1# Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2# All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3# 4# SPDX-License-Identifier: BSD-2-Clause 5# 6# This file is part of CEED: http://github.com/ceed 7 8CONFIG ?= config.mk 9-include $(CONFIG) 10COMMON ?= common.mk 11-include $(COMMON) 12 13ifeq (,$(filter-out undefined default,$(origin CC))) 14 CC = gcc 15endif 16ifeq (,$(filter-out undefined default,$(origin CXX))) 17 CXX = g++ 18endif 19ifeq (,$(filter-out undefined default,$(origin FC))) 20 FC = gfortran 21endif 22ifeq (,$(filter-out undefined default,$(origin LINK))) 23 LINK = $(CC) 24endif 25ifeq (,$(filter-out undefined default,$(origin AR))) 26 AR = ar 27endif 28ifeq (,$(filter-out undefined default,$(origin ARFLAGS))) 29 ARFLAGS = crD 30endif 31NVCC ?= $(CUDA_DIR)/bin/nvcc 32NVCC_CXX ?= $(CXX) 33HIPCC ?= $(HIP_DIR)/bin/hipcc 34SED ?= sed 35 36# ASAN must be left empty if you don't want to use it 37ASAN ?= 38 39# These are the values automatically detected here in the makefile. They are 40# augmented with LDFLAGS and LDLIBS from the environment/passed by command line, 41# if any. If the user sets CEED_LDFLAGS or CEED_LDLIBS, they are used *instead 42# of* what we populate here (thus that's advanced usage and not recommended). 43CEED_LDFLAGS ?= 44CEED_LDLIBS ?= 45 46UNDERSCORE ?= 1 47 48# Verbose mode, V or VERBOSE 49V ?= $(VERBOSE) 50 51# MFEM_DIR env variable should point to sibling directory 52ifneq ($(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 := $(subst oneAPI,icc,$(firstword $(filter gcc clang icc oneAPI XL,$(subst -, ,$(shell $(CC) --version))))) 94FC_VENDOR := $(if $(FC),$(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 -munsafe-fp-atomics 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 CEED_LDFLAGS += --coverage 146endif 147 148CFLAGS += $(if $(ASAN),$(AFLAGS)) 149FFLAGS += $(if $(ASAN),$(AFLAGS)) 150CEED_LDFLAGS += $(if $(ASAN),$(AFLAGS)) 151CPPFLAGS += -I./include 152CEED_LDLIBS = -lm 153OBJDIR := build 154for_install := $(filter install,$(MAKECMDGOALS)) 155LIBDIR := $(if $(for_install),$(OBJDIR),lib) 156 157 158# Installation variables 159prefix ?= /usr/local 160bindir = $(prefix)/bin 161libdir = $(prefix)/lib 162includedir = $(prefix)/include 163pkgconfigdir = $(libdir)/pkgconfig 164INSTALL = install 165INSTALL_PROGRAM = $(INSTALL) 166INSTALL_DATA = $(INSTALL) -m644 167 168# Get number of processors of the machine 169NPROCS := $(shell getconf _NPROCESSORS_ONLN) 170# prepare make options to run in parallel 171MFLAGS := -j $(NPROCS) --warn-undefined-variables \ 172 --no-print-directory --no-keep-going 173 174PYTHON ?= python3 175PROVE ?= prove 176PROVE_OPTS ?= -j $(NPROCS) 177DARWIN := $(filter Darwin,$(shell uname -s)) 178SO_EXT := $(if $(DARWIN),dylib,so) 179 180ceed.pc := $(LIBDIR)/pkgconfig/ceed.pc 181libceed.so := $(LIBDIR)/libceed.$(SO_EXT) 182libceed.a := $(LIBDIR)/libceed.a 183libceed := $(if $(STATIC),$(libceed.a),$(libceed.so)) 184CEED_LIBS = -lceed 185libceed.c := $(filter-out interface/ceed-cuda.c interface/ceed-hip.c interface/ceed-jit-source-root-$(if $(for_install),default,install).c, $(wildcard interface/ceed*.c backends/*.c gallery/*.c)) 186gallery.c := $(wildcard gallery/*/ceed*.c) 187libceed.c += $(gallery.c) 188libceeds = $(libceed) 189BACKENDS_BUILTIN := /cpu/self/ref/serial /cpu/self/ref/blocked /cpu/self/opt/serial /cpu/self/opt/blocked 190BACKENDS_MAKE := $(BACKENDS_BUILTIN) 191TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub 192 193# Tests 194tests.c := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.c)) 195tests.f := $(if $(FC),$(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 := $(if $(FC),$(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, memcheck, opt, avx, occa, magma] 220ref.c := $(sort $(wildcard backends/ref/*.c)) 221blocked.c := $(sort $(wildcard backends/blocked/*.c)) 222ceedmemcheck.c := $(sort $(wildcard backends/memcheck/*.c)) 223opt.c := $(sort $(wildcard backends/opt/*.c)) 224avx.c := $(sort $(wildcard backends/avx/*.c)) 225xsmm.c := $(sort $(wildcard backends/xsmm/*.c)) 226cuda.c := $(sort $(wildcard backends/cuda/*.c)) 227cuda.cpp := $(sort $(wildcard backends/cuda/*.cpp)) 228cuda-ref.c := $(sort $(wildcard backends/cuda-ref/*.c)) 229cuda-ref.cpp := $(sort $(wildcard backends/cuda-ref/*.cpp)) 230cuda-ref.cu := $(sort $(wildcard backends/cuda-ref/kernels/*.cu)) 231cuda-shared.c := $(sort $(wildcard backends/cuda-shared/*.c)) 232cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/kernels/*.cu)) 233cuda-gen.c := $(sort $(wildcard backends/cuda-gen/*.c)) 234cuda-gen.cpp := $(sort $(wildcard backends/cuda-gen/*.cpp)) 235cuda-gen.cu := $(sort $(wildcard backends/cuda-gen/kernels/*.cu)) 236occa.cpp := $(sort $(shell find backends/occa -type f -name *.cpp)) 237magma.c := $(sort $(wildcard backends/magma/*.c)) 238magma.cu := $(sort $(wildcard backends/magma/kernels/cuda/*.cu)) 239magma.hip := $(sort $(wildcard backends/magma/kernels/hip/*.hip.cpp)) 240hip.c := $(sort $(wildcard backends/hip/*.c)) 241hip.cpp := $(sort $(wildcard backends/hip/*.cpp)) 242hip-ref.c := $(sort $(wildcard backends/hip-ref/*.c)) 243hip-ref.cpp := $(sort $(wildcard backends/hip-ref/*.cpp)) 244hip-ref.hip := $(sort $(wildcard backends/hip-ref/kernels/*.hip.cpp)) 245hip-shared.c := $(sort $(wildcard backends/hip-shared/*.c)) 246hip-gen.c := $(sort $(wildcard backends/hip-gen/*.c)) 247hip-gen.cpp := $(sort $(wildcard backends/hip-gen/*.cpp)) 248 249# Quiet, color output 250quiet ?= $($(1)) 251 252# Cancel built-in and old-fashioned implicit rules which we don't use 253.SUFFIXES: 254 255.SECONDEXPANSION: # to expand $$(@D)/.DIR 256 257%/.DIR : 258 @mkdir -p $(@D) 259 @touch $@ 260 261.PRECIOUS: %/.DIR 262 263lib: $(libceed) $(ceed.pc) 264# run 'lib' target in parallel 265par:;@$(MAKE) $(MFLAGS) V=$(V) lib 266backend_status = $(if $(filter $1,$(BACKENDS_MAKE)), [backends: $1], [not found]) 267info: 268 $(info ------------------------------------) 269 $(info CC = $(CC)) 270 $(info CXX = $(CXX)) 271 $(info FC = $(FC)) 272 $(info CPPFLAGS = $(CPPFLAGS)) 273 $(info CFLAGS = $(value CFLAGS)) 274 $(info CXXFLAGS = $(value CXXFLAGS)) 275 $(info FFLAGS = $(value FFLAGS)) 276 $(info NVCCFLAGS = $(value NVCCFLAGS)) 277 $(info HIPCCFLAGS = $(value HIPCCFLAGS)) 278 $(info CEED_LDFLAGS = $(value CEED_LDFLAGS)) 279 $(info CEED_LDLIBS = $(value CEED_LDLIBS)) 280 $(info AR = $(AR)) 281 $(info ARFLAGS = $(ARFLAGS)) 282 $(info OPT = $(OPT)) 283 $(info AFLAGS = $(AFLAGS)) 284 $(info ASAN = $(or $(ASAN),(empty))) 285 $(info VERBOSE = $(or $(V),(empty)) [verbose=$(if $(V),on,off)]) 286 $(info ------------------------------------) 287 $(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,$(MEMCHK_BACKENDS))) 288 $(info AVX_STATUS = $(AVX_STATUS)$(call backend_status,$(AVX_BACKENDS))) 289 $(info XSMM_DIR = $(XSMM_DIR)$(call backend_status,$(XSMM_BACKENDS))) 290 $(info OCCA_DIR = $(OCCA_DIR)$(call backend_status,$(OCCA_BACKENDS))) 291 $(info MAGMA_DIR = $(MAGMA_DIR)$(call backend_status,$(MAGMA_BACKENDS))) 292 $(info CUDA_DIR = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS))) 293 $(info HIP_DIR = $(HIP_DIR)$(call backend_status,$(HIP_BACKENDS))) 294 $(info ------------------------------------) 295 $(info MFEM_DIR = $(MFEM_DIR)) 296 $(info NEK5K_DIR = $(NEK5K_DIR)) 297 $(info PETSC_DIR = $(PETSC_DIR)) 298 $(info ------------------------------------) 299 $(info prefix = $(prefix)) 300 $(info includedir = $(value includedir)) 301 $(info libdir = $(value libdir)) 302 $(info pkgconfigdir = $(value