1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
|
// Copyright (c) Meta Platforms, Inc. and affiliates.
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <caffe2/torch/csrc/jit/backends/xnnpack/xnnpack_graph_builder.h>
#include <torch/csrc/jit/runtime/graph_iterator.h>
#include <xnnpack.h>
// graph passes
#include <torch/csrc/jit/passes/constant_propagation.h>
#include <torch/csrc/jit/passes/dead_code_elimination.h>
#include <torch/csrc/jit/passes/frozen_graph_optimizations.h>
#include <torch/csrc/jit/passes/lower_tuples.h>
#include <torch/csrc/jit/passes/remove_mutation.h>
#include <torch/csrc/jit/runtime/jit_trace.h>
#include <torch/csrc/jit/tensorexpr/graph_opt.h>
namespace torch {
namespace jit {
namespace xnnpack {
namespace delegate {
std::shared_ptr<torch::jit::Graph> XNNGraph::optimizeAndTraceGraph(
std::shared_ptr<torch::jit::Graph> graph,
std::vector<c10::IValue>& example_inputs) {
OptimizeFrozenGraph(graph, true);
RemoveListMutation(graph);
RemoveTensorMutation(graph);
LowerAllTuples(graph);
ConstantPropagation(graph);
graph = TraceGraph(graph, example_inputs);
return graph;
}
void XNNGraph::buildXNNGraph(
std::shared_ptr<torch::jit::Graph>& graph,
std::vector<c10::IValue> example_inputs) {
graph = optimizeAndTraceGraph(graph, example_inputs);
checkOpsToDelegate(graph);
gatherTensorValues(graph);
// count unique input/outputs (some inputs can be outputs)
std::unordered_set<torch::jit::Value*> externals;
for (auto inp : _inputs) {
externals.insert(inp);
}
for (auto out : _outputs) {
externals.insert(out);
}
// create subgraph
xnn_status status = xnn_create_subgraph(
/*external_value_ids=*/externals.size(),
/*flags=*/0,
&_subgraph_ptr);
TORCH_CHECK(xnn_status_success == status, "Failed to create xnn subgraph");
defineAllTensorValues();
defineAllNodes(graph);
// at this point graph is complete, for the sake of testing preprocess at
// this point we will do runtime setup and run with some default values
}
void XNNGraph::runGraphOnInputs(
std::vector<at::Tensor> tensor_inputs,
std::vector<at::Tensor> tensor_outputs) {
TORCH_CHECK(
_subgraph_ptr != nullptr,
"run buildXNNGraph before running graph on inputs");
xnn_runtime_t runtime = nullptr;
xnn_status status =
xnn_create_runtime_v2(_subgraph_ptr, nullptr, /*flags=*/0, &runtime);
TORCH_CHECK(
xnn_status_success == status,
"failed to create runtime for running inputs");
// smart pointer for runtime
std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(
runtime, xnn_delete_runtime);
std::vector<xnn_external_value> external_values;
TORCH_CHECK(
tensor_inputs.size() == _inputs.size(),
"supplied inputs does not match expected inputs");
for (int i = 0; i < tensor_inputs.size(); i++) {
external_values.push_back(
{_val_to_ids[_inputs[i]], tensor_inputs[i].data_ptr<float>()});
}
TORCH_CHECK(
tensor_outputs.size() == _outputs.size(),
"supplied outputs does not match expected outputs");
for (int i = 0; i < tensor_outputs.size(); i++) {
external_values.push_back(
{_val_to_ids[_outputs[i]], tensor_outputs[i].data_ptr<float>()});
}
status = xnn_setup_runtime(
auto_runtime.get(), external_values.size(), external_values.data());
TORCH_CHECK(xnn_status_success == status, "runtime not properly setup");
TORCH_CHECK(xnn_status_success == xnn_invoke_runtime(auto_runtime.get()));
}
void XNNGraph::checkOpsToDelegate(std::shared_ptr<torch::jit::Graph>& graph) {
std::unordered_set<string> unsupported_ops;
DepthFirstGraphNodeIterator it(graph);
Node* node = nullptr;
while ((node = it.next()) != nullptr) {
switch (node->kind()) {
case prim::Constant:
case aten::add: {
break;
}
default: {
unsupported_ops.insert(node->kind().toDisplayString());
}
}
}
std::stringstream error;
for (auto itr = unsupported_ops.begin(); itr != unsupported_ops.end();
itr++) {
error << *itr << std::endl;
;
}
TORCH_CHECK(
unsupported_ops.empty(),
"the module contains the following unsupported ops:\n" + error.str());
}
std::string XNNGraph::serializedXNNGraph() {
std::vector<uint32_t> input_ids;
std::vector<uint32_t> output_ids;
std::unordered_set<uint32_t> num_externs;
for (auto val : _inputs) {
input_ids.push_back(_val_to_ids[val]);
num_externs.emplace(_val_to_ids[val]);
}
for (auto val : _outputs) {
output_ids.push_back(_val_to_ids[val]);
num_externs.emplace(_val_to_ids[val]);
}
return _serializer.finishAndSerialize(
input_ids, output_ids, num_externs.size());
}
std::vector<std::vector<long>> XNNGraph::getGraphOutputShapes() {
