Input data must have length TP_LEN × 4 samples for each vector (P and Q), regardless of TP_DIM value. The hardware processes data in 4-element groups, using only the first TP_DIM elements for calculations.
Configuration Example:
TP_LEN = 8, TP_DIM = 3, TP_IS_OUTPUT_SQUARED = 1 in_P[32] = [1,2,3,4, 5,6,7,8, 9,10,11,12, ..., 29,30,31,32] in_Q[32] = [32,31,30,29, 28,27,26,25, 24,23,22,21, ..., 4,3,2,1]
Data Organization:
4-Element Groups (TP_LEN=8):
P0:{1,2,3,4} P1:{5,6,7,8} P2:{9,10,11,12} ... P7:{29,30,31,32}
Q0:{32,31,30,29} Q1:{28,27,26,25} Q2:{24,23,22,21} ... Q7:{4,3,2,1}
When TP_DIM=3, only first 3 elements are used:
P0→{1,2,3} Q0→{32,31,30} (4th element ignored)
Data Fetching and Processing Pattern
The algorithm processes data in chunks of 4 elements (FIXED_DIM = 4) from each input vector. For TP_LEN = 8, this results in 8 processing iterations, where each iteration compares corresponding 4-element groups from vectors P and Q.
Memory Layout and Access Pattern
Memory Layout (FIXED_DIM = 4):
Vector P Memory:
[1][2][3][4] [5][6][7][8] [9][10][11][12] ... [29][30][31][32]
P0 P1 P2 P7
Vector Q Memory:
[32][31][30][29] [28][27][26][25] [24][23][22][21] ... [4][3][2][1]
Q0 Q1 Q2 Q7
Processing Flow:
Iteration 1: P0{1,2,3,4} <--> Q0{32,31,30,29}
Iteration 2: P1{5,6,7,8} <--> Q1{28,27,26,25}
...
Iteration 8: P7{29,30,31,32} <--> Q7{4,3,2,1}
Processing Algorithm (TP_DIM = 3 Example)
For each iteration, only the first TP_DIM elements from each 4-element group are used in calculations. When TP_DIM = 3, the 4th element in each group is ignored. The process repeats for all TP_LEN vector pairs (8 iterations in this example).
Calculation Process:
For each iteration i (0 to 7):
1. Extract TP_DIM elements: Pi[0:2] and Qi[0:2]
2. Compute differences: Diff = Qi - Pi (element-wise)
3. Square differences: Squared = Diff² (element-wise)
4. Sum squared differences: Sum = Σ(Squared)
5. Apply square root: Distance = √Sum (if TP_IS_OUTPUT_SQUARED = 0)
Concrete Examples:
D0: P0{1,2,3} vs Q0{32,31,30} → {31²+29²+27²} → 2531 → √2531 = 50.31
D1: P1{5,6,7} vs Q1{28,27,26} → {23²+21²+19²} → 1331 → √1331 = 36.48
D2: P2{9,10,11} vs Q2{24,23,22} → {15²+13²+11²} → 515 → √515 = 22.69
...
D7: P7{29,30,31} vs Q7{4,3,2} → {-25²,-27²,-29²} → 2195 → √2195 = 46.85
Key Points:
- Element Selection: Only first TP_DIM elements used (4th element ignored when TP_DIM = 3)
- Vectorized Operations: Differences, squaring, and summation use hardware vectorization
- Output Control: TP_IS_OUTPUT_SQUARED determines whether to apply square root
Important
Critical Memory Requirement: ED expects input data of P and Q to have TP_LEN × 4 samples, regardless of the user’s dimension (i.e., TP_DIM).
This means:
- Always allocate memory for 4 elements per vector group
- Even if TP_DIM = 1, 2, or 3, you still need 4 elements per group
- Total memory required: TP_LEN × 4 samples for each input vector
- Unused elements (when TP_DIM < 4) are ignored during calculation