def renderWithError(stepsize, prevImage):
rendererArgs.stepsize = stepsize
timer.start()
pyrenderer.render(STEPPING_KERNEL, volume, rendererArgs, output)
timer.stop()
outputStepping = np.array(output.copy_to_cpu())
timeStepping = timer.elapsed_ms()
outputSteppingBlended = blendToWhite(outputStepping)
error = np.max(np.abs(prevImage - outputSteppingBlended))
print("step size", stepsize, "-> error", error)
return error, timeStepping, outputSteppingBlended
def runTimings(mode: str):
# CONFIGURATION
SETTINGS_FILE = "../scenes/dvrSphere-scene1.json"
ROOT_PATH = ".."
RESOLUTION = (512, 512)
NUM_FRAMES = 1
VOLUME_RESOLUTION = 128
KERNEL_NAMES = [
("DVR: DDA - fixed step (control points)",
"dvr stepping 0.0001\n(Baseline)", 0.001),
("DVR: Fixed step size - trilinear", "stepping 0.1", 0.1),
("DVR: Fixed step size - trilinear", "stepping 0.01", 0.01),
("DVR: Fixed step size - trilinear", "stepping 0.001", 0.001),
("DVR: DDA - interval simple", "interval - simple", 1),
("DVR: DDA - interval stepping (3)", "interval - stepping-3", 1),
("DVR: DDA - interval trapezoid (2)", "interval - trapezoid-2", 1),
("DVR: DDA - interval trapezoid (4)", "interval - trapezoid-4", 1),
("DVR: DDA - interval trapezoid (10)", "interval - trapezoid-10", 1),
("DVR: DDA - interval Simpson (2)", "interval - Simpson-2", 1),
("DVR: DDA - interval Simpson (4)", "interval - Simpson-4", 1),
("DVR: DDA - interval Simpson (10)", "interval - Simpson-10", 1),
("DVR: DDA - interval Simpson adapt",
"interval - Simpson-adaptive e-3", 1e-3 * VOLUME_RESOLUTION),
("DVR: DDA - interval Simpson adapt",
"interval - Simpson-adaptive e-5", 1e-5 * VOLUME_RESOLUTION),
#("DVR: DDA - interval trapezoid var", "interval - trapezoid-var 0.1", 0.1),
#("DVR: DDA - interval trapezoid var", "interval - trapezoid-var 0.01", 0.01),
#("DVR: DDA - interval trapezoid var", "interval - trapezoid-var 0.001", 0.001),
("DVR: Marching Cubes", "marching cubes 1", 1 / 1 - 0.001
), # number of subdivisions
("DVR: Marching Cubes", "marching cubes 2", 1 / 2 - 0.001),
("DVR: Marching Cubes", "marching cubes 4", 1 / 4 - 0.001),
("DVR: Marching Cubes", "marching cubes 8", 1 / 8 - 0.001),
("DVR: Marching Cubes", "marching cubes 16", 1 / 16 - 0.001),
]
DENSITY_STEPS = 7
MIN_DENSITY_DIFFERENCE = 0.005 # minimal difference between min and max density
TIMING_STEPS = 50
OUTPUT_STATS_ALL = "../results/statistics/dvr-sphere/timings-all-%s.tsv"
OUTPUT_STATS_AVG = "../results/statistics/dvr-sphere/timings-avg-%s.tsv"
OUTPUT_HISTO_ALL = "../results/statistics/dvr-sphere/histograms-%s.tsv"
OUTPUT_HISTO_CFG = "../results/statistics/dvr-sphere/histogram-cfg-%s.tsv"
OUTPUT_STATS_USE_DOUBLE = False
OUTPUT_IMAGE_PATH = "../results/statistics/dvr-sphere/images/"
OUTPUT_INSTRUMENTATION = "../results/statistics/dvr-sphere/instrumentation.tsv"
HISTO_NUM_BINS = 100
HISTO_BIN_MIN = np.log10(1e-6)
HISTO_BIN_MAX = np.log10(1)
HISTO_BIN_EDGES = [0.0] + list(10**np.linspace(
HISTO_BIN_MIN, HISTO_BIN_MAX, HISTO_NUM_BINS))
print("histogram bins:", HISTO_BIN_EDGES)
pyrenderer.oit.set_fragment_buffer_size(2**26)
pyrenderer.oit.set_marching_cubes_mode(
pyrenderer.oit.MarchingCubesComputationMode.OnTheFly)
pyrenderer.oit.set_max_fragments_per_pixel(256)
pyrenderer.oit.set_tile_size(256)
os.makedirs(OUTPUT_IMAGE_PATH, exist_ok=True)
# load settings file
rendererArgs, camera, volumePath = pyrenderer.load_from_json(
SETTINGS_FILE, ROOT_PATH)
print("settings loaded")
rendererArgs.width = RESOLUTION[0]
rendererArgs.height = RESOLUTION[1]
base_min_density = rendererArgs.min_density
base_max_density = rendererArgs.max_density
base_opacity = rendererArgs.opacity_scaling
# create density+opacity test cases
end_max_density = 0.5 * (base_min_density +
base_max_density) + MIN_DENSITY_DIFFERENCE
max_densities = np.exp(
np.linspace(np.log(base_max_density), np.log(end_max_density),
DENSITY_STEPS))
scaling = (base_max_density - base_min_density) / (max_densities -
base_min_density)
end_min_density = 0.5 * (base_min_density +
base_max_density) - MIN_DENSITY_DIFFERENCE
min_densities = np.exp(
np.linspace(np.log(base_min_density), np.log(end_min_density),
DENSITY_STEPS))
scaling = (base_max_density - base_min_density) / (max_densities -
min_densities)
opacities = base_opacity * scaling
print("min_densities:", min_densities)
print("max_densities:", max_densities)
print("opacities:", opacities)
# create volume
print("Create Marschner Lobb")
volume = pyrenderer.Volume.create_implicit(
pyrenderer.ImplicitEquation.Sphere, VOLUME_RESOLUTION)
print("Loaded volumed of resolution", volume.resolution, "and world size",
volume.world_size)
volume.copy_to_gpu()
# allocate timing
timer = pyrenderer.GpuTimer()
times = [[None] * DENSITY_STEPS for i in range(len(KERNEL_NAMES))]
if mode == "visualize":
# allocate output
pyrenderer.reload_kernels(enableDebugging=False,
enableInstrumentation=False)
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [[None] * DENSITY_STEPS for i in range(len(KERNEL_NAMES))]
# render
camera.update_render_args(rendererArgs)
for j, (kernel_name, _, stepsize) in enumerate(KERNEL_NAMES):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
for i in range(DENSITY_STEPS):
rendererArgs.min_density = min_densities[i]
rendererArgs.max_density = max_densities[i]
rendererArgs.opacity_scaling = opacities[i]
timer.start()
pyrenderer.render(kernel_name, volume, rendererArgs, output)
timer.stop()
outputs[j][i] = np.array(output.copy_to_cpu())
times[j][i] = timer.elapsed_ms()
def slugify(value):
"""
Normalizes string, converts to lowercase, removes non-alpha characters,
and converts spaces to hyphens.
