xref: /libCEED/Makefile (revision fc39c77ef991b17a0c558c74fe821c6309b1c03f)
1# Copyright (c) 2017-2024, 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 ?= $(ROCM_DIR)/bin/hipcc
34SYCLCXX ?= $(CXX)
35SED ?= sed
36ifneq ($(EMSCRIPTEN),)
37  STATIC = 1
38  EXE_SUFFIX = .wasm
39  EM_LDFLAGS = -s TOTAL_MEMORY=256MB
40endif
41
42# ASAN must be left empty if you don't want to use it
43ASAN ?=
44
45# These are the values automatically detected here in the makefile. They are
46# augmented with LDFLAGS and LDLIBS from the environment/passed by command line,
47# if any. If the user sets CEED_LDFLAGS or CEED_LDLIBS, they are used *instead
48# of* what we populate here (thus that's advanced usage and not recommended).
49CEED_LDFLAGS ?=
50CEED_LDLIBS ?=
51
52UNDERSCORE ?= 1
53
54# Verbose mode, V or VERBOSE
55V ?= $(VERBOSE)
56
57# MFEM_DIR env variable should point to sibling directory
58ifneq ($(wildcard ../mfem/libmfem.*),)
59  MFEM_DIR ?= ../mfem
60endif
61
62# NEK5K_DIR env variable should point to sibling directory
63ifneq ($(wildcard ../Nek5000/*),)
64  NEK5K_DIR ?= $(abspath ../Nek5000)
65endif
66export NEK5K_DIR
67MPI ?= 1
68
69# DEAL_II_DIR env variable should point to sibling directory
70ifneq ($(wildcard ../dealii/install/lib/libdeal_II.*),)
71  DEAL_II_DIR ?= ../dealii/install
72endif
73export DEAL_II_DIR
74
75# CEED_DIR env for NEK5K testing
76export CEED_DIR = $(abspath .)
77
78# XSMM_DIR env variable should point to XSMM main (github.com/hfp/libxsmm)
79XSMM_DIR ?= ../libxsmm
80
81# OCCA_DIR env variable should point to OCCA main (github.com/libocca/occa)
82OCCA_DIR ?= ../occa/install
83
84# env variable MAGMA_DIR can be used too
85MAGMA_DIR ?= ../magma
86
87# Often /opt/cuda or /usr/local/cuda, but sometimes present on machines that don't support CUDA
88CUDA_DIR  ?=
89CUDA_ARCH ?=
90
91# Often /opt/rocm, but sometimes present on machines that don't support HIP
92ROCM_DIR ?=
93HIP_ARCH ?=
94
95# Check for PETSc in ../petsc
96ifneq ($(wildcard ../petsc/lib/libpetsc.*),)
97  PETSC_DIR ?= ../petsc
98endif
99
100# SmartSim testing
101SMARTREDIS_DIR ?=
102
103# PyTorch testing
104USE_TORCH ?=
105
106# Warning: SANTIZ options still don't run with /gpu/occa
107AFLAGS ?= -fsanitize=address #-fsanitize=undefined -fno-omit-frame-pointer
108
109# Note: Intel oneAPI C/C++ compiler is now icx/icpx
110CC_VENDOR := $(firstword $(filter gcc (GCC) clang icc icc_orig oneAPI XL emcc,$(subst -, ,$(shell $(CC) --version))))
111CC_VENDOR := $(subst (GCC),gcc,$(subst icc_orig,icc,$(CC_VENDOR)))
112FC_VENDOR := $(if $(FC),$(firstword $(filter GNU ifort ifx XL,$(shell $(FC) --version 2>&1 || $(FC) -qversion))))
113
114# Default extra flags by vendor
115MARCHFLAG.gcc           := -march=native
116MARCHFLAG.clang         := $(MARCHFLAG.gcc)
117MARCHFLAG.icc           :=
118MARCHFLAG.oneAPI        := $(MARCHFLAG.clang)
119OMP_SIMD_FLAG.gcc       := -fopenmp-simd
120OMP_SIMD_FLAG.clang     := $(OMP_SIMD_FLAG.gcc)
121OMP_SIMD_FLAG.icc       := -qopenmp-simd
122OMP_SIMD_FLAG.oneAPI    := $(OMP_SIMD_FLAG.icc)
123OMP_FLAG.gcc            := -fopenmp
124OMP_FLAG.clang          := $(OMP_FLAG.gcc)
125OMP_FLAG.icc            := -qopenmp
126OMP_FLAG.oneAPI         := $(OMP_FLAG.icc)
127SYCL_FLAG.gcc           :=
128SYCL_FLAG.clang         := -fsycl
129SYCL_FLAG.icc           :=
130SYCL_FLAG.oneAPI        := -fsycl -fno-sycl-id-queries-fit-in-int
131OPT.gcc                 := -g -ffp-contract=fast
132OPT.clang               := $(OPT.gcc)
133OPT.icc                 := $(OPT.gcc)
134OPT.oneAPI              := $(OPT.clang)
135OPT.emcc                :=
136CFLAGS.gcc              := $(if $(STATIC),,-fPIC) -std=c99 -Wall -Wextra -Wno-unused-parameter -MMD -MP
137CFLAGS.clang            := $(CFLAGS.gcc)
138CFLAGS.icc              := $(CFLAGS.gcc)
139CFLAGS.oneAPI           := $(CFLAGS.clang)
140CFLAGS.XL               := $(if $(STATIC),,-qpic) -MMD
141CFLAGS.emcc             := $(CFLAGS.clang)
142CXXFLAGS.gcc            := $(if $(STATIC),,-fPIC) -std=c++11 -Wall -Wextra -Wno-unused-parameter -MMD -MP
143CXXFLAGS.clang          := $(CXXFLAGS.gcc)
144CXXFLAGS.icc            := $(CXXFLAGS.gcc)
145CXXFLAGS.oneAPI         := $(CXXFLAGS.clang)
146CXXFLAGS.XL             := $(if $(STATIC),,-qpic) -std=c++11 -MMD
147CXXFLAGS.emcc           := $(CXXFLAGS.clang)
148FFLAGS.GNU              := $(if $(STATIC),,-fPIC) -cpp -Wall -Wextra -Wno-unused-parameter -Wno-unused-dummy-argument -MMD -MP
149FFLAGS.ifort            := $(if $(STATIC),,-fPIC) -cpp
150FFLAGS.ifx              := $(FFLAGS.ifort)
151FFLAGS.XL               := $(if $(STATIC),,-qpic) -ffree-form -qpreprocess -qextname -MMD
152
153# This check works with compilers that use gcc and clang.  It fails with some
154# compilers; e.g., xlc apparently ignores all options when -E is passed, thus
155# succeeds with any flags.  Users can pass MARCHFLAG=... if desired.
