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 35EXE_SUFFIX = $(if $(EMSCRIPTEN),.wasm) 36 37# ASAN must be left empty if you don't want to use it 38ASAN ?= 39 40# These are the values automatically detected here in the makefile. They are 41# augmented with LDFLAGS and LDLIBS from the environment/passed by command line, 42# if any. If the user sets CEED_LDFLAGS or CEED_LDLIBS, they are used *instead 43# of* what we populate here (thus that's advanced usage and not recommended). 44CEED_LDFLAGS ?= 45CEED_LDLIBS ?= 46 47UNDERSCORE ?= 1 48 49# Verbose mode, V or VERBOSE 50V ?= $(VERBOSE) 51 52# MFEM_DIR env variable should point to sibling directory 53ifneq ($(wildcard ../mfem/libmfem.*),) 54 MFEM_DIR ?= ../mfem 55endif 56 57# NEK5K_DIR env variable should point to sibling directory 58ifneq ($(wildcard ../Nek5000/*),) 59 NEK5K_DIR ?= $(abspath ../Nek5000) 60endif 61export NEK5K_DIR 62MPI ?= 1 63 64# CEED_DIR env for NEK5K testing 65export CEED_DIR = $(abspath .) 66 67# XSMM_DIR env variable should point to XSMM master (github.com/hfp/libxsmm) 68XSMM_DIR ?= ../libxsmm 69 70# OCCA_DIR env variable should point to OCCA master (github.com/libocca/occa) 71OCCA_DIR ?= ../occa 72 73# env variable MAGMA_DIR can be used too 74MAGMA_DIR ?= ../magma 75 76# Often /opt/cuda or /usr/local/cuda, but sometimes present on machines that don't support CUDA 77CUDA_DIR ?= 78CUDA_ARCH ?= 79 80# Often /opt/rocm, but sometimes present on machines that don't support HIP 81HIP_DIR ?= 82HIP_ARCH ?= 83 84# Check for PETSc in ../petsc 85ifneq ($(wildcard ../petsc/lib/libpetsc.*),) 86 PETSC_DIR ?= ../petsc 87endif 88 89# Warning: SANTIZ options still don't run with /gpu/occa 90# export LSAN_OPTIONS=suppressions=.asanignore 91AFLAGS = -fsanitize=address #-fsanitize=undefined -fno-omit-frame-pointer 92 93# Note: Intel oneAPI C/C++ compiler is now icx/icpx 94CC_VENDOR := $(subst icc_orig,icc,$(subst oneAPI,icc,$(firstword $(filter gcc clang icc icc_orig oneAPI XL,$(subst -, ,$(shell $(CC) --version)))))) 95FC_VENDOR := $(if $(FC),$(firstword $(filter GNU ifort XL,$(shell $(FC) --version 2>&1 || $(FC) -qversion)))) 96 97# Default extra flags by vendor 98MARCHFLAG.gcc := -march=native 99MARCHFLAG.clang := $(MARCHFLAG.gcc) 100MARCHFLAG.icc := 101OMP_SIMD_FLAG.gcc := -fopenmp-simd 102OMP_SIMD_FLAG.clang := $(OMP_SIMD_FLAG.gcc) 103OMP_SIMD_FLAG.icc := -qopenmp-simd 104OPT.gcc := -ffp-contract=fast 105OPT.clang := $(OPT.gcc) 106CFLAGS.gcc := -fPIC -std=c99 -Wall -Wextra -Wno-unused-parameter -MMD -MP 107CFLAGS.clang := $(CFLAGS.gcc) 108CFLAGS.icc := $(CFLAGS.gcc) 109CFLAGS.XL := -qpic -MMD 110CXXFLAGS.gcc := -fPIC -std=c++11 -Wall -Wextra -Wno-unused-parameter -MMD -MP 111CXXFLAGS.clang := $(CXXFLAGS.gcc) 112CXXFLAGS.icc := $(CXXFLAGS.gcc) 113CXXFLAGS.XL := -qpic -std=c++11 -MMD 114FFLAGS.GNU := -fPIC -cpp -Wall -Wextra -Wno-unused-parameter -Wno-unused-dummy-argument -MMD -MP 115FFLAGS.ifort := -fPIC -cpp 116FFLAGS.XL := -qpic -ffree-form -qpreprocess -qextname -MMD 117 118# This check works with compilers that use gcc and clang. It fails with some 119# compilers; e.g., xlc apparently ignores all options when -E is passed, thus 120# succeeds with any flags. Users can pass MARCHFLAG=... if desired. 121cc_check_flag = $(shell $(CC) -E -Werror $(1) -x c /dev/null > /dev/null 2>&1 && echo 1) 122MARCHFLAG := $(MARCHFLAG.$(CC_VENDOR)) 123MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG),-mcpu=native) 124MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG)) 125 126OMP_SIMD_FLAG := $(OMP_SIMD_FLAG.