pkgconfigdir)) 303 $(info ------------------------------------) 304 @true 305info-backends: 306 $(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS))) 307 @true 308info-backends-all: 309 $(info make: 'lib' with backends: $(filter-out $(TEST_BACKENDS),$(BACKENDS))) 310 @true 311 312$(libceed.so) : CEED_LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed.so))) 313 314# Standard Backends 315libceed.c += $(ref.c) 316libceed.c += $(blocked.c) 317libceed.c += $(opt.c) 318 319# Memcheck Backend 320MEMCHK_STATUS = Disabled 321MEMCHK := $(shell echo "$(HASH)include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1) 322MEMCHK_BACKENDS = /cpu/self/memcheck/serial /cpu/self/memcheck/blocked 323ifeq ($(MEMCHK),1) 324 MEMCHK_STATUS = Enabled 325 libceed.c += $(ceedmemcheck.c) 326 BACKENDS_MAKE += $(MEMCHK_BACKENDS) 327endif 328 329# AVX Backed 330AVX_STATUS = Disabled 331AVX_FLAG := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx) 332AVX := $(filter $(AVX_FLAG),$(shell $(CC) $(CFLAGS) -v -E -x c /dev/null 2>&1)) 333AVX_BACKENDS = /cpu/self/avx/serial /cpu/self/avx/blocked 334ifneq ($(AVX),) 335 AVX_STATUS = Enabled 336 libceed.c += $(avx.c) 337 BACKENDS_MAKE += $(AVX_BACKENDS) 338endif 339 340# Collect list of libraries and paths for use in linking and pkg-config 341PKG_LIBS = 342# Stubs that will not be RPATH'd 343PKG_STUBS_LIBS = 344 345# libXSMM Backends 346XSMM_BACKENDS = /cpu/self/xsmm/serial /cpu/self/xsmm/blocked 347ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),) 348 PKG_LIBS += -L$(abspath $(XSMM_DIR))/lib -lxsmm -ldl 349 MKL ?= 350 ifeq (,$(MKL)$(MKLROOT)) 351 BLAS_LIB = -lblas 352 else 353 ifneq ($(MKLROOT),) 354 # Some installs put everything inside an intel64 subdirectory, others not 355 MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*))) 356 MKL_LINK = -L$(MKL_LIBDIR) 357 PKG_LIB_DIRS += $(MKL_LIBDIR) 358 endif 359 BLAS_LIB = $(MKL_LINK) -Wl,--push-state,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl -Wl,--pop-state 360 endif 361 PKG_LIBS += $(BLAS_LIB) 362 libceed.c += $(xsmm.c) 363 $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include 364 BACKENDS_MAKE += $(XSMM_BACKENDS) 365endif 366 367# OCCA Backends 368OCCA_BACKENDS = /cpu/self/occa 369ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),) 370 OCCA_MODES := $(shell $(OCCA_DIR)/bin/occa modes) 371 OCCA_BACKENDS += $(if $(filter OpenMP,$(OCCA_MODES)),/cpu/openmp/occa) 372# OCCA_BACKENDS += $(if $(filter OpenCL,$(OCCA_MODES)),/gpu/opencl/occa) 373 OCCA_BACKENDS += $(if $(filter HIP,$(OCCA_MODES)),/gpu/hip/occa) 374 OCCA_BACKENDS += $(if $(filter CUDA,$(OCCA_MODES)),/gpu/cuda/occa) 375 376 $(libceeds) : CPPFLAGS += -I$(OCCA_DIR)/include 377 PKG_LIBS += -L$(abspath $(OCCA_DIR))/lib -locca 378 LIBCEED_CONTAINS_CXX = 1 379 libceed.cpp += $(occa.cpp) 380 BACKENDS_MAKE += $(OCCA_BACKENDS) 381endif 382 383# CUDA Backends 384CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(CUDA_DIR)/$d/libcudart.${SO_EXT})) 385CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR)))) 386CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs 387CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/shared /gpu/cuda/gen 388ifneq ($(CUDA_LIB_DIR),) 389 $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include 390 PKG_LIBS += -L$(abspath $(CUDA_LIB_DIR)) -lcudart -lnvrtc -lcuda -lcublas 391 PKG_STUBS_LIBS += -L$(CUDA_LIB_DIR_STUBS) 392 LIBCEED_CONTAINS_CXX = 1 393 libceed.c += interface/ceed-cuda.c 394 libceed.c += $(cuda.c) $(cuda-ref.c) $(cuda-shared.c) $(cuda-gen.c) 395 libceed.cpp += $(cuda.cpp) $(cuda-ref.cpp) $(cuda-gen.cpp) 396 libceed.cu += $(cuda-ref.cu) $(cuda-shared.cu) $(cuda-gen.cu) 397 BACKENDS_MAKE += $(CUDA_BACKENDS) 398endif 399 400# HIP Backends 401HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libamdhip64.