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 48# Allow users to customize LDFLAGS (libceed appends directly to this variable) 49override LDFLAGS := $(LDFLAGS) 50 51UNDERSCORE ?= 1 52 53# Verbose mode, V or VERBOSE 54V ?= $(VERBOSE) 55 56# MFEM_DIR env variable should point to sibling directory 57ifneq ($(wildcard ../mfem/libmfem.*),) 58 MFEM_DIR ?= ../mfem 59endif 60 61# NEK5K_DIR env variable should point to sibling directory 62ifneq ($(wildcard ../Nek5000/*),) 63 NEK5K_DIR ?= $(abspath ../Nek5000) 64endif 65export NEK5K_DIR 66MPI ?= 1 67 68# CEED_DIR env for NEK5K testing 69export CEED_DIR = $(abspath .) 70 71# XSMM_DIR env variable should point to XSMM master (github.com/hfp/libxsmm) 72XSMM_DIR ?= ../libxsmm 73 74# OCCA_DIR env variable should point to OCCA master (github.com/libocca/occa) 75OCCA_DIR ?= ../occa 76 77# env variable MAGMA_DIR can be used too 78MAGMA_DIR ?= ../magma 79 80# Often /opt/cuda or /usr/local/cuda, but sometimes present on machines that don't support CUDA 81CUDA_DIR ?= 82CUDA_ARCH ?= 83 84# Often /opt/rocm, but sometimes present on machines that don't support HIP 85HIP_DIR ?= 86HIP_ARCH ?= 87 88# Check for PETSc in ../petsc 89ifneq ($(wildcard ../petsc/lib/libpetsc.*),) 90 PETSC_DIR ?= ../petsc 91endif 92 93# Warning: SANTIZ options still don't run with /gpu/occa 94# export LSAN_OPTIONS=suppressions=.asanignore 95AFLAGS = -fsanitize=address #-fsanitize=undefined -fno-omit-frame-pointer 96 97# Note: Intel oneAPI C/C++ compiler is now icx/icpx 98CC_VENDOR := $(subst oneAPI,icc,$(firstword $(filter gcc clang icc oneAPI XL,$(subst -, ,$(shell $(CC) --version))))) 99FC_VENDOR := $(if $(FC),$(firstword $(filter GNU ifort XL,$(shell $(FC) --version 2>&1 || $(FC) -qversion)))) 100 101# Default extra flags by vendor 102MARCHFLAG.gcc := -march=native 103MARCHFLAG.clang := $(MARCHFLAG.gcc) 104MARCHFLAG.icc := 105OMP_SIMD_FLAG.gcc := -fopenmp-simd 106OMP_SIMD_FLAG.clang := $(OMP_SIMD_FLAG.gcc) 107OMP_SIMD_FLAG.icc := -qopenmp-simd 108OPT.gcc := -ffp-contract=fast 109OPT.clang := $(OPT.gcc) 110CFLAGS.gcc := -fPIC -std=c99 -Wall -Wextra -Wno-unused-parameter -MMD -MP 111CFLAGS.clang := $(CFLAGS.gcc) 112CFLAGS.icc := $(CFLAGS.gcc) 113CFLAGS.XL := -qpic -MMD 114CXXFLAGS.gcc := -fPIC -std=c++11 -Wall -Wextra -Wno-unused-parameter -MMD -MP 115CXXFLAGS.clang := $(CXXFLAGS.gcc) 116CXXFLAGS.icc := $(CXXFLAGS.gcc) 117CXXFLAGS.XL := -qpic -std=c++11 -MMD 118FFLAGS.GNU := -fPIC -cpp -Wall -Wextra -Wno-unused-parameter -Wno-unused-dummy-argument -MMD -MP 119FFLAGS.ifort := -fPIC -cpp 120FFLAGS.XL := -qpic -ffree-form -qpreprocess -qextname -MMD 121 122# This check works with compilers that use gcc and clang. It fails with some 123# compilers; e.g., xlc apparently ignores all options when -E is passed, thus 124# succeeds with any flags. Users can pass MARCHFLAG=... if desired. 125cc_check_flag = $(shell $(CC) -E -Werror $(1) -x c /dev/null > /dev/null 2>&1 && echo 1) 126MARCHFLAG := $(MARCHFLAG.$(CC_VENDOR)) 127MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG),-mcpu=native) 128MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG)) 129 130OMP_SIMD_FLAG := $(OMP_SIMD_FLAG.$(CC_VENDOR)) 131OMP_SIMD_FLAG := $(if $(call cc_check_flag,$(OMP_SIMD_FLAG)),$(OMP_SIMD_FLAG)) 132 133OPT ?= -O -g $(MARCHFLAG) $(OPT.$(CC_VENDOR)) $(OMP_SIMD_FLAG) 134CFLAGS ?= $(OPT) $(CFLAGS.$(CC_VENDOR)) 135CXXFLAGS ?= $(OPT) $(CXXFLAGS.$(CC_VENDOR)) 136LIBCXX ?= -lstdc++ 137NVCCFLAGS ?