SGX543

General Info
It looks like instructions are 8 bytes long. Roughly speaking - first 4 bytes contain opcode and addressing mode. Second 4 bytes contain operands encoding.

Bit encoding used in this reference:

Predicates
Not sure about predicates yet, but they are used to mask execution of certain instructions.

Notation is the following:

For example:

To reduce amount of examples - they are not listed with predicates.

It is assumed that all predicates are applicable to all instructions in the group unless opposite is indicated.

Operands
Different types of operands exist. They are described in further sections:


 * Registers


 * Immediates


 * Indexes


 * Constants

Instructions may have up to four operands specified.

In this documentation they will be encoded as:

Operand 0
Destination operand 0 can be encoded in different ways.

Usually the following fields are used to encode it:


 * alt_opt0 - alter opt0. this bit can be combined with opt0 to produce the following modes for op0:


 * opt0 - type of operand op0, encoded with Register Selector Indexable RSI2.

or selects other modes for encoding op0 if specified in alt_opt0.


 * op0 - encoded with Register R6.

or with Register Index Offset RIO6 using index1 mode if specified in opt0.

or with Constant CNST6 if specified in alt_opt0.

or with Index IDX6 if specified in alt_opt0.

or with Register Index Offset RIO6 using index2 mode if specified in alt_opt0.

Operand N
Source operand  can be encoded in different ways.

Usually the following fields are used to encode it:


 * alt_opt - alter opt. this bit can be combined with opt to produce the following modes for op:


 * opt - type of operand op, encoded with Register Selector RS2.

or selects other modes for encoding op if specified in alt_opt.


 * op - encoded with Register R6.

or with Register Index Offset RIO6 using index1 mode if specified in alt_opt.

or with CNST6 if specified in alt_opt.

or IMM6 if specified in alt_opt.

or with Register Index Offset RIO6 using index2 mode if specified in alt_opt.

Registers

 * pa - primary attribute register. 32 bit long.
 * sa - secondary attribute register. 32 bit long.
 * o - output register. 32 bit long.
 * r - temporary register. 32 bit long.
 * i - internal register. 128 bit long.

Register Selector RS2
This encoding uses 2 bits to encode register type.

selector is encoded as:

Note that internal registers are not encoded - they are reserved in Register R6

Register Selector Indexable RSI2
This encoding uses 2 bits to encode register type.

selector is encoded as:

When index mode is used - there has to be another field that encodes index expression with Register Index Offset RIO6

The way that index expression is buit:

Example:

Register R6
This encoding uses 6 bits to encode register index.

Specific type of register can be selected with Register Selector RS2.

For destination operand op0 specific type of register can be selected with Register Selector Indexable RSI2.

index is calculated as: value * 2

Register expression is built as:

Example:

Register Encoding - r
register is encoded as:

Last 4 values are reserved for internal registers i0, i1, i2, i3

Register Encoding - o, pa, sa
register is encoded as:

Register R7
This encoding uses 7 bits to encode register index.

Specific type of register can be selected with Register Selector RS2.

For destination operand op0 specific type of register can be selected with Register Selector Indexable RSI2.

index is calculated as: value * 2

Register expression is built as:

Example:

Register Encoding - r
register is encoded as:

Last 4 values are reserved for internal registers i0, i1, i2, i3

Register Encoding - o, pa, sa
register is encoded as:

Register RI2
This encoding uses 2 bits to encode internal register.

Register Index Offset RIO6
rt is encoded as Register Selector RS2

offset is calculated as: value * 2

offset is encoded as:

Immediate IMM6
Some operands may act as immediate values which are encoded using 6 bits.

Indexes
Not sure what  expression means.

It can be used in 2 places:


 * RSI2


 * Alternative mode for operand 0

Index IDX6
Uses 6 bits to encode  expression.

Index is calculated as value * 2. Max index is 126.

Example:

Constant CNST6
Some operands may act as constant values which are encoded using 6 bits.

Constants are taken from table below.

Constants differ in 32 and 16 bit mode.