156cc_check_flag = $(shell $(CC) -E -Werror $(1) -x c /dev/null > /dev/null 2>&1 && echo 1)
157MARCHFLAG := $(MARCHFLAG.$(CC_VENDOR))
158MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG),-mcpu=native)
159MARCHFLAG := $(if $(call cc_check_flag,$(MARCHFLAG)),$(MARCHFLAG))
160
161OMP_SIMD_FLAG := $(OMP_SIMD_FLAG.$(CC_VENDOR))
162OMP_SIMD_FLAG := $(if $(call cc_check_flag,$(OMP_SIMD_FLAG)),$(OMP_SIMD_FLAG))
163
164# Error checking flags
165PEDANTIC      ?=
166PEDANTICFLAGS ?= -Werror -pedantic
167
168OPT    ?= -O $(MARCHFLAG) $(OPT.$(CC_VENDOR)) $(OMP_SIMD_FLAG)
169CFLAGS ?= $(OPT) $(CFLAGS.$(CC_VENDOR)) $(if $(PEDANTIC),$(PEDANTICFLAGS))
170CXXFLAGS ?= $(OPT) $(CXXFLAGS.$(CC_VENDOR)) $(if $(PEDANTIC),$(PEDANTICFLAGS))
171FFLAGS ?= $(OPT) $(FFLAGS.$(FC_VENDOR))
172LIBCXX ?= -lstdc++
173NVCCFLAGS ?= -ccbin $(CXX) -Xcompiler "$(OPT)" -Xcompiler -fPIC
174ifneq ($(CUDA_ARCH),)
175  NVCCFLAGS += -arch=$(CUDA_ARCH)
176endif
177HIPCCFLAGS ?= $(filter-out $(OMP_SIMD_FLAG),$(OPT)) -fPIC -munsafe-fp-atomics
178ifneq ($(HIP_ARCH),)
179  HIPCCFLAGS += --amdgpu-target=$(HIP_ARCH)
180endif
181SYCL_FLAG := $(SYCL_FLAG.$(CC_VENDOR))
182SYCLFLAGS ?= $(SYCL_FLAG) -fPIC -std=c++17 $(filter-out -std=c++11,$(CXXFLAGS)) $(filter-out $(OMP_SIMD_FLAG),$(OPT))
183
184OPENMP ?=
185ifneq ($(OPENMP),)
186  OMP_FLAG := $(OMP_FLAG.$(CC_VENDOR))
187  OMP_FLAG := $(if $(call cc_check_flag,$(OMP_FLAG)),$(OMP_FLAG))
188  CFLAGS += $(OMP_FLAG)
189  CEED_LDFLAGS += $(OMP_FLAG)
190endif
191
192ifeq ($(COVERAGE), 1)
193  CFLAGS += --coverage
194  CXXFLAGS += --coverage
195  CEED_LDFLAGS += --coverage
196endif
197
198CFLAGS += $(if $(ASAN),$(AFLAGS))
199FFLAGS += $(if $(ASAN),$(AFLAGS))
200CEED_LDFLAGS += $(if $(ASAN),$(AFLAGS))
201CPPFLAGS += -I./include
202CEED_LDLIBS = -lm
203OBJDIR := build
204for_install := $(filter install,$(MAKECMDGOALS))
205LIBDIR := $(if $(for_install),$(OBJDIR),lib)
206
207
208# Installation variables
209prefix ?= /usr/local
210bindir = $(prefix)/bin
211libdir = $(prefix)/lib
212includedir = $(prefix)/include
213pkgconfigdir = $(libdir)/pkgconfig
214INSTALL = install
215INSTALL_PROGRAM = $(INSTALL)
216INSTALL_DATA = $(INSTALL) -m644
217
218# Get number of processors of the machine
219NPROCS := $(shell getconf _NPROCESSORS_ONLN)
220# prepare make options to run in parallel
221MFLAGS := -j $(NPROCS) --warn-undefined-variables \
222                       --no-print-directory --no-keep-going
223
224PYTHON ?= python3
225PROVE ?= prove
226PROVE_OPTS ?= -j $(NPROCS)
227DARWIN := $(filter Darwin,$(shell uname -s))
228SO_EXT := $(if $(DARWIN),dylib,so)
229
230ceed.pc := $(LIBDIR)/pkgconfig/ceed.pc
231libceed.so := $(LIBDIR)/libceed.$(SO_EXT)
232libceed.a := $(LIBDIR)/libceed.a
233libceed := $(if $(STATIC),$(libceed.a),$(libceed.so))
234CEED_LIBS = -lceed
235libceed.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))
236gallery.c := $(wildcard gallery/*/ceed*.c)
237libceed.c += $(gallery.c)
238libceeds = $(libceed)
239BACKENDS_BUILTIN := /cpu/self/ref/serial /cpu/self/ref/blocked /cpu/self/opt/serial /cpu/self/opt/blocked
240BACKENDS_MAKE := $(BACKENDS_BUILTIN)
241TEST_BACKENDS := /cpu/self/tmpl /cpu/self/tmpl/sub
242
243# Tests
244tests.c   := $(sort $(wildcard tests/t[0-9][0-9][0-9]-*.c))
245tests.f   := $(if $(FC),$(sort $(wildcard tests/t[0-9][0-9][0-9]-*.f90)))
246tests     := $(tests.c:tests/%.c=$(OBJDIR)/%$(EXE_SUFFIX))
247ctests    := $(tests)
248tests     += $(tests.f:tests/%.f90=$(OBJDIR)/%$(EXE_SUFFIX))
249# Examples
250examples.c := $(sort $(wildcard examples/ceed/*.c))
251examples.f := $(if $(FC),$(sort $(wildcard examples/ceed/*.f)))
252examples   := $(examples.c:examples/ceed/%.c=$(OBJDIR)/%$(EXE_SUFFIX))
253examples   += $(examples.f:examples/ceed/%.f=$(OBJDIR)/%$(EXE_SUFFIX))
254# MFEM Examples
255mfemexamples.cpp := $(sort $(wildcard examples/mfem/*.cpp))
256mfemexamples  := $(mfemexamples.cpp:examples/mfem/%.cpp=$(OBJDIR)/mfem-%)
257# Nek5K Examples
258nekexamples   := $(OBJDIR)/nek-bps
259# PETSc Examples
260petscexamples.c := $(wildcard examples/petsc/*.c)
261petscexamples   := $(petscexamples.c:examples/petsc/%.c=$(OBJDIR)/petsc-%)
262# deal.II Examples
263dealiiexamples  := $(OBJDIR)/dealii-bps
264# Fluid Dynamics Examples
265fluidsexamples.c  := $(sort $(wildcard examples/fluids/*.c))
266fluidsexamples.py := examples/fluids/smartsim_regression_framework.py
267fluidsexamples    := $(fluidsexamples.c:examples/fluids/%.c=$(OBJDIR)/fluids-%)
