Skip to content

Commit

Permalink
First cl of a series of cls to canonicalize all shape layouts to desc…
Browse files Browse the repository at this point in the history
…ending layout.

For now, we only support broadcast and transpose operations that appear in the entry computation.

PiperOrigin-RevId: 621941481
  • Loading branch information
tensorflower-gardener authored and copybara-github committed Apr 4, 2024
1 parent c2cc020 commit 9e7339a
Show file tree
Hide file tree
Showing 4 changed files with 356 additions and 0 deletions.
34 changes: 34 additions & 0 deletions xla/service/BUILD
Original file line number Diff line number Diff line change
Expand Up @@ -3287,6 +3287,40 @@ cc_library(
],
)

cc_library(
name = "layout_canonicalizer",
srcs = ["layout_canonicalizer.cc"],
hdrs = ["layout_canonicalizer.h"],
deps = [
":hlo_pass",
":layout_assignment",
"//xla:permutation_util",
"//xla:shape_util",
"//xla/hlo/ir:hlo",
"@com_google_absl//absl/algorithm:container",
"@com_google_absl//absl/container:flat_hash_set",
"@com_google_absl//absl/log",
"@com_google_absl//absl/log:check",
"@com_google_absl//absl/status:statusor",
"@com_google_absl//absl/strings",
"@com_google_absl//absl/strings:string_view",
"@com_google_absl//absl/types:span",
],
)

xla_cc_test(
name = "layout_canonicalizer_test",
srcs = ["layout_canonicalizer_test.cc"],
deps = [
":layout_canonicalizer",
"//xla/hlo/ir:hlo",
"//xla/tests:hlo_test_base",
"@com_google_absl//absl/log",
"@com_google_googletest//:gtest_main",
"@tsl//tsl/platform:statusor",
],
)

xla_cc_test(
name = "while_loop_simplifier_test",
srcs = ["while_loop_simplifier_test.cc"],
Expand Down
145 changes: 145 additions & 0 deletions xla/service/layout_canonicalizer.cc
Original file line number Diff line number Diff line change
@@ -0,0 +1,145 @@
/* Copyright 2024 The OpenXLA Authors.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "xla/service/layout_canonicalizer.h"

#include <cstdint>
#include <iterator>
#include <vector>

#include "absl/algorithm/container.h"
#include "absl/container/flat_hash_set.h"
#include "absl/log/check.h"
#include "absl/log/log.h"
#include "absl/status/statusor.h"
#include "absl/strings/str_join.h"
#include "absl/strings/string_view.h"
#include "absl/types/span.h"
#include "xla/hlo/ir/hlo_instruction.h"
#include "xla/hlo/ir/hlo_module.h"
#include "xla/hlo/ir/hlo_opcode.h"
#include "xla/permutation_util.h"
#include "xla/service/layout_assignment.h"
#include "xla/shape.h"
#include "xla/shape_util.h"

namespace xla {
namespace {

std::vector<int64_t> CanonicalizeInstructionLayout(HloInstruction* instr,
bool is_entry_root);

bool IsLayoutDescending(const Shape& shape) {
return absl::c_is_sorted(shape.layout().minor_to_major(),
[](int64_t a, int64_t b) { return a > b; });
}

// Given an instruction (with non-tuple output shape), this function updates the
// output shape such that the layout is descending. It returns the
// major-to-minor layout ordering which will be used when instr is used as an
// operand.
std::vector<int64_t> HandleOutput(HloInstruction* instr) {
CHECK(!instr->shape().IsTuple());
if (IsLayoutDescending(instr->shape())) {
return {};
}
// Create the major-to-minor ordering to construct the new logical dimensions
std::vector<int64_t> major_to_minor;
absl::c_reverse_copy(instr->shape().layout().minor_to_major(),
std::back_inserter(major_to_minor));

// Compose shape's dimensions with the major-to-minor layout
std::vector<int64_t> input_new_logical_dims =
ComposePermutations(instr->shape().dimensions(), major_to_minor);

// Update the shape
*instr->mutable_shape() = ShapeUtil::MakeShapeWithDescendingLayout(
instr->shape().element_type(), input_new_logical_dims);
return major_to_minor;
}

std::vector<int64_t> HandleBroadcast(HloInstruction* broadcast,
bool is_entry_root) {
VLOG(3) << "HandleBroadcast: " << broadcast->name();
// Handle broadcast input
HloInstruction* operand = broadcast->mutable_operand(0);
std::vector<int64_t> operand_map =
CanonicalizeInstructionLayout(operand, false);
VLOG(3) << "operand_map = " << absl::StrJoin(operand_map, ",");

// Handle output
std::vector<int64_t> output_map;
if (!is_entry_root) {
output_map = HandleOutput(broadcast);
}
VLOG(3) << "output_map = " << absl::StrJoin(output_map, ",");

