Example 08 — Managed Memory

Source: examples/08_managed_memory.zig

Use ManagedBuffer (CUDA Unified Memory) to share data between CPU and GPU without explicit copies.

Code

const std = @import("std");
const cuda = @import("cuda");

pub fn main() !void {
    var gpa = std.heap.GeneralPurposeAllocator(.{}){};
    const allocator = gpa.allocator();

    try cuda.init();
    defer cuda.deinit();

    const n = 1024;
    var managed = try cuda.ManagedBuffer(f32).init(allocator, n);
    defer managed.deinit();

    // Write from CPU — no explicit H2D copy needed
    const s = managed.slice();
    for (s, 0..) |*v, i| v.* = @floatFromInt(i);

    std.debug.print("CPU write: s[42] = {d}\n", .{s[42]});

    var stream = try cuda.Stream.init();
    defer stream.deinit();

    // Prefetch to GPU before kernel launch
    try managed.prefetchToDevice(0, stream);
    try stream.sync();
    std.debug.print("Prefetched to GPU.\n", .{});

    // (Kernel would run here and update managed data)

    // Prefetch back to CPU for inspection
    try managed.prefetchToHost(stream);
    try stream.sync();
    std.debug.print("Back on CPU: s[42] = {d}\n", .{managed.slice()[42]});

    // Set access advice
    try managed.advise(.PreferredLocation, 0);
    try managed.advise(.AccessedBy, 0);
    std.debug.print("Memory advice set.\n", .{});
}

Expected Output

CPU write: s[42] = 42.0
Prefetched to GPU.
Back on CPU: s[42] = 42.0
Memory advice set.

When to Use Managed Memory

Scenario	Use managed memory?
Small, infrequently transferred data	✅ Yes
Large data transferred once	❌ Prefer explicit DeviceBuffer
Complex access patterns (CPU + GPU interleaved)	✅ Yes
Maximum throughput DMA transfers	❌ Prefer HostBuffer + async copy

Run

zig build example-08