Study Plan

Shared-Memory Bank Conflict Analyzer

GPU Programming

Easy

Analyze shared-memory bank conflicts for a group of active lanes. For lane $i$ , $a_i$ is its byte address, $w$ is bank_width_bytes, $B$ is num_banks, and $b_i$ is its bank ID:

b_i=\left\lfloor\frac{a_i}{w}\right\rfloor\bmod B.

Floor rounds down, and modulo gives the remainder. In this model, a lane's conflict degree is the number of distinct addresses requested from its bank. Identical addresses share one broadcast and count once. Return a dictionary containing bank_ids and conflict_degree as one-dimensional NumPy int64 arrays, each matching the input length and lane order.

Loading visualization...

Examples

Input: byte_addresses = [0,4,8,12,16,20,24,28], num_banks = 32, bank_width_bytes = 4

Output: {"bank_ids":[0,1,2,3,4,5,6,7],"conflict_degree":[1,1,1,1,1,1,1,1]}

Explanation: Each four-byte word maps to a different bank, so every lane has conflict degree one.

Input: byte_addresses = [0,128,256,384], num_banks = 32, bank_width_bytes = 4

Output: {"bank_ids":[0,0,0,0],"conflict_degree":[4,4,4,4]}

Input: byte_addresses = [64,64,64,64], num_banks = 32, bank_width_bytes = 4

Output: {"bank_ids":[16,16,16,16],"conflict_degree":[1,1,1,1]}

Hint 1

np.unique finds distinct values in an array.

Hint 2

A Boolean mask selects the lanes belonging to one bank.

Requirements

Use NumPy and exact integer arithmetic.
Count repeated requests to the same address only once per bank.
Do not modify the input array.

Constraints

byte_addresses is a one-dimensional NumPy int64 array with 1 to 32 entries.
Every byte address is nonnegative.
num_banks and bank_width_bytes are positive Python integers, defaulting to 32 and 4.
The address array may be noncontiguous.

Try Similar Problems

Global-Memory Coalescing CounterEasy GPU Occupancy CalculatorMedium Measure Kernel Timeline UtilizationMedium Orthogonalize a Muon Momentum UpdateEasy Perform Stable Exact DeduplicationEasy Patchify a Batch of ImagesEasy Mixed-Precision Training Memory AccountantEasy Account for ZeRO Memory by StageEasy Implement Symmetric INT8 QuantizationEasy Estimate Model Memory and Minimum GPU CountEasy Implement Per-Channel Weight QuantizationMedium Analyze Prefill and Decode with the Roofline ModelMedium