[mlir][GPU] Improve constant sinking in kernel outlining

[herhut]

Summary

The previous implementation did not support sinking simple expressions. In particular,
it is often beneficial to sink dim operations.

[rriddle]

mlir/lib/Dialect/GPU/Transforms/KernelOutlining.cpp

76

Drop uses of mlir::, there are several in this revision.

83

nit: Spell out auto here.

[herhut]

Comments

[ftynse]

mlir/lib/Dialect/GPU/Transforms/KernelOutlining.cpp

61

This function deserves a doc comment

[herhut]

More comments.

mlir/lib/Dialect/GPU/Transforms/KernelOutlining.cpp

 #include "PassDetail.h"
 #include "mlir/Dialect/GPU/GPUDialect.h"
 #include "mlir/Dialect/GPU/Passes.h"
 #include "mlir/Dialect/GPU/Utils.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 #include "mlir/IR/BlockAndValueMapping.h"
 #include "mlir/IR/Builders.h"
 #include "mlir/IR/SymbolTable.h"
+#include "mlir/Support/LLVM.h"
 #include "mlir/Transforms/RegionUtils.h"
 
 using namespace mlir;
 
 template <typename OpTy>
 static void createForAllDimensions(OpBuilder &builder, Location loc,
                                    SmallVectorImpl<Value> &values) {
   for (StringRef dim : {"x", "y", "z"}) {
static void injectGpuIndexOperations(Location loc, Region &launchFuncOpBody,
   createForAllDimensions<gpu::BlockDimOp>(builder, loc, indexOps);
   // Replace the leading 12 function args with the respective thread/block index
   // operations. Iterate backwards since args are erased and indices change.
   for (auto indexOp : enumerate(indexOps))
     map.map(firstBlock.getArgument(indexOp.index()), indexOp.value());
 }
 
 static bool isSinkingBeneficiary(Operation *op) {
-  return isa<ConstantOp, DimOp>(op);
+  return isa<ConstantOp, DimOp, SelectOp, CmpIOp>(op);
+}
+
+static bool extractBeneficiaryOps(Operation *op,
+                                  llvm::SetVector<Operation *> &beneficiaryOps,
+                                  llvm::SetVector<Value> &availableValues) {
+  if (beneficiaryOps.count(op))
+    return true;
+
+  if (!isSinkingBeneficiary(op))
+    return false;
+
+  for (Value operand : op->getOperands()) {
+    // It is already visisble in the kernel, keep going.
+    if (availableValues.count(operand))
+      continue;
+    // Else check whether it can be made available via sinking.
+    Operation *definingOp = operand.getDefiningOp();
+    if (!definingOp ||
+        !extractBeneficiaryOps(definingOp, beneficiaryOps, availableValues))
+      return false;
+  }
+  // We will sink the operation, mark its results as now available.
+  beneficiaryOps.insert(op);
+  for (Value result : op->getResults())
+    availableValues.insert(result);
+  return true;
 }
 
 LogicalResult mlir::sinkOperationsIntoLaunchOp(gpu::LaunchOp launchOp) {
   Region &launchOpBody = launchOp.body();
 
   // Identify uses from values defined outside of the scope of the launch
   // operation.
   llvm::SetVector<Value> sinkCandidates;
   getUsedValuesDefinedAbove(launchOpBody, sinkCandidates);
 
-  llvm::SetVector<Value> sunkValues;
-  llvm::SetVector<Operation *> sunkOperations;
-  for (Value operand : sinkCandidates) {
+  SmallVector<Value, 4> worklist(sinkCandidates.begin(), sinkCandidates.end());
+  llvm::SetVector<Operation *> toBeSunk;
+  for (Value operand : worklist) {
     Operation *operandOp = operand.getDefiningOp();
-    if (!operandOp || !isSinkingBeneficiary(operandOp))
+    if (!operandOp)
       continue;
-    // Only sink operations that do not create new sinkCandidates.
-    if (!llvm::all_of(operandOp->getOperands(), [&sinkCandidates](Value value) {
-          return sinkCandidates.count(value);
-        }))
-      continue;
-    sunkValues.insert(operand);
-    sunkOperations.insert(operandOp);
+    extractBeneficiaryOps(operandOp, toBeSunk, sinkCandidates);
   }
 
