Line data Source code
1 : #pragma once 2 : 3 : #include <c10/core/impl/LocalDispatchKeySet.h> 4 : #include <c10/macros/Export.h> 5 : #include <c10/macros/Macros.h> 6 : 7 : // NOTE [Tracing Mode Switches] 8 : // 9 : // Historically, tracing function was controlled by two switches: 10 : // 11 : // - `AutoDispatchBelowADInplaceOrView` guard 12 : // 13 : // Tracing function used to be script-generated inside `VariableType_*.cpp` 14 : // kernels, sharing the same `Autograd` dispatch key with autograd function. 15 : // Therefore, before tracing function was moved out of VariableType, 16 : // `AutoDispatchBelowADInplaceOrView` guard can also disable tracing as a 17 : // side effect of disabling `Autograd` dispatching. 18 : // 19 : // - `setTracingState()` API in `torch/csrc/jit/frontend/tracer.h` 20 : // 21 : // It stores tracing data in a `TracingState` object in TLS. If the 22 : // `TracingState` object in TLS is `null`, then tracing is paused. 23 : // 24 : // The `TracingState` object is created in `tracer::trace()` - the main 25 : // entrance of tracing function. It's temporarily set to `null` inside 26 : // generated VariableType (now TraceType) to bypass tracing for intermediate 27 : // ops (ops being called by other ops). After the intermediate op call 28 : // finishes it's set back to the original `TracingState` object. 29 : // 30 : // The `TracingState` obect in TLS can also be read/written via its Python 31 : // binding in `python_tracer.cpp`, and `get/setTracingState()` C++ APIs, 32 : // which are also exposed as `TORCH_API`. 33 : // 34 : // Two new switches were introduced since tracing function was moved out of 35 : // VariableType: 36 : // 37 : // - `tracer::impl::set_dispatch_enabled()` API 38 : // 39 : // Unlike the special `Autograd` dispatch key which is included in dispatch 40 : // key set by default, `Tracer` dispatch key is off by default. The 41 : // dispatching switch can be toggled via this new API. 42 : // 43 : // - `tracer::impl::NoTracerDispatchMode` guard 44 : // 45 : // It's used to cover the old semantics of `AutoDispatchBelowADInplaceOrView` 46 : // after tracing was moved out of VariableType. 47 : // 48 : // Before tracing function was moved out of VariableType, tracing was enabled 49 : // when the following conditions are satisfied: 50 : // 51 : // 1) `TracingState` object in TLS != null; 52 : // - Either inside the execution scope of `tracer::trace()`, or 53 : // - Eagerly called `setTracingState()` with non-null object. 54 : // 2) Not inside `AutoDispatchBelowADInplaceOrView` scope; 55 : // 56 : // After: 57 : // 58 : // 1) `TracingState` object in TLS != null; 59 : // 2) Has called `tracer::impl::set_dispatch_enabled(true)`; 60 : // 3) Not inside `tracer::impl::NonDispatchGuard` scope; 61 : // 62 : // [TODOs] 63 : // 64 : // - `setTracingState()` v.s. `tracer::impl::set_dispatch_enabled()` 65 : // 66 : // Currently `set_dispatch_enabled()` is set/unset inside `setTracingState()` 67 : // to keep the semantics exactly the same as before - it's confusing to keep 68 : // both switches, though. We should consider simplifying/limiting the exposed 69 : // `setTracingState()` Python/C++ APIs (and other APIs calling it) so that 70 : // these two can be unified. 71 : // 72 : // - `AutoDispatchBelowADInplaceOrView` v.s. 73 : // `tracer::impl::NoTracerDispatchMode` 74 : // 75 : // We don't need to always set both guards together to keep semantics 76 : // unchanged. For the follow use cases of `AutoDispatchBelowADInplaceOrView` 77 : // we don't need set the new tracer guard: 78 : // 79 : // * Script-generated VariableType kernels. The guard is not necessary as 80 : // tracing is already disabled explicitly by `setTracingState(null)` in 81 : // generated TraceType kernels - we could keep it as is or use the new guard 82 : // instead. 83 : // 84 : // * Custom ops. Will be handled by fallback kernel for `Tracer`. 85 : // 86 : // * Functions that are not likely to be called in tracing context (no python 87 : // binding / not an operator), e.g.: all mobile forward() wrappers, test 88 : // binaries, and etc. 89 : // 90 : // * Where new threads are spawned, e.g.: ATen/native/ConvolutionMM2d.cpp. 91 : // It's not necessary as tracing is off by default. 92 : // 93 : // For the rest of cases we might need have both: 94 : // 95 : // * Functions that might be reachable from eager mode python (especially 96 : // factory methods), e.g.: 97 : // `internal_new_from_data()` in `torch/csrc/utils/tensor_new.cpp`. 98 : // Without the new guard it will add `aten::empty` to the traced graph. 99 : // 100 : // * Some manually maintained functions, e.g.: 101 : // `torch/csrc/autograd/VariableTypeManual.cpp`. 102 : // Set the new guard if it's not obvious whether `setTracingState(null)` 103 : // has been called before it reaches the `AutoDispatchBelowADInplaceOrView` 104 : // guard. 105 : // 106 : // We might need tweak the usage of the new guard to optimize/fix things. 107 : // It should only affect the correctness of tracing function, because the 108 : // guard is essentially no-op when the master `setTracingState()` switch is 109 : // off. 110 : 111 : // TODO: move this from `at::` to `jit::torch::` after 112 : // `aten/src/ATen/cpp_custom_type_hack.h` is removed. 113 : 114 : namespace at::tracer::impl { 115 : 116 : inline bool is_dispatch_enabled() { 117 : return c10::impl::tls_is_dispatch_key_included(at::DispatchKey::Tracer) && 118 : !c10::impl::tls_is_dispatch_key_excluded(at::DispatchKey::Tracer); 119 : } 120 : 121 : inline void set_dispatch_enabled(bool enabled) { 122 : TORCH_INTERNAL_ASSERT( 123 : !c10::impl::tls_is_dispatch_key_excluded(at::DispatchKey::Tracer), 124 : "Cannot enable tracing within the scope of NoTracerDispatchMode!"); 125 : c10::impl::tls_set_dispatch_key_included(at::DispatchKey::Tracer, enabled); 126 : } 127 : 128 46 : struct NoTracerDispatchMode { 129 : c10::impl::ExcludeDispatchKeyGuard guard_{at::DispatchKey::Tracer}; 130 : }; 131 : 132 : } // namespace at::tracer::impl