feat: move the acc case to a plugin that calls the libtorch kernel

Author: narendasan

core/plugins/BUILD

@@ -66,6 +66,7 @@ cc_library(
         "//conditions:default": [
             "impl/interpolate_plugin.cpp",
             "impl/normalize_plugin.cpp",
+            "impl/scatter_add_plugin.cpp",
             "register_plugins.cpp",
         ],
     }),
@@ -75,6 +76,7 @@ cc_library(
         "//conditions:default": [
             "impl/interpolate_plugin.h",
             "impl/normalize_plugin.h",
+            "impl/scatter_add_plugin.h",
             "plugins.h",
         ],
     }),
@@ -132,6 +134,7 @@ filegroup(
     srcs = [
         "impl/interpolate_plugin.h",
         "impl/normalize_plugin.h",
+        "impl/scatter_add_plugin.h",
     ],
     visibility = ["//visibility:public"],
 )

core/plugins/CMakeLists.txt

@@ -4,6 +4,7 @@ add_library(${lib_name} OBJECT)
 target_sources(${lib_name}
     PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}/impl/interpolate_plugin.cpp"
             "${CMAKE_CURRENT_SOURCE_DIR}/impl/normalize_plugin.cpp"
+            "${CMAKE_CURRENT_SOURCE_DIR}/impl/scatter_add_plugin.cpp"
             "${CMAKE_CURRENT_SOURCE_DIR}/register_plugins.cpp"
     PUBLIC  $<TARGET_OBJECTS:core_util>
 )
@@ -33,6 +34,7 @@ install(
     FILES
         "${CMAKE_CURRENT_SOURCE_DIR}/impl/interpolate_plugin.h"
         "${CMAKE_CURRENT_SOURCE_DIR}/impl/normalize_plugin.h"
+        "${CMAKE_CURRENT_SOURCE_DIR}/impl/scatter_add_plugin.h"
     DESTINATION
         "${CMAKE_INSTALL_INCLUDEDIR}/torch_tensorrt/core/plugins/impl"
 )

