Enable output allocation cache (#2010)

Fixes #2002

checks all IterDomain on outputs and disables verifies that no extent value is a consumer of fusion inputs.

torch/csrc/jit/codegen/cuda/executor.cpp

@@ -183,6 +183,35 @@ void FusionExecutor::compileFusion(
     TORCH_INTERNAL_ASSERT(
         out->getValType() == ValType::TensorView,
         "Output types from fusions that are not tensors are not supported at this point.");
+
+    const auto maybe_rfactor_domain =
+        out->as<TensorView>()->getMaybeRFactorDomain();
+    // walking through outputs to see if output shapes are dependent on
+    // non-tensor inputs. For which case, we should have disabled output
+    // allocation, since the caching id only looks at tensor shapes.
+    // See issue https://github.com/csarofeen/pytorch/issues/2002
+    std::vector<Val*> output_extents;
+    for (const auto id : maybe_rfactor_domain) {
+      Val* extent = nullptr;
+      if (id->isReduction() || id->isStride()) {
+        continue;
+      } else if (id->isBroadcast() && id->hasExpandedExtent()) {
+        extent = id->expandedExtent();
+      } else {
+        extent = id->extent();
+      }
+      output_extents.emplace_back(extent);
+    }
+    auto dependencies = InputsOf::outputs(fusion, output_extents);
+    if (std::any_of(dependencies.begin(), dependencies.end(), [](Val* val) {
+          return val->isFusionInput();
+        })) {
+      // TODO: parameter cache is too big a hammer here. We should consider
+      // separate the caching logic of output sizes & launch params. Since
+      // output size dependency should only invalidate the output sizes
+      disable_parameter_cache_ = true;
+      break;
+    }
   }
 
   if (isDebugDumpEnabled(DebugDumpOption::FusionIr)) {

torch/csrc/jit/codegen/cuda/executor.h

@@ -310,7 +310,10 @@ class TORCH_CUDA_CU_API FusionExecutor : public NonCopyable {
   // Profiling support: the last launch param used
   LaunchParams launch_params_;
 
-  // Profiling support: knob to disable caching of launch params
+  // Profiling support: disable caching of launch params and output allocation
+  // output allocation is also disable when output sizes are dependent on
+  // runtime scalar inputs, such as for the case of tensor factory. see
+  // https://github.com/csarofeen/pytorch/issues/2002
   bool disable_parameter_cache_ = false;
 
   // Profiling support: kept copy of the cuda kernel

torch/csrc/jit/codegen/cuda/executor_kernel_arg.h

@@ -301,7 +301,9 @@ class TORCH_CUDA_CU_API KernelArgumentHolder {
       : index_mode_(index_mode) {}
 
   KernelArgumentHolder(const KernelArgumentHolder& self)
-      : device_index_(self.getDeviceIndex()), index_mode_(self.getIndexMode()) {
+      : device_index_(self.getDeviceIndex()),
+        cache_id_(self.getCacheId()),
+        index_mode_(self.getIndexMode()) {
     for (const auto& arg : self.arguments_) {
       push(arg.get());
     }

torch/csrc/jit/codegen/cuda/kernel_cache.cpp

@@ -621,11 +621,16 @@ std::vector<at::Tensor> FusionKernelRuntime::runWithInput(
               << std::endl;
   }
 
+  // group should share cache id.
+  auto group_cache_id = args.getCacheId();
   for (auto group_to_run : runtime_workspace_.group_run_order) {
     // TODO: index mode should be updated per segmented kernel
     // Prepare input vector
     KernelArgumentHolder group_runtime_inputs(args.getIndexMode());
     group_runtime_inputs.setDeviceIndex(args.getDeviceIndex());
+    if (group_cache_id.has_value()) {
+      group_runtime_inputs.setCacheId(group_cache_id.value());
+    }
     for (auto input : group_to_run->inputs()) {
       group_runtime_inputs.push(tensor_map.at(input));
     }