embed libzdnn in model.so (llvm#1324)

* - Build libzdnn.a with -fPIC and embed in model.so when
  -maccel=NNPA specified
- Add CompilerConfigMap to store states associated with
  certain options
- Move options in main() to CompilerOptions.cpp
- Fix compiler warning in Stickify.cpp

Signed-off-by: Gong Su <[email protected]>

* - fix NNPA_ENABLED for lit test
- install zdnn.h so we no longer need third_party/zdnn-lib
- move options back to onnx-mlir.cpp::main (consolidation
  requires much more effort, deferred)
- use DEPENDS in add_onnx_mlir_library for libzdnn dependency

Signed-off-by: Gong Su <[email protected]>

* Remove zdnn-lib from .gitmodules

Signed-off-by: Gong Su <[email protected]>

* Make libzdnn ALL target so it gets built before other NNPA components

Signed-off-by: Gong Su <[email protected]>

* Fix libzdnn dependency for NNPA components

Signed-off-by: Gong Su <[email protected]>

* Build NNPA in dev image as well

Signed-off-by: Gong Su <[email protected]>

* Comment out BYPRODUCTS in target libzdnn since generator support
requires cmake 3.20+ which is not yet available on official
Ubuntu Focal

Signed-off-by: Gong Su <[email protected]>

* - Force setting cached MLIR_DIR to honor command line argument
- unset cached LLVM_DIR so it changes along with MLIR_DIR
- surround third_party with set(CMAKE_MESSAGE_LOG_LEVEL NOTICE) and
  set(CMAKE_MESSAGE_LOG_LEVEL STATUS) to mask out their cmake
  output so we can see more clearly output by onnx-mlir only.
  third_party cmake output can still be turned on by the --log-level
  command line option

Signed-off-by: Gong Su <[email protected]>

* revert set(CMAKE_MESSAGE_LOG_LEVEL NOTICE) and set(CMAKE_MESSAGE_LOG_LEVEL STATUS)
around third_party to make another PR

Signed-off-by: Gong Su <[email protected]>

* revert force setting cached MLIR_DIR and unsetting cached LLVM_DIR
to make another PR

Signed-off-by: Gong Su <[email protected]>

Co-authored-by: Charles Volzka <[email protected]>
Co-authored-by: Tung D. Le <[email protected]>

Author: 3 people

.gitmodules

@@ -10,6 +10,3 @@
 [submodule "third_party/benchmark"]
 	path = third_party/benchmark
 	url = https://github.com/google/benchmark.git
-[submodule "third_party/zdnn-lib"]
-	path = third_party/zdnn-lib
-	url = https://github.com/IBM/zDNN.git

CMakeLists.txt

@@ -36,29 +36,21 @@ if (MSVC)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -wd4125")
 endif()
 
-# Define preprocessor for accelerators.
-set(NNPA_ENABLED 0)
-if (ACCELERATORS_TO_BUILD)
-  foreach(t ${ACCELERATORS_TO_BUILD})
-    if (${t} STREQUAL "NNPA")
-      set(NNPA_ENABLED 1)
-    endif()
-  endforeach(t)
-endif(ACCELERATORS_TO_BUILD)
-
 # Enable warnings as errors
 # Leverage the imported LLVM_ENABLE_WERROR for compiler logic
 set(LLVM_ENABLE_WERROR ${ONNX_MLIR_ENABLE_WERROR})
 
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/$<CONFIG>/lib)
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/$<CONFIG>/lib)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/$<CONFIG>/bin)
+set(CMAKE_INCLUDE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/$<CONFIG>/include)
 
 set(ONNX_MLIR_SRC_ROOT ${CMAKE_CURRENT_SOURCE_DIR})
 set(ONNX_MLIR_BIN_ROOT ${CMAKE_CURRENT_BINARY_DIR})
 
 set(ONNX_MLIR_LIBRARY_PATH ${CMAKE_LIBRARY_OUTPUT_DIRECTORY})
 set(ONNX_MLIR_RUNTIME_PATH ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
+set(ONNX_MLIR_INCLUDE_PATH ${CMAKE_INCLUDE_OUTPUT_DIRECTORY})
 
 set(ONNX_MLIR_VENDOR ${PACKAGE_VENDOR} CACHE STRING
   "Vendor-specific text for showing with version information.")

