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 $(CONFIG) 19 20ifeq (,$(filter-out undefined default,$(origin CC))) 21 CC = gcc 22endif 23ifeq (,$(filter-out undefined default,$(origin CXX))) 24 CXX = g++ 25endif 26ifeq (,$(filter-out undefined default,$(origin FC))) 27 FC = gfortran 28endif 29ifeq (,$(filter-out undefined default,$(origin LINK))) 30 LINK = $(CC) 31endif 32NVCC ?= $(CUDA_DIR)/bin/nvcc 33NVCC_CXX ?= $(CXX) 34HIPCC ?= $(HIP_DIR)/bin/hipcc 35 36# ASAN must be left empty if you don't want to use it 37ASAN ?= 38 39LDFLAGS ?= 40UNDERSCORE ?= 1 41 42# MFEM_DIR env variable should point to sibling directory 43ifneq ($(wildcard ../mfem/libmfem.*),) 44 MFEM_DIR ?= ../mfem 45endif 46 47# NEK5K_DIR env variable should point to sibling directory 48ifneq ($(wildcard ../Nek5000/*),) 49 NEK5K_DIR ?= $(abspath ../Nek5000) 50endif 51export NEK5K_DIR 52MPI ?= 1 53 54# CEED_DIR env for NEK5K testing 55export CEED_DIR = $(abspath .) 56 57# XSMM_DIR env variable should point to XSMM master (github.com/hfp/libxsmm) 58XSMM_DIR ?= ../libxsmm 59 60# OCCA_DIR env variable should point to OCCA master (github.com/libocca/occa) 61OCCA_DIR ?= ../occa 62 63# env variable MAGMA_DIR can be used too 64#TODO: FIX THIS with hip detection... 65#MAGMA_DIR ?= ../magma 66# If CUDA_DIR is not set, check for nvcc, or resort to /usr/local/cuda 67#CUDA_DIR ?= $(or $(patsubst %/,%,$(dir $(patsubst %/,%,$(dir \ 68 $(shell which nvcc 2> /dev/null))))),/usr/local/cuda) 69#TODO: FIX THIS to not avoid cuda 70CUDA_DIR ?= . 71HIP_DIR ?= /opt/rocm/hip 72 73# Check for PETSc in ../petsc 74ifneq ($(wildcard ../petsc/lib/libpetsc.*),) 75 PETSC_DIR ?= ../petsc 76endif 77 78# Warning: SANTIZ options still don't run with /gpu/occa 79# export LSAN_OPTIONS=suppressions=.asanignore 80AFLAGS = -fsanitize=address #-fsanitize=undefined -fno-omit-frame-pointer 81 82CC_VENDOR := $(patsubst gcc%,gcc,$(firstword $(filter gcc% clang icc XL,$(shell $(CC) --version)))) 83FC_VENDOR := $(firstword $(filter GNU ifort XL,$(shell $(FC) --version 2>&1 || $(FC) -qversion))) 84 85# Default extra flags by vendor 86MARCHFLAG.gcc := -march=native 87MARCHFLAG.clang := $(MARCHFLAG.gcc) 88MARCHFLAG.icc := 89OMP_SIMD_FLAG.gcc := -fopenmp-simd 90OMP_SIMD_FLAG.clang := $(OMP_SIMD_FLAG.gcc) 91OMP_SIMD_FLAG.icc := -qopenmp-simd 92OPT.gcc := -ffp-contract=fast 93OPT.clang := $(OPT.gcc) 94CFLAGS.gcc := -fPIC -std=c99 -Wall -Wextra -Wno-unused-parameter -MMD -MP 95CFLAGS.clang := $(CFLAGS.gcc) 96CFLAGS.icc := $(CFLAGS.gcc) 97CFLAGS.XL := -qpic -MMD 98CXXFLAGS.gcc := -fPIC -std=c++11 -Wall -Wextra -Wno-unused-parameter -MMD -MP 99CXXFLAGS.clang := $(CXXFLAGS.gcc) 100CXXFLAGS.icc := $(CXXFLAGS.gcc) 101CXXFLAGS.XL := -qpic -MMD 102FFLAGS.GNU := -fPIC -cpp -Wall -Wextra -Wno-unused-parameter -Wno-unused-dummy-argument -MMD -MP 103FFLAGS.ifort := -fPIC -cpp 104FFLAGS.XL := -qpic -ffree-form -qpreprocess -qextname -MMD 105 106# This check works with compilers that use gcc and clang. It fails with some 107# compilers; e.g., xlc apparently ignores all options when -E is passed, thus 108# succeeds with any flags. Users can pass MARCHFLAG=... if desired. 109cc_check_flag = $(shell $(CC) -E -Werror $(1) -x c /dev/null > /dev/null 2>&1 && echo 1) 110MARCHFLAG := $(MARCHFLAG.$(CC_VENDOR)) 111MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG),-mcpu=native) 112MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG)) 113 114OMP_SIMD_FLAG := $(OMP_SIMD_FLAG.$(CC_VENDOR)) 115OMP_SIMD_FLAG := $(if $(call cc_check_flag,$(OMP_SIMD_FLAG)),$(OMP_SIMD_FLAG)) 116 117OPT ?