[Mlir-commits] [mlir] 72d76a2 - [mlir][bufferize] lower allocation alignment from 128 to 64 bytes

[Emilio Cota]

Author: Emilio Cota
Date: 2022-12-07T11:12:46-05:00
New Revision: 72d76a2403459a38a1d6daae62de6945097db8f9

URL: https://github.com/llvm/llvm-project/commit/72d76a2403459a38a1d6daae62de6945097db8f9
DIFF: https://github.com/llvm/llvm-project/commit/72d76a2403459a38a1d6daae62de6945097db8f9.diff

LOG: [mlir][bufferize] lower allocation alignment from 128 to 64 bytes

While it is unlikely to matter in practice, there is no reason
for this value to be larger than it should be. 64 bytes is the
size of a cache line in most machines, and we can fit a full
512-bit vector in it.

Reviewed By: springerm
Differential Revision: https://reviews.llvm.org/D139434

Added: 
    

Modified: 
    mlir/include/mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h
    mlir/test/Dialect/Bufferization/Transforms/one-shot-module-bufferize.mlir
    mlir/test/Dialect/Linalg/one-shot-bufferize.mlir
    mlir/test/Dialect/Tensor/bufferize.mlir
    mlir/test/Dialect/Tensor/one-shot-bufferize.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/include/mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h b/mlir/include/mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h
index 78f73870a6ba7..24754d5993a1f 100644
--- a/mlir/include/mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h
+++ b/mlir/include/mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h
@@ -292,7 +292,7 @@ struct BufferizationOptions {
   bool printConflicts = false;
 
   /// Buffer alignment for new memory allocations.
-  unsigned int bufferAlignment = 128;
+  unsigned int bufferAlignment = 64;
 
   /// Initializer functions for analysis state. These can be used to
   /// initialize dialect-specific analysis state.

diff  --git a/mlir/test/Dialect/Bufferization/Transforms/one-shot-module-bufferize.mlir b/mlir/test/Dialect/Bufferization/Transforms/one-shot-module-bufferize.mlir
index d873e0719ad25..d25a3745e3da5 100644
--- a/mlir/test/Dialect/Bufferization/Transforms/one-shot-module-bufferize.mlir
+++ b/mlir/test/Dialect/Bufferization/Transforms/one-shot-module-bufferize.mlir
@@ -436,9 +436,9 @@ func.func @main() {
   %v1 = arith.constant 1.0 : f32
   %v2 = arith.constant 2.0 : f32
 
-  // CHECK-NEXT:   %[[A:.*]] = memref.alloc() {alignment = 128 : i64} : memref<64xf32>
-  // CHECK-NEXT:   %[[B:.*]] = memref.alloc() {alignment = 128 : i64} : memref<64xf32>
-  // CHECK-NEXT:   %[[C:.*]] = memref.alloc() {alignment = 128 : i64} : memref<f32>
+  // CHECK-NEXT:   %[[A:.*]] = memref.alloc() {alignment = 64 : i64} : memref<64xf32>
+  // CHECK-NEXT:   %[[B:.*]] = memref.alloc() {alignment = 64 : i64} : memref<64xf32>
+  // CHECK-NEXT:   %[[C:.*]] = memref.alloc() {alignment = 64 : i64} : memref<f32>
   //  CHECK-DAG:   %[[cA:.*]] = memref.cast %[[A]] : memref<64xf32> to memref<64xf32, strided<[?], offset: ?>>
   //  CHECK-DAG:   %[[cB:.*]] = memref.cast %[[B]] : memref<64xf32> to memref<64xf32, strided<[?], offset: ?>>
   //  CHECK-DAG:   %[[cC:.*]] = memref.cast %[[C]] : memref<f32> to memref<f32, strided<[], offset: ?>>

diff  --git a/mlir/test/Dialect/Linalg/one-shot-bufferize.mlir b/mlir/test/Dialect/Linalg/one-shot-bufferize.mlir
index 424539b7e86f2..ef2d218db6438 100644
--- a/mlir/test/Dialect/Linalg/one-shot-bufferize.mlir
+++ b/mlir/test/Dialect/Linalg/one-shot-bufferize.mlir
@@ -44,7 +44,7 @@ func.func @not_inplace(
   %f0 = arith.constant 0.0 : f32
 
