[HARDWARE] VTA Specialised Accelerator Stack

CMakeLists.txt

@@ -8,11 +8,6 @@ include(cmake/util/FindVulkan.cmake)
 include(cmake/util/FindLLVM.cmake)
 include(cmake/util/FindROCM.cmake)
 
-
-if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/build/private/local_config.cmake)
-  include(${CMAKE_CURRENT_SOURCE_DIR}/build/private/local_config.cmake)
-endif()
-
 if(EXISTS ${CMAKE_CURRENT_BINARY_DIR}/config.cmake)
   include(${CMAKE_CURRENT_BINARY_DIR}/config.cmake)
 else()
@@ -40,6 +35,8 @@ tvm_option(USE_RTTI "Build with RTTI" ON)
 tvm_option(USE_MSVC_MT "Build with MT" OFF)
 tvm_option(INSTALL_DEV "Install compiler infrastructure" OFF)
 
+tvm_option(USE_VTA_CFG "Use a specific json file for VTA runtime" "")
+
 # Contrib library options
 tvm_option(USE_BLAS "The blas library to be linked" none)
 tvm_option(USE_MKL_PATH "MKL root path when use MKL blas" none)
@@ -52,8 +49,9 @@ tvm_option(USE_NNPACK "Build with nnpack support" OFF)
 tvm_option(USE_RANDOM "Build with random support" OFF)
 
 # include directories
-include_directories(BEFORE "nnvm/include")
 include_directories("include")
+include_directories("nnvm/include")
+include_directories("dmlc-core/include")
 include_directories("HalideIR/src")
 include_directories("dlpack/include")
 include_directories("topi/include")
@@ -148,20 +146,8 @@ if(USE_GRAPH_RUNTIME)
   endif(USE_GRAPH_RUNTIME_DEBUG)
 endif(USE_GRAPH_RUNTIME)
 
-if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/dmlc-core/CMakeLists.txt)
-  include_directories(${CMAKE_CURRENT_SOURCE_DIR}/dmlc-core/include)
-  if (INSTALL_DEV)
-    install(
-      DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/dmlc-core/include/." DESTINATION "include"
-      FILES_MATCHING
-      PATTERN "*.h"
-    )
-  endif()
-elseif(DMLC_CORE_PATH)
-  include_directories(${DMLC_CORE_PATH}/include)
-endif()
-
 # Module rules
+include(cmake/modules/VTA.cmake)
 include(cmake/modules/CUDA.cmake)
 include(cmake/modules/OpenCL.cmake)
 include(cmake/modules/OpenGL.cmake)
@@ -174,7 +160,6 @@ include(cmake/modules/contrib/Random.cmake)
 include(cmake/modules/contrib/Sort.cmake)
 include(cmake/modules/contrib/NNPack.cmake)
 
-# Target rrules
 add_library(tvm SHARED ${COMPILER_SRCS} ${RUNTIME_SRCS})
 add_library(tvm_topi SHARED ${TOPI_SRCS})
 add_library(tvm_runtime SHARED ${RUNTIME_SRCS})
@@ -207,7 +192,6 @@ endif()
 # Custom targets
 add_custom_target(runtime DEPENDS tvm_runtime)
 
-
 # Installation rulse
 install(TARGETS tvm_runtime DESTINATION lib${LIB_SUFFIX})
 if(WIN32)

Jenkinsfile

@@ -8,7 +8,7 @@ tvm_runtime = "build/libtvm_runtime.so, build/config.cmake"
 tvm_lib = "build/libtvm.so, " + tvm_runtime
 // LLVM upstream lib
 tvm_multilib = "build/libtvm.so, " +
-             "build/libtvm_topi.so, build/libnnvm_compiler.so, " + tvm_runtime
+             "build/libvta.so, build/libtvm_topi.so, build/libnnvm_compiler.so, " + tvm_runtime
 
