ailzhang

Functionalization

Link to prototype branch: https://github.com/ailzhang/pytorch/commit/83f647e9f14a89e61378c8e834aec4dfbcb74a00

Quick summary:

This prototype only focus on getting rid of aliasing in view ops, after this Func2 you can safely assume that after Func2 kernel you won't see any view ops, but only view_op_copy ops that returns completely new storage.

You can build this branch and check some examples by running python test/base_to_view.py and python test/view_to_base.py. But this branch is only a proof of concept and comes with a lot of hacks, and requires some careful design work described in the section below.

ailzhang

diff --git a/fbcode/caffe2/c10/core/DispatchKeySet.h b/fbcode/caffe2/c10/core/DispatchKeySet.h
--- a/fbcode/caffe2/c10/core/DispatchKeySet.h
+++ b/fbcode/caffe2/c10/core/DispatchKeySet.h
@@ -82,6 +82,10 @@
   DispatchKeySet operator-(DispatchKeySet other) const {
     return DispatchKeySet(repr_ & ~other.repr_);
   }
+  // Compute self ^ other
+  DispatchKeySet operator^(DispatchKeySet other) const {
+    return DispatchKeySet(repr_ ^ other.repr_);

ailzhang

diff --git a/fbcode/caffe2/c10/core/DispatchKeySet.h b/fbcode/caffe2/c10/core/DispatchKeySet.h
--- a/fbcode/caffe2/c10/core/DispatchKeySet.h
+++ b/fbcode/caffe2/c10/core/DispatchKeySet.h
@@ -82,6 +82,10 @@
   DispatchKeySet operator-(DispatchKeySet other) const {
     return DispatchKeySet(repr_ & ~other.repr_);
   }
+  // Compute self ^ other
+  DispatchKeySet operator^(DispatchKeySet other) const {
+    return DispatchKeySet(repr_ ^ other.repr_);

ailzhang

diff --git a/fbcode/caffe2/aten/src/ATen/core/dispatch/DispatchKeyExtractor.h b/fbcode/caffe2/aten/src/ATen/core/dispatch/DispatchKeyExtractor.h
--- a/fbcode/caffe2/aten/src/ATen/core/dispatch/DispatchKeyExtractor.h
+++ b/fbcode/caffe2/aten/src/ATen/core/dispatch/DispatchKeyExtractor.h
@@ -49,7 +49,8 @@
   // it's a bit troublesome, because fastpath TLS access requires the type of
   // the TLS in question to be zero-initialized, so you don't actually win
   // anyting in that case.
-  return (((ks | local.included_ | always_included) - local.excluded_) & key_mask);
+  // For the addtional XOR op, see note [TLS Initialization]
+  return (((ks | (local.included_ ^ c10::InplaceOrView_keyset) | always_included) - local.excluded_) & key_mask);

ailzhang

#include "torch/csrc/autograd/VariableTypeUtils.h"

#include <torch/library.h>

#include "torch/csrc/autograd/function.h"

#include <ATen/RedispatchFunctions.h>
#include "ATen/quantized/Quantizer.h"

// @generated from tools/autograd/templates/InplaceOrViewType.cpp

ailzhang

 # No difference in VariableType
 
 λ ~/pytorch033 cat diff_TypeDefault
--- TypeDefault_before  2020-08-03 16:23:14.155651681 -0700
+++ TypeDefault_after   2020-08-03 16:17:08.590947589 -0700
@@ -3904,6 +3904,8 @@
   m.impl("conv_transpose3d.input",
          torch::CppFunction::makeUnboxedOnly(&TypeDefault::conv_transpose3d_input));
   m.def("copy_(Tensor(a!) self, Tensor src, bool non_blocking=False) -> Tensor(a!)");
+  m.impl("copy_",

ailzhang

--- before42031/VariableTypeEverything.cpp	2020-07-31 13:16:02.450902000 -0700
+++ after42031/VariableTypeEverything.cpp	2020-07-31 13:06:18.799446000 -0700
@@ -46,22 +46,6 @@
 // Later when we merge the mobile op registration the anonymous namespace
 // will be restored.
 // namespace {
-Tensor __and___Scalar(const Tensor & self, Scalar other) {
-  auto result = TypeDefault::__and___Scalar(self, other);
-  return result;
-}

ailzhang

	jit_premul_bias + master	jit_premul_bias + Reduce time per guard
fuser=none executor=simple	13.53	12.79
fuser=none executor=profiling	19.52	13.88
fuser=te executor=profiling	13.44	12.14

ailzhang

0.Program arguments: /usr/lib/llvm-9/bin/clang -cc1 -triple x86_64-pc-linux-gnu -emit-obj -disable-free -disable-llvm-verifier -discard-value-names -main-file-name SobolEngineOps.cpp -mrelocation-model pic -pic-level 2 -mthread-mod
el posix -fno-trapping-math -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -coverage-notes-file /var/lib/jenkins/pytorch/bui
ld/caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/native/SobolEngineOps.cpp.gcno -resource-dir /usr/lib/llvm-9/lib/clang/9.0.0 -isystem /var/lib/jenkins/pytorch/build/third_party/gloo -isystem /var/lib/jenkins/pytorch/cmake/../th
ird_party/gloo -isystem /var/lib/jenkins/pytorch/cmake/../third_party/googletest/googlemock/include -isystem /var/lib/jenkins/pytorch/cmake/../third_party/googletest/googletest/include -isystem /var/lib/jenkins/pytorch/third_party/
protobuf/src -isystem /opt/conda/include -isystem /var/lib/jenkins/pytorch/third_party/gemmlowp

ailzhang

0.Program arguments: /usr/lib/llvm-9/bin/clang -cc1 -triple x86_64-pc-linux-gnu -emit-obj -disable-free -disable-llvm-verifier -discard-value-names -main-file-name SobolEngineOps.cpp -mrelocation-model pic -pic-level 2 -mthread-mod
el posix -fno-trapping-math -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -coverage-notes-file /var/lib/jenkins/pytorch/bui
ld/caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/native/SobolEngineOps.cpp.gcno -resource-dir /usr/lib/llvm-9/lib/clang/9.0.0 -isystem /var/lib/jenkins/pytorch/build/third_party/gloo -isystem /var/lib/jenkins/pytorch/cmake/../th
ird_party/gloo -isystem /var/lib/jenkins/pytorch/cmake/../third_party/googletest/googlemock/include -isystem /var/lib/jenkins/pytorch/cmake/../third_party/googletest/googletest/include -isystem /var/lib/jenkins/pytorch/third_party/
protobuf/src -isystem /opt/conda/include -isystem /var/lib/jenkins/pytorch/third_party/gemmlowp