"""
import unicodedata
import re
value = str(unicodedata.normalize(
'NFKD', value)) #.encode('ascii', 'ignore'))
value = re.sub('[^\w\s-]', '', value).strip().lower()
value = re.sub('[-\s]+', '-', value)
return value
# visualize
print("Visualize")
#fig, axes = plt.subplots(nrows=len(KERNEL_NAMES), ncols=DENSITY_STEPS)
for j, (kernel_name, human_kernel_name, _) in enumerate(KERNEL_NAMES):
for i in range(DENSITY_STEPS):
filename = os.path.join(
OUTPUT_IMAGE_PATH,
slugify("%s__%d" % (human_kernel_name, i)) + ".png")
imageio.imwrite(filename, outputs[j][i][:, :, 0:4])
#axes[j][i].imshow(outputs[j][i][:,:,0:4])
#axes[j][i].set_title("time=%.2fms"%times[j][i])
#if j==len(KERNEL_NAMES)-1:
# axes[j][i].set_xlabel("range=%.3f"%(
# max_densities[i]-base_min_density))
#if i==0:
# axes[j][i].set_ylabel(human_kernel_name)
# save to numpy
npz_output = {}
npz_output['kernels'] = KERNEL_NAMES
npz_output['densities'] = [
max_densities[i] - base_min_density for i in range(DENSITY_STEPS)
]
for j in range(len(KERNEL_NAMES)):
for i in range(DENSITY_STEPS):
npz_output['img_%d_%d' % (j, i)] = outputs[j][i]
np.savez(os.path.join(OUTPUT_IMAGE_PATH, "raw.npz"), **npz_output)
#plt.subplots_adjust(left=0.03, bottom=0.05, right=0.99, top=0.97, wspace=0.20, hspace=0.23)
#plt.show()
elif mode == "measure":
summed_times = [[0] * DENSITY_STEPS for i in range(len(KERNEL_NAMES))]
pyrenderer.reload_kernels(enableDebugging=False,
enableInstrumentation=False,
otherPreprocessorArguments=[
"-DKERNEL_USE_DOUBLE=%s" %
("1" if OUTPUT_STATS_USE_DOUBLE else "0")
])
# allocate output for baseline
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [[None] * TIMING_STEPS for i in range(DENSITY_STEPS)]
histograms = [[
np.zeros(HISTO_NUM_BINS, dtype=np.int64)
for i in range(DENSITY_STEPS)
] for j in range(len(KERNEL_NAMES))]
# render and write output
with open(
OUTPUT_STATS_ALL %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tTF-Range\tFrame\tTime (ms)\tPSNR (dB)\n")
for j, (kernel_name, human_kernel_name,
stepsize) in enumerate(KERNEL_NAMES):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
for i in range(DENSITY_STEPS):
rendererArgs.min_density = min_densities[i]
rendererArgs.max_density = max_densities[i]
rendererArgs.opacity_scaling = opacities[i]
histogram = histograms[j][i]
histogram_edges = None
for k in range(TIMING_STEPS):
camera.yaw = k * 360 / TIMING_STEPS
camera.update_render_args(rendererArgs)
timer.start()
pyrenderer.render(kernel_name, volume, rendererArgs,
output)
timer.stop()
out_img = np.array(output.copy_to_cpu())
if j == 0: # baseline
outputs[i][k] = out_img
psnr = 0
else:
# compute psnr
maxValue = np.max(outputs[i][k][:, :, 0:4])
mse = ((outputs[i][k][:, :, 0:4] -
out_img[:, :, 0:4])**2).mean(axis=None)
psnr = 20 * np.log10(maxValue) - 10 * np.log10(mse)
# compute histogram
diff = outputs[i][k][:, :, 0:4] - out_img[:, :,
0:4]
new_histogram, histogram_edges = np.histogram(
diff, bins=HISTO_BIN_EDGES)
histogram += new_histogram
t = timer.elapsed_ms()
summed_times[j][i] += t
f.write(
"%s\t%.4f\t%d\t%.4f\t%.4f\n" %
(human_kernel_name.replace("\n", " "),
max_densities[i] - base_min_density, k, t, psnr))
# write average stats
with open(
OUTPUT_STATS_AVG %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tTF-Range\tAvg-Time (ms)\n")
for j, (_, human_kernel_name, _) in enumerate(KERNEL_NAMES):
for i in range(DENSITY_STEPS):
f.write("%s\t%.4f\t%.4f\n" %
(human_kernel_name.replace("\n", " "),
max_densities[i] - base_min_density,
summed_times[j][i] / TIMING_STEPS))
# write histograms
with open(
OUTPUT_HISTO_ALL %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("BinStart\tBinEnd")
for j in range(len(KERNEL_NAMES)):
for i in range(DENSITY_STEPS):
f.write("\t%d-%d" % (j, i))
f.write("\n")
for b in range(HISTO_NUM_BINS):
f.write("%.10f\t%.10f" %
(HISTO_BIN_EDGES[b], HISTO_BIN_EDGES[b + 1]))
for j in range(len(KERNEL_NAMES)):
for i in range(DENSITY_STEPS):
f.write("\t%d" % histograms[j][i][b])
f.write("\n")
with open(
OUTPUT_HISTO_CFG %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tTF-Range\tConfig-ID\n")
for j, (_, human_kernel_name, _) in enumerate(KERNEL_NAMES):
for i in range(DENSITY_STEPS):
f.write("%s\t%.4f\t%s\n" %
(human_kernel_name.replace("\n", " "),
max_densities[i] - base_min_density, "%d-%d" %
(j, i)))
elif mode == "instrumentation":
# recompile with instrumentation
pyrenderer.reload_kernels(enableInstrumentation=True)
# allocate output
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [[None] * DENSITY_STEPS for i in range(len(KERNEL_NAMES))]
fields = [
"densityFetches", "tfFetches", "ddaSteps", "isoIntersections",
"intervalEval", "intervalStep", "intervalMaxStep"
]
# render
with open(OUTPUT_INSTRUMENTATION, "w") as f:
f.write(
"Kernel Name\tTF-Range\t%s\tavgIntervalSteps\tnumTriangles\tnumFragments\n"
% "\t".join(fields))
camera.update_render_args(rendererArgs)
for j, (kernel_name, human_kernel_name,
stepsize) in enumerate(KERNEL_NAMES):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
for i in range(DENSITY_STEPS):
rendererArgs.min_density = min_densities[i]
rendererArgs.max_density = max_densities[i]
rendererArgs.opacity_scaling = opacities[i]
instrumentations, globalInstrumentations = \
pyrenderer.render_with_instrumentation(
kernel_name, volume, rendererArgs, output)
f.write("%s\t%s" % (human_kernel_name.replace(
"\n", " "), max_densities[i] - base_min_density))
for field in fields:
f.write(
"\t%.2f" %
np.mean(instrumentations[field], dtype=np.float32))
avgIntervalSteps = np.mean(
instrumentations["intervalStep"].astype(np.float32) /
instrumentations["intervalEval"].astype(np.float32))
f.write(
"\t%.2f\t%d\t%d\n" %
(avgIntervalSteps, globalInstrumentations.numTriangles,
globalInstrumentations.numFragments))
f.flush()
def runTimings(mode: str):
# CONFIGURATION
SETTINGS_FILE = "../scenes/dvrLobb-scene4.json"
ROOT_PATH = ".."
