https://sankethbk.github.io/blog/tags/shared-memory/Recent content in Shared-Memory on Sanketh's BlogHugo -- 0.166.0en-usTue, 01 Sep 2026 00:00:00 +0000https://sankethbk.github.io/blog/posts/gpu/2026-09-01-1-cuda-shared-memory-tiling/Tue, 01 Sep 2026 00:00:00 +0000https://sankethbk.github.io/blog/posts/gpu/2026-09-01-1-cuda-shared-memory-tiling/<h1 id="cuda-shared-memory-tiling">CUDA Shared Memory Tiling</h1> <p>The <a href="https://sankethbk.github.io/blog/gpu/2026-08-30-3-cuda-memory-coalescing">previous note</a> showed that coalesced memory access is the key to memory-bound kernel performance. But some algorithms naturally want uncoalesced access patterns. The fix is <strong>shared-memory tiling</strong>: load data into fast on-chip SRAM in a coalesced way, rearrange it there, then write it back to global memory in a coalesced way.</p> <p>This note explains the tiling pattern, why <code>__syncthreads()</code> is essential, the bank-conflict problem, and how to choose tile sizes.</p>