Convolution Benchmarks

Pre-requisites

Create a virtual environment and install python dependencies.

python3 -m venv ~/venv/convperf
source ~/venv/convperf/bin/activate
pip install -r requirements.txt

For multi-threaded libxsmm, you will need to install OpenMP. You will have to install the appropriate package for your compiler. For example, if your compiler is clang++-14, then you can install OpenMP by doing the following.

sudo apt install libomp-14-dev

Build Instructions

cmake -GNinja -B build .
cmake --build build

Run Benchmarks

export NUM_THREADS=1
python convperf.py --benchmark_tool build/tools/benchmark_conv --runners iree,xsmm --benchmark_sizes benchmark_sizes/resnet50.json --num_threads=${NUM_THREADS}

This will run the benchmarks and write the results to runtimes.csv. This file can be visualized with the following command.

python convperf.py --visualize --runtimes_file runtimes.json

This will generate convs.png which will visualize the runtimes of the different methods.

Extracting generated artifacts

The mlir file that contains the convolutions can be found in the build directory in build/iree/convs.mlir. This file contains functions for all the sizes specified in the benchmark sizes file. Sample mlir is shown below.

func.func @conv2d_1x230x230x3_7x7x3x64(%arg0: tensor<1x230x230x3xf32>, %arg1: tensor<7x7x3x64xf32>) -> tensor<1x112x112x64xf32> {
  %cst_0 = arith.constant 0.000000e+00 : f32
  %0 = tensor.empty() : tensor<1x112x112x64xf32>
  %1 = linalg.fill ins(%cst_0 : f32) outs(%0 : tensor<1x112x112x64xf32>) -> tensor<1x112x112x64xf32>
  %2 = linalg.conv_2d_nhwc_hwcf {dilations = dense<1> : tensor<2xi64>, strides = dense<2> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x230x230x3xf32>, tensor<7x7x3x64xf32>) outs(%1 : tensor<1x112x112x64xf32>) -> tensor<1x112x112x64xf32>
  return %2 : tensor<1x112x112x64xf32>
}

And here is an example with padding.

func.func @conv2d_1x56x56x64_3x3x64x64(%arg0: tensor<1x56x56x64xf32>, %arg1: tensor<3x3x64x64xf32>) -> tensor<1x56x56x64xf32> {
  %cst_0 = arith.constant 0.000000e+00 : f32
  %0 = tensor.empty() : tensor<1x56x56x64xf32>
  %1 = linalg.fill ins(%cst_0 : f32) outs(%0 : tensor<1x56x56x64xf32>) -> tensor<1x56x56x64xf32>
  %2 = tensor.pad %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0] {
         ^bb0(%arg2: index, %arg3: index, %arg4: index, %arg5: index):
                tensor.yield %cst_0 : f32
       } : tensor<1x56x56x64xf32> to tensor<1x58x58x64xf32>
  %3 = linalg.conv_2d_nhwc_hwcf {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%2, %arg1 : tensor<1x58x58x64xf32>, tensor<3x3x64x64xf32>) outs(%1 : tensor<1x56x56x64xf32>) -> tensor<1x56x56x64xf32>
  return %3 : tensor<1x56x56x64xf32>
}