// Compose dimension map with the inverse of the output map.
if (!output_map.empty()) {
std::vector<int64_t> inverse_output_map = InversePermutation(output_map);
std::vector<int64_t> new_broadcast_dimensions;
new_broadcast_dimensions.reserve(broadcast->dimensions().size());
for (int64_t dim : broadcast->dimensions()) {
new_broadcast_dimensions.push_back(inverse_output_map[dim]);
}
VLOG(3) << "dimensions after applying output_map = "
<< absl::StrJoin(new_broadcast_dimensions, ",");
*broadcast->mutable_dimensions() = new_broadcast_dimensions;
}

// Compose dimension map with the operand map.
if (!operand_map.empty()) {
std::vector<int64_t> new_broadcast_dimensions =
ComposePermutations(broadcast->dimensions(), operand_map);
VLOG(3) << "dimensions after applying operand_map = "
<< absl::StrJoin(new_broadcast_dimensions, ",");
*broadcast->mutable_dimensions() = new_broadcast_dimensions;
}
VLOG(3) << "Broadcast after: " << broadcast->ToString();
return output_map;
}

std::vector<int64_t> CanonicalizeInstructionLayout(HloInstruction* instr,
bool is_entry_root) {
if (!LayoutAssignment::InstructionCanChangeLayout(instr)) {
return {};
}
// For now, we only handle broadcast and transpose. I will add other ops
// gradually.
switch (instr->opcode()) {
case HloOpcode::kBroadcast:
case HloOpcode::kTranspose:
return HandleBroadcast(instr, is_entry_root);
default:
break;
}
return {};
}
}; // namespace

absl::StatusOr<bool> LayoutCanonicalizer::Run(
HloModule* module,
const absl::flat_hash_set<absl::string_view>& execution_threads) {
VLOG(3) << "LayoutCanonicalizer::Run: \n" << module->ToString();
for (auto* comp : module->MakeNonfusionComputations(execution_threads)) {
// We only canonicalize the entry computation for now.
if (comp->IsEntryComputation()) {
CanonicalizeInstructionLayout(comp->root_instruction(), true);
}
}
return true;
}

} // namespace xla
43 changes: 43 additions & 0 deletions xla/service/layout_canonicalizer.h
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
/* Copyright 2024 The OpenXLA Authors.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#ifndef XLA_SERVICE_LAYOUT_CANONICALIZER_H_
#define XLA_SERVICE_LAYOUT_CANONICALIZER_H_

#include "absl/container/flat_hash_set.h"
#include "absl/status/statusor.h"
#include "absl/strings/string_view.h"
#include "xla/hlo/ir/hlo_instruction.h"
#include "xla/service/hlo_pass_interface.h"

namespace xla {

// HLO pass that canonicalizes all layouts (except input and output of module)
// to have descending layout by default. This is done by applying the layout
// order to the logical dimension ordering and transform each operation
// attributes according to the new logical shape.
class LayoutCanonicalizer : public HloModulePass {
public:
explicit LayoutCanonicalizer() = default;
~LayoutCanonicalizer() override = default;
absl::string_view name() const override { return "cononicalize_layout"; }
using HloPassInterface::Run;
absl::StatusOr<bool> Run(
HloModule* module,
const absl::flat_hash_set<absl::string_view>& execution_threads) override;
};
} // namespace xla

#endif // XLA_SERVICE_LAYOUT_CANONICALIZER_H_
134 changes: 134 additions & 0 deletions xla/service/layout_canonicalizer_test.cc
Original file line number Diff line number Diff line change
@@ -0,0 +1,134 @@
/* Copyright 2024 The OpenXLA Authors.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "xla/service/layout_canonicalizer.h"

#include <cstdint>
#include <string>
#include <vector>

#include <gtest/gtest.h>
#include "absl/log/log.h"
#include "xla/hlo/ir/hlo_instruction.h"
#include "xla/tests/hlo_test_base.h"
#include "tsl/platform/statusor.h"

namespace xla {
namespace {

using LayoutCanonicalizerTest = HloTestBase;

TEST_F(LayoutCanonicalizerTest, CanonicalizeBroadcast) {
const std::string hlo_string = R"(
HloModule broadcast_module
ENTRY %main {
%p0 = f32[2,6]{0,1} parameter(0)
%broadcast = f32[3,2,1,6]{0,1,2,3} broadcast(%p0), dimensions={1,3}
ROOT %output = f32[3,2,1,6]{3,2,1,0} broadcast(%broadcast), dimensions={0,1,2,3}
}
)";

TF_ASSERT_OK_AND_ASSIGN(auto m, ParseAndReturnVerifiedModule(hlo_string));
LayoutCanonicalizer canonicalizer;
TF_ASSERT_OK_AND_ASSIGN(bool changed, canonicalizer.Run(m.get()));
ASSERT_TRUE(changed);