= -ccbin $(CXX) -Xcompiler "$(OPT)" -Xcompiler -fPIC 138ifneq ($(CUDA_ARCH),) 139 NVCCFLAGS += -arch=$(CUDA_ARCH) 140endif 141HIPCCFLAGS ?= $(filter-out $(OMP_SIMD_FLAG),$(OPT)) -fPIC -munsafe-fp-atomics 142ifneq ($(HIP_ARCH),) 143 HIPCCFLAGS += --amdgpu-target=$(HIP_ARCH) 144endif 145FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR)) 146 147ifeq ($(COVERAGE), 1) 148 CFLAGS += --coverage 149 CXXFLAGS += --coverage 150 override LDFLAGS += --coverage 151endif 152 153CFLAGS += $(if $(ASAN),$(AFLAGS)) 154FFLAGS += $(if $(ASAN),$(AFLAGS)) 155override LDFLAGS += $(if $(ASAN),$(AFLAGS)) 156CPPFLAGS += -I./include 157LDLIBS = -lm 158OBJDIR := build 159LIBDIR := lib 160 161# Installation variables 162prefix ?= /usr/local 163bindir = $(prefix)/bin 164libdir = $(prefix)/lib 165includedir = $(prefix)/include 166pkgconfigdir = $(libdir)/pkgconfig 167INSTALL = install 168INSTALL_PROGRAM = $(INSTALL) 169INSTALL_DATA = $(INSTALL) -m644 170 171# Get number of processors of the machine 172NPROCS := $(shell getconf _NPROCESSORS_ONLN) 173# prepare make options to run in parallel 174MFLAGS := -j $(NPROCS) --warn-undefined-variables \ 175 --no-print-directory --no-keep-going 176 177PYTHON ?= python3 178PROVE ?= prove 179PROVE_OPTS ?= -j $(NPROCS) 180DARWIN := $(filter Darwin,$(shell uname -s)) 181SO_EXT := $(if $(DARWIN),dylib,so) 182 183ceed.pc := $(LIBDIR)/pkgconfig/ceed.pc 184libceed.so := $(LIBDIR)/libceed.$(SO_EXT) 185libceed.a := $(LIBDIR)/libceed.a 186libceed := $(if $(STATIC),$(libceed.a),$(libceed.so)) 187CEED_LIBS = -lceed 188libceed.c := $(filter-out interface/ceed-cuda.c interface/ceed-hip.c, $(wildcard interface/ceed*.c backends/*.c gallery/*.c)) 189gallery.c := $(wildcard gallery/*/ceed*.c) 190libceed.c += $(gallery.c) 191libceeds = $(libceed) 192BACKENDS_BUILTIN := /cpu/self/ref/serial /cpu/self/ref/blocked /cpu/self/opt/serial /cpu/self/opt/blocked 193BACKENDS_MAKE := $(BACKENDS_BUILTIN) 194TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub 195 196# Tests 197tests.c := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.c)) 198tests.f := $(if $(FC),$(sort $(wildcard tests/t[0-9][0-9][0-9]-*.f90))) 199tests := $(tests.c:tests/%.c=$(OBJDIR)/%) 200ctests := $(tests) 201tests += $(tests.f:tests/%.f90=$(OBJDIR)/%) 202# Examples 203examples.c := $(sort $(wildcard examples/ceed/*.c)) 204examples.f := $(if $(FC),$(sort $(wildcard examples/ceed/*.f))) 205examples := $(examples.c:examples/ceed/%.c=$(OBJDIR)/%) 206examples += $(examples.f:examples/ceed/%.f=$(OBJDIR)/%) 207# MFEM Examples 208mfemexamples.cpp := $(sort $(wildcard examples/mfem/*.cpp)) 209mfemexamples := $(mfemexamples.cpp:examples/mfem/%.cpp=$(OBJDIR)/mfem-%) 210# Nek5K Examples 211nekexamples := $(OBJDIR)/nek-bps 212# PETSc Examples 213petscexamples.c := $(wildcard examples/petsc/*.c) 214petscexamples := $(petscexamples.c:examples/petsc/%.c=$(OBJDIR)/petsc-%) 215# Fluid Dynamics Examples 216fluidsexamples.c := $(sort $(wildcard examples/fluids/*.c)) 217fluidsexamples := $(fluidsexamples.c:examples/fluids/%.c=$(OBJDIR)/fluids-%) 218# Solid Mechanics Examples 219solidsexamples.c := $(sort $(wildcard examples/solids/*.c)) 220solidsexamples := $(solidsexamples.c:examples/solids/%.c=$(OBJDIR)/solids-%) 221 222# Backends/[ref, blocked, memcheck, opt, avx, occa, magma] 223ref.c := $(sort $(wildcard backends/ref/*.c)) 224blocked.c := $(sort $(wildcard backends/blocked/*.c)) 225ceedmemcheck.c := $(sort $(wildcard backends/memcheck/*.c)) 226opt.c := $(sort $(wildcard backends/opt/*.c)) 227avx.c := $(sort $(wildcard backends/avx/*.c)) 228xsmm.c := $(sort $(wildcard backends/xsmm/*.c)) 229cuda.c := $(sort $(wildcard backends/cuda/*.c)) 230cuda-ref.c := $(sort $(wildcard backends/cuda-ref/*.c)) 231cuda-ref.cpp := $(sort $(wildcard backends/cuda-ref/*.cpp)) 232cuda-ref.cu := $(sort $(wildcard backends/cuda-ref/kernels/*.cu)) 233cuda-shared.c := $(sort $(wildcard backends/cuda-shared/*.c)) 234cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/kernels/*.cu)) 235cuda-gen.c := $(sort $(wildcard backends/cuda-gen/*.c)) 236cuda-gen.cpp := $(sort $(wildcard backends/cuda-gen/*.cpp)) 237cuda-gen.cu := $(sort $(wildcard backends/cuda-gen/kernels/*.cu)) 238occa.cpp := $(sort $(shell find backends/occa -type f -name *.cpp)) 239magma.c := $(sort $(wildcard backends/magma/*.c)) 240magma.cu := $(sort $(wildcard backends/magma/kernels/cuda/*.cu)) 241magma.hip := $(sort $(wildcard backends/magma/kernels/hip/*.hip.cpp)) 242hip.c := $(sort $(wildcard backends/hip/*.c)) 243hip.cpp := $(sort $(wildcard backends/hip/*.cpp)) 244hip-ref.c := $(sort $(wildcard backends/hip-ref/*.c)) 245hip-ref.cpp := $(sort $(wildcard backends/hip-ref/*.cpp)) 246hip-ref.hip := $(sort $(wildcard backends/hip-ref/kernels/*.hip.cpp)) 247hip-shared.c := $(sort $(wildcard backends/hip-shared/*.c)) 248hip-gen.c := $(sort $(wildcard backends/hip-gen/*.c)) 249hip-gen.cpp := $(sort $(wildcard backends/hip-gen/*.cpp)) 250 251# Quiet, color output 252quiet ?= $($(1)) 253 254# Cancel built-in and old-fashioned implicit rules which we don't use 255.SUFFIXES: 256 257.SECONDEXPANSION: # to expand $$(@D)/.DIR 258 259%/.DIR : 260 @mkdir -p $(@D) 261 @touch $@ 262 263.PRECIOUS: %/.DIR 264 265lib: $(libceed) $(ceed.pc) 266# run 'lib' target in parallel 267par:;@$(MAKE) $(MFLAGS) V=$(V) lib 268backend_status = $(if $(filter $1,$(BACKENDS_MAKE)), [backends: $1], [not found]) 269info: 270 $(info ------------------------------------) 271 $(info CC = $(CC)) 272 $(info CXX = $(CXX)) 273 $(info FC = $(FC)) 274 $(info CPPFLAGS = $(CPPFLAGS)) 275 $(info CFLAGS = $(value CFLAGS)) 276 $(info CXXFLAGS = $(value CXXFLAGS)) 277 $(info FFLAGS = $(value FFLAGS)) 278 $(info NVCCFLAGS = $(value NVCCFLAGS)) 279 $(info HIPCCFLAGS = $(value HIPCCFLAGS)) 280 $(info LDFLAGS = $(value LDFLAGS)) 281 $(info LDLIBS = $(LDLIBS)) 282 $(info AR = $(AR)) 283 $(info ARFLAGS = $(ARFLAGS)) 284 $(info OPT = $(OPT)) 285 $(info AFLAGS = $(AFLAGS)) 286 $(info ASAN = $(or $(ASAN),(empty))) 287 $(info VERBOSE = $(or $(V),(empty)) [verbose=$(if $(V),on,off)]) 288 $(info ------------------------------------) 289 $(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,$(MEMCHK_BACKENDS))) 290 $(info AVX_STATUS = $(AVX_STATUS)$(call backend_status,$(AVX_BACKENDS))) 291 $(info XSMM_DIR = $(XSMM_DIR)$(call backend_status,$(XSMM_BACKENDS))) 292 $(info OCCA_DIR = $(OCCA_DIR)$(call backend_status,$(OCCA_BACKENDS))) 293 $(info MAGMA_DIR = $(MAGMA_DIR)$(call backend_status,$(MAGMA_BACKENDS))) 294 $(info CUDA_DIR = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS))) 295 $(info HIP_DIR = $(HIP_DIR)$(call backend_status,$(HIP_BACKENDS))) 296 $(info ------------------------------------) 297 $(info MFEM_DIR = $(MFEM_DIR)) 298 $(info NEK5K_DIR = $(NEK5K_DIR)) 299 $(info PETSC_DIR = $(PETSC_DIR)) 300 $(info ------------------------------------) 301 $(info prefix = $(prefix)) 302 $(info includedir = $(value includedir)) 303 $(info libdir = $(value libdir)) 304 $(info pkgconfigdir = $(value pkgconfigdir)) 305 $(info ------------------------------------) 306 @true 307info-backends: 