Skip to content

Latest commit

 

History

History
1248 lines (844 loc) · 42.8 KB

AMDGPUModifierSyntax.rst

File metadata and controls

1248 lines (844 loc) · 42.8 KB

Syntax of AMDGPU Instruction Modifiers

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.

Specifies an immediate unsigned 8-bit offset, in bytes. The default value is 0.

Used with DS instructions which have 2 addresses.

Syntax Description
offset:{0..0xFF} Specifies an unsigned 8-bit offset as a positive :ref:`integer number <amdgpu_synid_integer_number>`.

Examples:

offset:255
offset:0xff

Specifies an immediate unsigned 16-bit offset, in bytes. The default value is 0.

Used with DS instructions which have 1 address.

Syntax Description
offset:{0..0xFFFF} Specifies an unsigned 16-bit offset as a positive :ref:`integer number <amdgpu_synid_integer_number>`.

Examples:

offset:65535
offset:0xffff

This is a special modifier which may be used with ds_swizzle_b32 instruction only. It specifies a swizzle pattern in numeric or symbolic form. The default value is 0.

See AMD documentation for more information.

Syntax Description
offset:{0..0xFFFF} Specifies a 16-bit swizzle pattern.
offset:swizzle(QUAD_PERM,{0..3},{0..3},{0..3},{0..3})

Specifies a quad permute mode pattern

Each number is a lane id.

offset:swizzle(BITMASK_PERM, "<mask>")

Specifies a bitmask permute mode pattern.

The pattern converts a 5-bit lane id to another lane id with which the lane interacts.

mask is a 5 character sequence which specifies how to transform the bits of the lane id.

The following characters are allowed:

  • "0" - set bit to 0.
  • "1" - set bit to 1.
  • "p" - preserve bit.
  • "i" - inverse bit.
offset:swizzle(BROADCAST,{2..32},{0..N})

Specifies a broadcast mode.

Broadcasts the value of any particular lane to all lanes in its group.

The first numeric parameter is a group size and must be equal to 2, 4, 8, 16 or 32.

The second numeric parameter is an index of the lane being broadcasted.

The index must not exceed group size.

offset:swizzle(SWAP,{1..16})

Specifies a swap mode.

Swaps the neighboring groups of 1, 2, 4, 8 or 16 lanes.

offset:swizzle(REVERSE,{2..32})

Specifies a reverse mode.

Reverses the lanes for groups of 2, 4, 8, 16 or 32 lanes.

Numeric parameters may be specified as either :ref:`integer numbers<amdgpu_synid_integer_number>` or :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.

Examples:

offset:255
offset:0xffff
offset:swizzle(QUAD_PERM, 0, 1, 2 ,3)
offset:swizzle(BITMASK_PERM, "01pi0")
offset:swizzle(BROADCAST, 2, 0)
offset:swizzle(SWAP, 8)
offset:swizzle(REVERSE, 30 + 2)

Specifies whether to use GDS or LDS memory (LDS is the default).

Syntax Description
gds Use GDS memory.

Specifies if this is the last export from the shader to the target. By default, current instruction does not finish an export sequence.

Syntax Description
done Indicates the last export operation.

Indicates if the data are compressed (data are not compressed by default).

Syntax Description
compr Data are compressed.

Specifies valid mask flag state (off by default).

Syntax Description
vm Set valid mask flag.

Specifies an immediate unsigned 12-bit offset, in bytes. The default value is 0.

Cannot be used with global/scratch opcodes. GFX9 only.

Syntax Description
offset:{0..4095} Specifies a 12-bit unsigned offset as a positive :ref:`integer number <amdgpu_synid_integer_number>`.

Examples:

offset:4095
offset:0xff

Specifies an immediate signed 13-bit offset, in bytes. The default value is 0.

Can be used with global/scratch opcodes only. GFX9 only.

Syntax Description
offset:{-4096..4095} Specifies a 13-bit signed offset as an :ref:`integer number <amdgpu_synid_integer_number>`.

Examples:

offset:-4000
offset:0x10

See a description :ref:`here<amdgpu_synid_glc>`.

See a description :ref:`here<amdgpu_synid_slc>`.

See a description :ref:`here<amdgpu_synid_tfe>`.

