1# Copyright (c) 2017-2018, Lawrence Livermore National Security, LLC. 2# Produced at the Lawrence Livermore National Laboratory. LLNL-CODE-734707. 3# All Rights reserved. See files LICENSE and NOTICE for details. 4# 5# This file is part of CEED, a collection of benchmarks, miniapps, software 6# libraries and APIs for efficient high-order finite element and spectral 7# element discretizations for exascale applications. For more information and 8# source code availability see http://github.com/ceed. 9# 10# The CEED research is supported by the Exascale Computing Project 17-SC-20-SC, 11# a collaborative effort of two U.S. Department of Energy organizations (Office 12# of Science and the National Nuclear Security Administration) responsible for 13# the planning and preparation of a capable exascale ecosystem, including 14# software, applications, hardware, advanced system engineering and early 15# testbed platforms, in support of the nation's exascale computing imperative. 16 17CONFIG ?= config.mk 18-include $(CONFIG) 19COMMON ?= common.mk 20-include $(COMMON) 21 22ifeq (,$(filter-out undefined default,$(origin CC))) 23 CC = gcc 24endif 25ifeq (,$(filter-out undefined default,$(origin CXX))) 26 CXX = g++ 27endif 28ifeq (,$(filter-out undefined default,$(origin FC))) 29 FC = gfortran 30endif 31ifeq (,$(filter-out undefined default,$(origin LINK))) 32 LINK = $(CC) 33endif 34ifeq (,$(filter-out undefined default,$(origin AR))) 35 AR = ar 36endif 37ifeq (,$(filter-out undefined default,$(origin ARFLAGS))) 38 ARFLAGS = crD 39endif 40NVCC ?= $(CUDA_DIR)/bin/nvcc 41NVCC_CXX ?= $(CXX) 42HIPCC ?= $(HIP_DIR)/bin/hipcc 43SED ?= sed 44 45# ASAN must be left empty if you don't want to use it 46ASAN ?= 47 48LDFLAGS ?= 49UNDERSCORE ?= 1 50 51# MFEM_DIR env variable should point to sibling directory 52ifneq ($(wildcard ../mfem/libmfem.*),) 53 MFEM_DIR ?= ../mfem 54endif 55 56# NEK5K_DIR env variable should point to sibling directory 57ifneq ($(wildcard ../Nek5000/*),) 58 NEK5K_DIR ?= $(abspath ../Nek5000) 59endif 60export NEK5K_DIR 61MPI ?= 1 62 63# CEED_DIR env for NEK5K testing 64export CEED_DIR = $(abspath .) 65 66# XSMM_DIR env variable should point to XSMM master (github.com/hfp/libxsmm) 67XSMM_DIR ?= ../libxsmm 68 69# OCCA_DIR env variable should point to OCCA master (github.com/libocca/occa) 70OCCA_DIR ?= ../occa 71 72# env variable MAGMA_DIR can be used too 73MAGMA_DIR ?= ../magma 74 75# Often /opt/cuda or /usr/local/cuda, but sometimes present on machines that don't support CUDA 76CUDA_DIR ?= 77CUDA_ARCH ?= 78 79# Often /opt/rocm, but sometimes present on machines that don't support HIP 80HIP_DIR ?= 81HIP_ARCH ?= 82 83# Check for PETSc in ../petsc 84ifneq ($(wildcard ../petsc/lib/libpetsc.*),) 85 PETSC_DIR ?= ../petsc 86endif 87 88# Warning: SANTIZ options still don't run with /gpu/occa 89# export LSAN_OPTIONS=suppressions=.asanignore 90AFLAGS = -fsanitize=address #-fsanitize=undefined -fno-omit-frame-pointer 91 92# Note: Intel oneAPI C/C++ compiler is now icx/icpx 93CC_VENDOR := $(subst oneAPI,icc,$(patsubst gcc%,gcc,$(firstword $(filter gcc% clang icc oneAPI XL,$(shell $(CC) --version))))) 94FC_VENDOR := $(firstword $(filter GNU ifort XL,$(shell $(FC) --version 2>&1 || $(FC) -qversion))) 95 96# Default extra flags by vendor 97MARCHFLAG.gcc := -march=native 98MARCHFLAG.clang := $(MARCHFLAG.gcc) 99MARCHFLAG.icc := 100OMP_SIMD_FLAG.gcc := -fopenmp-simd 101OMP_SIMD_FLAG.clang := $(OMP_SIMD_FLAG.gcc) 102OMP_SIMD_FLAG.icc := -qopenmp-simd 103OPT.gcc := -ffp-contract=fast 104OPT.clang := $(OPT.gcc) 105CFLAGS.gcc := -fPIC -std=c99 -Wall -Wextra -Wno-unused-parameter -MMD -MP 106CFLAGS.clang := $(CFLAGS.gcc) 107CFLAGS.icc := $(CFLAGS.gcc) 108CFLAGS.XL := -qpic -MMD 109CXXFLAGS.gcc := -fPIC -std=c++11 -Wall -Wextra -Wno-unused-parameter -MMD -MP 110CXXFLAGS.clang := $(CXXFLAGS.gcc) 111CXXFLAGS.icc := $(CXXFLAGS.gcc) 112CXXFLAGS.XL := -qpic -std=c++11 -MMD 113FFLAGS.GNU := -fPIC -cpp -Wall -Wextra -Wno-unused-parameter -Wno-unused-dummy-argument -MMD -MP 114FFLAGS.ifort := -fPIC -cpp 115FFLAGS.XL := -qpic -ffree-form -qpreprocess -qextname -MMD 116 117# This check works with compilers that use gcc and clang. It fails with some 118# compilers; e.g., xlc apparently ignores all options when -E is passed, thus 119# succeeds with any flags. Users can pass MARCHFLAG=... if desired. 120cc_check_flag = $(shell $(CC) -E -Werror $(1) -x c /dev/null > /dev/null 2>&1 && echo 1) 121MARCHFLAG := $(MARCHFLAG.$(CC_VENDOR)) 122MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG),-mcpu=native) 123MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG)) 124 125OMP_SIMD_FLAG := $(OMP_SIMD_FLAG.$(CC_VENDOR)) 126OMP_SIMD_FLAG := $(if $(call cc_check_flag,$(OMP_SIMD_FLAG)),$(OMP_SIMD_FLAG)) 127 128OPT ?= -O -g $(MARCHFLAG) $(OPT.$(CC_VENDOR)) $(OMP_SIMD_FLAG) 129CFLAGS ?= $(OPT) $(CFLAGS.$(CC_VENDOR)) 130CXXFLAGS ?= $(OPT) $(CXXFLAGS.$(CC_VENDOR)) 131LIBCXX ?= -lstdc++ 132NVCCFLAGS ?= -ccbin $(CXX) -Xcompiler "$(OPT)" -Xcompiler -fPIC 133ifneq ($(CUDA_ARCH),) 134 NVCCFLAGS += -arch=$(CUDA_ARCH) 135endif 136HIPCCFLAGS ?= $(filter-out $(OMP_SIMD_FLAG),$(OPT)) -fPIC 137ifneq ($(HIP_ARCH),) 138 HIPCCFLAGS += --amdgpu-target=$(HIP_ARCH) 139endif 140FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR)) 141 142ifeq ($(COVERAGE), 1) 143 CFLAGS += --coverage 144 CXXFLAGS += --coverage 145 LDFLAGS += --coverage 146endif 147 148CFLAGS += $(if $(ASAN),$(AFLAGS)) 149FFLAGS += $(if $(ASAN),$(AFLAGS)) 150LDFLAGS += $(if $(ASAN),$(AFLAGS)) 151CPPFLAGS += -I./include 152LDLIBS = -lm 153OBJDIR := build 154LIBDIR := lib 155 156# Installation variables 157prefix ?= /usr/local 158bindir = $(prefix)/bin 159libdir = $(prefix)/lib 160includedir = $(prefix)/include 161pkgconfigdir = $(libdir)/pkgconfig 162INSTALL = install 163INSTALL_PROGRAM = $(INSTALL) 164INSTALL_DATA = $(INSTALL) -m644 165 166# Get number of processors of the machine 167NPROCS := $(shell getconf _NPROCESSORS_ONLN) 168# prepare make options to run in parallel 169MFLAGS := -j $(NPROCS) --warn-undefined-variables \ 170 --no-print-directory --no-keep-going 171 172PYTHON ?= python3 173PROVE ?= prove 174PROVE_OPTS ?= -j $(NPROCS) 175DARWIN := $(filter Darwin,$(shell uname -s)) 176SO_EXT := $(if $(DARWIN),dylib,so) 177 178ceed.pc := $(LIBDIR)/pkgconfig/ceed.pc 179libceed.so := $(LIBDIR)/libceed.$(SO_EXT) 180libceed.a := $(LIBDIR)/libceed.a 181libceed := $(if $(STATIC),$(libceed.a),$(libceed.so)) 182CEED_LIBS = -lceed 183libceed.c := $(filter-out interface/ceed-cuda.c interface/ceed-hip.c, $(wildcard interface/ceed*.c backends/*.c gallery/*.c)) 184gallery.c := $(wildcard gallery/*/ceed*.c) 185libceed.c += $(gallery.c) 186libceed_test.so := $(LIBDIR)/libceed_test.$(SO_EXT) 187libceed_test.a := $(LIBDIR)/libceed_test.a 188libceed_test := $(if $(STATIC),$(libceed_test.a),$(libceed_test.so)) 189libceeds = $(libceed) $(libceed_test) 190BACKENDS_BUILTIN := /cpu/self/ref/serial /cpu/self/ref/blocked /cpu/self/opt/serial /cpu/self/opt/blocked 191BACKENDS := $(BACKENDS_BUILTIN) 192 193# Tests 194tests.c := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.c)) 195tests.f := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.f90)) 196tests := $(tests.c:tests/%.c=$(OBJDIR)/%) 197ctests := $(tests) 