MLIR.cmake

@@ -27,17 +27,17 @@ include_directories(${MLIR_INCLUDE_DIRS})
 add_definitions(${LLVM_DEFINITIONS})
 
 set(BUILD_SHARED_LIBS ${LLVM_ENABLE_SHARED_LIBS} CACHE BOOL "" FORCE)
-message(STATUS "BUILD_SHARED_LIBS       : " ${BUILD_SHARED_LIBS})
+message(STATUS "BUILD_SHARED_LIBS        : " ${BUILD_SHARED_LIBS})
 
 # onnx uses exceptions, so we need to make sure that LLVM_REQUIRES_EH is set to ON, so that
 # the functions from HandleLLVMOptions and AddLLVM don't disable exceptions.
 set(LLVM_REQUIRES_EH ON)
-message(STATUS "LLVM_REQUIRES_EH        : " ${LLVM_REQUIRES_EH})
+message(STATUS "LLVM_REQUIRES_EH         : " ${LLVM_REQUIRES_EH})
 
 # LLVM_HOST_TRIPLE is exported as part of the llvm config, so we should be able to leverage it.
 # If, for some reason, it is not set, default to an empty string which is the old default behavior of onnx-mlir.
 set(ONNX_MLIR_DEFAULT_TRIPLE "${LLVM_HOST_TRIPLE}" CACHE STRING "Default triple for onnx-mlir.")
-message(STATUS "ONNX_MLIR_DEFAULT_TRIPLE: " ${ONNX_MLIR_DEFAULT_TRIPLE})
+message(STATUS "ONNX_MLIR_DEFAULT_TRIPLE : " ${ONNX_MLIR_DEFAULT_TRIPLE})
 
 # If CMAKE_INSTALL_PREFIX was not provided explicitly and we are not using an install of
 # LLVM and a CMakeCache.txt exists,
@@ -50,7 +50,7 @@ if (CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT AND NOT LLVM_INSTALL_PREFIX)
     set(CMAKE_INSTALL_PREFIX ${prefix} CACHE PATH "" FORCE)
   endif()
 endif()
-message(STATUS "CMAKE_INSTALL_PREFIX    : " ${CMAKE_INSTALL_PREFIX})
+message(STATUS "CMAKE_INSTALL_PREFIX     : " ${CMAKE_INSTALL_PREFIX})
 
 # The tablegen functions below are modeled based on the corresponding functions
 # in mlir: https://github.com/llvm/llvm-project/blob/main/mlir/cmake/modules/AddMLIR.cmake

docker/Dockerfile.onnx-mlir-dev

@@ -30,8 +30,13 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
 # Build onnx-mlir and run tests
     && cd ${ONNX_MLIR_ROOT} \
     && rm -rf build && mkdir -p build && cd build \
+# Dev image is built with NNPA acclerator (currently on s390x only)
+    && BUILD_NNPA=${BUILD_NNPA:-$([ "$(uname -m)" = "s390x" ] && echo NNPA || \
+                                  [ "$(uname -m)" = "x86_64" ] &&  echo || \
+                                  [ "$(uname -m)" = "ppc64le" ] && echo || echo)} \
     && MLIR_DIR=${LLVM_PROJECT_ROOT}/build/lib/cmake/mlir \
-       cmake .. \
+       cmake -DCMAKE_BUILD_TYPE=Debug \
+             -DACCELERATORS_TO_BUILD=${BUILD_NNPA} .. \
     && make -j${NPROC} \
     && make -j${NPROC} LIT_OPTS=-v check-onnx-lit \
 # Dev image is built without SIMD, placeholder for easy SIMD enablement

include/onnx-mlir/Runtime/OMInstrument.h

@@ -36,7 +36,7 @@
 #include <malloc.h>
 #endif // #ifdef __APPLE__
 
-#include "include/onnx-mlir/Compiler/OMCompilerMacros.h"
+#include <onnx-mlir/Compiler/OMCompilerMacros.h>
 