// Layout should be descending.
HloInstruction* output = m->entry_computation()->root_instruction();
HloInstruction* broadcast = output->mutable_operand(0);
EXPECT_EQ(broadcast->shape().layout().minor_to_major(),
std::vector<int64_t>({3, 2, 1, 0}));

// Logical dimensions should be as follows.
EXPECT_EQ(broadcast->shape().dimensions(),
std::vector<int64_t>({6, 1, 2, 3}));

// Dimensions should change according to the new descending layout.
EXPECT_EQ(broadcast->dimensions(), std::vector<int64_t>({2, 0}));
EXPECT_EQ(output->dimensions(), std::vector<int64_t>({3, 2, 1, 0}));
VLOG(3) << "module after:\n" << m->ToString();
}

TEST_F(LayoutCanonicalizerTest, CanonicalizeBroadcast2) {
const std::string hlo_string = R"(
HloModule broadcast_module
ENTRY %main {
%p0 = f32[2,6]{0,1} parameter(0)
%broadcast = f32[3,2,1,6]{2,3,1,0} broadcast(%p0), dimensions={1,3}
ROOT %output = f32[3,5,2,1,6]{3,4,2,1,0} broadcast(%broadcast), dimensions={0,2,3,4}
}
)";

TF_ASSERT_OK_AND_ASSIGN(auto m, ParseAndReturnVerifiedModule(hlo_string));
LayoutCanonicalizer canonicalizer;
TF_ASSERT_OK_AND_ASSIGN(bool changed, canonicalizer.Run(m.get()));
ASSERT_TRUE(changed);

// Layout should be descending.
HloInstruction* output = m->entry_computation()->root_instruction();
HloInstruction* broadcast = output->mutable_operand(0);
EXPECT_EQ(broadcast->shape().layout().minor_to_major(),
std::vector<int64_t>({3, 2, 1, 0}));

// Logical dimensions should be as follows.
EXPECT_EQ(broadcast->shape().dimensions(),
std::vector<int64_t>({3, 2, 6, 1}));

// Dimensions should change according to the new descending layout.
EXPECT_EQ(broadcast->dimensions(), std::vector<int64_t>({1, 2}));
EXPECT_EQ(output->dimensions(), std::vector<int64_t>({0, 2, 4, 3}));
VLOG(3) << "module after:\n" << m->ToString();
}

TEST_F(LayoutCanonicalizerTest, CanonicalizeBroadcast3) {
const std::string hlo_string = R"(
HloModule broadcast_module
ENTRY %main {
%p0 = f32[2,6]{0,1} parameter(0)
%broadcast = f32[3,2,1,6]{2,3,0,1} broadcast(%p0), dimensions={1,3}
%broadcast2 = f32[3,5,2,1,6]{3,4,0,2,1} broadcast(f32[3,2,1,6]{2,3,0,1} %broadcast), dimensions={0,2,3,4}
ROOT %output = f32[3,5,2,1,6]{3,4,0,1,2} broadcast(f32[3,5,2,1,6]{3,4,0,2,1} %broadcast2), dimensions={0,1,2,3,4}
}
)";

TF_ASSERT_OK_AND_ASSIGN(auto m, ParseAndReturnVerifiedModule(hlo_string));
LayoutCanonicalizer canonicalizer;
TF_ASSERT_OK_AND_ASSIGN(bool changed, canonicalizer.Run(m.get()));
ASSERT_TRUE(changed);

// Layout should be descending.
HloInstruction* root = m->entry_computation()->root_instruction();
HloInstruction* broadcast2 = root->mutable_operand(0);
HloInstruction* broadcast = broadcast2->mutable_operand(0);
EXPECT_EQ(broadcast->shape().layout().minor_to_major(),
std::vector<int64_t>({3, 2, 1, 0}));
EXPECT_EQ(broadcast2->shape().layout().minor_to_major(),
std::vector<int64_t>({4, 3, 2, 1, 0}));

// Logical dimensions should be as follows.
EXPECT_EQ(broadcast->shape().dimensions(),
std::vector<int64_t>({2, 3, 6, 1}));
EXPECT_EQ(broadcast2->shape().dimensions(),
std::vector<int64_t>({5, 2, 3, 6, 1}));

// Dimensions should change according to the new descending layout.
EXPECT_EQ(broadcast->dimensions(), std::vector<int64_t>({0, 2}));
EXPECT_EQ(broadcast2->dimensions(), std::vector<int64_t>({1, 2, 3, 4}));
EXPECT_EQ(root->dimensions(), std::vector<int64_t>({1, 2, 0, 4, 3}));
VLOG(3) << "module after:\n" << m->ToString();
}

} // namespace
} // namespace xla

0 comments on commit 9e7339a

Please sign in to comment.