Merged PR 1583: ELL release v3.1.0

ELL release v3.1.0

- Move to VS 2019
- Fix a codegen error that was resulting in incorrect functional behavior
- Fix regressions in audio training tutorial (#232)
- Add importing of Sum nodes to ONNX importer
- Fix crash in LLVMContext::SetName
- Improved performance of CNN models on Pi3 with new implementations of spatial, pointwise and regular convolutions
- Improved performance of reorder node
- New nodes: ReorderDataCodeNode, SpatialConvolutionNode, MatrixMatrixMultiplyCodeNode
- Implement parallelization strategies for matrix multiplication nodes.
- Only enable new MatrixMatrixMultipleCodeNode path for select ARM targets like Pi, and not Intel/AMD CPUs
- Add the flag `--skip_ellcode` to `compile` and `wrap.py` tools to use OpenBLAS for linear algebra computations.

.gitattributes

@@ -3,6 +3,11 @@
 ###############################################################################
 * text=auto
 
+###############################################################################
+# Explicitly force .sh scripts to have LF line endings
+###############################################################################
+*.sh text eol=lf
+
 ###############################################################################
 # Set default behavior for command prompt diff.
 #

.gitignore

@@ -66,6 +66,7 @@ artifacts/
 *.pidb
 *.svclog
 *.scc
+*.ll
 
 # Chutzpah Test files
 _Chutzpah*

CMake/CommonInterfaces.cmake

@@ -8,11 +8,9 @@
 # On Linux and Mac, this can be done by call *make* on the specific language wrapper e.g.
 # make _ELL_python
 
-cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
-
 set(GLOBAL_BIN_DIR "${CMAKE_BINARY_DIR}/bin")
 if(WIN32)
-set(GLOBAL_BIN_DIR "${CMAKE_BINARY_DIR}/bin/release")
+  set(GLOBAL_BIN_DIR "${GLOBAL_BIN_DIR}/release")
 endif()
 
 #

CMake/CopySharedLibraries.cmake

@@ -3,21 +3,27 @@
 #
 
 # Copies necessary DLLs to global binary directory
-macro(copy_shared_libraries target_name)
+macro(copy_shared_libraries_to target_name target_location)
     if(WIN32)
-        set(target_location "${CMAKE_BINARY_DIR}/bin/$<CONFIG>/")
         if(EXISTS ${BLAS_DLL_DIR})
             set(command_target_name copy_dlls_to_${target_name})
             foreach(blas_dll ${BLAS_DLLS})
                 add_custom_command(TARGET ${target_name} POST_BUILD
-                    COMMAND ${CMAKE_COMMAND} -E make_directory ${target_location}
-                    COMMAND ${CMAKE_COMMAND} -E copy_if_different ${BLAS_DLL_DIR}/${blas_dll} ${target_location}
+                    COMMAND ${CMAKE_COMMAND} -E make_directory "${target_location}/$<CONFIG>"
+                    COMMAND ${CMAKE_COMMAND} -E copy_if_different ${BLAS_DLL_DIR}/${blas_dll} "${target_location}/$<CONFIG>"
                 )
             endforeach()
         endif()
     endif()
 endmacro()
 
+macro(copy_shared_libraries target_name)
+    if(WIN32)
+        set(target_location "${CMAKE_BINARY_DIR}/bin")
+        copy_shared_libraries_to(${target_name} ${target_location})
+    endif()
+endmacro()
+
 macro(set_test_library_path test_name)
     if(WIN32)
         set (GLOBAL_BIN_DIR ${CMAKE_BINARY_DIR}/bin)

CMake/LLVMSetup.cmake

@@ -66,7 +66,4 @@ foreach(DEFINITION ${LLVM_DEFINITIONS})
     add_definitions(${DEFINITION})
 endforeach()
 
-set(LLVM_LIBS ${LLVM_AVAILABLE_LIBS})
-list(FILTER LLVM_LIBS INCLUDE REGEX "LLVM.+")
-
 set_property(TARGET intrinsics_gen PROPERTY FOLDER "cmake_macros")

