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# Allow users to customize LDFLAGS (libceed appends directly to this variable) 40override LDFLAGS := $(LDFLAGS) 41 42UNDERSCORE ?= 1 43 44# Verbose mode, V or VERBOSE 45V ?= $(VERBOSE) 46 47# MFEM_DIR env variable should point to sibling directory 48ifneq ($(wildcard ../mfem/libmfem.*),) 49 MFEM_DIR ?= ../mfem 50endif 51 52# NEK5K_DIR env variable should point to sibling directory 53ifneq ($(wildcard ../Nek5000/*),) 54 NEK5K_DIR ?= $(abspath ../Nek5000) 55endif 56export NEK5K_DIR 57MPI ?= 1 58 59# CEED_DIR env for NEK5K testing 60export CEED_DIR = $(abspath .) 61 62# XSMM_DIR env variable should point to XSMM master (github.com/hfp/libxsmm) 63XSMM_DIR ?= ../libxsmm 64 65# OCCA_DIR env variable should point to OCCA master (github.com/libocca/occa) 66OCCA_DIR ?= ../occa 67 68# env variable MAGMA_DIR can be used too 69MAGMA_DIR ?= ../magma 70 71# Often /opt/cuda or /usr/local/cuda, but sometimes present on machines that don't support CUDA 72CUDA_DIR ?= 73CUDA_ARCH ?= 74 75# Often /opt/rocm, but sometimes present on machines that don't support HIP 76HIP_DIR ?= 77HIP_ARCH ?= 78 79# Check for PETSc in ../petsc 80ifneq ($(wildcard ../petsc/lib/libpetsc.*),) 81 PETSC_DIR ?= ../petsc 82endif 83 84# Warning: SANTIZ options still don't run with /gpu/occa 85# export LSAN_OPTIONS=suppressions=.asanignore 86AFLAGS = -fsanitize=address #-fsanitize=undefined -fno-omit-frame-pointer 87 88# Note: Intel oneAPI C/C++ compiler is now icx/icpx 89CC_VENDOR := $(subst oneAPI,icc,$(firstword $(filter gcc clang icc oneAPI XL,$(subst -, ,$(shell $(CC) --version))))) 90FC_VENDOR := $(if $(FC),$(firstword $(filter GNU ifort XL,$(shell $(FC) --version 2>&1 || $(FC) -qversion)))) 91 92# Default extra flags by vendor 93MARCHFLAG.gcc := -march=native 94MARCHFLAG.clang := $(MARCHFLAG.gcc) 95MARCHFLAG.icc := 96OMP_SIMD_FLAG.gcc := -fopenmp-simd 97OMP_SIMD_FLAG.clang := $(OMP_SIMD_FLAG.gcc) 98OMP_SIMD_FLAG.icc := -qopenmp-simd 99OPT.gcc := -ffp-contract=fast 100OPT.clang := $(OPT.gcc) 101CFLAGS.gcc := -fPIC -std=c99 -Wall -Wextra -Wno-unused-parameter -MMD -MP 102CFLAGS.clang := $(CFLAGS.gcc) 103CFLAGS.icc := $(CFLAGS.gcc) 104CFLAGS.XL := -qpic -MMD 105CXXFLAGS.gcc := -fPIC -std=c++11 -Wall -Wextra -Wno-unused-parameter -MMD -MP 106CXXFLAGS.clang := $(CXXFLAGS.gcc) 107CXXFLAGS.icc := $(CXXFLAGS.gcc) 108CXXFLAGS.XL := -qpic -std=c++11 -MMD 109FFLAGS.GNU := -fPIC -cpp -Wall -Wextra -Wno-unused-parameter -Wno-unused-dummy-argument -MMD -MP 110FFLAGS.ifort := -fPIC -cpp 111FFLAGS.XL := -qpic -ffree-form -qpreprocess -qextname -MMD 112 113# This check works with compilers that use gcc and clang. It fails with some 114# compilers; e.g., xlc apparently ignores all options when -E is passed, thus 115# succeeds with any flags. Users can pass MARCHFLAG=... if desired. 116cc_check_flag = $(shell $(CC) -E -Werror $(1) -x c /dev/null > /dev/null 2>&1 && echo 1) 117MARCHFLAG := $(MARCHFLAG.$(CC_VENDOR)) 118MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG),-mcpu=native) 119MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG)) 120 121OMP_SIMD_FLAG := $(OMP_SIMD_FLAG.$(CC_VENDOR)) 122OMP_SIMD_FLAG := $(if $(call cc_check_flag,$(OMP_SIMD_FLAG)),$(OMP_SIMD_FLAG)) 123 124OPT ?= -O -g $(MARCHFLAG) $(OPT.$(CC_VENDOR)) $(OMP_SIMD_FLAG) 125CFLAGS ?= $(OPT) $(CFLAGS.$(CC_VENDOR)) 126CXXFLAGS ?= $(OPT) $(CXXFLAGS.$(CC_VENDOR)) 127LIBCXX ?= -lstdc++ 128NVCCFLAGS ?= -ccbin $(CXX) -Xcompiler "$(OPT)" -Xcompiler -fPIC 129ifneq ($(CUDA_ARCH),) 130 NVCCFLAGS += -arch=$(CUDA_ARCH) 131endif 132HIPCCFLAGS ?