See a description :ref:`here<amdgpu_synid_nv>`.

Specifies which channels (image components) are used by the operation. By default, no channels are used.

Syntax Description
dmask:{0..15}

Specifies image channels as a positive :ref:`integer number <amdgpu_synid_integer_number>`.

Each bit corresponds to one of 4 image components (RGBA).

If the specified bit value is 0, the component is not used, value 1 means that the component is used.

This modifier has some limitations depending on instruction kind:

Instruction Kind Valid dmask Values
32-bit atomic cmpswap 0x3
32-bit atomic instructions except for cmpswap 0x1
64-bit atomic cmpswap 0xF
64-bit atomic instructions except for cmpswap 0x3
gather4 0x1, 0x2, 0x4, 0x8
Other instructions any value

Examples:

dmask:0xf
dmask:0b1111
dmask:3

Specifies whether the address is normalized or not (the address is normalized by default).

Syntax Description
unorm Force the address to be unnormalized.

See a description :ref:`here<amdgpu_synid_glc>`.

See a description :ref:`here<amdgpu_synid_slc>`.

Specifies texture resource size. The default size is 256 bits.

GFX7 and GFX8 only.

Syntax Description
r128 Specifies 128 bits texture resource size.

Warning

Using this modifier should descrease rsrc operand size from 8 to 4 dwords, but assembler does not currently support this feature.

See a description :ref:`here<amdgpu_synid_tfe>`.

Specifies LOD warning status (LOD warning is disabled by default).

Syntax Description
lwe Enables LOD warning.

Specifies if an array index must be sent to TA. By default, array index is not sent.

Syntax Description
da Send an array-index to TA.

Specifies data size: 16 or 32 bits (32 bits by default). Not supported by GFX7.

Syntax Description
d16

Enables 16-bits data mode.

On loads, convert data in memory to 16-bit format before storing it in VGPRs.

For stores, convert 16-bit data in VGPRs to 32 bits before going to memory.

Note that GFX8.0 does not support data packing. Each 16-bit data element occupies 1 VGPR.

GFX8.1 and GFX9 support data packing. Each pair of 16-bit data elements occupies 1 VGPR.

Specifies size of image address components: 16 or 32 bits (32 bits by default). GFX9 only.

Syntax Description
a16 Enables 16-bits image address components.

This modifier has different meaning for loads, stores, and atomic operations. The default value is off (0).

See AMD documentation for details.

Syntax Description
glc Set glc bit to 1.

Specifies cache policy. The default value is off (0).

See AMD documentation for details.

Syntax Description
slc Set slc bit to 1.

Controls access to partially resident textures. The default value is off (0).

See AMD documentation for details.

Syntax Description
tfe Set tfe bit to 1.

Specifies if instruction is operating on non-volatile memory. By default, memory is volatile.

GFX9 only.

Syntax Description
nv Indicates that instruction operates on non-volatile memory.

Specifies whether address components include an index. By default, no components are used.

Can be used together with :ref:`offen<amdgpu_synid_offen>`.

Cannot be used with :ref:`addr64<amdgpu_synid_addr64>`.

Syntax Description
idxen Address components include an index.

Specifies whether address components include an offset. By default, no components are used.

Can be used together with :ref:`idxen<amdgpu_synid_idxen>`.

Cannot be used with :ref:`addr64<amdgpu_synid_addr64>`.

Syntax Description
offen Address components include an offset.

Specifies whether a 64-bit address is used. By default, no address is used.

GFX7 only. Cannot be used with :ref:`offen<amdgpu_synid_offen>` and :ref:`idxen<amdgpu_synid_idxen>` modifiers.

Syntax Description
addr64 A 64-bit address is used.

Specifies an immediate unsigned 12-bit offset, in bytes. The default value is 0.

Syntax Description
offset:{0..0xFFF} Specifies a 12-bit unsigned offset as a positive :ref:`integer number <amdgpu_synid_integer_number>`.

Examples:

offset:0
offset:0x10

See a description :ref:`here<amdgpu_synid_glc>`.

See a description :ref:`here<amdgpu_synid_slc>`.

Specifies where to store the result: VGPRs or LDS (VGPRs by default).

Syntax Description
lds Store result in LDS.

