[Pallas TPU] Generalize while_loop lowering in Pallas -> Mosaic.

The existing lowering path supports only while_loops which can be converted to fori_loop.
That path makes it significantly easier to optimize and unroll, but cannot support a large class of interesting loop formulations.

This patch draws from the Pallas -> Triton while_loop lowering rule to support such loops in Pallas.
Matching is still performed against fori_loop, to lower via that mechanism if possible -- as it is likely more straightforwardly optimizable compared to general "while".

PiperOrigin-RevId: 626089180

Author: jax authors

jax/_src/pallas/mosaic/lowering.py

@@ -1833,28 +1833,23 @@ def _scan_lowering_rule(
 skip_mlir_conversions.add(lax.scan_p)
 
 
-def _while_lowering_rule(
+def _lower_while_via_fori(
     ctx: LoweringRuleContext,
     *args,
+    fori_jaxpr,
     cond_nconsts,
     cond_jaxpr,
     body_nconsts,
     body_jaxpr,
 ):
-  jaxpr, err = pallas_utils.pattern_match_while_to_fori_loop(
-      cond_jaxpr, cond_nconsts, body_jaxpr, body_nconsts
-  )
-  if jaxpr is None:
-    raise NotImplementedError(err)
-
   _, body_consts, carry = split_list(args, [cond_nconsts, body_nconsts])
   (lb, ub), args = carry[:2], carry[2:]
   for_out = _lower_jaxpr_to_for_loop(
       ctx.replace(
           block_shapes=ctx.block_shapes[: body_nconsts + 1]
           + ctx.block_shapes[body_nconsts + 2 :],
       ),
-      jaxpr,
+      fori_jaxpr,
       lb,
       arith.subi(ub, lb),
       body_consts,
@@ -1865,6 +1860,84 @@ def _while_lowering_rule(
   return [ub, ub, *for_out]
 
 
+def _while_lowering_rule(
+    ctx: LoweringRuleContext,
+    *args,
+    cond_nconsts,
+    cond_jaxpr,
+    body_nconsts,
+    body_jaxpr,
+):
+  # First try to lower via a simpler fori loop, which may optimize better.
+  fori_jaxpr, err = pallas_utils.pattern_match_while_to_fori_loop(
+      cond_jaxpr, cond_nconsts, body_jaxpr, body_nconsts
+  )
+  if fori_jaxpr is not None:
+    return _lower_while_via_fori(
+        ctx,
+        *args,
+        fori_jaxpr=fori_jaxpr,
+        cond_nconsts=cond_nconsts,
+        cond_jaxpr=cond_jaxpr,
+        body_nconsts=body_nconsts,
+        body_jaxpr=body_jaxpr,
+    )
+
+  # If we fail conversion to fori, fallback to an ordinary while loop.
+  cond_consts, body_consts, carry = split_list(
+      args, [cond_nconsts, body_nconsts]
+  )
+  cond_const_block_shapes, body_const_block_shapes, carry_block_shapes = (
+      split_list(ctx.block_shapes, [cond_nconsts, body_nconsts])
+  )
+  cond_const_types = [a.type for a in cond_consts]
+  body_const_types = [a.type for a in body_consts]
+  carry_types = [a.type for a in carry]
+  all_types = [*cond_const_types, *body_const_types, *carry_types]
+  while_op = scf.WhileOp(all_types, args)
+
+  before_block = while_op.before.blocks.append(*all_types)
+  cond_consts_, _, carry_ = split_list(
+      before_block.arguments,
+      [cond_nconsts, body_nconsts],
+  )
+  cond_args = [*cond_consts_, *carry_]
+  with ir.InsertionPoint.at_block_begin(before_block):
+    [cond] = jaxpr_subcomp(
+        ctx.lowering_context.replace(
+            block_shapes=[*cond_const_block_shapes, *carry_block_shapes]
+        ),
+        cond_jaxpr.jaxpr,
+        *cond_args,
+    )
+    scf.condition(cond, before_block.arguments)
+
+  after_block = while_op.after.blocks.append(*all_types)
+  cond_consts_, body_consts_, carry_ = split_list(
+      after_block.arguments,
+      [cond_nconsts, body_nconsts],
+  )
+  all_args = [*cond_consts_, *body_consts_, *carry_]
+  cond_const_args, body_const_args, carry_args = split_list(
+      all_args, [cond_nconsts, body_nconsts]
+  )
+  with ir.InsertionPoint.at_block_begin(after_block):
+    loop_out = jaxpr_subcomp(
+        ctx.lowering_context.replace(
+            block_shapes=[*body_const_block_shapes, *carry_block_shapes],
+        ),
+        body_jaxpr.jaxpr,
+        *body_const_args,
+        *carry_args,
+    )
+    all_handles = [*cond_const_args, *body_const_args, *loop_out]
+    if all_handles:
+      scf.yield_(all_handles)
+
+  all_out = list(while_op.results_)
+  return all_out[cond_nconsts + body_nconsts :]
+
+
 lowering_rules[lax.while_p] = _while_lowering_rule
 
