Skip to content

R2.18 rocm enhanced bd opt #2798

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
amd-jianli12 wants to merge 4 commits into
base: r2.18-rocm-enhanced
Choose a base branch
from
Open

R2.18 rocm enhanced bd opt #2798

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
93a9581
Non-atomic for UnsortedSegmentCustomKernel
amd-jianli12 Jan 8, 2025
a2c4ba4
Abandon cluster with small size regardless of _XlaCompile attr
amd-jianli12 Jan 8, 2025
fa1e9b5
Always inline functions with internal linkage
amd-jianli12 Feb 20, 2025
8d9feb8
Decluster SplitV operator if no use in current cluster
amd-jianli12 Feb 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 2 additions & 3 deletions tensorflow/compiler/jit/mark_for_compilation_pass.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -916,7 +916,7 @@ Status MarkForCompilationPassImpl::DeclusterNodes() {
// increasing its live range.
//
// See b/221997940 for a real-world example of this.
if (n->op_def().name() == "Fill" &&
if ((n->op_def().name() == "Fill" || n->def().op() == "SplitV") &&
n->out_nodes().begin() != n->out_nodes().end() &&
absl::c_all_of(n->out_nodes(), [&](Node* user) {
return GetClusterForNode(user) != cluster;
Expand Down Expand Up @@ -990,8 +990,7 @@ Status MarkForCompilationPassImpl::CreateClusters() {
// trouble.

if (cluster->effective_cluster_size() >= debug_options_.min_cluster_size ||
cluster->has_functional_control_flow() ||
cluster->is_xla_compile_attr_true()) {
cluster->has_functional_control_flow()) {
string& name = cluster_names[cluster->cycles_graph_node_id()];

if (name.empty()) {
Expand Down
2 changes: 1 addition & 1 deletion tensorflow/core/kernels/segment_reduction_ops_gpu.cu.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -902,7 +902,7 @@ struct UnsortedSegmentFunctor<GPUDevice, T, Index, InitialValueF, ReductionF> {
config = GetGpuLaunchConfig(data_size, d);
TF_CHECK_OK(GpuLaunchKernel(
UnsortedSegmentCustomKernel<
T, Index, typename ReduceUpdateOpFor<ReductionF>::atomic_op>,
T, Index, typename ReduceUpdateOpFor<ReductionF>::nonatomic_op>,
config.block_count, config.thread_per_block, 0, d.stream(),
input_outer_dim_size, input_inner_dim_size, output_outer_dim_size,
unsorted_segment_ids.data(), data.data(), output.data()));
Expand Down
8 changes: 6 additions & 2 deletions third_party/xla/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -947,12 +947,16 @@ absl::Status AMDGPUTargetModuleLinker(
// not has major impact as the hipcc path by default enables flush to zero for
// compilation.
// If ftz is enabled, set it as an attribute on every function in the module.
if (debug_options.xla_gpu_ftz()) {
for (llvm::Function& fn : *module) {
for (llvm::Function& fn : *module) {
if (debug_options.xla_gpu_ftz()) {
// may be necessary for the compiler to generate atomics (confirm!)
fn.addFnAttr("denormal-fp-math-f32", "preserve-sign");
fn.addFnAttr("amdgpu-unsafe-fp-atomics", "true");
}

if (!fn.isDeclaration() && fn.hasInternalLinkage()) {
fn.addFnAttr(llvm::Attribute::AttrKind::AlwaysInline);
}
}

return absl::OkStatus();
Expand Down