Integrates LLVM @ a758bcdbd92efb64a3482eb95d2769d74e33f5bb (#18783)

No local patch is carried. --------- Signed-off-by: yzhang93 <zhyuhang88@gmail.com>
iree-org · Oct 16, 2024 · f9fa934 · f9fa934
1 parent 2945399
commit f9fa934
Show file tree

Hide file tree

Showing 5 changed files with 10 additions and 10 deletions.
diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/LLVMGPUPromoteMatmulToFitMMA.cpp b/compiler/src/iree/compiler/Codegen/LLVMGPU/LLVMGPUPromoteMatmulToFitMMA.cpp
@@ -45,7 +45,7 @@ class LLVMGPUPromoteMatmulToFitMMAPass final
     OpBuilder::InsertionGuard guard(rewriter);
     rewriter.setInsertionPointAfter(op);
 
-    SmallVector<bool> packPaddings(op.getNumDpsInputs(), noFold);
+    SmallVector<bool> nofoldFlags(op.getNumDpsInputs(), noFold);
 
     SmallVector<Attribute> paddingValueAttributes;
     for (auto &operand : op->getOpOperands()) {
@@ -58,7 +58,7 @@ class LLVMGPUPromoteMatmulToFitMMAPass final
             .setPaddingDimensions(paddingDims)
             .setPaddingValues(paddingValueAttributes)
             .setPadToMultipleOf(padToMultipleOf)
-            .setPackPaddings(packPaddings)
+            .setPackPaddings(nofoldFlags)
             .setCopyBackOp(linalg::LinalgPaddingOptions::CopyBackOp::None);
 
     FailureOr<linalg::LinalgOp> result =

diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/test/set_transform_strategy_batch_matmul.mlir b/compiler/src/iree/compiler/Codegen/LLVMGPU/test/set_transform_strategy_batch_matmul.mlir
@@ -57,7 +57,7 @@ func.func @batch_matmul_dispatch_0_generic_128x80x320x32_f32() {
 // DEFAULT:   [0, 0, 0, 16]
 // OPTIONS:   [0, 0, 0, 8]
 // CHECK:   %[[PADDED:.+]], %{{.*}}, %{{.+}} = transform.structured.pad %tiled_linalg_op pad_to_multiple_of [1, 1, 1, 1]
-// CHECK:     pack_paddings = [1, 1, 1, 1], padding_dimensions = [0, 1, 2, 3]
+// CHECK:     nofold_flags = [1, 1, 1, 1], padding_dimensions = [0, 1, 2, 3]
 // CHECK:     padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
 // CHECK:   %[[V3:.+]] = transform.get_producer_of_operand %[[PADDED]][2]
 // CHECK:   transform.structured.hoist_pad %{{.*}} by 1 loops

diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/test/set_transform_strategy_convolution.mlir b/compiler/src/iree/compiler/Codegen/LLVMGPU/test/set_transform_strategy_convolution.mlir
@@ -36,7 +36,7 @@ func.func @nchw_convolution() {
 // CHECK: transform.structured.fuse_into_containing_op
 // CHECK: transform.structured.tile_using_for %{{.*}}[0, 0, 0, 16]
 // CHECK: transform.structured.fuse_into_containing_op
-// CHECK: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1, 1] {copy_back_op = "none", pack_paddings = [1, 0, 1], padding_dimensions = [0, 1, 2, 3], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
+// CHECK: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1, 1] {copy_back_op = "none", nofold_flags = [1, 0, 1], padding_dimensions = [0, 1, 2, 3], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
 // CHECK: transform.structured.match ops{["linalg.fill"]}
 // CHECK: %[[RES:.+]] = transform.get_producer_of_operand %{{.*}}[2]
 // CHECK: transform.structured.rewrite_in_destination_passing_style %[[RES]]
@@ -89,7 +89,7 @@ func.func @nhwc_convolution() {
 
