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
|
<!---======================= begin_copyright_notice ============================
Copyright (C) 2020-2021 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ==========================-->
## Opcode
SVM = 0x4e
SCATTER = 0x04
## Format
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| 0x4e(SVM) | 0x04(SCATTER) | Exec_size | Pred | Block_size | Num_blocks | Addresses | Src |
## Semantics
for (i = 0; i < exec_size; ++i) {
if (ChEn[i]) {
if (block_size == 4 || block_size == 8) {
for (j = 0; j < num_blocks; ++j) {
*(Addresses[i] + j*block_size) = src[j*exec_size+i]; //4 or 8 byte
}
}
else { // block_size == 1
for (j = 0; j < num_blocks; ++j) {
UD min_block_size = num_blocks < 4 ? 4 : num_blocks;
*(Addresses[i] + j) = src[i*min_block_size+j]; //1 byte
}
}
}
}
## Description
Performs 8 element scattered write to virtual addresses, taking the
values from <src>. Each element may have multiple blocks.
- **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
- **Block_size(ub):**
- Bit[1..0]: encodes the byte size of each block
- 0b00: 1 byte
- 0b01: 4 byte
- 0b10: 8 byte
- **Num_blocks(ub):**
- Bit[1..0]: encodes the number of blocks to read for each address
- 0b00: 1 block
- 0b01: 2 blocks
- 0b10: 4 blocks
- 0b11: 8 blocks. It is only valid for 4 byte blocks and execution size 8
- **Addresses(raw_operand):** The general variable storing the addresses to write. The first 8 elements of the operand will be used as the virtual addresses to write to, and they are in the unit of bytes. Each address must be aligned to the Block. Must have type UQ
- **Src(raw_operand):** The general variable providing the values to be written to each address. Its format depends on <block_size> and <num_blocks>
#### Properties
## Text
```
[(<P>)] SVM_SCATTER.<block_size>.<num_blocks> (<exec_size>) <addresses> <src>
```
## Notes
4 byte scatter:
+-----------------+-----------------+----+----+-----+----+----+-----------------+
| A0[0] | A1[0] | | | ... | | | A7[0] |
+-----------------+-----------------+----+----+-----+----+----+-----------------+
| ... |
+-----------------+-----------------+----+----+-----+----+----+-----------------+
| A0[#blocks-1] | A1[#blocks-1] | | | ... | | | A7[#blocks-1] |
+-----------------+-----------------+----+----+-----+----+----+-----------------+
8 byte scatter:
+-----------------+-----------------+-----------------+-----------------+
| A0[0] | A1[0] | A2[0] | A3[0] |
+-----------------+-----------------+-----------------+-----------------+
| A4[0] | A5[0] | A6[0] | A7[0] |
+-----------------+-----------------+-----------------+-----------------+
| ... |
+-----------------+-----------------+-----------------+-----------------+
| A0[#blocks-1] | A1[#blocks-1] | A2[#blocks-1] | A3[#blocks-1] |
+-----------------+-----------------+-----------------+-----------------+
| A4[#blocks-1] | A5[#blocks-1] | A6[#blocks-1] | A7[#blocks-1] |
+-----------------+-----------------+-----------------+-----------------+
1 byte scatter, 1/2/4 block (Ai[N] is ignored if N>= num_blocks)
+-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
| A0[0:3] | A1[0:3] | A2[0:3] | A3[0:3] | A4[0:3] | A5[0:3] | A6[0:3] | A7[0:3] |
+-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
1 byte scatter, 8 block
+-----------+-----------+-----------+-----------+
| A0[0:7] | A1[0:7] | A2[0:7] | A3[0:7] |
+-----------+-----------+-----------+-----------+
| A4[0:7] | A5[0:7] | A6[0:7] | A7[0:7] |
+-----------+-----------+-----------+-----------+
Each table row represents one GRF, and A0[0] corresponds to the
first element in <src> in all of the above tables. The operand's
type must be the same size as <block_size>.
|