std::vector<std::vector<long>> output_shapes;
for (auto val : _outputs) {
auto tensor_ptr = val->type()->cast<TensorType>();
std::vector<long> sizes = tensor_ptr->sizes().concrete_sizes().value();
output_shapes.push_back(sizes);
}
return output_shapes;
}
void XNNGraph::defineAllNodes(std::shared_ptr<torch::jit::Graph>& graph) {
DepthFirstGraphNodeIterator it(graph);
Node* node = nullptr;
while ((node = it.next()) != nullptr) {
switch (node->kind()) {
case prim::Constant: {
break;
}
case aten::add: {
// todo: handle alpha for aten::add
uint32_t input1_id = _val_to_ids[node->inputs()[0]];
uint32_t input2_id = _val_to_ids[node->inputs()[1]];
TORCH_CHECK(
node->inputs()[2]->type()->cast<IntType>() == 1,
"non-1 alpha values not supported");
uint32_t output_id = _val_to_ids[node->outputs()[0]];
xnn_status status = xnn_define_add2(
_subgraph_ptr,
output_min,
output_max,
input1_id,
input2_id,
output_id,
/*flags=*/0);
_serializer.serializeAddNode(input1_id, input2_id, output_id, 0);
TORCH_CHECK(status == xnn_status_success, "failed to create add node");
break;
}
default: {
throw std::exception();
TORCH_CHECK(
false,
"The node of ",
node->kind().toQualString(),
" is not supported yet");
break;
}
}
}
}
void XNNGraph::defineAllTensorValues() {
uint32_t external_id =
std::numeric_limits<decltype(XNN_INVALID_VALUE_ID)>::min();
for (auto val : _intermediate_tensors) {
if (_val_to_ids.find(val) == _val_to_ids.end()) {
uint32_t id = XNN_INVALID_VALUE_ID;
// cast value to tensortype
auto tensor_ptr = val->type()->cast<TensorType>();
auto num_dims = tensor_ptr->dim().value();
// create size_t* for tensor shape, casting must be done from long ->
// size_t
std::vector<long> sizes = tensor_ptr->sizes().concrete_sizes().value();
std::vector<size_t> tensor_shape;
tensor_shape.reserve(sizes.size());
for (auto dim : sizes) {
TORCH_CHECK(dim >= 0, "Input Dims should be unsigned");
tensor_shape.push_back(static_cast<size_t>(dim));
}
// ext_id value
uint32_t ext_id = XNN_INVALID_VALUE_ID;
// update flag for if tensor is either graph input/output
uint32_t flags = 0;
// Check if value was produced by prim::Constant
void* value_data = nullptr;
size_t buffer_idx = 0;
size_t num_bytes = 0;
if (val->node()->kind() == prim::Constant) {
std::optional<IValue> constant = val->node()->t(attr::value);
auto const_val = constant->toIValue().toTensor();
// Need tensor data to be contiguous for serialization
auto cont_const_val = const_val.contiguous();
value_data = cont_const_val.data_ptr();
num_bytes = const_val.storage().nbytes();
buffer_idx = _serializer.serializeData(
static_cast<const uint8_t*>(value_data), num_bytes);
}
if (isGraphInput(val) || isGraphOutput(val)) {
if (isGraphInput(val)) {
flags |= XNN_VALUE_FLAG_EXTERNAL_INPUT;
}
if (isGraphOutput(val)) {
flags |= XNN_VALUE_FLAG_EXTERNAL_OUTPUT;
}
ext_id = external_id++;
}
xnn_status status = xnn_define_tensor_value(
/*subgraph=*/_subgraph_ptr,
/*datatype=*/xnn_datatype_fp32,
/*num_dims=*/num_dims,
/*dims=*/tensor_shape.data(),
/*data=*/value_data,
/*external_id=*/ext_id,
/*flags=*/flags,
/*id_out=*/&id);
TORCH_CHECK(
status == xnn_status_success,
"failed to define xnn_tensor_id for: " + val->debugName());
_serializer.serializeTensorValue(
xnn_datatype_fp32,
num_dims,
tensor_shape,
buffer_idx,
ext_id,
flags,
id);
_val_to_ids.insert({val, id});
}
}
}
void XNNGraph::gatherTensorValues(std::shared_ptr<torch::jit::Graph>& graph) {
for (auto input : graph->inputs()) {
if (input->isCompleteTensor()) {
_intermediate_tensors.insert(input);
_inputs.push_back(input);
}
}
DepthFirstGraphNodeIterator it(graph);
Node* n = nullptr;
while ((n = it.next()) != nullptr) {
gatherNodeInputs(*n);
}
for (auto output : graph->outputs()) {
if (output->isCompleteTensor()) {
_intermediate_tensors.insert(output);
_outputs.push_back(output);
}
}
}
void XNNGraph::gatherNodeInputs(torch::jit::Node& node) {
switch (node.kind()) {
case aten::add: {
// this case will support all ops with only two inputs i.e. sub, add,
for (auto value : node.inputs()) {
if (value->isCompleteTensor()) {
_intermediate_tensors.insert(value);
}
}
}
}
}
bool XNNGraph::isGraphInput(torch::jit::Value* val) {
return std::count(_inputs.begin(), _inputs.end(), val) > 0;
};
bool XNNGraph::isGraphOutput(torch::jit::Value* val) {
return std::count(_outputs.begin(), _outputs.end(), val) > 0;
};
} // namespace delegate
} // namespace xnnpack
} // namespace jit
} // namespace torch
|