Gitiles
Code Review Sign In
review.blissroms.org / platform_external_llvm / refs/heads/r / . / docs / AMDGPUInstructionSyntax.rst
blob: 90ad54aade839479e026d730ef85805a8736c30c [file] [log] [blame]
Dmitry Preobrazhensky51120d72018-12-17 17:38:11 +0000[diff] [blame]1=========================
2AMDGPU Instruction Syntax
3=========================
4
5.. contents::
6 :local:
7
8.. _amdgpu_syn_instructions:
9
10Instructions
11============
12
13Syntax
14~~~~~~
15
16An instruction has the following syntax:
17
18 ``<``\ *opcode mnemonic*\ ``> <``\ *operand0*\ ``>, <``\ *operand1*\ ``>,... <``\ *modifier0*\ ``> <``\ *modifier1*\ ``>...``
19
20:doc:`Operands<AMDGPUOperandSyntax>` are normally comma-separated while
21:doc:`modifiers<AMDGPUModifierSyntax>` are space-separated.
22
23The order of *operands* and *modifiers* is fixed.
24Most *modifiers* are optional and may be omitted.
25
26.. _amdgpu_syn_instruction_mnemo:
27
28Opcode Mnemonic
29~~~~~~~~~~~~~~~
30
31Opcode mnemonic describes opcode semantics and may include one or more suffices in this order:
32
33* :ref:`Destination operand type suffix<amdgpu_syn_instruction_type>`.
34* :ref:`Source operand type suffix<amdgpu_syn_instruction_type>`.
35* :ref:`Encoding suffix<amdgpu_syn_instruction_enc>`.
36
37.. _amdgpu_syn_instruction_type:
38
39Type and Size Suffices
40~~~~~~~~~~~~~~~~~~~~~~
41
42Instructions which operate with data have an implied type of *data* operands.
43This data type is specified as a suffix of instruction mnemonic.
44
45There are instructions which have 2 type suffices:
46the first is the data type of the destination operand,
47the second is the data type of source *data* operand(s).
48
49Note that data type specified by an instruction does not apply
50to other kinds of operands such as *addresses*, *offsets* and so on.
51
52The following table enumerates the most frequently used type suffices.
53
54 ============================================ ======================= =================
55 Type Suffices Packed instruction? Data Type
56 ============================================ ======================= =================
57 _b512, _b256, _b128, _b64, _b32, _b16, _b8 No Bits.
58 _u64, _u32, _u16, _u8 No Unsigned integer.
59 _i64, _i32, _i16, _i8 No Signed integer.
60 _f64, _f32, _f16 No Floating-point.
61 _b16, _u16, _i16, _f16 Yes Packed.
62 ============================================ ======================= =================
63
64Instructions which have no type suffices are assumed to operate with typeless data.
65The size of data is specified by size suffices:
66
67 ================= =================== =====================================
68 Size Suffix Implied data type Required register size in dwords
69 ================= =================== =====================================
70 \- b32 1
71 x2 b64 2
72 x3 b96 3
73 x4 b128 4
74 x8 b256 8
75 x16 b512 16
76 x b32 1
77 xy b64 2
78 xyz b96 3
79 xyzw b128 4
80 d16_x b16 1
81 d16_xy b16x2 2 for GFX8.0, 1 for GFX8.1 and GFX9
82 d16_xyz b16x3 3 for GFX8.0, 2 for GFX8.1 and GFX9
83 d16_xyzw b16x4 4 for GFX8.0, 2 for GFX8.1 and GFX9
84 ================= =================== =====================================
85
86.. WARNING::
87 There are exceptions from rules described above.
88 Operands which have type different from type specified by the opcode are
89 :ref:`tagged<amdgpu_syn_instruction_operand_tags>` in the description.
90
91Examples of instructions with different types of source and destination operands:
92
Dmitry Preobrazhensky23da1102018-12-17 18:53:10 +0000[diff] [blame]93.. parsed-literal::
Dmitry Preobrazhensky51120d72018-12-17 17:38:11 +0000[diff] [blame]94
95 s_bcnt0_i32_b64
96 v_cvt_f32_u32
97
98Examples of instructions with one data type:
99
Dmitry Preobrazhensky23da1102018-12-17 18:53:10 +0000[diff] [blame]100.. parsed-literal::
Dmitry Preobrazhensky51120d72018-12-17 17:38:11 +0000[diff] [blame]101
102 v_max3_f32
103 v_max3_i16
104
105Examples of instructions which operate with packed data:
106
Dmitry Preobrazhensky23da1102018-12-17 18:53:10 +0000[diff] [blame]107.. parsed-literal::
Dmitry Preobrazhensky51120d72018-12-17 17:38:11 +0000[diff] [blame]108
109 v_pk_add_u16
110 v_pk_add_i16
111 v_pk_add_f16
112
113Examples of typeless instructions which operate on b128 data:
114
Dmitry Preobrazhensky23da1102018-12-17 18:53:10 +0000[diff] [blame]115.. parsed-literal::
Dmitry Preobrazhensky51120d72018-12-17 17:38:11 +0000[diff] [blame]116
117 buffer_store_dwordx4
118 flat_load_dwordx4
119
120.. _amdgpu_syn_instruction_enc:
121
122Encoding Suffices
123~~~~~~~~~~~~~~~~~
124
125Most *VOP1*, *VOP2* and *VOPC* instructions have several variants:
126they may also be encoded in *VOP3*, *DPP* and *SDWA* formats.
127
128The assembler will automatically use optimal encoding based on instruction operands.
129To force specific encoding, one can add a suffix to the opcode of the instruction:
130
131 =================================================== =================
132 Encoding Encoding Suffix
133 =================================================== =================
134 Native 32-bit encoding (*VOP1*, *VOP2* or *VOPC*) _e32
135 *VOP3* (64-bit) encoding _e64
136 *DPP* encoding _dpp
137 *SDWA* encoding _sdwa
138 =================================================== =================
139
140These suffices are used in this reference to indicate the assumed encoding.
141When no suffix is specified, a native encoding is implied.
142
143Operands
144========
145
146Syntax
147~~~~~~
148
149Syntax of most operands is described :doc:`in this document<AMDGPUOperandSyntax>`.
150
151For detailed information about operands follow *operand links* in GPU-specific documents:
152
153* :doc:`GFX7<AMDGPU/AMDGPUAsmGFX7>`
154* :doc:`GFX8<AMDGPU/AMDGPUAsmGFX8>`
155* :doc:`GFX9<AMDGPU/AMDGPUAsmGFX9>`
156
157Modifiers
158=========
159
160Syntax
161~~~~~~
162
163Syntax of modifiers is described :doc:`in this document<AMDGPUModifierSyntax>`.
164
165Information about modifiers supported for individual instructions may be found in GPU-specific documents:
166
167* :doc:`GFX7<AMDGPU/AMDGPUAsmGFX7>`
168* :doc:`GFX8<AMDGPU/AMDGPUAsmGFX8>`
169* :doc:`GFX9<AMDGPU/AMDGPUAsmGFX9>`
170