From 784755d888eb41e7d148cd9ce9b1bda2c6d6ebed Mon Sep 17 00:00:00 2001
From: mariecwhite <mariewhite@google.com>
Date: Sat, 7 Oct 2023 03:23:02 +0000
Subject: [PATCH] Benchmark TFLite

---
 .github/workflows/run_tflite_benchmark.yml    | 123 +++++++++++++
 .../comparative_suite/tf/model_definitions.py |   2 +-
 experimental/ggml/benchmark_ggml.sh           |   4 +-
 experimental/iree/benchmark_iree.sh           |  61 ++++++
 .../iree/set_android_scaling_governor.sh      |  51 ++++++
 experimental/tflite/benchmark_lib.py          | 151 +++++++++++++++
 experimental/tflite/benchmark_tflite.sh       |  96 ++++++++++
 experimental/tflite/requirements.txt          |   2 +
 experimental/tflite/run_benchmarks.py         | 137 ++++++++++++++
 experimental/tflite/run_benchmarks_android.py | 173 ++++++++++++++++++
 .../tflite/set_android_scaling_governor.sh    |  51 ++++++
 experimental/tflite/setup_venv.sh             |  32 ++++
 12 files changed, 880 insertions(+), 3 deletions(-)
 create mode 100644 .github/workflows/run_tflite_benchmark.yml
 create mode 100755 experimental/iree/benchmark_iree.sh
 create mode 100755 experimental/iree/set_android_scaling_governor.sh
 create mode 100644 experimental/tflite/benchmark_lib.py
 create mode 100755 experimental/tflite/benchmark_tflite.sh
 create mode 100644 experimental/tflite/requirements.txt
 create mode 100755 experimental/tflite/run_benchmarks.py
 create mode 100755 experimental/tflite/run_benchmarks_android.py
 create mode 100755 experimental/tflite/set_android_scaling_governor.sh
 create mode 100644 experimental/tflite/setup_venv.sh

diff --git a/.github/workflows/run_tflite_benchmark.yml b/.github/workflows/run_tflite_benchmark.yml
new file mode 100644
index 00000000..46839d08
--- /dev/null
+++ b/.github/workflows/run_tflite_benchmark.yml
@@ -0,0 +1,123 @@
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+#
+# TFLite Benchmarks Workflow.
+
+name: TFLite Benchmarks
+
+on:
+  workflow_dispatch:
+  pull_request:
+
+concurrency:
+  # A PR number if a pull request and otherwise the commit hash. This cancels
+  # queued and in-progress runs for the same PR (presubmit) or commit
+  # (postsubmit).
+  group: ${{ github.workflow }}-${{ github.event.number || github.sha }}
+  cancel-in-progress: true
+
+env:
+  GCS_DIR: gs://openxla-github-actions-${{ github.event_name == 'pull_request' && 'presubmit' || 'postsubmit' }}-artifacts/${{ github.run_id }}/${{ github.run_attempt }}
+
+jobs:
+  setup:
+    runs-on: ubuntu-22.04
+    outputs:
+      runner-group: ${{ steps.configure.outputs.runner-group }}
+      benchmark-gcs-dir: ${{ steps.configure.outputs.benchmark-gcs-dir }}
+    steps:
+      - name: "Checking out PR repository"
+        uses: actions/checkout@e2f20e631ae6d7dd3b768f56a5d2af784dd54791  # v2.5.0
+      - name: "Configuring CI options"
+        id: configure
+        env:
+          RUNNER_GROUP: ${{ github.event_name == 'pull_request' && 'presubmit' || 'postsubmit' }}
+        run: |
+          # Just informative logging. There should only be two commits in the
+          # history here, but limiting the depth helps when copying from a local
+          # repo instead of using checkout, e.g. with
+          # https://github.com/nektos/act where there will be more.
+          git log --oneline --graph --max-count=3
+          # Workflow jobs can't access `env` in `runs-on`, so we need to make
+          # `runner-group` a job output variable.
+          echo "runner-group=${RUNNER_GROUP}" > "${GITHUB_OUTPUT}"
+
+          # For presubmit testing, the result artifacts are uploaded to the
+          # temporary workflow GCS dir. In postsubmit, the result artifacts are
+          # uploaded to the comparative benchmark GCS dir.
