Syntax of AMDGPU Instruction Operands

Conventions

The following notation is used throughout this document:

Notation Description
{0..N} Any integer value in the range from 0 to N (inclusive).
<x> Syntax and meaning of x is explained elsewhere.

Operands

v

Vector registers. There are 256 32-bit vector registers.

A sequence of vector registers may be used to operate with more than 32 bits of data.

Assembler currently supports sequences of 1, 2, 3, 4, 8 and 16 vector registers.

Syntax Description
v<N>

A single 32-bit vector register.

N must be a decimal integer number.

v[<N>]

A single 32-bit vector register.

N may be specified as an integer number or an absolute expression.

v[<N>:<K>]

A sequence of (K-N+1) vector registers.

N and K may be specified as integer numbers or absolute expressions.

[v<N>, v<N+1>, … v<K>]

A sequence of (K-N+1) vector registers.

Register indices must be specified as decimal integer numbers.

Note: N and K must satisfy the following conditions:

  • N <= K.
  • 0 <= N <= 255.
  • 0 <= K <= 255.
  • K-N+1 must be equal to 1, 2, 3, 4, 8 or 16.

Examples:

v255
v[0]
v[0:1]
v[1:1]
v[0:3]
v[2*2]
v[1-1:2-1]
[v252]
[v252,v253,v254,v255]

GFX10 Image instructions may use special NSA (Non-Sequential Address) syntax for image addresses:

Syntax Description
[Vm, Vn, … Vk]

A sequence of 32-bit vector registers. Each register may be specified using a syntax defined above.

In contrast with standard syntax, registers in NSA sequence are not required to have consecutive indices. Moreover, the same register may appear in the list more than once.

Examples:

[v32,v1,v[2]]
[v[32],v[1:1],[v2]]
[v4,v4,v4,v4]

s

Scalar 32-bit registers. The number of available scalar registers depends on GPU:

GPU Number of scalar registers
GFX7 104
GFX8 102
GFX9 102
GFX10 106

A sequence of scalar registers may be used to operate with more than 32 bits of data. Assembler currently supports sequences of 1, 2, 4, 8 and 16 scalar registers.

Pairs of scalar registers must be even-aligned (the first register must be even). Sequences of 4 and more scalar registers must be quad-aligned.

Syntax Description
s<N>

A single 32-bit scalar register.

N must be a decimal integer number.

s[<N>]

A single 32-bit scalar register.

N may be specified as an integer number or an absolute expression.

s[<N>:<K>]

A sequence of (K-N+1) scalar registers.

N and K may be specified as integer numbers or absolute expressions.

[s<N>, s<N+1>, … s<K>]

A sequence of (K-N+1) scalar registers.

Register indices must be specified as decimal integer numbers.

Note: N and K must satisfy the following conditions:

  • N must be properly aligned based on sequence size.
  • N <= K.
  • 0 <= N < SMAX, where SMAX is the number of available scalar registers.
  • 0 <= K < SMAX, where SMAX is the number of available scalar registers.
  • K-N+1 must be equal to 1, 2, 4, 8 or 16.

Examples:

s0
s[0]
s[0:1]
s[1:1]
s[0:3]
s[2*2]
s[1-1:2-1]
[s4]
[s4,s5,s6,s7]

Examples of scalar registers with an invalid alignment:

s[1:2]
s[2:5]

trap

A set of trap handler registers:

ttmp

Trap handler temporary scalar registers, 32-bits wide. The number of available ttmp registers depends on GPU:

GPU Number of ttmp registers
GFX7 12
GFX8 12
GFX9 16
GFX10 16

A sequence of ttmp registers may be used to operate with more than 32 bits of data. Assembler currently supports sequences of 1, 2, 4, 8 and 16 ttmp registers.

Pairs of ttmp registers must be even-aligned (the first register must be even). Sequences of 4 and more ttmp registers must be quad-aligned.

Syntax Description
ttmp<N>

A single 32-bit ttmp register.

N must be a decimal integer number.

ttmp[<N>]

A single 32-bit ttmp register.

N may be specified as an integer number or an absolute expression.

ttmp[<N>:<K>]

A sequence of (K-N+1) ttmp registers.

