experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[XLA/GPU] Use SequentialHloOrdering for single stream modules.

Browse Source 
Also simplify tests in gpu_hlo_schedule_test.

PiperOrigin-RevId: 358025620
Change-Id: I1de2f85c91765e967865c6ea90bc629627146559
This commit is contained in:
A. Unique TensorFlower 2021-02-17 13:53:40 -08:00 committed by TensorFlower Gardener
parent 7716b719e5
commit 2c332fbc4b
5 changed files with 86 additions and 51 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
tensorflow/compiler/xla/service/gpu/gpu_compiler.cc
View File

@ -520,9 +520,9 @@ GpuCompiler::RunHloPassesAndBufferAssignement(
std::unique_ptr<StreamAssignment> stream_assignment =
AssignStreams(*hlo_module);
TF_ASSIGN_OR_RETURN(std::unique_ptr<GpuHloSchedule> hlo_schedule,
GpuHloSchedule::Build(hlo_module.get(),
*stream_assignment, pointer_size_));
TF_ASSIGN_OR_RETURN(
std::unique_ptr<GpuHloSchedule> hlo_schedule,
GpuHloSchedule::Build(*hlo_module, *stream_assignment, pointer_size_));
auto buffer_size_bytes_function =
[this](const BufferValue& buffer_value) -> int64 {
@ -565,7 +565,7 @@ static Status CompileModuleToLlvmIrImpl(
AssignStreams(*hlo_module);
TF_ASSIGN_OR_RETURN(
std::unique_ptr<GpuHloSchedule> hlo_schedule,
GpuHloSchedule::Build(hlo_module, *stream_assignment, pointer_size));
GpuHloSchedule::Build(*hlo_module, *stream_assignment, pointer_size));
auto buffer_size_bytes_function =
[pointer_size](const BufferValue& buffer_value) -> int64 {

25
tensorflow/compiler/xla/service/gpu/gpu_hlo_schedule.cc
View File

@ -190,29 +190,30 @@ GpuHloSchedule::GpuHloSchedule() {}
/* static */
StatusOr<std::unique_ptr<GpuHloSchedule>> GpuHloSchedule::Build(
HloModule* module, const StreamAssignment& stream_assignment,
const HloModule& module, const StreamAssignment& stream_assignment,
int64 pointer_size) {
std::unique_ptr<GpuHloSchedule> schedule(new GpuHloSchedule);
// Initialize thunk_launch_order_, the total order of thunk launches.
HloComputation* entry_computation = module->entry_computation();
HloComputation* entry_computation = module.entry_computation();
if (stream_assignment.StreamCount() == 1) {
// All kernels are launched on a single stream, so there's no loss of
// concurrency by optimizing for minimal memory usage.
TF_ASSIGN_OR_RETURN(
HloSchedule sequences,
ScheduleModule(module, [pointer_size](const BufferValue& buffer) {
return ShapeUtil::ByteSizeOf(buffer.shape(), pointer_size);
}));
schedule->thunk_launch_order_ =
sequences.sequence(entry_computation).instructions();
schedule->hlo_ordering_ =
absl::make_unique<SequentialHloOrdering>(sequences);
HloInstructionSequence sequence,
ScheduleComputation(
entry_computation, [pointer_size](const BufferValue& buffer) {
return ShapeUtil::ByteSizeOf(buffer.shape(), pointer_size);
}));
schedule->thunk_launch_order_ = sequence.instructions();
} else {
// BFS tends to increase concurrency, but also increases memory usage.
BFSLaunchOrder(entry_computation, &schedule->thunk_launch_order_);
schedule->hlo_ordering_ = absl::make_unique<GpuHloOrdering>(
module, stream_assignment, schedule->thunk_launch_order_);
}
schedule->hlo_ordering_ = absl::make_unique<GpuHloOrdering>(
&module, stream_assignment, schedule->thunk_launch_order_);
return std::move(schedule);
}