198tests += $(tests.f:tests/%.f90=$(OBJDIR)/%) 199# Examples 200examples.c := $(sort $(wildcard examples/ceed/*.c)) 201examples.f := $(sort $(wildcard examples/ceed/*.f)) 202examples := $(examples.c:examples/ceed/%.c=$(OBJDIR)/%) 203examples += $(examples.f:examples/ceed/%.f=$(OBJDIR)/%) 204# MFEM Examples 205mfemexamples.cpp := $(sort $(wildcard examples/mfem/*.cpp)) 206mfemexamples := $(mfemexamples.cpp:examples/mfem/%.cpp=$(OBJDIR)/mfem-%) 207# Nek5K Examples 208nekexamples := $(OBJDIR)/nek-bps 209# PETSc Examples 210petscexamples.c := $(wildcard examples/petsc/*.c) 211petscexamples := $(petscexamples.c:examples/petsc/%.c=$(OBJDIR)/petsc-%) 212# Fluid Dynamics Examples 213fluidsexamples.c := $(sort $(wildcard examples/fluids/*.c)) 214fluidsexamples := $(fluidsexamples.c:examples/fluids/%.c=$(OBJDIR)/fluids-%) 215# Solid Mechanics Examples 216solidsexamples.c := $(sort $(wildcard examples/solids/*.c)) 217solidsexamples := $(solidsexamples.c:examples/solids/%.c=$(OBJDIR)/solids-%) 218 219# Backends/[ref, blocked, template, memcheck, opt, avx, occa, magma] 220ref.c := $(sort $(wildcard backends/ref/*.c)) 221blocked.c := $(sort $(wildcard backends/blocked/*.c)) 222template.c := $(sort $(wildcard backends/template/*.c)) 223ceedmemcheck.c := $(sort $(wildcard backends/memcheck/*.c)) 224opt.c := $(sort $(wildcard backends/opt/*.c)) 225avx.c := $(sort $(wildcard backends/avx/*.c)) 226xsmm.c := $(sort $(wildcard backends/xsmm/*.c)) 227cuda.c := $(sort $(wildcard backends/cuda/*.c)) 228cuda.cpp := $(sort $(wildcard backends/cuda/*.cpp)) 229cuda.cu := $(sort $(wildcard backends/cuda/kernels/*.cu)) 230cuda-shared.c := $(sort $(wildcard backends/cuda-shared/*.c)) 231cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/kernels/*.cu)) 232cuda-gen.c := $(sort $(wildcard backends/cuda-gen/*.c)) 233cuda-gen.cpp := $(sort $(wildcard backends/cuda-gen/*.cpp)) 234cuda-gen.cu := $(sort $(wildcard backends/cuda-gen/kernels/*.cu)) 235occa.cpp := $(sort $(shell find backends/occa -type f -name *.cpp)) 236magma.c := $(sort $(wildcard backends/magma/*.c)) 237magma.cu := $(sort $(wildcard backends/magma/kernels/cuda/*.cu)) 238magma.hip := $(sort $(wildcard backends/magma/kernels/hip/*.hip.cpp)) 239hip.c := $(sort $(wildcard backends/hip/*.c)) 240hip.cpp := $(sort $(wildcard backends/hip/*.cpp)) 241hip.hip := $(sort $(wildcard backends/hip/kernels/*.hip.cpp)) 242hip-shared.c := $(sort $(wildcard backends/hip-shared/*.c)) 243hip-gen.c := $(sort $(wildcard backends/hip-gen/*.c)) 244hip-gen.cpp := $(sort $(wildcard backends/hip-gen/*.cpp)) 245 246# Quiet, color output 247quiet ?= $($(1)) 248 249# Cancel built-in and old-fashioned implicit rules which we don't use 250.SUFFIXES: 251 252.SECONDEXPANSION: # to expand $$(@D)/.DIR 253 254%/.DIR : 255 @mkdir -p $(@D) 256 @touch $@ 257 258.PRECIOUS: %/.DIR 259 260lib: $(libceed) $(ceed.pc) 261# run 'lib' target in parallel 262par:;@$(MAKE) $(MFLAGS) V=$(V) lib 263backend_status = $(if $(filter $1,$(BACKENDS)), [backends: $1], [not found]) 264info: 265 $(info ------------------------------------) 266 $(info CC = $(CC)) 267 $(info CXX = $(CXX)) 268 $(info FC = $(FC)) 269 $(info CPPFLAGS = $(CPPFLAGS)) 270 $(info CFLAGS = $(value CFLAGS)) 271 $(info CXXFLAGS = $(value CXXFLAGS)) 272 $(info FFLAGS = $(value FFLAGS)) 273 $(info NVCCFLAGS = $(value NVCCFLAGS)) 274 $(info HIPCCFLAGS = $(value HIPCCFLAGS)) 275 $(info LDFLAGS = $(value LDFLAGS)) 276 $(info LDLIBS = $(LDLIBS)) 277 $(info AR = $(AR)) 278 $(info ARFLAGS = $(ARFLAGS)) 279 $(info OPT = $(OPT)) 280 $(info AFLAGS = $(AFLAGS)) 281 $(info ASAN = $(or $(ASAN),(empty))) 282 $(info V = $(or $(V),(empty)) [verbose=$(if $(V),on,off)]) 283 $(info ------------------------------------) 284 $(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,$(MEMCHK_BACKENDS))) 285 $(info AVX_STATUS = $(AVX_STATUS)$(call backend_status,$(AVX_BACKENDS))) 286 $(info XSMM_DIR = $(XSMM_DIR)$(call backend_status,$(XSMM_BACKENDS))) 287 $(info OCCA_DIR = $(OCCA_DIR)$(call backend_status,$(OCCA_BACKENDS))) 288 $(info MAGMA_DIR = $(MAGMA_DIR)$(call backend_status,$(MAGMA_BACKENDS))) 289 $(info CUDA_DIR = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS))) 290 $(info HIP_DIR = $(HIP_DIR)$(call backend_status,$(HIP_BACKENDS))) 291 $(info ------------------------------------) 292 $(info MFEM_DIR = $(MFEM_DIR)) 293 $(info NEK5K_DIR = $(NEK5K_DIR)) 294 $(info PETSC_DIR = $(PETSC_DIR)) 295 $(info ------------------------------------) 296 $(info prefix = $(prefix)) 297 $(info includedir = $(value includedir)) 298 $(info libdir = $(value libdir)) 299 $(info pkgconfigdir = $(value pkgconfigdir)) 300 $(info ------------------------------------) 301 @true 302info-backends: 303 $(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS))) 304.PHONY: lib all par info info-backends 305 306$(libceed.so) : LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed.so))) 307$(libceed_test) : LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed_test))) 308 309# Standard Backends 310libceed.c += $(ref.c) 311libceed.c += $(blocked.c) 312libceed.c += $(opt.c) 313 314# Testing Backends 315test_backends.c := $(template.c) 316TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub 317 318# Memcheck Backend 319MEMCHK_STATUS = Disabled 320MEMCHK := $(shell echo "$(HASH)include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1) 321MEMCHK_BACKENDS = /cpu/self/memcheck/serial /cpu/self/memcheck/blocked 322ifeq ($(MEMCHK),1) 323 MEMCHK_STATUS = Enabled 324 libceed.c += $(ceedmemcheck.c) 325 BACKENDS += $(MEMCHK_BACKENDS) 326endif 327 328# AVX Backed 329AVX_STATUS = Disabled 330AVX_FLAG := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx) 331AVX := $(filter $(AVX_FLAG),$(shell $(CC) $(OPT) -v -E -x c /dev/null 2>&1)) 332AVX_BACKENDS = /cpu/self/avx/serial /cpu/self/avx/blocked 333ifneq ($(AVX),) 334 AVX_STATUS = Enabled 335 libceed.c += $(avx.c) 336 BACKENDS += $(AVX_BACKENDS) 337endif 338 339# Collect list of libraries and paths for use in linking and pkg-config 340PKG_LIBS = 341 342# libXSMM Backends 343XSMM_BACKENDS = /cpu/self/xsmm/serial /cpu/self/xsmm/blocked 344ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),) 345 PKG_LIBS += -L$(abspath $(XSMM_DIR))/lib -lxsmm -ldl 346 MKL ?= 347 ifeq (,$(MKL)$(MKLROOT)) 348 BLAS_LIB = -lblas 349 else 350 ifneq ($(MKLROOT),) 351 # Some installs put everything inside an intel64 subdirectory, others not 352 MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*))) 353 MKL_LINK = -L$(MKL_LIBDIR) 354 PKG_LIB_DIRS += $(MKL_LIBDIR) 355 endif 356 BLAS_LIB = $(MKL_LINK) -Wl,--push-state,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl -Wl,--pop-state 357 endif 358 PKG_LIBS += $(BLAS_LIB) 359 libceed.c += $(xsmm.c) 360 $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include 361 BACKENDS += $(XSMM_BACKENDS) 362endif 363 364# OCCA Backends 365OCCA_BACKENDS = /cpu/self/occa 366ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),) 367 OCCA_MODES := $(shell $(OCCA_DIR)/bin/occa modes) 368 OCCA_BACKENDS += $(if $(filter OpenMP,$(OCCA_MODES)),/cpu/openmp/occa) 369# OCCA_BACKENDS += $(if $(filter OpenCL,$(OCCA_MODES)),/gpu/opencl/occa) 370 OCCA_BACKENDS += $(if $(filter HIP,$(OCCA_MODES)),/gpu/hip/occa) 371 OCCA_BACKENDS += $(if $(filter CUDA,$(OCCA_MODES)),/gpu/cuda/occa) 372 373 $(libceeds) : CPPFLAGS += -I$(OCCA_DIR)/include 374 PKG_LIBS += -L$(abspath $(OCCA_DIR))/lib -locca 375 LIBCEED_CONTAINS_CXX = 1 376 libceed.cpp += $(occa.cpp) 377 BACKENDS += $(OCCA_BACKENDS) 378endif 379 380# CUDA Backends 381CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(CUDA_DIR)/$d/libcudart.