${SO_EXT})) 402HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 403HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen 404ifneq ($(HIP_LIB_DIR),) 405 $(libceeds) : HIPCCFLAGS += -I./include 406 ifneq ($(CXX), $(HIPCC)) 407 CPPFLAGS += $(subst =,,$(shell $(HIP_DIR)/bin/hipconfig -C)) 408 endif 409 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include 410 PKG_LIBS += -L$(abspath $(HIP_LIB_DIR)) -lamdhip64 -lhipblas 411 LIBCEED_CONTAINS_CXX = 1 412 libceed.c += interface/ceed-hip.c 413 libceed.c += $(hip.c) $(hip-ref.c) $(hip-shared.c) $(hip-gen.c) 414 libceed.cpp += $(hip.cpp) $(hip-ref.cpp) $(hip-gen.cpp) 415 libceed.hip += $(hip-ref.hip) 416 BACKENDS_MAKE += $(HIP_BACKENDS) 417endif 418 419# MAGMA Backend 420ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 421 MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas") 422 ifeq ($(MAGMA_ARCH), 0) #CUDA MAGMA 423 ifneq ($(CUDA_LIB_DIR),) 424 cuda_link = -Wl,-rpath,$(CUDA_LIB_DIR) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart 425 omp_link = -fopenmp 426 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link) 427 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 428 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 429 PKG_LIBS += $(magma_link) 430 libceed.c += $(magma.c) 431 libceed.cu += $(magma.cu) 432 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 433 $(magma.cu:%.cu=$(OBJDIR)/%.o) : CPPFLAGS += --compiler-options=-fPIC -DADD_ -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(CUDA_DIR)/include 434 MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det 435 endif 436 else # HIP MAGMA 437 ifneq ($(HIP_LIB_DIR),) 438 hip_link = -Wl,-rpath,$(HIP_LIB_DIR) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64 439 omp_link = -fopenmp 440 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link) 441 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 442 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 443 PKG_LIBS += $(magma_link) 444 libceed.c += $(magma.c) 445 libceed.hip += $(magma.hip) 446 ifneq ($(CXX), $(HIPCC)) 447 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 448 else 449 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 450 endif 451 $(magma.hip:%.hip.cpp=$(OBJDIR)/%.o) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 452 MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det 453 endif 454 endif 455 LIBCEED_CONTAINS_CXX = 1 456 BACKENDS_MAKE += $(MAGMA_BACKENDS) 457endif 458 459BACKENDS ?= $(BACKENDS_MAKE) 460export BACKENDS 461 462_pkg_ldflags = $(filter -L%,$(PKG_LIBS)) 463_pkg_ldlibs = $(filter-out -L%,$(PKG_LIBS)) 464$(libceeds) : CEED_LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 465$(libceeds) : CEED_LDLIBS += $(_pkg_ldlibs) 466ifeq ($(STATIC),1) 467$(examples) $(tests) : CEED_LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 468$(examples) $(tests) : CEED_LDLIBS += $(_pkg_ldlibs) 469endif 470 471pkgconfig-libs-private = $(PKG_LIBS) 472ifeq ($(LIBCEED_CONTAINS_CXX),1) 473 $(libceeds) : LINK = $(CXX) 474 ifeq ($(STATIC),1) 475 $(examples) $(tests) : CEED_LDLIBS += $(LIBCXX) 476 pkgconfig-libs-private += $(LIBCXX) 477 endif 478endif 479 480# File names *-weak.c contain weak symbol definitions, which must be listed last 481# when creating shared or static libraries. 