reintroduce the Threefry GPU kernel lowering, under a flag

On GPU, the Threefry PRNG implementation no longer lowers to a kernel call by default. This choice can improve runtime memory usage at a compile-time cost. Prior behavior, which produces a kernel call, can be recovered with:

   `jax.config.update('jax_threefry_gpu_kernel_lowering', True)`

PiperOrigin-RevId: 629763763

CHANGELOG.md

@@ -33,6 +33,12 @@ Remember to align the itemized text with the first line of an item within a list
     be created and threaded in and out of computations to build up dependency.
     The singleton object `core.token` has been removed, users now should create
     and use fresh `core.Token` objects instead.
+  * On GPU, the Threefry PRNG implementation no longer lowers to a kernel call
+    by default. This choice can improve runtime memory usage at a compile-time
+    cost. Prior behavior, which produces a kernel call, can be recovered with
+    `jax.config.update('jax_threefry_gpu_kernel_lowering', True)`. If the new
+    default causes issues, please file a bug. Otherwise, we intend to remove
+    this flag in a future release.
 
 * Deprecations & Removals
   * Pallas now exclusively uses XLA for compiling kernels on GPU. The old

jax/_src/config.py

@@ -210,6 +210,7 @@ def trace_context():
           dynamic_shapes.value, numpy_dtype_promotion.value,
           default_device.value, random_seed_offset.value,
           threefry_partitionable.value,
+          threefry_gpu_kernel_lowering.value,
           softmax_custom_jvp.value,
           enable_memories.value,
           disable_jit.value,
@@ -811,6 +812,7 @@ class _GlobalExtraJitContext(NamedTuple):
   dynamic_shapes: bool = False
   random_seed_offset: int = 0
   threefry_partitionable: bool = False
+  threefry_gpu_kernel_lowering: bool = False
   softmax_custom_jvp: bool = False
   xla_profile_version: int = 0
 
@@ -845,6 +847,7 @@ class _ThreadLocalExtraJitContext(NamedTuple):
   dynamic_shapes: bool | None = None
   random_seed_offset: int | None = None
   threefry_partitionable: bool | None = None
+  threefry_gpu_kernel_lowering: bool | None = None
   softmax_custom_jvp: bool | None = None
   xla_profile_version: int | None = None
 
@@ -1083,6 +1086,17 @@ def _update_jax_memories_thread_local(val):
     update_thread_local_hook=lambda val: update_thread_local_jit_state(
         threefry_partitionable=val))
 
+threefry_gpu_kernel_lowering = define_bool_state(
+    name='jax_threefry_gpu_kernel_lowering',
+    default=False,
+    help=('On GPU, lower threefry PRNG operations to a kernel implementation. '
+          'This makes compile times faster at a potential runtime memory '
+          'cost.'),
+    update_global_hook=lambda val: _update_global_jit_state(
+        threefry_gpu_kernel_lowering=val),
+    update_thread_local_hook=lambda val: update_thread_local_jit_state(
+        threefry_gpu_kernel_lowering=val))
+
 
 softmax_custom_jvp = define_bool_state(
     name='jax_softmax_custom_jvp',

jax/_src/prng.py

@@ -47,8 +47,9 @@
 from jax._src.interpreters import xla
 from jax._src.lax import lax as lax_internal
 from jax._src.lax import utils as lax_utils
-from jax._src.lib.mlir import ir
+from jax._src.lib import gpu_prng
 from jax._src.lib import xla_client as xc
+from jax._src.lib.mlir import ir
 from jax._src.lib.mlir.dialects import hlo
 from jax._src.numpy.array_methods import (
     _array_operators, _set_array_base_attributes, _IndexUpdateHelper)
@@ -1002,17 +1003,63 @@ def _threefry2x32_lowering(key1, key2, x1, x2, use_rolled_loops=True):
   return tuple(x)
 