32 bit mode
bank 1 is used for channel 1 of op0

for other operands - only bank 0 is used for each channel

16 bit mode
Table for 16 bit mode does not have accurate values.

bank 1, bank 2, bank 3 is used for channel 1, channel 2, channel 3 of op0

for other operands - only bank 0 is used for each channel

Swizzle notation
There are 2 notations:


 * text notation


 * constant notation

When some of channels have constants - text notation is used

When all channels have constants - constant notation is used

When channel is masked in text notation it is marked as

When channel is masked in constant notation it is replaced with zero

Register Swizzle RSWZ2
This encoding uses 2 bits to encode the swizzle.

Usually combinations are additinally controlled by 1 or more bits called swz_alt_op

This type of swizzling does not allow precise control on each channel as opposed to RSWZ3

Usually there is a predefined table of swizzles.

Swizzle expression is built as:

Example:

Register Swizzle RSWZ3
This encoding uses 3 bits to encode the mask.

channel is encoded as:

swizzle expression is built as:

Example:

Modifier and dest data format
At the moment it is not known which of the data format fields is dest and which is source.

This is the reason why term modifier is mixed with term dest data format.

Instructions
mad

Encoding
Higher 4 bytes

Lower 4 bytes

Fields - operands

 * swz_alt_op1 - alter op1 swizzle. consult Swizzles f32 or Swizzles f16.
 * alt_opt0 - consult Operand 0
 * abs_op1 - add abs modifier to op1. example:
 * alt_opt2 - consult Operand N.
 * alt_opt3 - consult Operand N.
 * swz_alt_op3 - alter op3 swizzle. consult Swizzles f32 or Swizzles f16.
 * op3_swz - op3 swizzle encoded with Register Swizzle RSWZ2.
 * swz_alt_op2 - alter op2 swizzle. consult Swizzles f32 or Swizzles f16.
 * swz_mask16 - mask swizzles. consult Swizzle masking.
 * swz_mask32 - mask swizzles. consult Swizzle masking.
 * swz_en - enables swizzling and controls swizzle masking. consult Swizzle masking.
 * abs_op2 - add abs modifier to op2. example:
 * neg_op2 - negate op2. example:
 * abs_op3 - add abs modifier to op3. example:
 * neg_op3 - negate op3. example:
 * opt1 - when enabled - selects pa register type. when disabled - selects r register type.
 * opt0 - consult Operand 0.
 * opt2 - consult Operand N.
 * opt3 - consult Operand N.
 * op0 - consult Operand 0.
 * op2_swz - op2 swizzle encoded with Register Swizzle RSWZ2. consult Swizzles f32 or Swizzles f16.
 * op1_swz - op1 swizzle encoded with Register Swizzle RSWZ2. consult Swizzles f32 or Swizzles f16.
 * op1 - encoded with Register R6
 * op2 - consult Operand N.
 * op3 - consult Operand N.

Constants
Specific operand may be used as float constant. This can be achieved with following groups of bits:


 * alt_opt0, opt0, op0
 * alt_opt2, opt2, op2
 * alt_opt3, opt3, op3

Float constants can only be used when swizzling is enabled for particular operand. Consider checking sections Swizzles_f32 and Swizzles_f16.

Constants are taken from tables Constants.

Constants differ between 32 and 16 bit mode.

Swizzle masking
Masking is controled by control bits:
 * control bits: swz_en, swz_mask32, swz_mask16

Each channel can be masked with control bits. Combinations of control bits produce the following masking table.

Encoding used in masking table:

Masking table 32 bit mode:

Masking table 16 bit mode:

Swizzles f32
Swizzles of operand 1, operand 2 and operand 3 can not be precisely controlled and have predefined combinations.

Swizzles are controlled with bits:


 * swizzle fields: op1_swz, op2_swz, op3_swz
 * control bits: swz_alt_op1, swz_alt_op2, swz_alt_op3

Swizzles of operand 0 can not be controlled.

Swizzles f16
Swizzles of operand 1, operand 2 and operand 3 can not be precisely controlled and have predefined combinations.