 // command to start a docker container
 docker_run = 'docker/build.sh'
@@ -134,6 +134,7 @@ stage('Build') {
         pack_lib('cpu', tvm_lib)
         timeout(time: max_time, unit: 'MINUTES') {
           sh "${docker_run} ci_cpu ./tests/scripts/task_cpp_unittest.sh"
+          sh "${docker_run} ci_cpu ./tests/scripts/task_python_vta.sh"
         }
       }
     }
@@ -179,6 +180,7 @@ stage('Unit Test') {
         timeout(time: max_time, unit: 'MINUTES') {
           sh "${docker_run} ci_i386 ./tests/scripts/task_python_unittest.sh"
           sh "${docker_run} ci_i386 ./tests/scripts/task_python_integration.sh"
+          sh "${docker_run} ci_i386 ./tests/scripts/task_python_vta.sh"
         }
       }
     }

Makefile

@@ -1,7 +1,7 @@
 ROOTDIR = $(CURDIR)
 
 .PHONY: clean all test doc pylint cpplint lint\
-	 cython cython2 cython3 web runtime
+	 cython cython2 cython3 web runtime vta
 
 ifndef DMLC_CORE_PATH
   DMLC_CORE_PATH = $(ROOTDIR)/dmlc-core
@@ -20,9 +20,11 @@ all:
 	@mkdir -p build && cd build && cmake .. && $(MAKE)
 
 runtime:
-
 	@mkdir -p build && cd build && cmake .. && $(MAKE) runtime
 
+vta:
+	@mkdir -p build && cd build && cmake .. && $(MAKE) vta
+
 cpptest:
 	@mkdir -p build && cd build && cmake .. && $(MAKE) cpptest
 
@@ -48,6 +50,7 @@ build/libtvm_web_runtime.js: build/libtvm_web_runtime.bc
 
 # Lint scripts
 cpplint:
+	python3 dmlc-core/scripts/lint.py vta cpp vta/include vta/src
 	python3 dmlc-core/scripts/lint.py topi cpp topi/include;
 	python3 dmlc-core/scripts/lint.py nnvm cpp nnvm/include nnvm/src;
 	python3 dmlc-core/scripts/lint.py tvm cpp include src verilog\
@@ -57,6 +60,7 @@ pylint:
 	python3 -m pylint python/tvm --rcfile=$(ROOTDIR)/tests/lint/pylintrc
 	python3 -m pylint topi/python/topi --rcfile=$(ROOTDIR)/tests/lint/pylintrc
 	python3 -m pylint nnvm/python/nnvm --rcfile=$(ROOTDIR)/tests/lint/pylintrc
+	python3 -m pylint vta/python/vta --rcfile=$(ROOTDIR)/tests/lint/pylintrc
 
 jnilint:
 	python3 dmlc-core/scripts/lint.py tvm4j-jni cpp jvm/native/src

apps/pynq_rpc/start_rpc_server.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+PROJROOT="$( cd "$( dirname "${BASH_SOURCE[0]}" )/../../" && pwd )"
+
+export PYTHONPATH=${PYTHONPATH}:${PROJROOT}/python:${PROJROOT}/vta/python
+python -m vta.exec.rpc_server