CMake/OpenBLASSetup.cmake

@@ -123,14 +123,17 @@ else()
             set(BLAS_LIB_SEARCH_PATHS ${BLAS_PACKAGE_DIR}/lib/)
             set(BLAS_FOUND TRUE)
         else()
-            # Known registry ID (family, model) settings for various CPU types
+            # Known registry ID (family, model) settings for various Intel CPU types
             #
             # Haswell: Family 6, model 60, 63, 69
             # Broadwell: Family 6, Model 70, 79 (compatible with Haswell)
             # Kaby Lake: Family 6, Model 78, 142, 158 (compatible with Haswell)
             # Sandybridge: Family 6, model 42, 45
             # Ivybridge: Family 6, model 58 (compatible with Sandybridge)
-            # Skylake: Family 6, model 42
+            # Skylake: Family 6, model 85
+            #
+            # Known registry ID (family, model) settings for various AMD CPU types
+            # Epyc: Family 23, model 1 (compatible with Haswell)
 
             # We can set up a mapping from a detected processor generation to the version of
             # the OpenBLAS libraries to use with the set_processor_mapping macro. For instance,
@@ -159,11 +162,15 @@ else()
                         set(processor_generation "sandybridge")
                     elseif(processor_model EQUAL 58)
                         set(processor_generation "sandybridge") # actually ivybridge, but it is compatible with sandybridge
-                    elseif(processor_model EQUAL 42)
+                    elseif(processor_model EQUAL 85)
                         set(processor_generation "sandybridge") # actually skylake, but it is compatible with sandybridge
                     else()
                         set(processor_generation "unknown")
                     endif()
+                elseif(processor_family EQUAL 23)
+                    if(processor_model EQUAL 1)
+                        set(processor_generation "haswell")
+                    endif()
                 endif()
             else()
                 set(processor_generation "${PROCESSOR_HINT}")

CMakeLists.txt

@@ -4,11 +4,18 @@
 
 cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
 
+# Error on non-existent dependency in add_dependencies.
+cmake_policy(SET CMP0046 NEW)
+
 # Include modules in the CMake directory.
-list(APPEND CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/CMake")
+set(ELL_ROOT "${CMAKE_CURRENT_SOURCE_DIR}")
+list(APPEND CMAKE_MODULE_PATH "${ELL_ROOT}/CMake")
 include(CompilerCache)
 
-project(ELL)
+project(ELL CXX ASM)
+if(MSVC)
+  enable_language(ASM_MASM)
+endif()
 
 file(STRINGS "VERSION" ELL_VERSION)
 message(STATUS "ELL version ${ELL_VERSION}")
@@ -41,11 +48,14 @@ option(DISABLE_PYTHON "Explicitly disable building python modules" OFF)
 option(CNTK "Enable CNTK importer and related unit tests (requires CNTK python module)" OFF)
 option(ONNX "Enable ONNX importer and related unit tests (requires PyTorch and ONNX python modules)" OFF)
 
-set(ELL_ROOT "${CMAKE_SOURCE_DIR}")
-set(FLAKE8_CONFIG "${CMAKE_SOURCE_DIR}/.flake8")
-set(TEST_MODELS_REPO "https://github.com/Microsoft/ell-test-models" CACHE DOCUMENTATION "URL to the git repo containing test models" )
+set(FLAKE8_CONFIG "${ELL_ROOT}/.flake8")
+set(TEST_MODELS_REPO "https://github.com/Microsoft/ell-test-models" CACHE STRING "URL to the git repo containing test models" )
 message(STATUS "Configuring tests to use TEST_MODELS_REPO at: ${TEST_MODELS_REPO}")
-set(EXTERNAL_DIR "${CMAKE_SOURCE_DIR}/external" CACHE DOCUMENTATION "Directory to install external dependencies" )
+
+if(NOT ELL_EXTERNAL_DIR)
+  set(ELL_EXTERNAL_DIR "${ELL_ROOT}/external" CACHE STRING "Directory to install external dependencies" )
+endif(NOT ELL_EXTERNAL_DIR)
+
 set(RPI_PASSWORD "$ENV{RPI_PASSWORD}")
 set(RPI_CLUSTER "$ENV{RPI_CLUSTER}")
 set(RPI_KEY "$ENV{RPI_APIKEY}")
@@ -106,7 +116,8 @@ endif()
 enable_testing()
 