308 $(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS))) 309 @true 310info-backends-all: 311 $(info make: 'lib' with backends: $(filter-out $(TEST_BACKENDS),$(BACKENDS))) 312 @true 313 314$(libceed.so) : override LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed.so))) 315 316# Standard Backends 317libceed.c += $(ref.c) 318libceed.c += $(blocked.c) 319libceed.c += $(opt.c) 320 321# Memcheck Backend 322MEMCHK_STATUS = Disabled 323MEMCHK := $(shell echo "$(HASH)include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1) 324MEMCHK_BACKENDS = /cpu/self/memcheck/serial /cpu/self/memcheck/blocked 325ifeq ($(MEMCHK),1) 326 MEMCHK_STATUS = Enabled 327 libceed.c += $(ceedmemcheck.c) 328 BACKENDS_MAKE += $(MEMCHK_BACKENDS) 329endif 330 331# AVX Backed 332AVX_STATUS = Disabled 333AVX_FLAG := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx) 334AVX := $(filter $(AVX_FLAG),$(shell $(CC) $(CFLAGS) -v -E -x c /dev/null 2>&1)) 335AVX_BACKENDS = /cpu/self/avx/serial /cpu/self/avx/blocked 336ifneq ($(AVX),) 337 AVX_STATUS = Enabled 338 libceed.c += $(avx.c) 339 BACKENDS_MAKE += $(AVX_BACKENDS) 340endif 341 342# Collect list of libraries and paths for use in linking and pkg-config 343PKG_LIBS = 344# Stubs that will not be RPATH'd 345PKG_STUBS_LIBS = 346 347# libXSMM Backends 348XSMM_BACKENDS = /cpu/self/xsmm/serial /cpu/self/xsmm/blocked 349ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),) 350 PKG_LIBS += -L$(abspath $(XSMM_DIR))/lib -lxsmm -ldl 351 MKL ?= 352 ifeq (,$(MKL)$(MKLROOT)) 353 BLAS_LIB = -lblas 354 else 355 ifneq ($(MKLROOT),) 356 # Some installs put everything inside an intel64 subdirectory, others not 357 MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*))) 358 MKL_LINK = -L$(MKL_LIBDIR) 359 PKG_LIB_DIRS += $(MKL_LIBDIR) 360 endif 361 BLAS_LIB = $(MKL_LINK) -Wl,--push-state,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl -Wl,--pop-state 362 endif 363 PKG_LIBS += $(BLAS_LIB) 364 libceed.c += $(xsmm.c) 365 $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include 366 BACKENDS_MAKE += $(XSMM_BACKENDS) 367endif 368 369# OCCA Backends 370OCCA_BACKENDS = /cpu/self/occa 371ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),) 372 OCCA_MODES := $(shell $(OCCA_DIR)/bin/occa modes) 373 OCCA_BACKENDS += $(if $(filter OpenMP,$(OCCA_MODES)),/cpu/openmp/occa) 374# OCCA_BACKENDS += $(if $(filter OpenCL,$(OCCA_MODES)),/gpu/opencl/occa) 375 OCCA_BACKENDS += $(if $(filter HIP,$(OCCA_MODES)),/gpu/hip/occa) 376 OCCA_BACKENDS += $(if $(filter CUDA,$(OCCA_MODES)),/gpu/cuda/occa) 377 378 $(libceeds) : CPPFLAGS += -I$(OCCA_DIR)/include 379 PKG_LIBS += -L$(abspath $(OCCA_DIR))/lib -locca 380 LIBCEED_CONTAINS_CXX = 1 381 libceed.cpp += $(occa.cpp) 382 BACKENDS_MAKE += $(OCCA_BACKENDS) 383endif 384 385# CUDA Backends 386CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(CUDA_DIR)/$d/libcudart.${SO_EXT})) 387CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR)))) 388CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs 389CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/shared /gpu/cuda/gen 390ifneq ($(CUDA_LIB_DIR),) 391 $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include 392 PKG_LIBS += -L$(abspath $(CUDA_LIB_DIR)) -lcudart -lnvrtc -lcuda -lcublas 393 PKG_STUBS_LIBS += -L$(CUDA_LIB_DIR_STUBS) 394 LIBCEED_CONTAINS_CXX = 1 395 libceed.c += interface/ceed-cuda.c 396 libceed.c += $(cuda.c) $(cuda-ref.c) $(cuda-shared.c) $(cuda-gen.c) 397 libceed.cpp += $(cuda-ref.cpp) $(cuda-gen.cpp) 398 libceed.cu += $(cuda-ref.cu) $(cuda-shared.cu) $(cuda-gen.cu) 399 BACKENDS_MAKE += $(CUDA_BACKENDS) 400endif 401 402# HIP Backends 403HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libamdhip64.