= $(filter-out $(OMP_SIMD_FLAG),$(OPT)) -fPIC -munsafe-fp-atomics 133ifneq ($(HIP_ARCH),) 134 HIPCCFLAGS += --amdgpu-target=$(HIP_ARCH) 135endif 136FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR)) 137 138ifeq ($(COVERAGE), 1) 139 CFLAGS += --coverage 140 CXXFLAGS += --coverage 141 override LDFLAGS += --coverage 142endif 143 144CFLAGS += $(if $(ASAN),$(AFLAGS)) 145FFLAGS += $(if $(ASAN),$(AFLAGS)) 146override LDFLAGS += $(if $(ASAN),$(AFLAGS)) 147CPPFLAGS += -I./include 148LDLIBS = -lm 149OBJDIR := build 150LIBDIR := lib 151 152# Installation variables 153prefix ?= /usr/local 154bindir = $(prefix)/bin 155libdir = $(prefix)/lib 156includedir = $(prefix)/include 157pkgconfigdir = $(libdir)/pkgconfig 158INSTALL = install 159INSTALL_PROGRAM = $(INSTALL) 160INSTALL_DATA = $(INSTALL) -m644 161 162# Get number of processors of the machine 163NPROCS := $(shell getconf _NPROCESSORS_ONLN) 164# prepare make options to run in parallel 165MFLAGS := -j $(NPROCS) --warn-undefined-variables \ 166 --no-print-directory --no-keep-going 167 168PYTHON ?= python3 169PROVE ?= prove 170PROVE_OPTS ?= -j $(NPROCS) 171DARWIN := $(filter Darwin,$(shell uname -s)) 172SO_EXT := $(if $(DARWIN),dylib,so) 173 174ceed.pc := $(LIBDIR)/pkgconfig/ceed.pc 175libceed.so := $(LIBDIR)/libceed.$(SO_EXT) 176libceed.a := $(LIBDIR)/libceed.a 177libceed := $(if $(STATIC),$(libceed.a),$(libceed.so)) 178CEED_LIBS = -lceed 179libceed.c := $(filter-out interface/ceed-cuda.c interface/ceed-hip.c, $(wildcard interface/ceed*.c backends/*.c gallery/*.c)) 180gallery.c := $(wildcard gallery/*/ceed*.c) 181libceed.c += $(gallery.c) 182libceeds = $(libceed) 183BACKENDS_BUILTIN := /cpu/self/ref/serial /cpu/self/ref/blocked /cpu/self/opt/serial /cpu/self/opt/blocked 184BACKENDS_MAKE := $(BACKENDS_BUILTIN) 185TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub 186 187# Tests 188tests.c := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.c)) 189tests.f := $(if $(FC),$(sort $(wildcard tests/t[0-9][0-9][0-9]-*.f90))) 190tests := $(tests.c:tests/%.c=$(OBJDIR)/%) 191ctests := $(tests) 192tests += $(tests.f:tests/%.f90=$(OBJDIR)/%) 193# Examples 194examples.c := $(sort $(wildcard examples/ceed/*.c)) 195examples.f := $(if $(FC),$(sort $(wildcard examples/ceed/*.f))) 196examples := $(examples.c:examples/ceed/%.c=$(OBJDIR)/%) 197examples += $(examples.f:examples/ceed/%.f=$(OBJDIR)/%) 198# MFEM Examples 199mfemexamples.cpp := $(sort $(wildcard examples/mfem/*.cpp)) 200mfemexamples := $(mfemexamples.cpp:examples/mfem/%.cpp=$(OBJDIR)/mfem-%) 201# Nek5K Examples 202nekexamples := $(OBJDIR)/nek-bps 203# PETSc Examples 204petscexamples.c := $(wildcard examples/petsc/*.c) 205petscexamples := $(petscexamples.c:examples/petsc/%.c=$(OBJDIR)/petsc-%) 206# Fluid Dynamics Examples 207fluidsexamples.c := $(sort $(wildcard examples/fluids/*.c)) 208fluidsexamples := $(fluidsexamples.c:examples/fluids/%.c=$(OBJDIR)/fluids-%) 209# Solid Mechanics Examples 210solidsexamples.c := $(sort $(wildcard examples/solids/*.c)) 211solidsexamples := $(solidsexamples.c:examples/solids/%.c=$(OBJDIR)/solids-%) 212 213# Backends/[ref, blocked, memcheck, opt, avx, occa, magma] 214ref.c := $(sort $(wildcard backends/ref/*.c)) 215blocked.c := $(sort $(wildcard backends/blocked/*.c)) 216ceedmemcheck.c := $(sort $(wildcard backends/memcheck/*.c)) 217opt.c := $(sort $(wildcard backends/opt/*.c)) 218avx.c := $(sort $(wildcard backends/avx/*.c)) 219xsmm.c := $(sort $(wildcard