1# Gemm Tuner 2 3## Introduction 4 5This is a set of tools for tuning the performance of OpenCL GEMM kernels. Specifically, we tune 3 GEMM kernels, each 6has a different implementation **strategy** of the GEMM operation: **native**, **reshaped**, **reshaped only rhs**. 7The details of these strategies can be found in the documentations of the corresponding kernels: 8**CLGEMMMatrixMultiplyNativeKernel**, **CLGEMMMatrixMultiplyReshapedKernel** and 9**CLGEMMMatrixMultiplyReshapedOnlyRHSKernel**. 10 11The Tuner consists of 2 scripts and 3 binaries: 12* benchmark_gemm_examples.sh and GemmTuner.py under examples/gemm_tuner, and 13* benchmark_cl_gemm_native, benchmark_cl_gemm_reshaped_rhs_only and benchmark_cl_gemm_reshaped under 14 build/tests/gemm_tuner (you'll need to build the library first) 15 16The inputs to the Tuner are a list of 4 valued tuples we call **GEMM shape** or **GEMMParam** (M, N, K, B, and possibly 17data type). They define the "shape" and other parameters (eg. data type) of a GEMM operation: 18``` 19LHS x RHS = DST 20``` 21Where LHS is of shape MxK, RHS is of shape KxN and DST is of shape MxN, and B is the batch size. 22 23The outputs of the tuning process are 4 json files: 241. gemm_type_selection.json: selects which kernel type is the best for each GEMMParam 252. gemm_config_native.json: selects a list of best **GEMMConfigs** of the native kernel for each GEMMParam 263. gemm_config_reshapedonlyrhs.json: selects a list of best GEMMConfigs of the reshaped_only_rhs kernel for each GEMMParam 274. gemm_config_reshaped.json: selects a list of best GEMMConfigs of the reshaped kernel for each GEMMParam 28 29These 4 files are the current representations we use for what we call the **heuristics** of a GEMM op: given a GEMMParam, 30what kernel and subsequently what configurations for that kernels are the most performant. 31 32## Step-by-step example 33 34### Step1: Prepare the shape and configs files 351. We first need to identify the shapes that we are interested in and store them in a csv file, say *gemm_shapes.csv*. 362. Then we need to specify a set of good GEMMConfig candidates for each kernel in 3 separate csv files (this requires 37 some prior heuristics, but can be provided by the Compute Library developers upon requests, based on your target device). 38 39 Say we have *gemm_configs_native.csv", "gemm_configs_reshaped.csv" and "gemm_configs_reshaped_only_rhs.csv". 40 41 Please refer to the Prerequisite section for more details 42 43### Step2: Push relevant files to the target device 44All the files that need to be present on the target device are: 45* benchmark script: \<ComputeLibrary\>/examples/gemm_tuner/benchmark_gemm_examples.sh 46* shapes and configs csv files: gemm_shapes.csv, gemm_configs_native.csv, gemm_configs_reshaped_only_rhs.csv, gemm_configs_reshaped.csv 47* Example benchmark binaries: \<ComputeLibrary\>/build/tests/gemm_tuner/benchmark_cl_gemm* 48 49### Step3: Collect benchmark data 50With these files on device, we can collect benchmark data using the script. Assume all the example binaries are pushed 51to a folder called *gemm_tuner*. While logged onto our device: 52``` 53# Native 54./benchmark_gemm_examples.sh -s native -e ./gemm_tuner -g ./gemm_shapes.csv -c ./gemm_configs_native.csv -o results/native 55# Reshaped Only RHS 56./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 57# Reshaped 58./benchmark_gemm_examples.sh -s reshaped -e ./gemm_tuner -g ./gemm_shapes.csv -c ./gemm_configs_reshaped.csv -o results/reshaped 59``` 60You can repeat the 3 commands above to have a bit redundancy in your benchmark data (as you can imagine, measurement is noisy), 61but you may need to change the output folder for each repeat 62 63### Step4: Generate the heuristics 641. After benchmarking, we pull the benchmark data, the *results* folder, from the target device to our host machine 652. We use the GemmTuner.py script to give us the heuristics 66 ``` 67 python3 <ComputeLibrary>/examples/gemm_tuner/GemmTuner.py -b ./results -o heuristics 68 ``` 69 When it's finished, there should be 4 json files in the *heuristics* folder 70 71One thing to notice is that the config heuristics might give more than 1 recommendations for each GEMMParam, because 72we accept all good GEMMConfigs with a tolerance. If you want fewer recommendations, you can decrease the tolerance by 73passing a lower value to *-t \<tolerance\>* to the GemmTuner.py script. 