Add epilogue subtiling

stack-info: PR: #948, branch: PaulZhang12/stack/14

Author: PaulZhang12

helion/_compiler/compile_environment.py

@@ -99,6 +99,7 @@ def __init__(self, device: torch.device, settings: Settings) -> None:
         self.device_load_count = (
             0  # Track number of loads in all device code for eviction policy tuning
         )
+        self.device_store_count = 0  # Track number of stores for subtiling
 
     def add_kernel_tensor_size(self, sizes: Sequence[int | torch.SymInt]) -> None:
         from .device_function import contains_only_block_size_symbols

helion/_compiler/device_function.py

@@ -250,6 +250,9 @@ def __init__(self, name: str, config: Config, codegen: GenerateAST) -> None:
         self.rng_seed_count = 0
         self.device_load_index = 0  # Track which load in device code we're generating (for eviction policy tuning)
         # Name of the RNG seed buffer parameter in kernel signature
+        self.device_store_index = (
+            0  # Track which store in device code we're generating (for subtiling)
+        )
         self.rng_seed_buffer_param_name = None
 
     def has_rng_ops(self) -> bool:
@@ -420,9 +423,15 @@ def tensor_arg(
     def tensor_descriptor_arg(
         self, fake_value: torch.Tensor, block_size: list[int | torch.SymInt]
     ) -> TensorDescriptorArg:
+        import re
+
         host_function = HostFunction.current()
         block_size_expr = ", ".join(map(self.literal_expr, block_size))
+        pattern = r"triton_helpers\.div_floor_integer\(([^,]+),\s*(\d+)\)"
+        replacement = r"\1 // \2"
+        block_size_expr = re.sub(pattern, replacement, block_size_expr)
         key = (fake_value, block_size_expr)
+
         if key not in self._tensor_descriptor_args:
             origin = host_function.tensor_to_origin[fake_value]
             desc_name = self.new_var(origin.suggest_var_name() + "_desc")

helion/_compiler/device_ir.py

@@ -1076,7 +1076,7 @@ def visit_For(self, node: ast.For) -> None:
             self.generic_visit(node)
 
 
-def _count_device_loads(device_ir: DeviceIR) -> int:
+def _count_device_loads_and_stores(device_ir: DeviceIR) -> int:
     """Count the number of load operations in all device code for eviction policy tuning."""
     from ..language import memory_ops
 
@@ -1087,26 +1087,29 @@ def _count_device_loads(device_ir: DeviceIR) -> int:
         if info.new_graph_id is not None
     }
 
-    load_count = 0
+    load_count, store_count = 0, 0
     # Walk all graphs except rolled duplicates
     for graph_info in device_ir.graphs:
         if graph_info.graph_id in rolled_graph_ids:
             continue
 
         for node in graph_info.graph.nodes:
             # Check if this is a load operation
-            if node.op == "call_function" and node.target is memory_ops.load:
-                # Only count loads without explicit eviction policy
-                # (user can still specify eviction_policy to override tuning)
-                # Check kwargs first, then check if 4th arg (eviction_policy) is None
-                eviction_policy_arg = node.kwargs.get("eviction_policy")
-                if eviction_policy_arg is None:
-                    # Check if eviction_policy was passed as positional arg (index 3)
-                    if len(node.args) >= 4:
-                        eviction_policy_arg = node.args[3]
+            if node.op == "call_function":
+                if node.target is memory_ops.load:
+                    # Only count loads without explicit eviction policy
+                    # (user can still specify eviction_policy to override tuning)
+                    # Check kwargs first, then check if 4th arg (eviction_policy) is None
+                    eviction_policy_arg = node.kwargs.get("eviction_policy")
                     if eviction_policy_arg is None:
-                        load_count += 1
-    return load_count
+                        # Check if eviction_policy was passed as positional arg (index 3)
+                        if len(node.args) >= 4:
+                            eviction_policy_arg = node.args[3]
+                        if eviction_policy_arg is None:
+                            load_count += 1
+                elif node.target is memory_ops.store:
+                    store_count += 1
+    return load_count, store_count
 
 
 def _register_eviction_policy_tunable(load_count: int) -> None:
@@ -1125,6 +1128,24 @@ def _register_eviction_policy_tunable(load_count: int) -> None:
     env.device_load_count = load_count
 