2
tensorflow/compiler/xla/service/gpu/gpu_hlo_schedule.h
View File

@ -41,7 +41,7 @@ class GpuHloSchedule {
// Constructs an GpuHloSchedule for the given module, based on the given
// stream assignment.
static StatusOr<std::unique_ptr<GpuHloSchedule>> Build(
HloModule* module, const StreamAssignment& stream_assignment,
const HloModule& module, const StreamAssignment& stream_assignment,
int64 pointer_size);
// Returns the total order of thunk launches, represented in terms of HLO

74
tensorflow/compiler/xla/service/gpu/gpu_hlo_schedule_test.cc
View File

@ -39,7 +39,7 @@ class GpuHloScheduleTest : public HloTestBase {
Shape f32_2x2_ = ShapeUtil::MakeShape(F32, {2, 2});
static std::unique_ptr<GpuHloSchedule> BuildGpuHloSchedule(
HloModule* module, const StreamAssignment& streams) {
const HloModule& module, const StreamAssignment& streams) {
return GpuHloSchedule::Build(module, streams, /*pointer_size=*/8)
.ConsumeValueOrDie();
}
@ -86,7 +86,7 @@ TEST_F(GpuHloScheduleTest, SequentialMatMul) {
EXPECT_EQ(streams->StreamNumberForHlo(*dot1),
streams->StreamNumberForHlo(*dot2));
auto schedule = BuildGpuHloSchedule(module.get(), *streams);
auto schedule = BuildGpuHloSchedule(*module, *streams);
// Remove parameters, which are unordered.
EXPECT_EQ(RemoveHlo(schedule->ThunkLaunchOrder(), {x, y, z}),
HloVec({dot1, dot2}));
@ -94,10 +94,32 @@ TEST_F(GpuHloScheduleTest, SequentialMatMul) {
// Parameters x,y,z are mutually unordered, while dot1 and dot2 are
// transitively ordered by operands.
auto order = schedule->ConsumeHloOrdering();
EXPECT_TRUE(order->ExecutesBefore(z, y));
EXPECT_TRUE(order->ExecutesBefore(y, x));
EXPECT_TRUE(order->ExecutesBefore(x, dot1));
EXPECT_TRUE(order->ExecutesBefore(x, dot2));
EXPECT_TRUE(order->ExecutesBefore(y, dot1));
EXPECT_TRUE(order->ExecutesBefore(y, dot2));
EXPECT_TRUE(order->ExecutesBefore(z, dot2));
EXPECT_TRUE(order->ExecutesBefore(dot1, dot2));
EXPECT_FALSE(order->ExecutesBefore(x, x));
EXPECT_FALSE(order->ExecutesBefore(x, y));
EXPECT_FALSE(order->ExecutesBefore(x, z));
EXPECT_FALSE(order->ExecutesBefore(y, x));
EXPECT_FALSE(order->ExecutesBefore(y, y));
EXPECT_FALSE(order->ExecutesBefore(y, z));
EXPECT_FALSE(order->ExecutesBefore(z, x));
EXPECT_FALSE(order->ExecutesBefore(z, y));
EXPECT_FALSE(order->ExecutesBefore(z, z));
EXPECT_FALSE(order->ExecutesBefore(z, dot1));
EXPECT_FALSE(order->ExecutesBefore(dot1, x));
EXPECT_FALSE(order->ExecutesBefore(dot1, y));
EXPECT_FALSE(order->ExecutesBefore(dot1, z));
EXPECT_FALSE(order->ExecutesBefore(dot1, dot1));
EXPECT_FALSE(order->ExecutesBefore(dot2, x));
EXPECT_FALSE(order->ExecutesBefore(dot2, y));
EXPECT_FALSE(order->ExecutesBefore(dot2, z));
EXPECT_FALSE(order->ExecutesBefore(dot2, dot1));
EXPECT_FALSE(order->ExecutesBefore(dot2, dot2));
}
// Test of a single stream, where data dependencies do not fully determine the
@ -126,7 +148,7 @@ TEST_F(GpuHloScheduleTest, SequentialAdd) {
EXPECT_EQ(streams->StreamNumberForHlo(*add1),
streams->StreamNumberForHlo(*add3));