74 75## Prerequisite 76* A target device to be tuned, plus the following on the device: 77 * Android or Linux OS 78 * Bash shell 79 * Built Compute Library with benchmark examples binaries 80 * benchmark_gemm_examples.sh script 81 * gemm shape file 82 83 A csv file containing the **GEMMParam search list**. This is the list of GEMMParams/gemm shapes that we're 84 interested in (For more details see Approach section). The default list is prepared by Compute Library developers in advance 85 and can be provided on request. 86 87 The format is described as: 88 89 A headerless csv file with fields separated by commas. 90 91 A gemm shape is a list of 4 positive integers \<M, N, K, B\> describing the shapes of the two matrices (LHS and 92 RHS) with: 93 94 M - Number of lhs matrix rows 95 N - Number of rhs matrix columns 96 K - Number of lhs matrix columns/rhs matrix rows 97 B - Batch size 98 99 An example gemm shape file looks like: 100 ``` 101 100,100,30,1 102 100,100,30,3 103 ... 104 ``` 105 * gemm config file 106 A csv file containing the **GEMMConfig search list**. This is the list of candidate GEMMConfigs among which we 107 search for the optimal one. **Note that we have a different list for each strategy.** 108 The default lists are prepared by Compute Library developers in advance and can be provided on request. 109 110 The format of the file for each strategy is the same: 111 112 A headerless csv file with fields separated by commas. 113 114 However the fields of GEMMConfig differ for each strategy: 115 116 * Strategy **native**: 117 A gemm config is a list of 3 positive integers \<m0, n0, k0\>, with: 118 119 m0 - Number of rows processed by the matrix multiplication 120 n0 - Number of columns processed by the matrix multiplication 121 k0 - Number of partial accumulations performed by the matrix multiplication 122 123 Only the following configurations of M0, N0 and K0 are currently supported: 124 125 M0 = 1, 2, 3, 4, 5, 6, 7, 8 126 N0 = 2, 3, 4, 8, 16 127 K0 = 2, 3, 4, 8, 16 128 129 An example gemm config file looks like: 130 ``` 131 1,4,4 132 2,3,8 133 ... 134 ``` 135 * Strategy **reshaped_rhs_only**: 136 A gemm config is a list of 4 positive integers <m0, n0, k0, h0> and 3 boolean values: 137 138 m0 - Number of rows processed by the matrix multiplication 139 n0 - Number of columns processed by the matrix multiplication 140 k0 - Number of partial accumulations performed by the matrix multiplication 141 h0 - Number of horizontal blocks of size (k0xn0) stored on the same output row 142 interleave_rhs - Interleave rhs matrix (1) / Do not interleave rhs matrix (0) 143 transpose_rhs - Transpose rhs matrix (1) / Do not transpose rhs matrix (0) 144 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 145 with certain combinations of the GEMMParams and other configs. Please refer to CLGEMMReshapeRHSMatrixKernel 146 for more details 147 148 Only the following configurations of M0, N0 and K0 are currently supported: 149 150 M0 = 1, 2, 3, 4, 5, 6, 7, 8 151 N0 = 2, 3, 4, 8, 16 152 K0 = 2, 3, 4, 8, 16 153 H0 >= 1 154 155 An example gemm config file looks like: 156 ``` 157 4,4,4,1,1,1,0 158 4,4,4,3,1,0,1 159 ... 