$(CC_VENDOR)) 127OMP_SIMD_FLAG := $(if $(call cc_check_flag,$(OMP_SIMD_FLAG)),$(OMP_SIMD_FLAG)) 128 129OPT ?= -O -g $(MARCHFLAG) $(OPT.$(CC_VENDOR)) $(OMP_SIMD_FLAG) 130CFLAGS ?= $(OPT) $(CFLAGS.$(CC_VENDOR)) 131CXXFLAGS ?= $(OPT) $(CXXFLAGS.$(CC_VENDOR)) 132LIBCXX ?= -lstdc++ 133NVCCFLAGS ?= -ccbin $(CXX) -Xcompiler "$(OPT)" -Xcompiler -fPIC 134ifneq ($(CUDA_ARCH),) 135 NVCCFLAGS += -arch=$(CUDA_ARCH) 136endif 137HIPCCFLAGS ?= $(filter-out $(OMP_SIMD_FLAG),$(OPT)) -fPIC -munsafe-fp-atomics 138ifneq ($(HIP_ARCH),) 139 HIPCCFLAGS += --amdgpu-target=$(HIP_ARCH) 140endif 141FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR)) 142 143ifeq ($(COVERAGE), 1) 144 CFLAGS += --coverage 145 CXXFLAGS += --coverage 146 CEED_LDFLAGS += --coverage 147endif 148 149CFLAGS += $(if $(ASAN),$(AFLAGS)) 150FFLAGS += $(if $(ASAN),$(AFLAGS)) 151CEED_LDFLAGS += $(if $(ASAN),$(AFLAGS)) 152CPPFLAGS += -I./include 153CEED_LDLIBS = -lm 154OBJDIR := build 155for_install := $(filter install,$(MAKECMDGOALS)) 156LIBDIR := $(if $(for_install),$(OBJDIR),lib) 157 158 159# Installation variables 160prefix ?= /usr/local 161bindir = $(prefix)/bin 162libdir = $(prefix)/lib 163includedir = $(prefix)/include 164pkgconfigdir = $(libdir)/pkgconfig 165INSTALL = install 166INSTALL_PROGRAM = $(INSTALL) 167INSTALL_DATA = $(INSTALL) -m644 168 169# Get number of processors of the machine 170NPROCS := $(shell getconf _NPROCESSORS_ONLN) 171# prepare make options to run in parallel 172MFLAGS := -j $(NPROCS) --warn-undefined-variables \ 173 --no-print-directory --no-keep-going 174 175PYTHON ?= python3 176PROVE ?= prove 177PROVE_OPTS ?= -j $(NPROCS) 178DARWIN := $(filter Darwin,$(shell uname -s)) 179SO_EXT := $(if $(DARWIN),dylib,so) 180 181ceed.pc := $(LIBDIR)/pkgconfig/ceed.pc 182libceed.so := $(LIBDIR)/libceed.$(SO_EXT) 183libceed.a := $(LIBDIR)/libceed.a 184libceed := $(if $(STATIC),$(libceed.a),$(libceed.so)) 185CEED_LIBS = -lceed 186libceed.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)) 187gallery.c := $(wildcard gallery/*/ceed*.c) 188libceed.c += $(gallery.c) 189libceeds = $(libceed) 190BACKENDS_BUILTIN := /cpu/self/ref/serial /cpu/self/ref/blocked /cpu/self/opt/serial /cpu/self/opt/blocked 191BACKENDS_MAKE := $(BACKENDS_BUILTIN) 192TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub 193 194# Tests 195tests.c := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.c)) 196tests.f := $(if $(FC),$(sort $(wildcard tests/t[0-9][0-9][0-9]-*.f90))) 197tests := $(tests.c:tests/%.c=$(OBJDIR)/%$(EXE_SUFFIX)) 198ctests := $(tests) 199tests += $(tests.f:tests/%.f90=$(OBJDIR)/%$(EXE_SUFFIX)) 200# Examples 201examples.c := $(sort $(wildcard examples/ceed/*.c)) 202examples.f := $(if $(FC),$(sort $(wildcard examples/ceed/*.f))) 203examples := $(examples.c:examples/ceed/%.c=$(OBJDIR)/%$(EXE_SUFFIX)) 204examples += $(examples.f:examples/ceed/%.f=$(OBJDIR)/%$(EXE_SUFFIX)) 205# MFEM Examples 206mfemexamples.cpp := $(sort $(wildcard examples/mfem/*.cpp)) 207mfemexamples := $(mfemexamples.cpp:examples/mfem/%.cpp=$(OBJDIR)/mfem-%) 208# Nek5K Examples 209nekexamples := $(OBJDIR)/nek-bps 210# PETSc Examples 211petscexamples.c := $(wildcard examples/petsc/*.c) 212petscexamples := $(petscexamples.c:examples/petsc/%.c=$(OBJDIR)/petsc-%) 213# Fluid Dynamics Examples 214fluidsexamples.c := $(sort $(wildcard examples/fluids/*.c)) 215fluidsexamples := $(fluidsexamples.c:examples/fluids/%.c=$(OBJDIR)/fluids-%) 216# Solid Mechanics Examples 217solidsexamples.c := $(sort $(wildcard examples/solids/*.c)) 218solidsexamples := $(solidsexamples.c:examples/solids/%.c=$(OBJDIR)/solids-%) 219 