[TKW] Hoist loop invariant reads (#296)

.github/workflows/ci-tk.yaml

@@ -57,13 +57,20 @@ jobs:
       run: |
         pytest -n 4 --capture=tee-sys -vv ./tests/kernel/wave/
 
-    - name: Run e2e tests on AMD GPU
-      if: "contains(matrix.os, 'amdgpu') && !cancelled()"
+    - name: Run e2e tests on AMD GPU MI300
+      if: "contains(matrix.os, 'mi300') && !cancelled()"
       run: |
         pip install --no-compile -r pytorch-rocm-requirements.txt
         export WAVE_RUN_E2E_TESTS=1
         pytest -n 4 --capture=tee-sys -vv ./tests/kernel/wave/
 
+    - name: Run e2e tests on AMD GPU MI250
+      if: "contains(matrix.os, 'mi250') && !cancelled()"
+      run: |
+        pip install --no-compile -r pytorch-rocm-requirements.txt
+        export WAVE_RUN_E2E_TESTS=1
+        pytest -n 2 --capture=tee-sys -vv ./tests/kernel/wave/
+
     - name: Run LIT tests
       if: ${{ !cancelled() }}
       run: |

iree/turbine/kernel/wave/hoisting.py

@@ -4,6 +4,8 @@
 # See https://llvm.org/LICENSE.txt for license information.
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
+from .constraints import Constraint
+from .utils import get_induction_variable
 from ...support.logging import get_logger
 from iree.turbine.kernel._support.tracing import CapturedTrace
 import torch.fx as fx
@@ -13,25 +15,87 @@
 logger = get_logger("turbine.wave.hoisting")
 
 
-def get_allocs(graph: fx.Graph) -> list[CustomOp]:
-    return [
-        custom_node
-        for node in graph.nodes
-        if isinstance((custom_node := get_custom(node)), Allocate)
-    ]
+def get_hoistable_ops(
+    graph: fx.Graph,
+    captured_vars: list[CustomOp],
+    induction_variable: IndexExpr,
+) -> list[CustomOp]:
+    """
+    Get hoistable ops. Currently only handle allocs and read who doesn't depends on
+    induction variables.
+
+    Note: For codegen to work properly, we'd need to hoist allocs first. This is to avoid
+    using alloc before defined/non-dominating behavior.
+    (e.g hoisting read from global to shared before shared alloc is defined.)
+    """
+    hoistable_allocs = []
+    hoistable_ops = []
+    for node in graph.nodes:
+        custom_node = get_custom(node)
+        if isinstance(custom_node, Allocate):
+            hoistable_allocs.append(custom_node)
+        elif isinstance(custom_node, Read):
+            if custom_node.index is None:
+                continue
+            # Only handle case where memory is captured var.
+            # i.e it has source value from root graph.
+            if not custom_node.memory in captured_vars:
+                continue
+            # Only handle case where we are not writing to the same memory.
+            # Counterproof: we may expect different read if we write to same memory.
+            if any(
+                isinstance(get_custom(mem_user), Write)
+                for mem_user in custom_node.memory.users
+            ):
+                continue
+            # Only hoist Read that is loop invariant.
+            if any(
+                ind.start.has(induction_variable) for ind in custom_node.index.values()
+            ):
+                continue
+            hoistable_ops.append(custom_node)
+        else:
+            continue
+    all_hoistables_ops = hoistable_allocs + hoistable_ops
+    return all_hoistables_ops
+
+
+def remove_unused_captured_vars(reduction: CustomOp, subgraph: fx.Graph):
+    captured_vars = reduction.captured_vars(subgraph)
+    new_implicit_captures = list(reduction.implicit_captures)
+    for captured_idx in reversed(range(len(captured_vars))):
+        if len(captured_vars[captured_idx].users) == 0:
+            get_custom(captured_vars[captured_idx]).erase()
+            new_implicit_captures.pop(captured_idx)
+            reduction.update_arg("implicit_captures", new_implicit_captures)
 