 #ifdef __cplusplus
 extern "C" {

src/Accelerators/CMakeLists.txt

@@ -8,6 +8,9 @@ if (ACCELERATORS_TO_BUILD)
     add_subdirectory(${t})
     list(APPEND ACCEL_LIST "${t}Accel")
   endforeach(t)
+  message(STATUS "Accelerators             : ${ACCEL_LIST}")
+else()
+  message(STATUS "Accelerators             : NONE")
 endif (ACCELERATORS_TO_BUILD)
 
 # Generate the Accelerators.inc file using a cmake script.

src/Accelerators/NNPA/CMakeLists.txt

@@ -1,10 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
 
+set(NNPA_ENABLED 1 CACHE BOOL "flag indicating NNPA enabled")
+
 set(NNPA_SRC_ROOT "${CMAKE_CURRENT_SOURCE_DIR}")
 set(NNPA_BIN_ROOT "${CMAKE_CURRENT_BINARY_DIR}")
 
 set(NNPA_LIBRARY_PATH ${CMAKE_LIBRARY_OUTPUT_DIRECTORY})
 set(NNPA_RUNTIME_PATH ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
+set(NNPA_INCLUDE_PATH ${CMAKE_INCLUDE_OUTPUT_DIRECTORY})
 
 set(NNPA_ONNX_MLIR_SRC_ROOT ${ONNX_MLIR_SRC_ROOT})
 set(NNPA_ONNX_MLIR_BIN_ROOT ${ONNX_MLIR_BIN_ROOT})
@@ -19,6 +22,9 @@ else()
   set(ZDNN_LIBRARY_ENABLED 0)
 endif()
 
+include(zdnn.cmake)
+setup_zdnn(v1.0.0)
+
 add_subdirectory(Dialect)
 add_subdirectory(Conversion)
 add_subdirectory(Support)
@@ -46,4 +52,3 @@ add_onnx_mlir_library(NNPAAccel
   OMZLowOps
   OMZLowToLLVM
   )
-

src/Accelerators/NNPA/Conversion/ONNXToZHigh/CMakeLists.txt

@@ -7,10 +7,14 @@ add_onnx_mlir_library(OMONNXToZHigh
 
   DEPENDS
   OMONNXONNXToZHighIncGen
+  libzdnn
 
   LINK_LIBS PUBLIC
   OMONNXToKrnl
   OMZHighOps
+
+  ACCEL_INCLUDE_DIRS PRIVATE
+  ${NNPA_INCLUDE_PATH}
   )
 
 add_onnx_mlir_rewriter(RewriteONNXForZHigh)
@@ -22,8 +26,12 @@ add_onnx_mlir_library(OMRewriteONNXForZHigh
 
   DEPENDS
   OMONNXRewriteONNXForZHighIncGen
+  libzdnn
 
   LINK_LIBS PUBLIC
   OMONNXToKrnl
   OMZHighOps
+
+  ACCEL_INCLUDE_DIRS PRIVATE
+  ${NNPA_INCLUDE_PATH}
   )

src/Accelerators/NNPA/Conversion/ZLowToLLVM/CMakeLists.txt

@@ -2,8 +2,14 @@ add_onnx_mlir_library(OMZLowToLLVM
   ZLowToLLVM.cpp
   ZLowToLLVMCommon.cpp
 
+  DEPENDS
+  libzdnn
+
   LINK_LIBS PUBLIC
   MLIRLLVMCommonConversion
   OMLayoutHelper
   OMZLowOps
+
+  ACCEL_INCLUDE_DIRS PRIVATE
+  ${NNPA_INCLUDE_PATH}
   )

src/Accelerators/NNPA/Conversion/ZLowToLLVM/ZLowToLLVM.cpp

@@ -18,7 +18,7 @@
 #include "src/Accelerators/NNPA/Conversion/ZLowToLLVM/ZLowToLLVMCommon.hpp"
 #include "src/Accelerators/NNPA/Dialect/ZLow/ZLowOps.hpp"
 #include "src/Accelerators/NNPA/Support/LayoutHelper.hpp"
-#include "third_party/zdnn-lib/zdnn/zdnn.h"
+#include "zdnn.h"
 
 using namespace mlir;
 using namespace onnx_mlir;

src/Accelerators/NNPA/Conversion/ZLowToLLVM/ZLowToLLVMCommon.cpp

@@ -19,7 +19,7 @@
 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 
-#include "third_party/zdnn-lib/zdnn/zdnn.h"
+#include "zdnn.h"
 
 using namespace mlir;
 

src/Accelerators/NNPA/Conversion/ZLowToLLVM/ZLowToLLVMCommon.hpp

@@ -18,7 +18,7 @@
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 #include "mlir/Transforms/DialectConversion.h"
 