+          if [[ "${RUNNER_GROUP}" == "presubmit" ]]; then
+            BENCHMARK_GCS_DIR="${GCS_DIR}/comparative-benchmark-artifacts"
+          else
+            BENCHMARK_GCS_DIR="gs://comparative-benchmark-artifacts/$(date +'%Y-%m-%d').$(date +'%s')"
+          fi
+          echo "benchmark-gcs-dir=${BENCHMARK_GCS_DIR}" >> "${GITHUB_OUTPUT}"
+
+  benchmark_on_c2-standard-16:
+    needs: [setup]
+    runs-on:
+      - self-hosted  # must come first
+      - runner-group=${{ needs.setup.outputs.runner-group }}
+      - environment=prod
+      - machine-type=c2-standard-16
+    env:
+      BENCHMARK_GCS_DIR: ${{ needs.setup.outputs.benchmark-gcs-dir }}
+      RESULTS_DIR: results-dir
+      TARGET_DEVICE: c2-standard-16
+      TFLITE_TOOL_DIR: tool-dir
+    steps:
+      - name: "Checking out PR repository"
+        uses: actions/checkout@e2f20e631ae6d7dd3b768f56a5d2af784dd54791  # v2.5.0
+      - name: "Setup"
+        id: setup
+        run: |
+          echo "results-gcs-dir=${BENCHMARK_GCS_DIR}/${TARGET_DEVICE}-results" >> "${GITHUB_OUTPUT}"
+          mkdir "${RESULTS_DIR}"
+          mkdir "${TFLITE_TOOL_DIR}"
+      - name: "Benchmarking TFLite"
+        env:
+          TFLITE_RESULTS_JSON: tflite.json
+          RESULTS_GCS_DIR: ${{ steps.setup.outputs.results-gcs-dir }}
+        run: |
+          RESULTS_PATH="${RESULTS_DIR}/${TFLITE_RESULTS_JSON}"
+          docker run --mount="type=bind,src="${PWD}",target=/work" --workdir="/work" \
+            "gcr.io/iree-oss/openxla-benchmark/base@sha256:1bf3e319465ec8fb465baae3f6ba9a5b09cb84a5349a675c671a552fc77f2251" \
+            ./experimental/tflite/benchmark_tflite.sh \
+              "${TARGET_DEVICE}" \
+              "${TFLITE_TOOL_DIR}" \
+              "${RESULTS_PATH}"
+          gcloud storage cp "${RESULTS_PATH}" "${RESULTS_GCS_DIR}/"
+
+  benchmark_on_pixel-6-pro:
+    needs: [setup]
+    runs-on:
+      - self-hosted  # must come first
+      - runner-group=${{ needs.setup.outputs.runner-group }}
+      - environment=prod
+      - machine-type=pixel-6-pro
+    env:
+      BENCHMARK_GCS_DIR: ${{ needs.setup.outputs.benchmark-gcs-dir }}
+      RESULTS_DIR: results-dir
+      TARGET_DEVICE: pixel-6-pro
+      TFLITE_TOOL_DIR: tool-dir
+    steps:
+      - name: "Checking out PR repository"
+        uses: actions/checkout@e2f20e631ae6d7dd3b768f56a5d2af784dd54791  # v2.5.0
+      - name: "Setup"
+        id: setup
+        run: |
+          echo "results-gcs-dir=${BENCHMARK_GCS_DIR}/${TARGET_DEVICE}-results" >> "${GITHUB_OUTPUT}"
+          mkdir "${RESULTS_DIR}"
+          mkdir "${TFLITE_TOOL_DIR}"
+      - name: "Benchmarking TFLite"
+        env:
+          TFLITE_RESULTS_JSON: tflite.json
+          RESULTS_GCS_DIR: ${{ steps.setup.outputs.results-gcs-dir }}
+        run: |
+          ./experimental/iree/benchmark_iree.sh
+
+          #RESULTS_PATH="${RESULTS_DIR}/${TFLITE_RESULTS_JSON}"
+          #./experimental/tflite/benchmark_tflite.sh "${TARGET_DEVICE}" "${TFLITE_TOOL_DIR}" "${RESULTS_PATH}"
+          #gcloud storage cp "${RESULTS_PATH}" "${RESULTS_GCS_DIR}/"
diff --git a/common_benchmark_suite/openxla/benchmark/comparative_suite/tf/model_definitions.py b/common_benchmark_suite/openxla/benchmark/comparative_suite/tf/model_definitions.py
index 5c048aea..79737c32 100644
--- a/common_benchmark_suite/openxla/benchmark/comparative_suite/tf/model_definitions.py
+++ b/common_benchmark_suite/openxla/benchmark/comparative_suite/tf/model_definitions.py
@@ -10,7 +10,7 @@
 from openxla.benchmark import def_types
 from openxla.benchmark.comparative_suite import utils
 
-PARENT_GCS_DIR = "https://storage.googleapis.com/iree-model-artifacts/tensorflow/tf_models_2.15.0.dev20230817_1692333975/"
+PARENT_GCS_DIR = "https://storage.googleapis.com/iree-model-artifacts/tensorflow/tf_models_2.15.0.dev20230829_1696537918/"
 ARTIFACTS_DIR_URL_TEMPLATE = string.Template(PARENT_GCS_DIR + "${name}")
 
 # T5-Large models.
diff --git a/experimental/ggml/benchmark_ggml.sh b/experimental/ggml/benchmark_ggml.sh
index de2778b9..5cdcf44d 100755
--- a/experimental/ggml/benchmark_ggml.sh
+++ b/experimental/ggml/benchmark_ggml.sh
@@ -88,8 +88,8 @@ if [[ "${TARGET_DEVICE_NAME}" =~ ^(pixel-4|pixel-6-pro|moto-edge-x30)$ ]]; then
   adb shell "su root sh /data/local/tmp/set_android_scaling_governor.sh performance"
 else
   BENCHMARK_SCRIPT="run_benchmarks.py"
-  # c2-standard-16 has 16 cores.
-  THREADS="1,8,16"
+  # c2-standard-16 has 8 cores.
+  THREADS="1,8"
 
   args+=(
      --threads "${THREADS}"
diff --git a/experimental/iree/benchmark_iree.sh b/experimental/iree/benchmark_iree.sh
new file mode 100755
index 00000000..a843f122
--- /dev/null
+++ b/experimental/iree/benchmark_iree.sh
@@ -0,0 +1,61 @@
+#!/bin/bash
+
+# This is a temporary hack to run IREE benchmarks on pixel-6-pro since
+# it's currently not working in the IREE repo.