= -O -g $(MARCHFLAG) $(OPT.$(CC_VENDOR)) $(OMP_SIMD_FLAG) 118CFLAGS ?= $(OPT) $(CFLAGS.$(CC_VENDOR)) 119CXXFLAGS ?= $(OPT) $(CXXFLAGS.$(CC_VENDOR)) 120NVCCFLAGS ?= -ccbin $(CXX) -Xcompiler "$(OPT)" -Xcompiler -fPIC 121HIPCCFLAGS ?= -g -ffp-contract=fast $(OMP_SIMD_FLAG) -Wextra -Wno-unused-parameter -fno-gpu-rdc -fPIC 122FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR)) 123 124ifeq ($(COVERAGE), 1) 125 CFLAGS += --coverage 126 LDFLAGS += --coverage 127endif 128 129CFLAGS += $(if $(ASAN),$(AFLAGS)) 130FFLAGS += $(if $(ASAN),$(AFLAGS)) 131LDFLAGS += $(if $(ASAN),$(AFLAGS)) 132CPPFLAGS += -I./include 133LDLIBS = -lm 134OBJDIR := build 135LIBDIR := lib 136 137# Installation variables 138prefix ?= /usr/local 139bindir = $(prefix)/bin 140libdir = $(prefix)/lib 141okldir = $(libdir)/okl 142includedir = $(prefix)/include 143pkgconfigdir = $(libdir)/pkgconfig 144INSTALL = install 145INSTALL_PROGRAM = $(INSTALL) 146INSTALL_DATA = $(INSTALL) -m644 147 148# Get number of processors of the machine 149NPROCS := $(shell getconf _NPROCESSORS_ONLN) 150# prepare make options to run in parallel 151MFLAGS := -j $(NPROCS) --warn-undefined-variables \ 152 --no-print-directory --no-keep-going 153 154PYTHON ?= python3 155PROVE ?= prove 156PROVE_OPTS ?= -j $(NPROCS) 157DARWIN := $(filter Darwin,$(shell uname -s)) 158SO_EXT := $(if $(DARWIN),dylib,so) 159 160ceed.pc := $(LIBDIR)/pkgconfig/ceed.pc 161libceed := $(LIBDIR)/libceed.$(SO_EXT) 162CEED_LIBS = -lceed 163libceed.c := $(wildcard interface/ceed*.c) 164gallery.c := $(wildcard gallery/*/ceed*.c) 165libceed.c += $(gallery.c) 166libceed_test := $(LIBDIR)/libceed_test.$(SO_EXT) 167libceeds = $(libceed) $(libceed_test) 168BACKENDS_BUILTIN := /cpu/self/ref/serial /cpu/self/ref/blocked /cpu/self/opt/serial /cpu/self/opt/blocked 169BACKENDS := $(BACKENDS_BUILTIN) 170 171# Tests 172tests.c := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.c)) 173tests.f := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.f90)) 174tests := $(tests.c:tests/%.c=$(OBJDIR)/%) 175ctests := $(tests) 176tests += $(tests.f:tests/%.f90=$(OBJDIR)/%) 177# Examples 178examples.c := $(sort $(wildcard examples/ceed/*.c)) 179examples.f := $(sort $(wildcard examples/ceed/*.f)) 180examples := $(examples.c:examples/ceed/%.c=$(OBJDIR)/%) 181examples += $(examples.f:examples/ceed/%.f=$(OBJDIR)/%) 182# MFEM Examples 183mfemexamples.cpp := $(sort $(wildcard examples/mfem/*.cpp)) 184mfemexamples := $(mfemexamples.cpp:examples/mfem/%.cpp=$(OBJDIR)/mfem-%) 185# Nek5K Examples 186nekexamples := $(OBJDIR)/nek-bps 187# PETSc Examples 188petscexamples.c := $(wildcard examples/petsc/*.c) 189petscexamples := $(petscexamples.c:examples/petsc/%.c=$(OBJDIR)/petsc-%) 190# Fluid Dynamics Examples 191fluidsexamples.c := $(sort $(wildcard examples/fluids/*.c)) 192fluidsexamples := $(fluidsexamples.c:examples/fluids/%.c=$(OBJDIR)/fluids-%) 193# Solid Mechanics Examples 194solidsexamples.c := $(sort $(wildcard examples/solids/*.c)) 195solidsexamples := $(solidsexamples.c:examples/solids/%.c=$(OBJDIR)/solids-%) 196 197# Backends/[ref, blocked, template, memcheck, opt, avx, occa, magma] 198ref.c := $(sort $(wildcard backends/ref/*.c)) 199blocked.c := $(sort $(wildcard backends/blocked/*.c)) 200template.c := $(sort $(wildcard backends/template/*.c)) 201ceedmemcheck.c := $(sort $(wildcard backends/memcheck/*.c)) 202opt.c := $(sort $(wildcard backends/opt/*.c)) 203avx.c := $(sort $(wildcard backends/avx/*.c)) 204xsmm.c := $(sort $(wildcard backends/xsmm/*.c)) 205cuda.c := $(sort $(wildcard backends/cuda/*.c)) 