Skip to content

Latest commit

 

History

History
103 lines (77 loc) · 4.91 KB

README.md

File metadata and controls

103 lines (77 loc) · 4.91 KB
Raw

GEMM Parameter Tuner

The GEMM Parameter Tuner is a standalone tool that can calculate the optimal parameters for a GEMM operation through experimentation.

The Tuner is provided M, N and K values, iterates through a number of potential configurations and then prints a list of them and their performance.

Building

  1. Clone the portBLAS repository, making sure to pass the --recursive option, in order to clone submodule(s).
  2. Create a build directory as tools/auto_tuner/build.
  3. Run CMake and Ninja from the build directory:
$ cmake -GNinja ../ [-DTUNING_TARGET=supported backend]
$ ninja

See the Setup section in this repository's main readme for more details.

Make options

CMake options are given using -D immediately followed by the option name, the symbol = and a value (ON and OFF can be used for boolean options and are equivalent to 1 and 0). Example: -DBLAS_ENABLE_TESTING=OFF

name value description
BLAS_MEMPOOL_BENCHMARK ON/OFF Enable the scratchpad memory pool, useful just in case of tall skinny matrices. OFF by default

Usage

Upon a successful build, five binaries will be created which will print the optimal (highest gflops) tile sizes for a specific transposition of A and B:

Binary Matrix A Matrix B
tune_nn Normal Normal
tune_tn Transposed Normal
tune_nt Normal Transposed
tune_tt Transposed Transposed

The tune_all binary runs through each combination in turn, printing out separate results for each of them.

All these binaries are invoked as follows:

$ tune M K N bs rep [batch_type]

Where the provided options mean the following:

Option Meaning
M, N, K Values for these parameters in the GEMM algorithm
bs The number of batches to use for batched GEMM. Set to 1 to use regular GEMM
rep The number of times to run GEMM for each combination. The mean average is taken off all executions
batch_type The type of batching to be used. It can be interleaved or strided. The default is strided.

This will execute GEMM on a number of different combinations depending on the current platform, and display the results of each in order from worst to best performance.

Configuration

A number of configurations are available in gen/ which describe all the combinations of tunable parameters which will be used. These are automatically used by the build system and integrated into the binary at compile time. As an end user, you will not need to modify any of these files.

However, if you want to add or modify these values, the various .json files can be used. If a new target is to be added, it should be included in the CMakeLists.txt file as appropriate.

The root of the json file is an object containing three arrays. Each array ( local, non_local, naive) is for a different GEMM algorithm, and contain a number of Configuration Generators.

A configuration Generator is an object, with all of its values being arrays. A generator adds all combinations (as a cartesian product) of its array elements to the list of Gemm parameters to try.

The parameters are listed as follows:

Parameter Algorithm Description
cache_line_size All The size of the cache line
work_item_tiles Not Naive The [rows, cols] processed by each work item
work_group_tiles Not Naive The number of item-level tiles within each [row, col] of block-level tile
block_level_tiles Local Only The number of block-level tiles within each [row, col] of top-level tile
batch_level_tiles non local Interleave The number of batch work_item and batch work group . default to 1
vectorization_size Not Naive the size of vector type
double_buffer Local Only Enable the use of double buffering
no_bank_conflict_a Local Only Avoids bank conflicts when accessing blocks of matrix A in local memory
no_bank_conflict_b Local Only Avoids bank conflicts when accessing blocks of matrix B in local memory