268fluidsexamples    += $(fluidsexamples.py:examples/fluids/%.py=$(OBJDIR)/fluids-py-%)
269# Solid Mechanics Examples
270solidsexamples.c  := $(sort $(wildcard examples/solids/*.c))
271solidsexamples    := $(solidsexamples.c:examples/solids/%.c=$(OBJDIR)/solids-%)
272
273# Backends/[ref, blocked, memcheck, opt, avx, occa, magma]
274ref.c          := $(sort $(wildcard backends/ref/*.c))
275blocked.c      := $(sort $(wildcard backends/blocked/*.c))
276ceedmemcheck.c := $(sort $(wildcard backends/memcheck/*.c))
277opt.c          := $(sort $(wildcard backends/opt/*.c))
278avx.c          := $(sort $(wildcard backends/avx/*.c))
279xsmm.c         := $(sort $(wildcard backends/xsmm/*.c))
280cuda.c         := $(sort $(wildcard backends/cuda/*.c))
281cuda.cpp       := $(sort $(wildcard backends/cuda/*.cpp))
282cuda-ref.c     := $(sort $(wildcard backends/cuda-ref/*.c))
283cuda-ref.cpp   := $(sort $(wildcard backends/cuda-ref/*.cpp))
284cuda-ref.cu    := $(sort $(wildcard backends/cuda-ref/kernels/*.cu))
285cuda-shared.c  := $(sort $(wildcard backends/cuda-shared/*.c))
286cuda-shared.cu := $(sort $(wildcard backends/cuda-shared/kernels/*.cu))
287cuda-gen.c     := $(sort $(wildcard backends/cuda-gen/*.c))
288cuda-gen.cpp   := $(sort $(wildcard backends/cuda-gen/*.cpp))
289cuda-gen.cu    := $(sort $(wildcard backends/cuda-gen/kernels/*.cu))
290occa.cpp       := $(sort $(shell find backends/occa -type f -name *.cpp))
291magma.c        := $(sort $(wildcard backends/magma/*.c))
292magma.cpp      := $(sort $(wildcard backends/magma/*.cpp))
293hip.c          := $(sort $(wildcard backends/hip/*.c))
294hip.cpp        := $(sort $(wildcard backends/hip/*.cpp))
295hip-ref.c      := $(sort $(wildcard backends/hip-ref/*.c))
296hip-ref.cpp    := $(sort $(wildcard backends/hip-ref/*.cpp))
297hip-ref.hip    := $(sort $(wildcard backends/hip-ref/kernels/*.hip.cpp))
298hip-shared.c   := $(sort $(wildcard backends/hip-shared/*.c))
299hip-gen.c      := $(sort $(wildcard backends/hip-gen/*.c))
300hip-gen.cpp    := $(sort $(wildcard backends/hip-gen/*.cpp))
301sycl-core.cpp  := $(sort $(wildcard backends/sycl/*.sycl.cpp))
302sycl-ref.cpp   := $(sort $(wildcard backends/sycl-ref/*.sycl.cpp))
303sycl-shared.cpp:= $(sort $(wildcard backends/sycl-shared/*.sycl.cpp))
304sycl-gen.cpp   := $(sort $(wildcard backends/sycl-gen/*.sycl.cpp))
305
306hip-all.c := interface/ceed-hip.c $(hip.c) $(hip-ref.c) $(hip-shared.c) $(hip-gen.c)
307hip-all.cpp := $(hip.cpp) $(hip-ref.cpp) $(hip-gen.cpp)
308
309# Quiet, color output
310quiet ?= $($(1))
311
312# Cancel built-in and old-fashioned implicit rules which we don't use
313.SUFFIXES:
314
315.SECONDEXPANSION: # to expand $$(@D)/.DIR
316
317%/.DIR :
318	@mkdir -p $(@D)
319	@touch $@
320
321.PRECIOUS: %/.DIR
322
323lib: $(libceed) $(ceed.pc)
324# run 'lib' target in parallel
325par:;@$(MAKE) $(MFLAGS) V=$(V) lib
326backend_status = $(if $(filter $1,$(BACKENDS_MAKE)), [backends: $1], [not found])
327info:
328	$(info ------------------------------------)
329	$(info CC            = $(CC))
330	$(info CXX           = $(CXX))
331	$(info FC            = $(FC))
332	$(info CPPFLAGS      = $(CPPFLAGS))
333	$(info CFLAGS        = $(CFLAGS))
334	$(info CXXFLAGS      = $(CXXFLAGS))
335	$(info FFLAGS        = $(FFLAGS))
336	$(info NVCCFLAGS     = $(NVCCFLAGS))
337	$(info HIPCCFLAGS    = $(HIPCCFLAGS))
338	$(info SYCLFLAGS     = $(SYCLFLAGS))
339	$(info CEED_LDFLAGS  = $(CEED_LDFLAGS))
340	$(info CEED_LDLIBS   = $(CEED_LDLIBS))
341	$(info AR            = $(AR))
342	$(info ARFLAGS       = $(ARFLAGS))
343	$(info OPT           = $(OPT))
344	$(info AFLAGS        = $(AFLAGS))
345	$(info ASAN          = $(or $(ASAN),(empty)))
346	$(info VERBOSE       = $(or $(V),(empty)) [verbose=$(if $(V),on,off)])
347	$(info ------------------------------------)
348	$(info MEMCHK_STATUS = $(MEMCHK_STATUS)$(call backend_status,$(MEMCHK_BACKENDS)))
349	$(info AVX_STATUS    = $(AVX_STATUS)$(call backend_status,$(AVX_BACKENDS)))
350	$(info XSMM_DIR      = $(XSMM_DIR)$(call backend_status,$(XSMM_BACKENDS)))
351	$(info OCCA_DIR      = $(OCCA_DIR)$(call backend_status,$(OCCA_BACKENDS)))
352	$(info MAGMA_DIR     = $(MAGMA_DIR)$(call backend_status,$(MAGMA_BACKENDS)))
353	$(info CUDA_DIR      = $(CUDA_DIR)$(call backend_status,$(CUDA_BACKENDS)))
354	$(info ROCM_DIR      = $(ROCM_DIR)$(call backend_status,$(HIP_BACKENDS)))
355	$(info SYCL_DIR      = $(SYCL_DIR)$(call backend_status,$(SYCL_BACKENDS)))
356	$(info ------------------------------------)
