cubecl-ir 0.11.0-pre.3

Intermediate representation for CubeCL
Documentation 
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
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177

use cubecl_macros_internal::{cube_op, op_traits};
use pliron::r#type::{TypeHandle, TypedHandle};

use crate::{
    CanMaterialize, HasSideEffects,
    attributes::{BoolAttr, IndexAttr},
    dialect::synchronization::SyncScope,
    interfaces::Synchronizes,
    prelude::*,
    types::{PointerType, barrier::BarrierTokenType},
};

#[cube_op(name = "barrier.init")]
#[result_ty(none)]
#[op_traits(CanMaterialize)]
pub struct InitOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
    pub arrival_count: Value,
}

#[cube_op(name = "barrier.memcpy_async")]
#[result_ty(none)]
#[op_traits(CanMaterialize)]
pub struct MemCopyAsyncOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
    #[operand(ptr_read)]
    pub source: Value,
    #[operand(ptr_write)]
    pub destination: Value,
    pub source_length: Value,
    pub cooperative: BoolAttr,
}

#[cube_op(name = "barrier.memcpy_async_tx")]
#[result_ty(none)]
#[op_traits(CanMaterialize)]
pub struct MemCopyAsyncTxOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
    #[operand(ptr_read)]
    pub source: Value,
    #[operand(ptr_write)]
    pub destination: Value,
    pub source_length: Value,
}

#[cube_op(name = "barrier.copy_async")]
#[result_ty(none)]
#[op_traits(CanMaterialize)]
pub struct CopyAsyncOp {
    #[operand(ptr_read)]
    pub source: Value,
    #[operand(ptr_write)]
    pub destination: Value,
    pub source_length: Value,
    pub copy_length: IndexAttr,
    pub checked: BoolAttr,
}

#[cube_op(name = "barrier.arrive")]
#[result_ty(from_inputs = token_ty)]
#[op_traits(CanMaterialize)]
pub struct ArriveOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
}

#[cube_op(name = "barrier.arrive_and_expect_tx")]
#[result_ty(from_inputs = |ctx, bar, _, _| token_ty(ctx, bar))]
#[op_traits(CanMaterialize)]
pub struct ArriveAndExpectTxOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
    pub arrive_count_update: Value,
    pub transaction_count_update: Value,
}

fn token_ty(ctx: &Context, barrier: &Value) -> TypeHandle {
    let bar_ptr = barrier.get_type(ctx).deref(ctx);
    let bar_ptr = bar_ptr.downcast_ref::<PointerType>().unwrap();
    let bar = TypedHandle::from_handle(bar_ptr.inner, ctx).expect("Should be barrier");
    BarrierTokenType::get(ctx, bar).into()
}

#[cube_op(name = "barrier.commit_copy_async")]
#[result_ty(none)]
#[op_traits(CanMaterialize)]
pub struct CommitCopyAsyncOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
}

#[cube_op(name = "barrier.expect_tx")]
#[result_ty(none)]
#[op_traits(CanMaterialize)]
pub struct ExpectTxOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
    pub transaction_count_update: Value,
}

#[cube_op(name = "barrier.wait")]
#[result_ty(none)]
#[op_traits(CanMaterialize, HasSideEffects)]
pub struct WaitOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
    pub token: Value,
}

/// Sync scope depends on init params and use, so can't be trivially analyzed. So just be
/// conservative.
#[op_interface_impl]
impl Synchronizes for WaitOp {
    fn minimum_scope(&self, _ctx: &Context) -> SyncScope {
        SyncScope::Unit
    }

    fn maximum_scope(&self, _ctx: &Context) -> SyncScope {
        SyncScope::Device
    }
}

#[cube_op(name = "barrier.wait_parity")]
#[result_ty(none)]
#[op_traits(CanMaterialize, HasSideEffects)]
pub struct WaitParityOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
    pub phase: Value,
}

/// Sync scope depends on init params and use, so can't be trivially analyzed. So just be
/// conservative.
#[op_interface_impl]
impl Synchronizes for WaitParityOp {
    fn minimum_scope(&self, _ctx: &Context) -> SyncScope {
        SyncScope::Unit
    }

    fn maximum_scope(&self, _ctx: &Context) -> SyncScope {
        SyncScope::Device
    }
}

#[cube_op(name = "barrier.arrive_and_wait")]
#[result_ty(none)]
#[op_traits(CanMaterialize, HasSideEffects)]
pub struct ArriveAndWaitOp {
    // Opaque so we can't know exact memory effects. Treat it as atomic read-update.
    #[operand(ptr_read, ptr_write)]
    pub barrier: Value,
}

/// Sync scope depends on init params and use, so can't be trivially analyzed. So just be
/// conservative.
#[op_interface_impl]
impl Synchronizes for ArriveAndWaitOp {
    fn minimum_scope(&self, _ctx: &Context) -> SyncScope {
        SyncScope::Unit
    }

    fn maximum_scope(&self, _ctx: &Context) -> SyncScope {
        SyncScope::Device
    }
}