   // Insert operations so that the defs get cloned before uses.
   BlockAndValueMapping map;
   OpBuilder builder(launchOpBody);
-  DenseSet<Operation *> processed;
-  SmallVector<Operation *, 2> clonedOps;
-  while (processed.size() != sunkOperations.size()) {
-    auto startSize = processed.size();
-    for (Operation *sunkOperation : sunkOperations) {
-      if (processed.count(sunkOperation))
-        continue;
-
-      // Operation cant be cloned yet if any of its operands is also being sunk,
-      // but isnt cloned yet.
-      if (llvm::any_of(
-              sunkOperation->getOperands(), [&sunkValues, &map](Value value) {
-                return sunkValues.count(value) && !map.lookupOrNull(value);
-              }))
-        continue;
-
-      Operation *clonedOp = builder.clone(*sunkOperation, map);
-      // Only replace uses within the launch op.
-      for (auto result : llvm::enumerate(sunkOperation->getResults())) {
-        auto replacement = clonedOp->getResult(result.index());
-        for (auto &use : llvm::make_early_inc_range(result.value().getUses()))
-          if (use.getOwner()->getParentOfType<gpu::LaunchOp>() == launchOp)
-            use.set(replacement);
-      }
-      processed.insert(sunkOperation);
-    }
-    if (startSize == processed.size())
-      return launchOp.emitError(
-          "found illegal cyclic dependency between operations while sinking");
+  for (Operation *op : toBeSunk) {
+    Operation *clonedOp = builder.clone(*op, map);
+    for (auto pair : llvm::zip(op->getResults(), clonedOp->getResults()))
+      replaceAllUsesInRegionWith(std::get<0>(pair), std::get<1>(pair),
+                                 launchOp.body());
   }
   return success();
 }
 
 // Outline the `gpu.launch` operation body into a kernel function. Replace
 // `gpu.terminator` operations by `gpu.return` in the generated function.
 static gpu::GPUFuncOp outlineKernelFuncImpl(gpu::LaunchOp launchOp,
                                             StringRef kernelFnName,

mlir/test/Dialect/GPU/outlining.mlir

 // CHECK-NEXT: ^[[BLOCK]]:
 // CHECK-NEXT: "use"(%[[KERNEL_ARG0]]) : (f32) -> ()
 // CHECK-NEXT: "some_op"(%[[BID]], %[[BDIM]]) : (index, index) -> ()
 // CHECK-NEXT: = load %[[KERNEL_ARG1]][%[[TID]]] : memref<?xf32, 1>
 
 // -----
 
 // CHECK: module attributes {gpu.container_module}
-
+// CHECK-LABEL: @multiple_launches
 func @multiple_launches() {
   // CHECK: %[[CST:.*]] = constant 8 : index
   %cst = constant 8 : index
   // CHECK: "gpu.launch_func"(%[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]]) {kernel = @multiple_launches_kernel::@multiple_launches_kernel} : (index, index, index, index, index, index) -> ()
   gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %cst, %grid_y = %cst,
                                        %grid_z = %cst)
              threads(%tx, %ty, %tz) in (%block_x = %cst, %block_y = %cst,
                                         %block_z = %cst) {

 
 // CHECK: module @multiple_launches_kernel
 // CHECK: func @multiple_launches_kernel
 // CHECK: module @multiple_launches_kernel_0
 // CHECK: func @multiple_launches_kernel
 
