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 34NVCC ?= $(CUDA_DIR)/bin/nvcc 35NVCC_CXX ?= $(CXX) 36HIPCC ?= $(HIP_DIR)/bin/hipcc 37 38# ASAN must be left empty if you don't want to use it 39ASAN ?= 40 41LDFLAGS ?= 42UNDERSCORE ?= 1 43 44# MFEM_DIR env variable should point to sibling directory 45ifneq ($(wildcard ../mfem/libmfem.*),) 46 MFEM_DIR ?= ../mfem 47endif 48 49# NEK5K_DIR env variable should point to sibling directory 50ifneq ($(wildcard ../Nek5000/*),) 51 NEK5K_DIR ?= $(abspath ../Nek5000) 52endif 53export NEK5K_DIR 54MPI ?= 1 55 56# CEED_DIR env for NEK5K testing 57export CEED_DIR = $(abspath .) 58 59# XSMM_DIR env variable should point to XSMM master (github.com/hfp/libxsmm) 60XSMM_DIR ?= ../libxsmm 61 62# OCCA_DIR env variable should point to OCCA master (github.com/libocca/occa) 63OCCA_DIR ?= ../occa 64 65# env variable MAGMA_DIR can be used too 66MAGMA_DIR ?= ../magma 67# If CUDA_DIR is not set, check for nvcc, or resort to /usr/local/cuda 68CUDA_DIR ?= $(or $(patsubst %/,%,$(dir $(patsubst %/,%,$(dir \ 69 $(shell which nvcc 2> /dev/null))))),/usr/local/cuda) 70HIP_DIR ?= /opt/rocm 71 72# Check for PETSc in ../petsc 73ifneq ($(wildcard ../petsc/lib/libpetsc.*),) 74 PETSC_DIR ?= ../petsc 75endif 76 77# Warning: SANTIZ options still don't run with /gpu/occa 78# export LSAN_OPTIONS=suppressions=.asanignore 79AFLAGS = -fsanitize=address #-fsanitize=undefined -fno-omit-frame-pointer 80 81# Note: Intel oneAPI C/C++ compiler is now icx/icpx 82CC_VENDOR := $(subst oneAPI,icc,$(patsubst gcc%,gcc,$(firstword $(filter gcc% clang icc oneAPI XL,$(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) -Wno-unused-function 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) -Wno-unused-function 100CXXFLAGS.icc := $(CXXFLAGS.gcc) 101CXXFLAGS.XL := -qpic -std=c++11 -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 ?= $(filter-out $(OMP_SIMD_FLAG),$(OPT)) -fPIC 122FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR)) 123 124ifeq ($(COVERAGE), 1) 125 CFLAGS += --coverage 126 CXXFLAGS += --coverage 127 LDFLAGS += --coverage 128endif 129 130CFLAGS += $(if $(ASAN),$(AFLAGS)) 131FFLAGS += $(if $(ASAN),$(AFLAGS)) 132LDFLAGS += $(if $(ASAN),$(AFLAGS)) 133CPPFLAGS += -I./include 134LDLIBS = -lm 135OBJDIR := build 136LIBDIR := lib 137 138# Installation variables 139prefix ?= /usr/local 140bindir = $(prefix)/bin 141libdir = $(prefix)/lib 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 := $(filter-out interface/ceed-cuda.c interface/ceed-hip.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/kernels/*.cu)) 208cuda-shared.c := $(sort $(wildcard backends/cuda-shared/*.c)) 209cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/kernels/*.cu)) 210cuda-gen.c := $(sort $(wildcard backends/cuda-gen/*.c)) 211cuda-gen.cpp := $(sort $(wildcard backends/cuda-gen/*.cpp)) 212cuda-gen.cu := $(sort $(wildcard backends/cuda-gen/kernels/*.cu)) 213occa.cpp := $(sort $(shell find backends/occa -type f -name *.cpp)) 214magma.c := $(sort $(wildcard backends/magma/*.c)) 215magma.cu := $(sort $(wildcard backends/magma/kernels/cuda/*.cu)) 216magma.hip := $(sort $(wildcard backends/magma/kernels/hip/*.hip.cpp)) 217hip.c := $(sort $(wildcard backends/hip/*.c)) 218hip.cpp := $(sort $(wildcard backends/hip/*.cpp)) 219hip.hip := $(sort $(wildcard backends/hip/kernels/*.hip.cpp)) 220hip-shared.c := $(sort $(wildcard backends/hip-shared/*.c)) 221hip-shared.hip := $(sort $(wildcard backends/hip-shared/kernels/*.hip.cpp)) 222hip-gen.c := $(sort $(wildcard backends/hip-gen/*.c)) 223hip-gen.cpp := $(sort $(wildcard backends/hip-gen/*.cpp)) 224 225# Quiet, color output 226quiet ?