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