+
+ROOT_DIR=/tmp/iree-benchmarks
+TD="$(cd $(dirname $0) && pwd)"
+
+rm -rf "${ROOT_DIR}"
+mkdir "${ROOT_DIR}"
+pushd "${ROOT_DIR}"
+
+# Download benchmark tool.
+gsutil cp "gs://iree-github-actions-presubmit-artifacts/6464567954/1/benchmark-tools/android-armv8.2-a-benchmark-tools.tar" .
+tar -xf "android-armv8.2-a-benchmark-tools.tar"
+adb push "android-armv8.2-a-benchmark-tools-dir/build/tools/iree-benchmark-module" "/data/local/tmp"
+adb shell "chmod +x /data/local/tmp/iree-benchmark-module"
+
+# Download vmfb's.
+
+
+# Setup environment.
+adb push "${TD}/set_android_scaling_governor.sh" "/data/local/tmp"
+adb shell "chmod +x /data/local/tmp/set_android_scaling_governor.sh"
+adb shell "su root sh /data/local/tmp/set_android_scaling_governor.sh performance"
+
+# Benchmark.
+ITERATIONS=10
+gsutil cp "gs://iree-github-actions-presubmit-artifacts/6464567954/1/e2e-test-artifacts/iree_module_BertLarge_Fp32_Batch1_tflite___armv8.2-a-generic-linux_android29-llvm_cpu__experimental-flags_data-tiling_ukernel_/module.vmfb" "BertLarge_Batch1.vmfb"
+adb push "BertLarge_Batch1.vmfb" "/data/local/tmp"
+adb shell "taskset f0 /data/local/tmp/iree-benchmark-module --function=main --input=1x384xi32=0 --input=1x384xi32=0 --device_allocator=caching --task_topology_group_count=4 --device=local-task --module=/data/local/tmp/BertLarge_Batch1.vmfb --time_unit=ns --benchmark_format=json --benchmark_out_format=json --print_statistics=true --benchmark_repetitions=${ITERATIONS}"
+adb shell "rm /data/local/tmp/BertLarge_Batch1.vmfb"
+rm "BertLarge_Batch1.vmfb"
+
+gsutil cp "gs://iree-github-actions-presubmit-artifacts/6464567954/1/e2e-test-artifacts/iree_module_BertLarge_Fp32_Batch16_tflite___armv8.2-a-generic-linux_android29-llvm_cpu__experimental-flags_data-tiling_ukernel_/module.vmfb" "BertLarge_Batch16.vmfb"
+adb push "BertLarge_Batch16.vmfb" "/data/local/tmp"
+adb shell "taskset f0 /data/local/tmp/iree-benchmark-module --function=main --input=16x384xi32=0 --input=16x384xi32=0 --device_allocator=caching --task_topology_group_count=4 --device=local-task --module=/data/local/tmp/BertLarge_Batch16.vmfb --time_unit=ns --benchmark_format=json --benchmark_out_format=json --print_statistics=true --benchmark_repetitions=${ITERATIONS}"
+adb shell "rm /data/local/tmp/BertLarge_Batch16.vmfb"
+rm "BertLarge_Batch16.vmfb"
+
+gsutil cp "gs://iree-github-actions-presubmit-artifacts/6464567954/1/e2e-test-artifacts/iree_module_BertLarge_Fp32_Batch24_tflite___armv8.2-a-generic-linux_android29-llvm_cpu__experimental-flags_data-tiling_ukernel_/module.vmfb" "BertLarge_Batch24.vmfb"
+adb push "BertLarge_Batch24.vmfb" "/data/local/tmp"
+adb shell "taskset f0 /data/local/tmp/iree-benchmark-module --function=main --input=24x384xi32=0 --input=24x384xi32=0 --device_allocator=caching --task_topology_group_count=4 --device=local-task --module=/data/local/tmp/BertLarge_Batch24.vmfb --time_unit=ns --benchmark_format=json --benchmark_out_format=json --print_statistics=true --benchmark_repetitions=${ITERATIONS}"
+adb shell "rm /data/local/tmp/BertLarge_Batch24.vmfb"
+rm "BertLarge_Batch24.vmfb"
+
+gsutil cp "gs://iree-github-actions-presubmit-artifacts/6464567954/1/e2e-test-artifacts/iree_module_BertLarge_Fp32_Batch32_tflite___armv8.2-a-generic-linux_android29-llvm_cpu__experimental-flags_data-tiling_ukernel_/module.vmfb" "BertLarge_Batch32.vmfb"
+adb push "BertLarge_Batch32.vmfb" "/data/local/tmp"
+adb shell "taskset f0 /data/local/tmp/iree-benchmark-module --function=main --input=32x384xi32=0 --input=32x384xi32=0 --device_allocator=caching --task_topology_group_count=4 --device=local-task --module=/data/local/tmp/BertLarge_Batch32.vmfb --time_unit=ns --benchmark_format=json --benchmark_out_format=json --print_statistics=true --benchmark_repetitions=${ITERATIONS}"
+adb shell "rm /data/local/tmp/BertLarge_Batch32.vmfb"
+rm "BertLarge_Batch32.vmfb"
+
+adb shell "rm -rf /data/local/tmp/*"
+
+popd
+rm -rf "${ROOT_DIR}"
+
+
+
+
+
diff --git a/experimental/iree/set_android_scaling_governor.sh b/experimental/iree/set_android_scaling_governor.sh
new file mode 100755
index 00000000..9f51e273
--- /dev/null
+++ b/experimental/iree/set_android_scaling_governor.sh
@@ -0,0 +1,51 @@
+#!/bin/sh
+
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+# Runs on an android device itself to set the frequency scaling governor for all
+# CPUs (default performance).