${SO_EXT})) 382CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR)))) 383CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs 384CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/shared /gpu/cuda/gen 385ifneq ($(CUDA_LIB_DIR),) 386 $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include 387 PKG_LIBS += -L$(abspath $(CUDA_LIB_DIR)) -lcudart -lnvrtc -lcuda -lcublas 388 LIBCEED_CONTAINS_CXX = 1 389 libceed.c += interface/ceed-cuda.c 390 libceed.c += $(cuda.c) $(cuda-shared.c) $(cuda-gen.c) 391 libceed.cpp += $(cuda.cpp) $(cuda-gen.cpp) 392 libceed.cu += $(cuda.cu) $(cuda-shared.cu) $(cuda-gen.cu) 393 BACKENDS += $(CUDA_BACKENDS) 394endif 395 396# HIP Backends 397HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libamdhip64.${SO_EXT})) 398HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 399HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen 400ifneq ($(HIP_LIB_DIR),) 401 $(libceeds) : HIPCCFLAGS += -I./include 402 ifneq ($(CXX), $(HIPCC)) 403 CPPFLAGS += $(subst =,,$(shell $(HIP_DIR)/bin/hipconfig -C)) 404 endif 405 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include 406 PKG_LIBS += -L$(abspath $(HIP_LIB_DIR)) -lamdhip64 -lhipblas 407 LIBCEED_CONTAINS_CXX = 1 408 libceed.c += interface/ceed-hip.c 409 libceed.c += $(hip.c) $(hip-shared.c) $(hip-gen.c) 410 libceed.cpp += $(hip.cpp) $(hip-gen.cpp) 411 libceed.hip += $(hip.hip) 412 BACKENDS += $(HIP_BACKENDS) 413endif 414 415# MAGMA Backend 416ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 417 MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas") 418 ifeq ($(MAGMA_ARCH), 0) #CUDA MAGMA 419 ifneq ($(CUDA_LIB_DIR),) 420 cuda_link = -Wl,-rpath,$(CUDA_LIB_DIR) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart 421 omp_link = -fopenmp 422 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link) 423 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 424 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 425 PKG_LIBS += $(magma_link) 426 libceed.c += $(magma.c) 427 libceed.cu += $(magma.cu) 428 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 429 $(magma.cu:%.cu=$(OBJDIR)/%.o) : CPPFLAGS += --compiler-options=-fPIC -DADD_ -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(MAGMA_DIR)/control -I$(CUDA_DIR)/include 430 MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det 431 endif 432 else # HIP MAGMA 433 ifneq ($(HIP_LIB_DIR),) 434 hip_link = -Wl,-rpath,$(HIP_LIB_DIR) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64 435 omp_link = -fopenmp 436 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link) 437 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 438 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 439 PKG_LIBS += $(magma_link) 440 libceed.c += $(magma.c) 441 libceed.hip += $(magma.hip) 442 ifneq ($(CXX), $(HIPCC)) 443 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 444 else 445 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 446 endif 447 $(magma.hip:%.hip.cpp=$(OBJDIR)/%.o) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(MAGMA_DIR)/control -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 448 MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det 449 endif 450 endif 451 LIBCEED_CONTAINS_CXX = 1 452 BACKENDS += $(MAGMA_BACKENDS) 453endif 454 455export