220# Backends/[ref, blocked, memcheck, opt, avx, occa, magma] 221ref.c := $(sort $(wildcard backends/ref/*.c)) 222blocked.c := $(sort $(wildcard backends/blocked/*.c)) 223ceedmemcheck.c := $(sort $(wildcard backends/memcheck/*.c)) 224opt.c := $(sort $(wildcard backends/opt/*.c)) 225avx.c := $(sort $(wildcard backends/avx/*.c)) 226xsmm.c := $(sort $(wildcard backends/xsmm/*.c)) 227cuda.c := $(sort $(wildcard backends/cuda/*.c)) 228cuda.cpp := $(sort $(wildcard backends/cuda/*.cpp)) 229cuda-ref.c := $(sort $(wildcard backends/cuda-ref/*.c)) 230cuda-ref.cpp := $(sort $(wildcard backends/cuda-ref/*.cpp)) 231cuda-ref.cu := $(sort $(wildcard backends/cuda-ref/kernels/*.cu)) 232cuda-shared.c := $(sort $(wildcard backends/cuda-shared/*.c)) 233cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/kernels/*.cu)) 234cuda-gen.c := $(sort $(wildcard backends/cuda-gen/*.c)) 235cuda-gen.cpp := $(sort $(wildcard backends/cuda-gen/*.cpp)) 236cuda-gen.cu := $(sort $(wildcard backends/cuda-gen/kernels/*.cu)) 237occa.cpp := $(sort $(shell find backends/occa -type f -name *.cpp)) 238magma.c := $(sort $(wildcard backends/magma/*.c)) 239magma.cpp := $(sort $(wildcard backends/magma/*.cpp)) 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 CEED_LDFLAGS = $(value CEED_LDFLAGS)) 281 $(info CEED_LDLIBS = $(value CEED_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) : CEED_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 386ifneq ($(CUDA_DIR),) 387 CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64 lib/x86_64-linux-gnu,$(CUDA_DIR)/$d/libcudart.${SO_EXT})) 388 CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR)))) 389endif 390CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs 391CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/shared /gpu/cuda/gen 392ifneq ($(CUDA_LIB_DIR),) 393 $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include 394 PKG_LIBS += -L$(abspath $(CUDA_LIB_DIR)) -lcudart -lnvrtc -lcuda -lcublas 395 PKG_STUBS_LIBS += -L$(CUDA_LIB_DIR_STUBS) 396 LIBCEED_CONTAINS_CXX = 1 397 libceed.c += interface/ceed-cuda.c 398 libceed.c += $(cuda.c) $(cuda-ref.c) $(cuda-shared.c) $(cuda-gen.c) 399 libceed.cpp += $(cuda.cpp) $(cuda-ref.cpp) $(cuda-gen.cpp) 400 libceed.cu += $(cuda-ref.cu) $(cuda-shared.cu) $(cuda-gen.cu) 401 BACKENDS_MAKE += $(CUDA_BACKENDS) 402endif 403 404# HIP Backends 405HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libamdhip64.${SO_EXT})) 406HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 407HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen 408ifneq ($(HIP_LIB_DIR),) 409 $(libceeds) : HIPCCFLAGS += -I./include 410 ifneq ($(CXX), $(HIPCC)) 411 CPPFLAGS += $(subst =,,$(shell $(HIP_DIR)/bin/hipconfig -C)) 412 endif 413 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include 414 PKG_LIBS += -L$(abspath $(HIP_LIB_DIR)) -lamdhip64 -lhipblas 415 LIBCEED_CONTAINS_CXX = 1 416 libceed.c += interface/ceed-hip.c 417 libceed.c += $(hip.c) $(hip-ref.c) $(hip-shared.c) $(hip-gen.c) 418 libceed.cpp += $(hip.cpp) $(hip-ref.cpp) $(hip-gen.cpp) 419 libceed.hip += $(hip-ref.hip) 420 BACKENDS_MAKE += $(HIP_BACKENDS) 421endif 422 423# MAGMA Backend 424ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 425 MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas") 426 ifeq ($(MAGMA_ARCH), 0) #CUDA MAGMA 427 ifneq ($(CUDA_LIB_DIR),) 428 cuda_link = $(if $(STATIC),,-Wl,-rpath,$(CUDA_LIB_DIR)) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart 429 omp_link = -fopenmp 430 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link) 431 magma_link_shared = -L$(MAGMA_DIR)/lib $(if $(STATIC),,-Wl,-rpath,$(abspath $(MAGMA_DIR)/lib)) -lmagma 432 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 433 PKG_LIBS += $(magma_link) 434 libceed.c += $(magma.c) 435 libceed.cpp += $(magma.cpp) 436 libceed.cu += $(magma.cu) 437 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 438 $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 439 $(magma.cu:%.cu=$(OBJDIR)/%.o) : CPPFLAGS += --compiler-options=-fPIC -DADD_ -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(CUDA_DIR)/include 440 MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det 441 endif 442 else # HIP MAGMA 443 ifneq ($(HIP_LIB_DIR),) 444 hip_link = $(if $(STATIC),,-Wl,-rpath,$(HIP_LIB_DIR)) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64 445 omp_link = -fopenmp 446 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link) 447 magma_link_shared = -L$(MAGMA_DIR)/lib $(if $(STATIC),,-Wl,-rpath,$(abspath $(MAGMA_DIR)/lib)) -lmagma 448 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 449 PKG_LIBS += $(magma_link) 450 libceed.c += $(magma.c) 451 libceed.cpp += $(magma.cpp) 452 libceed.hip += $(magma.hip) 453 ifneq ($(CXX), $(HIPCC)) 454 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 455 $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 456 else 457 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 458 $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 459 endif 460 $(magma.hip:%.hip.cpp=$(OBJDIR)/%.o) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(MAGMA_DIR)/magmablas -I$(HIP_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_ 461 MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det 462 endif 463 endif 464 LIBCEED_CONTAINS_CXX = 1 465 BACKENDS_MAKE += $(MAGMA_BACKENDS) 466endif 467 468BACKENDS ?= $(BACKENDS_MAKE) 469export BACKENDS 470 471_pkg_ldflags = $(filter -L%,$(PKG_LIBS)) 472_pkg_ldlibs = $(filter-out -L%,$(PKG_LIBS)) 473$(libceeds) : CEED_LDFLAGS += $(_pkg_ldflags) $(if $(STATIC),,$(_pkg_ldflags:-L%=-Wl,-rpath,%)) $(PKG_STUBS_LIBS) 474$(libceeds) : CEED_LDLIBS += $(_pkg_ldlibs) 475ifeq ($(STATIC),1) 476$(examples) $(tests) : CEED_LDFLAGS += $(_pkg_ldflags) $(if $(STATIC),,$(_pkg_ldflags:-L%=-Wl,-rpath,%)) $(PKG_STUBS_LIBS) 477$(examples) $(tests) : CEED_LDLIBS += $(_pkg_ldlibs) 478endif 479 480pkgconfig-libs-private = $(PKG_LIBS) 481ifeq ($(LIBCEED_CONTAINS_CXX),1) 482 $(libceeds) : LINK = $(CXX) 483 ifeq ($(STATIC),1) 484 $(examples) $(tests) : CEED_LDLIBS += $(LIBCXX) 485 pkgconfig-libs-private += $(LIBCXX) 486 endif 487endif 488 489# File names *-weak.c contain weak symbol definitions, which must be listed last 490# when creating shared or static libraries. 491weak_last = $(filter-out %-weak.o,$(1)) $(filter %-weak.o,$(1)) 492 493libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) 494$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 495$(libceed.o): | info-backends 496$(libceed.so) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 497 $(call quiet,LINK) $(LDFLAGS) $(CEED_LDFLAGS) -shared -o $@ $^ $(CEED_LDLIBS) $(LDLIBS) 498 499$(libceed.a) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR 500 $(call quiet,AR) $(ARFLAGS) $@ $^ 501 502$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 503 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 504 505$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 506 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 507 508$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 509 $(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<) 510 511$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR 512 