TensileCreateLibrary

TensileCreateLibrary#

TensileCreateLibrary is a command-line tool that generates libraries and code object files for a set of supplied logic files.

Syntax#

TensileCreateLibrary [OPTIONS...] <LOGIC DIRECTORY> <OUTPUT DIRECTORY> <RUNTIME LANGUAGE>

Required arguments#

Here is the list of required arguments for invoking the TensileCreateLibrary command.

Table 2 Required arguments#

Argument

Description

<LOGIC DIRECTORY>

Absolute path for logic files. The logic files are generally found in either of the following locations:
- The build directory under 3_LibraryLogic, when generated by the Tensile program.

- In a project that hosts pregenerated logic files, such as rocBLAS.

<OUTPUT DIRECTORY>

Absolute or relative path to the output directory where build artifacts are placed.

<RUNTIME LANGUAGE>

Runtime language out of OCL, HIP, or HSA

Options#

Here is the list of optional arguments for invoking the TensileCreateLibrary command.

Table 3 Options#

Option

Description

--architecture=ARCHITECTURE

Architectures to generate a library for. When specifying multiple options, use quoted and semicolon-delimited architectures such as --architecture=’gfx908;gfx1012’. Supported architectures include: all; gfx000; gfx803; gfx900; gfx900:xnack-; gfx906; gfx906:xnack+; gfx906:xnack-; gfx908; gfx908:xnack+; gfx908:xnack-; gfx90a; gfx90a:xnack+; gfx90a:xnack-; gfx940; gfx940:xnack+; gfx940:xnack-; gfx941; gfx941:xnack+; gfx941:xnack-; gfx942; gfx942:xnack+; gfx942:xnack-; gfx1010; gfx1011; gfx1012; gfx1030; gfx1031; gfx1032; gfx1034; gfx1035; gfx1100; gfx1101; gfx1102.

--build-client

Builds Tensile client executable that is used for stand alone benchmarking. This option is set by default.

--client-config

Creates best-solution.ini in the output directory for the library and generated code object files. This option is set by default.

--code-object-version={default,V4,V5}

HSA code object version.

--cxx-compiler={amdclang++, hipcc} or {clang++, hipcc} for Windows

C++ compiler used when generating binaries.

--embed-library=EMBEDLIBRARY

Specifies the library to embed into static variables.

--embed-library-key=EMBEDLIBRARYKEY

Access key for embedding library files.

--generate-manifest-and-exit

Similar to the dry-run option for make, this option computes the outputs of TensileCreateLibrary and writes the expected outputs to a manifest file but doesn’t exectue the commands to generate the output.

--generate-sources-and-exit

Skips building the source and assembly code object files. Outputs source files only and exits.

--ignore-asm-cap-cache

Ignores the asm capability cache and derives the asm capabilities at runtime.

--jobs=CPUTHREADS or -j CPUTHREADS

Number of parallel jobs to launch. If this option is supplied with a value higher than nproc, the number of parallel jobs will be the same as the number of cores. If this option is supplied with a value below 1 (0 or -1), the number of parallel jobs will be the same as the number of cores, up to a maximum of 64. The default value is -1.

--lazy-library-loading

Loads Tensile libraries only when needed.

--library-format={yaml,msgpack}

Specifies the library format to use. Default value: msgpack.

--no-enumerate

Prohibits rocm_agent_enumerator from running.

--no-merge-files

Stores every solution and kernel in separate files.

--no-short-file-names

Prohibits short files names.

--num-merged-files=NUMMERGEDFILES

Number of files the kernels must be written into.

--merge-files

Stores all solutions in a single file. This is set by default.

--short-file-names

Converts solution and kernel names to serial Ids if Windows kernel name is too long. The option is set by default.

--separate-architectures

Separates TensileLibrary file according to architecture to reduce the library file loading time. This option writes each architecture into a different TensileLibrary_gfxXXX.dat file.

--verbose=PRINTLEVEL or -v PRINTLEVEL

Sets printout verbosity level out of 0, 1, and 2.

--version=VERSION

Version string to embed into the library file.

--write-master-solution-index

Outputs master solution index including number of kernels per architecture post build in csv format.

Usage without options#

Here is how to run TensileCreateLibrary command without options. The specified logic directory contains the logic files and the generated artifacts are directed to the tensile-output directory that is created in the directory where the TensileCreateLibrary command is invoked. The runtime language is HIP.

TensileCreateLibrary /home/myuser/Logic tensile-output HIP

Usage with options#

Here is how to run TensileCreateLibrary command with options. The following example shows options with and without arguments.

TensileCreateLibrary --separate-architectures --jobs=32 /home/myuser/Logic tensile-output HIP