See a description :ref:`here<amdgpu_synid_tfe>`.

TBD

TBD

See a description :ref:`here<amdgpu_synid_glc>`.

See a description :ref:`here<amdgpu_synid_nv>`.

Specifies which half of the LDS word to use. Low half of LDS word is used by default. GFX9 only.

Syntax Description
high Use high half of LDS word.

GFX8 and GFX9 only.

Specifies how data are shared between threads. This is a mandatory modifier. There is no default value.

Note. The lanes of a wavefront are organized in four banks and four rows.

Syntax Description
quad_perm:[{0..3},{0..3},{0..3},{0..3}] Full permute of 4 threads.
row_mirror Mirror threads within row.
row_half_mirror Mirror threads within 1/2 row (8 threads).
row_bcast:15 Broadcast 15th thread of each row to next row.
row_bcast:31 Broadcast thread 31 to rows 2 and 3.
wave_shl:1 Wavefront left shift by 1 thread.
wave_rol:1 Wavefront left rotate by 1 thread.
wave_shr:1 Wavefront right shift by 1 thread.
wave_ror:1 Wavefront right rotate by 1 thread.
row_shl:{1..15} Row shift left by 1-15 threads.
row_shr:{1..15} Row shift right by 1-15 threads.
row_ror:{1..15} Row rotate right by 1-15 threads.

Note: Numeric parameters may be specified as either :ref:`integer numbers<amdgpu_synid_integer_number>` or :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.

Examples:

quad_perm:[0, 1, 2, 3]
row_shl:3

Controls which rows are enabled for data sharing. By default, all rows are enabled.

Note. The lanes of a wavefront are organized in four banks and four rows.

Syntax Description
row_mask:{0..15}

Specifies a row mask as a positive :ref:`integer number <amdgpu_synid_integer_number>`.

Each of 4 bits in the mask controls one row (0 - disabled, 1 - enabled).

Examples:

row_mask:0xf
row_mask:0b1010
row_mask:0b1111

Controls which banks are enabled for data sharing. By default, all banks are enabled.

Note. The lanes of a wavefront are organized in four banks and four rows.

Syntax Description
bank_mask:{0..15}

Specifies a bank mask as a positive :ref:`integer number <amdgpu_synid_integer_number>`.

Each of 4 bits in the mask controls one bank (0 - disabled, 1 - enabled).

Examples:

bank_mask:0x3
bank_mask:0b0011
bank_mask:0b1111

Controls data sharing when accessing an invalid lane. By default, data sharing with invalid lanes is disabled.

Syntax Description
bound_ctrl:0

Enables data sharing with invalid lanes.

Accessing data from an invalid lane will return zero.

GFX8 and GFX9 only.

See a description :ref:`here<amdgpu_synid_clamp>`.

See a description :ref:`here<amdgpu_synid_omod>`.

GFX9 only.

Selects which bits in the destination are affected. By default, all bits are affected.

Syntax Description
dst_sel:DWORD Use bits 31:0.
dst_sel:BYTE_0 Use bits 7:0.
dst_sel:BYTE_1 Use bits 15:8.
dst_sel:BYTE_2 Use bits 23:16.
dst_sel:BYTE_3 Use bits 31:24.
dst_sel:WORD_0 Use bits 15:0.
dst_sel:WORD_1 Use bits 31:16.

Controls what to do with the bits in the destination which are not selected by :ref:`dst_sel<amdgpu_synid_dst_sel>`. By default, unused bits are preserved.

Syntax Description
dst_unused:UNUSED_PAD Pad with zeros.
dst_unused:UNUSED_SEXT Sign-extend upper bits, zero lower bits.
dst_unused:UNUSED_PRESERVE Preserve bits.

Controls which bits in the src0 are used. By default, all bits are used.

Syntax Description
src0_sel:DWORD Use bits 31:0.
src0_sel:BYTE_0 Use bits 7:0.
src0_sel:BYTE_1 Use bits 15:8.
src0_sel:BYTE_2 Use bits 23:16.
src0_sel:BYTE_3 Use bits 31:24.
src0_sel:WORD_0 Use bits 15:0.
src0_sel:WORD_1 Use bits 31:16.

Controls which bits in the src1 are used. By default, all bits are used.