 # Set up global variables to help find NuGet projects
-set(PACKAGE_ROOT ${EXTERNAL_DIR})
+set(PACKAGE_ROOT ${ELL_EXTERNAL_DIR})
+
 include(OpenBLASSetup)
 include(LLVMSetup)
 include(SWIGSetup)
@@ -138,14 +149,16 @@ else()
   add_compile_options(-Wmissing-field-initializers)
   add_compile_options(-fvisibility-inlines-hidden)
   add_compile_options(-Wno-unknown-pragmas)
-  add_compile_options(-Wno-backslash-newline-escape)
   add_compile_options(-Wno-comment)
   set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -ggdb3 -O0")
   set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} -ggdb3 -O0")
   set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -ggdb3")
   set(CMAKE_C_FLAGS_RELWITHDEBINFO "${CMAKE_C_FLAGS_RELWITHDEBINFO} -ggdb3")
   if(${CMAKE_CXX_COMPILER_ID} STREQUAL Clang)
+    add_compile_options(-Wno-backslash-newline-escape)
     add_compile_options(-Wno-self-assign)
+  else() # GCC
+    add_compile_options(-Wno-ignored-attributes)
   endif()
 endif()
 
@@ -163,17 +176,17 @@ add_subdirectory(interfaces)
 add_subdirectory(examples)
 
 # Add user directories to ELL build if requested
-if(EXISTS "${CMAKE_SOURCE_DIR}/user")
+if(EXISTS "${ELL_ROOT}/user")
   # Add root user directory if it has a CMakeLists.txt file and INCLUDE_IN_ELL_BUILD.txt file
-  if(EXISTS"${CMAKE_SOURCE_DIR}/user/CMakeLists.txt" AND EXISTS "${CMAKE_SOURCE_DIR}/user/INCLUDE_IN_ELL_BUILD.txt")
+  if(EXISTS"${ELL_ROOT}/user/CMakeLists.txt" AND EXISTS "${ELL_ROOT}/user/INCLUDE_IN_ELL_BUILD.txt")
     message(STATUS "Adding user directory to ELL build")
     add_subdirectory(user)
   endif()
 
   # Now add all child directories that have CMakeLists.txt files and INCLUDE_IN_ELL_BUILD.txt file
-  file(GLOB children RELATIVE "${CMAKE_SOURCE_DIR}/user" "${CMAKE_SOURCE_DIR}/user/*")
+  file(GLOB children RELATIVE "${ELL_ROOT}/user" "${ELL_ROOT}/user/*")
   foreach(child ${children})
-    if(IS_DIRECTORY "${CMAKE_SOURCE_DIR}/user/${child}" AND EXISTS "${CMAKE_SOURCE_DIR}/user/${child}/CMakeLists.txt" AND EXISTS "${CMAKE_SOURCE_DIR}/user/${child}/INCLUDE_IN_ELL_BUILD.txt")
+    if(IS_DIRECTORY "${ELL_ROOT}/user/${child}" AND EXISTS "${ELL_ROOT}/user/${child}/CMakeLists.txt" AND EXISTS "${ELL_ROOT}/user/${child}/INCLUDE_IN_ELL_BUILD.txt")
       message(STATUS "Adding user directory ${child} to ELL build")
       add_subdirectory("user/${child}")
     endif()

History.md

@@ -1,3 +1,16 @@
+## 3.1.0
+- Move to VS 2019
+- Fix a codegen error that was resulting in incorrect functional behavior
+- Fix regressions in audio training tutorial (#232)
+- Add importing of Sum nodes to ONNX importer
+- Fix crash in LLVMContext::SetName
+- Improved performance of CNN models on Pi3 with new implementations of spatial, pointwise and regular convolutions
+- Improved performance of reorder node
+- New nodes: ReorderDataCodeNode, SpatialConvolutionNode, MatrixMatrixMultiplyCodeNode
+- Implement parallelization strategies for matrix multiplication nodes.
+- Only enable new MatrixMatrixMultipleCodeNode path for select ARM targets like Pi, and not Intel/AMD CPUs
+- Add the flag `--skip_ellcode` to `compile` and `wrap.py` tools to use OpenBLAS for linear algebra computations.
+
 ## 3.0.3
 - Fix VS 2019 build.
 

