Update on "quant docs: add and clean up InstanceNorm{n}d"

Fixes docstrings and adds to quantization docs for quantized InstanceNorm. Test plan: * build on Mac OS and inspect Differential Revision: [D22152637](https://our.internmc.facebook.com/intern/diff/D22152637) [ghstack-poisoned]
pytorch · Jun 23, 2020 · b08d1f4 · b08d1f4
2 parents 6855ab6 + aea3438
commit b08d1f4
Show file tree

Hide file tree

Showing 81 changed files with 1,478 additions and 1,931 deletions.
diff --git a/.circleci/cimodel/data/simple/docker_definitions.py b/.circleci/cimodel/data/simple/docker_definitions.py
@@ -5,16 +5,13 @@
 
 # TODO: make this generated from a matrix rather than just a static list
 IMAGE_NAMES = [
-    "pytorch-linux-bionic-cuda11.0-cudnn8-py3.6-gcc9",
-    "pytorch-linux-bionic-cuda11.0-cudnn8-py3.8-gcc9",
     "pytorch-linux-bionic-cuda10.2-cudnn7-py3.8-gcc9",
     "pytorch-linux-bionic-py3.6-clang9",
     "pytorch-linux-bionic-cuda10.2-cudnn7-py3.6-clang9",
     "pytorch-linux-bionic-py3.8-gcc9",
     "pytorch-linux-xenial-cuda10-cudnn7-py3-gcc7",
     "pytorch-linux-xenial-cuda10.1-cudnn7-py3-gcc7",
     "pytorch-linux-xenial-cuda10.2-cudnn7-py3-gcc7",
-    "pytorch-linux-xenial-cuda11.0-cudnn8-py3-gcc7",
     "pytorch-linux-xenial-cuda9.2-cudnn7-py3-gcc5.4",
     "pytorch-linux-xenial-cuda9.2-cudnn7-py3-gcc7",
     "pytorch-linux-xenial-py3-clang5-android-ndk-r19c",

diff --git a/.circleci/config.yml b/.circleci/config.yml
@@ -7946,12 +7946,6 @@ workflows:
               only:
                 - master
     jobs:
-      - docker_build_job:
-          name: "pytorch-linux-bionic-cuda11.0-cudnn8-py3.6-gcc9"
-          image_name: "pytorch-linux-bionic-cuda11.0-cudnn8-py3.6-gcc9"
-      - docker_build_job:
-          name: "pytorch-linux-bionic-cuda11.0-cudnn8-py3.8-gcc9"
-          image_name: "pytorch-linux-bionic-cuda11.0-cudnn8-py3.8-gcc9"
       - docker_build_job:
           name: "pytorch-linux-bionic-cuda10.2-cudnn7-py3.8-gcc9"
           image_name: "pytorch-linux-bionic-cuda10.2-cudnn7-py3.8-gcc9"
@@ -7973,9 +7967,6 @@ workflows:
       - docker_build_job:
           name: "pytorch-linux-xenial-cuda10.2-cudnn7-py3-gcc7"
           image_name: "pytorch-linux-xenial-cuda10.2-cudnn7-py3-gcc7"
-      - docker_build_job:
-          name: "pytorch-linux-xenial-cuda11.0-cudnn8-py3-gcc7"
-          image_name: "pytorch-linux-xenial-cuda11.0-cudnn8-py3-gcc7"
       - docker_build_job:
           name: "pytorch-linux-xenial-cuda9.2-cudnn7-py3-gcc5.4"
           image_name: "pytorch-linux-xenial-cuda9.2-cudnn7-py3-gcc5.4"

