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
|
#include <torch/csrc/jit/codegen/cuda/lower2device.h>
#include <torch/csrc/jit/codegen/cuda/fusion.h>
#include <torch/csrc/jit/codegen/cuda/instrumentation.h>
#include <torch/csrc/jit/codegen/cuda/ir_iostream.h>
#include <torch/csrc/jit/codegen/cuda/lower_index.h>
#include <torch/csrc/jit/codegen/cuda/lower_insert_syncs.h>
#include <torch/csrc/jit/codegen/cuda/lower_loops.h>
#include <torch/csrc/jit/codegen/cuda/lower_thread_predicate.h>
#include <torch/csrc/jit/codegen/cuda/lower_unroll.h>
#include <torch/csrc/jit/codegen/cuda/lower_utils.h>
#include <torch/csrc/jit/codegen/cuda/lower_validation.h>
namespace torch {
namespace jit {
namespace fuser {
// TODO(kir): revisit this
thread_local GpuLower* active_gpu_lower = nullptr;
void GpuLower::replaceSymbolicSizes() {
FUSER_PERF_SCOPE("replaceSymbolicSizes");
kir::IrBuilder ir_builder(kernel());
// Grab inputs and outputs
// TODO: Only run through inputs for the size map, outputs don't actually set
// any sizes of the problem.
std::vector<TensorView*> inputs_and_outputs;
for (auto val : fusion_->inputs()) {
if (ir_utils::isTV(val)) {
inputs_and_outputs.push_back(val->as<TensorView>());
}
}
for (auto val : fusion_->outputs()) {
if (ir_utils::isTV(val)) {
inputs_and_outputs.push_back(val->as<TensorView>());
}
}
// Run through inputs and outputs first. Since we're replacing full
// tensorviews their names are going to change. We need the new referenc
// name for the inputs/outputs. This way we won't reference the wrong tensor
// view. For example T0 may be translated to T9. We don't want our new
// variable to be T0->size[...] we need it to be T9->size[...]
for (TensorView* tv : inputs_and_outputs) {
// Replace the domain with one based on Ti.size[j]
std::vector<IterDomain*> new_domain_iters;
const std::vector<IterDomain*>& root_td = tv->getRootDomain();
size_t dim = 0;
for (auto id : root_td) {
const Val* orig_size = id->extent();
// Output sizes could have reduction axes, which isn't what gets output.
if (id->isReduction()) {
continue;
} else if (id->getIterType() == IterType::BroadcastWithoutStride) {
continue;
} else if (id->getIterType() == IterType::BroadcastWithStride) {
dim++;
continue;
} else if (orig_size->isConstScalar()) {
dim++;
continue;
}
// TODO(kir): consider a different implementation which doesn't
// hijack the kir_map_
if (kir_map_.find(orig_size) == kir_map_.end()) {
std::stringstream ss;
ss << "T" << tv->name() << ".size[" << dim++ << "]";
kir_map_[orig_size] = ir_builder.create<kir::NamedScalar>(
ss.str(), orig_size->getDataType().value());
}
}
}
}
void GpuLower::lower() {
FUSER_PERF_SCOPE("lower");
TORCH_INTERNAL_ASSERT(fusion_ != nullptr);
TORCH_INTERNAL_ASSERT(
active_gpu_lower == nullptr, "Nested lowering passes are not supported");
// TODO(kir): revisit this
struct LowerGuard {
LowerGuard(GpuLower* gpu_lower) {
active_gpu_lower = gpu_lower;
}
~LowerGuard() {
active_gpu_lower = nullptr;
}
} lower_guard(this);
FusionGuard fg(fusion_);
// Start with a fresh kernel
kernel_ = std::make_unique<Kernel>();
// prepare for lowering
validateIr(fusion_);
replaceSymbolicSizes();
// Compute thread predicates
ThreadPredicateMap preds(fusion_);
// Run our passes keeping the lowered expressions and forwarding them
const auto lowered_exprs =
LoopNestGenerator::loweredExprs(fusion_, preds, fusion_->exprs(true));
const auto unrolled_loops =
UnrollPass::runPass(fusion_, lowered_exprs, preds);
// Insert SyncThreads at end of for-loop to avoid WAR race condition
const auto sync_exprs = insertThreadSynchronization(fusion_, unrolled_loops);
const auto indexed_loops =
IndexLowering::getIndexedExprs(fusion_, sync_exprs);
// We now have the lowered expressions, finalize the kernel IR
kernel_->finalize(indexed_loops, preds);
// Set the kernel inputs & outputs
for (auto input : fusion_->inputs()) {
kernel_->addInput(GpuLower::lowerValue(input));
}