 
+def _register_epilogue_subtile_tunable(store_count: int) -> None:
+    """Register the epilogue subtile tunable for all device stores."""
+    if store_count == 0:
+        return
+
+    from ..autotuner.config_fragment import EnumFragment
+    from ..autotuner.config_fragment import ListOf
+    from ..autotuner.config_spec import VALID_EPILOGUE_SUBTILE_SIZES
+
+    env = CompileEnvironment.current()
+    # Register a tunable for epilogue subtile for all device stores
+    fragment = ListOf(
+        EnumFragment(choices=VALID_EPILOGUE_SUBTILE_SIZES), length=store_count
+    )
+    env.config_spec.epilogue_subtiling = fragment
+    env.device_store_count = store_count
+
+
 def lower_to_device_ir(func: HostFunction) -> DeviceIR:
     device_ir = DeviceIR()
     with func, device_ir, compile_lock:
@@ -1148,9 +1169,13 @@ def lower_to_device_ir(func: HostFunction) -> DeviceIR:
             CompileEnvironment.current().config_spec.disallow_pid_type("xyz")
 
         # Count all device loads and register eviction policy tunable
-        load_count = _count_device_loads(device_ir)
+        load_count, store_count = _count_device_loads_and_stores(device_ir)
         _register_eviction_policy_tunable(load_count)
 
+        # Epilogue subtiling only for Blackwell
+        if torch.cuda.get_device_capability() >= (10, 0):
+            _register_epilogue_subtile_tunable(store_count)
+
         return device_ir
 
 

helion/_compiler/indexing_strategy.py

@@ -15,6 +15,7 @@
 from .. import exc
 from .._compat import get_tensor_descriptor_fn_name
 from .ast_extension import expr_from_string
+from .ast_extension import statement_from_string
 from .compile_environment import CompileEnvironment
 from .device_function import DeviceFunction
 from .host_function import HostFunction
@@ -353,7 +354,6 @@ def codegen_load(
             )
         assert extra_mask is None
         indexing = BlockedSubscriptIndexing.create(state, fake_tensor, subscript)
-
         # Load from tensor descriptor with permuted offsets
         load_expr = expr_from_string(
             f"{indexing.tensor_descriptor(state)}.load({indexing.offsets_str_permuted(state)})"
@@ -383,10 +383,12 @@ def codegen_store(
             )
         assert extra_mask is None
         indexing = BlockedSubscriptIndexing.create(state, fake_tensor, subscript)
+        store_value = indexing.reshape_store(state, value)
 
+        config = DeviceFunction.current().config
+        epilogue_subtiles = state.config.epilogue_subtiling
         # Apply permutation to the value being stored if needed
         desc_arg = indexing.tensor_descriptor_arg(state)
-        store_value = indexing.reshape_store(state, value)
 
         if desc_arg.permutation is not None:
             # Apply permutation to the value
@@ -395,11 +397,204 @@ def codegen_store(
                 store_val=store_value,
             )
 
+        if (idx := state.device_function.device_store_index) < len(epilogue_subtiles):
+            subtile_split = epilogue_subtiles[idx]
+            state.device_function.device_store_index += 1
+
+            # Check if we should fuse a pointwise operation into the epilogue store
+            fused_pointwise_node = self._get_fusable_pointwise_node(state)
+
+            subtile_codegen = self._codegen_epilogue_subtile_store(
+                state,
+                fake_tensor,
+                indexing,
+                store_value,
+                subtile_split,
+                config,
+                fused_pointwise_node,
+            )
+            if subtile_codegen is not None:
+                return subtile_codegen
+
         return expr_from_string(
             f"{indexing.tensor_descriptor(state)}.store({indexing.offsets_str_permuted(state)}, {{value}})",
             value=store_value,
         )
 