Syntax Description
src1_sel:DWORD Use bits 31:0.
src1_sel:BYTE_0 Use bits 7:0.
src1_sel:BYTE_1 Use bits 15:8.
src1_sel:BYTE_2 Use bits 23:16.
src1_sel:BYTE_3 Use bits 31:24.
src1_sel:WORD_0 Use bits 15:0.
src1_sel:WORD_1 Use bits 31:16.

Operand modifiers are not used separately. They are applied to source operands.

GFX8 and GFX9 only.

See a description :ref:`here<amdgpu_synid_abs>`.

See a description :ref:`here<amdgpu_synid_neg>`.

Sign-extends value of a (sub-dword) operand to fill all 32 bits. Has no effect for 32-bit operands.

Valid for integer operands only.

Syntax Description
sext(<operand>) Sign-extend operand value.

Examples:

sext(v4)
sext(v255)

Selects the low [15:0] or high [31:16] operand bits for source and destination operands. By default, low bits are used for all operands.

The number of values specified with the op_sel modifier must match the number of instruction operands (both source and destination). First value controls src0, second value controls src1 and so on, except that the last value controls destination. The value 0 selects the low bits, while 1 selects the high bits.

Note. op_sel modifier affects 16-bit operands only. For 32-bit operands the value specified by op_sel must be 0.

GFX9 only.

Syntax Description
op_sel:[{0..1},{0..1}] Select operand bits for instructions with 1 source operand.
op_sel:[{0..1},{0..1},{0..1}] Select operand bits for instructions with 2 source operands.
op_sel:[{0..1},{0..1},{0..1},{0..1}] Select operand bits for instructions with 3 source operands.

Examples:

op_sel:[0,0]
op_sel:[0,1]

Clamp meaning depends on instruction.

For v_cmp instructions, clamp modifier indicates that the compare signals if a floating point exception occurs. By default, signaling is disabled. Not supported by GFX7.

For integer operations, clamp modifier indicates that the result must be clamped to the largest and smallest representable value. By default, there is no clamping. Integer clamping is not supported by GFX7.

For floating point operations, clamp modifier indicates that the result must be clamped to the range [0.0, 1.0]. By default, there is no clamping.

Note. Clamp modifier is applied after :ref:`output modifiers<amdgpu_synid_omod>` (if any).

Syntax Description
clamp Enables clamping (or signaling).

Specifies if an output modifier must be applied to the result. By default, no output modifiers are applied.

Note. Output modifiers are applied before :ref:`clamping<amdgpu_synid_clamp>` (if any).

Output modifiers are valid for f32 and f64 floating point results only. They must not be used with f16.

Note. v_cvt_f16_f32 is an exception. This instruction produces f16 result but accepts output modifiers.

Syntax Description
mul:2 Multiply the result by 2.
mul:4 Multiply the result by 4.
div:2 Multiply the result by 0.5.

Operand modifiers are not used separately. They are applied to source operands.

Computes absolute value of its operand. Applied before :ref:`neg<amdgpu_synid_neg>` (if any). Valid for floating point operands only.

Syntax Description
abs(<operand>) Get absolute value of operand.
|<operand>| The same as above.

Examples:

abs(v36)
|v36|

Computes negative value of its operand. Applied after :ref:`abs<amdgpu_synid_abs>` (if any). Valid for floating point operands only.

Syntax Description
neg(<operand>) Get negative value of operand.
-<operand> The same as above.

Examples:

neg(v[0])
-v4

This section describes modifiers of regular VOP3P instructions.

v_mad_mix_f32, v_mad_mixhi_f16 and v_mad_mixlo_f16 instructions use these modifiers :ref:`in a special manner<amdgpu_synid_mad_mix>`.

GFX9 only.

Selects the low [15:0] or high [31:16] operand bits as input to the operation which results in the lower-half of the destination. By default, low bits are used for all operands.

The number of values specified by the op_sel modifier must match the number of source operands. First value controls src0, second value controls src1 and so on.

The value 0 selects the low bits, while 1 selects the high bits.

Syntax Description
op_sel:[{0..1}] Select operand bits for instructions with 1 source operand.
op_sel:[{0..1},{0..1}] Select operand bits for instructions with 2 source operands.
op_sel:[{0..1},{0..1},{0..1}] Select operand bits for instructions with 3 source operands.

