Benchmark improvements.

- Use multiple devices if present
- FLAGS more agnostic to diverse platforms

PiperOrigin-RevId: 598990299
Change-Id: Ifbd454ca521a788324f2b62f1c79e7d4b8305a30

mjx/mujoco/mjx/benchmark/benchmark.py

@@ -27,76 +27,51 @@ from mujoco import mjx
 
 FLAGS = flags.FLAGS
 
-_PATHS = {
-    'humanoid': 'benchmark/model/humanoid/humanoid.xml',
-    'barkour': 'benchmark/model/barkour_v0/assets/barkour_v0_mjx.xml',
-    'shadow_hand': 'benchmark/model/shadow_hand/scene_right.xml',
-}
-
-_BATCH_SIZE = {
-    ('barkour', 'tpu_v5e'): 1024,
-    ('humanoid', 'tpu_v5e'): 1024,
-    ('shadow_hand', 'tpu_v5e'): 1024,
-    ('barkour', 'gpu_a100'): 8192,
-    ('humanoid', 'gpu_a100'): 8192,
-    ('shadow_hand', 'gpu_a100'): 4096,
-    ('barkour', 'cpu'): 64,
-    ('humanoid', 'cpu'): 64,
-    ('shadow_hand', 'cpu'): 64,
-}
-
-_SOLVER_CONFIG = {
-    ('barkour', 'tpu_v5e'): (mujoco.mjtSolver.mjSOL_CG, 4, 6),
-    ('humanoid', 'tpu_v5e'): (mujoco.mjtSolver.mjSOL_CG, 6, 6),
-    ('shadow_hand', 'tpu_v5e'): (mujoco.mjtSolver.mjSOL_CG, 8, 6),
-    ('humanoid', 'gpu_a100'): (mujoco.mjtSolver.mjSOL_NEWTON, 1, 4),
-    ('barkour', 'gpu_a100'): (mujoco.mjtSolver.mjSOL_NEWTON, 1, 4),
-    ('shadow_hand', 'gpu_a100'): (mujoco.mjtSolver.mjSOL_NEWTON, 1, 4),
-    ('barkour', 'cpu'): (mujoco.mjtSolver.mjSOL_NEWTON, 1, 4),
-    ('humanoid', 'cpu'): (mujoco.mjtSolver.mjSOL_NEWTON, 1, 4),
-    ('shadow_hand', 'cpu'): (mujoco.mjtSolver.mjSOL_NEWTON, 1, 4),
-}
+flags.DEFINE_string('mjcf', None, 'path to model', required=True)
+flags.DEFINE_integer('step_count', 1000, 'number of steps per rollout')
+flags.DEFINE_integer('batch_size', 1024, 'number of parallel rollouts')
+flags.DEFINE_integer('unroll', 1, 'loop unroll length')
+flags.DEFINE_enum('solver', 'cg', ['cg', 'newton'], 'constraint solver')
+flags.DEFINE_integer('iterations', 1, 'number of solver iterations')
+flags.DEFINE_integer('ls_iterations', 4, 'number of linesearch iterations')
 
 
-flags.DEFINE_string('model', 'humanoid', 'Model to benchmark')
-flags.DEFINE_enum('device', 'cpu', ('cpu', 'tpu_v5e', 'gpu_a100'),
-                  'Device benchmark is running on')
-
-
-def _measure_fn(state, init_fn, step_fn, batch_size: int = 1024) -> float:
+def _measure(state, init_fn, step_fn) -> float:
   """Reports jit time and op time for a function."""
 
-  step_count = 100 if FLAGS.device == 'cpu' else 1000
-
-  @jax.jit
+  @jax.pmap
   def run_batch(seed: jp.ndarray):
+    batch_size = FLAGS.batch_size // jax.device_count()
     rngs = jax.random.split(jax.random.PRNGKey(seed), batch_size)
-    init_state = jax.vmap(init_fn)(rngs)
+    state = jax.vmap(init_fn)(rngs)
 
     @jax.vmap
-    def run(state):
-      def step(state, _):
-        state = step_fn(state)
-        return state, ()
+    def step(state, _):
+      state = step_fn(state)
+      return state, None
 
-      return jax.lax.scan(step, state, (), length=step_count)
-
-    return run(init_state)
+    state, _ = jax.lax.scan(
+        step, state, None, length=FLAGS.step_count, unroll=FLAGS.unroll
+    )
+    return state
 
   # run once to jit
   beg = time.perf_counter()
-  jax.tree_util.tree_map(lambda x: x.block_until_ready(), run_batch(0))
+  seed = 0
+  seeds = jp.arange(seed, seed + jax.device_count(), dtype=int)
+  jax.tree_util.tree_map(lambda x: x.block_until_ready(), run_batch(seeds))
   first_t = time.perf_counter() - beg
 
   times = []
   while state:
+    seed += jax.device_count()
+    seeds = jp.arange(seed, seed + jax.device_count(), dtype=int)
     beg = time.perf_counter()
-    batch = run_batch(jp.array(len(times)))
-    jax.tree_util.tree_map(lambda x: x.block_until_ready(), batch)
+    jax.tree_util.tree_map(lambda x: x.block_until_ready(), run_batch(seeds))
     times.append(time.perf_counter() - beg)
 
   op_time = jp.mean(jp.array(times))
-  batch_sps = batch_size * step_count / op_time
+  batch_sps = FLAGS.batch_size * FLAGS.step_count / op_time
 
   state.counters['jit_time'] = first_t - op_time
   state.counters['batch_sps'] = batch_sps
@@ -106,11 +81,14 @@ def _measure_fn(state, init_fn, step_fn, batch_size: int = 1024) -> float:
 def _run(state: benchmark.State):
   """Benchmark a model."""
 
-  f = epath.resource_path('mujoco.mjx') / _PATHS[FLAGS.model]
+  f = epath.resource_path('mujoco.mjx') / 'benchmark/model' / FLAGS.mjcf
   m = mujoco.MjModel.from_xml_path(f.as_posix())
-  m.opt.solver, m.opt.iterations, m.opt.ls_iterations = _SOLVER_CONFIG[
-      (FLAGS.model, FLAGS.device)
-  ]
+  m.opt.solver = {
+      'cg': mujoco.mjtSolver.mjSOL_CG,
+      'newton': mujoco.mjtSolver.mjSOL_NEWTON,
+  }[FLAGS.solver.lower()]
+  m.opt.iterations = FLAGS.iterations
+  m.opt.ls_iterations = FLAGS.ls_iterations
   m = mjx.device_put(m)
 
   def init(rng):
@@ -122,11 +100,10 @@ def _run(state: benchmark.State):
   def step(d):
     return mjx.step(m, d)
 
-  batch_size = _BATCH_SIZE[(FLAGS.model, FLAGS.device)]
-  _measure_fn(state, init, step, batch_size=batch_size)
+  _measure(state, init, step)
 
 
 if __name__ == '__main__':
   FLAGS(sys.argv)
-  benchmark.register(_run, name=FLAGS.model + '_' + FLAGS.device)
+  benchmark.register(_run, name=sys.argv[0].split('/')[-1])
   benchmark.main()