GPU/Shader Instruction Set

Overview

A compiled shader binary is comprised of two parts : the main instruction sequence and the operand descriptor table. These are both sent to the GPU around the same time but using separate GPU Commands. Instructions (such as format 1 instruction) may reference operand descriptors. When such is the case, the operand descriptor ID is the offset, in words, of the descriptor within the table. Both instructions and descriptors are coded in little endian. Basic implementations of the following specification can be found at [1] and [2] Please note that this page is being written as the instruction set is reverse engineered; as such it may very well contain mistakes.

Instruction formats

Format 1 : (used for register instructions)

Offset	Size (bits)	Description
0x0	0x7	Operand descriptor ID (DESC)
0x7	0x5	Source 2 register (SRC2)
0xC	0x7	Source 1 register (SRC1)
0x13	0x2	Index (IDX)
0x15	0x5	Destination register (DST)
0x1A	0x6	Opcode

Format 2 : (used for flow control instructions)

Offset	Size (bits)	Description
0x0	0x8	Number of instructions (NUM)
0xA	0xC	Destination offset (in words) (DST)
0x1A	0x6	Opcode

Format 3 : (used for conditional flow control instructions)

Offset	Size (bits)	Description
0x0	0x8	Number of instructions ? (NUM)
0xA	0xC	Destination offset (in words) (DST)
0x16	0x4	Uniform boolean ID (BOOL)
0x1A	0x6	Opcode

Format 4 : (used for SETEMIT)

Offset	Size (bits)	Description
0x16	0x2	Primitive ID (PRIMID)
0x18	0x2	Vertex ID (VTXID)
0x1A	0x6	Opcode

Instructions

Opcode	Format	Name	Description
0x00	1	ADD	Adds two vectors component by component; DST[i] = SRC1[i]+SRC2[i] for all i (modulo destination component masking)
0x01	1	DP3	Computes dot product on 3-component vectors; DST = SRC1.SRC2
0x02	1	DP4	Computes dot product on 4-component vectors; DST = SRC1.SRC2
0x08	1	MUL	Multiplies two vectors component by component; DST[i] = SRC1[i].SRC2[i] for all i (modulo destination component masking)
0x0C	1	MAX	Takes the max of two vectors, component by component; DST[i] = MAX(SRC1[i], SRC2[i]) for all i (modulo destination component masking)
0x0D	1	MIN	Takes the min of two vectors, component by component; DST[i] = MIN(SRC1[i], SRC2[i]) for all i (modulo destination component masking)
0x0E	1	RCP	Computes the reciprocal of the vector, component by component; DST[i] = 1/SRC1[i] for all i (modulo destination component masking)
0x0F	1	RSQ	Computes the reciprocal of the square root of the vector, component by component; DST[i] = 1/sqrt(SRC1[i]) for all i (modulo destination component masking)
0x12	1	SETINDEX	Sets (idx1, idx2, _, _) to SRC1 (cast to integer).
0x13	1	MOV	Moves value from one register to another; DST = SRC1.
0x21	1	END2	?
0x22	1	END1	?
0x23	1	BREAK?	Conditionally breaks out of LOOP block ?
0x24	2	CALL	Jumps to DST and executes instructions until it reaches DST+NUM instructions
0x26	3	CALLC	Jumps to DST and executes instructions until it reaches DST+NUM instructions if BOOL is true
0x27	3	IFU	If condition BOOL is true, then executes instructions until DST, then jumps to DST+NUM; else, jumps to DST.
0x28	2	IF?	If condition (don't know how condition flags work yet) is true, then executes instructions until DST, then jumps to DST+NUM; else, jumps to DST
0x29	3	LOOP?	Loops over the code between itself and DST an unknown number of times. Possibly set by an integer uniform.
0x2A	0 (no param)	EMIT	(geometry shader only) Emits a vertex (and primitive if PRIMID is non-zero). SETEMIT must be called before this.
0x2B	4	SETEMIT	(geometry shader only) Sets VTXID and PRIMID for the next EMIT instruction. VTXID is the ID of the vertex about to be emitted within the primitive, while PRIMID is zero if we are just emitting a single vertex and non-zero if are emitting a vertex and primitive simultaneously. Note that the output vertex buffer (which holds 4 vertices) is not cleared when the primitive is emitted, meaning that vertices from the previous primitive can be reused for the current one. (this is still a working hypothesis and unconfirmed)
0x2C	2	JMPC?	Jumps to DST if condition (don't know how condition flags work yet) is true, else does nothing.
0x2E	1	CMP1	Presumably compares two vectors component by component and sets the appropriate flags. (unknown exactly how this works as of yet)
0x2F	1	CMP2	Presumably compares two vectors component by component and sets the appropriate flags. (unknown exactly how this works as of yet)

Operand descriptors

Sizes below are in bits, not bytes.

Offset	Size	Description
0x0	0x4	Destination component mask. Bit 3 = x, 2 = y, 1 = z, 0 = w.
0x5	0x8	Source 1 component selector
0xE	0x8	Source 2 component selector
0x1F	0x1	Flag

Component selector :

Offset	Size	Description
0x0	0x2	Component 3 value
0x2	0x2	Component 2 value
0x4	0x2	Component 1 value
0x6	0x2	Component 0 value

Value	Component
0x0	x
0x1	y
0x2	z
0x3	w

The component selector enables swizzling. For example, component selector 0x1B is equivalent to .xyzw, while 0x55 is equivalent to .yyyy.

Relative addressing

There are 3 global index registers : idx1, idx2 and idx3 = lcnt (loop counter). For format 1 instructions, when IDX != 0, the value of the corresponding index is added to SRC1's value.

For example, if IDX = 2, idx2 = 3 and SRC1 = c8, then instead SRC1+idx2 = c11 will be used for the instruction.

idx1 and idx2 can be set manually through the SETINDEX instruction. lcnt is set automatically by the LOOP instruction. Note that lcnt is still accessible and valid after exiting a LOOP block.

Registers

It is not yet fully understood how registers are organized. It does however seem that registers are separated into various banks, some RO, some WO and some RW. Because of this separation, a given register ID may not refer to the same register value when it is used as SRC or as DST.

Attribute (input, RO) registers are located within the 0x0-0x10 range. What data they are fed is specified by the CPU. Output (WO) registers are also located within the 0x0-0x10 range. What data they are contain is specified by the CPU. Registers within the 0x20-0x40 ranges seem to be RW. They contain uniforms, such as matrix data.

SRC2 being only 5 bits long rather than 7 bits like its friend SRC1, it can only access v (input attribute) and r (temporary) registers.

Registers in the 0x88-0x97 range are uniform booleans.

It appears that writing twice to the same output register can cause problems, such as the GPU hanging.

DST mapping :

DST raw value	Register name	Description
0x0-0x7	o0-o7	Output registers.
0x10-0x1F	r0-r15	Temporary registers.

SRC1 mapping :

SRC1 raw value	Register name	Description
0x0-0x7	v0-v7	Input attribute registers.
0x10-0x1F	r0-r15	Temporary registers.
0x20-0x7F	c0-c95	Vector uniform registers.

SRC2 mapping :

SRC2 raw value	Register name	Description
0x0-0x7	v0-v7	Input attribute registers.
0x10-0x1F	r0-r15	Temporary registers.