BACKENDS 456 457_pkg_ldflags = $(filter -L%,$(PKG_LIBS)) 458_pkg_ldlibs = $(filter-out -L%,$(PKG_LIBS)) 459$(libceeds) : LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) 460$(libceeds) : LDLIBS += $(_pkg_ldlibs) 461ifeq ($(STATIC),1) 462$(examples) $(tests) : LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) 463$(examples) $(tests) : LDLIBS += $(_pkg_ldlibs) 464endif 465 466pkgconfig-libs-private = $(PKG_LIBS) 467ifeq ($(LIBCEED_CONTAINS_CXX),1) 468 $(libceeds) : LINK = $(CXX) 469 ifeq ($(STATIC),1) 470 $(examples) $(tests) : LDLIBS += $(LIBCXX) 471 pkgconfig-libs-private += $(LIBCXX) 472 endif 473endif 474 475# File names *-weak.c contain weak symbol definitions, which must be listed last 476# when creating shared or static libraries. 477weak_last = $(filter-out %-weak.o,$(1)) $(filter %-weak.o,$(1)) 478 479libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) 480$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 481$(libceed.o): | info-backends 482$(libceed.so) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 483 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 484 485$(libceed.a) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 486 $(call quiet,AR) $(ARFLAGS) $@ $^ 487 488$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 489 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 490 491$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 492 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 493 494$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 495 $(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<) 496 497$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR 498 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 499 500$(OBJDIR)/% : tests/%.c | $$(@D)/.DIR 501 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 502 503$(OBJDIR)/% : tests/%.f90 | $$(@D)/.DIR 504 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 505 506$(OBJDIR)/% : examples/ceed/%.c | $$(@D)/.DIR 507 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 508 509$(OBJDIR)/% : examples/ceed/%.f | $$(@D)/.DIR 510 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 511 512$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 513 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 514 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 515 mv examples/mfem/$* $@ 516 517# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 518# this single Nek bps file to be built in parallel with other examples, 519# such as when calling `make prove-all -j2`. 520$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 521 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 522 mv examples/nek/build/bps $(OBJDIR)/bps 523 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 524 525$(OBJDIR)/petsc-% : examples/petsc/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 526 +$(MAKE) -C examples/petsc CEED_DIR=`pwd` \ 527 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 528 mv examples/petsc/$* $@ 529 530$(OBJDIR)/fluids-% : examples/fluids/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 531 +$(MAKE) -C examples/fluids CEED_DIR=`pwd` \ 532 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 533 mv examples/fluids/$* $@ 534 535$(OBJDIR)/solids-% : examples/solids/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 536 +$(MAKE) -C examples/solids CEED_DIR=`pwd` \ 537 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 538 mv examples/solids/$* $@ 539 540libceed_test.o = $(test_backends.c:%.c=$(OBJDIR)/%.o) 541$(libceed_test.so) : $(call weak_last,$(libceed.o) $(libceed_test.o)) | $$(@D)/.DIR 542 