mirror of
https://github.com/KrisKennaway/ii-sound.git
synced 2024-06-09 18:29:32 +00:00
Checkpoint working code that implements delta modulation with look-ahead.
I split out preprocess_audio because scipy/librosa don't run under pypy and I was hoping that would speed up the encoding, but it does not for some reason (maybe numpy interactions?) Anyway, I should be able to vectorize the slow path which will likely help a lot.
This commit is contained in:
parent
d283789199
commit
0f8ad1bf17
151
encode_audio.py
Normal file
151
encode_audio.py
Normal file
|
@ -0,0 +1,151 @@
|
|||
import sys
|
||||
import functools
|
||||
import numpy
|
||||
import soundfile
|
||||
from typing import List
|
||||
|
||||
PORT = 1977
|
||||
|
||||
OPCODES = {
|
||||
'tick': 0x00,
|
||||
'notick_page1': 0x08,
|
||||
'notick_page2': 0x10,
|
||||
'exit': 0x18,
|
||||
'slowpath': 0x28
|
||||
}
|
||||
|
||||
# Errors
|
||||
# 1000 LAH 1 = 704632.752909
|
||||
|
||||
# 100 LAH 1 Norm 0.5 = 6783135.377118
|
||||
|
||||
# 100 LAH 1 Norm 0.3 = 3251821.285253
|
||||
# 100 LAH 2 Norm 0.3 = 6446570.565330
|
||||
|
||||
# 100 LAH 1 Norm 0.2 = 1798442.832489
|
||||
# 100 LAH 7 Norm 0.2 = 1624865.370107
|
||||
|
||||
# 10 LAH 1 Norm 0.5 = 3251142.161679
|
||||
# 20 LAH 1 Norm 0.5 = 3518009.983899
|
||||
# 30 ... = 4137179.898981
|
||||
|
||||
# 40 LAH 1 Norm 0.5 = 4667834.501476
|
||||
# 40 LAH 7 Norm 0.5 = 3614627.124544
|
||||
|
||||
# 50 ... = 5118520.178948
|
||||
# 60 ... = 5513928.786507
|
||||
# 70 ... = 5872060.623297
|
||||
# 80 ... = 6198922.754212
|
||||
# 90 ... = 6501755.264692
|
||||
# 100 ... = 6782164.670392
|
||||
# 200 = 8708489.044314
|
||||
# 300 = 9686398.703931
|
||||
# 400 = 10233308.826519
|
||||
# 500 = 10567966.087857
|
||||
|
||||
# TODO: synchronize to VBL
|
||||
# TODO: tick now has space to also flip a soft-switch
|
||||
|
||||
|
||||
# @profile
|
||||
def lookahead(step_size: int, initial_position: float,
|
||||
initial_voltage: float, data: numpy.ndarray, offset: int,
|
||||
toggles: List):
|
||||
|
||||
# TODO: construct list of voltage values directly
|
||||
# - construct voltage trajectories from position and vectorize comparison
|
||||
voltage = initial_voltage
|
||||
position = initial_position
|
||||
|
||||
total_error = 0.0
|
||||
for i, t in enumerate(toggles):
|
||||
target_val = data[i+offset]
|
||||
if t:
|
||||
voltage = -voltage
|
||||
position += (voltage - position) / step_size
|
||||
err = position - target_val
|
||||
total_error += abs(err)
|
||||
return total_error
|
||||
|
||||
|
||||
# TODO: test slowpath
|
||||
@functools.lru_cache(None)
|
||||
def lookahead_patterns(lookahead: int, slowpath: int):
|
||||
patterns = []
|
||||
|
||||
num_bits = max(lookahead - slowpath, 0)
|
||||
for i in range(2 ** num_bits):
|
||||
pattern = []
|
||||
for j in range(num_bits):
|
||||
pattern.append(bool((i >> j) & 1))
|
||||
pattern.extend([True for _ in range(min(slowpath, lookahead))])
|
||||
patterns.append(pattern)
|
||||
|
||||
return patterns
|
||||
|
||||
def sample(data: numpy.ndarray, step: int, lookahead_steps: int):
|
||||
dlen = len(data)
|
||||
data = numpy.concatenate([data, numpy.zeros(lookahead_steps)])
|
||||
|
||||
voltage = -1.0
|
||||
position = -1.0
|
||||
|
||||
total_err = 0.0
|
||||
slowpath = 0
|
||||
cnt = 0
|
||||
for i, val in enumerate(data[:dlen]):
|
||||
if i % int((dlen / 100)) == 0:
|
||||
print("%d%% complete" % (i * 100 / dlen))
|
||||
if (cnt % 2048) == 2047:
|
||||
slowpath = 12
|
||||
|
||||
min_err = 1e9
|
||||
best_pattern = None
|
||||
|
||||
# TODO: double-check this
|
||||
slowpath_bits = max((cnt % 2048) + lookahead_steps - 2047, 0)
|
||||
for lh in lookahead_patterns(lookahead_steps, slowpath_bits):
|
||||