160 ``` 161 * Strategy **reshaped**: 162 A gemm config is a list of 5 positive integers <m0, n0, k0, v0, h0> and 4 boolean values: 163 164 m0 - Number of rows processed by the matrix multiplication 165 n0 - Number of columns processed by the matrix multiplication 166 k0 - Number of partial accumulations performed by the matrix multiplication 167 v0 - Number of vertical blocks of size (m0xk0) stored on the same output row 168 h0 - Number of horizontal blocks of size (k0xn0) stored on the same output row 169 interleave_lhs - Interleave lhs matrix (1) / Do not interleave lhs matrix (0) 170 interleave_rhs - Interleave rhs matrix (1) / Do not interleave rhs matrix (0) 171 transpose_rhs - Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do transpose lhs matrix (0) 172 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 173 with certain combinations of the GEMMParams and other configs. Please refer to CLGEMMReshapeRHSMatrixKernel 174 for more details 175 176 If rhs matrix is transposed only the following configurations are currently supported: 177 178 M0 = 2, 3, 4, 5, 6, 7, 8 179 N0 = 2, 3, 4, 8, 16 180 K0 = 2, 3, 4, 8, 16 181 V0 >= 1 182 H0 >= 1 183 184 If lhs matrix is transposed only the following configurations are currently supported: 185 186 M0 = 2, 3, 4, 8 187 N0 = 2, 3, 4, 8, 16 188 K0 = 2, 3, 4, 8, 16 189 V0 >= 1 190 H0 >= 1 191 192 An example gemm config file looks like: 193 ``` 194 4,4,4,1,3,1,1,1,0 195 4,4,4,3,3,1,1,0,1 196 ... 197 ``` 198* A host machine, plus these on the machine: 199 * python >= 3.6 200 * GemmTuner.py script 201 202## Usage 203The usage of the 2 scripts: 204 2051. benchmark_gemm_examples.sh 206 207 Run the shell script (**benchmark_gemm_examples.sh**) on your **target device**. Note that all the built benchmark 208 examples: build/tests/gemm_tuner/benchmark_cl_gemm*, have to be present on your target device prior to running. 209 The benchmark results will be saved to json files in an output directory. 210 ``` 211 Usage: benchmark_gemm_examples.sh [-h] -s \<strategy\> -e \<example_binary_dir\> -g \<gemm_shape_file\> 212 -c \<gemm_config_file\> [-d \<data_type\>] [-o \<out_dir\>] 213 214 Options: 215 -h 216 Print help messages. If a strategy is specified with -s <strategy>, then only display messages relevant to that 217 strategy. Otherwise if no strategy is specified, display messages for all available strategies. 218 219 -s <strategy> 220 Strategy option. 221 Options: ${ALL_STRATEGY_OPTIONS[@]}. 222 223 -e <example_binary_dir> 224 Path to directory that holds all example binaries 225 226 -g <gemm_shape_file> 227 Path to gemm shape csv file 228 229 -c <gemm_config_file> 230 Path to gemm config csv file 231 232 -d <data_type> 233 Data type option with which to run benchmark examples 234 Default: ${DEFAULT_DATA_TYPE} 235 Supported options: 236 Strategy : Data Types 237 Native : F32 238 Reshaped : F16, F32 239 Reshaped RHS Only : F16, F32 240 241 -o <out_dir> 242 Path to output directory that holds output json files 243 Default: ${DEFAULT_OUT_DIR} 244 ``` 2452. GemmTuner.py: 246 247 Run the python script (**GemmTuner.py**) on your **host machine**. 248 You'll need to transfer all the benchmark result json files generated from the previous step to your host machine 249 beforehand. The script will output the best kernel and gemm configurations for each gemm param in the 4 output json files 250 ``` 251 Usage: GemmTuner.py [-h] -b PATH [-o PATH] [-t TOLERANCE] [-D] 252 253 CL GEMM Tuner 254 optional arguments: 255 -h, --help show this help message and exit 256 -b PATH, --benchmark_results PATH 257 Path to benchmark result directory, where benchmark 258 result json files have a file extension of 259 'gemmtuner_benchmark' 260 -o PATH, --output_dir PATH 261 Path to directory that holds output json files. 262 -t TOLERANCE, --tolerance TOLERANCE 263 For testing if two GEMMConfigs are equivalent in terms 264 of performance. The tolerance is OpenCL timer in 265 milliseconds. Recommended value: <= 0.1 ms 266 -D, --debug Enable script debugging output 267 268 ```