Skip to content

Latest commit

 

History

History
105 lines (72 loc) · 4.19 KB

kernel_isa_gpu.md

File metadata and controls

105 lines (72 loc) · 4.19 KB
Raw

Kernel ISA

This document describes how to use the Intercept Layer for OpenCL Applications to retrieve kernel ISA for Intel GPU devices.

Tools

You will need:

  • The Intercept Layer for OpenCL Applications
  • A Kernel ISA Binary Disassembler

Process

Retrieving kernel ISA is a two step process: The first step is to dump kernel ISA in binary form. Then, the kernel ISA binaries can be disassembled to view the kernel ISA in text form.

Dumping Kernel ISA Binaries

Dumping kernel ISA binaries is the easy part! Simply install the Intercept Layer for OpenCL Applications, set DumpKernelISABinaries, and execute your application. Then, every time our application calls clBuildProgram, the Intercept Layer for OpenCL Applications will dump an ISA binary for each kernel in the program.

How does this work? Drivers for Intel GPU OpenCL devices support a kernel query for the kernel ISA binary.

NOTE: The control to dump kernel ISA binaries to disassemble for Intel GPU devices is DumpKernelISABinaries, which is different than the control to dump program binaries to disassemble for Intel CPU devices!

Building an Intel GPU ISA Disassembler

This section describes how to build an Intel GPU ISA disassembler, which can be used to disassemble a kernel ISA binary to view the kernel ISA in text form. The particular Intel GPU ISA disassembler we'll be using is IGA, which is part of the Intel Graphics Compiler (IGC).

The first step is to get the Intel Graphics Compiler source code. From a directory where you want to build IGA:

git clone https://github.com/intel/intel-graphics-compiler.git

We're only going to build IGA, and not all of IGC. So, change to the IGA directory in the repo we just cloned:

cd intel-graphics-compiler/visa/iga

IGA creates its build files using CMake, just like the Intercept Layer for OpenCL Applications. So, create a build directory for build files, and run cmake from this build directory:

mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release

This should create build files without errors or warnings. After build files have been created, either use them directly to build IGA, or invoke CMake to perform the build. For example:

cmake --build . --config Release

If all goes well, this will successfully build an IGA executable. Note that the exact location of the executable will be dependent on the operating system you are building on, and the configuration you are building.

Using the Intel GPU ISA Disassembler

Disassembling a kernel ISA binary using IGA is a mostly straightforward process. You'll want to tell IGA to disassemble your file (-d), the device your kernel ISA binary was compiled for (-p), and your kernel ISA binary file name. For example:

./iga32 -d -p 8 CLI_0000_3F40E1CD_0000_GPU_GenerateJuliaSet.isabin

The table below describes the mapping between Intel processor code names and GPU devices:

Processor Code Name Device
Broadwell BDW GEN8
Cherry Trail CHV GEN8LP
Skylake SKL GEN9
Apollo Lake BXT GEN9LP
Kaby Lake KBL GEN9.5
Ice Lake ICL GEN11
Tiger Lake TGL GEN12

Here is a helpful link that describes the Intel GPU ISA:

https://software.intel.com/en-us/articles/introduction-to-gen-assembly

Disassembling Many Kernel ISA Binaries

The Python script ../scripts/disassemble_all_gpu.py may be useful to invoke a disassembler on all ISA binaries in a specified directory.

* Other names and brands may be claimed as the property of others.

Copyright (c) 2018-2023, Intel(R) Corporation