err = lookahead(step, position, voltage, data, i, lh)
|
||||
# print(err, lh)
|
||||
if err < min_err:
|
||||
min_err = err
|
||||
best_pattern = lh
|
||||
# "BEST: %f, %s --> %s" % (err, lh, lh[0]))
|
||||
|
||||
if slowpath:
|
||||
if slowpath == 12:
|
||||
yield OPCODES['slowpath']
|
||||
cnt += 1
|
||||
slowpath -= 1
|
||||
elif best_pattern[0]:
|
||||
voltage = -voltage
|
||||
yield OPCODES['tick']
|
||||
cnt += 1
|
||||
else:
|
||||
yield OPCODES['notick_page1']
|
||||
cnt += 1
|
||||
position += (voltage - position) / step
|
||||
err = position - val
|
||||
total_err += abs(err)
|
||||
# print("State: ", sp.voltage, sp.position, val)
|
||||
|
||||
for _ in range(cnt % 2048, 2047):
|
||||
yield OPCODES['notick_page1']
|
||||
yield OPCODES['exit']
|
||||
print("Total error %f" % total_err)
|
||||
|
||||
|
||||
def main(argv):
|
||||
serve_file = argv[1]
|
||||
step = int(argv[2])
|
||||
lookahead_steps = int(argv[3])
|
||||
out = argv[4]
|
||||
|
||||
data, sr = soundfile.read(serve_file)
|
||||
with open(out, "wb+") as f:
|
||||
for b in sample(data, step, lookahead_steps):
|
||||
f.write(bytes([b]))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main(sys.argv)
|
|
@ -1,53 +1,14 @@
|
|||
import socketserver
|
||||
import sys
|
||||
import functools
|
||||
import random
|
||||
|
||||
import numpy
|
||||
from scipy.io import wavfile
|
||||
import scipy.signal as sps
|
||||
|
||||
PORT = 1977
|
||||
|
||||
OPCODES = {
|
||||
'tick': 0x00,
|
||||
'notick_page1': 0x08,
|
||||
'notick_page2': 0x10,
|
||||
'exit': 0x18,
|
||||
'slowpath': 0x28
|
||||
}
|
||||
|
||||
def serve(stream: bytes):
|
||||
cnt = 0
|
||||
state = False
|
||||
page = 0
|
||||
cycles = 0
|
||||
|
||||
direction = 100
|
||||
# TODO: synchronize to VBL
|
||||
# TODO: tick now has space to also flip a soft-switch
|
||||
# - i.e. we have 4 variants
|
||||
# can make sure we always flip at start/end of line
|
||||
for b in stream:
|
||||
for i in range(8):
|
||||
bit = (b >> i) & 0x1
|
||||
cycles += 13
|
||||
if bit != state:
|
||||
state = not state
|
||||
yield OPCODES['tick']
|
||||
else:
|
||||
# 18270
|
||||
if (cycles % 18800) < (13):
|
||||
yield OPCODES['notick_page1']
|
||||
else:
|
||||
yield OPCODES['notick_page2']
|
||||
page += 1
|
||||
cnt += 1
|
||||
if (cnt % 2048) == 2047:
|
||||
yield OPCODES['slowpath']
|
||||
cnt += 1
|
||||
cycles += 12*13
|
||||
# make sure we send enough data to fill player's receive buffer one last
|
||||
# time, so it doesn't loop forever
|
||||
for i in range(cnt % 2048, 2048):
|
||||
yield OPCODES['exit']
|
||||
|
||||
|
||||
def main(argv):
|
||||
serve_file = argv[1]
|
||||
|
||||
|
@ -56,8 +17,7 @@ def main(argv):
|
|||
def handle(self):
|
||||
with open(serve_file, "rb") as f:
|
||||
print("Sending...")
|
||||
for b in serve(f.read()):
|
||||
self.request.send(bytes([b]))
|
||||
self.request.sendall(f.read())
|
||||
|
||||
return ChunkHandler
|
||||
|
||||
|
|
27
preprocess_audio.py
Normal file
27
preprocess_audio.py
Normal file
|
@ -0,0 +1,27 @@
|
|||
import sys
|
||||
import librosa
|
||||
import numpy
|
||||
import soundfile as sf
|
||||
|
||||
|
||||
def preprocess(
|
||||
filename: str, target_sample_rate: int,
|
||||
normalize: float = 0.5) -> numpy.ndarray:
|
||||
data, _ = librosa.load(filename, sr=target_sample_rate, mono=True)
|
||||
|
||||
max_value = numpy.percentile(data, 90)
|
||||
data /= max_value
|
||||
data *= normalize
|
||||
|
||||
return data
|
||||
|
||||
def main(argv):
|
||||
serve_file = argv[1]
|
||||
out = argv[2]
|
||||
sample_rate = int(1024. * 1000 / 13)
|
||||
|
||||
sf.write(out, preprocess(serve_file, sample_rate), sample_rate)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main(sys.argv)
|
Loading…
Reference in New Issue
Block a user