Swizzles are controlled with bits:


 * swizzle fields: op1_swz, op2_swz, op3_swz
 * control bits: swz_alt_op1, swz_alt_op2, swz_alt_op3

Swizzles of operand 0 can not be controlled.

Instructions
mul.f32, add.f32, frc.f32, dsx.f32, dsy.f32, min.f32, max.f32, dot.f32

Encoding
Higher 4 bytes

Lower 4 bytes

Fields - operands

 * op1_swz_c3x - operand 1 swizzling channel 3 bit 1, 2. encoded as RSWZ3. consult Swizzles - operand 1.
 * alt_opt0 - consult Operand 0.
 * op1_swz_c30 - operand 1 swizzling channel 3 bit 0. encoded as RSWZ3. consult Swizzles - operand 1.
 * alt_opt1 - consult Operand N.
 * alt_opt2 - consult Operand N.
 * swz_alt_op2 - change op2 swizzle. consult Swizzles - operand 2.
 * op2_swz - op2 swizzle encoded with Register Swizzle RSWZ2. consult Swizzles - operand 2.
 * swz_mask3 - mask swizzle. consult Swizzle masking.
 * swz_mask2 - mask swizzle. consult Swizzle masking.
 * swz_mask1 - mask swizzle. consult Swizzle masking.
 * swz_en - enables usage of swizzling. consult Swizzle masking.
 * abs_op1 - add abs modifier to op1.
 * neg_op1 - negate op1.
 * abs_op2 - add abs modifier to op2.
 * op1_swz_c2x - operand 1 swizzling channel 2 bit 1, 2. encoded as RSWZ3. consult Swizzles - operand 1.
 * opt0 - consult Operand 0.
 * opt1 - consult Operand N.
 * opt2 - consult Operand N.
 * op0 - consult Operand 0.
 * op1_swz_c20 - operand 1 swizzling channel 2 bit 0. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1_swz_c1 - operand 1 swizzling channel 1. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1_swz_c0 - operand 1 swizzling channel 0. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1 - consult Operand N.
 * op2 - consult Operand N.

Constants
Specific operand may be used as float constant. This can be achieved with following groups of bits:


 * alt_opt0, opt0, op0
 * alt_opt1, opt1, op1
 * alt_opt2, opt2, op2

Float constants can only be used when swizzling is enabled for particular operand. Consider checking sections Swizzle_masking.

Constants are taken from tables Constants.

Constants correspond to table for 32 bit mode.

Swizzle masking
Masking is controled by control bits:
 * control bits: swz_en, swz_mask1, swz_mask2, swz_mask3

Each channel can be masked with control bits. Combinations of control bits produce the following masking table.

''dot.f32 instruction has explicit swizzling in operand 1 and operand 2 so masking does not apply to these operands. number of channels is always 4.''

Encoding used in masking table:

Masking table:

Swizzles - operand 0
Swizzles of operand 0 can not be controled and have predefined combinations described below:

Each channel can be masked with control bits. Masking is described in Swizzle_masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

Swizzles - operand 1
Each channel of operand 1 can be precisely controlled with swizzle fields encoded as RSWZ3.


 * op1_swz_c0, op1_swz_c1, op1_swz_c20, op1_swz_c2x, op1_swz_c30, op1_swz_c3x

Each channel can be masked with control bits. Masking is described in Swizzle_masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

masking does not apply for dot.f32 instruction

Swizzles - operand 2
Swizzles of operand 2 can not be precisely controlled and have predefined combinations described below and controlled by swizzle fields:


 * op2_swz, swz_alt_op2

Each channel can be masked with control bits. Masking is described in Swizzle_masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

masking does not apply for dot.f32 instruction

Instructions
mul.f16, add.f16, frc.f16, dsx.f16, dsy.f16, min.f16, max.f16, dot.f16