RESOLUTION = (512, 512)
NUM_FRAMES = 1
VOLUME_RESOLUTION = 128
KERNEL_NAMES_NORMAL, KERNEL_NAMES_TIMING = utils.getKernelNames(
VOLUME_RESOLUTION)
DENSITY_STEPS = 10
WIDTH_START = 0.05
WIDTH_END = 0.002
SCALING_START = 20.0
EXPONENT = 0.1
TIMING_STEPS = utils.TIMING_STEPS
OUTPUT_STATS_ALL = "../results/statistics/dvr-marschner-lobb/timings-all-%s.tsv"
OUTPUT_STATS_AVG = "../results/statistics/dvr-marschner-lobb/timings-avg-%s.tsv"
OUTPUT_HISTO_ALL = "../results/statistics/dvr-marschner-lobb/histograms-%s.tsv"
OUTPUT_HISTO_CFG = "../results/statistics/dvr-marschner-lobb/histogram-cfg-%s.tsv"
OUTPUT_STATS_USE_DOUBLE = False
OUTPUT_IMAGE_PATH = "../results/statistics/dvr-marschner-lobb/images/"
OUTPUT_INSTRUMENTATION = "../results/statistics/dvr-marschner-lobb/instrumentation.tsv"
HISTO_NUM_BINS = 100
HISTO_BIN_MIN = np.log10(1e-6)
HISTO_BIN_MAX = np.log10(1)
HISTO_BIN_EDGES = [0.0] + list(10**np.linspace(
HISTO_BIN_MIN, HISTO_BIN_MAX, HISTO_NUM_BINS))
print("histogram bins:", HISTO_BIN_EDGES)
pyrenderer.oit.set_fragment_buffer_size(2**26)
pyrenderer.oit.set_marching_cubes_mode(
pyrenderer.oit.MarchingCubesComputationMode.OnTheFly)
pyrenderer.oit.set_max_fragments_per_pixel(512)
pyrenderer.oit.set_tile_size(128)
os.makedirs(OUTPUT_IMAGE_PATH, exist_ok=True)
# load settings file
rendererArgs, camera, volumePath = pyrenderer.load_from_json(
SETTINGS_FILE, ROOT_PATH)
print("settings loaded")
rendererArgs.width = RESOLUTION[0]
rendererArgs.height = RESOLUTION[1]
camera.update_render_args(rendererArgs)
multiiso_tf = rendererArgs.get_multiiso_tf()
# create volume
print("Create Marschner Lobb")
volume = pyrenderer.Volume.create_implicit(
pyrenderer.ImplicitEquation.MarschnerLobb, VOLUME_RESOLUTION)
print("Loaded volumed of resolution", volume.resolution, "and world size",
volume.world_size)
volume.copy_to_gpu()
def computeSettings(time, rendererArgs):
settings = rendererArgs.clone()
width = pow(
utils.lerp(time, pow(WIDTH_START, EXPONENT),
pow(WIDTH_END, EXPONENT)), 1 / EXPONENT)
scaling = SCALING_START * WIDTH_START / width
settings.set_linear_tf(utils.multiIso2Linear(multiiso_tf, width))
settings.opacity_scaling = scaling
return settings, width
# allocate timing
timer = pyrenderer.GpuTimer()
times = [[None] * DENSITY_STEPS for i in range(len(KERNEL_NAMES_NORMAL))]
if mode == "visualize":
# allocate output
pyrenderer.reload_kernels(enableDebugging=False,
enableInstrumentation=False)
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [[None] * DENSITY_STEPS
for i in range(len(KERNEL_NAMES_NORMAL))]
# render
camera.update_render_args(rendererArgs)
for j, (kernel_name, _, stepsize) in enumerate(KERNEL_NAMES_NORMAL):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
for i in range(DENSITY_STEPS):
settings, width = computeSettings(i / (DENSITY_STEPS - 1),
rendererArgs)
timer.start()
pyrenderer.render(kernel_name, volume, settings, output)
timer.stop()
outputs[j][i] = np.array(output.copy_to_cpu())
times[j][i] = timer.elapsed_ms()
def slugify(value):
"""
Normalizes string, converts to lowercase, removes non-alpha characters,
and converts spaces to hyphens.