 def _cond_lowering_rule(ctx: LoweringRuleContext, *args, branches, linear):

tests/pallas/pallas_call_tpu_test.py

@@ -1751,6 +1751,199 @@ def body(i, _):
     )(*(jnp.array([[x]]) for x in (2, 6)))
     np.testing.assert_array_equal(r, 4)
 
+  def test_non_range_while_loop(self):
+    """Tests lowering of a while_loop which cannot reduce to a fori_loop."""
+
+    def kernel(x_ref, r_ref):
+      @pl.when(pl.program_id(0) == 0)
+      def _():
+        pl.store(r_ref, (0, 0), 0)
+
+      def cond(state):
+        i, s = state
+        return jnp.logical_and(i < 1024, s < 1024)
+
+      def body(state):
+        i, s = state
+        sl = sl = jax.lax.div(i, 128)
+        l = jax.lax.rem(i, 128)
+        v = pl.load(x_ref, (0, sl, l))
+        return i + 1, s + v
+
+      i = jnp.int32(0)
+      s = pl.load(r_ref, (0, 0))
+
+      i, s = jax.lax.while_loop(cond, body, (i, s))
+      pl.store(r_ref, (0, 0), s)
+
+    x = jnp.arange(4096)
+    x = jnp.reshape(x, [4, 8, 128])
+
+    r = pl.pallas_call(
+        kernel,
+        grid=(4,),
+        out_specs=pl.BlockSpec(block_shape=(1, 1), memory_space=pltpu.SMEM),
+        out_shape=jax.ShapeDtypeStruct([1, 1], jnp.int32),
+        in_specs=[
+            pl.BlockSpec(
+                lambda i: (i, 0, 0),
+                block_shape=(1, 8, 128),
+                memory_space=pltpu.SMEM,
+            )
+        ],
+    )(x)
+    np.testing.assert_array_equal(r, [[1035]])
+
+  def test_vector_carry_while_loop(self):
+    """Tests lowering of a while_loop which carries a vector quantity."""
+
+    def kernel(x_ref, r_ref):
+
+      def cond(v):
+        return v[0, 0] < 16
+
+      def body(v):
+        return v * 2
+
+      r_ref[:] = jax.lax.while_loop(cond, body, x_ref[:])
+
+    x = jnp.full((8, 128), 3, dtype=jnp.int32)
+    fn = pl.pallas_call(
+        kernel,
+        grid=(1,),
+        in_specs=[pl.BlockSpec(lambda i: (0, 0), (8, 128))],
+        out_specs=pl.BlockSpec(lambda i: (0, 0), (8, 128)),
+        out_shape=jax.ShapeDtypeStruct((8, 128), jnp.int32),
+    )
+    r = fn(x)
+    reduced = jnp.sum(r)
+    # 3 -> 6 -> 12 -> 24
+    np.testing.assert_array_equal(reduced, 1024 * 24)
+
+  @parameterized.named_parameters(
+      ('1x128', (1, 128)),
+      ('2x128', (2, 128)),
+      ('4x128', (4, 128)),
+      ('8x128', (8, 128)),
+      ('8x256', (8, 256)),
+  )
+  def test_while_loop_carry_memref(self, shape):
+    """Tests a while loop carrying a memref."""
+
+    # TODO(hmckenzie): Investigate further why this occurs.
+    if shape == (1, 128):
+      self.skipTest('memref<1x128> inexplicably doubles to 2x128.')
+
+    def kernel(out_ref, bound):
+      def cond(i):
+        return i < bound
+
+      def body(i):
+        out_ref[0, i] = 2
+        return i + 1
+
+      jax.lax.while_loop(cond, body, 0)
+
+    x = jnp.asarray([1, 1, 1, 1])
+    x = jnp.asarray(x)
+    x = jnp.pad(x, (0, np.prod(shape) - 4), constant_values=0)
+    x = jnp.reshape(x, shape)
+    kernel = partial(kernel, bound=x.shape[1])
+
+    fn = pl.pallas_call(
+        kernel,
+        grid=(1,),