+
+################################### WARNING ####################################
+# This will overheat the phone if it's not on a cooling plate, resulting in    #
+# thermal throttling. To prevent anything catching on fire, the actual CPU     #
+# frequencies will be throttled to below the maximum, skewing your results.    #
+################################################################################
+
+set -euo pipefail
+
+GOVERNOR="${1:-performance}"
+
+echo "CPU info (before changing governor):"
+echo 'cpu\tgovernor\tcur\tmin\tmax'
+echo "------------------------------------------------"
+for i in `cat /sys/devices/system/cpu/present | tr '-' ' ' | xargs seq`; do \
+    echo "cpu${i}" | paste \
+      - \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/scaling_governor" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_cur_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_min_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_max_freq"; \
+done
+
+echo "Setting CPU frequency governor to ${GOVERNOR}"
+
+for i in `cat /sys/devices/system/cpu/present | tr '-' ' ' | xargs seq`; do \
+  echo "${GOVERNOR}" > \
+    "/sys/devices/system/cpu/cpu${i?}/cpufreq/scaling_governor"; \
+done
+
+echo "CPU info (after changing governor):"
+echo 'cpu\tgovernor\tcur\tmin\tmax'
+echo "------------------------------------------------"
+for i in `cat /sys/devices/system/cpu/present | tr '-' ' ' | xargs seq`; do \
+    echo "cpu${i}" | paste \
+      - \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/scaling_governor" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_cur_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_min_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_max_freq"; \
+done
diff --git a/experimental/tflite/benchmark_lib.py b/experimental/tflite/benchmark_lib.py
new file mode 100644
index 00000000..48bd217f
--- /dev/null
+++ b/experimental/tflite/benchmark_lib.py
@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+import argparse
+import dataclasses
+import json
+import pathlib
+import re
+import subprocess
+import sys
+from typing import Sequence, List
+
+# Add common_benchmark_suite dir to the search path.
+sys.path.insert(
+    0, str(pathlib.Path(__file__).parents[2] / "common_benchmark_suite"))
+from openxla.benchmark import def_types, devices
+from openxla.benchmark.comparative_suite.tf import benchmark_definitions as tf_benchmark_definitions
+
+# Add common_benchmark_suite dir to the search path.
+sys.path.insert(
+    0, str(pathlib.Path(__file__).parents[2] / "comparative_benchmark"))
+import utils
+
+ALL_DEVICE_NAMES = [device.name for device in devices.ALL_DEVICES]
+TFLITE_FP32_FILENAME = "model_fp32.tflite"
+
+MIN_MAX_LATENCY_REGEXP = re.compile(
+    "INFO: count=\d+ first=\d+ curr=\d+ min=(.*) max=(.*) avg=(.*) std=(.*)")
+AVG_LATENCY_REGEXP = re.compile(
+    "INFO: Inference timings in us: .* Inference \(avg\): (.*)")
+PEAK_MEMORY_REGEXP = re.compile(
+    "INFO: Overall peak memory footprint \(MB\) via periodic monitoring: (.*)")
+
+
+def download_artifacts(benchmarks: Sequence[def_types.BenchmarkCase],
+                       root_dir: pathlib.Path,
+                       verbose: bool = False):
+  """Download benchmark artifacts."""
+  download_list = []
+  for benchmark in benchmarks:
+    model = benchmark.model
+    if (model.artifacts_dir_url is None or
+        def_types.ModelArtifactType.TFLITE_FP32
+        not in model.exported_model_types):
+      raise ValueError(f"XLA HLO dump isn't provided by '{model.name}'.")
+    model_url = model.artifacts_dir_url + "/" + TFLITE_FP32_FILENAME
+    model_path = root_dir / model.name / TFLITE_FP32_FILENAME
+    download_list.append((model_url, model_path))
+
+  utils.download_files(download_list, verbose=verbose)
+
+
+def configure_parser(parser: argparse.ArgumentParser):
+  parser.add_argument("-o",
+                      "--output",
+                      type=pathlib.Path,
+                      required=True,
+                      help="JSON file path to merge the results.")
+  parser.add_argument("-name",
+                      "--benchmark_name",
+                      required=True,
+                      help="The unique id that defines a benchmark.")
+  parser.add_argument("-device",
+                      "--target_device",
+                      dest="target_device_name",
+                      type=str,
+                      required=True,
+                      choices=ALL_DEVICE_NAMES,
+                      help="The target device to benchmark.")
+  parser.add_argument("--tflite-benchmark-tool",
+                      "--tflite_benchmark_tool",
+                      type=pathlib.Path,
+                      required=True,
+                      help="The path to the TFLite `benchmrk_model` tool.")
+  parser.add_argument("-t",
+                      "--threads",
+                      type=str,
+                      default="1,4",
+                      help="A comma-delimited list of threads.")
+  parser.add_argument("-iter",
+                      "--iterations",
+                      type=int,
+                      default=10,
+                      help="The number of iterations to benchmark.")
+  parser.add_argument("--root-dir",
+                      "--root_dir",
+                      type=pathlib.Path,
+                      default=pathlib.Path("/tmp/openxla-benchmark/tflite"),
+                      help="Root directory stores benchmark artifacts.")
+  parser.add_argument("--no-download",
+                      "--no_download",
+                      action="store_true",
+                      help="Don't automatically download benchmark artifacts.")