N and K may be specified as integer numbers or absolute expressions.

[ttmp<N>, ttmp<N+1>, … ttmp<K>]

A sequence of (K-N+1) ttmp registers.

Register indices must be specified as decimal integer numbers.

Note: N and K must satisfy the following conditions:

  • N must be properly aligned based on sequence size.
  • N <= K.
  • 0 <= N < TMAX, where TMAX is the number of available ttmp registers.
  • 0 <= K < TMAX, where TMAX is the number of available ttmp registers.
  • K-N+1 must be equal to 1, 2, 4, 8 or 16.

Examples:

ttmp0
ttmp[0]
ttmp[0:1]
ttmp[1:1]
ttmp[0:3]
ttmp[2*2]
ttmp[1-1:2-1]
[ttmp4]
[ttmp4,ttmp5,ttmp6,ttmp7]

Examples of ttmp registers with an invalid alignment:

ttmp[1:2]
ttmp[2:5]

tba

Trap base address, 64-bits wide. Holds the pointer to the current trap handler program.

Syntax Description Availability
tba 64-bit trap base address register. GFX7, GFX8
[tba] 64-bit trap base address register (an SP3 syntax). GFX7, GFX8
[tba_lo,tba_hi] 64-bit trap base address register (an SP3 syntax). GFX7, GFX8

High and low 32 bits of trap base address may be accessed as separate registers:

Syntax Description Availability
tba_lo Low 32 bits of trap base address register. GFX7, GFX8
tba_hi High 32 bits of trap base address register. GFX7, GFX8
[tba_lo] Low 32 bits of trap base address register (an SP3 syntax). GFX7, GFX8
[tba_hi] High 32 bits of trap base address register (an SP3 syntax). GFX7, GFX8

Note that tba, tba_lo and tba_hi are not accessible as assembler registers in GFX9 and GFX10, but tba is readable/writable with the help of s_get_reg and s_set_reg instructions.

tma

Trap memory address, 64-bits wide.

Syntax Description Availability
tma 64-bit trap memory address register. GFX7, GFX8
[tma] 64-bit trap memory address register (an SP3 syntax). GFX7, GFX8
[tma_lo,tma_hi] 64-bit trap memory address register (an SP3 syntax). GFX7, GFX8

High and low 32 bits of trap memory address may be accessed as separate registers:

Syntax Description Availability
tma_lo Low 32 bits of trap memory address register. GFX7, GFX8
tma_hi High 32 bits of trap memory address register. GFX7, GFX8
[tma_lo] Low 32 bits of trap memory address register (an SP3 syntax). GFX7, GFX8
[tma_hi] High 32 bits of trap memory address register (an SP3 syntax). GFX7, GFX8

Note that tma, tma_lo and tma_hi are not accessible as assembler registers in GFX9 and GFX10, but tma is readable/writable with the help of s_get_reg and s_set_reg instructions.

flat_scratch

Flat scratch address, 64-bits wide. Holds the base address of scratch memory.

Syntax Description
flat_scratch 64-bit flat scratch address register.
[flat_scratch] 64-bit flat scratch address register (an SP3 syntax).
[flat_scratch_lo,flat_scratch_hi] 64-bit flat scratch address register (an SP3 syntax).

High and low 32 bits of flat scratch address may be accessed as separate registers:

Syntax Description
flat_scratch_lo Low 32 bits of flat scratch address register.
flat_scratch_hi High 32 bits of flat scratch address register.
[flat_scratch_lo] Low 32 bits of flat scratch address register (an SP3 syntax).
[flat_scratch_hi] High 32 bits of flat scratch address register (an SP3 syntax).

xnack

Xnack mask, 64-bits wide. Holds a 64-bit mask of which threads received an XNACK due to a vector memory operation.

Warning

GFX7 does not support xnack feature. For availability of this feature in other GPUs, refer this table.

Syntax Description
xnack_mask 64-bit xnack mask register.
[xnack_mask] 64-bit xnack mask register (an SP3 syntax).
[xnack_mask_lo,xnack_mask_hi] 64-bit xnack mask register (an SP3 syntax).

