Gemm Tuner

Introduction

This is a set of tools for tuning the performance of OpenCL GEMM kernels. Specifically, we tune 3 GEMM kernels, each has a different implementation strategy of the GEMM operation: native, reshaped, reshaped only rhs. The details of these strategies can be found in the documentations of the corresponding kernels: CLGEMMMatrixMultiplyNativeKernel, CLGEMMMatrixMultiplyReshapedKernel and CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.

The Tuner consists of 2 scripts and 3 binaries:

• benchmark_gemm_examples.sh and GemmTuner.py under examples/gemm_tuner, and
• benchmark_cl_gemm_native, benchmark_cl_gemm_reshaped_rhs_only and benchmark_cl_gemm_reshaped under build/tests/gemm_tuner (you'll need to build the library first)

The inputs to the Tuner are a list of 4 valued tuples we call GEMM shape or GEMMParam (M, N, K, B, and possibly data type). They define the "shape" and other parameters (eg. data type) of a GEMM operation:

LHS x RHS = DST

Where LHS is of shape MxK, RHS is of shape KxN and DST is of shape MxN, and B is the batch size.

The outputs of the tuning process are 4 json files:

1. gemm_type_selection.json: selects which kernel type is the best for each GEMMParam
2. gemm_config_native.json: selects a list of best GEMMConfigs of the native kernel for each GEMMParam
3. gemm_config_reshapedonlyrhs.json: selects a list of best GEMMConfigs of the reshaped_only_rhs kernel for each GEMMParam
4. gemm_config_reshaped.json: selects a list of best GEMMConfigs of the reshaped kernel for each GEMMParam

These 4 files are the current representations we use for what we call the heuristics of a GEMM op: given a GEMMParam, what kernel and subsequently what configurations for that kernels are the most performant.

Step-by-step example

Step1: Prepare the shape and configs files

1. We first need to identify the shapes that we are interested in and store them in a csv file, say gemm_shapes.csv.

2. Then we need to specify a set of good GEMMConfig candidates for each kernel in 3 separate csv files (this requires some prior heuristics, but can be provided by the Compute Library developers upon requests, based on your target device).

   Say we have *gemm_configs_native.csv", "gemm_configs_reshaped.csv" and "gemm_configs_reshaped_only_rhs.csv".

   Please refer to the Prerequisite section for more details

Step2: Push relevant files to the target device

All the files that need to be present on the target device are:

• benchmark script: <ComputeLibrary>/examples/gemm_tuner/benchmark_gemm_examples.sh
• shapes and configs csv files: gemm_shapes.csv, gemm_configs_native.csv, gemm_configs_reshaped_only_rhs.csv, gemm_configs_reshaped.csv
• Example benchmark binaries: <ComputeLibrary>/build/tests/gemm_tuner/benchmark_cl_gemm*

Step3: Collect benchmark data

With these files on device, we can collect benchmark data using the script. Assume all the example binaries are pushed to a folder called gemm_tuner. While logged onto our device:

# Native
./benchmark_gemm_examples.sh -s native -e ./gemm_tuner -g ./gemm_shapes.csv -c ./gemm_configs_native.csv -o results/native
# Reshaped Only RHS
./benchmark_gemm_examples.sh -s reshaped_rhs_only -e ./gemm_tuner -g ./gemm_shapes.csv -c ./gemm_configs_reshaped_only_rhs.csv -o results/reshaped_only_rhs
# Reshaped
./benchmark_gemm_examples.sh -s reshaped -e ./gemm_tuner -g ./gemm_shapes.csv -c ./gemm_configs_reshaped.csv -o results/reshaped

You can repeat the 3 commands above to have a bit redundancy in your benchmark data (as you can imagine, measurement is noisy), but you may need to change the output folder for each repeat

Step4: Generate the heuristics

1. After benchmarking, we pull the benchmark data, the results folder, from the target device to our host machine
2. We use the GemmTuner.py script to give us the heuristics

   python3 <ComputeLibrary>/examples/gemm_tuner/GemmTuner.py -b ./results -o heuristics
   

   When it's finished, there should be 4 json files in the heuristics folder

One thing to notice is that the config heuristics might give more than 1 recommendations for each GEMMParam, because we accept all good GEMMConfigs with a tolerance. If you want fewer recommendations, you can decrease the tolerance by passing a lower value to -t <tolerance> to the GemmTuner.py script.