backends/xsmm/*.c)) 220cuda.c := $(sort $(wildcard backends/cuda/*.c)) 221cuda-ref.c := $(sort $(wildcard backends/cuda-ref/*.c)) 222cuda-ref.cpp := $(sort $(wildcard backends/cuda-ref/*.cpp)) 223cuda-ref.cu := $(sort $(wildcard backends/cuda-ref/kernels/*.cu)) 224cuda-shared.c := $(sort $(wildcard backends/cuda-shared/*.c)) 225cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/kernels/*.cu)) 226cuda-gen.c := $(sort $(wildcard backends/cuda-gen/*.c)) 227cuda-gen.cpp := $(sort $(wildcard backends/cuda-gen/*.cpp)) 228cuda-gen.cu := $(sort $(wildcard backends/cuda-gen/kernels/*.cu)) 229occa.cpp := $(sort $(shell find backends/occa -type f -name *.cpp)) 230magma.c := $(sort $(wildcard backends/magma/*.c)) 231magma.cu := $(sort $(wildcard backends/magma/kernels/cuda/*.cu)) 232magma.hip := $(sort $(wildcard backends/magma/kernels/hip/*.hip.cpp)) 233hip.c := $(sort $(wildcard backends/hip/*.c)) 234hip.cpp := $(sort $(wildcard backends/hip/*.cpp)) 235hip-ref.c := $(sort $(wildcard backends/hip-ref/*.c)) 236hip-ref.cpp := $(sort $(wildcard backends/hip-ref/*.cpp)) 237hip-ref.hip := $(sort $(wildcard backends/hip-ref/kernels/*.hip.cpp)) 238hip-shared.c := $(sort $(wildcard backends/hip-shared/*.c)) 239hip-gen.c := $(sort $(wildcard backends/hip-gen/*.c)) 240hip-gen.cpp := $(sort $(wildcard backends/hip-gen/*.cpp)) 241 242# Quiet, color output 243quiet ?= $($(1)) 244 245# Cancel built-in and old-fashioned implicit rules which we don't use 246.SUFFIXES: 247 248.SECONDEXPANSION: # to expand $$(@D)/.DIR 249 250%/.DIR : 251 @mkdir -p $(@D) 252 @touch $@ 253 254.PRECIOUS: %/.DIR 255 256lib: $(libceed) $(ceed.pc) 257# run 'lib' target in parallel 258par:;@$(MAKE) $(MFLAGS) V=$(V) lib 259backend_status = $(if $(filter $1,$(BACKENDS_MAKE)), [backends: $1], [not found]) 260info: 261 $(info ------------------------------------) 262 $(info CC = $(CC)) 263 $(info CXX = $(CXX)) 264 $(info FC = $(FC)) 265 $(info CPPFLAGS = $(CPPFLAGS)) 266 $(info CFLAGS = $(value CFLAGS)) 267 $(info CXXFLAGS = $(value CXXFLAGS)) 268 $(info FFLAGS = $(value FFLAGS)) 269 $(info NVCCFLAGS = $(value NVCCFLAGS)) 270 $(info HIPCCFLAGS = $(value HIPCCFLAGS)) 271 $(info LDFLAGS = $(value LDFLAGS)) 272 $(info LDLIBS = $(LDLIBS)) 273 $(info AR = $(AR)) 274 $(info ARFLAGS = $(ARFLAGS)) 275 $(info OPT = $(OPT)) 276 $(info AFLAGS = $(AFLAGS)) 277 $(info ASAN = $(or $(ASAN),(empty))) 278 $(info VERBOSE = $(or $(V),(empty)) [verbose=$(if $(V),on,off)]) 279 $(info ------------------------------------) 280 $(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,$(MEMCHK_BACKENDS))) 281 $(info AVX_STATUS = $(AVX_STATUS)$(call backend_status,$(AVX_BACKENDS))) 282 $(info XSMM_DIR = $(XSMM_DIR)$(call backend_status,$(XSMM_BACKENDS))) 283 $(info OCCA_DIR = $(OCCA_DIR)$(call backend_status,$(OCCA_BACKENDS))) 284 $(info MAGMA_DIR = $(MAGMA_DIR)$(call backend_status,$(MAGMA_BACKENDS))) 285 $(info CUDA_DIR = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS))) 286 $(info HIP_DIR = $(HIP_DIR)$(call backend_status,$(HIP_BACKENDS))) 287 $(info ------------------------------------) 288 $(info MFEM_DIR = $(MFEM_DIR)) 289 $(info NEK5K_DIR = $(NEK5K_DIR)) 290 $(info PETSC_DIR = $(PETSC_DIR)) 291 $(info ------------------------------------) 292 $(info prefix = $(prefix)) 293 $(info includedir = $(value includedir)) 294 $(info libdir = $(value libdir)) 295 $(info pkgconfigdir = $(value pkgconfigdir)) 296 $(info ------------------------------------) 297 @true 298info-backends: 299 $(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS))) 