diff --git a/.circleci/docker/build.sh b/.circleci/docker/build.sh
@@ -125,17 +125,6 @@ case "$image" in
     VISION=yes
     KATEX=yes
     ;;
-  pytorch-linux-xenial-cuda11.0-cudnn8-py3-gcc7)
-    UBUNTU_VERSION=16.04-rc
-    CUDA_VERSION=11.0
-    CUDNN_VERSION=8
-    ANACONDA_PYTHON_VERSION=3.6
-    GCC_VERSION=7
-    PROTOBUF=yes
-    DB=yes
-    VISION=yes
-    KATEX=yes
-    ;;
   pytorch-linux-xenial-py3-clang5-asan)
     ANACONDA_PYTHON_VERSION=3.6
     CLANG_VERSION=5.0
@@ -193,28 +182,6 @@ case "$image" in
     DB=yes
     VISION=yes
     ;;
-  pytorch-linux-bionic-cuda11.0-cudnn8-py3.6-gcc9)
-    UBUNTU_VERSION=18.04-rc
-    CUDA_VERSION=11.0
-    CUDNN_VERSION=8
-    ANACONDA_PYTHON_VERSION=3.6
-    GCC_VERSION=9
-    PROTOBUF=yes
-    DB=yes
-    VISION=yes
-    KATEX=yes
-    ;;
-  pytorch-linux-bionic-cuda11.0-cudnn8-py3.8-gcc9)
-    UBUNTU_VERSION=18.04-rc
-    CUDA_VERSION=11.0
-    CUDNN_VERSION=8
-    ANACONDA_PYTHON_VERSION=3.8
-    GCC_VERSION=9
-    PROTOBUF=yes
-    DB=yes
-    VISION=yes
-    KATEX=yes
-    ;;
   pytorch-linux-xenial-rocm3.3-py3.6)
     ANACONDA_PYTHON_VERSION=3.6
     PROTOBUF=yes

diff --git a/.circleci/docker/common/install_base.sh b/.circleci/docker/common/install_base.sh
@@ -2,16 +2,17 @@
 
 set -ex
 
-# NVIDIA dockers for RC releases use tag names like `11.0-cudnn8-devel-ubuntu18.04-rc`,
-# for this case we will set UBUNTU_VERSION to `18.04-rc` so that the Dockerfile could
-# find the correct image. As a result, here we have to check for
-#   "$UBUNTU_VERSION" == "18.04"*
-# instead of
-#   "$UBUNTU_VERSION" == "18.04"
-if [[ "$UBUNTU_VERSION" == "18.04"* ]]; then
+if [[ "$UBUNTU_VERSION" == "14.04" ]]; then
+  # cmake 2 is too old
+  cmake3=cmake3
+else
+  cmake3=cmake
+fi
+
+if [[ "$UBUNTU_VERSION" == "18.04" ]]; then
   cmake3="cmake=3.10*"
 else
-  cmake3="cmake=3.5*"
+  cmake3="${cmake3}=3.5*"
 fi
 
 # Install common dependencies

diff --git a/README.md b/README.md
@@ -353,8 +353,6 @@ We appreciate all contributions. If you are planning to contribute back bug-fixe
 If you plan to contribute new features, utility functions or extensions to the core, please first open an issue and discuss the feature with us.
 Sending a PR without discussion might end up resulting in a rejected PR, because we might be taking the core in a different direction than you might be aware of.
 
-To learn more about making a contribution to Pytorch, please see our [Contribution page](CONTRIBUTING.md).
-
 ## The Team
 
 PyTorch is a community driven project with several skillful engineers and researchers contributing to it.

diff --git a/android/pytorch_android/build.gradle b/android/pytorch_android/build.gradle
@@ -86,7 +86,7 @@ task sourcesJar(type: Jar) {
 
 def getLibtorchHeadersDir() {
   def abi = ABI_FILTERS.split(",")[0]
-  return "$rootDir/../build_android_$abi/install/include";
+  return "$rootDir/../build_android_$abi/include";
 }
 
 afterEvaluate {
@@ -106,18 +106,6 @@ afterEvaluate {
   }
 }
 
-tasks.whenTaskAdded { task ->
-  if (task.name.startsWith("bundle") && task.name.endsWith("Aar")) {
-    task.doLast {
-      addFolderToAarTask(
-        "addHeadersTo" + task.name,
-        task.archivePath,
-        getLibtorchHeadersDir(),
-        "headers").execute()
-    }
-  }
-}
-
 def addFolderToAarTask(taskName, aarPath, folderPath, folderPathInAar) {
   def tmpDir = file("${buildDir}/${taskName}")
   tmpDir.mkdir()