206cuda.cpp := $(sort $(wildcard backends/cuda/*.cpp)) 207cuda.cu := $(sort $(wildcard backends/cuda/*.cu)) 208cuda-reg.c := $(sort $(wildcard backends/cuda-reg/*.c)) 209cuda-reg.cu := $(sort $(wildcard backends/cuda-reg/*.cu)) 210cuda-shared.c := $(sort $(wildcard backends/cuda-shared/*.c)) 211cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/*.cu)) 212cuda-gen.c := $(sort $(wildcard backends/cuda-gen/*.c)) 213cuda-gen.cpp := $(sort $(wildcard backends/cuda-gen/*.cpp)) 214cuda-gen.cu := $(sort $(wildcard backends/cuda-gen/*.cu)) 215occa.c := $(sort $(wildcard backends/occa/*.c)) 216magma.c := $(sort $(wildcard backends/magma/*.c)) 217magma.cu := $(sort $(wildcard backends/magma/*.cu)) 218hip.hip := $(sort $(wildcard backends/hip/*.hip)) 219hip.cpp := $(sort $(wildcard backends/hip/*.cpp)) 220hip.c := $(sort $(wildcard backends/hip/*.c)) 221 222# Output using the 216-color rules mode 223rule_file = $(notdir $(1)) 224rule_path = $(patsubst %/,%,$(dir $(1))) 225last_path = $(notdir $(patsubst %/,%,$(dir $(1)))) 226ansicolor = $(shell echo $(call last_path,$(1)) | cksum | cut -b1-2 | xargs -IS expr 2 \* S + 17) 227emacs_out = @printf " %10s %s/%s\n" $(1) $(call rule_path,$(2)) $(call rule_file,$(2)) 228color_out = @if [ -t 1 ]; then \ 229 printf " %10s \033[38;5;%d;1m%s\033[m/%s\n" \ 230 $(1) $(call ansicolor,$(2)) \ 231 $(call rule_path,$(2)) $(call rule_file,$(2)); else \ 232 printf " %10s %s\n" $(1) $(2); fi 233# if TERM=dumb, use it, otherwise switch to the term one 234output = $(if $(TERM:dumb=),$(call color_out,$1,$2),$(call emacs_out,$1,$2)) 235 236# if V is set to non-nil, turn the verbose mode 237quiet = $(if $(V),$($(1)),$(call output,$1,$@);$($(1))) 238 239# Cancel built-in and old-fashioned implicit rules which we don't use 240.SUFFIXES: 241 242.SECONDEXPANSION: # to expand $$(@D)/.DIR 243 244%/.DIR : 245 @mkdir -p $(@D) 246 @touch $@ 247 248.PRECIOUS: %/.DIR 249 250lib: $(libceed) $(ceed.pc) 251# run 'lib' target in parallel 252par:;@$(MAKE) $(MFLAGS) V=$(V) lib 253backend_status = $(if $(filter $1,$(BACKENDS)), [backends: $1], [not found]) 254info: 255 $(info ------------------------------------) 256 $(info CC = $(CC)) 257 $(info CXX = $(CXX)) 258 $(info FC = $(FC)) 259 $(info CPPFLAGS = $(CPPFLAGS)) 260 $(info CFLAGS = $(value CFLAGS)) 261 $(info CXXFLAGS = $(value CXXFLAGS)) 262 $(info FFLAGS = $(value FFLAGS)) 263 $(info NVCCFLAGS = $(value NVCCFLAGS)) 264 $(info LDFLAGS = $(value LDFLAGS)) 265 $(info LDLIBS = $(LDLIBS)) 266 $(info OPT = $(OPT)) 267 $(info AFLAGS = $(AFLAGS)) 268 $(info ASAN = $(or $(ASAN),(empty))) 269 $(info V = $(or $(V),(empty)) [verbose=$(if $(V),on,off)]) 270 $(info ------------------------------------) 271 $(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,/cpu/self/memcheck/serial /cpu/sef/memcheck/blocked)) 272 $(info AVX_STATUS = $(AVX_STATUS)$(call backend_status,/cpu/self/avx/serial /cpu/self/avx/blocked)) 273 $(info XSMM_DIR = $(XSMM_DIR)$(call backend_status,/cpu/self/xsmm/serial /cpu/self/xsmm/blocked)) 274 $(info OCCA_DIR = $(OCCA_DIR)$(call backend_status,/cpu/occa /gpu/occa /omp/occa)) 275 $(info MAGMA_DIR = $(MAGMA_DIR)$(call backend_status,/gpu/magma /gpu/magma/det)) 276 $(info CUDA_DIR = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS))) 277 $(info HIP_DIR = $(HIP_DIR)$(call backend_status,$(HIP_BACKENDS))) 278 $(info ------------------------------------) 279 $(info MFEM_DIR = $(MFEM_DIR)) 280 $(info NEK5K_DIR = $(NEK5K_DIR)) 281 $(info PETSC_DIR = $(PETSC_DIR)) 282 $(info ------------------------------------) 283 $(info