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 := $(subst icc_orig,icc,$(subst oneAPI,icc,$(firstword $(filter gcc clang icc icc_orig oneAPI XL,$(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.cpp := $(sort $(wildcard backends/magma/*.cpp)) 239magma.cu := $(sort $(wildcard backends/magma/kernels/cuda/*.cu)) 240magma.hip := $(sort $(wildcard backends/magma/kernels/hip/*.hip.cpp)) 241hip.c := $(sort $(wildcard backends/hip/*.c)) 242hip.cpp := $(sort $(wildcard backends/hip/*.cpp)) 243hip-ref.c := $(sort $(wildcard backends/hip-ref/*.c)) 244hip-ref.cpp := $(sort $(wildcard backends/hip-ref/*.cpp)) 245hip-ref.hip := $(sort $(wildcard backends/hip-ref/kernels/*.hip.cpp)) 246hip-shared.c := $(sort $(wildcard backends/hip-shared/*.c)) 247hip-gen.c := $(sort $(wildcard backends/hip-gen/*.c)) 248hip-gen.cpp := $(sort $(wildcard backends/hip-gen/*.cpp)) 249 250# Quiet, color output 251quiet ?= $($(1)) 252 253# Cancel built-in and old-fashioned implicit rules which we don't use 254.SUFFIXES: 255 256.SECONDEXPANSION: # to expand $$(@D)/.DIR 257 258%/.DIR : 259 @mkdir -p $(@D) 260 @touch $@ 261 262.PRECIOUS: %/.DIR 263 264lib: $(libceed) $(ceed.pc) 265# run 'lib' target in parallel 266par:;@$(MAKE) $(MFLAGS) V=$(V) lib 267backend_status = $(if $(filter $1,$(BACKENDS_MAKE)), [backends: $1], [not found]) 268info: 269 $(info ------------------------------------) 270 $(info CC = $(CC)) 271 $(info CXX = $(CXX)) 272 $(info FC = $(FC)) 273 $(info CPPFLAGS = $(CPPFLAGS)) 274 $(info CFLAGS = $(value CFLAGS)) 275 $(info CXXFLAGS = $(value CXXFLAGS)) 276 $(info FFLAGS = $(value FFLAGS)) 277 $(info NVCCFLAGS = $(value NVCCFLAGS)) 278 $(info HIPCCFLAGS = $(value HIPCCFLAGS)) 279 $(info CEED_LDFLAGS = $(value CEED_LDFLAGS)) 280 $(info CEED_LDLIBS = $(value CEED_LDLIBS)) 281 $(info AR = $(AR)) 282 $(info ARFLAGS = $(ARFLAGS)) 283 $(info OPT = $(OPT)) 284 $(info AFLAGS = $(AFLAGS)) 285 $(info ASAN = $(or $(ASAN),(empty))) 286 $(info VERBOSE = $(or $(V),(empty)) [verbose=$(if $(V),on,off)]) 287 $(info ------------------------------------) 288 $(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,$(MEMCHK_BACKENDS))) 289 $(info AVX_STATUS = $(AVX_STATUS)$(call backend_status,$(AVX_BACKENDS))) 290 $(info XSMM_DIR = $(XSMM_DIR)$(call backend_status,$(XSMM_BACKENDS))) 291 $(info OCCA_DIR = $(OCCA_DIR)$(call backend_status,$(OCCA_BACKENDS))) 292 $(info MAGMA_DIR = $(MAGMA_DIR)$(call backend_status,$(MAGMA_BACKENDS))) 293 $(info CUDA_DIR = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS))) 294 $(info HIP_DIR = $(HIP_DIR)$(call backend_status,$(HIP_BACKENDS))) 295 $(info ------------------------------------) 296 $(info MFEM_DIR = $(MFEM_DIR)) 297 $(info NEK5K_DIR = $(NEK5K_DIR)) 298 $(info PETSC_DIR = $(PETSC_DIR)) 299 $(info ------------------------------------) 300 $(info prefix = $(prefix)) 301 $(info includedir = $(value includedir)) 302 $(info libdir = $(value libdir)) 303 $(info pkgconfigdir = $(value pkgconfigdir)) 304 $(info ------------------------------------) 305 @true 306info-backends: 307 $(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS))) 308 @true 309info-backends-all: 310 $(info make: 'lib' with backends: $(filter-out $(TEST_BACKENDS),$(BACKENDS))) 311 @true 312 313$(libceed.so) : CEED_LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed.so))) 314 315# Standard Backends 316libceed.c += $(ref.c) 317libceed.c += $(blocked.c) 318libceed.c += $(opt.c) 319 320# Memcheck Backend 321MEMCHK_STATUS = Disabled 322MEMCHK := $(shell echo "$(HASH)include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1) 323MEMCHK_BACKENDS = /cpu/self/memcheck/serial /cpu/self/memcheck/blocked 