VERSION

@@ -1 +1 @@
-3.0.3
+3.1.0

docs/gallery/ILSVRC2012/Asparagus.md

@@ -6,7 +6,7 @@ permalink: /gallery/ILSVRC2012/Asparagus
 
 [Back to Gallery](/ELL/gallery)
 
-## ILSVRC2012 Classification: 64x64x3 Convolutional Neural Network (52.07% top 1 accuracy, 76.40% top 5 accuracy, 108ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
+## ILSVRC2012 Classification: 64x64x3 Convolutional Neural Network (52.07% top 1 accuracy, 76.40% top 5 accuracy, 181ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
 
 <table class="table table-striped table-bordered">
     <tr>
@@ -19,7 +19,7 @@ permalink: /gallery/ILSVRC2012/Asparagus
     </tr>
     <tr>
         <td> Performance </td>
-        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz : 108ms/frame </td>
+        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz : 181ms/frame </td>
     </tr>
     <tr>
         <td> Uncompressed Size </td>

docs/gallery/ILSVRC2012/Bean.md

@@ -6,20 +6,20 @@ permalink: /gallery/ILSVRC2012/Bean
 
 [Back to Gallery](/ELL/gallery)
 
-## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (55.12% top 1 accuracy, 78.21% top 5 accuracy, 144ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
+## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (55.12% top 1 accuracy, 78.21% top 5 accuracy, 87ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
 
 <table class="table table-striped table-bordered">
     <tr>
         <td> Download </td>
-        <td> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Bean/Bean_pi3.ell.zip">Bean_pi3.ell.zip</a></td>
+        <td> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Bean/Bean.ell.zip">Bean.ell.zip</a></td>
     </tr>
     <tr>
         <td> Accuracy </td>
         <td colspan="3"> ILSVRC2012: 78.21% (Top 5), 55.12% (Top 1) </td>
     </tr>
     <tr>
         <td> Performance </td>
-        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz : 144ms/frame </td>
+        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz : 87ms/frame </td>
     </tr>
     <tr>
         <td> Uncompressed Size </td>

docs/gallery/ILSVRC2012/Buckthorn.md

@@ -6,20 +6,20 @@ permalink: /gallery/ILSVRC2012/Buckthorn
 
 [Back to Gallery](/ELL/gallery)
 
-## ILSVRC2012 Classification: 64x64x3 Convolutional Neural Network (57.57% top 1 accuracy, 80.55% top 5 accuracy, 171ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
+## ILSVRC2012 Classification: 64x64x3 Convolutional Neural Network (57.57% top 1 accuracy, 80.55% top 5 accuracy, 113ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
 
 <table class="table table-striped table-bordered">
     <tr>
         <td> Download </td>
-        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Buckthorn/Buckthorn_pi3.ell.zip">Buckthorn_pi3.ell.zip</a></td>
+        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Buckthorn/Buckthorn.ell.zip">Buckthorn.ell.zip</a></td>
     </tr>
     <tr>
         <td> Accuracy </td>
         <td colspan="3"> ILSVRC2012: 80.55% (Top 5), 57.57% (Top 1) </td>
     </tr>
     <tr>
         <td> Performance </td>
-        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz : 171ms/frame </td>
+        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz : 113ms/frame </td>
     </tr>
     <tr>
         <td> Uncompressed Size </td>

docs/gallery/ILSVRC2012/Carrot.md

@@ -6,20 +6,20 @@ permalink: /gallery/ILSVRC2012/Carrot
 
 [Back to Gallery](/ELL/gallery)
 
-## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (64.61% top 1 accuracy, 85.63% top 5 accuracy, 397ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
+## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (64.61% top 1 accuracy, 85.63% top 5 accuracy, 341ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
 