prefix = $(prefix)) 284 $(info includedir = $(value includedir)) 285 $(info libdir = $(value libdir)) 286 $(info okldir = $(value okldir)) 287 $(info pkgconfigdir = $(value pkgconfigdir)) 288 $(info ------------------------------------) 289 @true 290info-backends: 291 $(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS))) 292.PHONY: lib all par info info-backends 293 294$(libceed) : LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed))) 295$(libceed_test) : LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed_test))) 296 297# Standard Backends 298libceed.c += $(ref.c) 299libceed.c += $(blocked.c) 300libceed.c += $(opt.c) 301 302# Testing Backends 303test_backends.c := $(template.c) 304TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub 305 306# Memcheck Backend 307MEMCHK_STATUS = Disabled 308MEMCHK := $(shell echo "\#include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1) 309ifeq ($(MEMCHK),1) 310 MEMCHK_STATUS = Enabled 311 libceed.c += $(ceedmemcheck.c) 312 BACKENDS += /cpu/self/memcheck/serial /cpu/self/memcheck/blocked 313endif 314 315# AVX Backed 316AVX_STATUS = Disabled 317AVX_FLAG := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx) 318AVX := $(filter $(AVX_FLAG),$(shell $(CC) $(OPT) -v -E -x c /dev/null 2>&1)) 319ifneq ($(AVX),) 320 AVX_STATUS = Enabled 321 libceed.c += $(avx.c) 322 BACKENDS += /cpu/self/avx/serial /cpu/self/avx/blocked 323endif 324 325# libXSMM Backends 326ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),) 327 $(libceeds) : LDFLAGS += -L$(XSMM_DIR)/lib -Wl,-rpath,$(abspath $(XSMM_DIR)/lib) 328 $(libceeds) : LDLIBS += -lxsmm -ldl 329 MKL ?= 330 ifeq (,$(MKL)$(MKLROOT)) 331 BLAS_LIB = -lblas 332 else 333 ifneq ($(MKLROOT),) 334 # Some installs put everything inside an intel64 subdirectory, others not 335 MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*))) 336 MKL_LINK = -L$(MKL_LIBDIR) -Wl,-rpath,$(MKL_LIBDIR) 337 endif 338 BLAS_LIB = $(MKL_LINK) -Wl,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl 339 endif 340 $(libceeds) : LDLIBS += $(BLAS_LIB) 341 libceed.c += $(xsmm.c) 342 $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include 343 BACKENDS += /cpu/self/xsmm/serial /cpu/self/xsmm/blocked 344endif 345 346# OCCA Backends 347ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),) 348 $(libceeds) : LDFLAGS += -L$(OCCA_DIR)/lib -Wl,-rpath,$(abspath $(OCCA_DIR)/lib) 349 $(libceeds) : LDLIBS += -locca 350 libceed.c += $(occa.c) 351 $(occa.c:%.c=$(OBJDIR)/%.o) $(occa.c:%=%.tidy) : CPPFLAGS += -I$(OCCA_DIR)/include 352 BACKENDS += /cpu/occa /gpu/occa /omp/occa 353endif 354 355# CUDA Backends 356CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(CUDA_DIR)/$d/libcudart.${SO_EXT})) 357CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR)))) 358CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs 359CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/reg /gpu/cuda/shared /gpu/cuda/gen 360ifneq ($(CUDA_LIB_DIR),) 361 $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include 362 $(libceeds) : LDFLAGS += -L$(CUDA_LIB_DIR) -Wl,-rpath,$(abspath $(CUDA_LIB_DIR)) 363 $(libceeds) : LDLIBS += -lcudart -lnvrtc -lcuda 364 $(libceeds) : LINK = $(CXX) 365 libceed.c += $(cuda.c) $(cuda-reg.c) $(cuda-shared.c) $(cuda-gen.c) 366 libceed.cpp += $(cuda.cpp) $(cuda-gen.cpp) 367 libceed.cu += $(cuda.cu) $(cuda-reg.cu) $(cuda-shared.cu) $(cuda-gen.cu) 368 BACKENDS += $(CUDA_BACKENDS) 369endif 370 371 372# HIP Backends 373ROCM_DIR = $(HIP_DIR)/.. 374HIPBLASDIR = $(ROCM_DIR)/hipblas 375HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libhiprtc.