Skip to content

Benchmark suiteslink

IREE Benchmarks Suites is a collection of benchmarks for IREE developers to track performance improvements/regressions during development.

The benchmark suites are run for each commit on the main branch and the results are uploaded to for regression analysis (for the current supported targets). On pull requests, users can add labels benchmarks:* to trigger the benchmark runs. The results will be compared with and post in the comments.

Information about the definitions of the benchmark suites can be found in the IREE Benchmark Suites Configurations.

Running benchmark suites locallylink


Install iree-import-tf and iree-import-tflite in your Python environment (see Tensorflow Integration and TFLite Integration).

Choose benchmark presetslink

IREE Benchmark Suites contain many benchmarks for different devices and model sizes, which can take lots of space and time to build all of them. So benchmarks are grouped into presets to allow building and running only a subset of them. The available presets are:

Execution benchmarks:

  • android-cpu: benchmarks for mobile CPUs
  • android-gpu: benchmarks for mobile GPUs
  • cuda: benchmarks for CUDA with a small model set
  • cuda-large: benchmarks for CUDA with a large model set
  • vulkan-nvidia: benchmarks for Vulkan on NVIDIA graphics cards
  • x86_64: benchmarks for x86_64 CPUs with a small model set
  • x86_64-large: benchmarks for x86_64 with a large model set

Compilation benchmarks (to collect compilation statistics, such as module sizes):

  • comp-stats: compilation benchmarks with a small model set
  • comp-stats-large: compilation benchmark with a large model set

Note that *-large presets will download and build a few hundreds GBs of artifacts.

Set the environment variables of benchmark presets for the steps below, for example:


Build benchmark suiteslink


cmake -GNinja -B "${IREE_BUILD_DIR?}" -S "${IREE_REPO?}" \

If you only need the imported MLIR models:

cmake --build "${IREE_BUILD_DIR?}" --target \
  # For large benchmarks (this will take > 100G disk space)
  # iree-benchmark-import-models-large

Otherwise, compile the benchmark suites and tools for benchmarking:

cmake --build "${IREE_BUILD_DIR?}" --target \
  iree-benchmark-suites \
  # If any *-large preset is enabled, also build this target:
  # iree-benchmark-suites-large \
export E2E_TEST_ARTIFACTS_DIR="${IREE_BUILD_DIR?}/e2e_test_artifacts"

Run benchmarkslink

Export the execution benchmark config:

build_tools/benchmarks/ execution \
  --benchmark_presets="${EXECUTION_BENCHMARK_PRESETS?}" \
  > "${E2E_TEST_ARTIFACTS_DIR?}/exec_config.json"

Run benchmarks (currently only support running on a Linux host):

build_tools/benchmarks/ \
  --benchmark_tool_dir="${IREE_BUILD_DIR?}/tools" \
  --e2e_test_artifacts_dir="${E2E_TEST_ARTIFACTS_DIR?}" \
  --execution_benchmark_config="${E2E_TEST_ARTIFACTS_DIR?}/exec_config.json" \
  --target_device_name="<target_device_name, e.g. c2-standard-60>" \
  --output="${E2E_TEST_ARTIFACTS_DIR?}/benchmark_results.json" \
  --verbose \
  --cpu_uarch="<host CPU uarch, e.g. CascadeLake>"

Note that:

  • <target_device_name> selects a benchmark group targets a specific device:
  • To run x86_64 benchmarks, right now --cpu_uarch needs to be provided and only CascadeLake is available currently.

Filters can be used to select the benchmarks:

build_tools/benchmarks/ \
  --benchmark_tool_dir="${IREE_BUILD_DIR?}/tools" \
  --e2e_test_artifacts_dir="${E2E_TEST_ARTIFACTS_DIR?}" \
  --execution_benchmark_config="${E2E_TEST_ARTIFACTS_DIR?}/exec_config.json" \
  --target_device_name="c2-standard-60" \
  --output="${E2E_TEST_ARTIFACTS_DIR?}/benchmark_results.json" \
  --verbose \
  --cpu_uarch="CascadeLake" \
  --model_name_regex="MobileBert*" \
  --driver_filter_regex='local-task' \

Generate compilation statistics (compilation benchmarks)link

Export the compilation benchmark config:

build_tools/benchmarks/ compilation \
  --benchmark_presets="${COMPILATION_BENCHMARK_PRESETS?}" \
  > "${E2E_TEST_ARTIFACTS_DIR?}/comp_config.json"

Generate the compilation statistics:

build_tools/benchmarks/ \
  --compilation_benchmark_config=comp_config.json \
  --e2e_test_artifacts_dir="${E2E_TEST_ARTIFACTS_DIR?}" \
  --build_log="${IREE_BUILD_DIR?}/.ninja_log" \

Note that you need to use Ninja to build the benchmark suites as the tool collects information from its build log.