Encoding
Higher 4 bytes

Lower 4 bytes

Fields - operands

 * op1_swz_c3x - operand 1 swizzling channel 3 bit 1, 2. encoded as RSWZ3. consult Swizzles - operand 1.
 * alt_opt0 - consult Operand 0.
 * op1_swz_c30 - operand 1 swizzling channel 3 bit 0. encoded as RSWZ3. consult Swizzles - operand 1.
 * alt_opt1 - consult Operand N.
 * alt_opt2 - consult Operand N.
 * swz_alt_op2 - change op2 swizzle. consult Swizzles - operand 2.
 * op2_swz - op2 swizzle encoded with Register Swizzle RSWZ2. consult Swizzles - operand 2.
 * swz_mask3 - mask swizzle. consult Swizzle masking.
 * swz_mask2 - mask swizzle. consult Swizzle masking.
 * swz_mask1 - mask swizzle. consult Swizzle masking.
 * swz_en - enables usage of swizzling. consult Swizzle masking.
 * abs_op1 - add abs modifier to op1.
 * neg_op1 - negate op1.
 * abs_op2 - add abs modifier to op2.
 * op1_swz_c2x - operand 1 swizzling channel 2 bit 1, 2. encoded as RSWZ3. consult Swizzles - operand 1.
 * opt0 - consult Operand 0.
 * opt1 - consult Operand N.
 * opt2 - consult Operand N.
 * op0 - consult Operand 0.
 * op1_swz_c20 - operand 1 swizzling channel 2 bit 0. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1_swz_c1 - operand 1 swizzling channel 1. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1_swz_c0 - operand 1 swizzling channel 0. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1 - consult Operand N.
 * op2 - consult Operand N.

Constants
Specific operand may be used as float constant. This can be achieved with following groups of bits:


 * alt_opt0, opt0, op0
 * alt_opt1, opt1, op1
 * alt_opt2, opt2, op2

Float constants can only be used when swizzling is enabled for particular operand. Consider checking sections Swizzle_masking.

Constants are taken from tables Constants.

Constants correspond to table for 16 bit mode.

Swizzle masking
Masking is controled by control bits:
 * control bits: swz_en, swz_mask1, swz_mask2, swz_mask3

Each channel can be masked with control bits. Combinations of control bits produce the following masking table.

''dot.f16 instruction has explicit swizzling in operand 1 and operand 2 so masking does not apply to these operands. number of channels is always 4.''

Encoding used in masking table:

Masking table:

Swizzles - operand 0
Swizzles of operand 0 can not be controled and have predefined combinations described below:

Each channel can be masked with control bits. Masking is described in Swizzle_masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

Swizzles - operand 1
Each channel of operand 1 can be precisely controlled with swizzle fields encoded as RSWZ3.


 * op1_swz_c0, op1_swz_c1, op1_swz_c20, op1_swz_c2x, op1_swz_c30, op1_swz_c3x

Each channel can be masked with control bits. Masking is described in Swizzle_masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

masking does not apply for dot.f16 instruction

Swizzles - operand 2
Swizzles of operand 2 can not be precisely controlled and have predefined combinations described below and controlled by swizzle fields:


 * op2_swz, swz_alt_op2

Each channel can be masked with control bits. Masking is described in Swizzle_masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