High and low 32 bits of xnack mask may be accessed as separate registers:

Syntax Description
xnack_mask_lo Low 32 bits of xnack mask register.
xnack_mask_hi High 32 bits of xnack mask register.
[xnack_mask_lo] Low 32 bits of xnack mask register (an SP3 syntax).
[xnack_mask_hi] High 32 bits of xnack mask register (an SP3 syntax).

vcc

Vector condition code, 64-bits wide. A bit mask with one bit per thread; it holds the result of a vector compare operation.

Note that GFX10 H/W does not use high 32 bits of vcc in wave32 mode.

Syntax Description
vcc 64-bit vector condition code register.
[vcc] 64-bit vector condition code register (an SP3 syntax).
[vcc_lo,vcc_hi] 64-bit vector condition code register (an SP3 syntax).

High and low 32 bits of vector condition code may be accessed as separate registers:

Syntax Description
vcc_lo Low 32 bits of vector condition code register.
vcc_hi High 32 bits of vector condition code register.
[vcc_lo] Low 32 bits of vector condition code register (an SP3 syntax).
[vcc_hi] High 32 bits of vector condition code register (an SP3 syntax).

m0

A 32-bit memory register. It has various uses, including register indexing and bounds checking.

Syntax Description
m0 A 32-bit memory register.
[m0] A 32-bit memory register (an SP3 syntax).

exec

Execute mask, 64-bits wide. A bit mask with one bit per thread, which is applied to vector instructions and controls which threads execute and which ignore the instruction.

Note that GFX10 H/W does not use high 32 bits of exec in wave32 mode.

Syntax Description
exec 64-bit execute mask register.
[exec] 64-bit execute mask register (an SP3 syntax).
[exec_lo,exec_hi] 64-bit execute mask register (an SP3 syntax).

High and low 32 bits of execute mask may be accessed as separate registers:

Syntax Description
exec_lo Low 32 bits of execute mask register.
exec_hi High 32 bits of execute mask register.
[exec_lo] Low 32 bits of execute mask register (an SP3 syntax).
[exec_hi] High 32 bits of execute mask register (an SP3 syntax).

vccz

A single bit flag indicating that the vcc is all zeros.

Note: when GFX10 operates in wave32 mode, this register reflects state of vcc_lo.

execz

A single bit flag indicating that the exec is all zeros.

Note: when GFX10 operates in wave32 mode, this register reflects state of exec_lo.

scc

A single bit flag indicating the result of a scalar compare operation.

lds_direct

A special operand which supplies a 32-bit value fetched from LDS memory using m0 as an address.

null

This is a special operand which may be used as a source or a destination.

When used as a destination, the result of the operation is discarded.

When used as a source, it supplies zero value.

GFX10 only.

Warning

Due to a H/W bug, this operand cannot be used with VALU instructions in first generation of GFX10.

inline constant

An inline constant is an integer or a floating-point value encoded as a part of an instruction. Compare inline constants with literals.

Inline constants include:

If a number may be encoded as either a literal or a constant, assembler selects the latter encoding as more efficient.

iconst

An integer number or an absolute expression encoded as an inline constant.

Only a small fraction of integer numbers may be encoded as inline constants. They are enumerated in the table below. Other integer numbers have to be encoded as literals.

Value Note
{0..64} Positive integer inline constants.
{-16..-1} Negative integer inline constants.

Warning

GFX7 does not support inline constants for f16 operands.

fconst

A floating-point number encoded as an inline constant.

Only a small fraction of floating-point numbers may be encoded as inline constants. They are enumerated in the table below. Other floating-point numbers have to be encoded as literals.

Value Note Availability
0.0 The same as integer constant 0. All GPUs
0.5 Floating-point constant 0.5 All GPUs
1.0 Floating-point constant 1.0 All GPUs
2.0 Floating-point constant 2.0 All GPUs
4.0 Floating-point constant 4.0 All GPUs
-0.5 Floating-point constant -0.5 All GPUs
-1.0 Floating-point constant -1.0 All GPUs
-2.0 Floating-point constant -2.0 All GPUs
-4.0 Floating-point constant -4.0 All GPUs
0.1592 1.0/(2.0*pi). Use only for 16-bit operands. GFX8, GFX9, GFX10
0.15915494 1.0/(2.0*pi). Use only for 16- and 32-bit operands. GFX8, GFX9, GFX10
0.15915494309189532 1.0/(2.0*pi). GFX8, GFX9, GFX10

Warning

GFX7 does not support inline constants for f16 operands.

ival

A symbolic operand encoded as an inline constant. These operands provide read-only access to H/W registers.