+  parser.add_argument("--verbose",
+                      action="store_true",
+                      help="Show verbose messages.")
+
+
+def benchmark(benchmark_command: List[str], benchmark_definition: dict,
+              iterations: int, verbose: bool) -> utils.BenchmarkResult:
+  if verbose:
+    print(f"Run command: {benchmark_command}")
+
+  result = subprocess.run(benchmark_command,
+                          stdout=subprocess.PIPE,
+                          stderr=subprocess.STDOUT)
+  result_text = result.stdout.decode("utf-8")
+
+  if verbose:
+    print(result_text)
+
+  min_latency_ms = None
+  max_latency_ms = None
+  mean_latency_ms = None
+  stddev_latency_ms = None
+  device_memory_peak_mb = None
+
+  match = AVG_LATENCY_REGEXP.search(result_text)
+  if match:
+    mean_latency_ms = float(match.group(1)) * 1e-3
+
+  match = MIN_MAX_LATENCY_REGEXP.search(result_text)
+  if match:
+    min_latency_ms = float(match.group(1)) * 1e-3
+    max_latency_ms = float(match.group(2)) * 1e-3
+    stddev_latency_ms = float(match.group(4)) * 1e-3
+
+  match = PEAK_MEMORY_REGEXP.search(result_text)
+  if match:
+    device_memory_peak_mb = float(match.group(1))
+
+  metrics = {
+      "min_latency_ms": min_latency_ms,
+      "max_latency_ms": max_latency_ms,
+      "mean_latency_ms": mean_latency_ms,
+      "stddev_latency_ms": stddev_latency_ms,
+      "benchmark_iterations": iterations,
+      "device_memory_peak_mb": device_memory_peak_mb,
+  }
+
+  return utils.BenchmarkResult(
+      definition=benchmark_definition,
+      metrics={
+          "compiler_level": metrics,
+      },
+  )
diff --git a/experimental/tflite/benchmark_tflite.sh b/experimental/tflite/benchmark_tflite.sh
new file mode 100755
index 00000000..dd301bae
--- /dev/null
+++ b/experimental/tflite/benchmark_tflite.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+#
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+#
+# Environment variables:
+# PYTHON: Python interpreter, default: /usr/bin/python3
+# OOBI_TARGET_DEVICE: target benchmark device, can also be specified the first
+#   argument.
+# OOBI_TOOL_DIR: path to save benchmark tools.
+# OOBI_OUTPUT: path to output benchmark results, can also be specified the
+#   third argument.
+#
+# Example usage:
+# ./benchmark_tflite.sh <target-device> <tool-dir> <result-path>
+
+set -xeuo pipefail
+
+VENV_DIR="${OOBI_VENV_DIR:-tflite-benchmarks.venv}"
+PYTHON="${PYTHON:-"$(which python3)"}"
+TARGET_DEVICE="${1:-${OOBI_TARGET_DEVICE}}"
+TOOL_DIR="${2:-${OOBI_TOOL_DIR}}"
+OUTPUT_PATH="${3:-${OOBI_OUTPUT}}"
+
+# Setup virtual environment.
+TD="$(cd $(dirname $0) && pwd)"
+VENV_DIR="${VENV_DIR}" PYTHON="${PYTHON}" source "${TD}/setup_venv.sh"
+
+# Initialize results json.
+OUTPUT_PATH="$(realpath ${OUTPUT_PATH})"
+"${TD}/../../comparative_benchmark/scripts/create_results_json.sh" "${OUTPUT_PATH}"
+
+declare -a args=(
+  --tflite_benchmark_tool "${TOOL_DIR}/tflite_benchmark_model"
+  --output "${OUTPUT_PATH}"
+  --target_device "${TARGET_DEVICE}"
+  --verbose
+)
+
+# Download TFLite benchmark tool depending on target.
+if [[ "${TARGET_DEVICE}" =~ ^(pixel-4|pixel-6-pro|moto-edge-x30)$ ]]; then
+  wget -O "${TOOL_DIR}/tflite_benchmark_model" https://storage.googleapis.com/tensorflow-nightly-public/prod/tensorflow/release/lite/tools/nightly/latest/android_arm_benchmark_model_plus_flex
+
+  # Setup mobile device for benchmarking.
+  adb push "${TD}/set_android_scaling_governor.sh" "/data/local/tmp"
+  adb shell "chmod +x /data/local/tmp/set_android_scaling_governor.sh"
+  adb shell "su root sh /data/local/tmp/set_android_scaling_governor.sh performance"
+
+  BENCHMARK_SCRIPT="run_benchmarks_android.py"
+
+  # Pixel 6 has a maximum of 8 cores.
+  THREADS="1,4"
+  TASKSETS="80,f0"
+
+  args+=(
+     --threads "${THREADS}"
+     --tasksets "${TASKSETS}"
+     --iterations 10
+  )
+
+  declare -a BENCHMARK_NAMES=(
+    #"models/RESNET50_FP32_TF_.+_BATCH(1|8)/.+"
+    "models/BERT_LARGE_FP32_TF_.+_BATCH(1|16|24|32)/.+"
+    #"models/T5_LARGE_FP32_TF_.+_BATCH(1|16)/.+"
+    #"models/EFFICIENTNETB7_FP32_TF_.+_BATCH1/.+"
+  )
+else
+  wget -O "${TOOL_DIR}/tflite_benchmark_model" https://storage.googleapis.com/tensorflow-nightly-public/prod/tensorflow/release/lite/tools/nightly/latest/linux_x86-64_benchmark_model_plus_flex
+  chmod +x "${TOOL_DIR}/tflite_benchmark_model"
+
+  BENCHMARK_SCRIPT="run_benchmarks.py"
+
+  # c2-standard-16 has 8 cores.
