Delete test_overflow_chains.py with huggingface_hub
Browse files- test_overflow_chains.py +0 -423
test_overflow_chains.py
DELETED
|
@@ -1,423 +0,0 @@
|
|
| 1 |
-
"""
|
| 2 |
-
TEST #1: Arithmetic Overflow Chains
|
| 3 |
-
====================================
|
| 4 |
-
Chains 1000+ arithmetic operations, verifying every intermediate state.
|
| 5 |
-
Tests carry/borrow propagation across long sequences, not just single ops.
|
| 6 |
-
|
| 7 |
-
A skeptic would demand: "Prove your adder doesn't accumulate errors over
|
| 8 |
-
repeated use. Show me every intermediate value matches Python's arithmetic."
|
| 9 |
-
"""
|
| 10 |
-
|
| 11 |
-
import torch
|
| 12 |
-
from safetensors.torch import load_file
|
| 13 |
-
import random
|
| 14 |
-
|
| 15 |
-
# Load circuits
|
| 16 |
-
model = load_file('neural_computer.safetensors')
|
| 17 |
-
|
| 18 |
-
def heaviside(x):
|
| 19 |
-
return (x >= 0).float()
|
| 20 |
-
|
| 21 |
-
def int_to_bits_lsb(val, width=8):
|
| 22 |
-
"""Convert int to bits, LSB first (for arithmetic)."""
|
| 23 |
-
return torch.tensor([(val >> i) & 1 for i in range(width)], dtype=torch.float32)
|
| 24 |
-
|
| 25 |
-
def bits_to_int_lsb(bits):
|
| 26 |
-
"""Convert bits back to int, LSB first."""
|
| 27 |
-
return sum(int(bits[i].item()) * (2**i) for i in range(len(bits)))
|
| 28 |
-
|
| 29 |
-
def eval_xor(a, b, prefix='boolean.xor'):
|
| 30 |
-
"""Evaluate XOR gate."""
|
| 31 |
-
inp = torch.tensor([a, b], dtype=torch.float32)
|
| 32 |
-
w1_n1 = model[f'{prefix}.layer1.neuron1.weight']
|
| 33 |
-
b1_n1 = model[f'{prefix}.layer1.neuron1.bias']
|
| 34 |
-
w1_n2 = model[f'{prefix}.layer1.neuron2.weight']
|
| 35 |
-
b1_n2 = model[f'{prefix}.layer1.neuron2.bias']
|
| 36 |
-
w2 = model[f'{prefix}.layer2.weight']
|
| 37 |
-
b2 = model[f'{prefix}.layer2.bias']
|
| 38 |
-
h1 = heaviside(inp @ w1_n1 + b1_n1)
|
| 39 |
-
h2 = heaviside(inp @ w1_n2 + b1_n2)
|
| 40 |
-
hidden = torch.tensor([h1.item(), h2.item()])
|
| 41 |
-
return heaviside(hidden @ w2 + b2).item()
|
| 42 |
-
|
| 43 |
-
def eval_xor_arith(inp, prefix):
|
| 44 |
-
"""Evaluate XOR for arithmetic circuits (different naming)."""
|
| 45 |
-
w1_or = model[f'{prefix}.layer1.or.weight']
|
| 46 |
-
b1_or = model[f'{prefix}.layer1.or.bias']
|
| 47 |
-
w1_nand = model[f'{prefix}.layer1.nand.weight']
|
| 48 |
-
b1_nand = model[f'{prefix}.layer1.nand.bias']
|
| 49 |
-
w2 = model[f'{prefix}.layer2.weight']
|
| 50 |
-
b2 = model[f'{prefix}.layer2.bias']
|
| 51 |
-
h_or = heaviside(inp @ w1_or + b1_or)
|
| 52 |
-
h_nand = heaviside(inp @ w1_nand + b1_nand)
|
| 53 |
-
hidden = torch.tensor([h_or.item(), h_nand.item()])
|
| 54 |
-
return heaviside(hidden @ w2 + b2).item()
|
| 55 |
-
|
| 56 |
-
def eval_full_adder(a, b, cin, prefix):
|
| 57 |
-
"""Evaluate full adder, return (sum, carry_out)."""