${SO_EXT})) 376HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 377HIP_BACKENDS = /gpu/hip/ref 378ifneq ($(HIP_LIB_DIR),) 379 $(libceeds) : HIPCCFLAGS += -I$(HIP_DIR)/include -I./include -I$(ROCM_DIR)/include 380 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include -I$(ROCM_DIR)/include 381 ifneq ($(CXX), $(HIPCC)) 382 CPPFLAGS += -D__HIP_PLATFORM_HCC__ 383 endif 384 $(libceeds) : LDFLAGS += -L$(HIP_LIB_DIR) -Wl,-rpath,$(abspath $(HIP_LIB_DIR)) -L$(HIPBLASDIR)/lib -Wl,-rpath,$(abspath $(HIPBLASDIR)/lib) 385 $(libceeds) : LDLIBS += -lhip_hcc -lhiprtc -lhipblas 386 $(libceeds) : LINK = $(CXX) 387 libceed.hip += $(hip.hip) 388 libceed.cpp += $(hip.cpp) 389 libceed.c += $(hip.c) 390 $(hip.c:%.c=$(OBJDIR)/%.o) $(hip.c:%=%.tidy) : CPPFLAGS += -I$(HIP_DIR)/include -I$(ROCM_DIR)/include 391 BACKENDS += $(HIP_BACKENDS) 392endif 393 394# MAGMA Backend 395ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 396 ifneq ($(CUDA_LIB_DIR),) 397 cuda_link = -Wl,-rpath,$(CUDA_LIB_DIR) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart 398 omp_link = -fopenmp 399 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link) 400 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 401 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 402 $(libceeds) : LDLIBS += $(magma_link) 403 $(tests) $(examples) : LDLIBS += $(magma_link) 404 libceed.c += $(magma.c) 405 libceed.cu += $(magma.cu) 406 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include 407 $(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 408 BACKENDS += /gpu/magma /gpu/magma/det 409 endif 410endif 411 412export BACKENDS 413 414libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip=$(OBJDIR)/%.o) 415$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 416$(libceed.o): | info-backends 417$(libceed) : $(libceed.o) | $$(@D)/.DIR 418 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 419 420$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 421 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 422 423$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 424 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 425 426$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 427 $(call quiet,NVCC) $(CPPFLAGS) $(NVCCFLAGS) -c -o $@ $(abspath $<) 428 429$(OBJDIR)/%.o : $(CURDIR)/%.hip | $$(@D)/.DIR 430 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 431 432$(OBJDIR)/% : tests/%.c | $$(@D)/.DIR 433 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 434 435$(OBJDIR)/% : tests/%.f90 | $$(@D)/.DIR 436 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 437 438$(OBJDIR)/% : examples/ceed/%.c | $$(@D)/.DIR 439 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 440 441$(OBJDIR)/% : examples/ceed/%.f | $$(@D)/.DIR 442 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 443 444$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 445 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 446 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 447 mv examples/mfem/$* $@ 448 449# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 450# this single Nek bps file to be built in parallel with other examples, 451# such as when calling `make prove-all -j2`. 