masking does not apply for dot.f16 instruction

Instructions
dot.f32

Encoding
Higher 4 bytes

Lower 4 bytes

Fields - operands

 * c3_en - enable channel 3 for swizzles for op1 and op2. by default dot.f32 has only 3 channels for op1 and op2. consult Swizzles - operand 1.
 * alt_opt0 - consult Operand 0.
 * alt_opt1 - consult Operand N.
 * abs_op2 - add abs modifier to op2.
 * swz_en_strange1 - force overrides swizzle masking with single channel.
 * swz_en_strange0 - force overrides swizzle masking with single channel.
 * swz_mask3 - mask swizzle. consult Swizzle masking.
 * swz_mask2 - mask swizzle. consult Swizzle masking.
 * swz_mask1 - mask swizzle. consult Swizzle masking.
 * swz_en - enables usage of swizzling. consult Swizzle masking.
 * neg_op1 - negate op1.
 * abs_op1 - add abs modifier to op1.
 * opt0 - consult Operand 0.
 * opt1 - consult Operand N.
 * op2i - encoded with RI2
 * op0 - consult Operand 0.
 * swz_alt_op2 - change op2 swizzle. consult Swizzles - operand 2.
 * op2_swz - op2 swizzle encoded with Register Swizzle RSWZ2. consult Swizzles - operand 2.
 * op1_swz_c3 - operand 1 swizzling channel 3. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1_swz_c2 - operand 1 swizzling channel 2. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1_swz_c1 - operand 1 swizzling channel 1. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1_swz_c0 - operand 1 swizzling channel 0. encoded as RSWZ3. consult Swizzles - operand 1.
 * op1 - consult Operand N.

Constants
Specific operand may be used as float constant. This can be achieved with following groups of bits:


 * alt_opt0, opt0, op0
 * alt_opt1, opt1, op1

Float constants can only be used when swizzling is enabled for particular operand. Consider checking sections Swizzle_masking.

Constants are taken from tables Constants.

Constants correspond to table for 32 bit mode.

Swizzle masking
Masking is controled by control bits:
 * control bits: swz_en, swz_mask1, swz_mask2, swz_mask3

Each channel can be masked with control bits. Combinations of control bits produce the following masking table.

''dot.f32 instruction has explicit swizzling in operand 1 and operand 2 so masking does not apply to these operands. number of channels is 3 or 4 depending on c3_en.''

Encoding used in masking table:

Masking table operand 0:

Swizzles - operand 0
Swizzles of operand 0 can not be controled and have predefined combinations described below:

Each channel can be masked with control bits. Masking is described in Swizzle_masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

Swizzles - operand 1
Each channel of operand 1 can be precisely controlled with swizzle fields encoded as RSWZ3.


 * op1_swz_c0, op1_swz_c1, op1_swz_c2, op1_swz_c3

Channel 3 can be enabled with bit:


 * c3_en

Swizzles - operand 2
Swizzles of operand 2 can not be precisely controlled and have predefined combinations described below and controlled by swizzle fields:


 * op2_swz, swz_alt_op2

Channel 3 can be enabled with bit:


 * c3_en

Instructions
mad.f32

Encoding
Higher 4 bytes

Lower 4 bytes

Fields - operands

 * swz_alt_op3_2 - alter operand 3 swizzling. consult Swizzles - operand 3.
 * alt_opt0 - consult Operand 0.
 * alt_opt1 - consult Operand N.
 * abs_op2 - add abs modifier to op2.
 * op0_strange1 - sums with op0, adds 2. does not combine with op0_strange0.
 * op0_strange0 - sums with op0, adds 2. does not combine with op1_strange0.
 * swz_mask3 - mask swizzle. consult Swizzle masking.
 * swz_mask2 - mask swizzle. consult Swizzle masking.
 * swz_mask1 - mask swizzle. consult Swizzle masking.
 * swz_en - enables usage of swizzling. consult Swizzle masking.
 * neg_op1 - negate op1.
 * abs_op1 - add abs modifier to op1.
 * neg_op3 - negate op3.
 * abs_op3 - add abs modifier to op3.
 * swz_alt_op2_2 - alter operand 2 swizzling. consult Swizzles - operand 2.
 * opt0 - consult Operand 0.
 * opt1 - consult Operand N.
 * op2i - encoded with RI2.
 * op0 - consult Operand 0.
 * swz_alt_op2_x - alter operand 2 swizzling. consult Swizzles - operand 2.
 * op2_swz - operand 2 swizzle encoded with RSWZ2. consult Swizzles - operand 2.
 * swz_alt_op3_x - alter operand 3 swizzling. consult Swizzles - operand 3.
 * op3_swz - operand 3 swizzle encoded with RSWZ2. consult Swizzles - operand 3.
 * op3i - encoded with RI2.
 * swz_alt_op1 - alter operand 1 swizzling. consult Swizzles - operand 1.
 * op1_swz - operand 1 swizzle encoded with RSWZ2. consult Swizzles - operand 1.
 * op1 - consult Operand N.