|
| 58 |
-
inp_ab = torch.tensor([a, b], dtype=torch.float32)
|
| 59 |
-
|
| 60 |
-
# HA1: a XOR b
|
| 61 |
-
ha1_sum = eval_xor_arith(inp_ab, f'{prefix}.ha1.sum')
|
| 62 |
-
|
| 63 |
-
# HA1 carry: a AND b
|
| 64 |
-
w_c1 = model[f'{prefix}.ha1.carry.weight']
|
| 65 |
-
b_c1 = model[f'{prefix}.ha1.carry.bias']
|
| 66 |
-
ha1_carry = heaviside(inp_ab @ w_c1 + b_c1).item()
|
| 67 |
-
|
| 68 |
-
# HA2: ha1_sum XOR cin
|
| 69 |
-
inp_ha2 = torch.tensor([ha1_sum, cin], dtype=torch.float32)
|
| 70 |
-
ha2_sum = eval_xor_arith(inp_ha2, f'{prefix}.ha2.sum')
|
| 71 |
-
|
| 72 |
-
# HA2 carry
|
| 73 |
-
w_c2 = model[f'{prefix}.ha2.carry.weight']
|
| 74 |
-
b_c2 = model[f'{prefix}.ha2.carry.bias']
|
| 75 |
-
ha2_carry = heaviside(inp_ha2 @ w_c2 + b_c2).item()
|
| 76 |
-
|
| 77 |
-
# Carry out = ha1_carry OR ha2_carry
|
| 78 |
-
inp_cout = torch.tensor([ha1_carry, ha2_carry], dtype=torch.float32)
|
| 79 |
-
w_or = model[f'{prefix}.carry_or.weight']
|
| 80 |
-
b_or = model[f'{prefix}.carry_or.bias']
|
| 81 |
-
cout = heaviside(inp_cout @ w_or + b_or).item()
|
| 82 |
-
|
| 83 |
-
return int(ha2_sum), int(cout)
|
| 84 |
-
|
| 85 |
-
def add_8bit(a, b):
|
| 86 |
-
"""8-bit addition using ripple carry adder. Returns (result, carry)."""
|
| 87 |
-
carry = 0.0
|
| 88 |
-
result_bits = []
|
| 89 |
-
|
| 90 |
-
for i in range(8):
|
| 91 |
-
a_bit = (a >> i) & 1
|
| 92 |
-
b_bit = (b >> i) & 1
|
| 93 |
-
s, carry = eval_full_adder(float(a_bit), float(b_bit), carry,
|
| 94 |
-
f'arithmetic.ripplecarry8bit.fa{i}')
|
| 95 |
-
result_bits.append(s)
|
| 96 |
-
|
| 97 |
-
result = sum(result_bits[i] * (2**i) for i in range(8))
|
| 98 |
-
return result, int(carry)
|
| 99 |
-
|
| 100 |
-
def sub_8bit(a, b):
|
| 101 |
-
"""8-bit subtraction via two's complement: a - b = a + (~b) + 1."""
|
| 102 |
-
not_b = (~b) & 0xFF
|
| 103 |
-
temp, c1 = add_8bit(a, not_b)
|
| 104 |
-
result, c2 = add_8bit(temp, 1)
|
| 105 |
-
return result, c1 | c2
|
| 106 |
-
|
| 107 |
-
# =============================================================================
|
| 108 |
-
# TEST CHAINS
|
| 109 |
-
# =============================================================================
|
| 110 |
-
|
| 111 |
-
def test_chain_add_overflow():
|
| 112 |
-
"""
|
| 113 |
-
Start at 0, add 1 repeatedly until we wrap around multiple times.
|
| 114 |
-
Verify every single intermediate value.