${SO_EXT})) 404HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 405HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen 406ifneq ($(HIP_LIB_DIR),) 407 $(libceeds) : HIPCCFLAGS += -I./include 408 ifneq ($(CXX), $(HIPCC)) 409 CPPFLAGS += $(subst =,,$(shell $(HIP_DIR)/bin/hipconfig -C)) 410 endif 411 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include 412 PKG_LIBS += -L$(abspath $(HIP_LIB_DIR)) -lamdhip64 -lhipblas 413 LIBCEED_CONTAINS_CXX = 1 414 libceed.c += interface/ceed-hip.c 415 libceed.c += $(hip.c) $(hip-ref.c) $(hip-shared.c) $(hip-gen.c) 416 libceed.cpp += $(hip.cpp) $(hip-ref.cpp) $(hip-gen.cpp) 417 libceed.hip += $(hip-ref.hip) 418 BACKENDS_MAKE += $(HIP_BACKENDS) 419endif 420 421# MAGMA Backend 422ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 423 MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas") 424 ifeq ($(MAGMA_ARCH), 0) #CUDA MAGMA 425 ifneq ($(CUDA_LIB_DIR),) 426 cuda_link = -Wl,-rpath,$(CUDA_LIB_DIR) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart 427 omp_link = -fopenmp 428 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link) 429 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 430 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 431 PKG_LIBS += $(magma_link) 432 libceed.c += $(magma.c) 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.cu:%.cu=$(OBJDIR)/%.o) : CPPFLAGS += --compiler-options=-fPIC -DADD_ -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(CUDA_DIR)/include 436 MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det 437 endif 438 else # HIP MAGMA 439 ifneq ($(HIP_LIB_DIR),) 440 hip_link = -Wl,-rpath,$(HIP_LIB_DIR) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64 441 omp_link = -fopenmp 442 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link) 443 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 444 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 445 PKG_LIBS += $(magma_link) 446 libceed.c += $(magma.c) 447 libceed.hip += $(magma.hip) 448 ifneq ($(CXX), $(HIPCC)) 449 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 450 else 451 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 452 endif 453 $(magma.hip:%.hip.cpp=$(OBJDIR)/%.o) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 454 MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det 455 endif 456 endif 457 LIBCEED_CONTAINS_CXX = 1 458 BACKENDS_MAKE += $(MAGMA_BACKENDS) 459endif 460 461BACKENDS ?= $(BACKENDS_MAKE) 462export BACKENDS 463 464_pkg_ldflags = $(filter -L%,$(PKG_LIBS)) 465_pkg_ldlibs = $(filter-out -L%,$(PKG_LIBS)) 466$(libceeds) : override LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 467$(libceeds) : LDLIBS += $(_pkg_ldlibs) 468ifeq ($(STATIC),1) 469$(examples) $(tests) : override LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 470$(examples) $(tests) : LDLIBS += $(_pkg_ldlibs) 471endif 472 473pkgconfig-libs-private = $(PKG_LIBS) 474ifeq ($(LIBCEED_CONTAINS_CXX),1) 475 $(libceeds) : LINK = $(CXX) 476 ifeq ($(STATIC),1) 477 $(examples) $(tests) : LDLIBS += $(LIBCXX) 478 pkgconfig-libs-private += $(LIBCXX) 479 endif 480endif 481 482# File names *-weak.c contain weak symbol definitions, which must be listed last 483# when creating shared or static libraries. 