482weak_last = $(filter-out %-weak.o,$(1)) $(filter %-weak.o,$(1)) 483 484libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) 485$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 486$(libceed.o): | info-backends 487$(libceed.so) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 488 $(call quiet,LINK) $(LDFLAGS) $(CEED_LDFLAGS) -shared -o $@ $^ $(CEED_LDLIBS) $(LDLIBS) 489 490$(libceed.a) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 491 $(call quiet,AR) $(ARFLAGS) $@ $^ 492 493$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 494 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 495 496$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 497 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 498 499$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 500 $(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<) 501 502$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR 503 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 504 505$(OBJDIR)/% : tests/%.c | $$(@D)/.DIR 506 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 507 508$(OBJDIR)/% : tests/%.f90 | $$(@D)/.DIR 509 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 510 511$(OBJDIR)/% : examples/ceed/%.c | $$(@D)/.DIR 512 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 513 514$(OBJDIR)/% : examples/ceed/%.f | $$(@D)/.DIR 515 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 516 517$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 518 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 519 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 520 cp examples/mfem/$* $@ 521 522# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 523# this single Nek bps file to be built in parallel with other examples, 524# such as when calling `make prove-all -j2`. 525$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 526 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 527 mv examples/nek/build/bps $(OBJDIR)/bps 528 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 529 530# Several executables have common utilities, but we can't build the utilities 531# from separate submake invocations because they'll compete with each 532# other/corrupt output. So we put it in this utility library, but we don't want 533# to manually list source dependencies up at this level, so we'll just always 534# call recursive make to check that this utility is up to date. 535examples/petsc/libutils.a.PHONY: $(libceed) $(ceed.pc) 536 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` AR=$(AR) ARFLAGS=$(ARFLAGS) \ 537 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $(basename $(@F)) 538 539$(OBJDIR)/petsc-% : examples/petsc/%.c examples/petsc/libutils.a.PHONY $(libceed) $(ceed.pc) | $$(@D)/.DIR 540 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` \ 541 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 542 cp examples/petsc/$* $@ 543 544$(OBJDIR)/fluids-% : examples/fluids/%.c examples/fluids/src/*.c examples/fluids/*.h examples/fluids/problems/*.c examples/fluids/qfunctions/*.h $(libceed) $(ceed.pc) | $$(@D)/.DIR 545 +$(call quiet,MAKE) -C examples/fluids CEED_DIR=`pwd` \ 546 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 547 cp examples/fluids/$* $@ 548 549$(OBJDIR)/solids-% : examples/solids/%.c examples/solids/%.h \ 550 examples/solids/problems/*.c examples/solids/src/*.c \ 551 examples/solids/include/*.h examples/solids/problems/*.h examples/solids/qfunctions/*.h \ 552 $(libceed) $(ceed.pc) | $$(@D)/.DIR 553 +$(call quiet,MAKE) -C examples/solids CEED_DIR=`pwd` \ 554 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 555 cp examples/solids/$* $@ 556 557$(examples) : $(libceed) 558$(tests) : $(libceed) 559$(tests) $(examples) : override LDFLAGS += -Wl,-rpath,$(abspath $(LIBDIR)) -L$(LIBDIR) 560 561run-% : $(OBJDIR)/% 562 @tests/tap.sh $(<:$(OBJDIR)/%=%) 563 564external_examples := \ 565 $(if $(MFEM_DIR),$(mfemexamples)) \ 566 $(if $(PETSC_DIR),$(petscexamples)) \ 567 $(if $(NEK5K_DIR),$(nekexamples)) \ 568 $(if $(PETSC_DIR),$(fluidsexamples)) \ 569 $(if $(PETSC_DIR),$(solidsexamples)) 570 571allexamples = $(examples) $(external_examples) 572 573# The test and prove targets can be controlled via pattern searches. The 574# default is to run tests and those examples that have no external dependencies. 