|
| 115 |
-
"""
|
| 116 |
-
print("\n[TEST 1] Add-1 chain: 0 -> 255 -> 0 -> 255 (512 additions)")
|
| 117 |
-
print("-" * 60)
|
| 118 |
-
|
| 119 |
-
value = 0
|
| 120 |
-
errors = []
|
| 121 |
-
|
| 122 |
-
for i in range(512):
|
| 123 |
-
expected = (value + 1) % 256
|
| 124 |
-
result, carry = add_8bit(value, 1)
|
| 125 |
-
|
| 126 |
-
if result != expected:
|
| 127 |
-
errors.append((i, value, 1, expected, result))
|
| 128 |
-
|
| 129 |
-
# Check carry on overflow
|
| 130 |
-
if value == 255 and carry != 1:
|
| 131 |
-
errors.append((i, value, 1, "carry=1", f"carry={carry}"))
|
| 132 |
-
|
| 133 |
-
value = result
|
| 134 |
-
|
| 135 |
-
if errors:
|
| 136 |
-
print(f" FAILED: {len(errors)} errors")
|
| 137 |
-
for e in errors[:5]:
|
| 138 |
-
print(f" Step {e[0]}: {e[1]} + {e[2]} = {e[4]}, expected {e[3]}")
|
| 139 |
-
else:
|
| 140 |
-
print(f" PASSED: 512 additions, 2 full wraparounds verified")
|
| 141 |
-
|
| 142 |
-
return len(errors) == 0
|
| 143 |
-
|
| 144 |
-
def test_chain_sub_overflow():
|
| 145 |
-
"""
|
| 146 |
-
Start at 255, subtract 1 repeatedly until we wrap around.
|
| 147 |
-
"""
|
| 148 |
-
print("\n[TEST 2] Sub-1 chain: 255 -> 0 -> 255 (512 subtractions)")
|
| 149 |
-
print("-" * 60)
|
| 150 |
-
|
| 151 |
-
value = 255
|
| 152 |
-
errors = []
|
| 153 |
-
|
| 154 |
-
for i in range(512):
|
| 155 |
-
expected = (value - 1) % 256
|
| 156 |
-
result, _ = sub_8bit(value, 1)
|
| 157 |
-
|
| 158 |
-
if result != expected:
|
| 159 |
-
errors.append((i, value, 1, expected, result))
|
| 160 |
-
|
| 161 |
-
value = result
|
| 162 |
-
|
| 163 |
-
if errors:
|
| 164 |
-
print(f" FAILED: {len(errors)} errors")
|
| 165 |
-
for e in errors[:5]:
|
| 166 |
-
print(f" Step {e[0]}: {e[1]} - {e[2]} = {e[4]}, expected {e[3]}")
|
| 167 |
-
else:
|
| 168 |
-
print(f" PASSED: 512 subtractions verified")
|
| 169 |
-
|
| 170 |
-
return len(errors) == 0
|
| 171 |
-
|
| 172 |
-
def test_chain_mixed():
|
| 173 |
-
"""
|
| 174 |
-
Random mix of +1, -1, +k, -k operations. Verify all intermediates.
