cudaMalloc وcudaFree

Your GPU does not borrow the computer's main memory. It owns a separate pool called global memory that kernels read and write. 🧠

Before a kernel can use device memory, you reserve a chunk yourself. The function that does this reserving is cudaMalloc.

cudaMalloc takes the address of a pointer and a byte count. It fills that pointer with a fresh device address.

You pass &d_data so cudaMalloc can write the new address back into your variable. The return value is reserved for an error code.

cudaMalloc thinks in raw bytes, not elements. Multiply your element count by sizeof(type) so the buffer is the right size.

When you finish with a buffer, return it to the GPU using cudaFree. This releases the global memory for other work.

Treat cudaMalloc and cudaFree as a pair. Every allocation you make should have exactly one matching free.

Fresh memory from cudaMalloc holds garbage values. Reading it before you write or copy data gives unpredictable results.

If the GPU is out of memory, cudaMalloc returns an error instead of crashing. Always check that return code in real code.

Buffers from cudaMalloc live in global memory: gigabytes of space, yet the slowest tier on the chip. Use it generously but mind access patterns.

Nearly every CUDA program follows the same beat: cudaMalloc, copy data in, launch a kernel, copy results out, then cudaFree.

Let us confirm how cudaMalloc receives its target pointer.

You learned to reserve GPU memory with cudaMalloc, sized in bytes, and to release it with cudaFree in matched pairs. 🎉