auto schedule = BuildGpuHloSchedule(module.get(), *streams);
auto schedule = BuildGpuHloSchedule(*module, *streams);
// Remove parameters, which are unordered.
EXPECT_EQ(RemoveHlo(schedule->ThunkLaunchOrder(), {x, y, z}),
HloVec({add1, add2, add3}));
@ -134,11 +156,43 @@ TEST_F(GpuHloScheduleTest, SequentialAdd) {
// Parameters x,y,z are mutually unordered, while add1, add2 and add3 are
// transitively ordered by operands.
auto order = schedule->ConsumeHloOrdering();
EXPECT_TRUE(order->ExecutesBefore(y, z));
EXPECT_TRUE(order->ExecutesBefore(z, x));
EXPECT_TRUE(order->ExecutesBefore(x, add1));
EXPECT_TRUE(order->ExecutesBefore(add1, add2));
EXPECT_TRUE(order->ExecutesBefore(x, add2));
EXPECT_TRUE(order->ExecutesBefore(x, add3));
EXPECT_TRUE(order->ExecutesBefore(y, add1));
EXPECT_TRUE(order->ExecutesBefore(y, add2));
EXPECT_TRUE(order->ExecutesBefore(y, add3));
EXPECT_TRUE(order->ExecutesBefore(z, add2));
EXPECT_TRUE(order->ExecutesBefore(z, add3));
EXPECT_TRUE(order->ExecutesBefore(add1, add3));
EXPECT_TRUE(order->ExecutesBefore(add2, add3));
// The HLO graph does not define an ordering for add1 and add2, but their
// assignment onto the same stream does define an ordering.
if (order->ExecutesBefore(add1, add2)) {
EXPECT_FALSE(order->ExecutesBefore(add2, add1));
} else {
EXPECT_TRUE(order->ExecutesBefore(add2, add1));
EXPECT_FALSE(order->ExecutesBefore(add1, add2));
}
EXPECT_FALSE(order->ExecutesBefore(x, x));
EXPECT_FALSE(order->ExecutesBefore(x, y));
EXPECT_FALSE(order->ExecutesBefore(x, z));
EXPECT_FALSE(order->ExecutesBefore(y, x));
EXPECT_FALSE(order->ExecutesBefore(y, y));
EXPECT_FALSE(order->ExecutesBefore(y, z));
EXPECT_FALSE(order->ExecutesBefore(z, x));
EXPECT_FALSE(order->ExecutesBefore(z, y));
EXPECT_FALSE(order->ExecutesBefore(z, z));
EXPECT_FALSE(order->ExecutesBefore(z, add1));
EXPECT_FALSE(order->ExecutesBefore(add1, x));
EXPECT_FALSE(order->ExecutesBefore(add1, y));
EXPECT_FALSE(order->ExecutesBefore(add1, z));
EXPECT_FALSE(order->ExecutesBefore(add1, add1));
EXPECT_FALSE(order->ExecutesBefore(add2, x));
EXPECT_FALSE(order->ExecutesBefore(add2, y));
EXPECT_FALSE(order->ExecutesBefore(add2, z));
EXPECT_FALSE(order->ExecutesBefore(add2, add2));
}
// Test of two streams.
@ -162,7 +216,7 @@ TEST_F(GpuHloScheduleTest, ConcurrentMatMul) {
EXPECT_NE(streams->StreamNumberForHlo(*dot1),
streams->StreamNumberForHlo(*dot2));
auto schedule = BuildGpuHloSchedule(module.get(), *streams);
auto schedule = BuildGpuHloSchedule(*module, *streams);
// Remove parameters, which are unordered.
HloVec thunk_launch_order = RemoveHlo(schedule->ThunkLaunchOrder(), {x, y});
EXPECT_TRUE(thunk_launch_order == HloVec({dot1, dot2, add}) ||
@ -254,7 +308,7 @@ TEST_F(GpuHloScheduleTest, LatticeMatMul) {
// We don't check the thunk launch order, since there are many valid total
// orders, and it's annoying to express.
auto schedule = BuildGpuHloSchedule(module.get(), *streams);
auto schedule = BuildGpuHloSchedule(*module, *streams);
auto order = schedule->ConsumeHloOrdering();
const HloVec all_params(

