cubecl-llvm 0.11.0-pre.4

LLVM compiler for CubeCL
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

//! Target-specific synchronization.

use crate::prelude::*;
use cubecl_core::ir::dialect::synchronization::{SyncOp, SyncScope};

#[op_interface_impl]
impl LowerOp for SyncOp {
    fn lower(&self, scope: &Scope) -> Vec<Value> {
        let ctx = scope.ctx();
        let sync_scope = self.scope(ctx).0;
        let target = ctx.target();
        let op = self.get_operation();

        match sync_scope {
            SyncScope::Unit => {}
            SyncScope::Plane => match target {
                LlvmTarget::Cpu => {}
                #[cfg(feature = "amdgpu")]
                LlvmTarget::AmdGpu => crate::amdgpu::synchronization::lower_sync_plane(scope),
                #[cfg(feature = "nvptx")]
                LlvmTarget::Nvptx => crate::nvptx::synchronization::lower_sync_plane(scope),
            },
            // GPU synchronization is limited to the current cube.
            SyncScope::Cube | SyncScope::Device => match target {
                LlvmTarget::Cpu if sync_scope == SyncScope::Device => {
                    panic!("Device wide synchronization is not supported by the CPU runtime")
                }
                LlvmTarget::Cpu => crate::cpu::synchronization::lower_sync_cube(scope, op),
                #[cfg(feature = "amdgpu")]
                LlvmTarget::AmdGpu => crate::amdgpu::synchronization::lower_sync_cube(scope),
                #[cfg(feature = "nvptx")]
                LlvmTarget::Nvptx => crate::nvptx::synchronization::lower_sync_cube(scope),
            },
        }
        vec![]
    }
}