Questions
Read the code, predict what happens, then see why. Hand-written questions on C++, Rust, and the systems underneath them.
- A warp whose even and odd lanes take different branchesMedium
- One thread per byte across a 5-billion-byte bufferHard
- A 2-D grid launch that over-provisions threads for the image sizeMedium
- A launch configured with dim3(64, 32) threads per blockEasy
- One global atomicAdd per element versus one per blockHard
- cudaMemcpy from pageable host memory versus cudaMallocHost memoryMedium
- A parent kernel that reads its child grid's outputHard
- A cross-warp flag-and-payload spin-wait in shared memoryHard
- Warp-wide sum via a shared-memory tree versus __shfl_down_syncHard
- Fifty kernel passes over a 1 GB cudaHostAllocMapped bufferMedium
- Histogram binning with hist[bin]++ versus atomicAddMedium
- An atomic-counter barrier between two phases of one kernelHard
- cudaMemcpyPeerAsync versus staging through a pinned host bufferMedium
- An intra-warp shared-memory handoff across a divergent branchHard
- Two 1 KB host-device transfers around a 256-element add kernelEasy
- A per-thread SpMV loop with data-dependent trip countsHard
- The SIMT execution model compared with classic SIMDMedium
- A 32x32 shared-memory tile read with swapped indicesHard
- Unrolling a memory-bound loop at 32 versus 96 registers per threadHard
- Reading one constant-memory address versus 32 different ones per warpMedium
1–20 of 46
Page 1 of 3