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 513 514$(OBJDIR)/%$(EXE_SUFFIX) : tests/%.c | $$(@D)/.DIR 515 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 516 517$(OBJDIR)/%$(EXE_SUFFIX) : tests/%.f90 | $$(@D)/.DIR 518 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 519 520$(OBJDIR)/%$(EXE_SUFFIX) : examples/ceed/%.c | $$(@D)/.DIR 521 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 522 523$(OBJDIR)/%$(EXE_SUFFIX) : examples/ceed/%.f | $$(@D)/.DIR 524 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS) 525 526$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 527 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 528 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 529 cp examples/mfem/$* $@ 530 531# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 532# this single Nek bps file to be built in parallel with other examples, 533# such as when calling `make prove-all -j2`. 534$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 535 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 536 mv examples/nek/build/bps $(OBJDIR)/bps 537 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 538 539# Several executables have common utilities, but we can't build the utilities 540# from separate submake invocations because they'll compete with each 541# other/corrupt output. So we put it in this utility library, but we don't want 542# to manually list source dependencies up at this level, so we'll just always 543# call recursive make to check that this utility is up to date. 544examples/petsc/libutils.a.PHONY: $(libceed) $(ceed.pc) 545 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` AR=$(AR) ARFLAGS=$(ARFLAGS) \ 546 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $(basename $(@F)) 547 548$(OBJDIR)/petsc-% : examples/petsc/%.c examples/petsc/libutils.a.PHONY $(libceed) $(ceed.pc) | $$(@D)/.DIR 549 +$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` \ 550 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 551 cp examples/petsc/$* $@ 552 553$(OBJDIR)/fluids-% : examples/fluids/%.c examples/fluids/src/*.c examples/fluids/*.h examples/fluids/problems/*.c examples/fluids/qfunctions/*.h $(libceed) $(ceed.pc) | $$(@D)/.DIR 554 +$(call quiet,MAKE) -C examples/fluids CEED_DIR=`pwd` \ 555 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 556 cp examples/fluids/$* $@ 557 558$(OBJDIR)/solids-% : examples/solids/%.c examples/solids/%.h \ 559 examples/solids/problems/*.c examples/solids/src/*.c \ 560 examples/solids/include/*.h examples/solids/problems/*.h examples/solids/qfunctions/*.h \ 561 $(libceed) $(ceed.pc) | $$(@D)/.DIR 562 +$(call quiet,MAKE) -C examples/solids CEED_DIR=`pwd` \ 563 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 564 cp examples/solids/$* $@ 565 566$(examples) : $(libceed) 567$(tests) : $(libceed) 568$(tests) $(examples) : override LDFLAGS += $(if $(STATIC),,-Wl,-rpath,$(abspath $(LIBDIR))) -L$(LIBDIR) 569 570run-% : $(OBJDIR)/% 571 @tests/tap.sh $(<:$(OBJDIR)/%=%) 572 573external_examples := \ 574 $(if $(MFEM_DIR),$(mfemexamples)) \ 575 $(if $(PETSC_DIR),$(petscexamples)) \ 576 $(if $(NEK5K_DIR),$(nekexamples)) \ 577 $(if $(PETSC_DIR),$(fluidsexamples)) \ 578 $(if $(PETSC_DIR),$(solidsexamples)) 579 580allexamples = $(examples) $(external_examples) 581 582# The test and prove targets can be controlled via pattern searches. The 583# default is to run tests and those examples that have no external dependencies. 584# Examples of finer grained control: 585# 586# make test search='petsc mfem' # PETSc and MFEM examples 587# make prove search='t3' # t3xx series tests 588# make junit search='ex petsc' # core ex* and PETSc tests 589search ?