300 @true 301info-backends-all: 302 $(info make: 'lib' with backends: $(filter-out $(TEST_BACKENDS),$(BACKENDS))) 303 @true 304 305$(libceed.so) : override LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed.so))) 306 307# Standard Backends 308libceed.c += $(ref.c) 309libceed.c += $(blocked.c) 310libceed.c += $(opt.c) 311 312# Memcheck Backend 313MEMCHK_STATUS = Disabled 314MEMCHK := $(shell echo "$(HASH)include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1) 315MEMCHK_BACKENDS = /cpu/self/memcheck/serial /cpu/self/memcheck/blocked 316ifeq ($(MEMCHK),1) 317 MEMCHK_STATUS = Enabled 318 libceed.c += $(ceedmemcheck.c) 319 BACKENDS_MAKE += $(MEMCHK_BACKENDS) 320endif 321 322# AVX Backed 323AVX_STATUS = Disabled 324AVX_FLAG := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx) 325AVX := $(filter $(AVX_FLAG),$(shell $(CC) $(CFLAGS) -v -E -x c /dev/null 2>&1)) 326AVX_BACKENDS = /cpu/self/avx/serial /cpu/self/avx/blocked 327ifneq ($(AVX),) 328 AVX_STATUS = Enabled 329 libceed.c += $(avx.c) 330 BACKENDS_MAKE += $(AVX_BACKENDS) 331endif 332 333# Collect list of libraries and paths for use in linking and pkg-config 334PKG_LIBS = 335# Stubs that will not be RPATH'd 336PKG_STUBS_LIBS = 337 338# libXSMM Backends 339XSMM_BACKENDS = /cpu/self/xsmm/serial /cpu/self/xsmm/blocked 340ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),) 341 PKG_LIBS += -L$(abspath $(XSMM_DIR))/lib -lxsmm -ldl 342 MKL ?= 343 ifeq (,$(MKL)$(MKLROOT)) 344 BLAS_LIB = -lblas 345 else 346 ifneq ($(MKLROOT),) 347 # Some installs put everything inside an intel64 subdirectory, others not 348 MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*))) 349 MKL_LINK = -L$(MKL_LIBDIR) 350 PKG_LIB_DIRS += $(MKL_LIBDIR) 351 endif 352 BLAS_LIB = $(MKL_LINK) -Wl,--push-state,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl -Wl,--pop-state 353 endif 354 PKG_LIBS += $(BLAS_LIB) 355 libceed.c += $(xsmm.c) 356 $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include 357 BACKENDS_MAKE += $(XSMM_BACKENDS) 358endif 359 360# OCCA Backends 361OCCA_BACKENDS = /cpu/self/occa 362ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),) 363 OCCA_MODES := $(shell $(OCCA_DIR)/bin/occa modes) 364 OCCA_BACKENDS += $(if $(filter OpenMP,$(OCCA_MODES)),/cpu/openmp/occa) 365# OCCA_BACKENDS += $(if $(filter OpenCL,$(OCCA_MODES)),/gpu/opencl/occa) 366 OCCA_BACKENDS += $(if $(filter HIP,$(OCCA_MODES)),/gpu/hip/occa) 367 OCCA_BACKENDS += $(if $(filter CUDA,$(OCCA_MODES)),/gpu/cuda/occa) 368 369 $(libceeds) : CPPFLAGS += -I$(OCCA_DIR)/include 370 PKG_LIBS += -L$(abspath $(OCCA_DIR))/lib -locca 371 LIBCEED_CONTAINS_CXX = 1 372 libceed.cpp += $(occa.cpp) 373 BACKENDS_MAKE += $(OCCA_BACKENDS) 374endif 375 376# CUDA Backends 377CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(CUDA_DIR)/$d/libcudart.${SO_EXT})) 378CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR)))) 379CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs 380CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/shared /gpu/cuda/gen 381ifneq ($(CUDA_LIB_DIR),) 382 $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include 383 PKG_LIBS += -L$(abspath $(CUDA_LIB_DIR)) -lcudart -lnvrtc -lcuda -lcublas 384 PKG_STUBS_LIBS += -L$(CUDA_LIB_DIR_STUBS) 385 LIBCEED_CONTAINS_CXX = 1 386 libceed.c += interface/ceed-cuda.c 387 libceed.c += $(cuda.c) $(cuda-ref.c) $(cuda-shared.c) $(cuda-gen.c) 388 libceed.cpp += $(cuda-ref.cpp) $(cuda-gen.cpp) 389 libceed.cu += $(cuda-ref.cu) $(cuda-shared.cu) $(cuda-gen.cu) 390 BACKENDS_MAKE += $(CUDA_BACKENDS) 391endif 392 393# HIP Backends 394HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libamdhip64.