cmake/modules/VTA.cmake

@@ -0,0 +1,51 @@
+# CMake Build rules for VTA
+find_program(PYTHON python)
+
+if(MSVC)
+  message(STATUS "VTA build is skipped in Windows..")
+elseif(PYTHON)
+  set(VTA_CONFIG ${PYTHON} ${CMAKE_CURRENT_SOURCE_DIR}/vta/config/vta_config.py)
+
+  if(EXISTS ${CMAKE_CURRENT_BINARY_DIR}/vta_config.json)
+    message(STATUS "Use VTA config " ${CMAKE_CURRENT_BINARY_DIR}/vta_config.json)
+    set(VTA_CONFIG ${PYTHON} ${CMAKE_CURRENT_SOURCE_DIR}/vta/config/vta_config.py
+      --use-cfg=${CMAKE_CURRENT_BINARY_DIR}/vta_config.json)
+  endif()
+
+  execute_process(COMMAND ${VTA_CONFIG} --target OUTPUT_VARIABLE __vta_target)
+  string(STRIP ${__vta_target} VTA_TARGET)
+
+  message(STATUS "Build VTA runtime with target: " ${VTA_TARGET})
+
+  execute_process(COMMAND ${VTA_CONFIG} --defs OUTPUT_VARIABLE __vta_defs)
+
+  string(REGEX MATCHALL "(^| )-D[A-Za-z0-9_=.]*" VTA_DEFINITIONS "${__vta_defs}")
+
+  file(GLOB VTA_RUNTIME_SRCS vta/src/*.cc)
+  file(GLOB __vta_target_srcs vta/src/${VTA_TARGET}/*.cc)
+  list(APPEND VTA_RUNTIME_SRCS ${__vta_target_srcs})
+
+  add_library(vta SHARED ${VTA_RUNTIME_SRCS})
+
+  target_include_directories(vta PUBLIC vta/include)
+
+  foreach(__def ${VTA_DEFINITIONS})
+    string(SUBSTRING ${__def} 3 -1 __strip_def)
+    target_compile_definitions(vta PUBLIC ${__strip_def})
+  endforeach()
+
+  if(APPLE)
+    set_target_properties(vta PROPERTIES LINK_FLAGS "-undefined dynamic_lookup")
+  endif(APPLE)
+
+  # PYNQ rules
+  if(${VTA_TARGET} STREQUAL "pynq")
+    find_library(__sds_lib NAMES sds_lib PATHS /usr/lib)
+    find_library(__dma_lib NAMES dma PATHS
+      "/opt/python3.6/lib/python3.6/site-packages/pynq/drivers/"
+      "/opt/python3.6/lib/python3.6/site-packages/pynq/lib/")
+    target_link_libraries(vta ${__sds_lib} ${__dma_lib})
+  endif()
+else()
+  message(STATUS "Cannot found python in env, VTA build is skipped..")
+endif()

docs/Doxyfile

@@ -753,7 +753,7 @@ WARN_LOGFILE           =
 # spaces.
 # Note: If this tag is empty the current directory is searched.
 
-INPUT                  = include/tvm topi/include/topi nnvm/include/nnvm
+INPUT                  = include/tvm topi/include/topi nnvm/include/nnvm vta/include/vta
 
 # This tag can be used to specify the character encoding of the source files
 # that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses

docs/Makefile

@@ -46,6 +46,8 @@ help:
 clean:
 	rm -rf $(BUILDDIR)/*
 	rm -rf gen_modules
+	rm -rf tutorials
+	rm -rf vta/tutorials
 
 html:
 	$(SPHINXBUILD) -b html $(ALLSPHINXOPTS) $(BUILDDIR)/html

docs/README.txt

@@ -1,6 +1,28 @@
-The documentation of tvm is generated with recommonmark and sphinx.
+TVM Documentations
+==================
+This folder contains the source of TVM documents
 
 - A hosted version of doc is at http://docs.tvm.ai
-- pip install sphinx>=1.5.5 sphinx-gallery sphinx_rtd_theme matplotlib Image recommonmark
+- pip install sphinx>=1.5.5 sphinx-gallery sphinx_rtd_theme matplotlib Image recommonmark Pillow
 - Build tvm first in the root folder.
 - To build locally, you need to enable USE_CUDA, USE_OPENCL, LLVM_CONFIG in config.mk and then type "make html" in this folder.
+
+Only Execute Specified Tutorials
+--------------------------------
+The document build process will execute all the tutorials in the sphinx gallery.
+This will cause failure in some cases when certain machines do not have necessary
+environment. You can set ```TVM_TUTORIAL_EXEC_PATTERN``` to only execute
+the path that matches the regular expression pattern.
+
+For example, to only build tutorials under /vta/tutorials, run
+
+```bash
+TVM_TUTORIAL_EXEC_PATTERN=/vta/tutorials make html
+```
+
+To only build one specific file, do
+
+```bash
+# The slash \ is used to get . in regular expression
+TVM_TUTORIAL_EXEC_PATTERN=file_name\.py make html
+```

docs/api/python/hybrid.rst

@@ -6,10 +6,6 @@ tvm.hybrid
 
    tvm.hybrid.parse
    tvm.hybrid.script
-   tvm.hybrid.popcount
-   tvm.hybrid.sigmoid
 