-#include "third_party/zdnn-lib/zdnn/zdnn.h"
+#include "zdnn.h"
 
 namespace onnx_mlir {
 namespace zlow {

src/Accelerators/NNPA/Dialect/ZHigh/CMakeLists.txt

@@ -14,9 +14,13 @@ add_onnx_mlir_library(OMZHighOps
   OMZHighIncGen
   OMONNXZHighCombineIncGen
   OMShapeInferenceOpInterfaceIncGen
+  libzdnn
 
   LINK_LIBS PUBLIC
   OMONNXOps # Use ONNXShapeHelper 
   OMLayoutHelper
   MLIRIR
+
+  ACCEL_INCLUDE_DIRS PRIVATE
+  ${NNPA_INCLUDE_PATH}
   )

src/Accelerators/NNPA/Dialect/ZHigh/ZHighOps.cpp

@@ -33,7 +33,7 @@
 #include "src/Accelerators/NNPA/Dialect/ZHigh/ZHighOps.hpp"
 #include "src/Accelerators/NNPA/Dialect/ZHigh/ZHighShapeHelper.hpp"
 #include "src/Accelerators/NNPA/Support/LayoutHelper.hpp"
-#include "third_party/zdnn-lib/zdnn/zdnn.h"
+#include "zdnn.h"
 
 using namespace mlir;
 

src/Accelerators/NNPA/NNPAAccelerator.cpp

@@ -48,6 +48,7 @@ NNPAAccelerator *NNPAAccelerator::getInstance() {
 NNPAAccelerator::NNPAAccelerator() : Accelerator(Accelerator::Kind::NNPA) {
   LLVM_DEBUG(llvm::dbgs() << "Creating an NNPA accelerator\n");
   acceleratorTargets.push_back(this);
+  addCompilerConfig(CCM_SHARED_LIB_DEPS, {"zdnn"});
 };
 
 NNPAAccelerator::~NNPAAccelerator() { delete instance; }

src/Accelerators/NNPA/Support/CMakeLists.txt

@@ -1,12 +1,16 @@
 add_onnx_mlir_library(OMLayoutHelper
   LayoutHelper.cpp
 
+  DEPENDS
+  libzdnn
+
   LINK_LIBS PUBLIC
   MLIRIR
 
   ACCEL_INCLUDE_DIRS PRIVATE
   ${NNPA_SRC_ROOT}
   ${NNPA_BIN_ROOT}
+  ${NNPA_INCLUDE_PATH}
 )
 
 add_onnx_mlir_library(OMNNPASupport

src/Accelerators/NNPA/Support/LayoutHelper.hpp

@@ -13,7 +13,7 @@
 #pragma once
 
 #include "mlir/IR/BuiltinAttributes.h"
-#include "third_party/zdnn-lib/zdnn/zdnn.h"
+#include "zdnn.h"
 
 namespace onnx_mlir {
 

src/Accelerators/NNPA/Transform/ZHigh/CMakeLists.txt

@@ -8,6 +8,7 @@ add_onnx_mlir_library(OMZHighConstPropagation
 
   DEPENDS
   OMONNXZHighConstPropagationIncGen
+  libzdnn
 
   LINK_LIBS PUBLIC
   MLIRRewrite
@@ -16,6 +17,9 @@ add_onnx_mlir_library(OMZHighConstPropagation
   OMStickify
   OMZHighOps
   OMONNXOps
+
+  ACCEL_INCLUDE_DIRS PRIVATE
+  ${NNPA_INCLUDE_PATH}
   )
 
 add_onnx_mlir_rewriter(ZHighLayoutPropagation)

