From b06f06148e8adf156702e43e421d78a5bd2ca79b Mon Sep 17 00:00:00 2001
From: Adrian Kuegel <akuegel@google.com>
Date: Mon, 1 Apr 2019 05:13:27 -0700
Subject: [PATCH] Change loop order to major to minor instead of minor to
 major.

This matches the comment, and it should make the memory access potentially faster.

PiperOrigin-RevId: 241304479
---
 tensorflow/compiler/xla/service/llvm_ir/BUILD        | 1 +
 tensorflow/compiler/xla/service/llvm_ir/llvm_loop.cc | 3 +++
 2 files changed, 4 insertions(+)

diff --git a/tensorflow/compiler/xla/service/llvm_ir/BUILD b/tensorflow/compiler/xla/service/llvm_ir/BUILD
index 8b35f1b5ba6..ca85dd7647e 100644
--- a/tensorflow/compiler/xla/service/llvm_ir/BUILD
+++ b/tensorflow/compiler/xla/service/llvm_ir/BUILD
@@ -113,6 +113,7 @@ cc_library(
         "//tensorflow/compiler/xla:types",
         "//tensorflow/compiler/xla:xla_data_proto",
         "//tensorflow/core:lib",
+        "@com_google_absl//absl/algorithm:container",
         "@com_google_absl//absl/strings",
         "@com_google_absl//absl/types:span",
         "@llvm//:core",
diff --git a/tensorflow/compiler/xla/service/llvm_ir/llvm_loop.cc b/tensorflow/compiler/xla/service/llvm_ir/llvm_loop.cc
index e08248b9cc4..2e769b5588a 100644
--- a/tensorflow/compiler/xla/service/llvm_ir/llvm_loop.cc
+++ b/tensorflow/compiler/xla/service/llvm_ir/llvm_loop.cc
@@ -18,6 +18,7 @@ limitations under the License.
 #include <numeric>
 #include <vector>
 
+#include "absl/algorithm/container.h"
 #include "absl/strings/str_cat.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Function.h"
@@ -261,11 +262,13 @@ std::vector<llvm::Value*> ForLoopNest::EmitOperandArrayLoopNest(
   // 'dimension_to_skip' dimension.
   std::vector<int64> dimensions;
   const Shape& shape = operand_array.GetShape();
+  // Initially get the dimensions in minor to major order, then reverse them.
   for (int64 dimension : LayoutUtil::MinorToMajor(shape)) {
     if (dimension != dimension_to_skip) {
       dimensions.push_back(dimension);
     }
   }
+  absl::c_reverse(dimensions);
 
   // Create loop nest with one for-loop for each dimension of the
   // output.