= $($(1)) 227 228# Cancel built-in and old-fashioned implicit rules which we don't use 229.SUFFIXES: 230 231.SECONDEXPANSION: # to expand $$(@D)/.DIR 232 233%/.DIR : 234 @mkdir -p $(@D) 235 @touch $@ 236 237.PRECIOUS: %/.DIR 238 239lib: $(libceed) $(ceed.pc) 240# run 'lib' target in parallel 241par:;@$(MAKE) $(MFLAGS) V=$(V) lib 242backend_status = $(if $(filter $1,$(BACKENDS)), [backends: $1], [not found]) 243info: 244 $(info ------------------------------------) 245 $(info CC = $(CC)) 246 $(info CXX = $(CXX)) 247 $(info FC = $(FC)) 248 $(info CPPFLAGS = $(CPPFLAGS)) 249 $(info CFLAGS = $(value CFLAGS)) 250 $(info CXXFLAGS = $(value CXXFLAGS)) 251 $(info FFLAGS = $(value FFLAGS)) 252 $(info NVCCFLAGS = $(value NVCCFLAGS)) 253 $(info HIPCCFLAGS = $(value HIPCCFLAGS)) 254 $(info LDFLAGS = $(value LDFLAGS)) 255 $(info LDLIBS = $(LDLIBS)) 256 $(info OPT = $(OPT)) 257 $(info AFLAGS = $(AFLAGS)) 258 $(info ASAN = $(or $(ASAN),(empty))) 259 $(info V = $(or $(V),(empty)) [verbose=$(if $(V),on,off)]) 260 $(info ------------------------------------) 261 $(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,$(MEMCHK_BACKENDS))) 262 $(info AVX_STATUS = $(AVX_STATUS)$(call backend_status,$(AVX_BACKENDS))) 263 $(info XSMM_DIR = $(XSMM_DIR)$(call backend_status,$(XSMM_BACKENDS))) 264 $(info OCCA_DIR = $(OCCA_DIR)$(call backend_status,$(OCCA_BACKENDS))) 265 $(info MAGMA_DIR = $(MAGMA_DIR)$(call backend_status,$(MAGMA_BACKENDS))) 266 $(info CUDA_DIR = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS))) 267 $(info HIP_DIR = $(HIP_DIR)$(call backend_status,$(HIP_BACKENDS))) 268 $(info ------------------------------------) 269 $(info MFEM_DIR = $(MFEM_DIR)) 270 $(info NEK5K_DIR = $(NEK5K_DIR)) 271 $(info PETSC_DIR = $(PETSC_DIR)) 272 $(info ------------------------------------) 273 $(info prefix = $(prefix)) 274 $(info includedir = $(value includedir)) 275 $(info libdir = $(value libdir)) 276 $(info pkgconfigdir = $(value pkgconfigdir)) 277 $(info ------------------------------------) 278 @true 279info-backends: 280 $(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS))) 281.PHONY: lib all par info info-backends 282 283$(libceed) : LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed))) 284$(libceed_test) : LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed_test))) 285 286# Standard Backends 287libceed.c += $(ref.c) 288libceed.c += $(blocked.c) 289libceed.c += $(opt.c) 290 291# Testing Backends 292test_backends.c := $(template.c) 293TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub 294 295# Memcheck Backend 296MEMCHK_STATUS = Disabled 297MEMCHK := $(shell echo "\#include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1) 298MEMCHK_BACKENDS = /cpu/self/memcheck/serial /cpu/self/memcheck/blocked 299ifeq ($(MEMCHK),1) 300 MEMCHK_STATUS = Enabled 301 libceed.c += $(ceedmemcheck.c) 302 BACKENDS += $(MEMCHK_BACKENDS) 303endif 304 305# AVX Backed 306AVX_STATUS = Disabled 307AVX_FLAG := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx) 308AVX := $(filter $(AVX_FLAG),$(shell $(CC) $(OPT) -v -E -x c /dev/null 2>&1)) 309AVX_BACKENDS = /cpu/self/avx/serial /cpu/self/avx/blocked 310ifneq ($(AVX),) 311 AVX_STATUS = Enabled 312 libceed.c += $(avx.c) 313 BACKENDS += $(AVX_BACKENDS) 314endif 315 316# libXSMM Backends 317XSMM_BACKENDS = /cpu/self/xsmm/serial /cpu/self/xsmm/blocked 318ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),) 319 $(libceeds) : LDFLAGS += -L$(XSMM_DIR)/lib -Wl,-rpath,$(abspath $(XSMM_DIR)/lib) 320 $(libceeds) : LDLIBS += -lxsmm -ldl 321 MKL ?