484weak_last = $(filter-out %-weak.o,$(1)) $(filter %-weak.o,$(1)) 485 486libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) 487$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 488$(libceed.o): | info-backends 489$(libceed.so) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 490 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 491 492$(libceed.a) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 493 $(call quiet,AR) $(ARFLAGS) $@ $^ 494 495$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 496 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 497 498$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 499 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 500 501$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 502 $(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<) 503 504$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR 505 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 506 507$(OBJDIR)/% : tests/%.c | $$(@D)/.DIR 508 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 509 510$(OBJDIR)/% : tests/%.f90 | $$(@D)/.DIR 511 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 512 513$(OBJDIR)/% : examples/ceed/%.c | $$(@D)/.DIR 514 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 515 516$(OBJDIR)/% : examples/ceed/%.f | $$(@D)/.DIR 517 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 518 519$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 520 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 521 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 522 cp examples/mfem/$* $@ 523 524# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 525# this single Nek bps file to be built in parallel with other examples, 526# such as when calling `make prove-all -j2`. 527$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 528 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 529 mv examples/nek/build/bps $(OBJDIR)/bps 530 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 531 532# Several executables have common utilities, but we can't build the utilities 533# from separate submake invocations because they'll compete with each 534# other/corrupt output. So we put it in this utility library, but we don't want 535# to manually list source dependencies up at this level, so we'll just always 536# call recursive make to check that this utility is up to date. 537examples/petsc/libutils.a.PHONY: $(libceed) $(ceed.pc) 538 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` AR=$(AR) ARFLAGS=$(ARFLAGS) \ 539 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $(basename $(@F)) 540 541$(OBJDIR)/petsc-% : examples/petsc/%.c examples/petsc/libutils.a.PHONY $(libceed) $(ceed.pc) | $$(@D)/.DIR 542 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` \ 543 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 544 cp examples/petsc/$* $@ 545 546$(OBJDIR)/fluids-% : examples/fluids/%.c examples/fluids/src/*.c examples/fluids/*.h examples/fluids/problems/*.c examples/fluids/qfunctions/*.h $(libceed) $(ceed.pc) | $$(@D)/.DIR 547 +$(call quiet,MAKE) -C examples/fluids CEED_DIR=`pwd` \ 548 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 549 cp examples/fluids/$* $@ 550 551$(OBJDIR)/solids-% : examples/solids/%.c examples/solids/%.h \ 552 examples/solids/problems/*.c examples/solids/src/*.c \ 553 examples/solids/include/*.h examples/solids/problems/*.h examples/solids/qfunctions/*.h \ 554 $(libceed) $(ceed.pc) | $$(@D)/.DIR 555 +$(call quiet,MAKE) -C examples/solids CEED_DIR=`pwd` \ 556 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 557 cp