File: QW_GATHER.md

package info (click to toggle)
intel-graphics-compiler2 2.16.0-2
  • links: PTS, VCS
  • area: main
  • in suites: sid
  • size: 106,644 kB
  • sloc: cpp: 805,640; lisp: 287,672; ansic: 16,414; python: 3,952; yacc: 2,588; lex: 1,666; pascal: 313; sh: 186; makefile: 35
file content (110 lines) | stat: -rw-r--r-- 2,232 bytes parent folder | download | duplicates (2)
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
<!---======================= begin_copyright_notice ============================

Copyright (C) 2020-2022 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ==========================-->

## Opcode

  QW_GATHER = 0x86

## Format

| | | | | | | |
| --- | --- | --- | --- | --- | --- | --- |
| 0x86(QW_GATHER) | Exec_size | Pred | Num_blocks | Surface | Offset | Dst |


## Semantics


```

                    for (i = 0; i < exec_size; ++i) {
                        if (ChEn[i]) {
                            for (j = 0; j < num_blocks; ++j) {
                                dst[j*exec_size+i] = Surface[Offset[i] + j]; // 8 byte
                            }
                        }
                    }
```

## Description





```
    Performs <exec_size> element scattered read from <surface> and stores the
    result into <dst>. For each channel <num_blocks> * 8 bytes are read.
```


- **Exec_size(ub):** Execution size

  - Bit[2..0]: size of the region for source and destination operands

    - 0b000:  1 element (scalar)
    - 0b001:  2 elements
    - 0b010:  4 elements
    - 0b011:  8 elements
    - 0b100:  16 elements
  - Bit[7..4]: execution mask (explicit control over the enabled channels)

    - 0b0000:  M1
    - 0b0001:  M2
    - 0b0010:  M3
    - 0b0011:  M4
    - 0b0100:  M5
    - 0b0101:  M6
    - 0b0110:  M7
    - 0b0111:  M8
    - 0b1000:  M1_NM
    - 0b1001:  M2_NM
    - 0b1010:  M3_NM
    - 0b1011:  M4_NM
    - 0b1100:  M5_NM
    - 0b1101:  M6_NM
    - 0b1110:  M7_NM
    - 0b1111:  M8_NM

- **Pred(uw):** Predication control


- **Num_blocks(ub):**

  - Bit[1..0]: encodes the number of blocks to read for each address

    - 0b00:  1 block

- **Surface(ub):** Index of the surface variable. T0 (SLM) is supported


- **Offset(raw_operand):** The first Exec_size elements of the operand will be used as the byte offsets into the surface. Must have type UD


- **Dst(raw_operand):** The general variable storing the results of the read. Must have type Q,UQ,DF


#### Properties
- **Out-of-bound Access:** On read: zeroes are returned.




## Text
```



    [(<P>)] QW_GATHER.<num_blocks> (<exec_size>) <surface> <offset> <dst>
```
## Notes