Swizzle masking
Masking is controled by control bits:
 * control bits: swz_en, swz_mask1, swz_mask2, swz_mask3

Each channel can be masked with control bits. Combinations of control bits produce the following masking table.

Encoding used in masking table:

Masking table:

Swizzles - operand 0
Swizzles of operand 0 can not be controled and have predefined combinations described below:

Each channel can be masked with control bits. Masking is described in Swizzle masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

Swizzles - operand 1
Swizzles of operand 1 can not be precisely controlled and have predefined combinations described below and controlled by swizzle fields:


 * op1_swz, swz_alt_op1

Each channel can be masked with control bits. Masking is described in Swizzle masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

Swizzles - operand 2
Swizzles of operand 2 can not be precisely controlled and have predefined combinations described below and controlled by swizzle fields:


 * op2_swz, swz_alt_op2_x, swz_alt_op2_2

Each channel can be masked with control bits. Masking is described in Swizzle masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

Swizzles - operand 3
Swizzles of operand 3 can not be precisely controlled and have predefined combinations described below and controlled by swizzle fields:


 * op3_swz, swz_alt_op3_x, swz_alt_op3_2

Each channel can be masked with control bits. Masking is described in Swizzle masking.


 * swz_en, swz_mask1, swz_mask2, swz_mask3

Instructions
mad, dot, add, mul, subfl, exp, mov, log, rsq, rcp

Encoding
There is total of 10 instructions with 28 variations in this group. However instruction encoding is quite complex and is controlled by following fields:


 * op_sel2 - controls opcode page that is selected. there are 2 pages.
 * opcode2 - controls opcodes in the page. there are 8 opcodes.
 * gr_sel - controls validity of instructions
 * op_sel1 - controls explicit swizzling
 * opcode3 - controls validity of instructions

Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Allowed Instruction Encodings
Allowed encodings are defined by these tables.

opcode2, op_sel2, op_sel1 are a composite key for table Allowed Instructions

Join with this table to fill in  gaps.

Allowed Instructions
This table describes 10 instructions with 28 variations.

opcode2, op_sel2, op_sel1 is a composite key.

Encoding
Higher 4 bytes

Lower 4 bytes

Fields - operands Compatibility
One or multiple registers can change type between register R7 and internal register RI2.

Different combinations of operand types are controlled by.

Not all types of fields are applicable to each combination of registers.

Fields - operands

 * neg_op2i
 * swz_alt_op2i_2
 * swz_alt_opN
 * opN_swz
 * neg_opN
 * swz_alt_op1i
 * op1i_swz
 * swz_alt_op2i_x
 * op2i_swz
 * optN
 * op0i
 * op3i
 * op2i
 * op1i
 * swz_mask2
 * swz_mask1
 * swz_en
 * opN

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Instructions
mad.f32

Encoding
Higher 4 bytes

{|
 * - style="vertical-align:top;"


 * }

Lower 4 bytes

{|
 * - style="vertical-align:top;"


 * }
 * }

Fields - instruction
opcode 3:

Instructions
dot, mov, rsq, rcp, exp, log

Encoding
Higher 4 bytes

Lower 4 bytes

{|
 * - style="vertical-align:top;"




 * }

Fields - instruction
opcode2 (depends on op_sel):

0x28000000 - 0x30000000
Instructions: dot, mov, rsq, rcp

Encoding:

Notes:

Having bit 3 in byte 2 set to 0 produces invalid instruction

Fields:

opcode2:

Examples:

0x30000000 - 0x38000000
Instructions: rcp, rsq, log, exp

Encoding:

Notes:

should be omitted if  matches.

Fields:

Examples:

0x38000000 - 0x40000000
Instructions: mov, cmov, cmov8

Encoding:

Notes:

is only applicable to  and   since this is conditional move.