examples/solids/$* $@ 558 559$(examples) : $(libceed) 560$(tests) : $(libceed) 561$(tests) $(examples) : override LDFLAGS += -Wl,-rpath,$(abspath $(LIBDIR)) -L$(LIBDIR) 562 563run-% : $(OBJDIR)/% 564 @tests/tap.sh $(<:$(OBJDIR)/%=%) 565 566external_examples := \ 567 $(if $(MFEM_DIR),$(mfemexamples)) \ 568 $(if $(PETSC_DIR),$(petscexamples)) \ 569 $(if $(NEK5K_DIR),$(nekexamples)) \ 570 $(if $(PETSC_DIR),$(fluidsexamples)) \ 571 $(if $(PETSC_DIR),$(solidsexamples)) 572 573allexamples = $(examples) $(external_examples) 574 575# The test and prove targets can be controlled via pattern searches. The 576# default is to run tests and those examples that have no external dependencies. 577# Examples of finer grained control: 578# 579# make test search='petsc mfem' # PETSc and MFEM examples 580# make prove search='t3' # t3xx series tests 581# make junit search='ex petsc' # core ex* and PETSc tests 582search ?= t ex 583realsearch = $(search:%=%%) 584matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 585 586# Test core libCEED 587test : $(matched:$(OBJDIR)/%=run-%) 588 589# Run test target in parallel 590tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 591# CPU C tests only for backend % 592ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 593 594prove : $(matched) 595 $(info Testing backends: $(BACKENDS)) 596 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 597# Run prove target in parallel 598prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 599 600prove-all : 601 +$(MAKE) prove realsearch=% 602 603junit-% : $(OBJDIR)/% 604 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 605 606junit : $(matched:$(OBJDIR)/%=junit-%) 607 608all: $(alltests) 609 610examples : $(allexamples) 611ceedexamples : $(examples) 612nekexamples : $(nekexamples) 613mfemexamples : $(mfemexamples) 614petscexamples : $(petscexamples) 615 616# Benchmarks 617allbenchmarks = petsc-bps 618bench_targets = $(addprefix bench-,$(allbenchmarks)) 619.PHONY: $(bench_targets) benchmarks 620$(bench_targets): bench-%: $(OBJDIR)/% 621 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS_MAKE)" -r $(*).sh 622benchmarks: $(bench_targets) 623 624$(ceed.pc) : pkgconfig-prefix = $(abspath .) 625$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 626.INTERMEDIATE : $(OBJDIR)/ceed.pc 627%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 628 @$(SED) \ 629 -e "s:%prefix%:$(pkgconfig-prefix):" \ 630 -e "s:%libs_private%:$(pkgconfig-libs-private):" $< > $@ 631 632install : $(libceed) $(OBJDIR)/ceed.pc 633 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 634 "$(includedir)/ceed/" "$(libdir)" "$(pkgconfigdir)") 635 $(INSTALL_DATA) include/ceed/ceed.h "$(DESTDIR)$(includedir)/ceed/" 636 $(INSTALL_DATA) include/ceed/ceed-f32.h "$(DESTDIR)$(includedir)/ceed/" 637 $(INSTALL_DATA) include/ceed/ceed-f64.h "$(DESTDIR)$(includedir)/ceed/" 638 $(INSTALL_DATA) include/ceed/fortran.h "$(DESTDIR)$(includedir)/ceed/" 639 $(INSTALL_DATA) include/ceed/backend.h "$(DESTDIR)$(includedir)/ceed/" 640 $(INSTALL_DATA) include/ceed/cuda.h "$(DESTDIR)$(includedir)/ceed/" 641 $(INSTALL_DATA) include/ceed/hip.h "$(DESTDIR)$(includedir)/ceed/" 642 $(INSTALL_DATA) include/ceed/hash.h "$(DESTDIR)$(includedir)/ceed/" 643 $(INSTALL_DATA) include/ceed/khash.h "$(DESTDIR)$(includedir)/ceed/" 644 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 645 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 646 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 647 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 648 649.