|
| 175 |
-
"""
|
| 176 |
-
print("\n[TEST 3] Mixed chain: 1000 random +/- operations")
|
| 177 |
-
print("-" * 60)
|
| 178 |
-
|
| 179 |
-
random.seed(42) # Reproducible
|
| 180 |
-
|
| 181 |
-
value = 128 # Start in middle
|
| 182 |
-
python_value = 128
|
| 183 |
-
errors = []
|
| 184 |
-
|
| 185 |
-
for i in range(1000):
|
| 186 |
-
op = random.choice(['+1', '-1', '+k', '-k'])
|
| 187 |
-
|
| 188 |
-
if op == '+1':
|
| 189 |
-
result, _ = add_8bit(value, 1)
|
| 190 |
-
python_value = (python_value + 1) % 256
|
| 191 |
-
elif op == '-1':
|
| 192 |
-
result, _ = sub_8bit(value, 1)
|
| 193 |
-
python_value = (python_value - 1) % 256
|
| 194 |
-
elif op == '+k':
|
| 195 |
-
k = random.randint(1, 50)
|
| 196 |
-
result, _ = add_8bit(value, k)
|
| 197 |
-
python_value = (python_value + k) % 256
|
| 198 |
-
else: # '-k'
|
| 199 |
-
k = random.randint(1, 50)
|
| 200 |
-
result, _ = sub_8bit(value, k)
|
| 201 |
-
python_value = (python_value - k) % 256
|
| 202 |
-
|
| 203 |
-
if result != python_value:
|
| 204 |
-
errors.append((i, op, value, python_value, result))
|
| 205 |
-
|
| 206 |
-
value = result
|
| 207 |
-
|
| 208 |
-
if errors:
|
| 209 |
-
print(f" FAILED: {len(errors)} errors")
|
| 210 |
-
for e in errors[:5]:
|
| 211 |
-
print(f" Step {e[0]}: {e[1]} on {e[2]} = {e[4]}, expected {e[3]}")
|
| 212 |
-
else:
|
| 213 |
-
print(f" PASSED: 1000 random ops verified")
|
| 214 |
-
|
| 215 |
-
return len(errors) == 0
|
| 216 |
-
|
| 217 |
-
def test_chain_carry_stress():
|
| 218 |
-
"""
|
| 219 |
-
Worst-case carry propagation: repeatedly compute 127+128=255, 255+1=0.
|
| 220 |
-
"""
|
| 221 |
-
print("\n[TEST 4] Carry stress: 127+128 and 255+1 chains (500 each)")
|
| 222 |
-
print("-" * 60)
|
| 223 |
-
|
| 224 |
-
errors = []
|
| 225 |
-
|
| 226 |
-
# 127 + 128 = 255 (all bits flip via carry)
|
| 227 |
-
for i in range(500):
|
| 228 |
-
result, carry = add_8bit(127, 128)
|
| 229 |
-
if result != 255:
|
| 230 |
-
errors.append((i, '127+128', 255, result))
|
| 231 |
-
|
| 232 |
-
# 255 + 1 = 0 with carry out (8-bit carry chain)
|
| 233 |
-
for i in range(500):
|
| 234 |
-
result, carry = add_8bit(255, 1)
|
| 235 |
-
if result != 0 or carry != 1:
|
| 236 |
-
errors.append((i, '255+1', '0,c=1', f'{result},c={carry}'))
|
| 237 |
-
|
| 238 |
-
if errors:
|
| 239 |
-
print(f" FAILED: {len(errors)} errors")
|
| 240 |
-
for e in errors[:5]:
|
| 241 |
-
print(f" Iteration {e[0]}: {e[1]} = {e[3]}, expected {e[2]}")
|
| 242 |
-
else:
|
| 243 |
-
print(f" PASSED: 1000 worst-case carry operations")
|
| 244 |
-
|
| 245 |
-
return len(errors) == 0
|
| 246 |
-
|
| 247 |
-
def test_chain_accumulator():
|
| 248 |
-
"""
|
| 249 |
-
Accumulate: start at 0, add 1,2,3,...,100. Verify running sum at each step.
|
| 250 |
-
"""
|
| 251 |
-
print("\n[TEST 5] Accumulator: sum(1..100) with intermediate verification")
|
| 252 |
-
print("-" * 60)
|
| 253 |
-
|
| 254 |
-
acc = 0
|
| 255 |
-
errors = []
|
| 256 |
-
|
| 257 |
-
for i in range(1, 101):
|
| 258 |
-
result, _ = add_8bit(acc, i)
|
| 259 |
-
expected = (acc + i) % 256
|
| 260 |
-
|
| 261 |
-
if result != expected:
|
| 262 |
-
errors.append((i, acc, i, expected, result))
|
| 263 |
-
|
| 264 |
-
acc = result
|
| 265 |
-
|
| 266 |
-
# Final value: sum(1..100) = 5050, mod 256 = 5050 % 256 = 186
|
| 267 |
-
final_expected = sum(range(1, 101)) % 256
|
| 268 |
-
|
| 269 |
-
if acc != final_expected:
|
| 270 |
-
errors.append(('final', acc, final_expected))
|
| 271 |
-
|
| 272 |
-
if errors:
|
| 273 |
-
print(f" FAILED: {len(errors)} errors")
|
| 274 |
-
for e in errors[:5]:
|
| 275 |
-
print(f" {e}")
|
| 276 |
-
else:
|
| 277 |
-
print(f" PASSED: sum(1..100) mod 256 = {acc} verified at every step")
|
| 278 |
-
|
| 279 |
-
return len(errors) == 0
|
| 280 |
-
|
| 281 |
-
def test_chain_fibonacci():
|
| 282 |
-
"""
|
| 283 |
-
Compute Fibonacci sequence mod 256. Verify against Python.