Syntax Note Availability
shared_base Base address of shared memory region. GFX9, GFX10
shared_limit Address of the end of shared memory region. GFX9, GFX10
private_base Base address of private memory region. GFX9, GFX10
private_limit Address of the end of private memory region. GFX9, GFX10
pops_exiting_wave_id A dedicated counter for POPS. GFX9, GFX10

literal

A literal is a 64-bit value encoded as a separate 32-bit dword in the instruction stream. Compare literals with inline constants.

If a number may be encoded as either a literal or an inline constant, assembler selects the latter encoding as more efficient.

Literals may be specified as integer numbers, floating-point numbers, absolute expressions or relocatable expressions.

An instruction may use only one literal but several operands may refer the same literal.

uimm8

A 8-bit integer number or an absolute expression. The value must be in the range 0..0xFF.

uimm32

A 32-bit integer number or an absolute expression. The value must be in the range 0..0xFFFFFFFF.

uimm20

A 20-bit integer number or an absolute expression.

The value must be in the range 0..0xFFFFF.

uimm21

A 21-bit integer number or an absolute expression.

The value must be in the range 0..0x1FFFFF.

Warning

Assembler currently supports 20-bit offsets only. Use uimm20 as a replacement.

simm21

A 21-bit integer number or an absolute expression.

The value must be in the range -0x100000..0x0FFFFF.

Warning

Assembler currently supports 20-bit unsigned offsets only. Use uimm20 as a replacement.

off

A special entity which indicates that the value of this operand is not used.

Syntax Description
off Indicates an unused operand.

Numbers

Integer Numbers

Integer numbers are 64 bits wide. They are converted to expected operand type as described here.

Integer numbers may be specified in binary, octal, hexadecimal and decimal formats:

Format Syntax Example
Decimal [-]?[1-9][0-9]* -1234
Binary [-]?0b[01]+ 0b1010
Octal [-]?0[0-7]+ 010
Hexadecimal [-]?0x[0-9a-fA-F]+ 0xff
[-]?[0x]?[0-9][0-9a-fA-F]*[hH] 0ffh

Floating-Point Numbers

All floating-point numbers are handled as double (64 bits wide). They are converted to expected operand type as described here.

Floating-point numbers may be specified in hexadecimal and decimal formats:

Format Syntax Examples Note
Decimal [-]?[0-9]*[.][0-9]*([eE][+-]?[0-9]*)? -1.234, 234e2 Must include either a decimal separator or an exponent.
Hexadecimal [-]0x[0-9a-fA-F]*(.[0-9a-fA-F]*)?[pP][+-]?[0-9a-fA-F]+ -0x1afp-10, 0x.1afp10  

Expressions

An expression is evaluated to a 64-bit integer. Note that floating-point expressions are not supported.

There are two kinds of expressions:

Absolute Expressions

The value of an absolute expression does not change after program relocation. Absolute expressions must not include unassigned and relocatable values such as labels.

Absolute expressions are evaluated to 64-bit integer values and converted to expected operand type as described here.

Examples:

x = -1
y = x + 10

Relocatable Expressions

The value of a relocatable expression depends on program relocation.

Note that use of relocatable expressions is limited with branch targets and 32-bit integer operands.

A relocatable expression is evaluated to a 64-bit integer value which depends on operand kind and relocation type of symbol(s) used in the expression. For example, if an instruction refers a label, this reference is evaluated to an offset from the address after the instruction to the label address:

label:
v_add_co_u32_e32 v0, vcc, label, v1  // 'label' operand is evaluated to -4

Note that values of relocatable expressions are usually unknown at assembly time; they are resolved later by a linker and converted to expected operand type as described here.

Operands and Operations

Expressions are composed of 64-bit integer operands and operations. Operands include integer numbers and symbols.

