File: SVM_GATHER4_SCALED.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 (117 lines) | stat: -rw-r--r-- 3,131 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
111
112
113
114
115
116
117
<!---======================= begin_copyright_notice ============================

Copyright (C) 2020-2022 Intel Corporation

SPDX-License-Identifier: MIT

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

## Opcode

  SVM = 0x4e
  GATHER4_SCALED = 0x06

## Format

| | | | | | | |
| --- | --- | --- | --- | --- | --- | --- |
| 0x4e(SVM) | 0x06(GATHER4_SCALED) | Exec_size | Pred | Channels | Scale | Address | Element_offset |
|           |                      | Dst       |      |          |       |         |                |


## Semantics


```

                    UD ch_pos = 0;
                    for (c = 0; c < 4; ++c) {
                        if (ch_mask[c] == 1) {
                            for (i = 0; i < exec_size; ++i) {
                                if (ChEn[i]) {
                                    UD ch_start = ch_pos * max(exec_size, GRF_SIZE / 4);  // GRF_SIZE is the register size in byte
                                    dst[ch_start+i] = *(address + element_offset[i] + (c * 4)); //4 byte
                                }
                            }
                            ch_pos++;
                        }
                    }
```

## Description





```
    Performs <exec_size>*<num_enabled_channels> element scattered read
    from virtual addresses and stores the results into <dst>. The address to
    be read is determined by the offset, scaling pitch, and each element
    offset, and the final address must be dword-aligned.
```


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

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

    - 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


- **Channels(ub):**

  - Bit[3..0]: determines the write masks for the RGBA channel, with R being bit 0 and A bit 3. At least one channel must be enabled (i.e., "0000" is not allowed)


- **Scale(uw):** This field is ingored and scale is always zero


- **Address(scalar):** The global byte offset. Must have type UQ


- **Element_offset(raw_operand):** The first Exec_size elements of the operand will be used as the byte offsets for the corresponding channel. Must have type UQ


- **Dst(raw_operand):** The variable storing the results of the read. For each enabled channel in RGBA order, exec_size elements will be returned to dst subject to predication. The next enabled channel will have its return data starting at the next register. If a channel's return data do not occupy the entire register, the remaining part of the register has undefined values. Must have type UD,D,F


#### Properties




## Text
```



    [(<P>)] SVM_GATHER4_SCALED.<channels> (<exec_size>) <address> <element_offset> <dst>
```
## Notes