Add challenge 77: Integral Image (Medium)

[claude]

Summary

• Adds challenge 77: Integral Image (Summed Area Table), a medium-difficulty GPU programming problem
• Solvers compute output[i][j] = sum of input[0..i, 0..j] for an H×W float32 image
• The interesting GPU challenge: row-wise prefix scans are coalesced and embarrassingly parallel, but the column-wise scan pass has strided (non-coalesced) memory access in row-major layout — motivating the classic shared-memory transpose trick
• Validated against all functional tests and performance test on NVIDIA Tesla T4 ✓

What makes this interesting

This challenge requires thinking about:

1. 2D prefix scan decomposition — two independent passes (rows then columns)
2. Memory coalescing — row scan is fast (coalesced), column scan is slow (strided) without optimization
3. Shared memory tiling — the standard fix is to load tiles transposed into shared memory, scan, then write back
4. Real-world relevance — integral images underpin Viola-Jones face detection, fast box filtering, ambient occlusion, and more

Test plan

• All 10 functional tests pass (edge cases, powers-of-2, non-powers-of-2, realistic sizes, zeros, negatives)
• Performance test passes (8192×8192 on Tesla T4)
• pre-commit run --all-files passes (black, isort, flake8, clang-format)
• All 6 starter files present and correct format

🤖 Generated with Claude Code

[claude]

Challenge Review: Integral Image (Challenge 77)

Checklist Verification

challenge.html ✅

• Starts with <p> (not <h1>)
• Has <h2> sections for: Implementation Requirements, Example, Constraints
• First example matches generate_example_test() values (3×3 matrix with correct output [[1,3,6],[5,12,21],[12,27,45]])
• Examples use LaTeX \begin{bmatrix} for 2D data (appropriate for matrix output)
• Constraints includes performance test size: H = 8,192, W = 8,192

challenge.py ✅

• class Challenge inherits ChallengeBase
• __init__ has name, atol, rtol, num_gpus, access_tier — atol=1.0 is appropriate for large cumulative sums with float32 precision
• reference_impl has assertions on shape, dtype, and device
• All 6 methods present
• generate_functional_test returns 10 cases covering: edge (1×1, 1×4, 4×1, 16×16 zeros), power-of-2 (32×32, 128×128), non-power-of-2 (30×30, 100×100, 255×33), realistic (1024×1024), zeros, negatives, mixed
• generate_performance_test: 8192×8192 × 4B × 2 tensors = ~536MB (fits 5× within 16GB ✅)

Starter files ✅

• All 6 files present: .cu, .pytorch.py, .triton.py, .jax.py, .cute.py, .mojo
• CUDA: // input, output are device pointers (medium, no parenthetical) ✓
• Mojo: # input, output are device pointers (medium, no parenthetical) ✓
• PyTorch/Triton/CuTe: # input, output are tensors on the GPU ✓
• JAX: # input is a tensor on the GPU + # return output tensor directly ✓
• All starters have empty solve function (correct for medium difficulty)

Validation

Wrote and ran a CUDA reference solution (row-wise prefix scan + column-wise prefix scan) against the platform:

• ✅ All functional tests passed (10 tests)
• ✅ Performance test passed (8192×8192 on NVIDIA Tesla T4)

[claude]

Review: Challenge 77 — Integral Image

Ran full checklist verification and validated with a CUDA solution on NVIDIA Tesla T4.

challenge.html ✅

• Starts with <p> ✓
• <h2> sections: Implementation Requirements, Example, Constraints ✓
• First example matches generate_example_test() (3×3 input/output values) ✓
• 2D data uses LaTeX \begin{bmatrix} consistently ✓
• Performance bullet: H = 8,192, W = 8,192 matches generate_performance_test() ✓

challenge.py ✅

• Inherits ChallengeBase, all 6 methods present ✓
• reference_impl asserts shape, dtype, and device ✓
• 10 functional tests: edge cases (1×1, 1×4, 4×1), powers-of-2 (16×16, 32×32, 128×128), non-powers (30×30, 100×100, 255×33), realistic (1024×1024), zeros, negatives ✓
• atol=1.0 appropriate for float32 prefix-sum accumulation ✓
• Performance test (8192×8192, ~536MB) fits 5× in 16GB VRAM ✓

Starter files ✅

• All 6 present ✓
• CUDA/Mojo: "device pointers" (no parenthetical — correct for medium) ✓
• PyTorch/Triton/CuTe: "tensors on the GPU" ✓
• JAX: "tensor on the GPU" + "return output tensor directly" ✓
• No extra comments ✓

Validation ✅

Wrote and ran a CUDA solution (two-pass: row-wise then column-wise prefix scan). All 10 functional tests and performance test passed on Tesla T4.

[claude]

Challenge Review: Add challenge 77 — Integral Image (Medium)

Reviewed all files against the CLAUDE.md checklist. A correct CUDA solution was written and submitted — all functional tests passed and the performance test completed successfully on NVIDIA TESLA T4.

challenge.html ✅

• Starts with <p> ✓
• <h2> sections present: Implementation Requirements, Example, Constraints ✓
• First example matches generate_example_test() output (verified manually: cumsum gives [[1,3,6],[5,12,21],[12,27,45]]) ✓
• 2D data uses LaTeX \begin{bmatrix} ✓
• Performance constraint bullet: H = 8,192, W = 8,192 matches generate_performance_test() ✓

challenge.py ✅

• class Challenge inherits ChallengeBase ✓
• __init__ has all required fields (name, atol, rtol, num_gpus, access_tier) ✓
• reference_impl has shape/dtype/device assertions ✓
• All 6 required methods present ✓
• generate_functional_test returns 10 cases covering edge cases (1×1, 1×4, 4×1, 16×16), power-of-2 (32×32, 128×128), non-power-of-2 (30×30, 100×100, 255×33), realistic (1024×1024), zeros, negatives, mixed ✓
• generate_performance_test: 8192×8192 × 4 bytes × 2 tensors ≈ 512 MB; fits 5× in 16 GB ✓
• atol=1.0 is appropriate for float32 cumulative sum accumulation ✓

Starter files ✅

All 6 files present: .cu, .pytorch.py, .triton.py, .jax.py, .cute.py, .mojo ✓

• CUDA: // input, output are device pointers (medium, no parenthetical) ✓
• Mojo: # input, output are device pointers ✓
• PyTorch/Triton/CuTe: # input, output are tensors on the GPU ✓
• JAX: # input is a tensor on the GPU + # return output tensor directly ✓
• All starters run without error but produce no correct output ✓

LGTM! All checklist items pass.