 
-def hoist_allocs(trace: CapturedTrace):
-    """Hoists allocs from reduction subgraphs to outer root graph."""
+def hoist_loop_invariant_ops(trace: CapturedTrace, constraints: list[Constraint]):
+    """Hoists ops that are loop-invariant from reduction subgraphs to outer root graph."""
     root_graph = trace.get_root_graph()
     for node in root_graph.nodes:
         custom_node = get_custom(node)
         match custom_node:
             case Reduction():
                 with root_graph.inserting_before(custom_node.fx_node):
+                    induction_variable = get_induction_variable(
+                        custom_node, constraints
+                    )
                     subgraph = trace.get_subgraph(custom_node.subgraph_name)
-                    allocs = get_allocs(subgraph)
-                    for alloc in allocs:
-                        new_alloc = alloc.copy(new_graph=root_graph)
-                        alloc.replace_all_uses_with(new_alloc)
-                        alloc.erase()
+                    # Capture/root variables from outside the loop.
+                    implicit_captures = custom_node.implicit_captures
+                    # Captured variables from inside the loop.
+                    captured_vars = custom_node.captured_vars(subgraph)
+                    hoistable_ops = get_hoistable_ops(
+                        subgraph, captured_vars, induction_variable
+                    )
+                    for hoistable_op in hoistable_ops:
+                        new_op = hoistable_op.copy(new_graph=root_graph)
+                        hoistable_op.replace_all_uses_with(new_op)
+                        hoistable_op.erase()
+                        if isinstance(hoistable_op, Read):
+                            capture_arg = captured_vars.index(hoistable_op.memory)
+                            new_op.update_arg("memory", implicit_captures[capture_arg])
+                    # Clear/Remove unused captured var to correct codegen. Ops inside
+                    # scf.for will be indexing/loading from the wrong bindings otherwise.
+                    remove_unused_captured_vars(custom_node, subgraph)

iree/turbine/kernel/wave/wave.py

@@ -22,7 +22,7 @@
 from .codegen import WaveEmitter
 from .expansion import expand_graph
 from .promotion import promote_placeholders
-from .hoisting import hoist_allocs
+from .hoisting import hoist_loop_invariant_ops
 from .utils import (
     canonicalize_module,
     compile_and_invoke,
@@ -232,7 +232,6 @@ def _trace_and_get_kernel_signature(
 
         # Promote the placeholders to the appropriate address space.
         promote_placeholders(graph, self.constraints)
-        hoist_allocs(graph)
 
         # Set indices.
         set_node_indices(graph, self.constraints)
@@ -250,6 +249,7 @@ def _trace_and_get_kernel_signature(
         remove_chained_getresult(graph)
 
         # Optimizations.
+        hoist_loop_invariant_ops(graph, self.constraints)
         minimize_global_loads(graph, self.constraints)
 
         # Apply shared memory indexing corrections.

lit_tests/kernel/wave/attention.py

@@ -154,15 +154,15 @@ def repeat(
         output = torch.zeros(shape[0], shape[1], shape[2], dtype=torch.float32)
         print(dynamic_attention_pipelined(q, k, v, output).module_op)
 
-        # CHECK:            func.func @dynamic_attention_pipelined
-        # CHECK-COUNT-4:        {{.*}} = vector.maskedload {{.*}}
+        # CHECK-LABEL:       func.func @dynamic_attention_pipelined
+        # CHECK-COUNT-6:        {{.*}} = vector.maskedload {{.*}}
         # CHECK:                {{.*}} = scf.for
-        # CHECK-COUNT-4:            {{.*}} = vector.maskedload {{.*}}
-        # CHECK-COUNT-13:           {{.*}} = amdgpu.mfma
-        # CHECK-COUNT-2:            {{.*}} = gpu.shuffle xor {{.*}}
-        # CHECK-COUNT-8:            {{.*}} = amdgpu.mfma
-        # CHECK-COUNT-2:            {{.*}} = gpu.shuffle xor {{.*}}
-        # CHECK-COUNT-3:            {{.*}} = amdgpu.mfma
+        # CHECK-COUNT-2:            {{.*}} = vector.maskedload {{.*}}
+        # CHECK-COUNT-14:           {{.*}} = amdgpu.mfma
+        # CHECK-COUNT-1:            {{.*}} = gpu.shuffle xor {{.*}}
+        # CHECK-COUNT-7:            {{.*}} = amdgpu.mfma
+        # CHECK-COUNT-5:            {{.*}} = gpu.shuffle xor {{.*}}
+        # CHECK-COUNT-2:            {{.*}} = amdgpu.mfma
         # CHECK-COUNT-4:            {{.*}} = gpu.shuffle xor {{.*}}
         # CHECK-COUNT-16:       vector.maskedstore {{.*}}
 