${SO_EXT})) 395HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 396HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen 397ifneq ($(HIP_LIB_DIR),) 398 $(libceeds) : HIPCCFLAGS += -I./include 399 ifneq ($(CXX), $(HIPCC)) 400 CPPFLAGS += $(subst =,,$(shell $(HIP_DIR)/bin/hipconfig -C)) 401 endif 402 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include 403 PKG_LIBS += -L$(abspath $(HIP_LIB_DIR)) -lamdhip64 -lhipblas 404 LIBCEED_CONTAINS_CXX = 1 405 libceed.c += interface/ceed-hip.c 406 libceed.c += $(hip.c) $(hip-ref.c) $(hip-shared.c) $(hip-gen.c) 407 libceed.cpp += $(hip.cpp) $(hip-ref.cpp) $(hip-gen.cpp) 408 libceed.hip += $(hip-ref.hip) 409 BACKENDS_MAKE += $(HIP_BACKENDS) 410endif 411 412# MAGMA Backend 413ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 414 MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas") 415 ifeq ($(MAGMA_ARCH), 0) #CUDA MAGMA 416 ifneq ($(CUDA_LIB_DIR),) 417 cuda_link = -Wl,-rpath,$(CUDA_LIB_DIR) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart 418 omp_link = -fopenmp 419 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link) 420 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 421 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 422 PKG_LIBS += $(magma_link) 423 libceed.c += $(magma.c) 424 libceed.cu += $(magma.cu) 425 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 426 $(magma.cu:%.cu=$(OBJDIR)/%.o) : CPPFLAGS += --compiler-options=-fPIC -DADD_ -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(CUDA_DIR)/include 427 MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det 428 endif 429 else # HIP MAGMA 430 ifneq ($(HIP_LIB_DIR),) 431 hip_link = -Wl,-rpath,$(HIP_LIB_DIR) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64 432 omp_link = -fopenmp 433 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link) 434 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 435 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 436 PKG_LIBS += $(magma_link) 437 libceed.c += $(magma.c) 438 libceed.hip += $(magma.hip) 439 ifneq ($(CXX), $(HIPCC)) 440 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 441 else 442 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 443 endif 444 $(magma.hip:%.hip.cpp=$(OBJDIR)/%.o) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 445 MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det 446 endif 447 endif 448 LIBCEED_CONTAINS_CXX = 1 449 BACKENDS_MAKE += $(MAGMA_BACKENDS) 450endif 451 452BACKENDS ?= $(BACKENDS_MAKE) 453export BACKENDS 454 455_pkg_ldflags = $(filter -L%,$(PKG_LIBS)) 456_pkg_ldlibs = $(filter-out -L%,$(PKG_LIBS)) 457$(libceeds) : override LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 458$(libceeds) : LDLIBS += $(_pkg_ldlibs) 459ifeq ($(STATIC),1) 460$(examples) $(tests) : override LDFLAGS += $(_pkg_ldflags) $(_pkg_ldflags:-L%=-Wl,-rpath,%) $(PKG_STUBS_LIBS) 461$(examples) $(tests) : LDLIBS += $(_pkg_ldlibs) 462endif 463 464pkgconfig-libs-private = $(PKG_LIBS) 465ifeq ($(LIBCEED_CONTAINS_CXX),1) 466 $(libceeds) : LINK = $(CXX) 467 ifeq ($(STATIC),1) 468 $(examples) $(tests) : LDLIBS += $(LIBCXX) 469 pkgconfig-libs-private += $(LIBCXX) 470 endif 471endif 472 473# File names *-weak.c contain weak symbol definitions, which must be listed last 474# when creating shared or static libraries. 