"""
import unicodedata
import re
value = str(unicodedata.normalize(
'NFKD', value)) #.encode('ascii', 'ignore'))
value = re.sub('[^\w\s-]', '', value).strip().lower()
value = re.sub('[-\s]+', '-', value)
return value
# visualize
print("Visualize")
#fig, axes = plt.subplots(nrows=len(KERNEL_NAMES_NORMAL), ncols=DENSITY_STEPS)
for j, (kernel_name, human_kernel_name,
_) in enumerate(KERNEL_NAMES_NORMAL):
for i in range(DENSITY_STEPS):
filename = os.path.join(
OUTPUT_IMAGE_PATH,
slugify("%s__%d" % (human_kernel_name, i)) + ".png")
imageio.imwrite(filename, outputs[j][i][:, :, 0:4])
#axes[j][i].imshow(outputs[j][i][:,:,0:4])
#axes[j][i].set_title("time=%.2fms"%times[j][i])
#if j==len(KERNEL_NAMES_NORMAL)-1:
# axes[j][i].set_xlabel("range=%.3f"%(
# max_densities[i]-base_min_density))
#if i==0:
# axes[j][i].set_ylabel(human_kernel_name)
# save to numpy
npz_output = {}
npz_output['kernels'] = KERNEL_NAMES_NORMAL
npz_output['widths'] = [
computeSettings(i / (DENSITY_STEPS - 1), rendererArgs)[1]
for i in range(DENSITY_STEPS)
]
for j in range(len(KERNEL_NAMES_NORMAL)):
for i in range(DENSITY_STEPS):
npz_output['img_%d_%d' % (j, i)] = outputs[j][i]
np.savez(os.path.join(OUTPUT_IMAGE_PATH, "raw.npz"), **npz_output)
#plt.subplots_adjust(left=0.03, bottom=0.05, right=0.99, top=0.97, wspace=0.20, hspace=0.23)
#plt.show()
elif mode == "measure":
summed_times = [[0] * DENSITY_STEPS
for i in range(len(KERNEL_NAMES_NORMAL))]
pyrenderer.reload_kernels(enableDebugging=False,
enableInstrumentation=False,
otherPreprocessorArguments=[
"-DKERNEL_USE_DOUBLE=%s" %
("1" if OUTPUT_STATS_USE_DOUBLE else "0")
])
# allocate output for baseline
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [[None] * TIMING_STEPS for i in range(DENSITY_STEPS)]
histograms = [[
np.zeros(HISTO_NUM_BINS, dtype=np.int64)
for i in range(DENSITY_STEPS)
] for j in range(len(KERNEL_NAMES_NORMAL))]
# render and write output
with open(
OUTPUT_STATS_ALL %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tWidth\tFrame\tTime (ms)\tPSNR (dB)\n")
#for j, (kernel_name, human_kernel_name, stepsize) in enumerate(KERNEL_NAMES):
for j in range(len(KERNEL_NAMES_NORMAL)):
kernel_name_normal, human_kernel_name, stepsize = KERNEL_NAMES_NORMAL[
j]
kernel_name_timings, _, _ = KERNEL_NAMES_TIMING[j]
print("Render", kernel_name_normal, stepsize)
rendererArgs.stepsize = stepsize
for i in range(DENSITY_STEPS):
settings, width = computeSettings(i / (DENSITY_STEPS - 1),
rendererArgs)
print(" width=%.4f, scaling=%.4f" %
(width, settings.opacity_scaling))
histogram = histograms[j][i]
histogram_edges = None
for k in range(TIMING_STEPS):
camera.yaw = k * 360 / TIMING_STEPS
camera.update_render_args(settings)
timer.start()
pyrenderer.render(kernel_name_timings, volume,
settings, output)
timer.stop()
if kernel_name_timings != kernel_name_normal:
pyrenderer.render(kernel_name_normal, volume,
settings, output)
out_img = np.array(output.copy_to_cpu())
if j == 0: # baseline
outputs[i][k] = out_img
psnr = 0
else:
# compute psnr
maxValue = 1.0 #np.max(outputs[i][k][:,:,0:4])
mse = ((outputs[i][k][:, :, 0:4] -
out_img[:, :, 0:4])**2).mean(axis=None)
psnr = 20 * np.log10(maxValue) - 10 * np.log10(mse)
#psnr = np.mean(np.abs(outputs[i][k][:,:,0:4] - out_img[:,:,0:4]))
# compute histogram
diff = np.abs(outputs[i][k][:, :, 0:4] -
out_img[:, :, 0:4])
new_histogram, histogram_edges = np.histogram(
diff, bins=HISTO_BIN_EDGES)
histogram += new_histogram
t = timer.elapsed_ms()
summed_times[j][i] += t
f.write("%s\t%.4f\t%d\t%.4f\t%.4f\n" %
(human_kernel_name.replace(
"\n", " "), width, k, t, psnr))
# write average stats
with open(
OUTPUT_STATS_AVG %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tWidth\tAvg-Time (ms)\n")
for j, (_, human_kernel_name, _) in enumerate(KERNEL_NAMES_NORMAL):
for i in range(DENSITY_STEPS):
_, width = computeSettings(i / (DENSITY_STEPS - 1),
rendererArgs)
f.write("%s\t%.4f\t%.4f\n" % (human_kernel_name.replace(
"\n", " "), width, summed_times[j][i] / TIMING_STEPS))
# write histograms
with open(
OUTPUT_HISTO_ALL %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("BinStart\tBinEnd")
for j in range(len(KERNEL_NAMES_NORMAL)):
for i in range(DENSITY_STEPS):
f.write("\t%d-%d" % (j, i))
f.write("\n")
for b in range(HISTO_NUM_BINS):
f.write("%.10f\t%.10f" %
(HISTO_BIN_EDGES[b], HISTO_BIN_EDGES[b + 1]))
for j in range(len(KERNEL_NAMES_NORMAL)):
for i in range(DENSITY_STEPS):
f.write("\t%d" % histograms[j][i][b])
f.write("\n")
with open(
OUTPUT_HISTO_CFG %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tWidth\tConfig-ID\n")
for j, (_, human_kernel_name, _) in enumerate(KERNEL_NAMES_NORMAL):
for i in range(DENSITY_STEPS):
f.write("%s\t%.4f\t%s\n" %
(human_kernel_name.replace("\n", " "),
computeSettings(i / (DENSITY_STEPS - 1),
rendererArgs)[1], "%d-%d" %
(j, i)))
elif mode == "instrumentation":
# recompile with instrumentation
pyrenderer.reload_kernels(enableInstrumentation=True)
# allocate output
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [[None] * DENSITY_STEPS
for i in range(len(KERNEL_NAMES_NORMAL))]
fields = [
"densityFetches", "tfFetches", "ddaSteps", "isoIntersections",
"intervalEval", "intervalStep", "intervalMaxStep"
]
# render
with open(OUTPUT_INSTRUMENTATION, "w") as f:
f.write(
"Kernel Name\tWidth\t%s\tavgIntervalSteps\tnumTriangles\tnumFragments\n"
% "\t".join(fields))
camera.update_render_args(rendererArgs)
for j, (kernel_name, human_kernel_name,
stepsize) in enumerate(KERNEL_NAMES_NORMAL):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
for i in range(DENSITY_STEPS):
settings, width = computeSettings(i / (DENSITY_STEPS - 1),
rendererArgs)
instrumentations, globalInstrumentations = \
pyrenderer.render_with_instrumentation(
kernel_name, volume, settings, output)
f.write("%s\t%s" %
(human_kernel_name.replace("\n", " "), width))
for field in fields:
f.write(
"\t%.2f" %
np.mean(instrumentations[field], dtype=np.float32))
avgIntervalSteps = np.mean(
instrumentations["intervalStep"].astype(np.float32) /
instrumentations["intervalEval"].astype(np.float32))
f.write(
"\t%.2f\t%d\t%d\n" %
(avgIntervalSteps, globalInstrumentations.numTriangles,
globalInstrumentations.numFragments))
f.flush()
def runTimings(mode: str):
# CONFIGURATION
SETTINGS_FILE = "../scenes/bigBug1.json"
ROOT_PATH = ".."