 <table class="table table-striped table-bordered">
     <tr>
         <td> Download </td>
-        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Carrot/Carrot_pi3.ell.zip">Carrot_pi3.ell.zip</a></td>
+        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Carrot/Carrot.ell.zip">Carrot.ell.zip</a></td>
     </tr>
     <tr>
         <td> Accuracy </td>
         <td colspan="3"> ILSVRC2012: 85.63% (Top 5), 64.61% (Top 1) </td>
     </tr>
     <tr>
         <td> Performance </td>
-        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 397ms/frame </td>
+        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 341ms/frame </td>
     </tr>
     <tr>
         <td> Uncompressed Size </td>

docs/gallery/ILSVRC2012/CashewNut.md

@@ -6,20 +6,20 @@ permalink: /gallery/ILSVRC2012/CashewNut
 
 [Back to Gallery](/ELL/gallery)
 
-## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (60.22% top 1 accuracy, 82.44% top 5 accuracy, 178ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
+## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (60.22% top 1 accuracy, 82.44% top 5 accuracy, 106ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
 
 <table class="table table-striped table-bordered">
     <tr>
         <td> Download </td>
-        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/CashewNut/CashewNut_pi3.ell.zip">CashewNut_pi3.ell.zip</a></td>
+        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/CashewNut/CashewNut.ell.zip">CashewNut.ell.zip</a></td>
     </tr>
     <tr>
         <td> Accuracy </td>
         <td colspan="3"> ILSVRC2012: 82.44% (Top 5), 60.22% (Top 1) </td>
     </tr>
     <tr>
         <td> Performance </td>
-        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 178ms/frame </td>
+        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 106ms/frame </td>
     </tr>
     <tr>
         <td> Uncompressed Size </td>

docs/gallery/ILSVRC2012/Chalta.md

@@ -6,20 +6,20 @@ permalink: /gallery/ILSVRC2012/Chalta
 
 [Back to Gallery](/ELL/gallery)
 
-## ILSVRC2012 Classification: 64x64x3 Convolutional Neural Network (58.74% top 1 accuracy, 81.59% top 5 accuracy, 221ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
+## ILSVRC2012 Classification: 64x64x3 Convolutional Neural Network (58.74% top 1 accuracy, 81.59% top 5 accuracy, 147ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
 
 <table class="table table-striped table-bordered">
     <tr>
         <td> Download </td>
-        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Chalta/Chalta_pi3.ell.zip">Chalta_pi3.ell.zip</a></td>
+        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Chalta/Chalta.ell.zip">Chalta.ell.zip</a></td>
     </tr>
     <tr>
         <td> Accuracy </td>
         <td colspan="3"> ILSVRC2012: 81.59% (Top 5), 58.74% (Top 1) </td>
     </tr>
     <tr>
         <td> Performance </td>
-        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 221ms/frame </td>
+        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 147ms/frame </td>
     </tr>
     <tr>
         <td> Uncompressed Size </td>

docs/gallery/ILSVRC2012/Clary.md

@@ -6,20 +6,20 @@ permalink: /gallery/ILSVRC2012/Clary
 
 [Back to Gallery](/ELL/gallery)
 
-## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (66.65% top 1 accuracy, 87.17% top 5 accuracy, 506ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
+## ILSVRC2012 Classification: 128x128x3 Convolutional Neural Network (66.65% top 1 accuracy, 87.17% top 5 accuracy, 361ms/frame on Raspberry Pi 3 (Raspbian) @ 700MHz)
 
 <table class="table table-striped table-bordered">
     <tr>
         <td> Download </td>
-        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Clary/Clary_pi3.ell.zip">Clary_pi3.ell.zip</a></td>
+        <td colspan="3"> <a href="https://github.com/Microsoft/ELL-models/raw/master/models/ILSVRC2012/Clary/Clary.ell.zip">Clary.ell.zip</a></td>
     </tr>
     <tr>
         <td> Accuracy </td>
         <td colspan="3"> ILSVRC2012: 87.17% (Top 5), 66.65% (Top 1) </td>
     </tr>
     <tr>
         <td> Performance </td>
-        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 506ms/frame </td>
+        <td colspan="3"> Raspberry Pi 3 (Raspbian) @ 700MHz: 361ms/frame </td>
     </tr>
     <tr>
         <td> Uncompressed Size </td>