src/Accelerators/NNPA/Transform/ZHigh/Stickify/CMakeLists.txt

@@ -2,6 +2,12 @@ add_onnx_mlir_library(OMStickify
   Convert.cpp
   Stickify.cpp
 
+  DEPENDS
+  libzdnn
+
   LINK_LIBS PUBLIC
   LLVMSupport
+
+  ACCEL_INCLUDE_DIRS PRIVATE
+  ${NNPA_INCLUDE_PATH}
   )

src/Accelerators/NNPA/Transform/ZHigh/Stickify/Stickify.cpp

@@ -262,8 +262,8 @@ void free_aligned_4k(void *aligned_ptr) {
 // Functions from third_party/zdnn-lib/zdnn/utils.c
 uint64_t get_num_elements(const zdnn_ztensor *ztensor, elements_mode mode) {
   uint64_t num_elements = 1;
-  uint32_t *dims_ptr;
-  int i;
+  uint32_t *dims_ptr = NULL;
+  int i = 0;
 
   // Setup how to loop over the shape based on the mode.
   switch (mode) {

src/Accelerators/NNPA/Transform/ZHigh/Stickify/Stickify.hpp

@@ -14,7 +14,7 @@
 
 #pragma once
 
-#include "third_party/zdnn-lib/zdnn/zdnn.h"
+#include "zdnn.h"
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/SmallVector.h"
 

src/Accelerators/NNPA/zdnn.cmake

@@ -0,0 +1,46 @@
+# SPDX-License-Identifier: Apache-2.0
+
+function(setup_zdnn version)
+  set(ZDNN_GITHUB_URL https://github.com/IBM/zDNN)
+  set(ZDNN_PREFIX     ${CMAKE_CURRENT_BINARY_DIR}/zDNN)
+  set(ZDNN_TOPDIR     ${ZDNN_PREFIX}/src/zdnn)
+  set(ZDNN_OBJDIR     ${ZDNN_TOPDIR}/zdnn/obj)
+  set(ZDNN_LIBDIR     ${ZDNN_TOPDIR}/zdnn/lib)
+
+  ExternalProject_Add(zdnn
+    GIT_REPOSITORY ${ZDNN_GITHUB_URL}
+    GIT_TAG ${version}
+    PREFIX ${ZDNN_PREFIX}
+    BUILD_IN_SOURCE ON
+
+    # Skip autoconf and configure if config.make already exists
+    CONFIGURE_COMMAND sh -c "[ -f config.make ] || (autoconf && ./configure)"
+
+    # We build libzdnn.so so that obj/*.o are compiled with -fPIC
+    # Then we create libzdnn.a ourselves from these PIC .o since
+    # we want to embed libzdnn.a into model.so.
+    #
+    # Note we use sh to run the command. Otherwise, cmake will
+    # generate quotes around ${ZDNN_OBJDIR}/*.o due to the *
+    # in the path, which breaks ar.
+    #
+    # Set MAKEFLAGS to remove -j option passed down from the top
+    # level make which produces a warning about jobserver unavailable.
+    BUILD_COMMAND sh -c "MAKEFLAGS=--no-print-directory \
+                         make -j$(nproc) -C zdnn lib/libzdnn.so && \
+                         ar -rc ${ZDNN_LIBDIR}/libzdnn.a ${ZDNN_OBJDIR}/*.o"
+
+    INSTALL_COMMAND ""
+    )
+
+  add_custom_target(libzdnn
+    COMMAND ${CMAKE_COMMAND} -E copy_if_different
+            ${ZDNN_LIBDIR}/libzdnn.a ${NNPA_LIBRARY_PATH}/libzdnn.a
+    COMMAND ${CMAKE_COMMAND} -E copy_if_different
+            ${ZDNN_TOPDIR}/zdnn/zdnn.h ${NNPA_INCLUDE_PATH}/zdnn.h
+    DEPENDS zdnn
+    #BYPRODUCTS ${NNPA_LIBRARY_PATH}/libzdnn.a ${NNPA_INCLUDE_PATH}/zdnn.h
+    )
+
+  install(FILES ${NNPA_LIBRARY_PATH}/libzdnn.a DESTINATION lib)
+endfunction(setup_zdnn)