|
| 284 |
-
"""
|
| 285 |
-
print("\n[TEST 6] Fibonacci chain: F(0)..F(100) mod 256")
|
| 286 |
-
print("-" * 60)
|
| 287 |
-
|
| 288 |
-
a, b = 0, 1 # Circuit values
|
| 289 |
-
pa, pb = 0, 1 # Python values
|
| 290 |
-
errors = []
|
| 291 |
-
|
| 292 |
-
for i in range(100):
|
| 293 |
-
# Verify current values
|
| 294 |
-
if a != pa:
|
| 295 |
-
errors.append((i, 'a', pa, a))
|
| 296 |
-
if b != pb:
|
| 297 |
-
errors.append((i, 'b', pb, b))
|
| 298 |
-
|
| 299 |
-
# Compute next
|
| 300 |
-
next_val, _ = add_8bit(a, b)
|
| 301 |
-
next_python = (pa + pb) % 256
|
| 302 |
-
|
| 303 |
-
a, b = b, next_val
|
| 304 |
-
pa, pb = pb, next_python
|
| 305 |
-
|
| 306 |
-
if errors:
|
| 307 |
-
print(f" FAILED: {len(errors)} errors")
|
| 308 |
-
for e in errors[:5]:
|
| 309 |
-
print(f" F({e[0]}) {e[1]}: expected {e[2]}, got {e[3]}")
|
| 310 |
-
else:
|
| 311 |
-
print(f" PASSED: 100 Fibonacci terms verified")
|
| 312 |
-
|
| 313 |
-
return len(errors) == 0
|
| 314 |
-
|
| 315 |
-
def test_chain_alternating():
|
| 316 |
-
"""
|
| 317 |
-
Alternating +127/-127 to stress positive/negative boundaries.
|
| 318 |
-
"""
|
| 319 |
-
print("\n[TEST 7] Alternating +127/-127 (200 operations)")
|
| 320 |
-
print("-" * 60)
|
| 321 |
-
|
| 322 |
-
value = 0
|
| 323 |
-
python_value = 0
|
| 324 |
-
errors = []
|
| 325 |
-
|
| 326 |
-
for i in range(200):
|
| 327 |
-
if i % 2 == 0:
|
| 328 |
-
result, _ = add_8bit(value, 127)
|
| 329 |
-
python_value = (python_value + 127) % 256
|
| 330 |
-
else:
|
| 331 |
-
result, _ = sub_8bit(value, 127)
|
| 332 |
-
python_value = (python_value - 127) % 256
|
| 333 |
-
|
| 334 |
-
if result != python_value:
|
| 335 |
-
errors.append((i, value, python_value, result))
|
| 336 |
-
|
| 337 |
-
value = result
|
| 338 |
-
|
| 339 |
-
if errors:
|
| 340 |
-
print(f" FAILED: {len(errors)} errors")
|
| 341 |
-
for e in errors[:5]:
|
| 342 |
-
print(f" Step {e[0]}: from {e[1]}, expected {e[2]}, got {e[3]}")
|
| 343 |
-
else:
|
| 344 |
-
print(f" PASSED: 200 alternating ops verified")
|
| 345 |
-
|
| 346 |
-
return len(errors) == 0
|
| 347 |
-
|
| 348 |
-
def test_chain_powers_of_two():
|
| 349 |
-
"""
|
| 350 |
-
Add powers of 2: 1+2+4+8+...+128. Verify intermediate sums.