+        out_specs=[
+            pl.BlockSpec(
+                lambda i: (0, 0), block_shape=shape, memory_space=pltpu.SMEM
+            ),
+        ],
+        out_shape=[
+            jax.ShapeDtypeStruct(shape, jnp.int32),
+        ],
+    )
+    y = fn()[0]
+    np.testing.assert_array_equal(y[0, 0], 2)
+    np.testing.assert_array_equal(y[0, 1], 2)
+    np.testing.assert_array_equal(y[0, 2], 2)
+    np.testing.assert_array_equal(y[0, 3], 2)
+
+  def test_nested_while_loop(self):
+    """Tests lowering a nested while_loop."""
+
+    def kernel(in_key_ref, out_segment_count, out_size_ref, key_count):
+      # Compute the length of contiguous segments of keys.
+
+      def inner_cond(carry):
+        i, prev_key = carry
+        sl = sl = jax.lax.div(i, 128)
+        l = jax.lax.rem(i, 128)
+        key = jax.lax.cond(
+            i < key_count, lambda i: in_key_ref[sl, l], lambda i: -1, i
+        )
+        return jnp.logical_and(i < key_count, key == prev_key)
+
+      def inner_body(carry):
+        i, key = carry
+        return i + 1, key
+
+      def outer_cond(carry):
+        i, _ = carry
+        return i < key_count
+
+      def outer_body(carry):
+        i, next_out_idx = carry
+        sl = sl = jax.lax.div(i, 128)
+        l = jax.lax.rem(i, 128)
+        key = in_key_ref[sl, l]
+        end, _ = jax.lax.while_loop(inner_cond, inner_body, (i + 1, key))
+
+        sl = sl = jax.lax.div(next_out_idx, 128)
+        l = jax.lax.rem(next_out_idx, 128)
+        out_size_ref[sl, l] = end - i
+        return end, next_out_idx + 1
+
+      _, count = jax.lax.while_loop(outer_cond, outer_body, (0, 0))
+      out_segment_count[0, 0] = count
+
+    keys = [4, 4, 4, 3, 2, 2, 7, 7, 7, 7]
+    keys = jnp.asarray(keys)
+    real_keys = keys.shape[0]
+    key_count = 1024
+    keys = jnp.pad(keys, (0, key_count - real_keys), constant_values=32768)
+    keys = jnp.reshape(keys, (8, 128))
+    kernel_fn = partial(kernel, key_count=key_count)
+
+    fn = pl.pallas_call(
+        kernel_fn,
+        grid=(1,),
+        in_specs=[
+            # keys.
+            pl.BlockSpec(
+                lambda i: (0, 0),
+                block_shape=(8, 128),
+                memory_space=pltpu.SMEM,
+            ),
+        ],
+        out_specs=[
+            # Segments found.
+            pl.BlockSpec(block_shape=(1, 1), memory_space=pltpu.SMEM),
+            # Segment sizes.
+            pl.BlockSpec(block_shape=(8, 128), memory_space=pltpu.SMEM),
+        ],
+        out_shape=[
+            jax.ShapeDtypeStruct((1, 1), jnp.int32),
+            jax.ShapeDtypeStruct((8, 128), jnp.int32),
+        ],
+    )
+    count, sizes = fn(keys)
+    np.testing.assert_equal(count[0, 0], jnp.asarray(5))
+    np.testing.assert_equal(sizes[0, 0], jnp.asarray(3))
+    np.testing.assert_equal(sizes[0, 1], jnp.asarray(1))
+    np.testing.assert_equal(sizes[0, 2], jnp.asarray(2))
+    np.testing.assert_equal(sizes[0, 3], jnp.asarray(4))
+    np.testing.assert_equal(sizes[0, 4], jnp.asarray(key_count - real_keys))
+
 
 class PallasCallPipelineTest(parameterized.TestCase):