llvm-project/llvm/docs/AMDGPUInstructionSyntax.rst

209 lines
7.0 KiB
ReStructuredText

=========================
AMDGPU Instruction Syntax
=========================
.. contents::
:local:
.. _amdgpu_syn_instructions:
Instructions
============
Syntax
~~~~~~
Syntax of Regular Instructions
------------------------------
An instruction has the following syntax:
| ``<``\ *opcode mnemonic*\ ``> <``\ *operand0*\ ``>,
<``\ *operand1*\ ``>,... <``\ *modifier0*\ ``> <``\ *modifier1*\ ``>...``
:doc:`Operands<AMDGPUOperandSyntax>` are normally comma-separated, while
:doc:`modifiers<AMDGPUModifierSyntax>` are space-separated.
The order of *operands* and *modifiers* is fixed.
Most *modifiers* are optional and may be omitted.
Syntax of VOPD Instructions
---------------------------
*VOPDX* and *VOPDY* instructions must be concatenated with the :: operator to form a single *VOPD* instruction:
``<``\ *VOPDX instruction*\ ``> :: <``\ *VOPDY instruction*\ ``>``
An example:
.. parsed-literal::
v_dual_add_f32 v255, v255, v2 :: v_dual_fmaak_f32 v6, v2, v3, 1.0
Note that *VOPDX* and *VOPDY* instructions cannot be used as separate opcodes.
.. _amdgpu_syn_instruction_mnemo:
Opcode Mnemonic
~~~~~~~~~~~~~~~
Opcode mnemonic describes opcode semantics
and may include one or more suffices in this order:
* :ref:`Packing suffix<amdgpu_syn_instruction_pk>`.
* :ref:`Destination operand type suffix<amdgpu_syn_instruction_type>`.
* :ref:`Source operand type suffix<amdgpu_syn_instruction_type>`.
* :ref:`Encoding suffix<amdgpu_syn_instruction_enc>`.
.. _amdgpu_syn_instruction_pk:
Packing Suffix
~~~~~~~~~~~~~~
Most instructions which operate on packed data have a *_pk* suffix.
Unless otherwise :ref:`noted<amdgpu_syn_instruction_operand_tags>`,
these instructions operate on and produce packed data composed of
two values. The type of values is indicated by
:ref:`type suffices<amdgpu_syn_instruction_type>`.
For example, the following instruction sums up two pairs of f16 values
and produces a pair of f16 values:
.. parsed-literal::
v_pk_add_f16 v1, v2, v3 // Each operand has f16x2 type
.. _amdgpu_syn_instruction_type:
Type and Size Suffices
~~~~~~~~~~~~~~~~~~~~~~
Instructions which operate with data have an implied type of *data* operands.
This data type is specified as a suffix of instruction mnemonic.
There are instructions which have 2 type suffices:
the first is the data type of the destination operand,
the second is the data type of source *data* operand(s).
Note that data type specified by an instruction does not apply
to other kinds of operands such as *addresses*, *offsets* and so on.
The following table enumerates the most frequently used type suffices.
============================================ ======================= ============================
Type Suffices Packed instruction? Data Type
============================================ ======================= ============================
_b512, _b256, _b128, _b64, _b32, _b16, _b8 No Bits.
_u64, _u32, _u16, _u8 No Unsigned integer.
_i64, _i32, _i16, _i8 No Signed integer.
_f64, _f32, _f16 No Floating-point.
_b16, _u16, _i16, _f16 Yes Packed (b16x2, u16x2, etc).
============================================ ======================= ============================
Instructions which have no type suffices are assumed to operate with typeless data.
The size of typeless data is specified by size suffices:
================= =================== =====================================
Size Suffix Implied data type Required register size in dwords
================= =================== =====================================
\- b32 1
x2 b64 2
x3 b96 3
x4 b128 4
x8 b256 8
x16 b512 16
x b32 1
xy b64 2
xyz b96 3
xyzw b128 4
d16_x b16 1
d16_xy b16x2 2 for GFX8.0, 1 for GFX8.1 and GFX9+
d16_xyz b16x3 3 for GFX8.0, 2 for GFX8.1 and GFX9+
d16_xyzw b16x4 4 for GFX8.0, 2 for GFX8.1 and GFX9+
d16_format_x b16 1
d16_format_xy b16x2 1
d16_format_xyz b16x3 2
d16_format_xyzw b16x4 2
================= =================== =====================================
.. WARNING::
There are exceptions to the rules described above.
Operands which have a type different from the type specified by the opcode are
:ref:`tagged<amdgpu_syn_instruction_operand_tags>` in the description.
Examples of instructions with different types of source and destination operands:
.. parsed-literal::
s_bcnt0_i32_b64
v_cvt_f32_u32
Examples of instructions with one data type:
.. parsed-literal::
v_max3_f32
v_max3_i16
Examples of instructions which operate with packed data:
.. parsed-literal::
v_pk_add_u16
v_pk_add_i16
v_pk_add_f16
Examples of typeless instructions which operate on b128 data:
.. parsed-literal::
buffer_store_dwordx4
flat_load_dwordx4
.. _amdgpu_syn_instruction_enc:
Encoding Suffices
~~~~~~~~~~~~~~~~~
Most *VOP1*, *VOP2* and *VOPC* instructions have several variants:
they may also be encoded in *VOP3*, *DPP* and *SDWA* formats.
The assembler selects an optimal encoding automatically
based on instruction operands and modifiers,
unless a specific encoding is explicitly requested.
To force specific encoding, one can add a suffix to the opcode of the instruction:
=================================================== =================
Encoding Encoding Suffix
=================================================== =================
*VOP1*, *VOP2* and *VOPC* (32-bit) encoding _e32
*VOP3* (64-bit) encoding _e64
*DPP* encoding _dpp
*SDWA* encoding _sdwa
*VOP3 DPP* encoding _e64_dpp
=================================================== =================
This reference uses encoding suffices to specify which encoding is implied.
When no suffix is specified, native instruction encoding is assumed.
Operands
========
Syntax
~~~~~~
The syntax of generic operands is described :doc:`in this document<AMDGPUOperandSyntax>`.
For detailed information about operands, follow *operand links* in GPU-specific documents.
Modifiers
=========
Syntax
~~~~~~
The syntax of modifiers is described :doc:`in this document<AMDGPUModifierSyntax>`.
Information about modifiers supported for individual instructions
may be found in GPU-specific documents.