357	$(info MFEM_DIR      = $(MFEM_DIR))
358	$(info NEK5K_DIR     = $(NEK5K_DIR))
359	$(info PETSC_DIR     = $(PETSC_DIR))
360	$(info DEAL_II_DIR   = $(DEAL_II_DIR))
361	$(info ------------------------------------)
362	$(info prefix        = $(prefix))
363	$(info includedir    = $(value includedir))
364	$(info libdir        = $(value libdir))
365	$(info pkgconfigdir  = $(value pkgconfigdir))
366	$(info ------------------------------------)
367	@true
368info-backends:
369	$(info make: 'lib' with optional backends: $(filter-out $(BACKENDS_BUILTIN),$(BACKENDS)))
370	@true
371info-backends-all:
372	$(info make: 'lib' with backends: $(filter-out $(TEST_BACKENDS),$(BACKENDS)))
373	@true
374
375$(libceed.so) : CEED_LDFLAGS += $(if $(DARWIN), -install_name @rpath/$(notdir $(libceed.so)))
376
377# Standard Backends
378libceed.c += $(ref.c)
379libceed.c += $(blocked.c)
380libceed.c += $(opt.c)
381
382# Memcheck Backends
383MEMCHK_STATUS   = Disabled
384MEMCHK         := $(shell echo "$(HASH)include <valgrind/memcheck.h>" | $(CC) $(CPPFLAGS) -E - >/dev/null 2>&1 && echo 1)
385MEMCHK_BACKENDS = /cpu/self/memcheck/serial /cpu/self/memcheck/blocked
386ifeq ($(MEMCHK),1)
387  MEMCHK_STATUS = Enabled
388  libceed.c += $(ceedmemcheck.c)
389  BACKENDS_MAKE += $(MEMCHK_BACKENDS)
390endif
391
392# AVX Backeds
393AVX_STATUS   = Disabled
394AVX_FLAG    := $(if $(filter clang,$(CC_VENDOR)),+avx,-mavx)
395AVX         := $(filter $(AVX_FLAG),$(shell $(CC) $(CFLAGS:-M%=) -v -E -x c /dev/null 2>&1))
396AVX_BACKENDS = /cpu/self/avx/serial /cpu/self/avx/blocked
397ifneq ($(AVX),)
398  AVX_STATUS = Enabled
399  libceed.c += $(avx.c)
400  BACKENDS_MAKE += $(AVX_BACKENDS)
401endif
402
403# Collect list of libraries and paths for use in linking and pkg-config
404PKG_LIBS =
405# Stubs that will not be RPATH'd
406PKG_STUBS_LIBS =
407
408# libXSMM Backends
409XSMM_BACKENDS = /cpu/self/xsmm/serial /cpu/self/xsmm/blocked
410ifneq ($(wildcard $(XSMM_DIR)/lib/libxsmm.*),)
411  PKG_LIBS += -L$(abspath $(XSMM_DIR))/lib -lxsmm
412  MKL ?=
413  ifeq (,$(MKL)$(MKLROOT))
414    BLAS_LIB ?= -lblas -ldl
415  else
416    ifneq ($(MKLROOT),)
417      # Some installs put everything inside an intel64 subdirectory, others not
418      MKL_LIBDIR = $(dir $(firstword $(wildcard $(MKLROOT)/lib/intel64/libmkl_sequential.* $(MKLROOT)/lib/libmkl_sequential.*)))
419      MKL_LINK = -L$(MKL_LIBDIR)
420    endif
421    BLAS_LIB ?= $(MKL_LINK) -Wl,--push-state,--no-as-needed -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl -Wl,--pop-state
422  endif
423  PKG_LIBS += $(BLAS_LIB)
424  libceed.c += $(xsmm.c)
425  $(xsmm.c:%.c=$(OBJDIR)/%.o) $(xsmm.c:%=%.tidy) : CPPFLAGS += -I$(XSMM_DIR)/include
426  BACKENDS_MAKE += $(XSMM_BACKENDS)
427endif
428
429# OCCA Backends
430OCCA_BACKENDS = /cpu/self/occa
431ifneq ($(wildcard $(OCCA_DIR)/lib/libocca.*),)
432  OCCA_MODES := $(shell LD_LIBRARY_PATH=$(OCCA_DIR)/lib $(OCCA_DIR)/bin/occa modes)
433  OCCA_BACKENDS += $(if $(filter OpenMP,$(OCCA_MODES)),/cpu/openmp/occa)
434  OCCA_BACKENDS += $(if $(filter dpcpp,$(OCCA_MODES)),/gpu/dpcpp/occa)
435  OCCA_BACKENDS += $(if $(filter OpenCL,$(OCCA_MODES)),/gpu/opencl/occa)
436  OCCA_BACKENDS += $(if $(filter HIP,$(OCCA_MODES)),/gpu/hip/occa)
437  OCCA_BACKENDS += $(if $(filter CUDA,$(OCCA_MODES)),/gpu/cuda/occa)
438  $(libceeds) : CPPFLAGS += -I$(OCCA_DIR)/include
439  PKG_LIBS += -L$(abspath $(OCCA_DIR))/lib -locca
440  LIBCEED_CONTAINS_CXX = 1
441  libceed.cpp += $(occa.cpp)
442  BACKENDS_MAKE += $(OCCA_BACKENDS)
443endif
444
445# CUDA Backends
446ifneq ($(CUDA_DIR),)
447  CUDA_LIB_DIR := $(wildcard $(foreach d,lib lib64 lib/x86_64-linux-gnu,$(CUDA_DIR)/$d/libcudart.${SO_EXT}))
448  CUDA_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(CUDA_LIB_DIR))))
449endif
450CUDA_LIB_DIR_STUBS := $(CUDA_LIB_DIR)/stubs
451CUDA_BACKENDS = /gpu/cuda/ref /gpu/cuda/shared /gpu/cuda/gen
452ifneq ($(CUDA_LIB_DIR),)
453  $(libceeds) : CPPFLAGS += -I$(CUDA_DIR)/include
454  PKG_LIBS += -L$(abspath $(CUDA_LIB_DIR)) -lcudart -lnvrtc -lcuda -lcublas
455  PKG_STUBS_LIBS += -L$(CUDA_LIB_DIR_STUBS)
456  LIBCEED_CONTAINS_CXX = 1
457  libceed.c     += interface/ceed-cuda.c
458  libceed.c     += $(cuda.c) $(cuda-ref.c) $(cuda-shared.c) $(cuda-gen.c)
459  libceed.cpp   += $(cuda.cpp) $(cuda-ref.cpp) $(cuda-gen.cpp)
460  libceed.cu    += $(cuda-ref.cu) $(cuda-shared.cu) $(cuda-gen.cu)
461  BACKENDS_MAKE += $(CUDA_BACKENDS)
462endif
463
464# HIP Backends