Prerequisite

• A target device to be tuned, plus the following on the device:

  • Android or Linux OS

  • Bash shell

  • Built Compute Library with benchmark examples binaries

  • benchmark_gemm_examples.sh script

  • gemm shape file

    A csv file containing the GEMMParam search list. This is the list of GEMMParams/gemm shapes that we're interested in (For more details see Approach section). The default list is prepared by Compute Library developers in advance and can be provided on request.

    The format is described as:

    A headerless csv file with fields separated by commas.

    A gemm shape is a list of 4 positive integers <M, N, K, B> describing the shapes of the two matrices (LHS and RHS) with:

    M - Number of lhs matrix rows
    N - Number of rhs matrix columns
    K - Number of lhs matrix columns/rhs matrix rows
    B - Batch size

    An example gemm shape file looks like:

  100,100,30,1
  100,100,30,3
  ...
  

  • gemm config file
    A csv file containing the GEMMConfig search list. This is the list of candidate GEMMConfigs among which we search for the optimal one. Note that we have a different list for each strategy. The default lists are prepared by Compute Library developers in advance and can be provided on request.

    The format of the file for each strategy is the same:

    A headerless csv file with fields separated by commas.

    However the fields of GEMMConfig differ for each strategy:

    • Strategy native: A gemm config is a list of 3 positive integers <m0, n0, k0>, with:

      m0 - Number of rows processed by the matrix multiplication
      n0 - Number of columns processed by the matrix multiplication
      k0 - Number of partial accumulations performed by the matrix multiplication

      Only the following configurations of M0, N0 and K0 are currently supported:

      M0 = 1, 2, 3, 4, 5, 6, 7, 8
      N0 = 2, 3, 4, 8, 16
      K0 = 2, 3, 4, 8, 16

      An example gemm config file looks like:

  1,4,4
  2,3,8
  ...
  

  * Strategy **reshaped_rhs_only**:
    A gemm config is a list of 4 positive integers <m0, n0, k0, h0> and 3 boolean values:
  
    m0 - Number of rows processed by the matrix multiplication  
    n0 - Number of columns processed by the matrix multiplication  
    k0 - Number of partial accumulations performed by the matrix multiplication  
    h0 - Number of horizontal blocks of size (k0xn0) stored on the same output row  
    interleave_rhs - Interleave rhs matrix (1) / Do not interleave rhs matrix (0)  
    transpose_rhs - Transpose rhs matrix (1) / Do not transpose rhs matrix (0)  
    export_to_cl_image_rhs - Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0). Can only be true
                            with certain combinations of the GEMMParams and other configs. Please refer to CLGEMMReshapeRHSMatrixKernel
                            for more details
  
    Only the following configurations of M0, N0 and K0 are currently supported:
  
    M0 = 1, 2, 3, 4, 5, 6, 7, 8  
    N0 = 2, 3, 4, 8, 16  
    K0 = 2, 3, 4, 8, 16  
    H0 >= 1  
  
    An example gemm config file looks like:
  

  4,4,4,1,1,1,0
  4,4,4,3,1,0,1
  ...
  