RESOLUTION = (512, 512)
NUM_FRAMES = 1
VOLUME_RESOLUTION = 128
KERNEL_NAMES = [
("DVR: DDA - fixed step (control points)", "dvr stepping 0.0001\n(Baseline)", 0.0001)] + \
[("DVR: Fixed step size - trilinear", "stepping 2^-%d"%s, pow(2, -s)) for s in range(2,15)]
DENSITY_STEPS = 7
MIN_DENSITY_DIFFERENCE = 0.01 # minimal difference between min and max density
TIMING_STEPS = 1
OUTPUT_STATS_ALL = "../results/statistics/dvr-beetle/timingsConvergence-all-%s.tsv"
OUTPUT_STATS_AVG = "../results/statistics/dvr-beetle/timingsConvergence-avg-%s.tsv"
OUTPUT_HISTO_ALL = "../results/statistics/dvr-beetle/histogramsConvergence-%s.tsv"
OUTPUT_HISTO_CFG = "../results/statistics/dvr-beetle/histogramConvergence-cfg-%s.tsv"
OUTPUT_STATS_USE_DOUBLE = True
os.makedirs("../results/statistics/dvr-beetle/", exist_ok=True)
HISTO_NUM_BINS = 100
HISTO_BIN_MIN = np.log10(1e-6)
HISTO_BIN_MAX = np.log10(1)
HISTO_BIN_EDGES = [0.0] + list(10**np.linspace(
HISTO_BIN_MIN, HISTO_BIN_MAX, HISTO_NUM_BINS))
print("histogram bins:", HISTO_BIN_EDGES)
pyrenderer.oit.set_fragment_buffer_size(2**26)
pyrenderer.oit.set_marching_cubes_mode(
pyrenderer.oit.MarchingCubesComputationMode.OnTheFly)
pyrenderer.oit.set_max_fragments_per_pixel(512)
pyrenderer.oit.set_tile_size(128)
# load settings file
rendererArgs, camera, volumePath = pyrenderer.load_from_json(
SETTINGS_FILE, ROOT_PATH)
print("settings loaded")
rendererArgs.width = RESOLUTION[0]
rendererArgs.height = RESOLUTION[1]
base_min_density = rendererArgs.min_density
base_max_density = rendererArgs.max_density
base_opacity = rendererArgs.opacity_scaling
# create density+opacity test cases
end_max_density = base_min_density + MIN_DENSITY_DIFFERENCE
max_densities = np.exp(
np.linspace(np.log(base_max_density), np.log(end_max_density),
DENSITY_STEPS))
scaling = (base_max_density - base_min_density) / (max_densities -
base_min_density)
opacities = base_opacity * scaling
# create volume
print("Create Marschner Lobb")
volume = pyrenderer.Volume(volumePath)
print("Loaded volumed of resolution", volume.resolution, "and world size",
volume.world_size)
volume.copy_to_gpu()
# allocate timing
timer = pyrenderer.GpuTimer()
times = [[None] * DENSITY_STEPS for i in range(len(KERNEL_NAMES))]
if mode == "measure":
summed_times = [[0] * DENSITY_STEPS for i in range(len(KERNEL_NAMES))]
pyrenderer.reload_kernels(enableDebugging=False,
enableInstrumentation=False,
otherPreprocessorArguments=[
"-DKERNEL_USE_DOUBLE=%s" %
("1" if OUTPUT_STATS_USE_DOUBLE else "0")
])
# allocate output for baseline
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [[None] * TIMING_STEPS for i in range(DENSITY_STEPS)]
histograms = [[
np.zeros(HISTO_NUM_BINS, dtype=np.int64)
for i in range(DENSITY_STEPS)
] for j in range(len(KERNEL_NAMES))]
# render and write output
with open(
OUTPUT_STATS_ALL %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tTF-Range\tFrame\tTime (ms)\tPSNR (dB)\n")
for j, (kernel_name, human_kernel_name,
stepsize) in enumerate(KERNEL_NAMES):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
for i in range(DENSITY_STEPS):
rendererArgs.max_density = max_densities[i]
rendererArgs.opacity_scaling = opacities[i]
histogram = histograms[j][i]
histogram_edges = None
for k in range(TIMING_STEPS):
camera.yaw = k * 360 / TIMING_STEPS
camera.update_render_args(rendererArgs)
timer.start()
pyrenderer.render(kernel_name, volume, rendererArgs,
output)
timer.stop()
out_img = np.array(output.copy_to_cpu())
if j == 0: # baseline
outputs[i][k] = out_img
psnr = 0
else:
# compute psnr
maxValue = np.max(outputs[i][k][:, :, 0:4])
mse = ((outputs[i][k][:, :, 0:4] -
out_img[:, :, 0:4])**2).mean(axis=None)
psnr = 20 * np.log10(maxValue) - 10 * np.log10(mse)
# compute histogram
diff = outputs[i][k][:, :, 0:4] - out_img[:, :,
0:4]
new_histogram, histogram_edges = np.histogram(
diff, bins=HISTO_BIN_EDGES)
histogram += new_histogram
t = timer.elapsed_ms()
summed_times[j][i] += t
f.write(
"%s\t%.4f\t%d\t%.4f\t%.4f\n" %
(human_kernel_name.replace("\n", " "),
max_densities[i] - base_min_density, k, t, psnr))
# write average stats
with open(
OUTPUT_STATS_AVG %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tTF-Range\tAvg-Time (ms)\n")
for j, (_, human_kernel_name, _) in enumerate(KERNEL_NAMES):
for i in range(DENSITY_STEPS):
f.write("%s\t%.4f\t%.4f\n" %
(human_kernel_name.replace("\n", " "),
max_densities[i] - base_min_density,
summed_times[j][i] / TIMING_STEPS))
# write histograms
with open(
OUTPUT_HISTO_ALL %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("BinStart\tBinEnd")
for j in range(len(KERNEL_NAMES)):
for i in range(DENSITY_STEPS):
f.write("\t%d-%d" % (j, i))
f.write("\n")
for b in range(HISTO_NUM_BINS):
f.write("%.10f\t%.10f" %
(HISTO_BIN_EDGES[b], HISTO_BIN_EDGES[b + 1]))
for j in range(len(KERNEL_NAMES)):
for i in range(DENSITY_STEPS):
f.write("\t%d" % histograms[j][i][b])
f.write("\n")
with open(
OUTPUT_HISTO_CFG %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tTF-Range\tConfig-ID\n")
for j, (_, human_kernel_name, _) in enumerate(KERNEL_NAMES):
for i in range(DENSITY_STEPS):
f.write("%s\t%.4f\t%s\n" %
(human_kernel_name.replace("\n", " "),
max_densities[i] - base_min_density, "%d-%d" %
(j, i)))
def renderImage(sceneFile: str, name: str, scaleIndependent: bool = False):
# load scene file
RESOLUTION = (1920, 1080)
ROOT_PATH = ".."