465HIP_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(ROCM_DIR)/$d/libamdhip64.${SO_EXT}))
466HIP_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(HIP_LIB_DIR))))
467HIP_BACKENDS = /gpu/hip/ref /gpu/hip/shared /gpu/hip/gen
468ifneq ($(HIP_LIB_DIR),)
469  HIPCONFIG_CPPFLAGS := $(subst =,,$(shell $(ROCM_DIR)/bin/hipconfig -C))
470  $(hip-all.c:%.c=$(OBJDIR)/%.o) $(hip-all.c:%=%.tidy): CPPFLAGS += $(HIPCONFIG_CPPFLAGS)
471  ifneq ($(CXX), $(HIPCC))
472    $(hip-all.cpp:%.cpp=$(OBJDIR)/%.o) $(hip-all.cpp:%=%.tidy): CPPFLAGS += $(HIPCONFIG_CPPFLAGS)
473  endif
474  PKG_LIBS += -L$(abspath $(HIP_LIB_DIR)) -lamdhip64 -lhipblas
475  LIBCEED_CONTAINS_CXX = 1
476  libceed.c     += $(hip-all.c)
477  libceed.cpp   += $(hip-all.cpp)
478  libceed.hip   += $(hip-ref.hip)
479  BACKENDS_MAKE += $(HIP_BACKENDS)
480endif
481
482# SYCL Backends
483SYCL_BACKENDS = /gpu/sycl/ref /gpu/sycl/shared /gpu/sycl/gen
484ifneq ($(SYCL_DIR),)
485  SYCL_LIB_DIR := $(wildcard $(foreach d,lib lib64,$(SYCL_DIR)/$d/libsycl.${SO_EXT}))
486  SYCL_LIB_DIR := $(patsubst %/,%,$(dir $(firstword $(SYCL_LIB_DIR))))
487endif
488ifneq ($(SYCL_LIB_DIR),)
489  PKG_LIBS += $(SYCL_FLAG) -lze_loader
490  LIBCEED_CONTAINS_CXX = 1
491  libceed.sycl += $(sycl-core.cpp) $(sycl-ref.cpp) $(sycl-shared.cpp) $(sycl-gen.cpp)
492  BACKENDS_MAKE += $(SYCL_BACKENDS)
493endif
494
495# MAGMA Backends
496ifneq ($(wildcard $(MAGMA_DIR)/lib/libmagma.*),)
497  MAGMA_ARCH=$(shell nm -g $(MAGMA_DIR)/lib/libmagma.* | grep -c "hipblas")
498  ifeq ($(MAGMA_ARCH), 0)  # CUDA MAGMA
499    ifneq ($(CUDA_LIB_DIR),)
500      cuda_link = $(if $(STATIC),,-Wl,-rpath,$(CUDA_LIB_DIR)) -L$(CUDA_LIB_DIR) -lcublas -lcusparse -lcudart
501      omp_link = -fopenmp
502      magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(cuda_link) $(omp_link)
503      magma_link_shared = -L$(MAGMA_DIR)/lib $(if $(STATIC),,-Wl,-rpath,$(abspath $(MAGMA_DIR)/lib)) -lmagma
504      magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static))
505      PKG_LIBS += $(magma_link)
506      libceed.c   += $(magma.c)
507      libceed.cpp += $(magma.cpp)
508      $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include
509      $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : CPPFLAGS += -DADD_ -I$(MAGMA_DIR)/include -I$(CUDA_DIR)/include
510      MAGMA_BACKENDS = /gpu/cuda/magma /gpu/cuda/magma/det
511    endif
512  else  # HIP MAGMA
513    ifneq ($(HIP_LIB_DIR),)
514      omp_link = -fopenmp
515      hip_link = $(if $(STATIC),,-Wl,-rpath,$(HIP_LIB_DIR)) -L$(HIP_LIB_DIR) -lhipblas -lhipsparse -lamdhip64
516      magma_link_static = -L$(MAGMA_DIR)/lib -lmagma $(hip_link) $(omp_link)
517      magma_link_shared = -L$(MAGMA_DIR)/lib $(hip_link) $(omp_link) $(if $(STATIC),,-Wl,-rpath,$(abspath $(MAGMA_DIR)/lib)) -lmagma
518      magma_link := $(if $(wildcard $(MAGMA_DIR)/lib/libmagma.${SO_EXT}),$(magma_link_shared),$(magma_link_static))
519      PKG_LIBS += $(magma_link)
520      libceed.c   += $(magma.c)
521      libceed.cpp += $(magma.cpp)
522      $(magma.c:%.c=$(OBJDIR)/%.o) $(magma.c:%=%.tidy) : CPPFLAGS += $(HIPCONFIG_CPPFLAGS) -I$(MAGMA_DIR)/include -I$(ROCM_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_
523      $(magma.cpp:%.cpp=$(OBJDIR)/%.o) $(magma.cpp:%=%.tidy) : CPPFLAGS += $(HIPCONFIG_CPPFLAGS) -I$(MAGMA_DIR)/include -I$(ROCM_DIR)/include -DCEED_MAGMA_USE_HIP -DADD_
524      MAGMA_BACKENDS = /gpu/hip/magma /gpu/hip/magma/det
525    endif
526  endif
527  LIBCEED_CONTAINS_CXX = 1
528  BACKENDS_MAKE += $(MAGMA_BACKENDS)
529endif
530
531BACKENDS ?= $(BACKENDS_MAKE)
532export BACKENDS
533
534_pkg_ldflags = $(filter -L%,$(PKG_LIBS))
535_pkg_ldlibs = $(filter-out -L%,$(PKG_LIBS))
536$(libceeds) : CEED_LDFLAGS += $(_pkg_ldflags) $(if $(STATIC),,$(_pkg_ldflags:-L%=-Wl,-rpath,%)) $(PKG_STUBS_LIBS)
537$(libceeds) : CEED_LDLIBS += $(_pkg_ldlibs)
538ifeq ($(STATIC),1)
539  $(examples) $(tests) : CEED_LDFLAGS += $(EM_LDFLAGS) $(_pkg_ldflags) $(if $(STATIC),,$(_pkg_ldflags:-L%=-Wl,-rpath,%)) $(PKG_STUBS_LIBS)
540  $(examples) $(tests) : CEED_LDLIBS += $(_pkg_ldlibs)
541endif
542
543pkgconfig-libs-private = $(PKG_LIBS)
544ifeq ($(LIBCEED_CONTAINS_CXX),1)
545  $(libceeds) : LINK = $(CXX)
546  ifeq ($(STATIC),1)
547    $(examples) $(tests) : CEED_LDLIBS += $(LIBCXX)
548    pkgconfig-libs-private += $(LIBCXX)
549  endif
550endif
551
552# File names *-weak.c contain weak symbol definitions, which must be listed last
553# when creating shared or static libraries.