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 650 651cln clean : 652 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 653 $(call quiet,MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 654 $(call quiet,MAKE) -C python/tests clean 655 $(RM) benchmarks/*output.txt 656 657distclean : clean 658 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 659 660DOXYGEN ?= doxygen 661doxygen : 662 $(DOXYGEN) -q Doxyfile 663 664doc-html doc-latexpdf doc-epub doc-livehtml : doc-% : doxygen 665 make -C doc/sphinx $* 666 667doc : doc-html 668 669style-c : 670 @astyle --options=.astylerc \ 671 $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), \ 672 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 673 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 674 675AUTOPEP8 = autopep8 676style-py : AUTOPEP8_ARGS = --in-place --aggressive 677style-py : 678 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py python/tests/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 679 680style : style-c style-py 681 682CLANG_TIDY ?= clang-tidy 683 684%.c.tidy : %.c 685 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include -I$(HIP_DIR)/include 686 687%.cpp.tidy : %.cpp 688 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include -I$(HIP_DIR)/include 689 690tidy_c : $(libceed.c:%=%.tidy) 691tidy_cpp : $(libceed.cpp:%=%.tidy) 692 693tidy : tidy_c tidy_cpp 694 695ifneq ($(wildcard ../iwyu/*),) 696 IWYU_DIR ?= ../iwyu 697 IWYU_CC ?= $(IWYU_DIR)/build/bin/include-what-you-use 698endif 699 700iwyu : CC=$(IWYU_CC) 701iwyu : lib 702 703print : 704 @echo $(VAR)=$($(VAR)) 705 706print-% : 707 $(info [ variable name]: $*) 708 $(info [ origin]: $(origin $*)) 709 $(info [ flavor]: $(flavor $*)) 710 $(info [ value]: $(value $*)) 711 $(info [expanded value]: $($*)) 712 $(info ) 713 @true 714 715# "make configure" detects any variables passed on the command line or 716# previously set in config.mk, caching them in config.mk as simple 717# (:=) variables. Variables set in config.mk or on the command line 718# take precedence over the defaults provided in the file. Typical 719# usage: 720# 721# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 722# make 723# make prove 724# 725# The values in the file can be updated by passing them on the command 726# line, e.g., 727# 728# make configure CC=/path/to/other/clang 729 730# All variables to consider for caching 731CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 732 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 733 AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \ 734 MAGMA_DIR OCCA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR HIP_ARCH 735 736# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 737# was set on the command line or in config.mk (where it will appear as 738# a simple variable). 739needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 740 741configure : 742 $(file > $(CONFIG)) 743 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 744 @echo "Configuration cached in $(CONFIG):" 745 @cat $(CONFIG) 746 747wheel : export MARCHFLAG = -march=generic 748wheel : export WHEEL_PLAT = manylinux2010_x86_64 749wheel : 750 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 751 -e MARCHFLAG -e WHEEL_PLAT \ 752 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 753 754.PHONY : configure wheel 755 756# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 757-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 758