diff --git a/aten/src/ATen/Declarations.cwrap b/aten/src/ATen/Declarations.cwrap
@@ -210,6 +210,36 @@
     - bool largest
     - bool sorted
 ]]
+[[
+  name: _th_var
+  types:
+    - floating_point
+  backends:
+    - CPU
+    - CUDA
+  variants: function
+  options:
+    - cname: var_all
+      return: accreal
+      arguments:
+        - THTensor* self
+        - bool unbiased
+]]
+[[
+  name: _th_std
+  types:
+    - floating_point
+  backends:
+    - CPU
+    - CUDA
+  variants: function
+  options:
+    - cname: std_all
+      return: accreal
+      arguments:
+        - THTensor* self
+        - bool unbiased
+]]
 [[
   name: _th_renorm
   cname: renorm
@@ -345,10 +375,10 @@
   name: _th_addmm
   cname: addmm
   cpu_bfloat16: True
+  cuda_bfloat16: True
   variants:
     - function
   return: argument 0
-  backends: [CPU]
   options:
     - arguments:
       - arg: THTensor* result
@@ -365,7 +395,6 @@
   cuda_bfloat16: True
   variants: [function]
   return: self
-  backends: [CPU]
   options:
     - cname: addmm
       arguments:
@@ -380,9 +409,9 @@
   name: _th_addr
   cname: addr
   cpu_bfloat16: True
+  cuda_bfloat16: True
   variants: function
   return: argument 0
-  backends: [CPU]
   arguments:
     - arg: THTensor* result
       output: True
@@ -395,10 +424,10 @@
 [[
   name: _th_addr_
   cpu_bfloat16: True
+  cuda_bfloat16: True
   cname: addr
   return: self
   variants: function
-  backends: [CPU]
   arguments:
     - THTensor* self
     - THTensor* self
@@ -428,10 +457,10 @@
 ]]
 [[
   name: _th_addbmm
+  cuda_bfloat16: True
   cname: addbmm
   variants:
     - function
-  backends: [CPU]
   return: argument 0
   arguments:
     - arg: THTensor* result
@@ -444,10 +473,10 @@
 ]]
 [[
   name: _th_addbmm_
+  cuda_bfloat16: True
   cname: addbmm
   variants: function
   return: self
-  backends: [CPU]
   arguments:
     - THTensor* self
     - THTensor* self

diff --git a/aten/src/ATen/NamedTensorUtils.cpp b/aten/src/ATen/NamedTensorUtils.cpp
@@ -384,21 +384,6 @@ void propagate_names_for_addmv(
 }
 
 void propagate_names_for_addmm(
-    Tensor& result,
-    const Tensor& m1,
-    const Tensor& m2,
-    const Tensor& bias) {
-  if (!m1.has_names() && !m2.has_names() &&
-      !bias.has_names() && !result.has_names()) {
-    return;
-  }
-
-  auto mm_outnames = compute_matmul_outnames(m1.names(), m2.names());
-  auto add_outnames = unify_from_right(mm_outnames, bias.names());
-  propagate_names(result, add_outnames);
-}
-
-void propagate_names_for_addmm_legacy(
     TensorImpl* result,
     TensorImpl* m1,
     TensorImpl* m2,

diff --git a/aten/src/ATen/NamedTensorUtils.h b/aten/src/ATen/NamedTensorUtils.h
@@ -141,13 +141,6 @@ CAFFE2_API void propagate_names(TensorImpl* result, /*const */TensorImpl* src);
 
 // result = m1 @ m2 + bias
 CAFFE2_API void propagate_names_for_addmm(
-    Tensor& result,
-    const Tensor& m1,
-    const Tensor& m2,
-    const Tensor& bias);
-
-// result = m1 @ m2 + bias
-CAFFE2_API void propagate_names_for_addmm_legacy(
     TensorImpl* result,
     /*const*/TensorImpl* m1,
     /*const*/TensorImpl* m2,

diff --git a/aten/src/ATen/core/NamedRegistrations.cpp b/aten/src/ATen/core/NamedRegistrations.cpp
@@ -18,6 +18,8 @@ TORCH_LIBRARY_IMPL(aten, Named, m) {
   m.impl("_sparse_log_softmax.int", CppFunction::makeFallthrough());
   m.impl("_sparse_softmax.Dimname", CppFunction::makeFallthrough());
   m.impl("_sparse_softmax.int", CppFunction::makeFallthrough());
+  m.impl("_std", CppFunction::makeFallthrough());
+  m.impl("_var", CppFunction::makeFallthrough());
   m.impl("abs", CppFunction::makeFallthrough());
   m.impl("abs.out", CppFunction::makeFallthrough());
   m.impl("abs_", CppFunction::makeFallthrough());