28
tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc
View File

@ -3829,20 +3829,6 @@ Status CheckConditionalBuffersShareAllocation(
return Status::OK();
}
Status AcceptMaybeOrdered(HloComputation* computation,
IrEmitterUnnested* emitter,
const BufferAssignment& buffer_assignment) {
const auto& debug_options = computation->parent()->config().debug_options();
if (debug_options.xla_gpu_disable_multi_streaming()) {
const HloInstructionSequence* sequence =
buffer_assignment.hlo_ordering().SequentialOrder(*computation);
// Always expect a sequential ordering for single-stream programs.
TF_RET_CHECK(sequence);
return computation->AcceptOrdered(emitter, sequence->instructions());
}
return computation->Accept(emitter);
}
} // namespace
StatusOr<std::unique_ptr<Thunk>> IrEmitterUnnested::BuildWhileThunk(
@ -3856,18 +3842,14 @@ StatusOr<std::unique_ptr<Thunk>> IrEmitterUnnested::BuildWhileThunk(
TF_ASSIGN_OR_RETURN(auto ir_emitter_condition,
IrEmitterUnnested::Create(hlo_module_config_, condition,
ir_emitter_context_));
TF_RETURN_IF_ERROR(
AcceptMaybeOrdered(condition, ir_emitter_condition.get(),
ir_emitter_context_->buffer_assignment()));
TF_RETURN_IF_ERROR(condition->Accept(ir_emitter_condition.get()));
// Generate thunk sequence for while 'body'.
HloComputation* body = hlo->while_body();
TF_ASSIGN_OR_RETURN(
auto ir_emitter_body,
IrEmitterUnnested::Create(hlo_module_config_, body, ir_emitter_context_));
TF_RETURN_IF_ERROR(AcceptMaybeOrdered(
body, ir_emitter_body.get(), ir_emitter_context_->buffer_assignment()));
TF_RETURN_IF_ERROR(body->Accept(ir_emitter_body.get()));
const auto* index_map = ir_emitter_context_->profile_index_map();
absl::optional<size_t> condition_profile_index, body_profile_index;
@ -3895,8 +3877,7 @@ StatusOr<std::unique_ptr<Thunk>> IrEmitterUnnested::BuildForThunk(
TF_ASSIGN_OR_RETURN(
auto ir_emitter_body,
IrEmitterUnnested::Create(hlo_module_config_, body, ir_emitter_context_));
TF_RETURN_IF_ERROR(AcceptMaybeOrdered(
body, ir_emitter_body.get(), ir_emitter_context_->buffer_assignment()));
TF_RETURN_IF_ERROR(body->Accept(ir_emitter_body.get()));
const auto* index_map = ir_emitter_context_->profile_index_map();
absl::optional<size_t> body_profile_index;
@ -3933,8 +3914,7 @@ StatusOr<std::unique_ptr<Thunk>> IrEmitterUnnested::BuildConditionalThunk(
auto ir_emitter,
IrEmitterUnnested::Create(hlo_module_config_, branch_computation,
ir_emitter_context_));
TF_CHECK_OK(AcceptMaybeOrdered(branch_computation, ir_emitter.get(),
ir_emitter_context_->buffer_assignment()));
TF_CHECK_OK(branch_computation->Accept(ir_emitter.get()));
branch_thunks.push_back(std::move(*ir_emitter->ConsumeThunkSequence()));
absl::optional<size_t> profile_index;