 
+_threefry2x32_lowering_rule = mlir.lower_fun(
+    partial(_threefry2x32_lowering, use_rolled_loops=False),
+    multiple_results=True)
+
+_threefry2x32_cpu_lowering_rule = mlir.lower_fun(
+    partial(_threefry2x32_lowering, use_rolled_loops=True),
+    multiple_results=True)
+
+
+def _threefry2x32_gpu_lowering_rule(lowering_func, ctx, k1, k2, x1, x2):
+  if not config.threefry_gpu_kernel_lowering.value:  # back to default lowering
+    return _threefry2x32_lowering_rule(ctx, k1, k2, x1, x2)
+
+  aval_out, aval_out_2 = ctx.avals_out
+  assert aval_out == aval_out_2
+  k1_aval, k2_aval, x1_aval, x2_aval = ctx.avals_in
+  rank = len(aval_out.shape)
+  if 0 in aval_out.shape:
+    zeros = mlir.full_like_aval(ctx, 0, aval_out)
+    return [zeros, zeros]
+  def _broadcast(x, aval):
+    return mlir.broadcast_in_dim(ctx, x, aval_out,
+                                 broadcast_dimensions=range(rank - len(aval.shape), rank))
+
+  out_len = reduce(op.mul, aval_out.shape, 1)
+  if not core.is_constant_dim(out_len):
+    length = mlir.eval_dynamic_shape_as_tensor(ctx, [out_len])
+    length = mlir.hlo.convert(
+        ir.RankedTensorType.get((1,), ir.IntegerType.get_signless(64)),
+        length)
+    output_shape = mlir.eval_dynamic_shape_as_tensor(ctx, aval_out.shape)
+  else:
+    length = int(out_len)  # will be passed statically
+    output_shape = None
+
+  return lowering_func(
+          (_broadcast(k1, k1_aval), _broadcast(k2, k2_aval)),
+          (_broadcast(x1, x1_aval), _broadcast(x2, x2_aval)), length,
+          output_shape)
+
 threefry2x32_p = core.Primitive("threefry2x32")
 threefry2x32_p.multiple_results = True
 threefry2x32_p.def_impl(partial(dispatch.apply_primitive, threefry2x32_p))
 threefry2x32_p.def_abstract_eval(_threefry2x32_abstract_eval)
 batching.defbroadcasting(threefry2x32_p)
-mlir.register_lowering(threefry2x32_p, mlir.lower_fun(
-    partial(_threefry2x32_lowering, use_rolled_loops=False),
-    multiple_results=True))
-mlir.register_lowering(threefry2x32_p, mlir.lower_fun(
-    partial(_threefry2x32_lowering, use_rolled_loops=True),
-    multiple_results=True), platform='cpu')
+mlir.register_lowering(
+    threefry2x32_p, _threefry2x32_lowering_rule)
+mlir.register_lowering(
+    threefry2x32_p, _threefry2x32_cpu_lowering_rule, platform='cpu')
+mlir.register_lowering(
+    threefry2x32_p,
+    partial(_threefry2x32_gpu_lowering_rule, gpu_prng.cuda_threefry2x32),
+    platform='cuda')
+mlir.register_lowering(
+    threefry2x32_p,
+    partial(_threefry2x32_gpu_lowering_rule, gpu_prng.rocm_threefry2x32),
+    platform='rocm')
 
 
 def iota_2x32_shape(shape):

jax/experimental/export/_export.py

@@ -793,7 +793,7 @@ def _check_lowering(lowering) -> None:
 # Their backwards compatibility is tested by back_compat_test.py.
 _CUSTOM_CALL_TARGETS_GUARANTEED_STABLE = {
     "Sharding", "SPMDFullToShardShape", "SPMDShardToFullShape",
-    "dynamic_ducc_fft",
+    "dynamic_ducc_fft", "cu_threefry2x32",
     # cholesky on CPU
     "lapack_spotrf", "lapack_dpotrf", "lapack_cpotrf", "lapack_zpotrf",
     # eigh on CPU

tests/export_back_compat_test.py

@@ -65,7 +65,8 @@
 
 
 @jtu.with_config(jax_legacy_prng_key='allow',
-                 jax_debug_key_reuse=False)
+                 jax_debug_key_reuse=False,
+                 jax_threefry_gpu_kernel_lowering=True)
 class CompatTest(bctu.CompatTestBase):
   def test_dummy(self):
     # Tests the testing mechanism. Let this test run on all platforms
@@ -573,12 +574,11 @@ def func():
     self.run_one_test(func, data)
 
   def test_cuda_threefry2x32(self):
-    # TODO(frostig): remove after 2024-11-01
     def func(x):
       return jax.random.uniform(x, (2, 4), dtype=np.float32)
 
     data = self.load_testdata(cuda_threefry2x32.data_2023_03_15)
-    self.run_one_test(func, data, expect_current_custom_calls=[])
+    self.run_one_test(func, data)
 
   def test_sharding(self):
     # Tests "Sharding", "SPMDShardToFullShape", "SPMDFullToShardShape" on TPU

tests/random_test.py

@@ -385,6 +385,16 @@ def testPRNGValues(self, make_key):
         random.key_data(random.fold_in(make_key(seed), 4)),
         np.array([2285895361,  433833334], dtype='uint32'))
 
+  @jtu.run_on_devices("gpu")
+  def test_threefry_gpu_kernel_lowering(self):
+    f = lambda key: jax.random.uniform(key, (1,))
+    with jax._src.config.threefry_gpu_kernel_lowering(False):
+      hlo_text = jax.jit(f).lower(jax.random.key(17)).as_text()
+      self.assertNotIn("cu_threefry2x32", hlo_text)
+    with jax._src.config.threefry_gpu_kernel_lowering(True):
+      hlo_text = jax.jit(f).lower(jax.random.key(17)).as_text()
+      self.assertIn("cu_threefry2x32", hlo_text)
+
   @parameterized.parameters([{'make_key': ctor} for ctor in KEY_CTORS])
   def test_random_seed_offset(self, make_key):
     k1 = make_key(17)