Overview

Like most D3D11-era GPUs the Tegra X1 has 5 pipeline shader stages: Vertex, tessellation control, tessellation evaluation, geometry and fragment. Maxwell GPUs have two vertex shader stages: A and B, using both is not mandatory or common. There are also compute shaders or kernels which run on a separate channel.

Header

Pipeline stage shaders unlike compute kernels have a header documented by Nvidia[1]. It is a 0x50 byte structure that describes shader's behavior and enabled features that don't belong to the code itself (e.g. local memory size, whether it writes global memory). Some of these entries are not mandatory and are not emitted by Nvidia's compiler.

It is located at the beginning of the program address. Program code starts right after it (at address 0x50, address 0 on compute kernels).

Text

Every 4 instructions starting from the first one there's a scheduler instruction. All instructions are 64 bits long, including scheduler instructions.

Shaders emitted by Nvidia's compiler include a branching instruction pointing to itself after the final unconditional exit. The intention of this "sign" is unknown.

Registers

Maxwell GPUs have 255 type-less general purpose registers and one special register with id 255, nvdisasm shows it as RZ and envydis as 0x0. Writing here is a no-op unless there are side effects. Reading from RZ returns zero. The fewer registers a shader uses, the more it can be parallelized.

General purpose registers or GPRs are 32 bits long and are the same for all operations, these are given meaning on the instructions. Half float instructions are SIMD and operate on 16 bit pairs, meanwhile double instructions take two registers to operate. uint64 values are emulated using uint32 instructions extending their domain through condition codes; when an instruction has to read an uint64 value it reads two subsequent registers.

It is a common technique to read or write subsequent registers with a single instruction. For example TEXS (used to sample a texture) reads the texture coordinates from Ra and Ra+1, although it gets more complex with other layouts like 3D textures (it's not Ra+2). The result of the sample is given in an Rd and their subsequent registers.

Predicates

There are 6 general purpose "1-bit" predicates. These can be used to conditionally execute a given instructions. There is an extra predicate that always evaluates as true (and false when negated), writing here is a no-op unless there are side-effects. It is shown as "PT" in nvdisasm.

Most of the time predicates can be negated.

Condition codes

These mostly match x86 condition codes. They are explicitly set by instructions with a ".CC" flag or by special instructions like R2P.

These can be analyzed with the P2R instruction (which moves predicates or condition codes to a register).

Analysis Tools

Disassembling

To disassemble a Maxwell shader or compute kernel there are two applications:

nvdisasm

Nvidia's disassembler. Since it's made by Nvidia it usually knows more instruction variants. It tends to print the intention of the instructions rather than their internal encoding, helping visual inspection. It can show the binary representation with an option.

One of its major problems is that the input file has to be perfectly aligned with a stride of 0x20 bytes or it will otherwise fail. Appending zeroes to the file will make it happy since those are NOP instructions. Invalid instructions will also abort the disassemble.

To disassemble a Maxwell shader or kernel in a file this command can be used:

$ nvdisasm -b SM53 <file>

For more information about this program it's recommended to read its help entry and CUDA Toolkit Documentation[4].

envydis

envydis is nouveau's assembler and disassembler. Unlike Nvidia's disassembler this one doesn't care about alignment or invalid instructions, so it is useful to analyze a raw memory dump while searching for a shader or kernel. Its declarative-style C file[5] describes instructions internal encoding.

To disassemble with it:

$ envydis -m gm107 -i <file>

envydis also has an assembler variant: envyas. To edit existing shaders or write shaders from scratch this is easier than assembling the bits manually.

Since it's the product of reverse engineering the GPU and nvdisasm itself, it might contain errors.