for (auto output : fusion_->outputs()) {
kernel_->addOutput(GpuLower::lowerValue(output));
}
}
Kernel* GpuLower::kernel() const {
TORCH_CHECK(kernel_);
return kernel_.get();
}
// Maps Fusion IR nodes to the Kernel IR counterparts
//
// TODO(kir): this is a interim solution for easing the Kernel IR splitting
//
class TORCH_CUDA_API GpuLower::KernelIrMapper : private OptInConstDispatch {
public:
explicit KernelIrMapper(GpuLower* gpu_lower)
: gpu_lower_(gpu_lower), ir_builder_(gpu_lower->kernel()) {}
Val* lower(const Val* value) {
const auto it = gpu_lower_->kir_map_.find(value);
if (it != gpu_lower_->kir_map_.end()) {
return it->second;
} else {
handle(value);
const auto lowered_node = gpu_lower_->kir_map_[value];
TORCH_CHECK(lowered_node != nullptr);
TORCH_CHECK(kir::isLoweredVal(lowered_node));
// Lower the arithmetic expression defining the value, if any
if (value->isScalar()) {
if (auto def = value->getOrigin()) {
lowerDefinition(lowered_node, def);
}
}
return lowered_node;
}
}
private:
// TODO(kir): rewrite this
void lowerDefinition(Val* lowered_value, const Expr* def) {
switch (def->type()) {
case ExprType::UnaryOp: {
const auto op = def->as<fuser::UnaryOp>();
ir_builder_.create<kir::UnaryOp>(
op->getUnaryOpType(), lowered_value, lower(op->in()));
break;
}
case ExprType::BinaryOp: {
const auto op = def->as<fuser::BinaryOp>();
ir_builder_.create<kir::BinaryOp>(
op->getBinaryOpType(),
lowered_value,
lower(op->lhs()),
lower(op->rhs()));
break;
}
case ExprType::TernaryOp: {
const auto op = def->as<fuser::TernaryOp>();
ir_builder_.create<kir::TernaryOp>(
op->getTernaryOpType(),
lowered_value,
lower(op->in1()),
lower(op->in2()),
lower(op->in3()));
break;
}
default:
TORCH_CHECK(false, "Unexpected expression type");
}
}
void handle(const Statement* node) override {
OptInConstDispatch::handle(node);
}
void handle(const Val* node) override {
OptInConstDispatch::handle(node);
}
void handle(const Expr* node) override {
OptInConstDispatch::handle(node);
}
void handle(const TensorDomain* node) override {
const auto lowered_node = ir_builder_.create<kir::TensorDomain>(node);
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
void handle(const IterDomain* node) override {
const auto lowered_node = ir_builder_.create<kir::IterDomain>(node);
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
void handle(const TensorView* node) override {
const auto lowered_node = ir_builder_.create<kir::TensorView>(node);
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
void handle(const Bool* node) override {
const auto lowered_node = ir_builder_.create<kir::Bool>(node);
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
void handle(const Float* node) override {
const auto lowered_node = ir_builder_.create<kir::Float>(node);
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
void handle(const Half* node) override {
const auto lowered_node = ir_builder_.create<kir::Half>(node);
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
void handle(const Int* node) override {
const auto lowered_node = ir_builder_.create<kir::Int>(node, false);
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
void handle(const NamedScalar* node) override {
const auto lowered_node = ir_builder_.create<kir::NamedScalar>(
node->name(), node->getDataType().value());
TORCH_CHECK(gpu_lower_->kir_map_.insert({node, lowered_node}).second);
}
private:
GpuLower* gpu_lower_ = nullptr;
kir::IrBuilder ir_builder_;
};
Val* GpuLower::lowerValue(const Val* val) {
TORCH_INTERNAL_ASSERT(!kir::isLoweredVal(val));
TORCH_INTERNAL_ASSERT(active_gpu_lower != nullptr);
KernelIrMapper kir_mapper(active_gpu_lower);
return kir_mapper.lower(val);
}
Val* GpuLower::getLowerValue(const Val* val) {
KernelIrMapper kir_mapper(this);
return kir_mapper.lower(val);
}
GpuLower* GpuLower::current() {
TORCH_INTERNAL_ASSERT(active_gpu_lower != nullptr);
return active_gpu_lower;
}
} // namespace fuser
} // namespace jit
} // namespace torch
|