Merge branch 'main' into ksuvorov/fix800

Author: Retribution98

benchmarks/triton_kernels_benchmark/benchmark_testing.py

@@ -213,16 +213,18 @@ def do_bench_upstream_pytorch_profiler(fn, n_warmup=25, n_repeat=100, grad_to_no
 
     function_events = prof.events()
 
-    functions = []
+    all_functions = []
     if isinstance(kernel_name, str):
         kernel_name = [kernel_name]
     for ker_name in kernel_name:
-        functions.extend(list(filter(lambda x: x.name.startswith(ker_name), function_events)))  # pylint: disable=cell-var-from-loop
+        functions = list(filter(lambda x: x.name.startswith(ker_name), function_events))  # pylint: disable=cell-var-from-loop
+        assert len(functions) == n_repeat, f"the profiling number for kernel: '{ker_name}' not match, {len(functions)}"
+        all_functions.append(functions)
     # profiling_func_filter = filter(lambda x: x.name.startswith("__profile_kernel_of_func"), function_events)
 
-    assert len(functions) == n_repeat, f"the profiling number not match, {len(functions)}"
     # Make the time to the milliseconds.
-    times = torch.tensor([f.self_device_time_total * 1e-3 for f in functions], dtype=torch.float)
+    times = torch.tensor([sum(map(lambda elem: elem.self_device_time_total, f)) * 1e-3 for f in zip(*all_functions)],
+                         dtype=torch.float)
     return _summarize_statistics(times, quantiles, return_mode)
 
 

benchmarks/triton_kernels_benchmark/gemm_streamk_benchmark.py

@@ -293,9 +293,8 @@ def benchmark(M, N, K, provider):
         torch_fn = lambda: torch.matmul(a, b).to(torch.float32)
 
         # benchmark_suit.assert_close(xetla_fn(), torch_fn(), atol=1e-4, rtol=1.0, err_msg='xetla to torch')
-        _, min_ms, max_ms, mean_ms, cv = benchmark_suit.do_bench(
-            xetla_fn, n_warmup=10, n_repeat=10, quantiles=quantiles,
-            kernel_name='gpu::xetla::kernel::gemm_universal_t<dispatch_stream_k')
+        _, min_ms, max_ms, mean_ms, cv = benchmark_suit.do_bench(xetla_fn, n_warmup=10, n_repeat=10,
+                                                                 quantiles=quantiles, kernel_name='stream_k_gemm_run')
     else:
         raise NotImplementedError(f'Unsupported provider {provider}')
 

third_party/intel/lib/Analysis/AxisInfo.cpp

@@ -105,8 +105,8 @@ class BinaryOpVisitorImpl : public AxisInfoVisitorImpl<OpTy> {
   AxisInfo
   getAxisInfo(OpTy op,
               ArrayRef<const dataflow::Lattice<AxisInfo> *> operands) override {
-    auto lhsInfo = operands[0]->getValue();
-    auto rhsInfo = operands[1]->getValue();
+    const auto &lhsInfo = operands[0]->getValue();
+    const auto &rhsInfo = operands[1]->getValue();
     auto rank = lhsInfo.getRank();
     assert(operands.size() == 2 && "Expected two operands");
     AxisInfo::DimVectorT contiguity;
@@ -658,8 +658,8 @@ class CmpOpAxisInfoVisitor final : public AxisInfoVisitorImpl<OpTy> {
       return AxisInfo();
     auto shape = resTy.getShape();
     short rank = resTy.getRank();
-    auto lhsInfo = operands[0]->getValue();
-    auto rhsInfo = operands[1]->getValue();
+    const auto &lhsInfo = operands[0]->getValue();
+    const auto &rhsInfo = operands[1]->getValue();
 
     AxisInfo::DimVectorT contiguity, divisibility, constancy;
     std::optional<int64_t> constantValue;
@@ -782,8 +782,8 @@ class SelectOpAxisInfoVisitor final : public AxisInfoVisitorImpl<OpTy> {
   getAxisInfo(OpTy op,
               ArrayRef<const dataflow::Lattice<AxisInfo> *> operands) override {
     auto condConstancy = operands[0]->getValue().getConstancy();
-    auto lhsInfo = operands[1]->getValue();
-    auto rhsInfo = operands[2]->getValue();
+    const auto &lhsInfo = operands[1]->getValue();
+    const auto &rhsInfo = operands[2]->getValue();
     auto rank = lhsInfo.getRank();
 
     AxisInfo::DimVectorT contiguity, divisibility, constancy;
@@ -967,8 +967,8 @@ class MaxMinOpAxisInfoVisitor final : public AxisInfoVisitorImpl<OpTy> {
   AxisInfo
   getAxisInfo(OpTy op,
               ArrayRef<const dataflow::Lattice<AxisInfo> *> operands) override {
-    auto lhsInfo = operands[0]->getValue();
-    auto rhsInfo = operands[1]->getValue();
+    const auto &lhsInfo = operands[0]->getValue();
+    const auto &rhsInfo = operands[1]->getValue();
     auto rank = lhsInfo.getRank();
     std::optional<int64_t> constantValue;
     if (lhsInfo.getConstantValue().has_value() &&
@@ -1140,8 +1140,8 @@ LogicalResult AxisInfoAnalysis::visitOperation(
 
 void AxisInfoAnalysis::visitForOpInductionVar(
     scf::ForOp op, ArrayRef<dataflow::Lattice<AxisInfo> *> argLattices) {
-  auto lb = getLatticeElementFor(op, op.getLowerBound())->getValue();
-  auto step = getLatticeElementFor(op, op.getStep())->getValue();
+  const auto &lb = getLatticeElementFor(op, op.getLowerBound())->getValue();
+  const auto &step = getLatticeElementFor(op, op.getStep())->getValue();
 
   AxisInfo::DimVectorT knownContiguity(1, 1);
   AxisInfo::DimVectorT knownDivisibility(1, 1);
@@ -1337,7 +1337,7 @@ void ModuleAxisInfoAnalysis::initialize(FunctionOpInterface funcOp) {
     return;
   auto *axisInfoMap = getFuncData(funcOp);
   auto updateAxisInfoMap = [&](Value value) {
-    auto axisInfo = analysis->getLatticeElement(value)->getValue();
+    const auto &axisInfo = analysis->getLatticeElement(value)->getValue();
     AxisInfo curAxisInfo;
     if (axisInfoMap->count(value)) {
       curAxisInfo = AxisInfo::join(axisInfo, axisInfoMap->lookup(value));