452$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 453 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 454 mv examples/nek/build/bps $(OBJDIR)/bps 455 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 456 457$(OBJDIR)/petsc-% : examples/petsc/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 458 +$(MAKE) -C examples/petsc CEED_DIR=`pwd` \ 459 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 460 mv examples/petsc/$* $@ 461 462$(OBJDIR)/fluids-% : examples/fluids/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 463 +$(MAKE) -C examples/fluids CEED_DIR=`pwd` \ 464 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 465 mv examples/fluids/$* $@ 466 467$(OBJDIR)/solids-% : examples/solids/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 468 +$(MAKE) -C examples/solids CEED_DIR=`pwd` \ 469 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 470 mv examples/solids/$* $@ 471 472libceed_test.o = $(test_backends.c:%.c=$(OBJDIR)/%.o) 473$(libceed_test) : $(libceed.o) $(libceed_test.o) | $$(@D)/.DIR 474 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 475 476$(examples) : $(libceed) 477$(tests) : $(libceed_test) 478$(tests) : CEED_LIBS = -lceed_test 479$(tests) $(examples) : LDFLAGS += -Wl,-rpath,$(abspath $(LIBDIR)) -L$(LIBDIR) 480 481run-t% : BACKENDS += $(TEST_BACKENDS) 482run-% : $(OBJDIR)/% 483 @tests/tap.sh $(<:$(OBJDIR)/%=%) 484 485external_examples := \ 486 $(if $(MFEM_DIR),$(mfemexamples)) \ 487 $(if $(PETSC_DIR),$(petscexamples)) \ 488 $(if $(NEK5K_DIR),$(nekexamples)) \ 489 $(if $(PETSC_DIR),$(fluidsexamples)) \ 490 $(if $(PETSC_DIR),$(solidsexamples)) 491 492allexamples = $(examples) $(external_examples) 493 494# The test and prove targets can be controlled via pattern searches. The 495# default is to run tests and those examples that have no external dependencies. 496# Examples of finer grained control: 497# 498# make test search='petsc mfem' # PETSc and MFEM examples 499# make prove search='t3' # t3xx series tests 500# make junit search='ex petsc' # core ex* and PETSc tests 501search ?= t ex 502realsearch = $(search:%=%%) 503matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 504 505# Test core libCEED 506test : $(matched:$(OBJDIR)/%=run-%) 507 508# Run test target in parallel 509tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 510# CPU C tests only for backend % 511ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 512 513prove : BACKENDS += $(TEST_BACKENDS) 514prove : $(matched) 515 $(info Testing backends: $(BACKENDS)) 516 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 517# Run prove target in parallel 518prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 519 520prove-all : 521 +$(MAKE) prove realsearch=% 522 523junit-t% : BACKENDS += $(TEST_BACKENDS) 524junit-% : $(OBJDIR)/% 525 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 526 527junit : $(matched:$(OBJDIR)/%=junit-%) 528 529all: $(alltests) 530 531examples : $(allexamples) 532ceedexamples : $(examples) 533nekexamples : $(nekexamples) 534mfemexamples : $(mfemexamples) 535petscexamples : $(petscexamples) 536 537# Benchmarks 538allbenchmarks = petsc-bps 539bench_targets = $(addprefix bench-,$(allbenchmarks)) 540.PHONY: $(bench_targets) benchmarks 541$(bench_targets): bench-%: $(OBJDIR)/% 542 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS)" -r $(*).sh 543benchmarks: $(bench_targets) 544 545$(ceed.pc) : pkgconfig-prefix = $(abspath .) 