= t ex 590realsearch = $(search:%=%%) 591matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 592 593# Test core libCEED 594test : $(matched:$(OBJDIR)/%=run-%) 595 596# Run test target in parallel 597tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 598# CPU C tests only for backend % 599ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 600 601prove : $(matched) 602 $(info Testing backends: $(BACKENDS)) 603 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 604# Run prove target in parallel 605prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 606 607prove-all : 608 +$(MAKE) prove realsearch=% 609 610junit-% : $(OBJDIR)/% 611 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 612 613junit : $(matched:$(OBJDIR)/%=junit-%) 614 615all: $(alltests) 616 617examples : $(allexamples) 618ceedexamples : $(examples) 619nekexamples : $(nekexamples) 620mfemexamples : $(mfemexamples) 621petscexamples : $(petscexamples) 622 623# Benchmarks 624allbenchmarks = petsc-bps 625bench_targets = $(addprefix bench-,$(allbenchmarks)) 626.PHONY: $(bench_targets) benchmarks 627$(bench_targets): bench-%: $(OBJDIR)/% 628 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS_MAKE)" -r $(*).sh 629benchmarks: $(bench_targets) 630 631$(ceed.pc) : pkgconfig-prefix = $(abspath .) 632$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 633.INTERMEDIATE : $(OBJDIR)/ceed.pc 634%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 635 @$(SED) \ 636 -e "s:%prefix%:$(pkgconfig-prefix):" \ 637 -e "s:%libs_private%:$(pkgconfig-libs-private):" $< > $@ 638 639$(OBJDIR)/interface/ceed-jit-source-root-default.o : CPPFLAGS += -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath ./include)/\"" 640$(OBJDIR)/interface/ceed-jit-source-root-install.o : CPPFLAGS += -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath $(includedir))/\"" 641 642install : $(libceed) $(OBJDIR)/ceed.pc 643 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 644 "$(includedir)/ceed/" "$(includedir)/ceed/jit-source/"\ 645 "$(includedir)/ceed/jit-source/cuda/" "$(includedir)/ceed/jit-source/hip/"\ 646 "$(includedir)/ceed/jit-source/gallery/" "$(libdir)" "$(pkgconfigdir)") 647 $(INSTALL_DATA) include/ceed/ceed.h "$(DESTDIR)$(includedir)/ceed/" 648 $(INSTALL_DATA) include/ceed/types.h "$(DESTDIR)$(includedir)/ceed/" 649 $(INSTALL_DATA) include/ceed/ceed-f32.h "$(DESTDIR)$(includedir)/ceed/" 650 $(INSTALL_DATA) include/ceed/ceed-f64.h "$(DESTDIR)$(includedir)/ceed/" 651 $(INSTALL_DATA) include/ceed/fortran.h "$(DESTDIR)$(includedir)/ceed/" 652 $(INSTALL_DATA) include/ceed/backend.h "$(DESTDIR)$(includedir)/ceed/" 653 $(INSTALL_DATA) include/ceed/cuda.h "$(DESTDIR)$(includedir)/ceed/" 654 $(INSTALL_DATA) include/ceed/hip.h "$(DESTDIR)$(includedir)/ceed/" 655 $(INSTALL_DATA) include/ceed/hash.h "$(DESTDIR)$(includedir)/ceed/" 656 $(INSTALL_DATA) include/ceed/khash.h "$(DESTDIR)$(includedir)/ceed/" 657 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 658 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 659 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 660 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 661 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/cuda/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/cuda/" 662 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/hip/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/hip/" 663 $(INSTALL_DATA) $(wildcard include/ceed/jit-source/gallery/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/gallery/" 664 665.