core/plugins/impl/scatter_add_plugin.cpp

@@ -0,0 +1,271 @@
+#include "core/plugins/impl/scatter_add_plugin.h"
+#include "core/plugins/plugins.h"
+#include "core/util/prelude.h"
+
+#include <ATen/cuda/CUDAContext.h>
+#include <ATen/cuda/CUDAEvent.h>
+#include <c10/cuda/CUDAStream.h>
+
+namespace torch_tensorrt {
+namespace core {
+namespace plugins {
+namespace impl {
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+
+// ---------------------------------------------------------------------------
+// ScatterAddPlugin — construction
+// ---------------------------------------------------------------------------
+
+ScatterAddPlugin::ScatterAddPlugin() = default;
+
+// ---------------------------------------------------------------------------
+// IPluginV3
+// ---------------------------------------------------------------------------
+
+nvinfer1::IPluginCapability* ScatterAddPlugin::getCapabilityInterface(
+    nvinfer1::PluginCapabilityType type) noexcept {
+  switch (type) {
+    case nvinfer1::PluginCapabilityType::kCORE:
+      return static_cast<nvinfer1::IPluginV3OneCore*>(this);
+    case nvinfer1::PluginCapabilityType::kBUILD:
+      return static_cast<nvinfer1::IPluginV3OneBuild*>(this);
+    case nvinfer1::PluginCapabilityType::kRUNTIME:
+      return static_cast<nvinfer1::IPluginV3OneRuntime*>(this);
+    default:
+      return nullptr;
+  }
+}
+
+nvinfer1::IPluginV3* ScatterAddPlugin::clone() noexcept {
+  return new ScatterAddPlugin(*this);
+}
+
+// ---------------------------------------------------------------------------
+// IPluginV3OneCore
+// ---------------------------------------------------------------------------
+
+const char* ScatterAddPlugin::getPluginName() const noexcept {
+  return "ScatterAdd";
+}
+
+const char* ScatterAddPlugin::getPluginVersion() const noexcept {
+  return "1";
+}
+
+const char* ScatterAddPlugin::getPluginNamespace() const noexcept {
+  return "torch_tensorrt";
+}
+
+// ---------------------------------------------------------------------------
+// IPluginV3OneBuild
+// ---------------------------------------------------------------------------
+
+int32_t ScatterAddPlugin::getNbOutputs() const noexcept {
+  return 1;
+}
+
+int32_t ScatterAddPlugin::getOutputDataTypes(
+    nvinfer1::DataType* outputTypes,
+    int32_t nbOutputs,
+    const nvinfer1::DataType* inputTypes,
+    int32_t nbInputs) const noexcept {
+  // Output has the same dtype as src (input 0).
+  outputTypes[0] = inputTypes[0];
+  return 0;
+}
+
+int32_t ScatterAddPlugin::getOutputShapes(
+    const nvinfer1::DimsExprs* inputs,
+    int32_t nbInputs,
+    const nvinfer1::DimsExprs* /*shapeInputs*/,
+    int32_t /*nbShapeInputs*/,
+    nvinfer1::DimsExprs* outputs,
+    int32_t /*nbOutputs*/,
+    nvinfer1::IExprBuilder& /*exprBuilder*/) noexcept {
+  // Output shape == src shape (input 0).
+  outputs[0] = inputs[0];
+  return 0;
+}
+
+bool ScatterAddPlugin::supportsFormatCombination(
+    int32_t pos,
+    const nvinfer1::DynamicPluginTensorDesc* inOut,
+    int32_t nbInputs,
+    int32_t nbOutputs) noexcept {
+  const auto& desc = inOut[pos];
+
+  // All tensors must be row-major (linear) layout.
+  if (desc.desc.format != nvinfer1::TensorFormat::kLINEAR) {
+    return false;
+  }
+
+  // Positions 1 through nbInputs-2 are index tensors: int32 or int64.
+  if (pos >= 1 && pos <= nbInputs - 2) {
+    return desc.desc.type == nvinfer1::DataType::kINT32 ||
+           desc.desc.type == nvinfer1::DataType::kINT64;
+  }
+
+  // pos 0 (src), pos nbInputs-1 (values), pos nbInputs (output):
+  // float32 / float16 / bfloat16, all sharing the same type.
+  const bool float_type = desc.desc.type == nvinfer1::DataType::kFLOAT ||
+                          desc.desc.type == nvinfer1::DataType::kHALF ||
+                          desc.desc.type == nvinfer1::DataType::kBF16;
+  if (!float_type) {
+    return false;
+  }
+
+  // src, values and output must have the same dtype.
+  if (pos == 0) {
+    return true;
+  }
+  return desc.desc.type == inOut[0].desc.type;
+}
+
+int32_t ScatterAddPlugin::configurePlugin(
+    const nvinfer1::DynamicPluginTensorDesc* in,
+    int32_t nbInputs,
+    const nvinfer1::DynamicPluginTensorDesc* /*out*/,
+    int32_t /*nbOutputs*/) noexcept {
+  dtype_ = in[0].desc.type;
+  n_indices_ = nbInputs - 2;  // exclude src and values
+  idx_dtypes_.resize(n_indices_);
+  for (int i = 0; i < n_indices_; ++i) {
+    idx_dtypes_[i] = in[1 + i].desc.type;
+  }
+  return 0;
+}
+
+size_t ScatterAddPlugin::getWorkspaceSize(
+    const nvinfer1::DynamicPluginTensorDesc* /*inputs*/,
