From 3f4461c266db39997935fec5d7dc14c91a610c31 Mon Sep 17 00:00:00 2001
From: Yifan Zhao <yifanz16@illinois.edu>
Date: Tue, 19 Jan 2021 06:50:58 -0600
Subject: [PATCH] Use CUDA-lang support to automate nvcc-gcc version check
---
hpvm/projects/hpvm-tensor-rt/CMakeLists.txt | 185 +++++++-------------
1 file changed, 63 insertions(+), 122 deletions(-)
diff --git a/hpvm/projects/hpvm-tensor-rt/CMakeLists.txt b/hpvm/projects/hpvm-tensor-rt/CMakeLists.txt
index fc4497ef91..492fdbacb4 100644
--- a/hpvm/projects/hpvm-tensor-rt/CMakeLists.txt
+++ b/hpvm/projects/hpvm-tensor-rt/CMakeLists.txt
@@ -1,50 +1,8 @@
cmake_minimum_required(VERSION 3.17)
-project(hpvm-tensor-rt)
-find_package(CUDA 9.1 REQUIRED)
-set(CUDA_SEPARABLE_COMPILATION ON CACHE BOOL "")
-set(CUDA_PROPAGATE_HOST_FLAGS OFF)
-
-if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
- # gcc > 8 are not supported
- if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 8)
- message(FATAL_ERROR "GCC versions later than 8 are not supported")
- endif()
-elseif("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
- # clang < 3.2 || clang >= 9 unsupported
- set(clang_v ${CMAKE_CXX_COMPILER_VERSION})
- if (clang_v VERSION_GREATER_EQUAL 9 OR clang_v VERSION_LESS_EQUAL 3.2)
- message(FATAL_ERROR "Clang<3.2 or clang>=9 are not supported")
- endif()
-endif()
-# Addresses a bug where code is not compiled as C++11 in non-CUDA code and older g++ versions
-# Edit: using c++14 now
+project(hpvm-tensor-rt CUDA CXX)
set(CMAKE_CXX_STANDARD 14)
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++14 -I/")
-set(
- CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};
- -gencode;arch=compute_60,code=sm_60;
- -gencode;arch=compute_60,code=compute_60;
- -std=c++14 --expt-relaxed-constexpr -maxrregcount 32 # These are for image ops
-)
-if(CMAKE_BUILD_TYPE STREQUAL "Debug")
- message("Debug mode")
- set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-g;-lineinfo;-Xcompiler;-ggdb)
-else()
- set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-DNDEBUG;-Xcompiler;-DNDEBUG)
-endif()
-# Default options
-if(USE_GFLAGS)
- add_definitions(-DUSE_GFLAGS)
-endif()
-if(USE_AUTOTUNER)
- remove_definitions(-DNO_INJECTION)
-endif()
-add_definitions(-DNO_INJECTION)
-add_definitions(-DPROMISE_TUNER_ENABLED)
-add_definitions(-DSIMULATION_MODE=true)
-
-# Config path configuration file
+# -- Config path configuration file
if(NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/global_knobs.txt)
message(FATAL_ERROR "global_knobs.txt not found")
endif()
@@ -56,76 +14,90 @@ configure_file(
${CMAKE_CURRENT_BINARY_DIR}/tensor_runtime/include/config.h
)
-# Default include/link directories
+# -- Default include directories
set(
INCLUDES
- $ENV{CUDNN_PATH} $ENV{CUDNN_PATH}/include
- ${CUDA_INCLUDE_DIRS}
+ ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
./tensor_runtime/include ${CMAKE_CURRENT_BINARY_DIR}/tensor_runtime/include
./dnn_sources/include
../gpu_profiler/include ../soc_simulator/include
)
-set(
- LINK_DIRS
- ${CUDA_TOOLKIT_ROOT_DIR}/lib64 $ENV{CUDNN_PATH}
- $ENV{CUDNN_PATH}/lib $ENV{CUDNN_PATH}/lib64
-)
-include_directories(${INCLUDES})
-link_directories(${LINK_DIRS})
-# Source files of runtime
+# -- Link libraries
+# Configure gpu_profiler and soc_simulator, and setup all libs to link to
+# Conditionally add gpu_profiler project if we're building independently
+# (not building the whole hpvm)
+if(NOT LLVM_BUILD_DIR) # Defined in ../CMakeLists.txt. This means we're compiling in LLVM
+ message(STATUS "Compiling hpvm-tensor-rt independently")
+ message(STATUS "Also compiling gpu_profiler and soc_simulator")
+ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../gpu_profiler ${CMAKE_CURRENT_BINARY_DIR}/gpu_profiler)
+ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../soc_simulator ${CMAKE_CURRENT_BINARY_DIR}/soc_simulator)
+endif()
+set(LINK_LIBS gpu_profiler promise_profiler stdc++fs cudnn curand cublas)
+if(USE_GFLAGS)
+ list(APPEND LINK_LIBS gflags)
+endif()
+find_package(OpenMP REQUIRED) # Provides ${OpenMP_CXX_FLAGS}
+
+# -- Definitions
+set(DEFS -DPROMISE_TUNER_ENABLED -DSIMULATION_MODE=true)
+if(USE_GFLAGS)
+ list(APPEND DEFS -DUSE_GFLAGS)
+endif()
+
+# -- Sources of runtime
set(
RUNTIME_SRCS_FILENAME
- approx_simulation.cu
- group_conv.cu
- approx_techniques.cu
+ approx_knobs_utils.cc approx_simulation.cu approx_techniques.cu
configuration.cpp
- debug.cc
- debug.cpp
- device_math.cu
+ debug.cc debug.cpp device_math.cu
error.cu
- tensor_cpu_runtime.cc
- fp16_gemm.cu
- global_data.cc
- half_precision_api.cu
- hpvm-rt-controller.cpp
+ fp16_gemm.cu freq_utils.cc
+ global_data.cc group_conv.cu
+ half_precision_api.cu hpvm-rt-controller.cpp
+ init_api.cc
op_overheads.cc
profiling.cc
- tensor_runtime.cu
- tensor_utils.cu
+ tensor_cpu_runtime.cc tensor_runtime.cu tensor_utils.cu
wrapper_runtime.cu
- approx_knobs_utils.cc
- init_api.cc
)
foreach(FILE ${RUNTIME_SRCS_FILENAME})
list(APPEND RUNTIME_SRCS "tensor_runtime/src/${FILE}")
+ # Some files doesn't end in .cu or .cuh, but we know they are still CUDA files
+ set_source_files_properties("tensor_runtime/src/${FILE}" PROPERTIES LANGUAGE CUDA)
endforeach()
-# Compile gpu_profiler and soc_simulator
-# Conditionally add gpu_profiler project if we're building independently
-# (not building the whole hpvm)
-get_filename_component(root_dir ${CMAKE_SOURCE_DIR} REALPATH)
-get_filename_component(our_dir ${CMAKE_CURRENT_SOURCE_DIR} REALPATH)
-if(${root_dir} STREQUAL ${our_dir})
- message(STATUS "Compiling hpvm-tensor-rt independently")
- message(STATUS "Also compiling gpu_profiler and soc_simulator")
- add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../gpu_profiler ${CMAKE_CURRENT_BINARY_DIR}/gpu_profiler)
- add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../soc_simulator ${CMAKE_CURRENT_BINARY_DIR}/soc_simulator)
-endif()
-set(LINK_LIBS gpu_profiler promise_profiler cudnn cufft stdc++fs curand)
-if(USE_GFLAGS)
- list(APPEND LINK_LIBS gflags)
-endif()
+# -- Adding tensor_runtime targets
+function(add_tensor_runtime target_name)
+ add_library(${target_name} ${RUNTIME_SRCS})
+ set_property(TARGET ${target_name} PROPERTY CUDA_ARCHITECTURES 60)
+ target_compile_options(
+ ${target_name} PRIVATE
+ $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr -maxrregcount 3>
+ $<$<AND:$<COMPILE_LANGUAGE:CUDA>,$<CONFIG:DEBUG>>:-lineinfo -Xcompiler -ggdb>
+ $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=${OpenMP_CXX_FLAGS}>
+ )
+ target_include_directories(${target_name} PUBLIC ${INCLUDES})
+ target_link_libraries(${target_name} PUBLIC ${LINK_LIBS})
+ target_compile_definitions(${target_name} PRIVATE ${DEFS} ${ARGN})
+endfunction(add_tensor_runtime)
# Adding new rule for building a cuDNN runtime library
# Offline version
-find_package(OpenMP REQUIRED)
-cuda_add_library(tensor_runtime ${RUNTIME_SRCS})
-cuda_add_cublas_to_target(tensor_runtime ${OpenMP_CXX_FLAGS})
-target_compile_options(tensor_runtime PRIVATE ${OpenMP_CXX_FLAGS})
-target_link_libraries(tensor_runtime ${LINK_LIBS} ${OpenMP_CXX_FLAGS})
-target_compile_definitions(tensor_runtime PRIVATE -DONLINE_PROFILING=false -DFP16_tuning=true)
+add_tensor_runtime(tensor_runtime -DONLINE_PROFILING=false -DFP16_tuning=true)
+# Online version
+add_tensor_runtime(tensor_runtime_online -DONLINE_PROFILING=true -DFP16_tuning=false)
+# tensor_runtime_online is built AFTER tensor_runtime because of a nvcc bug (bug?)