475weak_last = $(filter-out %-weak.o,$(1)) $(filter %-weak.o,$(1)) 476 477libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) 478$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 479$(libceed.o): | info-backends 480$(libceed.so) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 481 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 482 483$(libceed.a) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 484 $(call quiet,AR) $(ARFLAGS) $@ $^ 485 486$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 487 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 488 489$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 490 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 491 492$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 493 $(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<) 494 495$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR 496 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 497 498$(OBJDIR)/% : tests/%.c | $$(@D)/.DIR 499 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 500 501$(OBJDIR)/% : tests/%.f90 | $$(@D)/.DIR 502 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 503 504$(OBJDIR)/% : examples/ceed/%.c | $$(@D)/.DIR 505 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 506 507$(OBJDIR)/% : examples/ceed/%.f | $$(@D)/.DIR 508 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 509 510$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 511 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 512 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 513 cp examples/mfem/$* $@ 514 515# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 516# this single Nek bps file to be built in parallel with other examples, 517# such as when calling `make prove-all -j2`. 518$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 519 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 520 mv examples/nek/build/bps $(OBJDIR)/bps 521 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 522 523# Several executables have common utilities, but we can't build the utilities 524# from separate submake invocations because they'll compete with each 525# other/corrupt output. So we put it in this utility library, but we don't want 526# to manually list source dependencies up at this level, so we'll just always 527# call recursive make to check that this utility is up to date. 528examples/petsc/libutils.a.PHONY: $(libceed) $(ceed.pc) 529 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` AR=$(AR) ARFLAGS=$(ARFLAGS) \ 530 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $(basename $(@F)) 531 532$(OBJDIR)/petsc-% : examples/petsc/%.c examples/petsc/libutils.a.PHONY $(libceed) $(ceed.pc) | $$(@D)/.DIR 533 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` \ 534 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 535 cp examples/petsc/$* $@ 536 537$(OBJDIR)/fluids-% : examples/fluids/%.c examples/fluids/src/*.c examples/fluids/*.h examples/fluids/problems/*.c examples/fluids/qfunctions/*.h $(libceed) $(ceed.pc) | $$(@D)/.DIR 538 +$(call quiet,MAKE) -C examples/fluids CEED_DIR=`pwd` \ 539 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 540 cp examples/fluids/$* $@ 541 542$(OBJDIR)/solids-% : examples/solids/%.c examples/solids/%.h \ 543 examples/solids/problems/*.c examples/solids/src/*.c \ 544 examples/solids/include/*.h examples/solids/problems/*.h examples/solids/qfunctions/*.h \ 545 $(libceed) $(ceed.pc) | $$(@D)/.DIR 546 +$(call quiet,MAKE) -C examples/solids CEED_DIR=`pwd` \ 547 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 