554weak_last = $(filter-out %-weak.o,$(1)) $(filter %-weak.o,$(1))
555
556libceed.o = $(libceed.c:%.c=$(OBJDIR)/%.o) $(libceed.cpp:%.cpp=$(OBJDIR)/%.o) $(libceed.cu:%.cu=$(OBJDIR)/%.o) $(libceed.hip:%.hip.cpp=$(OBJDIR)/%.o) $(libceed.sycl:%.sycl.cpp=$(OBJDIR)/%.o)
557$(filter %fortran.o,$(libceed.o)) : CPPFLAGS += $(if $(filter 1,$(UNDERSCORE)),-DUNDERSCORE)
558$(libceed.o): | info-backends
559$(libceed.so) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR
560	$(call quiet,LINK) $(LDFLAGS) $(CEED_LDFLAGS) -shared -o $@ $^ $(CEED_LDLIBS) $(LDLIBS)
561
562$(libceed.a) : $(call weak_last,$(libceed.o)) | $$(@D)/.DIR
563	$(call quiet,AR) $(ARFLAGS) $@ $^
564
565$(OBJDIR)/%.o : $(CURDIR)/%.c | $$(@D)/.DIR
566	$(call quiet,CC) $(CPPFLAGS) $(CFLAGS) -c -o $@ $(abspath $<)
567
568$(OBJDIR)/%.o : $(CURDIR)/%.cpp | $$(@D)/.DIR
569	$(call quiet,CXX) $(CPPFLAGS) $(CXXFLAGS) -c -o $@ $(abspath $<)
570
571$(OBJDIR)/%.o : $(CURDIR)/%.cu | $$(@D)/.DIR
572	$(call quiet,NVCC) $(filter-out -Wno-unused-function, $(CPPFLAGS)) $(NVCCFLAGS) -c -o $@ $(abspath $<)
573
574$(OBJDIR)/%.o : $(CURDIR)/%.hip.cpp | $$(@D)/.DIR
575	$(call quiet,HIPCC) $(CPPFLAGS) $(HIPCCFLAGS) -c -o $@ $(abspath $<)
576
577$(OBJDIR)/%.o : $(CURDIR)/%.sycl.cpp | $$(@D)/.DIR
578	$(call quiet,SYCLCXX) $(SYCLFLAGS) $(CPPFLAGS) -c -o $@ $(abspath $<)
579
580$(OBJDIR)/%.o : $(CURDIR)/%.sycl.cpp | $$(@D)/.DIR
581	$(call quiet,SYCLCXX) $(SYCLFLAGS) $(CPPFLAGS) -c -o $@ $(abspath $<)
582
583$(OBJDIR)/%$(EXE_SUFFIX) : tests/%.c | $$(@D)/.DIR
584	$(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS)
585
586$(OBJDIR)/%$(EXE_SUFFIX) : tests/%.f90 | $$(@D)/.DIR
587	$(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS)
588
589$(OBJDIR)/%$(EXE_SUFFIX) : examples/ceed/%.c | $$(@D)/.DIR
590	$(call quiet,LINK.c) $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS)
591
592$(OBJDIR)/%$(EXE_SUFFIX) : examples/ceed/%.f | $$(@D)/.DIR
593	$(call quiet,LINK.F) -DSOURCE_DIR='"$(abspath $(<D))/"' $(CEED_LDFLAGS) -o $@ $(abspath $<) $(CEED_LIBS) $(CEED_LDLIBS) $(LDLIBS)
594
595$(OBJDIR)/mfem-% : examples/mfem/%.cpp $(libceed) | $$(@D)/.DIR
596	+$(MAKE) -C examples/mfem CEED_DIR=`pwd` \
597	  MFEM_DIR="$(abspath $(MFEM_DIR))" CXX=$(CXX) $*
598	cp examples/mfem/$* $@
599
600# Note: Multiple Nek files cannot be built in parallel. The '+' here enables
601#       this single Nek bps file to be built in parallel with other examples,
602#       such as when calling `make prove-all -j2`.
603$(OBJDIR)/nek-bps : examples/nek/bps/bps.usr examples/nek/nek-examples.sh $(libceed) | $$(@D)/.DIR
604	+$(MAKE) -C examples MPI=$(MPI) CEED_DIR=`pwd` NEK5K_DIR="$(abspath $(NEK5K_DIR))" nek
605	mv examples/nek/build/bps $(OBJDIR)/bps
606	cp examples/nek/nek-examples.sh $(OBJDIR)/nek-bps
607
608# Note: Invoking deal.II's CMAKE build system here
609$(OBJDIR)/dealii-bps : examples/deal.II/*.cc examples/deal.II/*.h $(libceed) | $$(@D)/.DIR
610	mkdir -p examples/deal.II/build
611	cmake -B examples/deal.II/build -S examples/deal.II -DDEAL_II_DIR=$(DEAL_II_DIR) -DCEED_DIR=$(PWD)
612	+$(call quiet,MAKE) -C examples/deal.II/build
613	cp examples/deal.II/build/bps $(OBJDIR)/dealii-bps
614
615# Several executables have common utilities, but we can't build the utilities
616# from separate submake invocations because they'll compete with each
617# other/corrupt output. So we put it in this utility library, but we don't want
618# to manually list source dependencies up at this level, so we'll just always
619# call recursive make to check that this utility is up to date.
620examples/petsc/libutils.a.PHONY: $(libceed) $(ceed.pc)
621	+$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` AR=$(AR) ARFLAGS=$(ARFLAGS) \
622	  PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $(basename $(@F))
623
624$(OBJDIR)/petsc-% : examples/petsc/%.c examples/petsc/libutils.a.PHONY $(libceed) $(ceed.pc) | $$(@D)/.DIR
625	+$(call quiet,MAKE) -C examples/petsc CEED_DIR=`pwd` \
626	  PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $*
627	cp examples/petsc/$* $@
628
629$(OBJDIR)/fluids-% : examples/fluids/%.c examples/fluids/src/*.c examples/fluids/*.h examples/fluids/problems/*.c examples/fluids/problems/torch/*.cpp examples/fluids/qfunctions/*.h examples/fluids/src/smartsim/*.c $(libceed) $(ceed.pc) examples/fluids/Makefile | $$(@D)/.DIR
630	+$(call quiet,MAKE) -C examples/fluids CEED_DIR=`pwd` \
631	  PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $*
632	cp examples/fluids/$* $@
633
634$(OBJDIR)/fluids-py-% : examples/fluids/%.py $(OBJDIR)/fluids-navierstokes
635	cp $< $@
636
637$(OBJDIR)/solids-% : examples/solids/%.c examples/solids/%.h \
638    examples/solids/problems/*.c examples/solids/src/*.c \
639    examples/solids/include/*.h examples/solids/problems/*.h examples/solids/qfunctions/*.h \
640    $(libceed) $(ceed.pc) | $$(@D)/.DIR
641	+$(call quiet,MAKE) -C examples/solids CEED_DIR=`pwd` \
642	  PETSC_DIR="$(abspath $(PETSC_DIR))" OPT="$(OPT)" $*
643	cp examples/solids/$* $@
644
645$(examples) : $(libceed)
646$(tests) : $(libceed)
647$(tests) $(examples) : override LDFLAGS += $(if $(STATIC),,-Wl,-rpath,$(abspath $(LIBDIR))) -L$(LIBDIR)
648
649# Set number processes for testing
650NPROC_TEST ?= 1
651export NPROC_TEST
652
653# Set pool size for testing
654NPROC_POOL ?= 1
655export NPROC_POOL
656
657run-% : $(OBJDIR)/%
658	@$(PYTHON) tests/junit.py --mode tap --ceed-backends $(BACKENDS) $(if $(SMARTREDIS_DIR),--smartredis_dir $(SMARTREDIS_DIR) ) $(if $(USE_TORCH),--has_torch $(USE_TORCH) )--nproc $(NPROC_TEST) --pool-size $(NPROC_POOL) $(<:$(OBJDIR)/%=%)
659
660external_examples := \
661	$(if $(MFEM_DIR),$(mfemexamples)) \
662	$(if $(PETSC_DIR),$(petscexamples)) \
663	$(if $(NEK5K_DIR),$(nekexamples)) \
664	$(if $(DEAL_II_DIR),$(dealiiexamples)) \
665	$(if $(PETSC_DIR),$(fluidsexamples)) \
666	$(if $(PETSC_DIR),$(solidsexamples))
667
668allexamples = $(examples) $(external_examples)
669
670# The test and prove targets can be controlled via pattern searches.  The
671# default is to run tests and those examples that have no external dependencies.