+    int32_t /*nbInputs*/,
+    const nvinfer1::DynamicPluginTensorDesc* /*outputs*/,
+    int32_t /*nbOutputs*/) const noexcept {
+  return 0;
+}
+
+// ---------------------------------------------------------------------------
+// IPluginV3OneRuntime
+// ---------------------------------------------------------------------------
+
+int32_t ScatterAddPlugin::onShapeChange(
+    const nvinfer1::PluginTensorDesc* in,
+    int32_t nbInputs,
+    const nvinfer1::PluginTensorDesc* /*out*/,
+    int32_t /*nbOutputs*/) noexcept {
+  dtype_ = in[0].type;
+  n_indices_ = nbInputs - 2;
+  src_shape_ = util::toVec(in[0].dims);
+  val_shape_ = util::toVec(in[nbInputs - 1].dims);
+  idx_dtypes_.resize(n_indices_);
+  idx_shapes_.resize(n_indices_);
+  for (int i = 0; i < n_indices_; ++i) {
+    idx_dtypes_[i] = in[1 + i].type;
+    idx_shapes_[i] = util::toVec(in[1 + i].dims);
+  }
+  return 0;
+}
+
+int32_t ScatterAddPlugin::enqueue(
+    const nvinfer1::PluginTensorDesc* inputDesc,
+    const nvinfer1::PluginTensorDesc* outputDesc,
+    const void* const* inputs,
+    void* const* outputs,
+    void* /*workspace*/,
+    cudaStream_t stream) noexcept {
+  const at::ScalarType float_dtype = util::TRTDataTypeToScalarType(dtype_);
+  const auto float_opts = at::TensorOptions().device(at::kCUDA).dtype(float_dtype);
+
+  at::Tensor src = at::from_blob(const_cast<void*>(inputs[0]), src_shape_, float_opts);
+  at::Tensor val = at::from_blob(
+      const_cast<void*>(inputs[n_indices_ + 1]), val_shape_, float_opts);
+
+  // Build the indices list — one entry per index tensor, all cast to int64
+  // as required by ATen's index_put kernel.
+  c10::List<std::optional<at::Tensor>> indices;
+  indices.reserve(n_indices_);
+  for (int i = 0; i < n_indices_; ++i) {
+    const at::ScalarType int_dtype = util::TRTDataTypeToScalarType(idx_dtypes_[i]);
+    const auto int_opts = at::TensorOptions().device(at::kCUDA).dtype(int_dtype);
+    at::Tensor idx = at::from_blob(
+        const_cast<void*>(inputs[1 + i]), idx_shapes_[i], int_opts);
+    indices.push_back(std::optional<at::Tensor>(idx.to(torch::kLong)));
+  }
+
+  // Use a separate PyTorch CUDA stream and synchronise with the TRT stream via
+  // CUDA events — same pattern as interpolate_plugin.cpp.
+  at::cuda::CUDAStream torch_stream = at::cuda::getStreamFromPool();
+  at::cuda::CUDAStreamGuard torch_guard(torch_stream);
+
+  cudaEvent_t start_event;
+  cudaEventCreate(&start_event);
+  cudaEventRecord(start_event, stream);
+  cudaStreamWaitEvent(torch_stream.stream(), start_event, 0);
+
+  // index_put with accumulate=true calls the atomicAdd-based CUDA kernel.
+  at::Tensor result = at::index_put(src, indices, val, /*accumulate=*/true);
+
+  at::Tensor out_t = at::from_blob(outputs[0], src_shape_, float_opts);
+  out_t.copy_(result);
+
+  cudaEvent_t done_event;
+  cudaEventCreate(&done_event);
+  cudaEventRecord(done_event, torch_stream.stream());
+  cudaStreamWaitEvent(stream, done_event, 0);
+
+  cudaEventDestroy(start_event);
+  cudaEventDestroy(done_event);
+
+  return 0;
+}
+
+nvinfer1::IPluginV3* ScatterAddPlugin::attachToContext(
+    nvinfer1::IPluginResourceContext* /*context*/) noexcept {
+  return clone();
+}
+
+const nvinfer1::PluginFieldCollection* ScatterAddPlugin::getFieldsToSerialize() noexcept {
+  // No configuration attributes to serialize — shapes and dtype are captured
+  // from the tensor descriptors at runtime.
+  return &empty_fc_;
+}
+
+// ---------------------------------------------------------------------------
+// ScatterAddPluginCreator
+// ---------------------------------------------------------------------------
+
+ScatterAddPluginCreator::ScatterAddPluginCreator() = default;
+
+const char* ScatterAddPluginCreator::getPluginName() const noexcept {
+  return "ScatterAdd";
+}
+
+const char* ScatterAddPluginCreator::getPluginVersion() const noexcept {
+  return "1";
+}
+
+const char* ScatterAddPluginCreator::getPluginNamespace() const noexcept {
+  return "torch_tensorrt";
+}
+
+const nvinfer1::PluginFieldCollection* ScatterAddPluginCreator::getFieldNames() noexcept {
+  return &fc_;
+}
+
+nvinfer1::IPluginV3* ScatterAddPluginCreator::createPlugin(
+    const char* /*name*/,
+    const nvinfer1::PluginFieldCollection* /*fc*/,
+    nvinfer1::TensorRTPhase /*phase*/) noexcept {
+  return new ScatterAddPlugin();
+}
+
+// Register with the torch_tensorrt namespace.
+REGISTER_TORCHTRT_PLUGIN(ScatterAddPluginCreator);
+
+} // namespace impl
+} // namespace plugins
+} // namespace core
+} // namespace torch_tensorrt