548 cp examples/solids/$* $@ 549 550$(examples) : $(libceed) 551$(tests) : $(libceed) 552$(tests) $(examples) : override LDFLAGS += -Wl,-rpath,$(abspath $(LIBDIR)) -L$(LIBDIR) 553 554run-% : $(OBJDIR)/% 555 @tests/tap.sh $(<:$(OBJDIR)/%=%) 556 557external_examples := \ 558 $(if $(MFEM_DIR),$(mfemexamples)) \ 559 $(if $(PETSC_DIR),$(petscexamples)) \ 560 $(if $(NEK5K_DIR),$(nekexamples)) \ 561 $(if $(PETSC_DIR),$(fluidsexamples)) \ 562 $(if $(PETSC_DIR),$(solidsexamples)) 563 564allexamples = $(examples) $(external_examples) 565 566# The test and prove targets can be controlled via pattern searches. The 567# default is to run tests and those examples that have no external dependencies. 568# Examples of finer grained control: 569# 570# make test search='petsc mfem' # PETSc and MFEM examples 571# make prove search='t3' # t3xx series tests 572# make junit search='ex petsc' # core ex* and PETSc tests 573search ?= t ex 574realsearch = $(search:%=%%) 575matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 576 577# Test core libCEED 578test : $(matched:$(OBJDIR)/%=run-%) 579 580# Run test target in parallel 581tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 582# CPU C tests only for backend % 583ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 584 585prove : $(matched) 586 $(info Testing backends: $(BACKENDS)) 587 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 588# Run prove target in parallel 589prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 590 591prove-all : 592 +$(MAKE) prove realsearch=% 593 594junit-% : $(OBJDIR)/% 595 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 596 597junit : $(matched:$(OBJDIR)/%=junit-%) 598 599all: $(alltests) 600 601examples : $(allexamples) 602ceedexamples : $(examples) 603nekexamples : $(nekexamples) 604mfemexamples : $(mfemexamples) 605petscexamples : $(petscexamples) 606 607# Benchmarks 608allbenchmarks = petsc-bps 609bench_targets = $(addprefix bench-,$(allbenchmarks)) 610.PHONY: $(bench_targets) benchmarks 611$(bench_targets): bench-%: $(OBJDIR)/% 612 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS_MAKE)" -r $(*).sh 613benchmarks: $(bench_targets) 614 615$(ceed.pc) : pkgconfig-prefix = $(abspath .) 616$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 617.INTERMEDIATE : $(OBJDIR)/ceed.pc 618%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 619 @$(SED) \ 620 -e "s:%prefix%:$(pkgconfig-prefix):" \ 621 -e "s:%libs_private%:$(pkgconfig-libs-private):" $< > $@ 622 623install : $(libceed) $(OBJDIR)/ceed.pc 624 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 625 "$(includedir)/ceed/" "$(libdir)" "$(pkgconfigdir)") 626 $(INSTALL_DATA) include/ceed/ceed.h "$(DESTDIR)$(includedir)/ceed/" 627 $(INSTALL_DATA) include/ceed/ceed-f32.h "$(DESTDIR)$(includedir)/ceed/" 628 $(INSTALL_DATA) include/ceed/ceed-f64.h "$(DESTDIR)$(includedir)/ceed/" 629 $(INSTALL_DATA) include/ceed/fortran.h "$(DESTDIR)$(includedir)/ceed/" 630 $(INSTALL_DATA) include/ceed/backend.h "$(DESTDIR)$(includedir)/ceed/" 631 $(INSTALL_DATA) include/ceed/cuda.h "$(DESTDIR)$(includedir)/ceed/" 632 $(INSTALL_DATA) include/ceed/hip.h "$(DESTDIR)$(includedir)/ceed/" 633 $(INSTALL_DATA) include/ceed/hash.h "$(DESTDIR)$(includedir)/ceed/" 634 $(INSTALL_DATA) include/ceed/khash.h "$(DESTDIR)$(includedir)/ceed/" 635 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 636 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 637 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 638 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 639 640.