diff --git a/aten/src/ATen/core/aten_interned_strings.h b/aten/src/ATen/core/aten_interned_strings.h
@@ -730,10 +730,6 @@ _(aten, where) \
 _(aten, zero) \
 _(aten, zeros) \
 _(aten, zeros_like) \
-_(aten, real) \
-_(aten, imag) \
-_(aten, view_as_real) \
-_(aten, view_as_complex) \
 /* nothing */
 
 #define FORALL_ATTR_BASE_SYMBOLS(_) \

diff --git a/aten/src/ATen/cuda/CUDABlas.cpp b/aten/src/ATen/cuda/CUDABlas.cpp
@@ -392,47 +392,6 @@ void gemv<at::BFloat16>(CUDABLAS_GEMV_ARGTYPES(at::BFloat16)) {
 }
 #endif
 
-namespace {
-template<typename scalar_t>
-cublasStatus_t cublasGer(const cublasHandle_t &handle, int64_t m, int64_t n, scalar_t *alpha, scalar_t *x, int64_t incx, scalar_t *y, int64_t incy, scalar_t *a, int64_t lda) {
-  TORCH_CHECK(false, "cublas ger is defined only for float and double");
-  return {};
-}
-template<>
-cublasStatus_t cublasGer<float>(const cublasHandle_t &handle, int64_t m, int64_t n, float *alpha, float *x, int64_t incx, float *y, int64_t incy, float *a, int64_t lda) {
-  return cublasSger(handle, m, n, alpha, x, incx, y, incy, a, lda);
-}
-template<>
-cublasStatus_t cublasGer<double>(const cublasHandle_t &handle, int64_t m, int64_t n, double *alpha, double *x, int64_t incx, double *y, int64_t incy, double *a, int64_t lda) {
-  return cublasDger(handle, m, n, alpha, x, incx, y, incy, a, lda);
-}
-} // anonymous namespace
-
-template<typename scalar_t>
-void ger(int64_t m, int64_t n, scalar_t alpha, scalar_t *x, int64_t incx, scalar_t *y, int64_t incy, scalar_t *a, int64_t lda)
-{
-  _cublasAdjustLdLevel2(m, n, &lda);
-  TORCH_CHECK((m <= INT_MAX) &&
-              (n <= INT_MAX) &&
-              (lda <= INT_MAX) &&
-              (incx <= INT_MAX) &&
-              (incy <= INT_MAX),
-              "cublasSger/cublasDger only supports m, n, lda, incx, incy with "
-              "the bound [val] <= %d", INT_MAX);
-  int i_m = (int)m;
-  int i_n = (int)n;
-  int i_lda = (int)lda;
-  int i_incx = (int)incx;
-  int i_incy = (int)incy;
-
-  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
-  TORCH_CUDABLAS_CHECK(cublasGer<scalar_t>(
-    handle, i_m, i_n, &alpha, x, i_incx, y, i_incy, a, i_lda));
-}
-template void ger<float>(int64_t m, int64_t n, float alpha, float *x, int64_t incx, float *y, int64_t incy, float *a, int64_t lda);
-template void ger<double>(int64_t m, int64_t n, double alpha, double *x, int64_t incx, double *y, int64_t incy, double *a, int64_t lda);
-
-
 } // namespace blas
 } // namespace cuda
 } // namespace at
diff --git a/aten/src/ATen/cuda/CUDABlas.h b/aten/src/ATen/cuda/CUDABlas.h
@@ -69,9 +69,6 @@ template <>
 void gemv<at::BFloat16>(CUDABLAS_GEMV_ARGTYPES(at::BFloat16));
 #endif
 
-template<typename Dtype>
-void ger(int64_t m, int64_t n, Dtype alpha, Dtype *x, int64_t incx, Dtype *y, int64_t incy, Dtype *a, int64_t lda);
-
 } // namespace blas
 } // namespace cuda
 } // namespace at