324ifeq ($(MEMCHK),1) 325 MEMCHK_STATUS = Enabled 326 libceed.c += $(ceedmemcheck.c) 327 BACKENDS_MAKE += $(MEMCHK_BACKENDS) 328endif 329 330# AVX Backed 331AVX_STATUS = Disabled 332AVX_FLAG := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx) 333AVX := $(filter $(AVX_FLAG),$(shell $(CC) $(CFLAGS) -v -E -x c /dev/null 2>&1)) 334AVX_BACKENDS = /cpu/self/avx/serial /cpu/self/avx/blocked 335ifneq ($(AVX),) 336 AVX_STATUS = Enabled 337 libceed.c += $(avx.c) 338 BACKENDS_MAKE += $(AVX_BACKENDS) 339endif 340 341# Collect list of libraries and paths for use in linking and pkg-config 342PKG_LIBS = 343# Stubs that will not be RPATH'd 344PKG_STUBS_LIBS = 345 346# libXSMM Backends 347XSMM_BACKENDS = /cpu/self/xsmm/serial /cpu/self/xsmm/blocked 348ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),) 349 PKG_LIBS += -L$(abspath $(XSMM_DIR))/lib -lxsmm -ldl 350 MKL ?= 351 ifeq (,$(MKL)$(MKLROOT)) 352 BLAS_LIB = -lblas 353 else 354 ifneq ($(MKLROOT),) 355 # Some installs put everything inside an intel64 subdirectory, others not 356 MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*))) 357 MKL_LINK = -L$(MKL_LIBDIR) 358 PKG_LIB_DIRS += $(MKL_LIBDIR) 359 endif 360 BLAS_LIB = $(MKL_LINK) -Wl,--push-state,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl -Wl,--pop-state 361 endif 362 PKG_LIBS += $(BLAS_LIB) 363 libceed.c += $(xsmm.c) 364 $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include 365 BACKENDS_MAKE += $(XSMM_BACKENDS) 366endif 367 368# OCCA Backends 369OCCA_BACKENDS = /cpu/self/occa 370ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),) 371 OCCA_MODES := $(shell $(OCCA_DIR)/bin/occa modes) 372 OCCA_BACKENDS += $(if $(filter OpenMP,$(OCCA_MODES)),/cpu/openmp/occa) 373# OCCA_BACKENDS += $(if $(filter OpenCL,$(OCCA_MODES)),/gpu/opencl/occa) 374 OCCA_BACKENDS += $(if $(filter HIP,$(OCCA_MODES)),/gpu/hip/occa) 375 OCCA_BACKENDS += $(if $(filter CUDA,$(OCCA_MODES)),/gpu/cuda/occa) 376 377 $(libceeds) : CPPFLAGS += -I$(OCCA_DIR)/include 378 PKG_LIBS += -L$(abspath $(OCCA_DIR))/lib -locca 379 LIBCEED_CONTAINS_CXX = 1 380 libceed.cpp += $(occa.cpp) 381 BACKENDS_MAKE += $(OCCA_BACKENDS) 382endif 383 384# CUDA Backends 385CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(CUDA_DIR)/$d/libcudart.${SO_EXT})) 386CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR)))) 387CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs 388CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/shared /gpu/cuda/gen 389ifneq ($(CUDA_LIB_DIR),) 390 $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include 391 PKG_LIBS += -L$(abspath $(CUDA_LIB_DIR)) -lcudart -lnvrtc -lcuda -lcublas 392 PKG_STUBS_LIBS += -L$(CUDA_LIB_DIR_STUBS) 393 LIBCEED_CONTAINS_CXX = 1 394 libceed.c += interface/ceed-cuda.c 395 libceed.c += $(cuda.c) $(cuda-ref.c) $(cuda-shared.c) $(cuda-gen.c) 396 libceed.cpp += $(cuda.cpp) $(cuda-ref.cpp) $(cuda-gen.cpp) 397 libceed.cu += $(cuda-ref.cu) $(cuda-shared.cu) $(cuda-gen.cu) 398 BACKENDS_MAKE += $(CUDA_BACKENDS) 399endif 400 401# HIP Backends 402HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libamdhip64.