 // CHECK: transform.named_sequence
 // CHECK: transform.structured.tile_using_forall %{{.*}}   tile_sizes [1, 128, 128](mapping = [#gpu.block<z>, #gpu.block<y>, #gpu.block<x>])
-// CHECK: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1, 1] {copy_back_op = "none", pack_paddings = [0, 1, 1], padding_dimensions = [0, 1, 2, 3], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
+// CHECK: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1, 1] {copy_back_op = "none", nofold_flags = [0, 1, 1], padding_dimensions = [0, 1, 2, 3], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
 // CHECK: %[[RES:.+]] = transform.get_producer_of_operand %{{.*}}[2]
 // CHECK: transform.structured.rewrite_in_destination_passing_style %[[RES]]
 // CHECK: %[[LHS:.+]] = transform.get_producer_of_operand %{{.*}}[0]

diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/test/set_transform_strategy_matmul.mlir b/compiler/src/iree/compiler/Codegen/LLVMGPU/test/set_transform_strategy_matmul.mlir
@@ -71,7 +71,7 @@ func.func @matmul_1() {
 // CHECK: transform.structured.fuse_into_containing_op
 // CHECK: transform.iree.populate_workgroup_count_region_using_num_threads_slice
 // CHECK: transform.structured.tile_using_for %{{.*}}[0, 0, 16]
-// CHECK: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1] {copy_back_op = "none", pack_paddings = [1, 1, 1], padding_dimensions = [0, 1, 2], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
+// CHECK: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1] {copy_back_op = "none", nofold_flags = [1, 1, 1], padding_dimensions = [0, 1, 2], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
 // CHECK: transform.structured.hoist_pad %{{.}} by 1 loops
 // CHECK: transform.structured.insert_slice_to_copy %{{.*}} : (!transform.any_op) -> !transform.any_op
 // CHECK: transform.structured.tile_using_forall %{{.*}}   num_threads [32, 4](mapping = [#gpu.thread<linear_dim_1>, #gpu.thread<linear_dim_0>])
@@ -133,7 +133,7 @@ func.func @matmul_1() {
 // WITH_OPTIONS: transform.iree.populate_workgroup_count_region_using_num_threads_slice
 // The tiling is affected by td-matmul-strategy-reduc-size: 8.
 // WITH_OPTIONS: transform.structured.tile_using_for %{{.*}}[0, 0, 8]
-// WITH_OPTIONS: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1] {copy_back_op = "none", pack_paddings = [1, 1, 1], padding_dimensions = [0, 1, 2], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
+// WITH_OPTIONS: transform.structured.pad %{{.*}} pad_to_multiple_of [1, 1, 1] {copy_back_op = "none", nofold_flags = [1, 1, 1], padding_dimensions = [0, 1, 2], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]}
 // WITH_OPTIONS: transform.structured.hoist_pad %{{.}} by 1 loops
 // WITH_OPTIONS: transform.structured.insert_slice_to_copy %{{.*}} : (!transform.any_op) -> !transform.any_op
 // WITH_OPTIONS: transform.structured.tile_using_forall %{{.*}}   num_threads [64, 2](mapping = [#gpu.thread<linear_dim_1>, #gpu.thread<linear_dim_0>])
@@ -308,7 +308,7 @@ func.func @matmul_4_partially_unaligned() {
 // Make sure we do not canonicalize because the result is still aligned.
 // CHECK-NEXT: transform.structured.pad %tiled_linalg_op
 // CHECK-SAME:   copy_back_op = "none"
-// CHECK-SAME:   pack_paddings = [1, 1, 1]
+// CHECK-SAME:   nofold_flags = [1, 1, 1]
 // CHECK-SAME:   padding_dimensions = [0, 1, 2]
 // CHECK-SAME:   padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]
 // CHECK:      apply_patterns to %{{.*}} {
@@ -375,7 +375,7 @@ func.func @aligned_matmul() {
 // Make sure we do not canonicalize if the result is aligned to avoid folding the extract_slice on the iterator.
 // CHECK-NEXT: transform.structured.pad %tiled_linalg_op
 // CHECK-SAME:   copy_back_op = "none"
-// CHECK-SAME:   pack_paddings = [1, 1, 1]
+// CHECK-SAME:   nofold_flags = [1, 1, 1]
 // CHECK-SAME:   padding_dimensions = [0, 1, 2]
 // CHECK-SAME:   padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]
 

diff --git a/third_party/llvm-project b/third_party/llvm-project