672# Examples of finer grained control:
673#
674#   make test search='petsc mfem'      # PETSc and MFEM examples
675#   make prove search='t3'             # t3xx series tests
676#   make junit search='ex petsc'       # core ex* and PETSc tests
677search ?= t ex
678realsearch = $(search:%=%%)
679matched = $(foreach pattern,$(realsearch),$(filter $(OBJDIR)/$(pattern),$(tests) $(allexamples)))
680JUNIT_BATCH ?= ''
681
682# Test core libCEED
683test : $(matched:$(OBJDIR)/%=run-%)
684
685# Run test target in parallel
686tst : ;@$(MAKE) $(MFLAGS) V=$(V) test
687# CPU C tests only for backend %
688ctc-% : $(ctests);@$(foreach tst,$(ctests),$(tst) /cpu/$*;)
689
690prove : $(matched)
691	$(info Testing backends: $(BACKENDS))
692	$(PROVE) $(PROVE_OPTS) --exec '$(PYTHON) tests/junit.py --mode tap --ceed-backends $(BACKENDS) $(if $(SMARTREDIS_DIR),--smartredis_dir $(SMARTREDIS_DIR) ) $(if $(USE_TORCH),--has_torch $(USE_TORCH) )--nproc $(NPROC_TEST) --pool-size $(NPROC_POOL)' $(matched:$(OBJDIR)/%=%)
693# Run prove target in parallel
694prv : ;@$(MAKE) $(MFLAGS) V=$(V) prove
695
696prove-all :
697	+$(MAKE) prove realsearch=%
698
699junit-% : $(OBJDIR)/%
700	@printf "  %10s %s\n" TEST $(<:$(OBJDIR)/%=%); $(PYTHON) tests/junit.py --ceed-backends $(BACKENDS) $(if $(SMARTREDIS_DIR),--smartredis_dir $(SMARTREDIS_DIR) ) $(if $(USE_TORCH),--has_torch $(USE_TORCH) )--nproc $(NPROC_TEST) --pool-size $(NPROC_POOL) --junit-batch $(JUNIT_BATCH) $(<:$(OBJDIR)/%=%)
701
702junit : $(matched:$(OBJDIR)/%=junit-%)
703
704all: $(alltests)
705
706examples : $(allexamples)
707ceedexamples : $(examples)
708nekexamples : $(nekexamples)
709mfemexamples : $(mfemexamples)
710petscexamples : $(petscexamples)
711
712# Benchmarks
713allbenchmarks = petsc-bps
714bench_targets = $(addprefix bench-,$(allbenchmarks))
715.PHONY: $(bench_targets) benchmarks
716$(bench_targets): bench-%: $(OBJDIR)/%
717	cd benchmarks && ./benchmark.sh --ceed "$(BACKENDS_MAKE)" -r $(*).sh
718benchmarks: $(bench_targets)
719
720$(ceed.pc) : pkgconfig-prefix = $(abspath .)
721$(OBJDIR)/ceed.pc : pkgconfig-prefix = $(prefix)
722.INTERMEDIATE : $(OBJDIR)/ceed.pc
723%/ceed.pc : ceed.pc.template | $$(@D)/.DIR
724	@$(SED) \
725	    -e "s:%prefix%:$(pkgconfig-prefix):" \
726	    -e "s:%libs_private%:$(pkgconfig-libs-private):" $< > $@
727
728$(OBJDIR)/interface/ceed-jit-source-root-default.o : CPPFLAGS += -DCEED_JIT_SOURCE_ROOT_DEFAULT="\"$(abspath ./include)/\""
729$(OBJDIR)/interface/ceed-jit-source-root-install.o : CPPFLAGS += -DCEED_JIT_SOURCE_ROOT_DEFAULT="\"$(abspath $(includedir))/\""
730
731install : $(libceed) $(OBJDIR)/ceed.pc
732	$(INSTALL) -d $(addprefix $(if $(DESTDIR),"$(DESTDIR)"),"$(includedir)"\
733	  "$(includedir)/ceed/" "$(includedir)/ceed/jit-source/"\
734	  "$(includedir)/ceed/jit-source/cuda/" "$(includedir)/ceed/jit-source/hip/"\
735	  "$(includedir)/ceed/jit-source/gallery/" "$(includedir)/ceed/jit-source/magma/"\
736	  "$(includedir)/ceed/jit-source/sycl/" "$(libdir)" "$(pkgconfigdir)")
737	$(INSTALL_DATA) include/ceed/ceed.h "$(DESTDIR)$(includedir)/ceed/"
738	$(INSTALL_DATA) include/ceed/types.h "$(DESTDIR)$(includedir)/ceed/"
739	$(INSTALL_DATA) include/ceed/ceed-f32.h "$(DESTDIR)$(includedir)/ceed/"
740	$(INSTALL_DATA) include/ceed/ceed-f64.h "$(DESTDIR)$(includedir)/ceed/"
741	$(INSTALL_DATA) include/ceed/fortran.h "$(DESTDIR)$(includedir)/ceed/"
742	$(INSTALL_DATA) include/ceed/backend.h "$(DESTDIR)$(includedir)/ceed/"
743	$(INSTALL_DATA) include/ceed/cuda.h "$(DESTDIR)$(includedir)/ceed/"
744	$(INSTALL_DATA) include/ceed/hip.h "$(DESTDIR)$(includedir)/ceed/"
745	$(INSTALL_DATA) $(libceed) "$(DESTDIR)$(libdir)/"
746	$(INSTALL_DATA) $(OBJDIR)/ceed.pc "$(DESTDIR)$(pkgconfigdir)/"
747	$(INSTALL_DATA) include/ceed.h "$(DESTDIR)$(includedir)/"
748	$(INSTALL_DATA) include/ceedf.h "$(DESTDIR)$(includedir)/"
749	$(INSTALL_DATA) $(wildcard include/ceed/jit-source/cuda/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/cuda/"
750	$(INSTALL_DATA) $(wildcard include/ceed/jit-source/hip/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/hip/"
751	$(INSTALL_DATA) $(wildcard include/ceed/jit-source/gallery/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/gallery/"
752	$(INSTALL_DATA) $(wildcard include/ceed/jit-source/magma/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/magma/"
753	$(INSTALL_DATA) $(wildcard include/ceed/jit-source/sycl/*.h) "$(DESTDIR)$(includedir)/ceed/jit-source/sycl/"
754
