colin
/
mlir-hlo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[kernel_gen] Lower max rank specialization from 6 to 5 

We don't care much about rank 6 broadcasting operations and this lowers compile times significantly.

PiperOrigin-RevId: 345466476
This commit is contained in:
Tres Popp 2020-12-03 09:18:39 -08:00 committed by TensorFlow MLIR Team
parent d2e3797d7d
commit 7c3f049c8e
2 changed files with 12 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
lib/Dialect/mhlo/transforms/transform_unranked_hlo.cc
View File

@ -409,7 +409,7 @@ struct ConvertUnrankedDynamicBroadcastBinaryOp
// Put each subsequent rank specialization inside the else statement of the // Put each subsequent rank specialization inside the else statement of the
// previous one. // previous one.
OpBuilder else_builder = if_op.getElseBodyBuilder(rewriter.getListener()); OpBuilder else_builder = if_op.getElseBodyBuilder(rewriter.getListener());
constexpr int kMaxRankSpecialization = 6; constexpr int kMaxRankSpecialization = 5;
for (int i = 2; i < kMaxRankSpecialization; i++) { for (int i = 2; i < kMaxRankSpecialization; i++) {
auto inner_if = createIfOpForRankSpecializedBroadcastAndOp( auto inner_if = createIfOpForRankSpecializedBroadcastAndOp(
else_builder, op, greater_rank, i); else_builder, op, greater_rank, i);