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 543 544$(libceed_test.a) : $(call weak_last,$(libceed.o) $(libceed_test.o)) | $$(@D)/.DIR 545 $(call quiet,AR) $(ARFLAGS) $@ $^ 546 547$(examples) : $(libceed) 548$(tests) : $(libceed_test) 549$(tests) : CEED_LIBS = -lceed_test 550$(tests) $(examples) : LDFLAGS += -Wl,-rpath,$(abspath $(LIBDIR)) -L$(LIBDIR) 551 552run-t% : BACKENDS += $(TEST_BACKENDS) 553run-% : $(OBJDIR)/% 554 @tests/tap.sh $(<:$(OBJDIR)/%=%) 555 556external_examples := \ 557 $(if $(MFEM_DIR),$(mfemexamples)) \ 558 $(if $(PETSC_DIR),$(petscexamples)) \ 559 $(if $(NEK5K_DIR),$(nekexamples)) \ 560 $(if $(PETSC_DIR),$(fluidsexamples)) \ 561 $(if $(PETSC_DIR),$(solidsexamples)) 562 563allexamples = $(examples) $(external_examples) 564 565# The test and prove targets can be controlled via pattern searches. The 566# default is to run tests and those examples that have no external dependencies. 567# Examples of finer grained control: 568# 569# make test search='petsc mfem' # PETSc and MFEM examples 570# make prove search='t3' # t3xx series tests 571# make junit search='ex petsc' # core ex* and PETSc tests 572search ?= t ex 573realsearch = $(search:%=%%) 574matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 575 576# Test core libCEED 577test : $(matched:$(OBJDIR)/%=run-%) 578 579# Run test target in parallel 580tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 581# CPU C tests only for backend % 582ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 583 584prove : BACKENDS += $(TEST_BACKENDS) 585prove : $(matched) 586 $(info Testing backends: $(BACKENDS)) 587 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 588# Run prove target in parallel 589prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 590 591prove-all : 592 +$(MAKE) prove realsearch=% 593 594junit-t% : BACKENDS += $(TEST_BACKENDS) 595junit-% : $(OBJDIR)/% 596 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 597 598junit : $(matched:$(OBJDIR)/%=junit-%) 599 600all: $(alltests) 601 602examples : $(allexamples) 603ceedexamples : $(examples) 604nekexamples : $(nekexamples) 605mfemexamples : $(mfemexamples) 606petscexamples : $(petscexamples) 607 608# Benchmarks 609allbenchmarks = petsc-bps 610bench_targets = $(addprefix bench-,$(allbenchmarks)) 611.PHONY: $(bench_targets) benchmarks 612$(bench_targets): bench-%: $(OBJDIR)/% 613 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS)" -r $(*).sh 614benchmarks: $(bench_targets) 615 616$(ceed.pc) : pkgconfig-prefix = $(abspath .) 617$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 618.INTERMEDIATE : $(OBJDIR)/ceed.pc 619%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 620 @$(SED) \ 621 -e "s:%prefix%:$(pkgconfig-prefix):" \ 622 -e "s:%libs_private%:$(pkgconfig-libs-private):" $< > $@ 623 624install : $(libceed) $(OBJDIR)/ceed.pc 625 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 626 "$(includedir)/ceed/" "$(libdir)" "$(pkgconfigdir)") 627 $(INSTALL_DATA) include/ceed/ceed.h "$(DESTDIR)$(includedir)/ceed/" 628 $(INSTALL_DATA) include/ceed/fortran.h "$(DESTDIR)$(includedir)/ceed/" 629 $(INSTALL_DATA) include/ceed/backend.h "$(DESTDIR)$(includedir)/ceed/" 630 $(INSTALL_DATA) include/ceed/cuda.h "$(DESTDIR)$(includedir)/ceed/" 631 $(INSTALL_DATA) include/ceed/hip.h "$(DESTDIR)$(includedir)/ceed/" 632 $(INSTALL_DATA) include/ceed/hash.h "$(DESTDIR)$(includedir)/ceed/" 633 $(INSTALL_DATA) include/ceed/khash.h "$(DESTDIR)$(includedir)/ceed/" 634 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 635 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 636 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 637 