From a8aa64733cf621b7c5ef83070d2a1bc409234e39 Mon Sep 17 00:00:00 2001
From: Yifan Zhao <yifanz16@illinois.edu>
Date: Mon, 1 Feb 2021 17:48:13 -0600
Subject: [PATCH] Moved hpvm-c test to use approxhpvm.py to compile as well
---
.../test/dnn_benchmarks/hpvm-c/CMakeLists.txt | 136 ++++--------------
1 file changed, 26 insertions(+), 110 deletions(-)
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/CMakeLists.txt b/hpvm/test/dnn_benchmarks/hpvm-c/CMakeLists.txt
index 436c88790c..00e5ba507d 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/CMakeLists.txt
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/CMakeLists.txt
@@ -1,117 +1,35 @@
-# Look for CUDA again (already done in hpvm-tensor-rt) so we can include its header
-# This is not the best practice,
-# but easier than having the tensor runtime tell us which CUDA it used.
-find_package(CUDA REQUIRED)
+# First get approxhpvm.py which we then use to compile benchmarks.
+get_filename_component(APPROXHPVM_PY ${PROJECT_BINARY_DIR}/bin/approxhpvm.py REALPATH)
-get_filename_component(LLVM_BIN_DIR ${PROJECT_BINARY_DIR}/bin REALPATH)
-set(LLVM_OPT "${LLVM_BIN_DIR}/opt")
-set(LLVM_LINK "${LLVM_BIN_DIR}/llvm-link")
-set(CMAKE_CXX_COMPILER "${LLVM_BIN_DIR}/clang++")
-
-# Configure config.h which tells the benchmarks where's the build directory.
-# We can also use the one in tensor_runtime, but we're avoiding that trying to
+# Configure config.h which tells the benchmarks where's the model parameter directory.
+# We can also use the one in tensor_runtime, but we're avoiding that so as to
# decouple things.
set(MODEL_PARAMS_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../model_params/")
configure_file(
"include/config.h.in"
"${CMAKE_CURRENT_BINARY_DIR}/include/config.h"
)
+# This will be an extra include directory (specific to these benchmarks)
+# and we'll give this to approxhpvm.py
set(CONFIG_INCLUDE_DIR "${CMAKE_CURRENT_BINARY_DIR}/include")
-# Directories to include
-set(HPVM_PROJECTS ${PROJECT_SOURCE_DIR}/tools/hpvm/projects)
-set(HPVM_INCLUDE_DIR ${PROJECT_SOURCE_DIR}/tools/hpvm/include)
-set(BENCHMARK_INCLUDE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../dnn_benchmarks/include)
-set(TENSOR_INCLUDE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/include)
-set(TENSOR_RT_INCLUDE_DIR ${HPVM_PROJECTS}/hpvm-tensor-rt/tensor_runtime/include)
-set(
- INCLUDES
- ${CONFIG_INCLUDE_DIR} ${HPVM_INCLUDE_DIR} ${BENCHMARK_INCLUDE_DIR}
- ${TENSOR_INCLUDE_DIR} ${TENSOR_RT_INCLUDE_DIR} ${CUDA_INCLUDE_DIRS}
-)
-foreach(dir ${INCLUDES})
- list(APPEND INCLUDE_COMPILER_STRINGS "-I${dir}")
-endforeach()
-
-# Built-in libraries to link
-list(
- APPEND LINKER_FLAGS
- "-L${CUDA_TOOLKIT_ROOT_DIR}/lib64"
- -lpthread -lcudart -lcurand -lcudnn -lcublas -lcufft -lOpenCL -lstdc++fs -lomp -lm
-)
-
-# The hpvm-rt runtime
-# This has to be explicitly set as hpvm-rt.bc is created in a custom_target
-# and does not export its file location.
-# Keep this in sync with hpvm/projects/hpvm-rt/CMakeLists.txt.
-set(HPVM_RT_PATH ${PROJECT_BINARY_DIR}/tools/hpvm/projects/hpvm-rt/hpvm-rt.bc)
-
-# Compile flags (clang++)
-set(CLANG_FLAGS -fno-exceptions -std=c++11 -O3)
-
-# All compilation uses HPVM_DEFAULT_PASSES.