Expressions may also use “.” which is a reference to the current PC (program counter).

Unary and binary operations produce 64-bit integer results.

Syntax of Expressions

The syntax of expressions is shown below:

expr ::= expr binop expr | primaryexpr ;

primaryexpr ::= '(' expr ')' | symbol | number | '.' | unop primaryexpr ;

binop ::= '&&'
        | '||'
        | '|'
        | '^'
        | '&'
        | '!'
        | '=='
        | '!='
        | '<>'
        | '<'
        | '<='
        | '>'
        | '>='
        | '<<'
        | '>>'
        | '+'
        | '-'
        | '*'
        | '/'
        | '%' ;

unop ::= '~'
       | '+'
       | '-'
       | '!' ;

Binary Operators

Binary operators are described in the following table. They operate on and produce 64-bit integers. Operators with higher priority are performed first.

Operator Priority Meaning
* 5 Integer multiplication.
/ 5 Integer division.
% 5 Integer signed remainder.
+ 4 Integer addition.
- 4 Integer subtraction.
<< 3 Integer shift left.
>> 3 Logical shift right.
== 2 Equality comparison.
!= 2 Inequality comparison.
<> 2 Inequality comparison.
< 2 Signed less than comparison.
<= 2 Signed less than or equal comparison.
> 2 Signed greater than comparison.
>= 2 Signed greater than or equal comparison.
| 1 Bitwise or.
^ 1 Bitwise xor.
& 1 Bitwise and.
&& 0 Logical and.
|| 0 Logical or.

Unary Operators

Unary operators are described in the following table. They operate on and produce 64-bit integers.

Operator Meaning
! Logical negation.
~ Bitwise negation.
+ Integer unary plus.
- Integer unary minus.

Symbols

A symbol is a named 64-bit integer value, representing a relocatable address or an absolute (non-relocatable) number.

Symbol names have the following syntax:
[a-zA-Z_.][a-zA-Z0-9_$.@]*

The table below provides several examples of syntax used for symbol definition.

Syntax Meaning
.globl <S> Declares a global symbol S without assigning it a value.
.set <S>, <E> Assigns the value of an expression E to a symbol S.
<S> = <E> Assigns the value of an expression E to a symbol S.
<S>: Declares a label S and assigns it the current PC value.

A symbol may be used before it is declared or assigned; unassigned symbols are assumed to be PC-relative.

Additional information about symbols may be found here.

Type and Size Conversion

This section describes what happens when a 64-bit integer number, a floating-point number or an expression is used for an operand which has a different type or size.

Conversion of Integer Values

Instruction operands may be specified as 64-bit integer numbers or absolute expressions. These values are converted to the expected operand type using the following steps:

1. Validation. Assembler checks if the input value may be truncated without loss to the required truncation width (see the table below). There are two cases when this operation is enabled:

  • The truncated bits are all 0.
  • The truncated bits are all 1 and the value after truncation has its MSB bit set.

In all other cases assembler triggers an error.

2. Conversion. The input value is converted to the expected type as described in the table below. Depending on operand kind, this conversion is performed by either assembler or AMDGPU H/W (or both).

Expected type Truncation Width Conversion Description
i16, u16, b16 16 num.u16 Truncate to 16 bits.
i32, u32, b32 32 num.u32 Truncate to 32 bits.
i64 32 {-1,num.i32} Truncate to 32 bits and then sign-extend the result to 64 bits.
u64, b64 32 {0,num.u32} Truncate to 32 bits and then zero-extend the result to 64 bits.
f16 16 num.u16 Use low 16 bits as an f16 value.
f32 32 num.u32 Use low 32 bits as an f32 value.
f64 32 {num.u32,0} Use low 32 bits of the number as high 32 bits of the result; low 32 bits of the result are zeroed.