${SO_EXT})) 403HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 404HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen 405ifneq ($(HIP_LIB_DIR),) 406 $(libceeds) : HIPCCFLAGS += -I./include 407 ifneq ($(CXX), $(HIPCC)) 408 CPPFLAGS += $(subst =,,$(shell $(HIP_DIR)/bin/hipconfig -C)) 409 endif 410 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include 411 PKG_LIBS += -L$(abspath $(HIP_LIB_DIR)) -lamdhip64 -lhipblas 412 LIBCEED_CONTAINS_CXX = 1 413 libceed.c += interface/ceed-hip.c 414 libceed.c += $(hip.c) $(hip-ref.c) $(hip-shared.c) $(hip-gen.c) 415 libceed.cpp += $(hip.cpp) $(hip-ref.cpp) $(hip-gen.cpp) 416 libceed.hip += $(hip-ref.hip) 417 BACKENDS_MAKE += $(HIP_BACKENDS) 418endif 419 420# MAGMA Backend 421ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 422 MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas") 423 ifeq ($(MAGMA_ARCH), 0) #CUDA MAGMA 424 ifneq ($(CUDA_LIB_DIR),) 425 cuda_link = -Wl,-rpath,$(CUDA_LIB_DIR) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart 426 omp_link = -fopenmp 427 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link) 428 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 429 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 430 PKG_LIBS += $(magma_link) 431 libceed.c += $(magma.c) 432 libceed.cpp += $(magma.cpp) 433 libceed.cu += $(magma.cu) 434 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 435 $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 436 $(magma.cu:%.cu=$(OBJDIR)/%.o) : CPPFLAGS += --compiler-options=-fPIC -DADD_ -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(CUDA_DIR)/include 437 MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det 438 endif 439 else # HIP MAGMA 440 ifneq ($(HIP_LIB_DIR),) 441 hip_link = -Wl,-rpath,$(HIP_LIB_DIR) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64 442 omp_link = -fopenmp 443 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link) 444 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 445 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 446 PKG_LIBS += $(magma_link) 447 libceed.c += $(magma.c) 448 libceed.cpp += $(magma.cpp) 449 libceed.hip += $(magma.hip) 450 ifneq ($(CXX), $(HIPCC)) 451 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 452 $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 453 else 454 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 455 $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 456 endif 457 $(magma.hip:%.hip.cpp=$(OBJDIR)/%.o) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 458 MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det 459 endif 460 endif 461 LIBCEED_CONTAINS_CXX = 1 462 BACKENDS_MAKE += $(MAGMA_BACKENDS) 463endif 464 465BACKENDS ?= $(BACKENDS_MAKE) 466export BACKENDS 467 468_pkg_ldflags = $(filter -L%,$(PKG_LIBS)) 469_pkg_ldlibs = $(filter-out -L%,$(PKG_LIBS)) 470$(libceeds) : CEED_LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 471$(libceeds) : CEED_LDLIBS += $(_pkg_ldlibs) 472ifeq ($(STATIC),1) 473$(examples) $(tests) : CEED_LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 474$(examples) $(tests) : CEED_LDLIBS += $(_pkg_ldlibs) 475endif 476 477pkgconfig-libs-private = $(PKG_LIBS) 478ifeq ($(LIBCEED_CONTAINS_CXX),1) 479 $(libceeds) : LINK = $(CXX) 480 ifeq ($(STATIC),1) 481 $(examples) $(tests) : CEED_LDLIBS += $(LIBCXX) 482 pkgconfig-libs-private += $(LIBCXX) 483 endif 484endif 485 486# File names *-weak.c contain weak symbol definitions, which must be listed last 487# when creating shared or static libraries. 488weak_last = $(filter-out %-weak.o,$(1)) $(filter %-weak.o,$(1)) 489 490libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) 491$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 492$(libceed.o): | info-backends 493$(libceed.so) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 494 $(call quiet,LINK) $(LDFLAGS) $(CEED_LDFLAGS) -shared -o $@ $^ $(CEED_LDLIBS) $(LDLIBS) 495 496$(libceed.a) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 497 $(call quiet,AR) $(ARFLAGS) $@ $^ 498 499$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 500 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 501 502$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 