Show execution / compilation benchmark resultslink

If you want to generate a comparison report locally, you can use script to compare two result json files and generate the report. For example:

build_tools/benchmarks/ \
  --base "${E2E_TEST_ARTIFACTS_DIR?}/before_benchmark_results.json" \
  --target "${E2E_TEST_ARTIFACTS_DIR?}/after_benchmark_results.json" \

An example that compares compilation statistics:

build_tools/benchmarks/ \
  --base-compile-stats "${E2E_TEST_ARTIFACTS_DIR?}/before_compile_stats_results.json" \
  --target-compile-stats "${E2E_TEST_ARTIFACTS_DIR?}/after_compile_stats_results.json" \

Find compile and run commands to reproduce benchmarkslink

Each benchmark has its benchmark ID in the benchmark suites, you will see a benchmark ID at:

  • In the serie's URL of
    • Execution benchmark:<benchmark_id>
    • Compilation benchmark:<benchmark_id>-<metric_id>
  • In benchmark_results.json and compile_stats_results.json
    • Execution benchmark result has a field run_config_id
    • Compilation benchmark result has a field gen_config_id
  • In PR benchmark summary or the markdown generated by, each benchmark has the link to its URL, which includes the benchmark ID.

If you don't have artifacts locally, see Fetching Benchmark Artifacts from CI to find the GCS directory of the CI artifacts. Then fetch the needed files:

# Get ${E2E_TEST_ARTIFACTS_DIR_URL} from "Fetching Benchmark Artifacts from CI".
export E2E_TEST_ARTIFACTS_DIR="e2e_test_artifacts"

# Download all artifacts
gcloud storage cp -r "${E2E_TEST_ARTIFACTS_DIR_URL?}" "${E2E_TEST_ARTIFACTS_DIR?}"

Run the helper tool to dump benchmark commands from benchmark configs:

build_tools/benchmarks/ dump-cmds \
  --execution_benchmark_config="${E2E_TEST_ARTIFACTS_DIR?}/execution-benchmark-config.json" \
  --compilation_benchmark_config="${E2E_TEST_ARTIFACTS_DIR?}/compilation-benchmark-config.json" \
  --e2e_test_artifacts_dir="${E2E_TEST_ARTIFACTS_DIR?}" \

Get full list of benchmarkslink

The commands below output the full list of execution and compilation benchmarks, including the benchmark names and their flags:

build_tools/benchmarks/ execution > "${E2E_TEST_ARTIFACTS_DIR?}/exec_config.json"
build_tools/benchmarks/ compilation > "${E2E_TEST_ARTIFACTS_DIR?}/comp_config.json"
build_tools/benchmarks/ dump-cmds \
  --execution_benchmark_config="${E2E_TEST_ARTIFACTS_DIR?}/exec_config.json" \

Fetching benchmark Artifacts from CIlink

1. Find the corresponding CI workflow runlink

On the commit of the benchmark run, you can find the list of the workflow jobs by clicking the green check mark. Click any job starts with CI /:


2. Get URLs of GCS artifactslink

On the CI page, click Summary on the top-left to open the summary page. Scroll down and the links to artifacts are listed in a section titled "Artifact Links". Paste the content in your shell to define all needed variables for the following steps:


3. Fetch the benchmark artifactslink

To fetch files from the GCS URL, the gcloud CLI tool ( can list the directory contents and download files (see for more usages). If you want to use CI artifacts to reproduce benchmarks locally, see Find Compile and Run Commands to Reproduce Benchmarks.

Assume you get the GCS URL variables from Get URLs of GCS artifacts.

Download artifacts:

# The GCS directory has the same structure as your local ${IREE_BUILD_DIR?}/e2e_test_artifacts.
gcloud storage ls "${E2E_TEST_ARTIFACTS_DIR_URL?}"

# Download all source and imported MLIR files:
gcloud storage cp "${E2E_TEST_ARTIFACTS_DIR_URL?}/*.mlir" "<target_dir>"

Execution and compilation benchmark configs can be downloaded at:

# Execution benchmark config:
gcloud storage cp \
  "${E2E_TEST_ARTIFACTS_DIR_URL?}/execution-benchmark-config.json" \

# Compilation benchmark config:
gcloud storage cp \
  "${E2E_TEST_ARTIFACTS_DIR_URL?}/compilation-benchmark-config.json" \

Benchmark raw results can be downloaded at:

# Execution benchmark raw results
gcloud storage cp "${EXECUTION_BENCHMARK_RESULTS_DIR_URL?}/benchmark-results-*.json" .

# Optional: Merge raw results into a single file
build_tools/benchmarks/ merge-results benchmark-results-*.json > benchmark_results.json

# Compilation benchmark results