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 638 $(INSTALL_DATA) include/ceed-backend.h "$(DESTDIR)$(includedir)/" 639 $(INSTALL_DATA) include/ceed-hash.h "$(DESTDIR)$(includedir)/" 640 $(INSTALL_DATA) include/ceed-khash.h "$(DESTDIR)$(includedir)/" 641 642.PHONY : cln clean doxygen doc lib install all print test tst prove prv prove-all junit examples style style-c style-py tidy info info-backends 643 644cln clean : 645 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 646 $(MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 647 $(MAKE) -C python/tests clean 648 $(RM) benchmarks/*output.txt 649 650distclean : clean 651 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 652 653DOXYGEN ?= doxygen 654doxygen : 655 $(DOXYGEN) Doxyfile 656 657doc-html doc-latexpdf doc-epub doc-livehtml : doc-% : doxygen 658 make -C doc/sphinx $* 659 660doc : doc-html 661 662style-c : 663 @astyle --options=.astylerc \ 664 $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), \ 665 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 666 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 667 668AUTOPEP8 = autopep8 669style-py : AUTOPEP8_ARGS = --in-place --aggressive 670style-py : 671 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py python/tests/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 672 673style : style-c style-py 674 675CLANG_TIDY ?= clang-tidy 676 677%.c.tidy : %.c 678 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include -I$(HIP_DIR)/include 679 680%.cpp.tidy : %.cpp 681 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include -I$(HIP_DIR)/include 682 683tidy_c : $(libceed.c:%=%.tidy) 684tidy_cpp : $(libceed.cpp:%=%.tidy) 685 686tidy : tidy_c tidy_cpp 687 688print : 689 @echo $(VAR)=$($(VAR)) 690 691print-% : 692 $(info [ variable name]: $*) 693 $(info [ origin]: $(origin $*)) 694 $(info [ flavor]: $(flavor $*)) 695 $(info [ value]: $(value $*)) 696 $(info [expanded value]: $($*)) 697 $(info ) 698 @true 699 700# "make configure" detects any variables passed on the command line or 701# previously set in config.mk, caching them in config.mk as simple 702# (:=) variables. Variables set in config.mk or on the command line 703# take precedence over the defaults provided in the file. Typical 704# usage: 705# 706# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 707# make 708# make prove 709# 710# The values in the file can be updated by passing them on the command 711# line, e.g., 712# 713# make configure CC=/path/to/other/clang 714 715# All variables to consider for caching 716CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 717 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 718 AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \ 719 MAGMA_DIR OCCA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR HIP_ARCH 720 721# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 722# was set on the command line or in config.mk (where it will appear as 723# a simple variable). 724needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 725 726configure : 727 $(file > $(CONFIG)) 728 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 729 @echo "Configuration cached in $(CONFIG):" 730 @cat $(CONFIG) 731 732wheel : export MARCHFLAG = -march=generic 733wheel : export WHEEL_PLAT = manylinux2010_x86_64 734wheel : 735 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 736 -e MARCHFLAG -e WHEEL_PLAT \ 737 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 738 739.PHONY : configure wheel 740 741# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 742-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 743