= 322 ifeq (,$(MKL)$(MKLROOT)) 323 BLAS_LIB = -lblas 324 else 325 ifneq ($(MKLROOT),) 326 # Some installs put everything inside an intel64 subdirectory, others not 327 MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*))) 328 MKL_LINK = -L$(MKL_LIBDIR) -Wl,-rpath,$(MKL_LIBDIR) 329 endif 330 BLAS_LIB = $(MKL_LINK) -Wl,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl 331 endif 332 $(libceeds) : LDLIBS += $(BLAS_LIB) 333 libceed.c += $(xsmm.c) 334 $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include 335 BACKENDS += $(XSMM_BACKENDS) 336endif 337 338# OCCA Backends 339OCCA_BACKENDS = /cpu/self/occa 340ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),) 341 OCCA_MODES := $(shell $(OCCA_DIR)/bin/occa modes) 342 OCCA_BACKENDS += $(if $(filter OpenMP,$(OCCA_MODES)),/cpu/openmp/occa) 343# OCCA_BACKENDS += $(if $(filter OpenCL,$(OCCA_MODES)),/gpu/opencl/occa) 344 OCCA_BACKENDS += $(if $(filter HIP,$(OCCA_MODES)),/gpu/hip/occa) 345 OCCA_BACKENDS += $(if $(filter CUDA,$(OCCA_MODES)),/gpu/cuda/occa) 346 347 $(libceeds) : CPPFLAGS += -I$(OCCA_DIR)/include 348 $(libceeds) : LDFLAGS += -L$(OCCA_DIR)/lib -Wl,-rpath,$(abspath $(OCCA_DIR)/lib) 349 $(libceeds) : LDLIBS += -locca 350 $(libceeds) : LINK = $(CXX) 351 libceed.cpp += $(occa.cpp) 352 BACKENDS += $(OCCA_BACKENDS) 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/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 -lcublas 364 $(libceeds) : LINK = $(CXX) 365 libceed.c += interface/ceed-cuda.c 366 libceed.c += $(cuda.c) $(cuda-shared.c) $(cuda-gen.c) 367 libceed.cpp += $(cuda.cpp) $(cuda-gen.cpp) 368 libceed.cu += $(cuda.cu) $(cuda-shared.cu) $(cuda-gen.cu) 369 BACKENDS += $(CUDA_BACKENDS) 370endif 371 372# HIP Backends 373HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(HIP_DIR)/$d/libamdhip64.${SO_EXT})) 374HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR)))) 375HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen 376ifneq ($(HIP_LIB_DIR),) 377 $(libceeds) : HIPCCFLAGS += -I./include 378 ifneq ($(CXX), $(HIPCC)) 379 CPPFLAGS += $(subst =,,$(shell $(HIP_DIR)/bin/hipconfig -C)) 380 endif 381 $(libceeds) : CPPFLAGS += -I$(HIP_DIR)/include 382 $(libceeds) : LDFLAGS += -L$(HIP_LIB_DIR) -Wl,-rpath,$(abspath $(HIP_LIB_DIR)) 383 $(libceeds) : LDLIBS += -lamdhip64 -lhipblas 384 $(libceeds) : LINK = $(CXX) 385 libceed.c += interface/ceed-hip.c 386 libceed.c += $(hip.c) $(hip-shared.c) $(hip-gen.c) 387 libceed.cpp += $(hip.cpp) $(hip-gen.cpp) 388 libceed.hip += $(hip.hip) $(hip-shared.hip) 389 BACKENDS += $(HIP_BACKENDS) 390endif 391 392# MAGMA Backend 393ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),) 394 MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas") 395 ifeq ($(MAGMA_ARCH), 0) #CUDA MAGMA 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 MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det 409 endif 410 else # HIP MAGMA 411 