+# that doesn't allow compiling the same file from multiple targets at once.
+# Same for tensor_runtime_online.
+add_dependencies(tensor_runtime_online tensor_runtime)
+
+# Adding rule for the debugging source
+add_executable(unit_tests dnn_sources/src/unit_tests.cc)
+target_link_libraries(unit_tests tensor_runtime_online ${GPU_PROFILER_LIB} ${SOC_SIMULATOR_LIB})
+
+# -- Compile tensor_runtime.ll if possible
if(LLVM_BUILD_DIR) # Defined in ../CMakeLists.txt. This means we're compiling in LLVM
get_filename_component(LLVM_CLANG_XX ${LLVM_BUILD_DIR}/bin/clang++ REALPATH)
# It's important that tensor_runtime.ll goes here if we're compiling with LLVM
@@ -158,31 +130,6 @@ if(LLVM_CLANG_XX)
)
endif()
-# Install version (also offline)
-cuda_add_library(tensor_runtime_install ${RUNTIME_SRCS})
-cuda_add_cublas_to_target(tensor_runtime_install)
-# tensor_runtime_install is built AFTER tensor_runtime because of a nvcc bug (bug?)
-# that doesn't allow compiling the same file from multiple targets at once.
-# Same for tensor_runtime_online.
-add_dependencies(tensor_runtime_install tensor_runtime)
-target_link_libraries(tensor_runtime_install ${LINK_LIBS})
-target_compile_definitions(tensor_runtime_install PRIVATE -DONLINE_PROFILING=false -DFP16_tuning=true)
-
-# Online version
-cuda_add_library(tensor_runtime_online ${RUNTIME_SRCS})
-cuda_add_cublas_to_target(tensor_runtime_online ${OpenMP_CXX_FLAGS})
-target_compile_options(tensor_runtime_online PRIVATE ${OpenMP_CXX_FLAGS})
-add_dependencies(tensor_runtime_online tensor_runtime)
-target_link_libraries(tensor_runtime_online ${LINK_LIBS} ${OpenMP_CXX_FLAGS})
-target_compile_definitions(tensor_runtime_online PRIVATE -DONLINE_PROFILING=true -DFP16_tuning=false)
-
-
-
-# -------------- Unit Test Source ----------------
-
-add_executable(unit_tests dnn_sources/src/unit_tests.cc)
-target_link_libraries(unit_tests tensor_runtime_online ${GPU_PROFILER_LIB} ${SOC_SIMULATOR_LIB})
-
#**************** FP32 TensorRT Source Builds ***********
@@ -216,9 +163,6 @@ target_link_libraries(vgg16_imagenet_fp32 tensor_runtime_online ${GPU_PROFILER
add_executable(resnet50_imagenet_fp32 dnn_sources/src/fp32/resnet50_imagenet.cc)
target_link_libraries(resnet50_imagenet_fp32 tensor_runtime_online ${GPU_PROFILER_LIB} ${SOC_SIMULATOR_LIB})
-
-
-
#********* FP16 TensorRT Source Builds ******
add_executable(lenet_mnist_fp16 dnn_sources/src/fp16/lenet_mnist_half.cc)
@@ -243,12 +187,9 @@ add_executable(mobilenet_cifar10_fp16 dnn_sources/src/fp16/mobilenet_half.cc)
target_link_libraries(mobilenet_cifar10_fp16 tensor_runtime_online ${GPU_PROFILER_LIB} ${SOC_SIMULATOR_LIB})
-
-
file(GLOB files "dnn_sources/src/dynamic/*.cpp")
foreach(file ${files})
get_filename_component(stem ${file} NAME_WE)
add_executable(${stem} ${file})
target_link_libraries(${stem} tensor_runtime ${GPU_PROFILER_LIB} ${SOC_SIMULATOR_LIB})
endforeach()
-
--
GitLab