755.PHONY : all cln clean doxygen doc format lib install par print test tst prove prv prove-all junit examples tidy iwyu info info-backends info-backends-all
756
757cln clean :
758	$(RM) -r $(OBJDIR) $(LIBDIR) dist *egg* .pytest_cache *cffi*
759	$(call quiet,MAKE) -C examples clean NEK5K_DIR="$(abspath $(NEK5K_DIR))"
760	$(call quiet,MAKE) -C python/tests clean
761	$(RM) benchmarks/*output.txt
762
763distclean : clean
764	$(RM) -r doc/html doc/sphinx/build $(CONFIG)
765
766# Documentation
767DOXYGEN ?= doxygen
768
769doxygen :
770	$(DOXYGEN) -q Doxyfile
771
772doc-html doc-latexpdf doc-epub doc-livehtml : doc-% : doxygen
773	make -C doc/sphinx $*
774
775doc : doc-html
776
777# Style/Format
778CLANG_FORMAT      ?= clang-format
779CLANG_FORMAT_OPTS += -style=file -i
780AUTOPEP8          ?= autopep8
781AUTOPEP8_OPTS     += --in-place --aggressive --max-line-length 120
782
783format.ch := $(filter-out include/ceedf.h $(wildcard tests/t*-f.h), $(shell git ls-files '*.[ch]pp' '*.[ch]'))
784format.py := $(filter-out tests/junit-xml/junit_xml/__init__.py, $(shell git ls-files '*.py'))
785format.ot := $(filter-out doc/sphinx/source/CODE_OF_CONDUCT.md doc/sphinx/source/CONTRIBUTING.md, $(shell git ls-files '*.md' '*.f90'))
786
787format-c  :
788	$(CLANG_FORMAT) $(CLANG_FORMAT_OPTS) $(format.ch)
789
790format-py :
791	$(AUTOPEP8) $(AUTOPEP8_OPTS) $(format.py)
792
793format-ot:
794	@$(SED) -r 's/\s+$$//' -i $(format.ot)
795
796format    : format-c format-py format-ot
797
798# Vermin - python version requirements
799VERMIN            ?= vermin
800VERMIN_OPTS       += -t=3.7- --violations
801
802vermin    :
803	$(VERMIN) $(VERMIN_OPTS) $(format.py)
804
805# Tidy
806CLANG_TIDY ?= clang-tidy
807
808%.c.tidy : %.c
809	$(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c99 -I$(CUDA_DIR)/include -I$(ROCM_DIR)/include -DCEED_JIT_SOURCE_ROOT_DEFAULT="\"$(abspath ./include)/\""
810
811%.cpp.tidy : %.cpp
812	$(CLANG_TIDY) $(TIDY_OPTS) $^ -- $(CPPFLAGS) --std=c++11 -I$(CUDA_DIR)/include -I$(OCCA_DIR)/include -I$(ROCM_DIR)/include
813
814tidy-c   : $(libceed.c:%=%.tidy)
815tidy-cpp : $(libceed.cpp:%=%.tidy)
816
817tidy : tidy-c tidy-cpp
818
819# Include-What-You-Use
820ifneq ($(wildcard ../iwyu/*),)
821  IWYU_DIR ?= ../iwyu
822  IWYU_CC  ?= $(IWYU_DIR)/build/bin/include-what-you-use
823endif
824iwyu :
825	$(MAKE) -B CC=$(IWYU_CC)
826
827print :
828	@echo $(VAR)=$($(VAR))
829
830print-% :
831	$(info [ variable name]: $*)
832	$(info [        origin]: $(origin $*))
833	$(info [        flavor]: $(flavor $*))
834	$(info [         value]: $(value $*))
835	$(info [expanded value]: $($*))
836	$(info )
837	@true
838
839# "make configure" detects any variables passed on the command line or
840# previously set in config.mk, caching them in config.mk as simple
841# (:=) variables.  Variables set in config.mk or on the command line
842# take precedence over the defaults provided in the file.  Typical
843# usage:
844#
845#   make configure CC=/path/to/my/cc CUDA_DIR=/opt/cuda
846#   make
847#   make prove
848#
849# The values in the file can be updated by passing them on the command
850# line, e.g.,
851#
852#   make configure CC=/path/to/other/clang
853
854# All variables to consider for caching
855CONFIG_VARS = CC CXX FC NVCC NVCC_CXX HIPCC \
856  OPT CFLAGS CPPFLAGS CXXFLAGS FFLAGS NVCCFLAGS HIPCCFLAGS SYCLFLAGS \
857  AR ARFLAGS LDFLAGS LDLIBS LIBCXX SED \
858  MAGMA_DIR OCCA_DIR XSMM_DIR CUDA_DIR CUDA_ARCH MFEM_DIR PETSC_DIR NEK5K_DIR ROCM_DIR HIP_ARCH SYCL_DIR SMARTREDIS_DIR USE_TORCH
859
860# $(call needs_save,CFLAGS) returns true (a nonempty string) if CFLAGS
861# was set on the command line or in config.mk (where it will appear as
862# a simple variable).
863needs_save = $(or $(filter command line,$(origin $(1))),$(filter simple,$(flavor $(1))))
864
865configure :
866	$(file > $(CONFIG))
867	$(foreach v,$(CONFIG_VARS),$(if $(call needs_save,$(v)),$(file >> $(CONFIG),$(v) := $($(v)))))
868	@echo "Configuration cached in $(CONFIG):"
869	@cat $(CONFIG)
870
871wheel : export MARCHFLAG = -march=generic
872wheel : export WHEEL_PLAT = manylinux2010_x86_64
873wheel :
874	docker run -it --user $(shell id -u):$(shell id -g) --rm -v $(PWD):/io -w /io \
875	  -e MARCHFLAG -e WHEEL_PLAT \
876	  quay.io/pypa/$(WHEEL_PLAT) python/make-wheels.sh
877
878.PHONY : configure wheel
879
880# Include *.d deps when not -B = --always-make: useful if the paths are wonky in a container
881-include $(if $(filter B,$(MAKEFLAGS)),,$(libceed.c:%.c=$(OBJDIR)/%.d) $(tests.c:tests/%.c=$(OBJDIR)/%.d))
882