ifneq ($(HIP_LIB_DIR),) 412 hip_link = -Wl,-rpath,$(HIP_LIB_DIR) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64 413 omp_link = -fopenmp 414 magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link) 415 magma_link_shared = -L$(MAGMA_DIR)/lib -Wl,-rpath,$(abspath $(MAGMA_DIR)/lib) -lmagma 416 magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static)) 417 $(libceeds) : LDLIBS += $(magma_link) 418 $(tests) $(examples) : LDLIBS += $(magma_link) 419 libceed.c += $(magma.c) 420 libceed.hip += $(magma.hip) 421 ifneq ($(CXX), $(HIPCC)) 422 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 423 else 424 $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : HIPCCFLAGS += -I$(MAGMA_DIR)/include -I$(HIP_DIR)/include -DHAVE_HIP -DADD_ 425 endif 426 $(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_ 427 MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det 428 endif 429 endif 430 BACKENDS += $(MAGMA_BACKENDS) 431endif 432 433export BACKENDS 434 435libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) 436$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE) 437$(libceed.o): | info-backends 438$(libceed) : $(libceed.o) | $$(@D)/.DIR 439 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 440 441$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR 442 $(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<) 443 444$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR 445 $(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<) 446 447$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR 448 $(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<) 449 450$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR 451 $(call quiet,HIPCC) $(HIPCCFLAGS) -c -o $@ $(abspath $<) 452 453$(OBJDIR)/% : tests/%.c | $$(@D)/.DIR 454 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 455 456$(OBJDIR)/% : tests/%.f90 | $$(@D)/.DIR 457 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 458 459$(OBJDIR)/% : examples/ceed/%.c | $$(@D)/.DIR 460 $(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 461 462$(OBJDIR)/% : examples/ceed/%.f | $$(@D)/.DIR 463 $(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(LDLIBS) 464 465$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR 466 +$(MAKE) -C examples/mfem CEED_DIR=`pwd` \ 467 MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $* 468 mv examples/mfem/$* $@ 469 470# Note: Multiple Nek files cannot be built in parallel. The '+' here enables 471# this single Nek bps file to be built in parallel with other examples, 472# such as when calling `make prove-all -j2`. 473$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR 474 +$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek 475 mv examples/nek/build/bps $(OBJDIR)/bps 476 cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps 477 478$(OBJDIR)/petsc-% : examples/petsc/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 479 +$(MAKE) -C examples/petsc CEED_DIR=`pwd` \ 480 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 481 mv examples/petsc/$* $@ 482 483$(OBJDIR)/fluids-% : examples/fluids/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 484 +$(MAKE) -C examples/fluids CEED_DIR=`pwd` \ 485 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 486 mv examples/fluids/$* $@ 487 488$(OBJDIR)/solids-% : examples/solids/%.c $(libceed) $(ceed.pc) | $$(@D)/.DIR 