Examples:

op_sel:[0,0]
op_sel:[0,1,0]

Selects the low [15:0] or high [31:16] operand bits as input to the operation which results in the upper-half of the destination. By default, high bits are used for all operands.

The number of values specified by the op_sel_hi modifier must match the number of source operands. First value controls src0, second value controls src1 and so on.

The value 0 selects the low bits, while 1 selects the high bits.

Syntax Description
op_sel_hi:[{0..1}] Select operand bits for instructions with 1 source operand.
op_sel_hi:[{0..1},{0..1}] Select operand bits for instructions with 2 source operands.
op_sel_hi:[{0..1},{0..1},{0..1}] Select operand bits for instructions with 3 source operands.

Examples:

op_sel_hi:[0,0]
op_sel_hi:[0,0,1]

Specifies whether to change sign of operand values selected by :ref:`op_sel<amdgpu_synid_op_sel>`. These values are then used as input to the operation which results in the upper-half of the destination.

The number of values specified by this modifier must match the number of source operands. First value controls src0, second value controls src1 and so on.

The value 0 indicates that the corresponding operand value is used unmodified, the value 1 indicates that negative value of the operand must be used.

By default, operand values are used unmodified.

This modifier is valid for floating point operands only.

Syntax Description
neg_lo:[{0..1}] Select affected operands for instructions with 1 source operand.
neg_lo:[{0..1},{0..1}] Select affected operands for instructions with 2 source operands.
neg_lo:[{0..1},{0..1},{0..1}] Select affected operands for instructions with 3 source operands.

Examples:

neg_lo:[0]
neg_lo:[0,1]

Specifies whether to change sign of operand values selected by :ref:`op_sel_hi<amdgpu_synid_op_sel_hi>`. These values are then used as input to the operation which results in the upper-half of the destination.

The number of values specified by this modifier must match the number of source operands. First value controls src0, second value controls src1 and so on.

The value 0 indicates that the corresponding operand value is used unmodified, the value 1 indicates that negative value of the operand must be used.

By default, operand values are used unmodified.

This modifier is valid for floating point operands only.

Syntax Description
neg_hi:[{0..1}] Select affected operands for instructions with 1 source operand.
neg_hi:[{0..1},{0..1}] Select affected operands for instructions with 2 source operands.
neg_hi:[{0..1},{0..1},{0..1}] Select affected operands for instructions with 3 source operands.

Examples:

neg_hi:[1,0]
neg_hi:[0,1,1]

See a description :ref:`here<amdgpu_synid_clamp>`.

v_mad_mix_f32, v_mad_mixhi_f16 and v_mad_mixlo_f16 instructions use op_sel and op_sel_hi modifiers in a manner different from regular VOP3P instructions.

See a description below.

GFX9 only.

This operand has meaning only for 16-bit source operands as indicated by :ref:`m_op_sel_hi<amdgpu_synid_mad_mix_op_sel_hi>`. It specifies to select either the low [15:0] or high [31:16] operand bits as input to the operation.

The number of values specified by the op_sel modifier must match the number of source operands. First value controls src0, second value controls src1 and so on.

The value 0 indicates the low bits, the value 1 indicates the high 16 bits.

By default, low bits are used for all operands.

Syntax Description
op_sel:[{0..1},{0..1},{0..1}] Select location of each 16-bit source operand.

Examples:

op_sel:[0,1]

Selects the size of source operands: either 32 bits or 16 bits. By default, 32 bits are used for all source operands.

The number of values specified by the op_sel_hi modifier must match the number of source operands. First value controls src0, second value controls src1 and so on.

The value 0 indicates 32 bits, the value 1 indicates 16 bits.

The location of 16 bits in the operand may be specified by :ref:`m_op_sel<amdgpu_synid_mad_mix_op_sel>`.

Syntax Description
op_sel_hi:[{0..1},{0..1},{0..1}] Select size of each source operand.

Examples:

op_sel_hi:[1,1,1]

See a description :ref:`here<amdgpu_synid_abs>`.

See a description :ref:`here<amdgpu_synid_neg>`.

See a description :ref:`here<amdgpu_synid_clamp>`.