rendererArgs, camera, volumePath = pyrenderer.load_from_json(
sceneFile, ROOT_PATH)
print("\n\n============================\n")
print("settings loaded:", sceneFile)
rendererArgs.width = RESOLUTION[0]
rendererArgs.height = RESOLUTION[1]
volume = pyrenderer.Volume(volumePath)
print("Loaded volumed of resolution", volume.resolution, "and world size",
volume.world_size)
volume.copy_to_gpu()
# configuration
ERROR = 1 / 256.0
if scaleIndependent:
STEPPING_KERNEL = "DVR: Scale invariant - trilinear"
ANALYTIC_KERNEL = "DVR: Scale invariant - Simpson"
elif rendererArgs.dvrUseShading:
STEPPING_KERNEL = "DVR: Fixed step size - trilinear"
ANALYTIC_KERNEL = "DVR: DDA - interval Simpson shaded"
else:
STEPPING_KERNEL = "DVR: Fixed step size - trilinear"
ANALYTIC_KERNEL = "DVR: DDA - interval Simpson adapt"
print("stepping kernel:", STEPPING_KERNEL)
print("analytic kernel:", ANALYTIC_KERNEL)
INITIAL_STEPSIZE = 0.25
MIN_STEPSIZE = 1e-5
BINARY_SEARCH_STEPS = 10
os.makedirs(OUTPUT_IMAGE_PATH, exist_ok=True)
outputImagePath = OUTPUT_IMAGE_PATH + os.path.splitext(
os.path.basename(sceneFile))[0] + ".png"
outputTexPath = OUTPUT_IMAGE_PATH + os.path.splitext(
os.path.basename(sceneFile))[0] + ".tex"
# allocate outputs
timer = pyrenderer.GpuTimer()
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
camera.update_render_args(rendererArgs)
# render analytic
rendererArgs.stepsize = ERROR
timer.start()
pyrenderer.render(ANALYTIC_KERNEL, volume, rendererArgs, output)
timer.stop()
outputAnalytic = np.array(output.copy_to_cpu())
outputAnalyticBlended = blendToWhite(outputAnalytic)
timeAnalytic = timer.elapsed_ms()
# and save
imageio.imwrite(outputImagePath, outputAnalyticBlended)
analyticResult = StepErrorTime(-1, ERROR, timeAnalytic)
def renderWithError(stepsize, prevImage):
rendererArgs.stepsize = stepsize
timer.start()
pyrenderer.render(STEPPING_KERNEL, volume, rendererArgs, output)
timer.stop()
outputStepping = np.array(output.copy_to_cpu())
timeStepping = timer.elapsed_ms()
outputSteppingBlended = blendToWhite(outputStepping)
error = np.max(np.abs(prevImage - outputSteppingBlended))
print("step size", stepsize, "-> error", error)
return error, timeStepping, outputSteppingBlended
# render stepping (coarse)
errorSteppingCoarse, timeSteppingCoarse, outputSteppingCoarse = renderWithError(
INITIAL_STEPSIZE, outputAnalyticBlended)
previousError = errorSteppingCoarse
errorSteppingCoarse = np.mean(
np.abs(outputSteppingCoarse - outputAnalyticBlended))
coarseResult = StepErrorTime(INITIAL_STEPSIZE, errorSteppingCoarse,
timeSteppingCoarse)
# search until maximal error of ERROR is reached
stepsizeUpper = INITIAL_STEPSIZE
stepsizeLower = INITIAL_STEPSIZE / 2
previousImage = outputSteppingCoarse
currentTime = None
while True:
error, currentTime, currentImage = renderWithError(
stepsizeLower, previousImage)
if error < ERROR:
break
if error > previousError * 1.2:
print("ERROR: error increased, cancel binary search")
stepsizeMid = stepsizeUpper
error = previousError
currentImage = previousImage
BINARY_SEARCH_STEPS = 0
break
stepsizeUpper = stepsizeLower
stepsizeLower /= 2
previousImage = currentImage
previousError = error
# binary search
for i in range(BINARY_SEARCH_STEPS):
stepsizeMid = pow(
2, 0.5 * (np.log2(stepsizeUpper) + np.log2(stepsizeLower)))
#print(stepsizeUpper, stepsizeLower, "->", stepsizeMid)
error, currentTime, currentImage = renderWithError(
stepsizeMid, previousImage)
if error < ERROR: # increase stepsize (coarsen)
stepsizeLower = stepsizeMid
else: # decrease stepsize (finer)
stepsizeUpper = stepsizeMid
previousImage = currentImage
finalError = np.mean(np.abs(currentImage - outputAnalyticBlended))
convergedResult = StepErrorTime(stepsizeMid, error, currentTime)
print("Final stepsize:", stepsizeMid, "with an error of", error,
"and a time of", currentTime, "ms")
Results.append(
SceneResult(name, analyticResult, coarseResult, convergedResult))
def runTimings(mode: str):
# CONFIGURATION
SETTINGS_FILE = "../scenes/isoLobb-scene1.json"
ROOT_PATH = ".."