489 +$(MAKE) -C examples/solids CEED_DIR=`pwd` \ 490 PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $* 491 mv examples/solids/$* $@ 492 493libceed_test.o = $(test_backends.c:%.c=$(OBJDIR)/%.o) 494$(libceed_test) : $(libceed.o) $(libceed_test.o) | $$(@D)/.DIR 495 $(call quiet,LINK) $(LDFLAGS) -shared -o $@ $^ $(LDLIBS) 496 497$(examples) : $(libceed) 498$(tests) : $(libceed_test) 499$(tests) : CEED_LIBS = -lceed_test 500$(tests) $(examples) : LDFLAGS += -Wl,-rpath,$(abspath $(LIBDIR)) -L$(LIBDIR) 501 502run-t% : BACKENDS += $(TEST_BACKENDS) 503run-% : $(OBJDIR)/% 504 @tests/tap.sh $(<:$(OBJDIR)/%=%) 505 506external_examples := \ 507 $(if $(MFEM_DIR),$(mfemexamples)) \ 508 $(if $(PETSC_DIR),$(petscexamples)) \ 509 $(if $(NEK5K_DIR),$(nekexamples)) \ 510 $(if $(PETSC_DIR),$(fluidsexamples)) \ 511 $(if $(PETSC_DIR),$(solidsexamples)) 512 513allexamples = $(examples) $(external_examples) 514 515# The test and prove targets can be controlled via pattern searches. The 516# default is to run tests and those examples that have no external dependencies. 517# Examples of finer grained control: 518# 519# make test search='petsc mfem' # PETSc and MFEM examples 520# make prove search='t3' # t3xx series tests 521# make junit search='ex petsc' # core ex* and PETSc tests 522search ?= t ex 523realsearch = $(search:%=%%) 524matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples))) 525 526# Test core libCEED 527test : $(matched:$(OBJDIR)/%=run-%) 528 529# Run test target in parallel 530tst : ;@$(MAKE) $(MFLAGS) V=$(V) test 531# CPU C tests only for backend % 532ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;) 533 534prove : BACKENDS += $(TEST_BACKENDS) 535prove : $(matched) 536 $(info Testing backends: $(BACKENDS)) 537 $(PROVE) $(PROVE_OPTS) --exec 'tests/tap.sh' $(matched:$(OBJDIR)/%=%) 538# Run prove target in parallel 539prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove 540 541prove-all : 542 +$(MAKE) prove realsearch=% 543 544junit-t% : BACKENDS += $(TEST_BACKENDS) 545junit-% : $(OBJDIR)/% 546 @printf " %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py $(<:$(OBJDIR)/%=%) 547 548junit : $(matched:$(OBJDIR)/%=junit-%) 549 550all: $(alltests) 551 552examples : $(allexamples) 553ceedexamples : $(examples) 554nekexamples : $(nekexamples) 555mfemexamples : $(mfemexamples) 556petscexamples : $(petscexamples) 557 558# Benchmarks 559allbenchmarks = petsc-bps 560bench_targets = $(addprefix bench-,$(allbenchmarks)) 561.PHONY: $(bench_targets) benchmarks 562$(bench_targets): bench-%: $(OBJDIR)/% 563 cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS)" -r $(*).sh 564benchmarks: $(bench_targets) 565 566$(ceed.pc) : pkgconfig-prefix = $(abspath .) 567$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix) 568.INTERMEDIATE : $(OBJDIR)/ceed.pc 569%/ceed.pc : ceed.pc.template | $$(@D)/.DIR 570 @sed "s:%prefix%:$(pkgconfig-prefix):" $< > $@ 571 572install : $(libceed) $(OBJDIR)/ceed.pc 573 $(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\ 574 "$(libdir)" "$(pkgconfigdir)") 575 $(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/" 576 $(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/" 577 $(INSTALL_DATA) include/ceed-hash.h "$(DESTDIR)$(includedir)/" 578 $(INSTALL_DATA) include/ceed-khash.h "$(DESTDIR)$(includedir)/" 579 $(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/" 580 $(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/" 581 582.