W1028

Warning

W1028 — The working set of a memory space exceeds capacity, but the space is declared overcommit, so the cumulative-budget errors (E6010, and the cross-call E6027/E6028) are downgraded to this warning.

What the compiler reports

Warning[W1028] at {{.*}}.vx:[[@LINE+1]]:20: this channel holds `Tensor<f32, [128], Memory::O>`, which lives in memory space 'O', and its room is not known at compile time. {{.*}} This is allowed because 'O' is `overcommit`
{{^}}2 0{{$}}

The fragments the test suite holds the compiler to. An ellipsis marks text the test does not constrain; the emitted message also carries a source location.

A program that triggers it

//
// Part of the Vx Project, under the Apache License v2.0 with LLVM Exceptions.
// See LICENSE for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
// The capacity check counts the room of a channel of tensors in a memory with a capacity. Each
// tensor here is 512 bytes, and `W` and `O` hold 4096 each.
//
// - `fits`: a room of 3 tensors and one more tensor, 2048 bytes, fit in `W`.
// - `room_not_known`: a room that is a parameter has no bound. `O` is `overcommit`, so this is
//   a warning.
// - `on_the_host`: host memory has no capacity, so a room of 1000 tensors is not counted.
//

import std::sync::channel;

Memory CPU_DRAM {}
Memory W {
  within: Memory::CPU_DRAM, capacity: 4 KiB, managed: explicit, scope: device
}
Memory O {
  within: Memory::CPU_DRAM, capacity: 4 KiB, managed: explicit, scope: device, overcommit
}
Topology Dev {
  arch: nvptx64,
  memory: Memory::W,
  visible: [Memory::W, Memory::O],
  transfer Memory::CPU_DRAM -> Memory::W : 10,
  transfer Memory::W -> Memory::CPU_DRAM : 10,
  transfer Memory::CPU_DRAM -> Memory::O : 10,
  transfer Memory::O -> Memory::CPU_DRAM : 10
}

fn fits() -> f32 {
  let (tx, rx) = channel<Tensor<f32, [128], Memory::W>>(3);
  let _ = tx.send(transfer(Tensor<f32, [128]>::fill(2.0), Memory::W));
  let mut got = 0.0;
  match rx.recv() {
    Result<Tensor<f32, [128], Memory::W>, RecvError>::Ok(b) => {
      let h = transfer(b, Memory::CPU_DRAM);
      got = h[0];
    }
    Result<Tensor<f32, [128], Memory::W>, RecvError>::Err(_e) => {}
  }
  return got;
}

fn room_not_known(n : i64) -> void {
  let (_tx, _rx) = channel<Tensor<f32, [128], Memory::O>>(n);
}

fn on_the_host() -> i32 {
  let (_tx, _rx) = channel<Tensor<f32, [128]>>(1000);
  return 0;
}

fn main() -> i32 {
  room_not_known(2);
  print!("{} {}", fits(), on_the_host());
  return 0;
}

From tests/optimizations/pass/a_channel_of_gpu_tensors_that_fits.vx, which asserts this diagnostic on every commit.

Related