575# Examples of finer grained control: 576# 577# make test search='petsc mfem' # PETSc and MFEM examples 578# make prove search='t3' # t3xx series tests 579# make junit search='ex petsc' # core ex* and PETSc tests 580search ?= t ex 581realsearch = $(search:%=%%) 582matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 583 584# Test core libCEED 585test : $(matched:$(OBJDIR)/%=run-%) 586 587# Run test target in parallel 588tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 589# CPU C tests only for backend % 590ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 591 592prove : $(matched) 593 $(info Testing backends: $(BACKENDS)) 594 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 595# Run prove target in parallel 596prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 597 598prove-all : 599 +$(MAKE) prove realsearch=% 600 601junit-% : $(OBJDIR)/% 602 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 603 604junit : $(matched:$(OBJDIR)/%=junit-%) 605 606all: $(alltests) 607 608examples : $(allexamples) 609ceedexamples : $(examples) 610nekexamples : $(nekexamples) 611mfemexamples : $(mfemexamples) 612petscexamples : $(petscexamples) 613 614# Benchmarks 615allbenchmarks = petsc-bps 616bench_targets = $(addprefix bench-,$(allbenchmarks)) 617.PHONY: $(bench_targets) benchmarks 618$(bench_targets): bench-%: $(OBJDIR)/% 619 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS_MAKE)" -r $(*).sh 620benchmarks: $(bench_targets) 621 622$(ceed.pc) : pkgconfig-prefix = $(abspath .) 623$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 624.INTERMEDIATE : $(OBJDIR)/ceed.pc 625%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 626 @$(SED) \ 627 -e "s:%prefix%:$(pkgconfig-prefix):" \ 628 -e "s:%libs_private%:$(pkgconfig-libs-private):" $< > $@ 629 630$(OBJDIR)/interface/ceed-jit-source-root-default.o : CPPFLAGS += -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath ./include)/\"" 631$(OBJDIR)/interface/ceed-jit-source-root-install.o : CPPFLAGS += -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath $(includedir))/\"" 632 633install : $(libceed) $(OBJDIR)/ceed.pc 634 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 635 "$(includedir)/ceed/" "$(includedir)/ceed/jit-source/"\ 636 "$(includedir)/ceed/jit-source/cuda/" "$(includedir)/ceed/jit-source/hip/"\ 637 "$(includedir)/ceed/jit-source/gallery/" "$(libdir)" "$(pkgconfigdir)") 638 $(INSTALL_DATA) include/ceed/ceed.h "$(DESTDIR)$(includedir)/ceed/" 639 $(INSTALL_DATA) include/ceed/types.h "$(DESTDIR)$(includedir)/ceed/" 640 $(INSTALL_DATA) include/ceed/ceed-f32.h "$(DESTDIR)$(includedir)/ceed/" 641 $(INSTALL_DATA) include/ceed/ceed-f64.h "$(DESTDIR)$(includedir)/ceed/" 642 $(INSTALL_DATA) include/ceed/fortran.h "$(DESTDIR)$(includedir)/ceed/" 643 $(INSTALL_DATA) include/ceed/backend.h "$(DESTDIR)$(includedir)/ceed/" 644 $(INSTALL_DATA) include/ceed/cuda.h "$(DESTDIR)$(includedir)/ceed/" 645 $(INSTALL_DATA) include/ceed/hip.h "$(DESTDIR)$(includedir)/ceed/" 646 $(INSTALL_DATA) include/ceed/hash.h "$(DESTDIR)$(includedir)/ceed/" 647 $(INSTALL_DATA) include/ceed/khash.h "$(DESTDIR)$(includedir)/ceed/" 648 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 649 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 650 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 651 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 652 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/cuda/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/cuda/" 653 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/hip/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/hip/" 654 