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 583 584cln clean : 585 $(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi* 586 $(MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))" 587 $(MAKE) -C tests/python clean 588 $(RM) benchmarks/*output.txt 589 590distclean : clean 591 $(RM) -r doc/html doc/sphinx/build $(CONFIG) 592 593DOXYGEN ?= doxygen 594doxygen : 595 $(DOXYGEN) Doxyfile 596 597doc-html doc-latexpdf doc-epub : doc-% : doxygen 598 make -C doc/sphinx $* 599 600doc : doc-html 601 602style-c : 603 @astyle --options=.astylerc \ 604 $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), \ 605 $(wildcard include/*.h interface/*.[ch] tests/*.[ch] backends/*/*.[ch] \ 606 examples/*/*/*.[ch] examples/*/*.[ch] examples/*/*.[ch]pp gallery/*/*.[ch])) 607 608AUTOPEP8 = autopep8 609style-py : AUTOPEP8_ARGS = --in-place --aggressive 610style-py : 611 @$(AUTOPEP8) $(AUTOPEP8_ARGS) $(wildcard *.py python**/*.py tests/python**/*.py examples**/*.py doc/sphinx/source**/*.py benchmarks/*.py) 612 613style : style-c style-py 614 615CLANG_TIDY ?= clang-tidy 616 617%.c.tidy : %.c 618 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include 619 620%.cpp.tidy : %.cpp 621 $(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include 622 623tidy_c : $(libceed.c:%=%.tidy) 624tidy_cpp : $(libceed.cpp:%=%.tidy) 625 626tidy : tidy_c tidy_cpp 627 628print : 629 @echo $(VAR)=$($(VAR)) 630 631print-% : 632 $(info [ variable name]: $*) 633 $(info [ origin]: $(origin $*)) 634 $(info [ flavor]: $(flavor $*)) 635 $(info [ value]: $(value $*)) 636 $(info [expanded value]: $($*)) 637 $(info ) 638 @true 639 640# "make configure" detects any variables passed on the command line or 641# previously set in config.mk, caching them in config.mk as simple 642# (:=) variables. Variables set in config.mk or on the command line 643# take precedence over the defaults provided in the file. Typical 644# usage: 645# 646# make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda 647# make 648# make prove 649# 650# The values in the file can be updated by passing them on the command 651# line, e.g., 652# 653# make configure CC=/path/to/other/clang 654 655# All variables to consider for caching 656CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \ 657 OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS \ 658 LDFLAGS LDLIBS \ 659 MAGMA_DIR XSMM_DIR CUDA_DIR MFEM_DIR PETSC_DIR NEK5K_DIR HIP_DIR 660 661# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS 662# was set on the command line or in config.mk (where it will appear as 663# a simple variable). 664needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1)))) 665 666configure : 667 $(file > $(CONFIG)) 668 $(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v))))) 669 @echo "Configuration cached in $(CONFIG):" 670 @cat $(CONFIG) 671 672wheel : export MARCHFLAG = -march=generic 673wheel : export WHEEL_PLAT = manylinux2010_x86_64 674wheel : 675 docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \ 676 -e MARCHFLAG -e WHEEL_PLAT \ 677 quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh 678 679.PHONY : configure wheel 680 681# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container 682-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d)) 683