PR #17754: [NFC] Extract common builder code into hlo_creation_utils

Imported from GitHub PR #17754 Added `XlaComputationToHloComputation` helper function to `hlo_creation_utils` and replace the uses. Copybara import of the project: -- f9cbfd2 by Sergey Kozub <[email protected]>: Extract common builder code into hlo_creation_utils Merging this change closes #17754 FUTURE_COPYBARA_INTEGRATE_REVIEW=#17754 from openxla:skozub/xla_builder f9cbfd2 PiperOrigin-RevId: 680912058
openxla · Oct 4, 2024 · a0844a4 · a0844a4
1 parent 92b82de
commit a0844a4
Show file tree

Hide file tree

Showing 15 changed files with 61 additions and 119 deletions.
diff --git a/xla/service/BUILD b/xla/service/BUILD
@@ -2575,6 +2575,7 @@ cc_library(
     srcs = ["triangular_solve_expander.cc"],
     hdrs = ["triangular_solve_expander.h"],
     deps = [
+        ":hlo_creation_utils",
         ":hlo_module_config",
         ":op_expander_pass",
         "//xla:shape_util",
@@ -2617,6 +2618,7 @@ cc_library(
     srcs = ["cholesky_expander.cc"],
     hdrs = ["cholesky_expander.h"],
     deps = [
+        ":hlo_creation_utils",
         ":op_expander_pass",
         "//xla:literal",
         "//xla:shape_util",
@@ -2640,6 +2642,7 @@ cc_library(
     srcs = ["qr_expander.cc"],
     hdrs = ["qr_expander.h"],
     deps = [
+        ":hlo_creation_utils",
         ":op_expander_pass",
         "//xla:literal",
         "//xla:shape_util",
@@ -2695,6 +2698,7 @@ cc_library(
     srcs = ["eigh_expander.cc"],
     hdrs = ["eigh_expander.h"],
     deps = [
+        ":hlo_creation_utils",
         ":op_expander_pass",
         "//xla:literal_util",
         "//xla:shape_util",
@@ -3071,6 +3075,7 @@ cc_library(
     srcs = ["bitcast_dtypes_expander.cc"],
     hdrs = ["bitcast_dtypes_expander.h"],
     deps = [
+        ":hlo_creation_utils",
         ":hlo_module_config",
         ":op_expander_pass",
         "//xla:literal_util",
@@ -7483,6 +7488,7 @@ cc_library(
     srcs = ["rng_bit_generator_expander.cc"],
     hdrs = ["rng_bit_generator_expander.h"],
     deps = [
+        ":hlo_creation_utils",
         ":op_expander_pass",
         "//xla:shape_util",
         "//xla:util",
@@ -7597,6 +7603,7 @@ cc_library(
     srcs = ["topk_rewriter.cc"],
     hdrs = ["topk_rewriter.h"],
     deps = [
+        ":hlo_creation_utils",
         ":pattern_matcher",
         "//xla:shape_util",
         "//xla:util",

diff --git a/xla/service/bitcast_dtypes_expander.cc b/xla/service/bitcast_dtypes_expander.cc
@@ -26,6 +26,7 @@ limitations under the License.
 #include "xla/hlo/ir/hlo_instruction.h"
 #include "xla/hlo/ir/hlo_opcode.h"
 #include "xla/primitive_util.h"
+#include "xla/service/hlo_creation_utils.h"
 #include "xla/service/hlo_module_config.h"
 #include "xla/shape.h"
 #include "xla/shape_util.h"
@@ -104,14 +105,8 @@ absl::StatusOr<HloInstruction*> BitcastDtypesExpander::ExpandInstruction(
     BitcastConvertType(input, to_shape.element_type());
 
     TF_ASSIGN_OR_RETURN(XlaComputation xla_computation, b.Build());
-    TF_ASSIGN_OR_RETURN(ProgramShape program_shape,
-                        xla_computation.GetProgramShape());
-    HloModuleConfig config(program_shape);
-    TF_ASSIGN_OR_RETURN(auto new_module, HloModule::CreateFromProto(
-                                             xla_computation.proto(), config));
-    HloCloneContext context(module);
-    computation =
-        module->DeepCloneComputation(new_module->entry_computation(), &context);
+    TF_ASSIGN_OR_RETURN(
+        computation, XlaComputationToHloComputation(xla_computation, module));
   }
 
   return instruction->parent()->AddInstruction(HloInstruction::CreateCall(

diff --git a/xla/service/cholesky_expander.cc b/xla/service/cholesky_expander.cc
@@ -28,6 +28,7 @@ limitations under the License.
 #include "xla/hlo/builder/xla_builder.h"
 #include "xla/literal.h"
 #include "xla/primitive_util.h"
+#include "xla/service/hlo_creation_utils.h"
 #include "xla/shape_util.h"
 #include "xla/status_macros.h"
 #include "xla/util.h"
@@ -248,15 +249,8 @@ absl::StatusOr<HloInstruction*> CholeskyExpander::ExpandInstruction(
     MaybeTransposeInMinorDims(l, !options.lower());
 
     TF_ASSIGN_OR_RETURN(XlaComputation xla_computation, builder.Build());
-
-    TF_ASSIGN_OR_RETURN(ProgramShape program_shape,
-                        xla_computation.GetProgramShape());
-    HloModuleConfig config(program_shape);
-    TF_ASSIGN_OR_RETURN(auto new_module, HloModule::CreateFromProto(
-                                             xla_computation.proto(), config));
-    HloCloneContext context(module);
-    computation =
-        module->DeepCloneComputation(new_module->entry_computation(), &context);
+    TF_ASSIGN_OR_RETURN(
+        computation, XlaComputationToHloComputation(xla_computation, module));
   }
 
   return instruction->parent()->AddInstruction(HloInstruction::CreateCall(

diff --git a/xla/service/eigh_expander.cc b/xla/service/eigh_expander.cc
@@ -34,6 +34,7 @@ limitations under the License.
 #include "xla/hlo/builder/xla_builder.h"
 #include "xla/literal_util.h"
 #include "xla/primitive_util.h"
+#include "xla/service/hlo_creation_utils.h"
 #include "xla/shape_util.h"
 #include "xla/status_macros.h"
 #include "xla/util.h"
@@ -582,15 +583,8 @@ absl::StatusOr<HloInstruction*> EighExpander::ExpandInstruction(
     }
     XlaOp result = BuildEigh(a, lower, max_iter, tol, sort_eigenvalues);
     TF_ASSIGN_OR_RETURN(XlaComputation xla_computation, builder.Build(result));
-
-    TF_ASSIGN_OR_RETURN(ProgramShape program_shape,
-                        xla_computation.GetProgramShape());
-    HloModuleConfig config(program_shape);
-    TF_ASSIGN_OR_RETURN(auto new_module, HloModule::CreateFromProto(
-                                             xla_computation.proto(), config));
-    HloCloneContext context(module);
-    computation =
-        module->DeepCloneComputation(new_module->entry_computation(), &context);
+    TF_ASSIGN_OR_RETURN(
+        computation, XlaComputationToHloComputation(xla_computation, module));
   }
 
   return instruction->parent()->AddInstruction(HloInstruction::CreateCall(

diff --git a/xla/service/gpu/transforms/BUILD b/xla/service/gpu/transforms/BUILD
@@ -1191,6 +1191,7 @@ cc_library(
         "//xla/client:xla_computation",
         "//xla/hlo/ir:hlo",
         "//xla/hlo/pass:hlo_pass",
+        "//xla/service:hlo_creation_utils",
         "//xla/service:hlo_module_config",
         "//xla/service/gpu:backend_configs_cc",
         "//xla/service/gpu:cublas_cudnn",

diff --git a/xla/service/gpu/transforms/cudnn_vectorize_convolutions.cc b/xla/service/gpu/transforms/cudnn_vectorize_convolutions.cc
@@ -40,6 +40,7 @@ limitations under the License.
 #include "xla/service/gpu/cublas_cudnn.h"
 #include "xla/service/gpu/cudnn_support_utils.h"
 #include "xla/service/gpu/stream_executor_util.h"
+#include "xla/service/hlo_creation_utils.h"
 #include "xla/service/hlo_module_config.h"
 #include "xla/shape.h"
 #include "xla/shape_util.h"
@@ -96,24 +97,6 @@ static std::vector<HloCustomCallInstruction*> GetRelevantConvs(
   return convs;
 }
 
-// Converts an XlaBuilder into an HloComputation in the same module as
-// `sibling_computation`.
-//
-// Yes, we serialize/deserialize as a proto.  :)
-static absl::StatusOr<HloComputation*> BuilderToHloComputation(
-    XlaBuilder& b, XlaOp root, HloComputation* sibling_computation) {
-  TF_ASSIGN_OR_RETURN(XlaComputation comp, b.Build(root));
-  TF_ASSIGN_OR_RETURN(ProgramShape program_shape, comp.GetProgramShape());
-  HloModuleConfig config(program_shape);
-  TF_ASSIGN_OR_RETURN(auto new_module,
-                      HloModule::CreateFromProto(comp.proto(), config));
-
-  HloModule* dest_module = sibling_computation->parent();
-  HloCloneContext context(dest_module);
-  return dest_module->DeepCloneComputation(new_module->entry_computation(),
-                                           &context);
-}
-
 // Reshapes `instr` so that it has an extra dimension of size `vect_size` right
 // after `dim`.
 static XlaOp SplitAtDim(XlaOp instr, int64_t dim, int64_t vect_size) {
@@ -460,11 +443,11 @@ static absl::StatusOr<bool> TryRevectorizeConv(
       new_conv_result, dnums->output_feature_dimension(), *output_vect_dim,
       /*orig_vect_size=*/output_shape.dimensions(*output_vect_dim));
 
+  XlaOp root = Tuple(&b, {new_conv_result_unrevectorized, new_conv_scratch});
+  TF_ASSIGN_OR_RETURN(XlaComputation comp, b.Build(root));
   TF_ASSIGN_OR_RETURN(
       HloComputation * new_conv_comp,
-      BuilderToHloComputation(
-          b, Tuple(&b, {new_conv_result_unrevectorized, new_conv_scratch}),
-          conv->parent()));
+      XlaComputationToHloComputation(comp, conv->parent()->parent()));
 
   // Set the name on the new conv.  This is purely cosmetic, but we attempt to
   // preserve e.g. "cudnn-conv.42" instead of "custom-call.42".
@@ -599,11 +582,11 @@ static absl::StatusOr<bool> TryVectorizeConv(
       Collapse(new_conv_result, {dnums->output_feature_dimension(),
                                  dnums->output_feature_dimension() + 1});
 
+  XlaOp root = Tuple(&b, {conv_result_collapsed, new_conv_scratch});
+  TF_ASSIGN_OR_RETURN(XlaComputation comp, b.Build(root));
   TF_ASSIGN_OR_RETURN(
       HloComputation * new_conv_comp,
-      BuilderToHloComputation(
-          b, Tuple(&b, {conv_result_collapsed, new_conv_scratch}),
-          conv->parent()));
+      XlaComputationToHloComputation(comp, conv->parent()->parent()));
 
   // Create a tuple and replace the old conv with it!
   VLOG(1) << "Vectorized conv to: " << new_conv_comp->ToString();

diff --git a/xla/service/hlo_creation_utils.cc b/xla/service/hlo_creation_utils.cc
@@ -597,12 +597,22 @@ HloInstruction* MaybeMakeTuple(absl::Span<HloInstruction* const> operands) {
       HloInstruction::CreateTuple(operands));
 }
 
+absl::StatusOr<HloComputation*> XlaComputationToHloComputation(
+    XlaComputation& src_comp, HloModule* dest_module) {
+  TF_ASSIGN_OR_RETURN(ProgramShape program_shape, src_comp.GetProgramShape());
+  HloModuleConfig config(program_shape);
+  TF_ASSIGN_OR_RETURN(auto new_module,
+                      HloModule::CreateFromProto(src_comp.proto(), config));
+  HloCloneContext context(dest_module);
+  return dest_module->DeepCloneComputation(new_module->entry_computation(),
+                                           &context);
+}
+
 absl::StatusOr<HloInstruction*> MakeSortHlo(
     const Shape& sort_shape, absl::Span<HloInstruction* const> operands,
     int64_t dimension_to_sort, bool is_stable, HloComputation::Builder* builder,
     HloModule* module, const OpMetadata* metadata) {
   CHECK(!operands.empty()) << "Sort Hlo requires at least one operand.";
-  HloComputation* compare_computation;
   XlaBuilder b("Sort.Compare");
   if (metadata != nullptr) {
     b.SetOpMetadata(*metadata);
@@ -612,13 +622,8 @@ absl::StatusOr<HloInstruction*> MakeSortHlo(
     operand_types[i] = operands[i]->shape().element_type();
   }
   XlaComputation comparator = CreateScalarLtComputation(operand_types, &b);
-  TF_ASSIGN_OR_RETURN(ProgramShape program_shape, comparator.GetProgramShape());
-  HloModuleConfig config(program_shape);
-  TF_ASSIGN_OR_RETURN(auto new_module,
-                      HloModule::CreateFromProto(comparator.proto(), config));
-  HloCloneContext context(module);
-  compare_computation =
-      module->DeepCloneComputation(new_module->entry_computation(), &context);
+  TF_ASSIGN_OR_RETURN(HloComputation * compare_computation,
+                      XlaComputationToHloComputation(comparator, module));
   return builder->AddInstruction(HloInstruction::CreateSort(
       sort_shape, dimension_to_sort, operands, compare_computation, is_stable));
 }

diff --git a/xla/service/hlo_creation_utils.h b/xla/service/hlo_creation_utils.h
@@ -22,6 +22,7 @@ limitations under the License.
 #include <vector>
 
 #include "absl/types/span.h"
+#include "xla/hlo/builder/xla_computation.h"
 #include "xla/hlo/ir/hlo_computation.h"
 #include "xla/hlo/ir/hlo_instruction.h"
 #include "xla/literal_util.h"
@@ -257,6 +258,11 @@ absl::StatusOr<HloInstruction*> MakeSelectHlo(
 // instruction with all the operands. Crashes if `operands` is empty.
 HloInstruction* MaybeMakeTuple(absl::Span<HloInstruction* const> operands);
 
+// Creates a HloComputation in the destination module from a builder's
+// XlaComputation.
+absl::StatusOr<HloComputation*> XlaComputationToHloComputation(
+    XlaComputation& src_comp, HloModule* dest_module);
+
 // Creates a Sort HLO instruction and adds it to the computation containing the
 // operands. All operands must be in the same computation. Also creates a
 // default compare sub-computation which sorts the first operand into ascending

diff --git a/xla/service/qr_expander.cc b/xla/service/qr_expander.cc
@@ -29,6 +29,7 @@ limitations under the License.
 #include "xla/hlo/builder/xla_builder.h"
 #include "xla/literal.h"
 #include "xla/primitive_util.h"
+#include "xla/service/hlo_creation_utils.h"
 #include "xla/shape_util.h"
 #include "xla/status_macros.h"
 #include "xla/util.h"
@@ -551,15 +552,8 @@ absl::StatusOr<HloInstruction*> QrExpander::ExpandInstruction(
     }
 
     TF_ASSIGN_OR_RETURN(XlaComputation xla_computation, builder.Build(result));
-
-    TF_ASSIGN_OR_RETURN(ProgramShape program_shape,
-                        xla_computation.GetProgramShape());
-    HloModuleConfig config(program_shape);
-    TF_ASSIGN_OR_RETURN(auto new_module, HloModule::CreateFromProto(
-                                             xla_computation.proto(), config));
-    HloCloneContext context(module);
-    computation =
-        module->DeepCloneComputation(new_module->entry_computation(), &context);
+    TF_ASSIGN_OR_RETURN(
+        computation, XlaComputationToHloComputation(xla_computation, module));
   }
 
   return instruction->parent()->AddInstruction(HloInstruction::CreateCall(

diff --git a/xla/service/rng_bit_generator_expander.cc b/xla/service/rng_bit_generator_expander.cc
@@ -23,6 +23,7 @@ limitations under the License.
 #include "xla/hlo/ir/hlo_instruction.h"
 #include "xla/hlo/ir/hlo_instructions.h"
 #include "xla/hlo/ir/hlo_module.h"
+#include "xla/service/hlo_creation_utils.h"
 #include "xla/shape.h"
 #include "xla/shape_util.h"
 #include "xla/util.h"
@@ -86,15 +87,8 @@ RngBitGeneratorExpander::GetGeneratorComputation(const Shape& data_shape,
       ConcatInDim(&builder, {Reshape(key_op, {1}), output.state}, 0);
   Tuple(&builder, {final_state, output.value});
   TF_ASSIGN_OR_RETURN(XlaComputation xla_computation, builder.Build());
-
-  TF_ASSIGN_OR_RETURN(ProgramShape program_shape,
-                      xla_computation.GetProgramShape());
-  HloModuleConfig config(program_shape);
-  TF_ASSIGN_OR_RETURN(auto new_module, HloModule::CreateFromProto(
-                                           xla_computation.proto(), config));
-  HloCloneContext context(module);
-  HloComputation* new_computation =
-      module->DeepCloneComputation(new_module->entry_computation(), &context);
+  TF_ASSIGN_OR_RETURN(HloComputation * new_computation,
+                      XlaComputationToHloComputation(xla_computation, module));
   computation_cache_.emplace(cache_key, new_computation);
   return new_computation;
 }

diff --git a/xla/service/rng_expander.cc b/xla/service/rng_expander.cc
@@ -111,16 +111,7 @@ absl::StatusOr<HloComputation*> GetComputationForRng(HloInstruction* rng) {
   }
 
   TF_ASSIGN_OR_RETURN(XlaComputation xla_computation, builder.Build());
-
-  TF_ASSIGN_OR_RETURN(ProgramShape program_shape,
-                      xla_computation.GetProgramShape());
-  HloModuleConfig config(program_shape);
-  TF_ASSIGN_OR_RETURN(auto new_module, HloModule::CreateFromProto(
-                                           xla_computation.proto(), config));
-  HloModule* module = rng->GetModule();
-  HloCloneContext context(module);
-  return module->DeepCloneComputation(new_module->entry_computation(),
-                                      &context);
+  return XlaComputationToHloComputation(xla_computation, rng->GetModule());
 }
 
 }  // namespace

diff --git a/xla/service/spmd/BUILD b/xla/service/spmd/BUILD
@@ -61,6 +61,7 @@ cc_library(
         "//xla/service:custom_call_sharding_helper",
         "//xla/service:dot_as_convolution_util",
         "//xla/service:flatten_call_graph",
+        "//xla/service:hlo_creation_utils",
         "//xla/service:hlo_cse",
         "//xla/service:hlo_dce",
         "//xla/service:hlo_lexer",

diff --git a/xla/service/spmd/custom_call_handler.cc b/xla/service/spmd/custom_call_handler.cc
@@ -43,6 +43,7 @@ limitations under the License.
 #include "xla/hlo/utils/hlo_sharding_util.h"
 #include "xla/literal_util.h"
 #include "xla/service/custom_call_sharding_helper.h"
+#include "xla/service/hlo_creation_utils.h"
 #include "xla/service/hlo_lexer.h"
 #include "xla/service/hlo_module_config.h"
 #include "xla/service/host_memory_offload_annotations.h"
@@ -207,13 +208,8 @@ absl::Status SpmdPartitioningVisitor::HandleCustomCallTopK(
   XlaComputation comparator = CreateScalarComparisonComputation(
       "compare-value-and-index", {input->shape().element_type(), S32}, {Gt, Lt},
       &b);
-  TF_ASSIGN_OR_RETURN(ProgramShape program_shape, comparator.GetProgramShape());
-  HloModuleConfig config(program_shape);
-  TF_ASSIGN_OR_RETURN(auto new_module,
-                      HloModule::CreateFromProto(comparator.proto(), config));
-  HloCloneContext context(module_);
-  auto compare_computation =
-      module_->DeepCloneComputation(new_module->entry_computation(), &context);
+  TF_ASSIGN_OR_RETURN(HloComputation * compare_computation,
+                      XlaComputationToHloComputation(comparator, module_));
   // Each partition needs to do TopK separately, thus the base shape for sort
   // becomes [ceil(batch_size / batch_dim_partition), k * shard_count].
   const Shape sort_shape = ShapeUtil::MakeTupleShape(