503 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 504 505$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 506 $(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<) 507 508$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR 509 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 510 511$(OBJDIR)/% : tests/%.c | $$(@D)/.DIR 512 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 513 514$(OBJDIR)/% : tests/%.f90 | $$(@D)/.DIR 515 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 516 517$(OBJDIR)/% : examples/ceed/%.c | $$(@D)/.DIR 518 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 519 520$(OBJDIR)/% : examples/ceed/%.f | $$(@D)/.DIR 521 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 522 523$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 524 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 525 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 526 cp examples/mfem/$* $@ 527 528# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 529# this single Nek bps file to be built in parallel with other examples, 530# such as when calling `make prove-all -j2`. 531$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 532 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 533 mv examples/nek/build/bps $(OBJDIR)/bps 534 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 535 536# Several executables have common utilities, but we can't build the utilities 537# from separate submake invocations because they'll compete with each 538# other/corrupt output. So we put it in this utility library, but we don't want 539# to manually list source dependencies up at this level, so we'll just always 540# call recursive make to check that this utility is up to date. 541examples/petsc/libutils.a.PHONY: $(libceed) $(ceed.pc) 542 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` AR=$(AR) ARFLAGS=$(ARFLAGS) \ 543 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $(basename $(@F)) 544 545$(OBJDIR)/petsc-% : examples/petsc/%.c examples/petsc/libutils.a.PHONY $(libceed) $(ceed.pc) | $$(@D)/.DIR 546 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` \ 547 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 548 cp examples/petsc/$* $@ 549 550$(OBJDIR)/fluids-% : examples/fluids/%.c examples/fluids/src/*.c examples/fluids/*.h examples/fluids/problems/*.c examples/fluids/qfunctions/*.h $(libceed) $(ceed.pc) | $$(@D)/.DIR 551 +$(call quiet,MAKE) -C examples/fluids CEED_DIR=`pwd` \ 552 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 553 cp examples/fluids/$* $@ 554 555$(OBJDIR)/solids-% : examples/solids/%.c examples/solids/%.h \ 556 examples/solids/problems/*.c examples/solids/src/*.c \ 557 examples/solids/include/*.h examples/solids/problems/*.h examples/solids/qfunctions/*.h \ 558 $(libceed) $(ceed.pc) | $$(@D)/.DIR 559 +$(call quiet,MAKE) -C examples/solids CEED_DIR=`pwd` \ 560 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 561 cp examples/solids/$* $@ 562 563$(examples) : $(libceed) 564$(tests) : $(libceed) 565$(tests) $(examples) : override LDFLAGS += -Wl,-rpath,$(abspath $(LIBDIR)) -L$(LIBDIR) 566 567run-% : $(OBJDIR)/% 568 @tests/tap.sh $(<:$(OBJDIR)/%=%) 569 570external_examples := \ 571 $(if $(MFEM_DIR),$(mfemexamples)) \ 572 $(if $(PETSC_DIR),$(petscexamples)) \ 573 $(if $(NEK5K_DIR),$(nekexamples)) \ 574 $(if $(PETSC_DIR),$(fluidsexamples)) \ 575 $(if $(PETSC_DIR),$(solidsexamples)) 576 577allexamples = $(examples) $(external_examples) 578 579# The test and prove targets can be controlled via pattern searches. The 580# default is to run tests and those examples that have no external dependencies. 