Transmitindo blocos para imagens grandes

A huge image may not fit in GPU memory, or copying it whole stalls the pipeline. The fix is to split it into tiles and process them in turns. 🖼️

Slice the image into row bands or rectangular tiles. Each tile is small enough to upload, process, and download without straining device memory.

Every tile follows the same rhythm: copy it to the device, run the kernel, copy the result back. Repeat until the whole image is done.

Done one at a time, the GPU sits idle during every copy. To win, you must overlap a tile's transfer with another tile's compute.

Issue each tile's work in its own stream. The driver can then run one tile's kernel while another tile's copy is still in flight.

Plain cudaMemcpy blocks the host and kills overlap. Use cudaMemcpyAsync on a stream so the copy queues without stopping everything.

Async transfers need pinned host memory to use DMA. Allocate the tile staging buffers with cudaMallocHost or overlap silently falls back to blocking.

Keep two tile buffers and alternate streams. While tile N computes, tile N+1 uploads, hiding transfers behind work in a smooth pipeline.

A blur near a tile border needs pixels from the neighbor. Include a halo margin of overlap so edge results stay correct.

Async work is not finished when the call returns. Call cudaStreamSynchronize on each stream before you read a tile's result back on the host.

With streamed, double-buffered tiles, the GPU stays fed no matter the image size. Transfers vanish behind compute and throughput stays high. ⚡

You stream tiles in separate streams but see no overlap. Which mistake is most likely?

You split big images into tiles, used streams with async copies and pinned memory to overlap, double-buffered the pipeline, and added halos for correct edges. 🎯