@@ -281,13 +281,13 @@ def repeat(
         output = torch.zeros(shape[0], shape[1], shape[2], dtype=torch.float32)
         print(base_attention_pipelined(q, k, v, output).module_op)
 
-        # CHECK:            func.func @base_attention_pipelined
+        # CHECK-LABEL:       func.func @base_attention_pipelined
         # CHECK:                {{.*}} = scf.for
-        # CHECK-COUNT-13:           {{.*}} = amdgpu.mfma
-        # CHECK-COUNT-2:            {{.*}} = gpu.shuffle xor {{.*}}
-        # CHECK-COUNT-8:            {{.*}} = amdgpu.mfma
+        # CHECK-COUNT-14:           {{.*}} = amdgpu.mfma
+        # CHECK-COUNT-1:            {{.*}} = gpu.shuffle xor {{.*}}
+        # CHECK-COUNT-7:            {{.*}} = amdgpu.mfma
         # CHECK-COUNT-2:            {{.*}} = gpu.shuffle xor {{.*}}
-        # CHECK-COUNT-3:            {{.*}} = amdgpu.mfma
+        # CHECK-COUNT-2:            {{.*}} = amdgpu.mfma
         # CHECK-COUNT-4:            {{.*}} = gpu.shuffle xor {{.*}}
 
 
@@ -401,7 +401,7 @@ def repeat(
         output = torch.zeros(shape[0], shape[1], shape[2], dtype=torch.float32)
         print(base_attention_32x32x8(q, k, v, output).module_op)
 
-        # CHECK:            func.func @base_attention_32x32x8
+        # CHECK-LABEL:      func.func @base_attention_32x32x8
         # CHECK:                {{.*}} = scf.for
         # CHECK-COUNT-8:           {{.*}} = amdgpu.mfma
         # CHECK-COUNT-2:            {{.*}} = gpu.shuffle xor {{.*}}
@@ -524,7 +524,7 @@ def repeat(
         output = torch.zeros(shape[0], shape[1], shape[2], dtype=torch.float32)
         print(base_attention(q, k, v, output).module_op)
 
-        # CHECK:            func.func @base_attention
+        # CHECK-LABEL:       func.func @base_attention
         # CHECK:                {{.*}} = scf.for
         # CHECK-COUNT-16:           {{.*}} = amdgpu.mfma
         # CHECK-COUNT-8:            {{.*}} = gpu.shuffle xor {{.*}}

lit_tests/kernel/wave/barriers.py

@@ -12,7 +12,7 @@
 )
 from iree.turbine.kernel.wave.promotion import promote_node, promote_placeholders
 from iree.turbine.kernel.wave.barriers import add_shared_memory_barriers
-from iree.turbine.kernel.wave.hoisting import hoist_allocs
+from iree.turbine.kernel.wave.hoisting import hoist_loop_invariant_ops
 from iree.turbine.kernel.wave.expansion import expand_graph
 from iree.turbine.kernel.wave.type_inference import infer_types
 from iree.turbine.kernel.lang.global_symbols import *
@@ -177,11 +177,11 @@ def test_gemm():
         read_nodes = get_read_nodes(graph)
         for read_node in read_nodes:
             promote_node(read_node, SHARED_ADDRESS_SPACE, constraints)
-        hoist_allocs(trace)
         set_node_indices(trace, constraints)
         expand_graph(trace, constraints)
         set_post_expansion_indices(trace, constraints)
         tweak_index(graph)
+        hoist_loop_invariant_ops(trace, constraints)
         add_shared_memory_barriers(trace)
         print_trace(trace, False)
         # Root graph:

lit_tests/kernel/wave/codegen.py

@@ -1004,7 +1004,7 @@ def test_chained_gemm():
 
     @tkw.wave(constraints)
     def chained_gemm(
-        q: tkl.Memory[B, M, K1, ADDRESS_SPACE, tkl.f16],
+        q: tkl.Memory[B, M, K1, ADDRESS_SPACE_0, tkl.f16],
         k: tkl.Memory[B, K2, K1, ADDRESS_SPACE, tkl.f16],
         v: tkl.Memory[B, N, K2, ADDRESS_SPACE, tkl.f16],
         c: tkl.Memory[B, M, N, ADDRESS_SPACE_0, tkl.f32],
@@ -1051,8 +1051,13 @@ def repeat(
         output = torch.zeros(8, 64, 128, dtype=torch.float32)
         print(chained_gemm(q, k, v, output).module_op)
 