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 666 667cln clean : 668 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 669 $(call quiet,MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 670 $(call quiet,MAKE) -C python/tests clean 671 $(RM) benchmarks/*output.txt 672 673distclean : clean 674 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 675 676DOXYGEN ?= doxygen 677doxygen : 678 $(DOXYGEN) -q Doxyfile 679 680doc-html doc-latexpdf doc-epub doc-livehtml : doc-% : doxygen 681 make -C doc/sphinx $* 682 683doc : doc-html 684 685style-c : 686 @astyle --options=.astylerc \ 687 $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), \ 688 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 689 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 690 691AUTOPEP8 = autopep8 692style-py : AUTOPEP8_ARGS = --in-place --aggressive 693style-py : 694 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py python/tests/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 695 696style : style-c style-py 697 698CLANG_TIDY ?= clang-tidy 699 700%.c.tidy : %.c 701 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include -I$(HIP_DIR)/include -DCEED_JIT_SOUCE_ROOT_DEFAULT="\"$(abspath ./include)/\"" 702 703%.cpp.tidy : %.cpp 704 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include -I$(HIP_DIR)/include 705 706tidy_c : $(libceed.c:%=%.tidy) 707tidy_cpp : $(libceed.cpp:%=%.tidy) 708 709tidy : tidy_c tidy_cpp 710 711ifneq ($(wildcard ../iwyu/*),) 712 IWYU_DIR ?= ../iwyu 713 IWYU_CC ?= $(IWYU_DIR)/build/bin/include-what-you-use 714endif 715 716iwyu : CC=$(IWYU_CC) 717iwyu : lib 718 719print : 720 @echo $(VAR)=$($(VAR)) 721 722print-% : 723 $(info [ variable name]: $*) 724 $(info [ origin]: $(origin $*)) 725 $(info [ flavor]: $(flavor $*)) 726 $(info [ value]: $(value $*)) 727 $(info [expanded value]: $($*)) 728 $(info ) 729 @true 730 731# "make configure" detects any variables passed on the command line or 732# previously set in config.mk, caching them in config.mk as simple 733# (:=) variables. Variables set in config.mk or on the command line 734# take precedence over the defaults provided in the file. Typical 735# usage: 736# 737# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 738# make 739# make prove 740# 741# The values in the file can be updated by passing them on the command 742# line, e.g., 743# 744# make configure CC=/path/to/other/clang 745 746# All variables to consider for caching 747CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 748 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 749 AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \ 750 MAGMA_DIR OCCA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR HIP_ARCH 751 752# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 753# was set on the command line or in config.mk (where it will appear as 754# a simple variable). 755needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 756 757configure : 758 $(file > $(CONFIG)) 759 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 760 @echo "Configuration cached in $(CONFIG):" 761 @cat $(CONFIG) 762 763wheel : export MARCHFLAG = -march=generic 764wheel : export WHEEL_PLAT = manylinux2010_x86_64 765wheel : 766 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 767 -e MARCHFLAG -e WHEEL_PLAT \ 768 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 769 770.PHONY : configure wheel 771 772# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 773-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 774