  * Strategy **reshaped**:
    A gemm config is a list of 5 positive integers <m0, n0, k0, v0, h0> and 4 boolean values:
  
    m0 - Number of rows processed by the matrix multiplication  
    n0 - Number of columns processed by the matrix multiplication  
    k0 - Number of partial accumulations performed by the matrix multiplication  
    v0 - Number of vertical blocks of size (m0xk0) stored on the same output row  
    h0 - Number of horizontal blocks of size (k0xn0) stored on the same output row  
    interleave_lhs - Interleave lhs matrix (1) / Do not interleave lhs matrix (0)  
    interleave_rhs - Interleave rhs matrix (1) / Do not interleave rhs matrix (0)  
    transpose_rhs - Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do transpose lhs matrix (0)  
    export_to_cl_image_rhs - Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0). Can only be true
                            with certain combinations of the GEMMParams and other configs. Please refer to CLGEMMReshapeRHSMatrixKernel
                            for more details
  
    If rhs matrix is transposed only the following configurations are currently supported:
  
    M0 = 2, 3, 4, 5, 6, 7, 8  
    N0 = 2, 3, 4, 8, 16  
    K0 = 2, 3, 4, 8, 16  
    V0 >= 1  
    H0 >= 1  
  
    If lhs matrix is transposed only the following configurations are currently supported:
  
    M0 = 2, 3, 4, 8  
    N0 = 2, 3, 4, 8, 16  
    K0 = 2, 3, 4, 8, 16  
    V0 >= 1  
    H0 >= 1  
  
    An example gemm config file looks like:
  

  4,4,4,1,3,1,1,1,0
  4,4,4,3,3,1,1,0,1
  ...
  

• A host machine, plus these on the machine:
  • python >= 3.6
  • GemmTuner.py script

Usage

The usage of the 2 scripts:

1. benchmark_gemm_examples.sh

   Run the shell script (benchmark_gemm_examples.sh) on your target device. Note that all the built benchmark examples: build/tests/gemm_tuner/benchmark_cl_gemm*, have to be present on your target device prior to running. The benchmark results will be saved to json files in an output directory.

   Usage: benchmark_gemm_examples.sh [-h] -s \<strategy\> -e \<example_binary_dir\> -g \<gemm_shape_file\>
   -c \<gemm_config_file\> [-d \<data_type\>] [-o \<out_dir\>]
   
   Options:
           -h
           Print help messages. If a strategy is specified with -s <strategy>, then only display messages relevant to that
           strategy. Otherwise if no strategy is specified, display messages for all available strategies.
   
           -s <strategy>
           Strategy option.
           Options: ${ALL_STRATEGY_OPTIONS[@]}.
   
           -e <example_binary_dir>
           Path to directory that holds all example binaries
   
           -g <gemm_shape_file>
           Path to gemm shape csv file
   
           -c <gemm_config_file>
           Path to gemm config csv file
   
           -d <data_type>
           Data type option with which to run benchmark examples
           Default: ${DEFAULT_DATA_TYPE}
           Supported options:
           Strategy            :    Data Types
           Native              :    F32
           Reshaped            :    F16, F32
           Reshaped RHS Only   :    F16, F32
   
           -o <out_dir>
           Path to output directory that holds output json files
           Default: ${DEFAULT_OUT_DIR}
   

2. GemmTuner.py:

Run the python script (GemmTuner.py) on your host machine. You'll need to transfer all the benchmark result json files generated from the previous step to your host machine beforehand. The script will output the best kernel and gemm configurations for each gemm param in the 4 output json files

Usage: GemmTuner.py [-h] -b PATH [-o PATH] [-t TOLERANCE] [-D]

CL GEMM Tuner
optional arguments:
  -h, --help            show this help message and exit
  -b PATH, --benchmark_results PATH
                        Path to benchmark result directory, where benchmark
                        result json files have a file extension of
                        'gemmtuner_benchmark'
  -o PATH, --output_dir PATH
                        Path to directory that holds output json files.
  -t TOLERANCE, --tolerance TOLERANCE
                        For testing if two GEMMConfigs are equivalent in terms
                        of performance. The tolerance is OpenCL timer in
                        milliseconds. Recommended value: <= 0.1 ms
  -D, --debug           Enable script debugging output