Merge triton-mlir
branch - Complete rewrite of the backend from scratch (#1004)
This PR merges the `triton-mlir` branch, in which we have been quietly rewriting the Triton backend from scratch to increase maintainability, stability and ultimately performance. Changes to the runtime are minimal, and this new version aims to remain backward-compatible with the previous commit. The legacy backend is now officially deprecated, but can still be accessed via the `legacy-backend` tag. Co-authored-by: Keren Zhou <kerenzhou@openai.com> Co-authored-by: Yan Chunwei <yanchunwei@outlook.com> Co-authored-by: goostavz <109190422+goostavz@users.noreply.github.com> Co-authored-by: Shintaro Iwasaki <siwasaki@fb.com> Co-authored-by: Yan Da <dyanab@connect.ust.hk> Co-authored-by: Jun Yang <yangjunpro@gmail.com> Co-authored-by: Ian Bearman <ianb@microsoft.com> Co-authored-by: Jason Ansel <jansel@jansel.net> Co-authored-by: Qingyi Liu <qingyil@nvidia.com> Co-authored-by: ben-zhang-609 <110140741+ben-zhang-609@users.noreply.github.com> Co-authored-by: Chenggang Zhao <lyricz@yeah.net> Co-authored-by: ben-zhang-609 <benzh609@gmail.com> Co-authored-by: dongdongl <dongdongl@nvidia.com>
This commit is contained in:
9
test/lib/Analysis/CMakeLists.txt
Normal file
9
test/lib/Analysis/CMakeLists.txt
Normal file
@@ -0,0 +1,9 @@
|
||||
add_mlir_library(TritonTestAnalysis
|
||||
TestAlias.cpp
|
||||
TestAxisInfo.cpp
|
||||
TestAllocation.cpp
|
||||
TestMembar.cpp
|
||||
|
||||
LINK_LIBS PUBLIC
|
||||
TritonAnalysis
|
||||
)
|
92
test/lib/Analysis/TestAlias.cpp
Normal file
92
test/lib/Analysis/TestAlias.cpp
Normal file
@@ -0,0 +1,92 @@
|
||||
#include "mlir/IR/AsmState.h"
|
||||
#include "mlir/Pass/Pass.h"
|
||||
#include "triton/Analysis/Alias.h"
|
||||
#include "triton/Analysis/Utility.h"
|
||||
#include "triton/Dialect/TritonGPU/IR/Dialect.h"
|
||||
|
||||
using namespace mlir;
|
||||
|
||||
namespace {
|
||||
|
||||
struct TestAliasPass
|
||||
: public PassWrapper<TestAliasPass, OperationPass<FuncOp>> {
|
||||
|
||||
// LLVM15+
|
||||
// MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(TestAliasPass);
|
||||
static void print(StringRef name, SmallVector<std::string, 4> &vals,
|
||||
raw_ostream &os) {
|
||||
if (vals.empty())
|
||||
return;
|
||||
os << name << " -> ";
|
||||
size_t i = 0;
|
||||
for (auto val : vals) {
|
||||
if (i != 0)
|
||||
os << ",";
|
||||
os << val;
|
||||
++i;
|
||||
}
|
||||
os << "\n";
|
||||
}
|
||||
|
||||
StringRef getArgument() const final { return "test-print-alias"; }
|
||||
StringRef getDescription() const final {
|
||||
return "print the result of the alias analysis pass";
|
||||
}
|
||||
|
||||
void runOnOperation() override {
|
||||
Operation *operation = getOperation();
|
||||
auto &os = llvm::errs();
|
||||
auto op_name = SymbolTable::getSymbolName(operation).getValue().str();
|
||||
os << op_name << "\n";
|
||||
|
||||
SharedMemoryAliasAnalysis analysis(&getContext());
|
||||
analysis.run(operation);
|
||||
|
||||
AsmState state(operation->getParentOfType<ModuleOp>());
|
||||
// Get operation ids of value's aliases
|
||||
auto getAllocOpNames = [&](Value value) {
|
||||
LatticeElement<AliasInfo> *latticeElement =
|
||||
analysis.lookupLatticeElement(value);
|
||||
SmallVector<std::string, 4> opNames;
|
||||
if (latticeElement) {
|
||||
auto &info = latticeElement->getValue();
|
||||
if (!info.getAllocs().empty()) {
|
||||
for (auto &alias : info.getAllocs()) {
|
||||
auto opName =
|
||||
getValueOperandName(alias.getDefiningOp()->getResult(0), state);
|
||||
opNames.push_back(std::move(opName));
|
||||
}
|
||||
}
|
||||
}
|
||||