+  THREADS="1,8"
+  args+=(
+     --threads "${THREADS}"
+     --iterations 20
+  )
+
+  declare -a BENCHMARK_NAMES=(
+    #"models/RESNET50_FP32_TF_.+_BATCH(1|64|128)/.+"
+    "models/BERT_LARGE_FP32_TF_.+_BATCH(1|16|24|32)/.+"
+    #"models/T5_LARGE_FP32_TF_.+_BATCH(1|16|24|32)/.+"
+    #"models/EFFICIENTNETB7_FP32_TF_.+_BATCH(1|64|128)/.+"
+  )
+fi
+
+for i in ${!BENCHMARK_NAMES[@]}; do
+  args+=(
+    --benchmark_name "${BENCHMARK_NAMES[$i]}"
+  )
+  "${TD}/${BENCHMARK_SCRIPT}" "${args[@]}"
+done
diff --git a/experimental/tflite/requirements.txt b/experimental/tflite/requirements.txt
new file mode 100644
index 00000000..945b4703
--- /dev/null
+++ b/experimental/tflite/requirements.txt
@@ -0,0 +1,2 @@
+numpy
+requests
diff --git a/experimental/tflite/run_benchmarks.py b/experimental/tflite/run_benchmarks.py
new file mode 100755
index 00000000..9550caa3
--- /dev/null
+++ b/experimental/tflite/run_benchmarks.py
@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+import argparse
+import dataclasses
+import json
+import pathlib
+import re
+import subprocess
+import sys
+from typing import Sequence
+
+import benchmark_lib
+
+# Add common_benchmark_suite dir to the search path.
+sys.path.insert(
+    0, str(pathlib.Path(__file__).parents[2] / "common_benchmark_suite"))
+from openxla.benchmark import def_types, devices
+from openxla.benchmark.comparative_suite.tf import benchmark_definitions as tf_benchmark_definitions
+
+# Add common_benchmark_suite dir to the search path.
+sys.path.insert(
+    0, str(pathlib.Path(__file__).parents[2] / "comparative_benchmark"))
+import utils
+
+ALL_DEVICE_NAMES = [device.name for device in devices.ALL_DEVICES]
+TFLITE_FP32_FILENAME = "model_fp32.tflite"
+
+LATENCY_REGEXP = re.compile(
+    "INFO: count=\d+ first=\d+ curr=\d+ min=(.*) max=(.*) avg=(.*) std=(.*)")
+PEAK_MEMORY_REGEXP = re.compile(
+    "INFO: Overall peak memory footprint \(MB\) via periodic monitoring: (.*)")
+
+
+def _run(
+    benchmark: def_types.BenchmarkCase,
+    target_device: def_types.DeviceSpec,
+    iterations: int,
+    num_threads: str,
+    tflite_benchmark_tool: pathlib.Path,
+    tflite_model_path: pathlib.Path,
+    verbose: bool,
+) -> utils.BenchmarkResult:
+  model = benchmark.model
+  data_type = model.model_parameters["data_type"]
+  batch_size = model.model_parameters["batch_size"]
+  benchmark_definition = {
+      "benchmark_name": benchmark.name,
+      "framework": str(model.model_impl.framework_type),
+      "data_type": data_type,
+      "batch_size": batch_size,
+      "compiler": "TFLite",
+      "device": target_device.name,
+      "num_threads": num_threads,
+      "num_iterations": iterations,
+      "tags": model.model_impl.tags + model.tags,
+  }
+  cmd = [
+      tflite_benchmark_tool,
+      f"--graph={tflite_model_path}",
+      f"--num_runs={iterations}",
+      f"--num_threads={num_threads}",
+      f"--report_peak_memory_footprint=true",
+  ]
+
+  return benchmark_lib.benchmark(cmd, benchmark_definition, iterations, verbose)
+
+
+def _parse_arguments() -> argparse.Namespace:
+  parser = argparse.ArgumentParser(description="Run TFLite benchmarks.")
+  benchmark_lib.configure_parser(parser)
+  return parser.parse_args()
+
+
+def main(
+    benchmark_name: str,
+    target_device_name: str,
+    output: pathlib.Path,
+    root_dir: pathlib.Path,
+    threads: str,
+    tflite_benchmark_tool: pathlib.Path,
+    iterations: int,
+    no_download: bool,
+    verbose: bool,
+):
+  name_pattern = re.compile(f"^{benchmark_name}$")
+  all_benchmarks = tf_benchmark_definitions.ALL_BENCHMARKS
+  benchmarks = [
+      benchmark for benchmark in all_benchmarks
+      if name_pattern.match(benchmark.name)
+  ]
+
+  if not benchmarks:
+    all_benchmark_names = "\n".join(
+        benchmark.name for benchmark in all_benchmarks)
+    raise ValueError(f'No benchmark matches "{benchmark_name}".'
+                     f' Available benchmarks:\n{all_benchmark_names}')
+
+  try:
+    target_device = next(device for device in devices.ALL_DEVICES
+                         if device.name == target_device_name)
+  except StopIteration:
+    raise ValueError(f'Target device "{target_device_name}" is not defined.'