546$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 547.INTERMEDIATE : $(OBJDIR)/ceed.pc 548%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 549 @sed "s:%prefix%:$(pkgconfig-prefix):" $< > $@ 550 551OCCA := $(OCCA_DIR)/bin/occa 552OKL_KERNELS := $(wildcard backends/occa/*.okl) 553 554okl-cache : 555 $(OCCA) cache ceed $(OKL_KERNELS) 556 557okl-clear: 558 $(OCCA) clear -y -l ceed 559 560install : $(libceed) $(OBJDIR)/ceed.pc 561 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 562 "$(libdir)" "$(pkgconfigdir)" $(if $(OCCA_ON),"$(okldir)")) 563 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 564 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 565 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 566 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 567 $(if $(OCCA_ON),$(INSTALL_DATA) $(OKL_KERNELS) "$(DESTDIR)$(okldir)/") 568 569.PHONY : cln clean doxygen doc lib install all print test tst prove prv prove-all junit examples style style-c style-py tidy okl-cache okl-clear info info-backends 570 571cln clean : 572 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 573 $(MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 574 $(MAKE) -C tests/python clean 575 $(RM) benchmarks/*output.txt 576 577distclean : clean 578 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 579 580DOXYGEN ?= doxygen 581doxygen : 582 $(DOXYGEN) Doxyfile 583 584doc-html doc-latexpdf doc-epub : doc-% : doxygen 585 make -C doc/sphinx $* 586 587doc : doc-html 588 589style-c : 590 @astyle --options=.astylerc \ 591 $(filter-out include/ceedf.h tests/t320-basis-f.h, \ 592 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 593 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 594 595AUTOPEP8 = autopep8 596style-py : AUTOPEP8_ARGS = --in-place --aggressive 597style-py : 598 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py tests/python**/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 599 600style : style-c style-py 601 602CLANG_TIDY ?= clang-tidy 603%.c.tidy : %.c 604 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include 605 606tidy : $(libceed.c:%=%.tidy) 607 608print : 609 @echo $(VAR)=$($(VAR)) 610 611print-% : 612 $(info [ variable name]: $*) 613 $(info [ origin]: $(origin $*)) 614 $(info [ flavor]: $(flavor $*)) 615 $(info [ value]: $(value $*)) 616 $(info [expanded value]: $($*)) 617 $(info ) 618 @true 619 620# "make configure" detects any variables passed on the command line or 621# previously set in config.mk, caching them in config.mk as simple 622# (:=) variables. Variables set in config.mk or on the command line 623# take precedence over the defaults provided in the file. Typical 624# usage: 625# 626# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 627# make 628# make prove 629# 630# The values in the file can be updated by passing them on the command 631# line, e.g., 632# 633# make configure CC=/path/to/other/clang 634 635# All variables to consider for caching 636CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 637 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 638 LDFLAGS LDLIBS \ 639 MAGMA_DIR XSMM_DIR CUDA_DIR MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR 640 641# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 642# was set on the command line or in config.mk (where it will appear as 643# a simple variable). 644needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 645 646configure : 647 $(file > $(CONFIG)) 648 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 649 @echo "Configuration cached in $(CONFIG):" 650 @cat $(CONFIG) 651 652wheel : export MARCHFLAG = -march=generic 653wheel : export WHEEL_PLAT = manylinux2010_x86_64 654wheel : 655 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 656 -e MARCHFLAG -e WHEEL_PLAT \ 657 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 658 659.PHONY : configure wheel 660 661# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 662-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 663