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/gallery/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/gallery/" 655 656.PHONY : all cln clean doxygen doc lib install par print test tst prove prv prove-all junit examples style style-c style-py tidy iwyu info info-backends info-backends-all 657 658cln clean : 659 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 660 $(call quiet,MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 661 $(call quiet,MAKE) -C python/tests clean 662 $(RM) benchmarks/*output.txt 663 664distclean : clean 665 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 666 667DOXYGEN ?= doxygen 668doxygen : 669 $(DOXYGEN) -q Doxyfile 670 671doc-html doc-latexpdf doc-epub doc-livehtml : doc-% : doxygen 672 make -C doc/sphinx $* 673 674doc : doc-html 675 676style-c : 677 @astyle --options=.astylerc \ 678 $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), \ 679 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 680 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 681 682AUTOPEP8 = autopep8 683style-py : AUTOPEP8_ARGS = --in-place --aggressive 684style-py : 685 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py python/tests/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 686 687style : style-c style-py 688 689CLANG_TIDY ?= clang-tidy 690 691%.c.tidy : %.c 692 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include -I$(HIP_DIR)/include -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath ./include)/\"" 693 694%.cpp.tidy : %.cpp 695 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include -I$(HIP_DIR)/include 696 697tidy_c : $(libceed.c:%=%.tidy) 698tidy_cpp : $(libceed.cpp:%=%.tidy) 699 700tidy : tidy_c tidy_cpp 701 702ifneq ($(wildcard ../iwyu/*),) 703 IWYU_DIR ?= ../iwyu 704 IWYU_CC ?= $(IWYU_DIR)/build/bin/include-what-you-use 705endif 706 707iwyu : CC=$(IWYU_CC) 708iwyu : lib 709 710print : 711 @echo $(VAR)=$($(VAR)) 712 713print-% : 714 $(info [ variable name]: $*) 715 $(info [ origin]: $(origin $*)) 716 $(info [ flavor]: $(flavor $*)) 717 $(info [ value]: $(value $*)) 718 $(info [expanded value]: $($*)) 719 $(info ) 720 @true 721 722# "make configure" detects any variables passed on the command line or 723# previously set in config.mk, caching them in config.mk as simple 724# (:=) variables. Variables set in config.mk or on the command line 725# take precedence over the defaults provided in the file. Typical 726# usage: 727# 728# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 729# make 730# make prove 731# 732# The values in the file can be updated by passing them on the command 733# line, e.g., 734# 735# make configure CC=/path/to/other/clang 736 737# All variables to consider for caching 738CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 739 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 740 AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \ 741 MAGMA_DIR OCCA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR HIP_ARCH 742 743# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 744# was set on the command line or in config.mk (where it will appear as 745# a simple variable). 746needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 747 748configure : 749 $(file > $(CONFIG)) 750 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 751 @echo "Configuration cached in $(CONFIG):" 752 @cat $(CONFIG) 753 754wheel : export MARCHFLAG = -march=generic 755wheel : export WHEEL_PLAT = manylinux2010_x86_64 756wheel : 757 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 758 -e MARCHFLAG -e WHEEL_PLAT \ 759 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 760 761.PHONY : configure wheel 762 763# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 764-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 765