+                     f' Available device options:\n{ALL_DEVICE_NAMES}')
+
+  if not no_download:
+    benchmark_lib.download_artifacts(benchmarks=benchmarks,
+                                     root_dir=root_dir,
+                                     verbose=verbose)
+
+  threads = threads.split(",")
+  for benchmark in benchmarks:
+    tflite_model_path = root_dir / benchmark.model.name / TFLITE_FP32_FILENAME
+    if not tflite_model_path.exists():
+      raise ValueError(f"TFLite model not found: '{tflite_model_path}'.")
+
+    for num_threads in threads:
+      result = _run(benchmark=benchmark,
+                    target_device=target_device,
+                    iterations=iterations,
+                    num_threads=num_threads,
+                    tflite_benchmark_tool=tflite_benchmark_tool,
+                    tflite_model_path=tflite_model_path,
+                    verbose=verbose)
+      if verbose:
+        print(json.dumps(dataclasses.asdict(result), indent=2))
+
+      utils.append_benchmark_result(output, result)
+
+
+if __name__ == "__main__":
+  main(**vars(_parse_arguments()))
diff --git a/experimental/tflite/run_benchmarks_android.py b/experimental/tflite/run_benchmarks_android.py
new file mode 100755
index 00000000..ea742d5b
--- /dev/null
+++ b/experimental/tflite/run_benchmarks_android.py
@@ -0,0 +1,173 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+import argparse
+import dataclasses
+import json
+import pathlib
+import re
+import subprocess
+import sys
+from typing import Sequence
+
+import benchmark_lib
+
+# Add common_benchmark_suite dir to the search path.
+sys.path.insert(
+    0, str(pathlib.Path(__file__).parents[2] / "common_benchmark_suite"))
+from openxla.benchmark import def_types, devices
+from openxla.benchmark.comparative_suite.tf import benchmark_definitions as tf_benchmark_definitions
+
+# Add common_benchmark_suite dir to the search path.
+sys.path.insert(
+    0, str(pathlib.Path(__file__).parents[2] / "comparative_benchmark"))
+import utils
+
+ALL_DEVICE_NAMES = [device.name for device in devices.ALL_DEVICES]
+TFLITE_FP32_FILENAME = "model_fp32.tflite"
+
+LATENCY_REGEXP = re.compile(
+    "INFO: count=\d+ first=\d+ curr=\d+ min=(.*) max=(.*) avg=(.*) std=(.*)")
+PEAK_MEMORY_REGEXP = re.compile(
+    "INFO: Overall peak memory footprint \(MB\) via periodic monitoring: (.*)")
+
+
+def _run(
+    benchmark: def_types.BenchmarkCase,
+    target_device: def_types.DeviceSpec,
+    iterations: int,
+    num_threads: str,
+    taskset: str,
+    tflite_benchmark_tool: pathlib.Path,
+    tflite_model_path: pathlib.Path,
+    verbose: bool,
+) -> utils.BenchmarkResult:
+  model = benchmark.model
+  data_type = model.model_parameters["data_type"]
+  batch_size = model.model_parameters["batch_size"]
+  benchmark_definition = {
+      "benchmark_name": benchmark.name,
+      "framework": str(model.model_impl.framework_type),
+      "data_type": data_type,
+      "batch_size": batch_size,
+      "compiler": "TFLite",
+      "device": target_device.name,
+      "num_threads": num_threads,
+      "num_iterations": iterations,
+      "tags": model.model_impl.tags + model.tags,
+  }
+  cmd = [
+      "adb",
+      "shell",
+      "taskset",
+      taskset,
+      tflite_benchmark_tool,
+      f"--graph={tflite_model_path}",
+      f"--num_runs={iterations}",
+      f"--num_threads={num_threads}",
+      f"--report_peak_memory_footprint=true",
+  ]
+
+  return benchmark_lib.benchmark(cmd, benchmark_definition, iterations, verbose)
+
+
+def _parse_arguments() -> argparse.Namespace:
+  parser = argparse.ArgumentParser(description="Run TFLite benchmarks.")
+  parser.add_argument(
+      "--tasksets",
+      type=str,
+      default="f0",
+      help=
+      "A comma-separated list of tasksets to run under each thread configuration."
+  )
+  benchmark_lib.configure_parser(parser)
+  return parser.parse_args()
+
+
+def main(
+    benchmark_name: str,
+    target_device_name: str,
+    output: pathlib.Path,
+    root_dir: pathlib.Path,
+    threads: str,
+    tasksets: str,
+    tflite_benchmark_tool: pathlib.Path,
+    iterations: int,
+    no_download: bool,
+    verbose: bool,
+):
+  name_pattern = re.compile(f"^{benchmark_name}$")
+  all_benchmarks = tf_benchmark_definitions.ALL_BENCHMARKS
+  benchmarks = [
+      benchmark for benchmark in all_benchmarks
+      if name_pattern.match(benchmark.name)
+  ]
+
+  if not benchmarks:
+    all_benchmark_names = "\n".join(
+        benchmark.name for benchmark in all_benchmarks)
+    raise ValueError(f'No benchmark matches "{benchmark_name}".'
+                     f' Available benchmarks:\n{all_benchmark_names}')
+
+  try:
+    target_device = next(device for device in devices.ALL_DEVICES
+                         if device.name == target_device_name)
+  except StopIteration:
+    raise ValueError(f'Target device "{target_device_name}" is not defined.'