-set(
- HPVM_DEFAULT_PASSES
- LLVMBuildDFG
- LLVMInPlaceDFGAnalysis
- LLVMDFG2LLVM_CPU
- LLVMFuseHPVMTensorNodes
- LLVMClearDFG
- LLVMGenHPVM
-)
-
-set(WORK_DIR ${CMAKE_CURRENT_BINARY_DIR})
set(test_compile_targets "")
-function(compile_single_benchmark target src_file extra_passes extra_dfg_flags)
- foreach(pass ${HPVM_DEFAULT_PASSES} ${extra_passes})
- list(APPEND LOAD_FILE_FLAGS "-load" "${pass}.so")
- endforeach()
- set(
- HPVM_PASSES ${LOAD_FILE_FLAGS}
- -buildDFG -inplace -hpvm-fuse ${extra_dfg_flags} -dfg2llvm-cpu -clearDFG
- )
-
+function(approxhpvm_py_codegen bin_filename src_filepath codegen_target)
add_custom_command(
- OUTPUT "${WORK_DIR}/${target}.ll" DEPENDS ${src_file} clang
- COMMAND ${CMAKE_CXX_COMPILER} ${INCLUDE_COMPILER_STRINGS} ${CLANG_FLAGS} -emit-llvm -S ${src_file}
- -o ${WORK_DIR}/${target}.ll
+ OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${bin_filename}
+ DEPENDS ${src_filepath} approxhpvm.py
+ COMMAND ${APPROXHPVM_PY}
+ ${src_filepath} ${CMAKE_CURRENT_BINARY_DIR}/${bin_filename}
+ -t ${codegen_target} -I ${CONFIG_INCLUDE_DIR} ${ARGV}
)
- add_custom_command(
- OUTPUT "${WORK_DIR}/${target}.hpvm.ll"
- DEPENDS "${WORK_DIR}/${target}.ll" opt LLVMGenHPVM
- COMMAND ${LLVM_OPT} -load LLVMGenHPVM.so -genhpvm -globaldce -S ${WORK_DIR}/${target}.ll
- -o ${WORK_DIR}/${target}.hpvm.ll
- )
- add_custom_command(
- OUTPUT "${WORK_DIR}/${target}.llvm.ll"
- DEPENDS "${WORK_DIR}/${target}.hpvm.ll" opt ${HPVM_DEFAULT_PASSES} ${extra_passes}
- COMMAND ${LLVM_OPT} ${HPVM_PASSES} -S ${WORK_DIR}/${target}.hpvm.ll
- -o ${WORK_DIR}/${target}.llvm.ll
- )
- add_custom_command(
- OUTPUT "${WORK_DIR}/${target}.linked.bc"
- DEPENDS "${WORK_DIR}/${target}.llvm.ll" hpvm-rt.bc llvm-link
- COMMAND ${LLVM_LINK} ${WORK_DIR}/${target}.llvm.ll ${HPVM_RT_PATH}
- -o ${WORK_DIR}/${target}.linked.bc
- )
- add_custom_command(
- OUTPUT "${WORK_DIR}/${target}"
- DEPENDS "${WORK_DIR}/${target}.linked.bc" tensor_runtime gpu_profiler promise_profiler
- COMMAND ${CMAKE_CXX_COMPILER}
- ${WORK_DIR}/${target}.linked.bc
- $<TARGET_FILE:tensor_runtime> $<TARGET_FILE:gpu_profiler> $<TARGET_FILE:promise_profiler>
- -o ${WORK_DIR}/${target} ${LINKER_FLAGS}
- )
- add_custom_target(${target} DEPENDS "${WORK_DIR}/${target}")
-
- set(test_compile_targets ${test_compile_targets} ${target} PARENT_SCOPE)
-endfunction(compile_single_benchmark)
+ add_custom_target(${bin_filename} DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${bin_filename})
+ set(test_compile_targets ${test_compile_targets} ${bin_filename} PARENT_SCOPE)
+endfunction(approxhpvm_py_codegen)
set(test_run_targets "")
function(run_single_benchmark run_target benchmark)
add_custom_target(
- ${run_target}
- COMMAND ${WORK_DIR}/${benchmark}
+ ${run_target} COMMAND ${CMAKE_CURRENT_BINARY_DIR}/${benchmark}
)
add_dependencies(${run_target} ${benchmark})
set(test_run_targets ${test_run_targets} ${run_target} PARENT_SCOPE)
@@ -120,22 +38,20 @@ endfunction(run_single_benchmark)
file(GLOB entries ./benchmarks/*)
foreach(dir ${entries})
get_filename_component(dirname "${dir}" NAME)
- set(
- loop_extra_flags
- -dfg2llvm-wrapperapi
- -quantization-levels-filename=${dir}/data/quant_ranges_rt.txt
- -configuration-inputs-filename=${dir}/data/tuner_confs.txt
- )
- compile_single_benchmark(
- ${dirname} ${dir}/${dirname}.cpp
- LLVMDFG2LLVM_WrapperAPI "${loop_extra_flags}"
+ # Generate "tensor"-targeted code
+ approxhpvm_py_codegen(
+ ${dirname} ${dir}/${dirname}.cpp tensor
+ --quant-file ${dir}/data/quant_ranges_rt.txt
+ --config-file ${dir}/data/tuner_confs.txt
)
+ # Run tensor binary
run_single_benchmark(run_${dirname} ${dirname})
- compile_single_benchmark(
- ${dirname}_cudnn ${dir}/${dirname}_cudnn.cpp LLVMDFG2LLVM_CUDNN -dfg2llvm-cudnn
- )
+ # Generate "cudnn"-targeted code
+ approxhpvm_py_codegen(${dirname}_cudnn ${dir}/${dirname}_cudnn.cpp cudnn)
+ # Run cudnn binary
run_single_benchmark(run_${dirname}_cudnn ${dirname}_cudnn)
endforeach(dir)
+
message(STATUS "List of test dnn benchmarks: ${test_compile_targets}")
add_custom_target(dnn_benchmarks DEPENDS ${test_compile_targets})
message(STATUS "Target name for compiling all dnn benchmarks: dnn_benchmarks")
--
GitLab