 .. autofunction:: tvm.hybrid.parse
 .. autofunction:: tvm.hybrid.script
-.. autofunction:: tvm.hybrid.popcount
-.. autofunction:: tvm.hybrid.sigmoid

docs/api/python/index.rst

@@ -20,5 +20,6 @@ Python API
    contrib
    dev
    topi
+   vta/index
    nnvm/index
    hybrid

docs/api/python/vta/index.rst

@@ -0,0 +1,28 @@
+VTA API
+=======
+
+This document contains the python API to VTA compiler toolchain.
+
+.. automodule:: vta
+
+Hardware Information
+--------------------
+
+.. autofunction:: vta.Environment
+.. autofunction:: vta.get_env
+
+RPC Utilities
+-------------
+
+.. autofunction:: vta.reconfig_runtime
+.. autofunction:: vta.program_fpga
+
+
+Compiler API
+------------
+We program VTA using TVM, so the compiler API in vta package
+is only a thin wrapper to provide VTA specific extensions.
+
+.. autofunction:: vta.build_config
+.. autofunction:: vta.build
+.. autofunction:: vta.lower

docs/conf.py

@@ -26,6 +26,7 @@
 sys.path.insert(0, os.path.join(curr_path, '../python/'))
 sys.path.insert(0, os.path.join(curr_path, '../topi/python'))
 sys.path.insert(0, os.path.join(curr_path, '../nnvm/python'))
+sys.path.insert(0, os.path.join(curr_path, '../vta/python'))
 
 # -- General configuration ------------------------------------------------
 
@@ -184,16 +185,17 @@ def run_doxygen(folder):
 
 from sphinx_gallery.sorting import ExplicitOrder
 
-examples_dirs = ['../tutorials/']
-gallery_dirs = ['tutorials']
+examples_dirs = ["../tutorials/", "../vta/tutorials/"]
+gallery_dirs = ["tutorials", "vta/tutorials"]
+
 subsection_order = ExplicitOrder(
     ['../tutorials/language',
      '../tutorials/optimize',
+     '../tutorials/vta',
      '../tutorials/topi',
      '../tutorials/deployment',
      '../tutorials/nnvm'])
 
-
 def generate_doxygen_xml(app):
     """Run the doxygen make commands if we're on the ReadTheDocs server"""
     run_doxygen('..')
@@ -220,7 +222,7 @@ def setup(app):
     'examples_dirs': examples_dirs,
     'gallery_dirs': gallery_dirs,
     'subsection_order': subsection_order,
+    'filename_pattern': os.environ.get("TVM_TUTORIAL_EXEC_PATTERN", ".py"),
     'find_mayavi_figures': False,
-    'filename_pattern': '.py',
     'expected_failing_examples': []
 }

docs/index.rst

@@ -8,6 +8,7 @@ Get Started
 
    install/index
    tutorials/index
+   vta/index
    deploy/index
    contribute/index
    faq

docs/install/index.rst

@@ -1,6 +1,5 @@
 Installation
 ============
-
 To install TVM, please read :ref:`install-from-source`.
 If you are interested in deploying to mobile/embedded devices,
 you do not need to install the entire tvm stack on your device,

docs/vta/.gitignore

@@ -0,0 +1 @@
+tutorials

docs/vta/hardware.rst

@@ -0,0 +1,2 @@
+VTA Hardware Design Overview
+============================

docs/vta/index.rst

@@ -0,0 +1,23 @@
+VTA: Deep Learning Accelerator Stack
+====================================
+Specialized accelerators are key enablers of future deep learning workloads. TVM stack targets specialized accelerators.
+VTA(versatile tensor accelerator) is a generic, modular open-source deep learning accelerator.
+This page contains links to all the resources related to VTA:
+
+.. toctree::
+   :maxdepth: 1
+
+   install
+   tutorials/index
+   hardware
+
+
+Features
+--------
+VTA have the following key features:
+
+- Generic, modular open-source hardware
+- Streamlined workflow to deploy to FPGAs.
+- Simulator support to protoype compilation passes on regular workstations.
+- Driver and JIT runtime for both simulated and FPGA hardware backend.
+- End to end TVM stack integration