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
|
<!---======================= begin_copyright_notice ============================
Copyright (C) 2020-2022 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ==========================-->
## Opcode
LSC_FENCE = 0x8b
## Format
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| 0x8b(LSC_FENCE) | ExecSize | Pred | LscSFID | LscFenceOp | LscScope |
## Semantics
```
wait_for_pending_accesses_to_complete(LscSFID, LscFenceOp, LscScope)
Dest.data = 0 // Dest. is implicit within vISA
```
## Description
Wait until all previous accesses by this thread to this dataport are observable in the scope specified. Then optionally flush the cache.
- **ExecSize(ub):** Execution size. Should always be 1 and no mask
- **Pred(uw):** Predication control
- **LscSFID(ub):** The specific LSC shared function to fence
.. _table_LSC_SFID:
.. table:: **LSC_SFID**
| Value | Symbol | vISA Enum | Notes |
| --- | ---| ---| ---| ---|
| 0x0 | ugm | LSC_UGM | Untyped global memory |
| 0x1 | ugml | LSC_UGML | Low-bandwidth untyped global memory (cross tile) |
| 0x2 | tgm | LSC_TGM | Typed global memory |
| 0x3 | slm | LSC_SLM | Shared local memory |
- **LscFenceOp(ub):** The fence operation to apply
.. _table_LSC_FENCE_OP:
.. table:: **LSC_FENCE_OP**
| Value | Symbol | vISA Enum | Notes |
| --- | ---| ---| ---| ---|
| 0x0 | none | LSC_FENCE_OP_NONE | No operation |
| 0x1 | evict | LSC_FENCE_OP_EVICT | Dirty lines will be evicted and invalidated from the L1 cache. All clean lines will be invalidated |
| 0x2 | invalidate | LSC_FENCE_OP_INVALIDATE | Invalidate all Clean lines in the cache. Don't evict the dirty lines (they stay in M state) |
| 0x3 | discard | LSC_FENCE_OP_DISCARD | Direct and clean lines are discarded (invalidated) without eviction |
| 0x4 | clean | LSC_FENCE_OP_CLEAN | Dirty lines are written to memory, but retained in cache in the "Clean" state. Clean lines are not invalidated |
| 0x5 | flushl3 | LSC_FENCE_OP_FLUSHL3 | Flush L3 only |
- **LscScope(ub):** The scope that this operation should apply to
.. _table_LSC_SCOPE:
.. table:: **LSC_SCOPE**
| Value | Symbol | vISA Enum | Notes |
| --- | ---| ---| ---| ---|
| 0x0 | group | LSC_SCOPE_GROUP | Flush out to the threadgroup's scope |
| 0x1 | local | LSC_SCOPE_LOCAL | Flush out to the local scope (DSSs) |
| 0x2 | tile | LSC_SCOPE_TILE | Tile (out to several DSSs) |
| 0x3 | gpu | LSC_SCOPE_GPU | The entire GPU (out to the GPUs LLC) |
| 0x4 | gpus | LSC_SCOPE_GPUS | All GPUs in the system (memory shared by all GPUs) |
| 0x5 | system | LSC_SCOPE_SYSTEM | Wait until all previous memory transactions from this thread are observed at the system level |
| 0x6 | sysacq | LSC_SCOPE_SYSACQ | For GPUs that do not follow PCIe write ordering for downstream writes targeting device memory, a fence message with this scope will commit to device memory all downstream and peer writes that have reached the device |
#### Properties
## Text
```
LSC_FENCE.<LscSFID>.<LscFenceOp>.<LscScope>
```
## Examples
```
// Fences across the entire system for untyped global memory
lsc_fence.ugm.clean.sysrel
// Fences SLM
lsc_fence.slm.clean.group
```
## Notes
|
