// playground

Research Playground

1. Read the baseline code and task info. 2. Write your hypothesis — what to change and why. 3. Click Run and watch the agent iterate: propose, test, measure, revise.

starter tasks (free tier)

solve.c

solve.c (baseline)c

1#include "solve.h"
2#include <stdlib.h>
3#include <string.h>
4
5// Comparison function for qsort
6static int compare_uint32(const void *a, const void *b) {
7    uint32_t va = *(const uint32_t *)a;
8    uint32_t vb = *(const uint32_t *)b;
9    if (va < vb) return -1;
10    if (va > vb) return  1;
11    return 0;
12}
13
14// Baseline: stdlib qsort on 50M uint32 values
15// This is O(n log n) but with high constant factors
16// due to comparison function pointer overhead and
17// poor cache locality during partitioning.
18void radix_sort(uint32_t *data, size_t n) {
19    qsort(data, n, sizeof(uint32_t), compare_uint32);
20}

markdown

hypothesis

What is your core insight?

approach

What specific changes would you make to solve.c?

expected impact

What improvement do you expect on runtime?

guiding questions

1.Why is qsort slow for 50M integers despite being O(n log n)?
2.How many passes does an LSD radix sort need for 32-bit keys?
3.What radix (bits per pass) minimizes total cache misses?
4.How can you avoid branch mispredictions in the counting step?

solve.c (baseline)c

1#include "solve.h"
2#include <stdlib.h>
3#include <string.h>
4
5// Comparison function for qsort
6static int compare_uint32(const void *a, const void *b) {
7    uint32_t va = *(const uint32_t *)a;
8    uint32_t vb = *(const uint32_t *)b;
9    if (va < vb) return -1;
10    if (va > vb) return  1;
11    return 0;
12}
13
14// Baseline: stdlib qsort on 50M uint32 values
15// This is O(n log n) but with high constant factors
16// due to comparison function pointer overhead and
17// poor cache locality during partitioning.
18void radix_sort(uint32_t *data, size_t n) {
19    qsort(data, n, sizeof(uint32_t), compare_uint32);
20}

Free tier uses Haiku 4.5. Select a different model and bring your own API key for stronger results.