Examples of enabled conversions:

// GFX9

v_add_u16 v0, -1, 0                   // src0 = 0xFFFF
v_add_f16 v0, -1, 0                   // src0 = 0xFFFF (NaN)
                                      //
v_add_u32 v0, -1, 0                   // src0 = 0xFFFFFFFF
v_add_f32 v0, -1, 0                   // src0 = 0xFFFFFFFF (NaN)
                                      //
v_add_u16 v0, 0xff00, v0              // src0 = 0xff00
v_add_u16 v0, 0xffffffffffffff00, v0  // src0 = 0xff00
v_add_u16 v0, -256, v0                // src0 = 0xff00
                                      //
s_bfe_i64 s[0:1], 0xffefffff, s3      // src0 = 0xffffffffffefffff
s_bfe_u64 s[0:1], 0xffefffff, s3      // src0 = 0x00000000ffefffff
v_ceil_f64_e32 v[0:1], 0xffefffff     // src0 = 0xffefffff00000000 (-1.7976922776554302e308)
                                      //
x = 0xffefffff                        //
s_bfe_i64 s[0:1], x, s3               // src0 = 0xffffffffffefffff
s_bfe_u64 s[0:1], x, s3               // src0 = 0x00000000ffefffff
v_ceil_f64_e32 v[0:1], x              // src0 = 0xffefffff00000000 (-1.7976922776554302e308)

Examples of disabled conversions:

// GFX9

v_add_u16 v0, 0x1ff00, v0               // truncated bits are not all 0 or 1
v_add_u16 v0, 0xffffffffffff00ff, v0    // truncated bits do not match MSB of the result

Conversion of Floating-Point Values

Instruction operands may be specified as 64-bit floating-point numbers. These values are converted to the expected operand type using the following steps:

1. Validation. Assembler checks if the input f64 number can be converted to the required floating-point type (see the table below) without overflow or underflow. Precision lost is allowed. If this conversion is not possible, assembler triggers an error.

2. Conversion. The input value is converted to the expected type as described in the table below. Depending on operand kind, this is performed by either assembler or AMDGPU H/W (or both).

Expected type Required FP Type Conversion Description
i16, u16, b16 f16 f16(num) Convert to f16 and use bits of the result as an integer value.
i32, u32, b32 f32 f32(num) Convert to f32 and use bits of the result as an integer value.
i64, u64, b64 - - Conversion disabled.
f16 f16 f16(num) Convert to f16.
f32 f32 f32(num) Convert to f32.
f64 f64 {num.u32.hi,0}

Use high 32 bits of the number as high 32 bits of the result; zero-fill low 32 bits of the result.

Note that the result may differ from the original number.

Examples of enabled conversions:

// GFX9

v_add_f16 v0, 1.0, 0        // src0 = 0x3C00 (1.0)
v_add_u16 v0, 1.0, 0        // src0 = 0x3C00
                            //
v_add_f32 v0, 1.0, 0        // src0 = 0x3F800000 (1.0)
v_add_u32 v0, 1.0, 0        // src0 = 0x3F800000

                            // src0 before conversion:
                            //   1.7976931348623157e308 = 0x7fefffffffffffff
                            // src0 after conversion:
                            //   1.7976922776554302e308 = 0x7fefffff00000000
v_ceil_f64 v[0:1], 1.7976931348623157e308

v_add_f16 v1, 65500.0, v2   // ok for f16.
v_add_f32 v1, 65600.0, v2   // ok for f32, but would result in overflow for f16.

Examples of disabled conversions:

// GFX9

v_add_f16 v1, 65600.0, v2    // overflow

Conversion of Relocatable Values

Relocatable expressions may be used with 32-bit integer operands and jump targets.

When the value of a relocatable expression is resolved by a linker, it is converted as needed and truncated to the operand size. The conversion depends on relocation type and operand kind.

For example, when a 32-bit operand of an instruction refers a relocatable expression expr, this reference is evaluated to a 64-bit offset from the address after the instruction to the address being referenced, counted in bytes. Then the value is truncated to 32 bits and encoded as a literal:

expr = .
v_add_co_u32_e32 v0, vcc, expr, v1  // 'expr' operand is evaluated to -4
                                    // and then truncated to 0xFFFFFFFC

As another example, when a branch instruction refers a label, this reference is evaluated to an offset from the address after the instruction to the label address, counted in dwords. Then the value is truncated to 16 bits:

label:
s_branch label  // 'label' operand is evaluated to -1 and truncated to 0xFFFF