19
tests/hlo-transform-unranked.mlir
View File

@ -262,7 +262,7 @@ func @addUnrankedUnranked(
// CHECK-NEXT: %[[C5:.*]] = constant 5 : index // CHECK-NEXT: %[[C5:.*]] = constant 5 : index
// CHECK-NEXT: %[[GREATEST_RANK_IS_5:.*]] = cmpi "eq", %[[GREATEST_RANK]], %[[C5]] : index // CHECK-NEXT: %[[GREATEST_RANK_IS_5:.*]] = cmpi "eq", %[[GREATEST_RANK]], %[[C5]] : index
// Handle rank 5 specialization // Handle rank 5 specialization
// CHECK-NEXT: %[[VAL_50:.*]] = scf.if %[[GREATEST_RANK_IS_5]] -> (tensor<*xf32>) { // CHECK-NEXT: assert %[[GREATEST_RANK_IS_5]]
// CHECK-NEXT: %[[CONST_SHAPE_5:.*]] = shape.const_shape [1, 1, 1, 1, 1] // CHECK-NEXT: %[[CONST_SHAPE_5:.*]] = shape.const_shape [1, 1, 1, 1, 1]
// CHECK-NEXT: %[[BROADCASTED_LHS_5:.*]] = shape.broadcast %[[LHS_SHAPE]], %[[CONST_SHAPE_5]] : tensor<?xindex>, tensor<5xindex> -> tensor<?xindex> // CHECK-NEXT: %[[BROADCASTED_LHS_5:.*]] = shape.broadcast %[[LHS_SHAPE]], %[[CONST_SHAPE_5]] : tensor<?xindex>, tensor<5xindex> -> tensor<?xindex>
// CHECK-NEXT: %[[CASTED_LHS_5:.*]] = tensor_cast %[[BROADCASTED_LHS_5]] : tensor<?xindex> to tensor<5xindex> // CHECK-NEXT: %[[CASTED_LHS_5:.*]] = tensor_cast %[[BROADCASTED_LHS_5]] : tensor<?xindex> to tensor<5xindex>
@ -273,23 +273,6 @@ func @addUnrankedUnranked(
// CHECK-NEXT: %[[RESULT_RANK_5:.*]] = chlo.broadcast_add %[[RESHAPED_LHS_5]], %[[RESHAPED_RHS_5]] : (tensor<?x?x?x?x?xf32>, tensor<?x?x?x?x?xf32>) -> tensor<?x?x?x?x?xf32> // CHECK-NEXT: %[[RESULT_RANK_5:.*]] = chlo.broadcast_add %[[RESHAPED_LHS_5]], %[[RESHAPED_RHS_5]] : (tensor<?x?x?x?x?xf32>, tensor<?x?x?x?x?xf32>) -> tensor<?x?x?x?x?xf32>
// CHECK-NEXT: %[[RESULT_5:.*]] = tensor_cast %[[RESULT_RANK_5]] : tensor<?x?x?x?x?xf32> to tensor<*xf32> // CHECK-NEXT: %[[RESULT_5:.*]] = tensor_cast %[[RESULT_RANK_5]] : tensor<?x?x?x?x?xf32> to tensor<*xf32>
// CHECK-NEXT: scf.yield %[[RESULT_5]] : tensor<*xf32> // CHECK-NEXT: scf.yield %[[RESULT_5]] : tensor<*xf32>
// CHECK-NEXT: } else {
// CHECK-NEXT: %[[C6:.*]] = constant 6 : index
// CHECK-NEXT: %[[GREATEST_RANK_IS_6:.*]] = cmpi "eq", %[[GREATEST_RANK]], %[[C6]] : index
// CHECK-NEXT: assert %[[GREATEST_RANK_IS_6]]
// Handle rank 6 specialization
// CHECK-NEXT: %[[CONST_SHAPE_6:.*]] = shape.const_shape [1, 1, 1, 1, 1, 1]
// CHECK-NEXT: %[[BROADCASTED_LHS_6:.*]] = shape.broadcast %[[LHS_SHAPE]], %[[CONST_SHAPE_6]] : tensor<?xindex>, tensor<6xindex> -> tensor<?xindex>
// CHECK-NEXT: %[[CASTED_LHS_6:.*]] = tensor_cast %[[BROADCASTED_LHS_6]] : tensor<?xindex> to tensor<6xindex>
// CHECK-NEXT: %[[BROADCASTED_RHS_6:.*]] = shape.broadcast %[[RHS_SHAPE]], %[[CONST_SHAPE_6]] : tensor<?xindex>, tensor<6xindex> -> tensor<?xindex>
// CHECK-NEXT: %[[CASTED_RHS_6:.*]] = tensor_cast %[[BROADCASTED_RHS_6]] : tensor<?xindex> to tensor<6xindex>
// CHECK-NEXT: %[[RESHAPED_LHS_6:.*]] = "mhlo.dynamic_reshape"(%[[LHS]], %[[CASTED_LHS_6]]) : (tensor<*xf32>, tensor<6xindex>) -> tensor<?x?x?x?x?x?xf32>
// CHECK-NEXT: %[[RESHAPED_RHS_6:.*]] = "mhlo.dynamic_reshape"(%[[RHS]], %[[CASTED_RHS_6]]) : (tensor<*xf32>, tensor<6xindex>) -> tensor<?x?x?x?x?x?xf32>
// CHECK-NEXT: %[[RESULT_RANK_6:.*]] = chlo.broadcast_add %[[RESHAPED_LHS_6]], %[[RESHAPED_RHS_6]] : (tensor<?x?x?x?x?x?xf32>, tensor<?x?x?x?x?x?xf32>) -> tensor<?x?x?x?x?x?xf32>
// CHECK-NEXT: %[[RESULT_6:.*]] = tensor_cast %[[RESULT_RANK_6]] : tensor<?x?x?x?x?x?xf32> to tensor<*xf32>
// CHECK-NEXT: scf.yield %[[RESULT_6]] : tensor<*xf32>
// CHECK-NEXT: }
// CHECK-NEXT: scf.yield %[[VAL_65:.*]] : tensor<*xf32>
// CHECK-NEXT: } // CHECK-NEXT: }
// CHECK-NEXT: scf.yield %[[VAL_66:.*]] : tensor<*xf32> // CHECK-NEXT: scf.yield %[[VAL_66:.*]] : tensor<*xf32>
// CHECK-NEXT: } // CHECK-NEXT: }