File size: 6,167 Bytes
2166d7e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 | """
FlashAttention RTL - Test Data Generation & Utilities
Generates reference test vectors for hardware simulation
"""
import numpy as np
from pathlib import Path
import json
class TestDataGenerator:
"""
Generate realistic test data with known properties for RTL verification.
Supports multiple data types and numeric ranges.
"""
def __init__(self, seed: int = 42, dtype: np.dtype = np.float32):
self.seed = seed
self.dtype = dtype
np.random.seed(seed)
def generate_attention_inputs(
self,
seq_len: int,
d_model: int,
scale: float = None,
low: float = -1.0,
high: float = 1.0
) -> dict:
"""
Generate Q, K, V inputs for attention computation.
Args:
seq_len: Sequence length N
d_model: Embedding dimension
scale: Optional scaling factor (default: 1/sqrt(d_model))
low, high: Range for random values
Returns:
dict with Q, K, V arrays
"""
if scale is None:
scale = 1.0 / np.sqrt(d_model)
Q = np.random.uniform(low, high, (seq_len, d_model)).astype(self.dtype)
K = np.random.uniform(low, high, (seq_len, d_model)).astype(self.dtype)
V = np.random.uniform(low, high, (seq_len, d_model)).astype(self.dtype)
return {
'Q': Q,
'K': K,
'V': V,
'scale': scale,
'seq_len': seq_len,
'd_model': d_model
}
def generate_fixed_point_inputs(
self,
seq_len: int,
d_model: int,
int_bits: int = 4,
frac_bits: int = 4
) -> dict:
"""
Generate test data for fixed-point arithmetic (Q4.4 format).
Args:
seq_len, d_model: Dimensions
int_bits: Number of integer bits
frac_bits: Number of fractional bits
Returns:
dict with fixed-point Q, K, V
"""
# Fixed-point range: [-2^(int_bits-1), 2^(int_bits-1) - 1] / 2^frac_bits
max_val = (2 ** (int_bits - 1) - 1) / (2 ** frac_bits)
min_val = -(2 ** (int_bits - 1)) / (2 ** frac_bits)
Q = np.random.uniform(min_val, max_val, (seq_len, d_model)).astype(self.dtype)
K = np.random.uniform(min_val, max_val, (seq_len, d_model)).astype(self.dtype)
V = np.random.uniform(min_val, max_val, (seq_len, d_model)).astype(self.dtype)
return {
'Q': Q,
'K': K,
'V': V,
'int_bits': int_bits,
'frac_bits': frac_bits,
'seq_len': seq_len,
'd_model': d_model
}
def save_test_vectors(self, data: dict, filename: str):
"""Save test vectors to NPZ file for RTL simulation."""
output_path = Path("../sim") / filename
output_path.parent.mkdir(parents=True, exist_ok=True)
# Convert to float64 for higher precision in storage
data_f64 = {k: v.astype(np.float64) if isinstance(v, np.ndarray) else v
for k, v in data.items()}
np.savez_compressed(output_path, **data_f64)
print(f"✓ Saved test vectors: {output_path}")
def generate_golden_reference(
self,
Q: np.ndarray,
K: np.ndarray,
V: np.ndarray,
scale: float
) -> np.ndarray:
"""
Generate golden reference output using naive attention.
Args:
Q, K, V: Input matrices
scale: Attention scaling factor
Returns:
Output matrix (N, d)
"""
scores = Q @ K.T * scale
scores_shifted = scores - np.max(scores, axis=-1, keepdims=True)
attn_weights = np.exp(scores_shifted)
attn_weights = attn_weights / np.sum(attn_weights, axis=-1, keepdims=True)
output = attn_weights @ V
return output.astype(self.dtype)
def generate_standard_test_suite():
"""
Generate a standard test suite for RTL verification.
Creates multiple test cases with different characteristics.
"""
print("\n" + "=" * 70)
print("Generating Standard Test Suite for RTL Verification")
print("=" * 70 + "\n")
test_cases = [
{'name': 'tiny', 'seq_len': 8, 'd_model': 16, 'dtype': np.float32},
{'name': 'small', 'seq_len': 32, 'd_model': 32, 'dtype': np.float32},
{'name': 'medium', 'seq_len': 64, 'd_model': 64, 'dtype': np.float32},
{'name': 'large', 'seq_len': 128, 'd_model': 128, 'dtype': np.float32},
]
for i, config in enumerate(test_cases):
print(f"Test {i+1}/{len(test_cases)}: {config['name']}")
print(f" Seq Len: {config['seq_len']}, D Model: {config['d_model']}")
generator = TestDataGenerator(seed=42, dtype=config['dtype'])
# Generate floating-point test vectors
inputs_fp = generator.generate_attention_inputs(
config['seq_len'],
config['d_model']
)
# Generate reference output
output = generator.generate_golden_reference(
inputs_fp['Q'],
inputs_fp['K'],
inputs_fp['V'],
inputs_fp['scale']
)
inputs_fp['output_golden'] = output
# Save
generator.save_test_vectors(inputs_fp, f"test_vectors_{config['name']}_fp32.npz")
# Also generate fixed-point versions
inputs_fp8 = generator.generate_fixed_point_inputs(
config['seq_len'],
config['d_model'],
int_bits=4,
frac_bits=4
)
inputs_fp8['output_golden'] = output # Use same golden reference
generator.save_test_vectors(inputs_fp8, f"test_vectors_{config['name']}_fp8.npz")
print(f" ✓ Generated FP32 and FP8 test vectors\n")
print("=" * 70)
print("Test Suite Generation Complete!")
print("=" * 70 + "\n")
if __name__ == "__main__":
generate_standard_test_suite()
|