+                     f' Available device options:\n{ALL_DEVICE_NAMES}')
+
+  if not no_download:
+    benchmark_lib.download_artifacts(benchmarks=benchmarks,
+                                     root_dir=root_dir,
+                                     verbose=verbose)
+
+  threads = threads.split(",")
+  tasksets = tasksets.split(",")
+  if len(threads) != len(tasksets):
+    raise ValueError(
+        "The number of tasksets specified must be equal to the number of threads."
+    )
+
+  # Push artifacts to the Android device.
+  subprocess.run(["adb", "push", tflite_benchmark_tool, "/data/local/tmp"])
+  subprocess.run([
+      "adb", "shell", "chmod", "+x",
+      f"/data/local/tmp/{tflite_benchmark_tool.name}"
+  ])
+
+  for benchmark in benchmarks:
+    tflite_model_path = root_dir / benchmark.model.name / TFLITE_FP32_FILENAME
+    if not tflite_model_path.exists():
+      raise ValueError(f"TFLite model not found: '{tflite_model_path}'.")
+    subprocess.run(["adb", "push", tflite_model_path, "/data/local/tmp"])
+
+    for taskset, num_threads in zip(tasksets, threads):
+      result = _run(benchmark=benchmark,
+                    target_device=target_device,
+                    iterations=iterations,
+                    num_threads=num_threads,
+                    taskset=taskset,
+                    tflite_benchmark_tool=pathlib.Path(
+                        f"/data/local/tmp/{tflite_benchmark_tool.name}"),
+                    tflite_model_path=pathlib.Path(
+                        f"/data/local/tmp/{tflite_model_path.name}"),
+                    verbose=verbose)
+      if verbose:
+        print(json.dumps(dataclasses.asdict(result), indent=2))
+
+      utils.append_benchmark_result(output, result)
+
+    subprocess.run(
+        ["adb", "shell", "rm", f"/data/local/tmp/{tflite_model_path.name}"])
+
+  subprocess.run(
+      ["adb", "shell", "rm", f"/data/local/tmp/{tflite_benchmark_tool.name}"])
+
+
+if __name__ == "__main__":
+  main(**vars(_parse_arguments()))
diff --git a/experimental/tflite/set_android_scaling_governor.sh b/experimental/tflite/set_android_scaling_governor.sh
new file mode 100755
index 00000000..9f51e273
--- /dev/null
+++ b/experimental/tflite/set_android_scaling_governor.sh
@@ -0,0 +1,51 @@
+#!/bin/sh
+
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+# Runs on an android device itself to set the frequency scaling governor for all
+# CPUs (default performance).
+
+################################### WARNING ####################################
+# This will overheat the phone if it's not on a cooling plate, resulting in    #
+# thermal throttling. To prevent anything catching on fire, the actual CPU     #
+# frequencies will be throttled to below the maximum, skewing your results.    #
+################################################################################
+
+set -euo pipefail
+
+GOVERNOR="${1:-performance}"
+
+echo "CPU info (before changing governor):"
+echo 'cpu\tgovernor\tcur\tmin\tmax'
+echo "------------------------------------------------"
+for i in `cat /sys/devices/system/cpu/present | tr '-' ' ' | xargs seq`; do \
+    echo "cpu${i}" | paste \
+      - \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/scaling_governor" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_cur_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_min_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_max_freq"; \
+done
+
+echo "Setting CPU frequency governor to ${GOVERNOR}"
+
+for i in `cat /sys/devices/system/cpu/present | tr '-' ' ' | xargs seq`; do \
+  echo "${GOVERNOR}" > \
+    "/sys/devices/system/cpu/cpu${i?}/cpufreq/scaling_governor"; \
+done
+
+echo "CPU info (after changing governor):"
+echo 'cpu\tgovernor\tcur\tmin\tmax'
+echo "------------------------------------------------"
+for i in `cat /sys/devices/system/cpu/present | tr '-' ' ' | xargs seq`; do \
+    echo "cpu${i}" | paste \
+      - \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/scaling_governor" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_cur_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_min_freq" \
+      "/sys/devices/system/cpu/cpu${i}/cpufreq/cpuinfo_max_freq"; \
+done
diff --git a/experimental/tflite/setup_venv.sh b/experimental/tflite/setup_venv.sh
new file mode 100644
index 00000000..e7bd3960
--- /dev/null
+++ b/experimental/tflite/setup_venv.sh
@@ -0,0 +1,32 @@
+#!/bin/bash
+#
+# Copyright 2023 The OpenXLA Authors
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+#
+# Sets up a virtual environment suitable for running TFLite benchmarks.
+#
+# Environment variables:
+#   VENV_DIR=tflite-benchmarks.venv
+#   PYTHON=/usr/bin/python3.10
+
+set -euo pipefail
+
+TD="$(cd $(dirname $0) && pwd)"
+VENV_DIR="${VENV_DIR:-tflite-benchmarks.venv}"
+PYTHON="${PYTHON:-"$(which python3)"}"
+
+echo "Setting up venv dir: ${VENV_DIR}"
+
+"${PYTHON}" -m venv "${VENV_DIR}" || echo "Could not create venv."
+source "${VENV_DIR}/bin/activate" || echo "Could not activate venv"
+
+# Upgrade pip and install requirements. 'python' is used here in order to
+# reference to the python executable from the venv.
+python -m pip install --upgrade pip || echo "Could not upgrade pip"
+python -m pip install --upgrade -r "${TD}/requirements.txt"
+
+echo "Activate venv with:"
+echo "  source ${VENV_DIR}/bin/activate"