+    def _get_fusable_pointwise_node(self, state: CodegenState) -> torch.fx.Node | None:
+        """Find a pointwise node feeding into this store that can be fused.
+
+        Returns the pointwise FX node if found, None otherwise.
+        """
+        if state.fx_node is None:
+            return None
+
+        # Get the value being stored (3rd argument to store)
+        if len(state.fx_node.args) < 3:
+            return None
+
+        value_node = state.fx_node.args[2]
+        if not isinstance(value_node, torch.fx.Node):
+            return None
+
+        # Check if this is a pointwise node
+        from .inductor_lowering import PointwiseLowering
+
+        lowering = value_node.meta.get("lowering")
+        if not isinstance(lowering, PointwiseLowering):
+            return None
+
+        # Check if this node only has one user (the store)
+        if len(list(value_node.users)) != 1:
+            return None
+
+        return value_node
+
+    def _apply_pointwise_to_subtile(
+        self, state: CodegenState, pointwise_node: torch.fx.Node, subtile_value: ast.AST
+    ) -> ast.AST:
+        """Apply a pointwise operation to a subtile value.
+
+        Args:
+            state: The codegen state
+            pointwise_node: The FX node representing the pointwise operation
+            subtile_value: The AST for the subtile value to apply the operation to
+
+        Returns:
+            AST for the result after applying the pointwise operation
+        """
+        from torch._inductor import ir
+
+        from .inductor_lowering import PointwiseLowering
+        from .inductor_lowering import install_inductor_kernel_handlers
+
+        lowering = pointwise_node.meta["lowering"]
+        assert isinstance(lowering, PointwiseLowering)
+
+        # Get the pointwise buffer
+        buffer = lowering.buffer
+        assert isinstance(buffer.data, ir.Pointwise)
+
+        # Create a temporary variable for the subtile
+        codegen = state.codegen
+        subtile_var = codegen.lift(subtile_value, prefix="subtile")
+
+        # Set up the inductor kernel handlers with the subtile as input
+        with install_inductor_kernel_handlers(
+            codegen, {lowering.input_names[0]: subtile_var}
+        ):
+            # Generate the pointwise operation
+            indices = [sympy.Symbol(f"i{n}") for n in range(len(buffer.data.ranges))]
+            from .inductor_lowering import _unpack_opsvalue
+
+            result_name = _unpack_opsvalue(buffer.data.inner_fn(indices))
+            return expr_from_string(result_name)
+
+    def _codegen_epilogue_subtile_store(
+        self,
+        state: CodegenState,
+        fake_tensor: torch.Tensor,
+        indexing: BlockedSubscriptIndexing,
+        store_value: ast.AST,
+        subtile_split: int,
+        config: Config,
+        fused_pointwise_node: torch.fx.Node | None = None,
+    ) -> ast.AST | None:
+        # Currently support 2D tiles without permutations
+        if (
+            len(indexing.block_shape) != 2
+            or len(indexing.offsets) != 2
+            or subtile_split == 0
+        ):
+            return None
+
+        env = CompileEnvironment.current()
+        block_m, block_n = indexing.block_shape
+        try:
+            block_n_hint = env.size_hint(block_n)
+            block_idx = env.get_block_id(block_n)
+            block_size = env.block_sizes[block_idx].from_config(config)
+        except Exception:
+            return None
+
+        if block_n_hint % 2 != 0 or block_size <= 16:
+            return None
+
+        device_fn = state.device_function
+        codegen = state.codegen
+
+        block_m_str = device_fn.literal_expr(block_m)
+        block_n_str = device_fn.literal_expr(block_n)
+        indexing.block_shape[1] //= subtile_split
+
+        # TODO(PaulZhang12): Support more epilogue subtile configs besides 2
+        block_n_half_str = f"({block_n_str} // {subtile_split})"
+
+        # If we have a fused pointwise operation, mark it to skip normal codegen
+        # and get its input value instead
+        if fused_pointwise_node is not None:
+            fused_pointwise_node.meta["fused_into_store"] = True
+
+        # Lift the store value into a temporary variable for reuse
+        acc_var = codegen.lift(store_value, prefix="acc")
+
+        reshape_expr = expr_from_string(
+            "tl.reshape({acc}, [{dim_m}, 2, {dim_half}]).permute(0, 2, 1)",
+            acc=acc_var,
+            dim_m=expr_from_string(block_m_str),
+            dim_half=expr_from_string(block_n_half_str),
+        )
+        reshape_var = codegen.lift(reshape_expr, prefix="acc")
+
+        acc0_name = codegen.tmpvar(prefix="acc")
+        acc1_name = codegen.tmpvar(prefix="acc")
+        codegen.add_statement(
+            statement_from_string(
+                f"{acc0_name}, {acc1_name} = tl.split({{acc}})",
+                acc=reshape_var,
+            )
+        )
+
+        # Now apply the pointwise operation per-subtile if we have one
+        if fused_pointwise_node is not None:
+            acc0 = self._apply_pointwise_to_subtile(
+                state, fused_pointwise_node, expr_from_string(acc0_name)
+            )
+            acc1 = self._apply_pointwise_to_subtile(
+                state, fused_pointwise_node, expr_from_string(acc1_name)
+            )
+        else:
+            acc0 = expr_from_string(acc0_name)
+            acc1 = expr_from_string(acc1_name)
+
+        desc_name = indexing.tensor_descriptor(state)
+        offset0 = expr_from_string(indexing.offsets[0])
+        offset1 = expr_from_string(indexing.offsets[1])
+
+        # First subtile store
+        codegen.add_statement(
+            statement_from_string(
+                f"{desc_name}.store([{{off0}}, {{off1}}], {{value}})",
+                off0=offset0,
+                off1=offset1,
+                value=acc0,
+            )
+        )
+
+        offset1_shifted = expr_from_string(
+            "({offset} + {half})",
+            offset=expr_from_string(indexing.offsets[1]),
+            half=expr_from_string(block_n_half_str),
+        )
+
+        # Emit second subtile store as the expression returned to the caller
+        return expr_from_string(
+            f"{desc_name}.store([{{off0}}, {{off1}}], {{value}})",
+            off0=offset0,
+            off1=offset1_shifted,
+            value=acc1,
+        )
+
 
 class StackIndexingStrategy:
     """