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 641 642cln clean : 643 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 644 $(call quiet,MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 645 $(call quiet,MAKE) -C python/tests clean 646 $(RM) benchmarks/*output.txt 647 648distclean : clean 649 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 650 651DOXYGEN ?= doxygen 652doxygen : 653 $(DOXYGEN) -q Doxyfile 654 655doc-html doc-latexpdf doc-epub doc-livehtml : doc-% : doxygen 656 make -C doc/sphinx $* 657 658doc : doc-html 659 660style-c : 661 @astyle --options=.astylerc \ 662 $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), \ 663 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 664 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 665 666AUTOPEP8 = autopep8 667style-py : AUTOPEP8_ARGS = --in-place --aggressive 668style-py : 669 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py python/tests/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 670 671style : style-c style-py 672 673CLANG_TIDY ?= clang-tidy 674 675%.c.tidy : %.c 676 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include -I$(HIP_DIR)/include 677 678%.cpp.tidy : %.cpp 679 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include -I$(HIP_DIR)/include 680 681tidy_c : $(libceed.c:%=%.tidy) 682tidy_cpp : $(libceed.cpp:%=%.tidy) 683 684tidy : tidy_c tidy_cpp 685 686ifneq ($(wildcard ../iwyu/*),) 687 IWYU_DIR ?= ../iwyu 688 IWYU_CC ?= $(IWYU_DIR)/build/bin/include-what-you-use 689endif 690 691iwyu : CC=$(IWYU_CC) 692iwyu : lib 693 694print : 695 @echo $(VAR)=$($(VAR)) 696 697print-% : 698 $(info [ variable name]: $*) 699 $(info [ origin]: $(origin $*)) 700 $(info [ flavor]: $(flavor $*)) 701 $(info [ value]: $(value $*)) 702 $(info [expanded value]: $($*)) 703 $(info ) 704 @true 705 706# "make configure" detects any variables passed on the command line or 707# previously set in config.mk, caching them in config.mk as simple 708# (:=) variables. Variables set in config.mk or on the command line 709# take precedence over the defaults provided in the file. Typical 710# usage: 711# 712# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 713# make 714# make prove 715# 716# The values in the file can be updated by passing them on the command 717# line, e.g., 718# 719# make configure CC=/path/to/other/clang 720 721# All variables to consider for caching 722CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 723 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 724 AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \ 725 MAGMA_DIR OCCA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR HIP_ARCH 726 727# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 728# was set on the command line or in config.mk (where it will appear as 729# a simple variable). 730needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 731 732configure : 733 $(file > $(CONFIG)) 734 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 735 @echo "Configuration cached in $(CONFIG):" 736 @cat $(CONFIG) 737 738wheel : export MARCHFLAG = -march=generic 739wheel : export WHEEL_PLAT = manylinux2010_x86_64 740wheel : 741 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 742 -e MARCHFLAG -e WHEEL_PLAT \ 743 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 744 745.PHONY : configure wheel 746 747# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 748-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 749