RESOLUTION = (512, 512)
NUM_FRAMES = 1
VOLUME_RESOLUTION = 128
KERNEL_NAMES = [
("Iso: Fixed step size - nearest", "stepping nearest 0.1", 0.1),
("Iso: Fixed step size - trilinear", "stepping linear 0.1", 0.1),
("Iso: Fixed step size - tricubic", "stepping cubic 0.1", 0.1),
("Iso: Fixed step size - tricubic", "stepping cubic 0.01", 0.01),
("Iso: Fixed step size - tricubic", "stepping cubic 0.001", 0.001),
("Iso: DDA - [num] Marmitt (float, stable)", "DDA linear Marmitt", 1),
("Iso: Cubic DDA - fixed step (no poly)",
"DDA cubic fixed (no poly) 0.1", 0.1),
("Iso: Cubic DDA - fixed step (no poly)",
"DDA cubic fixed (no poly) 0.01", 0.01),
("Iso: Cubic DDA - fixed step (no poly)",
"DDA cubic fixed (no poly) 0.001", 0.001),
("Iso: Cubic DDA - fixed step (loop)", "DDA cubic fixed (loop) 0.1",
0.1),
("Iso: Cubic DDA - fixed step (loop)", "DDA cubic fixed (loop) 0.01",
0.01),
("Iso: Cubic DDA - fixed step (loop)", "DDA cubic fixed (loop) 0.001",
0.001),
("Iso: Cubic DDA - fixed step (explicit)",
"DDA cubic fixed (explicit) 0.1", 0.1),
("Iso: Cubic DDA - fixed step (explicit)",
"DDA cubic fixed (explicit) 0.01", 0.01),
("Iso: Cubic DDA - fixed step (explicit)",
"DDA cubic fixed (explicit) 0.001", 0.001),
("Iso: Cubic DDA - Sphere Simple (loop)",
"DDA cubic Sphere Simple (loop)", 1),
("Iso: Cubic DDA - Sphere Bernstein (loop)",
"DDA cubic Sphere Bernstein (loop)", 1),
("Iso: Cubic DDA - Sphere Simple (explicit)",
"DDA cubic Sphere Simple (explicit)", 1),
("Iso: Cubic DDA - Sphere Bernstein (explicit)",
"DDA cubic Sphere Bernstein (explicit)", 1),
]
KERNEL_NAMES_MEASURE = [
("Iso: Cubic DDA - fixed step (no poly)", "Baseline 0.001", 0.001),
("Iso: Cubic DDA - fixed step (no poly)",
"DDA cubic fixed (no poly) 0.1", 0.1),
("Iso: Cubic DDA - fixed step (no poly)",
"DDA cubic fixed (no poly) 0.01", 0.01),
("Iso: Cubic DDA - fixed step (loop)", "DDA cubic fixed (loop) 0.1",
0.1),
("Iso: Cubic DDA - fixed step (loop)", "DDA cubic fixed (loop) 0.01",
0.01),
("Iso: Cubic DDA - fixed step (explicit)",
"DDA cubic fixed (explicit) 0.1", 0.1),
("Iso: Cubic DDA - fixed step (explicit)",
"DDA cubic fixed (explicit) 0.01", 0.01),
("Iso: Cubic DDA - Sphere Simple (loop)",
"DDA cubic Sphere Simple (loop)", 1),
("Iso: Cubic DDA - Sphere Bernstein (loop)",
"DDA cubic Sphere Bernstein (loop)", 1),
("Iso: Cubic DDA - Sphere Simple (explicit)",
"DDA cubic Sphere Simple (explicit)", 1),
("Iso: Cubic DDA - Sphere Bernstein (explicit)",
"DDA cubic Sphere Bernstein (explicit)", 1),
]
TIMING_STEPS = 50
OUTPUT_STATS_ALL = "../results/statistics/iso-marschner-lobb/timings-all-%s.tsv"
OUTPUT_STATS_AVG = "../results/statistics/iso-marschner-lobb/timings-avg-%s.tsv"
OUTPUT_HISTO_ALL = "../results/statistics/iso-marschner-lobb/histograms-%s.tsv"
OUTPUT_HISTO_CFG = "../results/statistics/iso-marschner-lobb/histogram-cfg-%s.tsv"
OUTPUT_STATS_USE_DOUBLE = False
OUTPUT_IMAGE_PATH = "../results/statistics/iso-marschner-lobb/images/"
OUTPUT_INSTRUMENTATION = "../results/statistics/iso-marschner-lobb/instrumentation.tsv"
HISTO_NUM_BINS = 100
HISTO_BIN_MIN = np.log10(1e-6)
HISTO_BIN_MAX = np.log10(1)
HISTO_BIN_EDGES = [0.0] + list(10**np.linspace(
HISTO_BIN_MIN, HISTO_BIN_MAX, HISTO_NUM_BINS))
print("histogram bins:", HISTO_BIN_EDGES)
os.makedirs(OUTPUT_IMAGE_PATH, exist_ok=True)
# load settings file
rendererArgs, camera, volumePath = pyrenderer.load_from_json(
SETTINGS_FILE, ROOT_PATH)
print("settings loaded")
rendererArgs.width = RESOLUTION[0]
rendererArgs.height = RESOLUTION[1]
# create volume
print("Create Marschner Lobb")
volume = pyrenderer.Volume.create_marschner_lobb(VOLUME_RESOLUTION)
print("Loaded volumed of resolution", volume.resolution, "and world size",
volume.world_size)
volume.copy_to_gpu()
# allocate timing
timer = pyrenderer.GpuTimer()
times = [None] * len(KERNEL_NAMES)
if mode == "visualize":
# allocate output
pyrenderer.reload_kernels(enableDebugging=False,
enableInstrumentation=False)
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [None] * len(KERNEL_NAMES)
# render
camera.update_render_args(rendererArgs)
for j, (kernel_name, _, stepsize) in enumerate(KERNEL_NAMES):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
timer.start()
pyrenderer.render(kernel_name, volume, rendererArgs, output)
timer.stop()
outputs[j] = np.array(output.copy_to_cpu())
times[j] = timer.elapsed_ms()
def slugify(value):
"""
Normalizes string, converts to lowercase, removes non-alpha characters,
and converts spaces to hyphens.