   //     CHECK: %[[D0:.*]] = memref.dim %[[A]], {{.*}} : memref<?xf32, strided<[?], offset: ?>>
-  //     CHECK: %[[ALLOC:.*]] = memref.alloc(%[[D0]]) {alignment = 128 : i64} : memref<?xf32>
+  //     CHECK: %[[ALLOC:.*]] = memref.alloc(%[[D0]]) {alignment = 64 : i64} : memref<?xf32>
   //     CHECK: linalg.fill ins(%[[F0]] : f32) outs(%[[ALLOC]] : memref<?xf32>)
   %r = linalg.fill ins(%f0 : f32) outs(%A : tensor<?xf32>) -> tensor<?xf32>
 
@@ -161,7 +161,7 @@ func.func @matmul(
   %c16 = arith.constant 16 : index
 
   // Hoisted alloc.
-  // CHECK: %[[ALLOC:.*]] = memref.alloc() {alignment = 128 : i64} : memref<128x192xf32>
+  // CHECK: %[[ALLOC:.*]] = memref.alloc() {alignment = 64 : i64} : memref<128x192xf32>
   // CHECK: memref.copy %[[C]], %[[ALLOC]]
 
   // CHECK: scf.for %[[I:.*]] =

diff  --git a/mlir/test/Dialect/Tensor/bufferize.mlir b/mlir/test/Dialect/Tensor/bufferize.mlir
index 82226bc1b23a9..49e29d0fec668 100644
--- a/mlir/test/Dialect/Tensor/bufferize.mlir
+++ b/mlir/test/Dialect/Tensor/bufferize.mlir
@@ -520,7 +520,7 @@ func.func @tensor.reshape(%t1: tensor<?x10xf32>) -> tensor<2x2x5xf32> {
   // CHECK: %[[five:.*]] = arith.constant 5 : i64
   %five = arith.constant 5 : i64
 
-  // CHECK: %[[alloc:.*]] = memref.alloc() {alignment = 128 : i64} : memref<3xi64>
+  // CHECK: %[[alloc:.*]] = memref.alloc() {alignment = 64 : i64} : memref<3xi64>
   // CHECK: %[[zero_idx:.*]] = arith.constant 0 : index
   // CHECK: %[[one_idx:.*]] = arith.constant 1 : index
   // CHECK: %[[two_idx:.*]] = arith.constant 2 : index

diff  --git a/mlir/test/Dialect/Tensor/one-shot-bufferize.mlir b/mlir/test/Dialect/Tensor/one-shot-bufferize.mlir
index 405f112782a74..e1473c844dfe8 100644
--- a/mlir/test/Dialect/Tensor/one-shot-bufferize.mlir
+++ b/mlir/test/Dialect/Tensor/one-shot-bufferize.mlir
@@ -114,7 +114,7 @@ func.func @insert_slice_fun_not_inplace(
     %t : tensor<4xf32> {bufferization.writable = false})
   -> tensor<?xf32>
 {
-  //      CHECK: %[[ALLOC:.*]] = memref.alloc(%{{.*}}) {alignment = 128 : i64} : memref<?xf32>
+  //      CHECK: %[[ALLOC:.*]] = memref.alloc(%{{.*}}) {alignment = 64 : i64} : memref<?xf32>
   //      CHECK: memref.copy %[[A]], %[[ALLOC]] : memref<?xf32{{.*}} to memref<?xf32>
   //      CHECK: %[[SV:.*]] = memref.subview %[[ALLOC]][0] [4] [1] : memref<?xf32> to memref<4xf32, strided<[1]>>
   //      CHECK: memref.copy %[[t]], %[[SV]] : memref<4xf32, strided{{.*}}> to memref<4xf32, strided<[1]>>