581# Examples of finer grained control: 582# 583# make test search='petsc mfem' # PETSc and MFEM examples 584# make prove search='t3' # t3xx series tests 585# make junit search='ex petsc' # core ex* and PETSc tests 586search ?= t ex 587realsearch = $(search:%=%%) 588matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 589 590# Test core libCEED 591test : $(matched:$(OBJDIR)/%=run-%) 592 593# Run test target in parallel 594tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 595# CPU C tests only for backend % 596ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 597 598prove : $(matched) 599 $(info Testing backends: $(BACKENDS)) 600 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 601# Run prove target in parallel 602prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 603 604prove-all : 605 +$(MAKE) prove realsearch=% 606 607junit-% : $(OBJDIR)/% 608 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 609 610junit : $(matched:$(OBJDIR)/%=junit-%) 611 612all: $(alltests) 613 614examples : $(allexamples) 615ceedexamples : $(examples) 616nekexamples : $(nekexamples) 617mfemexamples : $(mfemexamples) 618petscexamples : $(petscexamples) 619 620# Benchmarks 621allbenchmarks = petsc-bps 622bench_targets = $(addprefix bench-,$(allbenchmarks)) 623.PHONY: $(bench_targets) benchmarks 624$(bench_targets): bench-%: $(OBJDIR)/% 625 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS_MAKE)" -r $(*).sh 626benchmarks: $(bench_targets) 627 628$(ceed.pc) : pkgconfig-prefix = $(abspath .) 629$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 630.INTERMEDIATE : $(OBJDIR)/ceed.pc 631%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 632 @$(SED) \ 633 -e "s:%prefix%:$(pkgconfig-prefix):" \ 634 -e "s:%libs_private%:$(pkgconfig-libs-private):" $< > $@ 635 636$(OBJDIR)/interface/ceed-jit-source-root-default.o : CPPFLAGS += -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath ./include)/\"" 637$(OBJDIR)/interface/ceed-jit-source-root-install.o : CPPFLAGS += -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath $(includedir))/\"" 638 639install : $(libceed) $(OBJDIR)/ceed.pc 640 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 641 "$(includedir)/ceed/" "$(includedir)/ceed/jit-source/"\ 642 "$(includedir)/ceed/jit-source/cuda/" "$(includedir)/ceed/jit-source/hip/"\ 643 "$(includedir)/ceed/jit-source/gallery/" "$(libdir)" "$(pkgconfigdir)") 644 $(INSTALL_DATA) include/ceed/ceed.h "$(DESTDIR)$(includedir)/ceed/" 645 $(INSTALL_DATA) include/ceed/types.h "$(DESTDIR)$(includedir)/ceed/" 646 $(INSTALL_DATA) include/ceed/ceed-f32.h "$(DESTDIR)$(includedir)/ceed/" 647 $(INSTALL_DATA) include/ceed/ceed-f64.h "$(DESTDIR)$(includedir)/ceed/" 648 $(INSTALL_DATA) include/ceed/fortran.h "$(DESTDIR)$(includedir)/ceed/" 649 $(INSTALL_DATA) include/ceed/backend.h "$(DESTDIR)$(includedir)/ceed/" 650 $(INSTALL_DATA) include/ceed/cuda.h "$(DESTDIR)$(includedir)/ceed/" 651 $(INSTALL_DATA) include/ceed/hip.h "$(DESTDIR)$(includedir)/ceed/" 652 $(INSTALL_DATA) include/ceed/hash.h "$(DESTDIR)$(includedir)/ceed/" 653 $(INSTALL_DATA) include/ceed/khash.h "$(DESTDIR)$(includedir)/ceed/" 654 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 655 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 656 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 657 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 658 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/cuda/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/cuda/" 659 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/hip/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/hip/" 660 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/gallery/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/gallery/" 661 662.