"""
import unicodedata
import re
value = str(unicodedata.normalize(
'NFKD', value)) #.encode('ascii', 'ignore'))
value = re.sub('[^\w\s-]', '', value).strip().lower()
value = re.sub('[-\s]+', '-', value)
return value
# visualize
print("Visualize")
fig, axes = plt.subplots(ncols=len(KERNEL_NAMES), nrows=1)
for j, (kernel_name, human_kernel_name, _) in enumerate(KERNEL_NAMES):
img = screenShading(outputs[j])
filename = os.path.join(OUTPUT_IMAGE_PATH,
slugify(human_kernel_name) + ".png")
imageio.imwrite(filename, img)
axes[j].imshow(img)
axes[j].set_xlabel(human_kernel_name)
# save to numpy
npz_output = {}
npz_output['kernels'] = KERNEL_NAMES
for j in range(len(KERNEL_NAMES)):
npz_output['img_%d' % j] = outputs[j]
np.savez(os.path.join(OUTPUT_IMAGE_PATH, "raw.npz"), **npz_output)
plt.subplots_adjust(left=0.03,
bottom=0.05,
right=0.99,
top=0.97,
wspace=0.20,
hspace=0.23)
plt.show()
elif mode == "measure":
summed_times = [0] * len(KERNEL_NAMES_MEASURE)
pyrenderer.reload_kernels(enableDebugging=False,
enableInstrumentation=False,
otherPreprocessorArguments=[
"-DKERNEL_USE_DOUBLE=%s" %
("1" if OUTPUT_STATS_USE_DOUBLE else "0")
])
# allocate output for baseline
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
gt_outputs = [None] * TIMING_STEPS
# render and write output
with open(
OUTPUT_STATS_ALL %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write(
"Kernel Name\tFrame\tTime (ms)\tNum False Positives\tNum False Negatives\tMean Abs Depth Error\tVar Abs Depth Error\n"
)
for j, (kernel_name, human_kernel_name,
stepsize) in enumerate(KERNEL_NAMES_MEASURE):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
for k in range(TIMING_STEPS):
camera.yaw = k * 360 / TIMING_STEPS
camera.update_render_args(rendererArgs)
timer.start()
pyrenderer.render(kernel_name, volume, rendererArgs,
output)
timer.stop()
out_img = np.array(output.copy_to_cpu())
if j == 0: # baseline
gt_outputs[k] = out_img
falsePositives = 0
falseNegatives = 0
meanDepthError = 0
varDepthError = 0
else:
# compute false positives and negatives
falsePositives = np.sum(
(gt_outputs[k][:, :, 0] < out_img[:, :, 0]) * 1.0)
falseNegatives = np.sum(
(gt_outputs[k][:, :, 0] > out_img[:, :, 0]) * 1.0)
# compute mean depth error
mask = (out_img[:, :, 0] > 0) & (gt_outputs[k][:, :, 0]
> 0)
depthDiff = np.ma.masked_array(
np.abs(gt_outputs[k][:, :, 4] - out_img[:, :, 4]),
mask=mask)
meanDepthError = depthDiff.mean()
varDepthError = depthDiff.var()
t = timer.elapsed_ms()
summed_times[j] += t
f.write("%s\t%d\t%.4f\t%d\t%d\t%.4f\t%.4f\n" %
(human_kernel_name.replace(
"\n", " "), k, t, falsePositives,
falseNegatives, meanDepthError, varDepthError))
# write average stats
with open(
OUTPUT_STATS_AVG %
("double" if OUTPUT_STATS_USE_DOUBLE else "float"), "w") as f:
f.write("Kernel Name\tAvg-Time (ms)\n")
for j, (_, human_kernel_name,
_) in enumerate(KERNEL_NAMES_MEASURE):
f.write("%s\t%.4f\n" % (human_kernel_name.replace(
"\n", " "), summed_times[j] / TIMING_STEPS))
elif mode == "instrumentation":
# recompile with instrumentation
pyrenderer.reload_kernels(enableInstrumentation=True)
# allocate output
output = pyrenderer.allocate_output(rendererArgs.width,
rendererArgs.height,
rendererArgs.render_mode)
outputs = [None] * len(KERNEL_NAMES)
fields = [
"densityFetches-avg", "densityFetches-std", "ddaSteps-avg",
"ddaSteps-std", "intervalEval-avg", "intervalEval-std",
"intervalStep-avg", "intervalStep-std", "intervalMaxStep",
"timeDensityFetch-avg", "timeDensityFetch-std",
"timePolynomialCreation-avg", "timePolynomialCreation-std",
"timePolynomialSolution-avg", "timePolynomialSolution-std",
"timeTotal-avg", "timeTotal-std"
]
# render
with open(OUTPUT_INSTRUMENTATION, "w") as f:
f.write("Kernel Name\t%s\n" % "\t".join(fields))
camera.update_render_args(rendererArgs)
for j, (kernel_name, human_kernel_name,
stepsize) in enumerate(KERNEL_NAMES):
print("Render", kernel_name, stepsize)
rendererArgs.stepsize = stepsize
instrumentations = \
pyrenderer.render_with_instrumentation(
kernel_name, volume, rendererArgs, output)
densityFetches = avg_and_std(
instrumentations["densityFetches"])
ddaSteps = avg_and_std(instrumentations["ddaSteps"])
intervalEval = avg_and_std(instrumentations["intervalEval"])
intervalStep = weighted_avg_and_std(
instrumentations["intervalStep"],
instrumentations["intervalEval"])
intervalMaxStep = np.max(instrumentations["intervalMaxStep"])
timeDensityFetch = weighted_avg_and_std(
instrumentations["timeDensityFetch"],
instrumentations["timeDensityFetch_NumSamples"])
timePolynomialCreation = weighted_avg_and_std(
instrumentations["timePolynomialCreation"],
instrumentations["timePolynomialCreation_NumSamples"])
timePolynomialSolution = weighted_avg_and_std(
instrumentations["timePolynomialSolution"],
instrumentations["timePolynomialSolution_NumSamples"])
timeTotal = avg_and_std(instrumentations["timeTotal"])
f.write(
"%s\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%d\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\n"
% (human_kernel_name.replace("\n", " "), densityFetches[0],
densityFetches[1], ddaSteps[0], ddaSteps[1],
intervalEval[0], intervalEval[1], intervalStep[0],
intervalStep[1], intervalMaxStep, timeDensityFetch[0],
timeDensityFetch[1], timePolynomialCreation[0],
timePolynomialCreation[1], timePolynomialSolution[0],
timePolynomialSolution[1], timeTotal[0], timeTotal[1]))
f.flush()