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 ACL 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 and commas only (there cannot be whitespaces around each field). Note also comments and extraneous empty lines are not permitted.
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 2 boolean values interleave_rhs and transpose_rhs, 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 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) 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 4,4,4,3,1,0 ...
* Strategy **reshaped**: A gemm config is a list of 5 positive integers \<m0, n0, k0, v0, h0\> and 3 boolean values interleave_lhs, interleave_rhs and transpose_rhs, 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 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) * 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 4,4,4,3,3,1,1,0 ...
The tuning stage consists of 2 steps:
Run benchmarks:
Run the shell script (benchmark_gemm_examples.sh) on your target device. Note that all the built benchmark examples 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\> [-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: native reshaped_rhs_only reshaped. -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 -o \<out_dir\> Path to output directory that holds output json files Default: out
Run analyser:
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 configuration, along with some analysis statistics for each strategy, and optionally save the parsed benchmark results into csv files (one for each strategy) for further analysis.
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 csv files. One per strategy -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
This section gives a brief description and rationale of the approach adopted by the current version of GEMM Tuner.
As explained in the Introduction section, the outputs of the tuner are 1 optimal GEMMConfig for each strategy. This is because we can only integrate 1 GEMMConfig for each strategy in ACL at compile time. In theory, however, the optimal GEMMConfig also depends on different parameters of GEMM (called GEMM Parameter or GEMMParam, e.g.: the shape of the operation); thus ideally, for each strategy, the optimal configurations should be a mapping from GEMMParam to GEMMConfig instead of a single GEMMConfig.
To address this issue, we ensure the one single optimal GEMMConfig can generalise well to all potential GEMMParams (or at least the ones that we care about). The approach we adopt involves a preliminary stage where a collection of common GEMMParams (GEMM shapes from popular networks) are compiled. Then, to reduce the final tuning time, rather contradictorily, we spend a lot of time searching for near-optimal GEMMConfigs for each GEMMParam first, and then discard redundant GEMMParams which share similar optimal GEMMConfigs with others. The resultant list of GEMMParams is called a GEMMParam search list, as in these GEMMParams are typical enough to capture the space of GEMMParams that we care about.
During this preliminary stage we also produce a list of good GEMMConfigs that can be used to search for the optimal one in the actual tuning stage. This, again, is to reduce the tuning time, and the resultant list is called a GEMMConfig search list.
The GEMMParam search list and the GEMMConfig search list are investigated and prepared by the developers; the users of GEMM tuner need not worry about producing them, but they need to obtain them prior to running the tuner.
Once these two lists (2 for each strategy, so 6 in total) are obtained, they can be fed to the tuner, to produce the optimal GEMMConfig(s).