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 663 664cln clean : 665 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 666 $(call quiet,MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 667 $(call quiet,MAKE) -C python/tests clean 668 $(RM) benchmarks/*output.txt 669 670distclean : clean 671 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 672 673DOXYGEN ?= doxygen 674doxygen : 675 $(DOXYGEN) -q Doxyfile 676 677doc-html doc-latexpdf doc-epub doc-livehtml : doc-% : doxygen 678 make -C doc/sphinx $* 679 680doc : doc-html 681 682style-c : 683 @astyle --options=.astylerc \ 684 $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), \ 685 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 686 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 687 688AUTOPEP8 = autopep8 689style-py : AUTOPEP8_ARGS = --in-place --aggressive 690style-py : 691 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py python/tests/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 692 693style : style-c style-py 694 695CLANG_TIDY ?= clang-tidy 696 697%.c.tidy : %.c 698 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include -I$(HIP_DIR)/include -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath ./include)/\"" 699 700%.cpp.tidy : %.cpp 701 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include -I$(HIP_DIR)/include 702 703tidy_c : $(libceed.c:%=%.tidy) 704tidy_cpp : $(libceed.cpp:%=%.tidy) 705 706tidy : tidy_c tidy_cpp 707 708ifneq ($(wildcard ../iwyu/*),) 709 IWYU_DIR ?= ../iwyu 710 IWYU_CC ?= $(IWYU_DIR)/build/bin/include-what-you-use 711endif 712 713iwyu : CC=$(IWYU_CC) 714iwyu : lib 715 716print : 717 @echo $(VAR)=$($(VAR)) 718 719print-% : 720 $(info [ variable name]: $*) 721 $(info [ origin]: $(origin $*)) 722 $(info [ flavor]: $(flavor $*)) 723 $(info [ value]: $(value $*)) 724 $(info [expanded value]: $($*)) 725 $(info ) 726 @true 727 728# "make configure" detects any variables passed on the command line or 729# previously set in config.mk, caching them in config.mk as simple 730# (:=) variables. Variables set in config.mk or on the command line 731# take precedence over the defaults provided in the file. Typical 732# usage: 733# 734# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 735# make 736# make prove 737# 738# The values in the file can be updated by passing them on the command 739# line, e.g., 740# 741# make configure CC=/path/to/other/clang 742 743# All variables to consider for caching 744CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 745 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 746 AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \ 747 MAGMA_DIR OCCA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR HIP_ARCH 748 749# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 750# was set on the command line or in config.mk (where it will appear as 751# a simple variable). 752needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 753 754configure : 755 $(file > $(CONFIG)) 756 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 757 @echo "Configuration cached in $(CONFIG):" 758 @cat $(CONFIG) 759 760wheel : export MARCHFLAG = -march=generic 761wheel : export WHEEL_PLAT = manylinux2010_x86_64 762wheel : 763 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 764 -e MARCHFLAG -e WHEEL_PLAT \ 765 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 766 767.PHONY : configure wheel 768 769# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 770-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 771