|
| 351 |
-
"""
|
| 352 |
-
print("\n[TEST 8] Powers of 2: 1+2+4+8+16+32+64+128")
|
| 353 |
-
print("-" * 60)
|
| 354 |
-
|
| 355 |
-
acc = 0
|
| 356 |
-
errors = []
|
| 357 |
-
|
| 358 |
-
for i in range(8):
|
| 359 |
-
power = 2 ** i
|
| 360 |
-
result, _ = add_8bit(acc, power)
|
| 361 |
-
expected = (acc + power) % 256
|
| 362 |
-
|
| 363 |
-
if result != expected:
|
| 364 |
-
errors.append((i, acc, power, expected, result))
|
| 365 |
-
|
| 366 |
-
acc = result
|
| 367 |
-
|
| 368 |
-
# Final: 1+2+4+8+16+32+64+128 = 255
|
| 369 |
-
if acc != 255:
|
| 370 |
-
errors.append(('final', 255, acc))
|
| 371 |
-
|
| 372 |
-
if errors:
|
| 373 |
-
print(f" FAILED: {len(errors)} errors")
|
| 374 |
-
for e in errors[:5]:
|
| 375 |
-
print(f" {e}")
|
| 376 |
-
else:
|
| 377 |
-
print(f" PASSED: 2^0 + 2^1 + ... + 2^7 = {acc}")
|
| 378 |
-
|
| 379 |
-
return len(errors) == 0
|
| 380 |
-
|
| 381 |
-
# =============================================================================
|
| 382 |
-
# MAIN
|
| 383 |
-
# =============================================================================
|
| 384 |
-
|
| 385 |
-
if __name__ == "__main__":
|
| 386 |
-
print("=" * 70)
|
| 387 |
-
print(" TEST #1: ARITHMETIC OVERFLOW CHAINS")
|
| 388 |
-
print(" Verifying every intermediate state across 3000+ chained operations")
|
| 389 |
-
print("=" * 70)
|
| 390 |
-
|
| 391 |
-
results = []
|
| 392 |
-
|
| 393 |
-
results.append(("Add-1 chain", test_chain_add_overflow()))
|
| 394 |
-
results.append(("Sub-1 chain", test_chain_sub_overflow()))
|
| 395 |
-
results.append(("Mixed random", test_chain_mixed()))
|
| 396 |
-
results.append(("Carry stress", test_chain_carry_stress()))
|
| 397 |
-
results.append(("Accumulator", test_chain_accumulator()))
|
| 398 |
-
results.append(("Fibonacci", test_chain_fibonacci()))
|
| 399 |
-
results.append(("Alternating", test_chain_alternating()))
|
| 400 |
-
results.append(("Powers of 2", test_chain_powers_of_two()))
|
| 401 |
-
|
| 402 |
-
print("\n" + "=" * 70)
|
| 403 |
-
print(" SUMMARY")
|
| 404 |
-
print("=" * 70)
|
| 405 |
-
|
| 406 |
-
passed = sum(1 for _, r in results if r)
|
| 407 |
-
total = len(results)
|
| 408 |
-
|
| 409 |
-
for name, r in results:
|
| 410 |
-
status = "PASS" if r else "FAIL"
|
| 411 |
-
print(f" {name:20s} [{status}]")
|
| 412 |
-
|
| 413 |
-
print(f"\n Total: {passed}/{total} tests passed")
|
| 414 |
-
|
| 415 |
-
total_ops = 512 + 512 + 1000 + 1000 + 100 + 100 + 200 + 8 # ~3400
|
| 416 |
-
print(f" Operations verified: ~{total_ops}")
|
| 417 |
-
|
| 418 |
-
if passed == total:
|
| 419 |
-
print("\n STATUS: ALL CHAINS VERIFIED - NO ACCUMULATED ERRORS")
|
| 420 |
-
else:
|
| 421 |
-
print("\n STATUS: FAILURES DETECTED")
|
| 422 |
-
|
| 423 |
-
print("=" * 70)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|