape[0])
width = np.int32(image.shape[1])
size = np.int32(height * width)
params = OrderedDict()
params["block_size_x"] = 512
params["num_blocks"] = 64
num_blocks = params["num_blocks"]
output = np.zeros(num_blocks, dtype=np.float32)
args = [size, output, image]
problem_size = ("num_blocks", 1)
answer = run_kernel("computeVarianceZeroMean",
kernel_string,
problem_size,
args,
params,
grid_div_x=[])
print("answer:")
ans = np.sum(answer[1])
print(ans, answer[1])
print("reference:")
reference = np.sum(image * image)
print(reference)
assert np.isclose(ans, reference, atol=1e-6)