-        # CHECK-LABEL:     func.func @chained_gemm(
+        # CHECK-LABEL:     func.func @chained_gemm
+        # CHECK-SAME:        (%[[ARG0:.*]]: !stream.binding, %{{.+}}: !stream.binding, %{{.+}}: !stream.binding, %{{.+}}: !stream.binding)
+        # CHECK:             %[[ALLOC:.+]] = memref.alloc() : memref<1x32x36xf16, #gpu.address_space<workgroup>>
+        # CHECK:             %[[GLOBAL_0:.+]] = stream.binding.subspan %[[ARG0]]
+        # CHECK-COUNT-4:     vector.load %[[GLOBAL_0]]
         # CHECK:             {{.*}} = scf.for
+        # CHECK-COUNT-4:       {{.*}} = vector.load %[[ALLOC]]
         # CHECK-COUNT-8:       {{.*}} = amdgpu.mfma
         # CHECK-COUNT-4:       {{.*}} = arith.truncf
         # CHECK-COUNT-8:       {{.*}} = amdgpu.mfma
@@ -1131,7 +1136,10 @@ def repeat(
         output = torch.zeros(8, 64, 128, dtype=torch.float32)
         print(chained_gemm_32x32x8(q, k, v, output).module_op)
 
-        # CHECK-LABEL:     func.func @chained_gemm_32x32x8(
+        # CHECK-LABEL:     func.func @chained_gemm_32x32x8
+        # CHECK-SAME:        (%[[ARG0:.*]]: !stream.binding, %{{.+}}: !stream.binding, %{{.+}}: !stream.binding, %{{.+}}: !stream.binding)
+        # CHECK:             %[[GLOBAL_0:.+]] = stream.binding.subspan %[[ARG0]]
+        # CHECK:             %[[GLOBAL_READ_0:.+]] = vector.load %[[GLOBAL_0]]
         # CHECK:             {{.*}} = scf.for
         # CHECK-COUNT-4:       {{.*}} = amdgpu.mfma
         # CHECK-COUNT-1:       {{.*}} = arith.truncf

lit_tests/kernel/wave/index_sequence_analysis.py

@@ -6,7 +6,7 @@
 import iree.turbine.kernel.lang as tkl
 import iree.turbine.kernel.wave as tkw
 from iree.turbine.kernel.wave.promotion import promote_placeholders
-from iree.turbine.kernel.wave.hoisting import hoist_allocs
+from iree.turbine.kernel.wave.hoisting import hoist_loop_invariant_ops
 from iree.turbine.kernel.wave.expansion import expand_graph
 from iree.turbine.kernel.wave.type_inference import infer_types
 from iree.turbine.kernel.lang.global_symbols import *
@@ -87,10 +87,10 @@ def test_gemm():
         IndexingContext.current().finalize()
         infer_types(trace)
         promote_placeholders(trace, constraints)
-        hoist_allocs(trace)
         set_node_indices(trace, constraints)
         expand_graph(trace, constraints)
         set_post_expansion_indices(trace, constraints)
+        hoist_loop_invariant_ops(trace, constraints)
         minimize_global_loads(trace, constraints)
         apply_shared_memory_indexing_corrections(trace, constraints)
         partition_strided_operators(trace, constraints)

lit_tests/kernel/wave/minimize_global_loads.py

@@ -7,7 +7,7 @@
 import iree.turbine.kernel.lang as tkl
 import iree.turbine.kernel.wave as tkw
 from iree.turbine.kernel.wave.promotion import promote_placeholders
-from iree.turbine.kernel.wave.hoisting import hoist_allocs
+from iree.turbine.kernel.wave.hoisting import hoist_loop_invariant_ops
 from iree.turbine.kernel.wave.barriers import add_shared_memory_barriers
 from iree.turbine.kernel.wave.expansion import expand_graph
 from iree.turbine.kernel.wave.type_inference import infer_types
@@ -90,12 +90,12 @@ def test_gemm():
         IndexingContext.current().finalize()
         infer_types(trace)
         promote_placeholders(trace, constraints)
-        hoist_allocs(trace)
         set_node_indices(trace, constraints)
         expand_graph(trace, constraints)
         set_post_expansion_indices(trace, constraints)
         if visualize:
             visualize_graph(trace.get_subgraph("region_0"), "before.png")
+        hoist_loop_invariant_ops(trace, constraints)
         minimize_global_loads(trace, constraints)
         apply_shared_memory_indexing_corrections(trace, constraints)
         if visualize:

lit_tests/kernel/wave/promotion.py

@@ -6,7 +6,7 @@
 import iree.turbine.kernel.lang as tkl
 import iree.turbine.kernel.wave as tkw
 from iree.turbine.kernel.wave.promotion import promote_node, promote_placeholders
-from iree.turbine.kernel.wave.hoisting import hoist_allocs
+from iree.turbine.kernel.wave.hoisting import hoist_loop_invariant_ops
 from iree.turbine.kernel.wave.type_inference import infer_types
 from iree.turbine.kernel.lang.global_symbols import *
 from iree.turbine.kernel._support.tracing import CapturedTrace
@@ -35,6 +35,9 @@ def get_read_nodes(graph: fx.Graph) -> list[CustomOp]:
 ADDRESS_SPACE_0 = tkl.sym.ADDRESS_SPACE_0
 ADDRESS_SPACE_1 = tkl.sym.ADDRESS_SPACE_1
 
+# Induction variable for dimension K
+ARGK = tkl.sym.ARGK
+
 
 @tkw.wave_trace_only()
 def read_write_same_size(
@@ -160,6 +163,7 @@ def test_gemm():
     constraints: list[tkw.Constraint] = [tkw.WorkgroupConstraint(M, BLOCK_M, 0)]
     constraints += [tkw.WorkgroupConstraint(N, BLOCK_N, 1)]
     constraints += [tkw.WorkgroupConstraint(K, BLOCK_K, 2)]
+    constraints += [tkw.TilingConstraint(K, BLOCK_K, ARGK)]
     constraints += [
         tkw.HardwareConstraint(threads_per_wave=64, waves_per_block=(1, 1, 1))
     ]
@@ -177,7 +181,7 @@ def test_gemm():
         infer_types(trace)
         for read_node in read_nodes:
             promote_node(read_node, SHARED_ADDRESS_SPACE, constraints)
-        hoist_allocs(trace)
+        hoist_loop_invariant_ops(trace, constraints)
         print_trace(trace, False)
         # Root graph:
         # CHECK: %a

lit_tests/kernel/wave/scheduling.py

@@ -6,7 +6,7 @@
 import iree.turbine.kernel.lang as tkl
 import iree.turbine.kernel.wave as tkw
 from iree.turbine.kernel.wave.promotion import promote_placeholders
-from iree.turbine.kernel.wave.hoisting import hoist_allocs
+from iree.turbine.kernel.wave.hoisting import hoist_loop_invariant_ops
 from iree.turbine.kernel.wave.expansion import expand_graph
 from iree.turbine.kernel.wave.type_inference import infer_types
 from iree.turbine.kernel.lang.global_symbols import *
@@ -99,10 +99,10 @@ def test_gemm_pipelined():
         IndexingContext.current().finalize()
         infer_types(trace)
         promote_placeholders(trace, constraints)
-        hoist_allocs(trace)
         set_node_indices(trace, constraints)
         expand_graph(trace, constraints)
         set_post_expansion_indices(trace, constraints)
+        hoist_loop_invariant_ops(trace, constraints)
         minimize_global_loads(trace, constraints)
         apply_shared_memory_indexing_corrections(trace, constraints)
         schedule_graph(trace, constraints, True)

tests/kernel/wave/scheduling_test.py

@@ -28,7 +28,7 @@
 from iree.turbine.kernel._support.tracing import CapturedTrace
 from iree.turbine.kernel._support.indexing import IndexingContext
 from iree.turbine.kernel.wave.promotion import promote_placeholders
-from iree.turbine.kernel.wave.hoisting import hoist_allocs
+from iree.turbine.kernel.wave.hoisting import hoist_loop_invariant_ops
 from iree.turbine.kernel.wave.expansion import expand_graph
 from iree.turbine.kernel.wave.type_inference import infer_types
 from iree.turbine.kernel.wave.minimize_global_loads import minimize_global_loads
@@ -285,7 +285,7 @@ def repeat(acc: tkl.Register[M, N, tkl.f32]) -> tkl.Register[M, N, tkl.f32]:
             IndexingContext.current().finalize()
             infer_types(trace)
             promote_placeholders(trace, constraints)
-            hoist_allocs(trace)
+            hoist_loop_invariant_ops(trace, constraints)
             set_node_indices(trace, constraints)
             expand_graph(trace, constraints)
             set_post_expansion_indices(trace, constraints)