 // -----
 
-func @extra_constants(%arg0 : memref<?xf32>) {
+// CHECK-LABEL: @extra_constants_not_inlined
+func @extra_constants_not_inlined(%arg0: memref<?xf32>) {
+  // CHECK: %[[CST:.*]] = constant 8 : index
+  %cst = constant 8 : index
+  %cst2 = constant 2 : index
+  %c0 = constant 0 : index
+  %cst3 = "secret_constant"() : () -> index
+  // CHECK: "gpu.launch_func"(%[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %{{.*}}, %{{.*}}) {kernel = @extra_constants_not_inlined_kernel::@extra_constants_not_inlined_kernel} : (index, index, index, index, index, index, memref<?xf32>, index) -> ()
+  gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %cst, %grid_y = %cst,
+                                       %grid_z = %cst)
+             threads(%tx, %ty, %tz) in (%block_x = %cst, %block_y = %cst,
+                                        %block_z = %cst) {
+    "use"(%cst2, %arg0, %cst3) : (index, memref<?xf32>, index) -> ()
+    gpu.terminator
+  }
+  return
+}
+
+// CHECK-LABEL: func @extra_constants_not_inlined_kernel(%{{.*}}: memref<?xf32>, %{{.*}}: index)
+// CHECK: constant 2
+
+// -----
+
+// CHECK-LABEL: @extra_constants
+// CHECK-SAME: %[[ARG0:.*]]: memref<?xf32>
+func @extra_constants(%arg0: memref<?xf32>) {
   // CHECK: %[[CST:.*]] = constant 8 : index
   %cst = constant 8 : index
   %cst2 = constant 2 : index
   %c0 = constant 0 : index
   %cst3 = dim %arg0, %c0 : memref<?xf32>
-  // CHECK: "gpu.launch_func"(%[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %{{.*}}, %{{.*}}) {kernel = @extra_constants_kernel::@extra_constants_kernel} : (index, index, index, index, index, index, memref<?xf32>, index) -> ()
+  // CHECK: "gpu.launch_func"(%[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[ARG0]]) {kernel = @extra_constants_kernel::@extra_constants_kernel} : (index, index, index, index, index, index, memref<?xf32>) -> ()
+  gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %cst, %grid_y = %cst,
+                                       %grid_z = %cst)
+             threads(%tx, %ty, %tz) in (%block_x = %cst, %block_y = %cst,
+                                        %block_z = %cst) {
+    "use"(%cst2, %arg0, %cst3) : (index, memref<?xf32>, index) -> ()
+    gpu.terminator
+  }
+  return
+}
+
+// CHECK-LABEL: func @extra_constants_kernel
+// CHECK-SAME: %[[KARG0:.*]]: memref<?xf32>
+// CHECK: constant 2
+// CHECK: constant 0
+// CHECK: dim %[[KARG0]]
+
+// -----
+
+// CHECK-LABEL: @extra_constants_noarg
+// CHECK-SAME: %[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: memref<?xf32>
+func @extra_constants_noarg(%arg0: memref<?xf32>, %arg1: memref<?xf32>) {
+  // CHECK: %[[CST:.*]] = constant 8 : index
+  %cst = constant 8 : index
+  %cst2 = constant 2 : index
+  %c0 = constant 0 : index
+  // CHECK: dim %[[ARG1]]
+  %cst3 = dim %arg1, %c0 : memref<?xf32>
+  // CHECK: "gpu.launch_func"(%[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[CST]], %[[ARG0]], %{{.*}}) {kernel = @extra_constants_noarg_kernel::@extra_constants_noarg_kernel} : (index, index, index, index, index, index, memref<?xf32>, index) -> ()
   gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %cst, %grid_y = %cst,
                                        %grid_z = %cst)
              threads(%tx, %ty, %tz) in (%block_x = %cst, %block_y = %cst,
                                         %block_z = %cst) {
     "use"(%cst2, %arg0, %cst3) : (index, memref<?xf32>, index) -> ()
     gpu.terminator
   }
   return
 }
 
-// CHECK-LABEL: func @extra_constants_kernel(%{{.*}}: memref<?xf32>, %{{.*}}: index)
-// CHECK: constant
-// CHECK: constant
+// CHECK-LABEL: func @extra_constants_noarg_kernel
+// CHECK-SAME: %[[KARG0:.*]]: memref<?xf32>, %[[KARG1:.*]]: index
+// CHECK: %[[KCST:.*]] = constant 2
+// CHECK: "use"(%[[KCST]], %[[KARG0]], %[[KARG1]])
 
 // -----
 
 func @multiple_uses(%arg0 : memref<?xf32>) {
   %c1 = constant 1 : index
   %c2 = constant 2 : index
   // CHECK: gpu.func {{.*}} {
   // CHECK: %[[C2:.*]] = constant 2 : index
func @multiple_uses(%arg0 : memref<?xf32>) {
   }
   return
 }
 
 // -----
 
 llvm.mlir.global internal @global(42 : i64) : !llvm.i64
 
+//CHECK-LABEL: @function_call
 func @function_call(%arg0 : memref<?xf32>) {
   %cst = constant 8 : index
   gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %cst, %grid_y = %cst,
                                        %grid_z = %cst)
              threads(%tx, %ty, %tz) in (%block_x = %cst, %block_y = %cst,
                                         %block_z = %cst) {
     call @device_function() : () -> ()
     call @device_function() : () -> ()