Fields:

Examples:

0x40000000 - 0x48000000
Instructions: pack, (mov)

Encoding:

Notes:

when  matches   it shall be omitted since it has no effect in terms of packing.

furthermore instruction mnemonic shall be replaced to

Fields:

Examples:

0x48000000 - 0x50000000
Instructions: this group only contains illegal instructions

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

0x50000000 - 0x58000000
Instructions: and.u32

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

Fields:

predicate:

Examples:

0x58000000 - 0x60000000
Instructions: xor.u32

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

Fields:

predicate:

Examples:

0x60000000 - 0x68000000
Instructions: shl.u32

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

Fields:

predicate:

Examples:

0x68000000 - 0x70000000
Instructions: shr.u32

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

Fields:

predicate:

Examples:

0x70000000 - 0x78000000
Instructions: rlp.u32

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

Fields:

predicate:

Examples:

0x78000000 - 0x80000000
Instructions: this group only contains illegal instructions

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

0x80000000 - 0x88000000
Instructions: add.fx8

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

Fields:

predicate:

Examples:

0x88000000 - 0x90000000
Instructions: add.fx8, sub.fx8

Encoding:

Notes:

Having bits 2, 3 in byte 2 set to 1 produces invalid instruction

Fields:

Examples:

0x90000000 - 0x98000000
Instructions: add.fx8, sub.fx8, min.fx8, max.fx8

Encoding:

Notes:

Having bit 0 in byte 2 set to 1 produces invalid instruction

Fields:

Examples:

0x98000000 - 0xA0000000
Instructions: mad.u8

Encoding:

Fields:

Examples:

0xA0000000 - 0xA8000000
Instructions: mad

Encoding:

Fields:

Examples:

0xA8000000 - 0xB0000000
Instructions: mad

Encoding:

Fields:

Examples:

0xB0000000 - 0xB8000000
Instructions: this group only contains illegal instructions

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

0xB8000000 - 0xC0000000
Instructions: this group only contains illegal instructions

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

0xC0000000 - 0xC8000000
Instructions: this group only contains illegal instructions

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

0xC8000000 - 0xD0000000
Instructions: mad.u8

Encoding:

Fields:

Examples:

0xD0000000 - 0xD8000000
Instructions: mad

Encoding:

Notes:

Having bit 5 in byte 1 set to 1 produces invalid instruction

Fields:

Examples:

0xD8000000 - 0xE0000000
Instructions: this group only contains illegal instructions

Encoding:

{|
 * - style="vertical-align:top;"


 * }
 * }

0xE0000000 - 0xE8000000
Instructions: tex

Encoding:

Fields:

Examples:

0xE8000000 - 0xF0000000
Instructions: lda32, ldl32, ldt32

Encoding:

Notes:

is only applicable when  modifier is specified

Fields:

Examples:

0xF0000000 - 0xF8000000
Instructions: sta32, stl32, stt32

Encoding:

Notes:

is only applicable when  modifier is specified

Fields:

Examples:

0xF8000000 - 0xFF000000
Notes:

this instruction group is much more complex than others so description is given in form of "glued" truth tables instead of independent truth tables.

predicate 000
Instructions:

Encoding:

Fields:

opcode2:

Examples:

predicate 001
Instructions:

Encoding

Notes:

predicate does not apply to all instructions

Fields

opcode2:

Examples:

predicate 010
Instructions:

Encoding:

Notes:

predicate does not apply to all instructions

Fields

opcode2

Examples:

predicate 011
Instructions:

Encoding

Notes:

predicate does not apply to all instructions

Fields

opcode2

Examples:

predicate 100
Instructions:

Encoding

Notes:

predicate does not apply to all instructions

Fields

opcode2:

Examples:

predicate 101
Instructions:

Encoding

Notes:

predicate does not apply to all instructions

Fields

opcode2:

Examples:

predicate 110
Instructions:

Encoding

Notes:

predicate does not apply to all instructions

Fields

opcode2:

Examples:

predicate 111
Instructions:

Encoding

Notes:

predicate does not apply to all instructions

Fields

opcode2:

Examples: