PR #20794: [gpu][ds-fusion] Add handling for offset module in ds-fusi…

…on thunk Imported from GitHub PR openxla/xla#20794 This patch adds support for offset modules in ds fusion thunk. It also moves the `ResourceRequests` structure from `gpu_executable.cc` to `gpu_executable.h` because a valid implementation of the abstract class `Thunk::ResourceRequests` is required for calling `Thunk::Prepare()`. This is split from #20332 as per request. Copybara import of the project: -- 53226e9428dff816819c192735b9569ef3d309ea by Shraiysh Vaishay <[email protected]>: [gpu][ds-fusion] Add handling for offset module in ds-fusion thunk This patch adds support for offset modules in ds fusion thunk. It also moves the `ResourceRequests` structure from `gpu_executable.cc` to `gpu_executable.h` because a valid implementation of the abstract class `Thunk::ResourceRequests` is required for calling `Thunk::Prepare()`. -- e554ac6b09b5023082210c5100a1f1a86c1c6605 by Shraiysh Vaishay <[email protected]>: Address comments and rebase -- 4ac8efa0c9885b0f21f526072f821ace40be4043 by Shraiysh Vaishay <[email protected]>: Address comments -- 8c23e2a4e2ff66fb7361c48c9d988ebb2261bc41 by Shraiysh Vaishay <[email protected]>: Addressed comments. -- efbe6bfa6f10f1ee998683161e398ae3adc0f183 by Shraiysh Vaishay <[email protected]>: Addressed comments Merging this change closes #20794 FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#20794 from shraiysh:ds_fusion_thunk_changes efbe6bfa6f10f1ee998683161e398ae3adc0f183 PiperOrigin-RevId: 718748248
tensorflow · Jan 27, 2025 · ec6cec9 · ec6cec9
1 parent 411b653
commit ec6cec9
Show file tree

Hide file tree

Showing 4 changed files with 352 additions and 15 deletions.
diff --git a/third_party/xla/xla/backends/gpu/runtime/BUILD b/third_party/xla/xla/backends/gpu/runtime/BUILD
@@ -198,6 +198,7 @@ cc_library(
         "//xla:literal",
         "//xla:shape_util",
         "//xla:status_macros",
+        "//xla/hlo/evaluator:hlo_evaluator",
         "//xla/service:buffer_assignment",
         "//xla/service/gpu:buffer_allocations",
         "//xla/service/gpu:ir_emission_utils",
@@ -243,10 +244,12 @@ xla_test(
         "//xla/ffi:ffi_api",
         "//xla/service:buffer_assignment",
         "//xla/service:executable",
+        "//xla/service:hlo_runner",
         "//xla/service:platform_util",
         "//xla/service/gpu:buffer_allocations",
         "//xla/service/gpu:launch_dimensions",
         "//xla/service/gpu:matmul_utils",
+        "//xla/service/gpu:resource_requests",
         "//xla/stream_executor:blas",
         "//xla/stream_executor:command_buffer",
         "//xla/stream_executor:device_memory",
@@ -258,6 +261,7 @@ xla_test(
         "//xla/stream_executor:stream_executor_memory_allocator",
         "//xla/stream_executor/gpu:gpu_test_kernels",
         "//xla/stream_executor/gpu:gpu_types_header",
+        "//xla/tests:hlo_runner_agnostic_test_base",
         "//xla/tsl/lib/core:status_test_util",
         "@com_google_absl//absl/algorithm:container",
         "@com_google_absl//absl/status",

diff --git a/third_party/xla/xla/backends/gpu/runtime/dynamic_slice_thunk.cc b/third_party/xla/xla/backends/gpu/runtime/dynamic_slice_thunk.cc
@@ -31,6 +31,7 @@ limitations under the License.
 #include "xla/backends/gpu/runtime/sequential_thunk.h"
 #include "xla/backends/gpu/runtime/thunk.h"
 #include "xla/backends/gpu/runtime/while_thunk.h"
+#include "xla/hlo/evaluator/hlo_evaluator.h"
 #include "xla/service/buffer_assignment.h"
 #include "xla/service/gpu/buffer_allocations.h"
 #include "xla/service/gpu/ir_emission_utils.h"
@@ -47,14 +48,35 @@ limitations under the License.
 namespace xla {
 namespace gpu {
 
+namespace {
+
+// Indvar is a thread-local map that stores the induction variable for each
+// dynamic slice thunk. The same thunk object in the memory is shared by
+// multiple replicas of the same computation. So, each replica should have its
+// own tracking of the induction variable (threadlocal). With threadlocal, we
+// cannot embed this inside the dynamic slice thunk object, and so we have a
+// static map. There could be multiple dynamic slice thunks in the same module,
+// and so we need a map to store the induction variable for each thunk. The
+// usage of threadlocal in this context is similar to `LoopCounters` in
+// while_thunk.cc (b/343294327).
+Literal& Indvar(DynamicSliceThunk* thunk) {
+  static thread_local absl::flat_hash_map<DynamicSliceThunk*, Literal>
+      indvar_map;
+  return indvar_map[thunk];
+}
+
+}  // namespace
+
 DynamicSliceThunk::DynamicSliceThunk(
     ThunkInfo thunk_info, std::unique_ptr<ThunkSequence> embedded_thunk,
     std::vector<std::optional<BufferAllocation::Slice>> arguments,
     std::vector<std::unique_ptr<BufferAllocation>> fake_allocations,
     std::vector<std::optional<std::vector<Offset>>> offsets,
     std::vector<std::optional<Shape>> orig_shapes,
     std::vector<std::optional<Shape>> sliced_shapes,
-    std::vector<std::optional<uint64_t>> offset_byte_sizes)
+    std::vector<std::optional<uint64_t>> offset_byte_sizes,
+    std::optional<OffsetAsFunctionOfIndvarModulesMetadata>
+        offset_as_function_of_indvar_metadata)
     : Thunk(Kind::kDynamicSlice, thunk_info),
       embedded_thunk_(std::make_unique<SequentialThunk>(
           ThunkInfo(), std::move(*embedded_thunk))),
@@ -63,7 +85,9 @@ DynamicSliceThunk::DynamicSliceThunk(
       offsets_(offsets),
       orig_shapes_(orig_shapes),
       sliced_shapes_(sliced_shapes),
-      offset_byte_sizes_(offset_byte_sizes) {
+      offset_byte_sizes_(offset_byte_sizes),
+      offset_as_function_of_indvar_metadata_(
+          std::move(offset_as_function_of_indvar_metadata)) {
   // Zip all arguments together to create a list of SliceDef.
   for (auto [arg, offsets, orig_shape, sliced_shape, offset_byte_size] :
        llvm::zip_equal(arguments, offsets, orig_shapes, sliced_shapes,
@@ -105,6 +129,16 @@ absl::Status DynamicSliceThunk::Prepare(
   }
 
   TF_RETURN_IF_ERROR(embedded_thunk_->Prepare(params, resource_requests));
+
+  if (offset_as_function_of_indvar_metadata_ != std::nullopt) {
+    Indvar(this) =
+        HloEvaluator()
+            .Evaluate(
+                /*module=*/*offset_as_function_of_indvar_metadata_->indvar_init,
+                /*arg_literals=*/{})
+            .value();
+    VLOG(2) << "Indvar = " << Indvar(this).ToString();
+  }
   return absl::OkStatus();
 }
 
@@ -182,6 +216,20 @@ absl::Status DynamicSliceThunk::ExecuteOnStream(const ExecuteParams& params) {
                 << "]: constant offset = " << *const_offset;
         offset_value(argument_idx, offset_idx) = *const_offset;
 
+      } else if (HloModule** offset_module = std::get_if<HloModule*>(&offset)) {
+        TF_ASSIGN_OR_RETURN(
+            Literal offset,
+            HloEvaluator().Evaluate(**offset_module, {&Indvar(this)}));
+        auto offset_int = LiteralUtil::LiteralAsScalarInt64(offset);
+        if (offset_int.has_value()) {
+          offset_value(argument_idx, offset_idx) = *offset_int;
+        } else {
+          return absl::InternalError(
+              absl::StrFormat("Unhandled type returned from offset module: %s",
+                              offset.shape().ToString()));
+        }
+        VLOG(2) << "Offset value = " << offset_value(argument_idx, offset_idx);
+
       } else {
         // Transfer slice offset value from device to host.
         auto alloc_slice = std::get<BufferAllocation::Slice>(offset);
@@ -252,6 +300,15 @@ absl::Status DynamicSliceThunk::ExecuteOnStream(const ExecuteParams& params) {
   // Execute the underlying custom call thunk with the new buffers.
   TF_RETURN_IF_ERROR(embedded_thunk_->ExecuteOnStream(new_params));
 
+  if (offset_as_function_of_indvar_metadata_ != std::nullopt) {
+    Indvar(this) =
+        HloEvaluator()
+            .Evaluate(*offset_as_function_of_indvar_metadata_->indvar_update,
+                      {&Indvar(this)})
+            .value();
+    VLOG(2) << "Indvar = " << Indvar(this).ToString();
+  }
+
   return absl::OkStatus();
 }
 

diff --git a/third_party/xla/xla/backends/gpu/runtime/dynamic_slice_thunk.h b/third_party/xla/xla/backends/gpu/runtime/dynamic_slice_thunk.h
@@ -32,6 +32,7 @@ limitations under the License.
 #include "xla/literal.h"
 #include "xla/service/buffer_assignment.h"
 #include "xla/shape.h"
+#include "xla/shape_util.h"
 #include "xla/stream_executor/memory_allocation.h"
 #include "xla/stream_executor/stream_executor.h"
 
@@ -47,8 +48,59 @@ class DynamicSliceThunk : public Thunk {
  public:
   // Dynamic slice offset can be either: (1) a statically known constant value
   // or (2) a truly dynamic offset that is computed on device and have to be
-  // transferred to host.
-  using Offset = std::variant<int64_t, BufferAllocation::Slice>;
+  // transferred to host or (3) a temporary HloModule that computes the offset
+  // with a single induction variable as the input.
+  using Offset = std::variant<int64_t, BufferAllocation::Slice, HloModule*>;
+
+  struct OffsetAsFunctionOfIndvarModulesMetadata {
+    // These two modules help keep track of the induction variable. The module
+    // `indvar_init_` is a module with prototype `() -> integer[]`. It takes
+    // no input, and returns the initial value of the induction variable. The
+    // module `indvar_update_` is a module with prototype `(integer[]) ->
+    // integer[]`. It takes the current value of the induction variable, and
+    // returns the next value of the induction variable.
+    std::unique_ptr<HloModule> indvar_init, indvar_update;
+
+    // Extracted HloModules for computing dynamic offsets. The modules are
+    // not used here, this is solely for keeping the modules alive and maintain
+    // their ownership with the thunk while their raw pointers would be used
+    // during execution from the `offsets_` vector. These modules are with
+    // signature `(integer[]) -> integer[]`, where the input is the current
+    // value of the loop induction variable, and the output is the offset value
+    // for that iteration.
+    std::vector<std::unique_ptr<HloModule>> extracted_offset_modules;
+
+    OffsetAsFunctionOfIndvarModulesMetadata(
+        std::unique_ptr<HloModule> indvar_init,
+        std::unique_ptr<HloModule> indvar_update,
+        std::vector<std::unique_ptr<HloModule>> extracted_offset_modules)
+        : indvar_init(std::move(indvar_init)),
+          indvar_update(std::move(indvar_update)),
+          extracted_offset_modules(std::move(extracted_offset_modules)) {
+      CHECK(this->indvar_init != nullptr && this->indvar_update != nullptr);
+      Shape init_output_shape =
+          this->indvar_init->entry_computation()->root_instruction()->shape();
+      CHECK(this->indvar_init->entry_computation()->num_parameters() == 0 &&
+            init_output_shape.IsInteger() &&
+            ShapeUtil::IsScalar(init_output_shape))
+          << "Induction variable init module expected with signature `() -> "
+             "integer[]`.";
+      Shape update_output_shape =
+          this->indvar_update->entry_computation()->root_instruction()->shape();
+      CHECK(this->indvar_update->entry_computation()->num_parameters() == 1 &&
+            update_output_shape.IsInteger() &&
+            ShapeUtil::IsScalar(update_output_shape))
+          << "Induction variable update module expected with signature "
+             "`(integer[]) -> integer[]`.";
+      Shape update_input_shape = this->indvar_update->entry_computation()
+                                     ->parameter_instruction(0)
+                                     ->shape();
+      CHECK(ShapeUtil::IsScalar(update_input_shape) &&
+            update_input_shape.IsInteger())
+          << "Induction variable update module expected with signature "
+             "`(integer[]) -> integer[]`.";
+    }
+  };
 
   DynamicSliceThunk(
       ThunkInfo thunk_info, std::unique_ptr<ThunkSequence> embedded_thunk,
@@ -57,8 +109,9 @@ class DynamicSliceThunk : public Thunk {
       std::vector<std::optional<std::vector<Offset>>> offsets,
       std::vector<std::optional<Shape>> orig_shapes,
       std::vector<std::optional<Shape>> sliced_shapes,
-      std::vector<std::optional<uint64_t>> offset_byte_sizes);
-
+      std::vector<std::optional<uint64_t>> offset_byte_sizes,
+      std::optional<OffsetAsFunctionOfIndvarModulesMetadata>
+          offset_as_function_of_indvar_metadata = std::nullopt);
   DynamicSliceThunk(const DynamicSliceThunk&) = delete;
   DynamicSliceThunk& operator=(const DynamicSliceThunk&) = delete;
 
@@ -132,6 +185,12 @@ class DynamicSliceThunk : public Thunk {
 
   // A mapping from argument index to the base offset in the `offsets_allocs_`.
   std::vector<int64_t> offsets_allocs_base_;
+
+  // This structure holds the metadata for offset computations on host. It
+  // stores a single induction variable initialization module, its update module
+  // and the offsets that are a function of the induction variable.
+  std::optional<OffsetAsFunctionOfIndvarModulesMetadata>
+      offset_as_function_of_indvar_metadata_;
 };
 
 }  // namespace gpu