// Ensure deterministic output
|
||||
std::sort(opNames.begin(), opNames.end());
|
||||
return opNames;
|
||||
};
|
||||
|
||||
operation->walk<WalkOrder::PreOrder>([&](Operation *op) {
|
||||
if (op->getNumResults() < 1)
|
||||
return;
|
||||
if (auto forOp = dyn_cast<scf::ForOp>(op)) {
|
||||
for (auto arg : llvm::enumerate(forOp.getRegionIterArgs())) {
|
||||
auto operand = forOp.getOpOperandForRegionIterArg(arg.value()).get();
|
||||
auto opNames = getAllocOpNames(operand);
|
||||
auto argName = getValueOperandName(arg.value(), state);
|
||||
print(argName, opNames, os);
|
||||
}
|
||||
}
|
||||
for (auto result : llvm::enumerate(op->getResults())) {
|
||||
auto opNames = getAllocOpNames(result.value());
|
||||
auto resultName = getValueOperandName(result.value(), state);
|
||||
print(resultName, opNames, os);
|
||||
}
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace
|
||||
|
||||
namespace mlir {
|
||||
namespace test {
|
||||
void registerTestAliasPass() { PassRegistration<TestAliasPass>(); }
|
||||
} // namespace test
|
||||
} // namespace mlir
|
54
test/lib/Analysis/TestAllocation.cpp
Normal file
54
test/lib/Analysis/TestAllocation.cpp
Normal file
@@ -0,0 +1,54 @@
|
||||
#include "mlir/Pass/Pass.h"
|
||||
#include "triton/Analysis/Allocation.h"
|
||||
|
||||
using namespace mlir;
|
||||
|
||||
namespace {
|
||||
|
||||
struct TestAllocationPass
|
||||
: public PassWrapper<TestAllocationPass, OperationPass<FuncOp>> {
|
||||
|
||||
// LLVM15+
|
||||
// MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(TestAllocationPass);
|
||||
|
||||
StringRef getArgument() const final { return "test-print-allocation"; }
|
||||
StringRef getDescription() const final {
|
||||
return "print the result of the allocation pass";
|
||||
}
|
||||
|
||||
void runOnOperation() override {
|
||||
Operation *operation = getOperation();
|
||||
auto &os = llvm::errs();
|
||||
// Convert to std::string can remove quotes from op_name
|
||||
auto op_name = SymbolTable::getSymbolName(operation).getValue().str();
|
||||
os << op_name << "\n";
|
||||
Allocation allocation(operation);
|
||||
operation->walk([&](Operation *op) {
|
||||
auto scratchBufferId = allocation.getBufferId(op);
|
||||
if (scratchBufferId != Allocation::InvalidBufferId) {
|
||||
size_t offset = allocation.getOffset(scratchBufferId);
|
||||
size_t size = allocation.getAllocatedSize(scratchBufferId);
|
||||
os << "scratch offset = " << offset << ", size = " << size << "\n";
|
||||
}
|
||||
if (op->getNumResults() < 1)
|
||||
return;
|
||||
for (Value result : op->getResults()) {
|
||||
auto bufferId = allocation.getBufferId(result);
|
||||
if (bufferId != Allocation::InvalidBufferId) {
|
||||
size_t offset = allocation.getOffset(bufferId);
|
||||
size_t size = allocation.getAllocatedSize(bufferId);
|
||||
os << "offset = " << offset << ", size = " << size << "\n";
|
||||
}
|
||||
}
|
||||
});
|
||||
os << "size = " << allocation.getSharedMemorySize() << "\n";
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace
|
||||
|
||||
namespace mlir {
|
||||
namespace test {
|
||||
void registerTestAllocationPass() { PassRegistration<TestAllocationPass>(); }
|
||||
} // namespace test
|
||||
} // namespace mlir
|
69
test/lib/Analysis/TestAxisInfo.cpp
Normal file
69
test/lib/Analysis/TestAxisInfo.cpp
Normal file
@@ -0,0 +1,69 @@
|
||||
#include "mlir/Pass/Pass.h"
|
||||
#include "triton/Analysis/AxisInfo.h"
|
||||
|
||||
using namespace mlir;
|
||||
|
||||
namespace {
|
||||
|
||||
struct TestAxisInfoPass
|
||||
: public PassWrapper<TestAxisInfoPass, OperationPass<FuncOp>> {
|
||||
|
||||
// LLVM15+
|
||||
// MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(TestAlignmentPass);
|
||||
|
||||
void print(const std::string &name, raw_ostream &os, ArrayRef<int> vals) {
|
||||
os << name << ": [";
|
||||
for (size_t d = 0; d < vals.size(); d++) {
|
||||
if (d != 0)
|
||||
os << ", ";
|
||||
os << vals[d];
|
||||
}
|
||||
os << "]";
|
||||
}
|
||||
|
||||
StringRef getArgument() const final { return "test-print-alignment"; }
|
||||
StringRef getDescription() const final {
|
||||
return "print the result of the alignment analysis pass";
|
||||
}
|
||||
|
||||
void runOnOperation() override {
|
||||
Operation *operation = getOperation();
|
||||
auto &os = llvm::errs();
|
||||
os << "Testing: " << operation->getName() << "\n";
|
||||
AxisInfoAnalysis analysis(&getContext());
|
||||
analysis.run(operation);
|
||||
operation->walk([&](Operation *op) {
|
||||
if (op->getNumResults() < 1)
|
||||
return;
|
||||
for (Value result : op->getResults()) {
|
||||
// std::ostringstream oss;
|
||||
// result.print(oss);
|
||||
// os << " => ";
|
||||
LatticeElement<AxisInfo> *latticeElement =
|
||||
analysis.lookupLatticeElement(result);
|
||||
if (!latticeElement) {
|
||||
os << "None\n";
|
||||
return;
|
||||
}
|
||||
AxisInfo &info = latticeElement->getValue();
|
||||
print("Contiguity", os, info.getContiguity());
|
||||
os << " ; ";
|
||||
print("Divisibility", os, info.getDivisibility());
|
||||
os << " ; ";
|
||||
print("Constancy", os, info.getConstancy());
|
||||
os << " ( ";
|
||||
result.print(os);
|
||||
os << " ) ";
|
||||
os << "\n";
|
||||
}
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace
|
||||
|
||||
namespace mlir {
|
||||
namespace test {
|
||||
void registerTestAlignmentPass() { PassRegistration<TestAxisInfoPass>(); }
|
||||
} // namespace test
|
||||
} // namespace mlir
|
52
test/lib/Analysis/TestMembar.cpp
Normal file
52
test/lib/Analysis/TestMembar.cpp
Normal file
@@ -0,0 +1,52 @@
|
||||
#include "mlir/Dialect/GPU/GPUDialect.h"
|
||||
#include "mlir/IR/Dialect.h"
|
||||
#include "mlir/Pass/Pass.h"
|
||||
#include "triton/Analysis/Allocation.h"
|
||||
#include "triton/Analysis/Membar.h"
|
||||
|
||||
using namespace mlir;
|
||||
|
||||
namespace {
|
||||
|
||||
struct TestMembarPass
|
||||
: public PassWrapper<TestMembarPass, OperationPass<FuncOp>> {
|
||||
|
||||
// LLVM15+
|
||||
// MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(TestMembarPass);
|
||||
|
||||
StringRef getArgument() const final { return "test-print-membar"; }
|
||||
StringRef getDescription() const final {
|
||||
return "print the result of the allocation pass";
|
||||
}
|
||||
|
||||
void runOnOperation() override {
|
||||
Operation *operation = getOperation();
|
||||
auto &os = llvm::errs();
|
||||
// Convert to std::string can remove quotes from op_name
|
||||
auto op_name = SymbolTable::getSymbolName(operation).getValue().str();
|
||||
os << op_name << "\n";
|
||||
Allocation allocation(operation);
|
||||
MembarAnalysis membarPass(&allocation);
|
||||
membarPass.run();
|
||||
|
||||
size_t operationId = 0;
|
||||
operation->walk<WalkOrder::PreOrder>([&](Operation *op) {
|
||||
if (isa<gpu::BarrierOp>(op)) {
|
||||
os << "Membar " << operationId << "\n";
|
||||
}
|
||||
if (op->getNumRegions() == 0) {
|
||||
// Don't count parent Operation to simplify the test.
|
||||
operationId++;
|
||||
}
|
||||
return;
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace
|
||||
|
||||
namespace mlir {
|
||||
namespace test {
|
||||
void registerTestMembarPass() { PassRegistration<TestMembarPass>(); }
|
||||
} // namespace test
|
||||
} // namespace mlir
|
Reference in New Issue
Block a user