scs 0.5.6 → 0.6.0

data/vendor/scs/linsys/scs_matrix.h

@@ -1,3 +1,8 @@
+/*
+ * Sparse matrix operations: normalization/equilibration, copying, validation,
+ * and matrix-vector products (A*x, A'*x, P*x) for CSC-format matrices.
+ */
+
 #ifndef SCS_MATRIX_H_GUARD
 #define SCS_MATRIX_H_GUARD
 
@@ -6,19 +11,14 @@ extern "C" {
 #endif
 
 #include "glbopts.h"
-#include "scs.h"
 #include "scs_work.h"
 
 /* Normalization routines, used if d->NORMALIZE is true */
 /* normalizes A matrix, sets scal->E and scal->D diagonal scaling matrices,
  * A -> D*A*E. D and E must be all positive entries, D must satisfy cone
  * boundaries */
-ScsScaling *SCS(normalize_a_p)(ScsMatrix *P, ScsMatrix *A, ScsConeWork *cone);
-
-/* unnormalizes A matrix, unnormalizes by w->D and w->E */
-/* void SCS(un_normalize_a_p)(ScsMatrix *A, ScsMatrix *P, const ScsScaling
- * *scal);
- */
+ScsScaling *SCS(normalize_a_p)(ScsMatrix *P, ScsMatrix *A, const scs_float *b,
+                               const scs_float *c, ScsConeWork *cone);
 
 /* to free the memory allocated in a ScsMatrix (called on A and P at finish) */
 void SCS(free_scs_matrix)(ScsMatrix *A);
@@ -27,8 +27,6 @@ void SCS(free_scs_matrix)(ScsMatrix *A);
  * allocates memory for dstp	*/
 scs_int SCS(copy_matrix)(ScsMatrix **dstp, const ScsMatrix *src);
 
-scs_float SCS(cumsum)(scs_int *p, scs_int *c, scs_int n);
-
 /**
  * Validate the linear system inputs, returns < 0 if not valid inputs.
  *

data/vendor/scs/scs.mk

@@ -4,10 +4,8 @@ else
 UNAME = $(shell uname -s)
 endif
 
-# CC = clang -fsanitize=address -fsanitize=undefined -fsanitize=float-divide-by-zero -fsanitize=float-cast-overflow
-# For cross-compiling with mingw use these.
-#CC = i686-w64-mingw32-gcc -m32
-#CC = x86_64-w64-mingw32-gcc-4.8
+# To enable sanitizers or cross-compile, override CC on the command line:
+#   make CC="clang -fsanitize=address,undefined"
 
 # For GPU must add cuda libs to path, e.g.
 # export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:$DYLD_LIBRARY_PATH
@@ -30,26 +28,37 @@ endif
 endif
 endif
 
+# ABI version embedded in the shared-library soname/install_name, derived from
+# SCS_VERSION (keep project() in CMakeLists.txt in step). Bump on ABI breaks,
+# never on patch releases.
+SCS_SOVERSION := $(shell awk -F'"' '/define SCS_VERSION/ {split($$2, v, "."); print v[1] "." v[2]}' include/glbopts.h)
+
+# versioned(name): the file a shared library is linked as; the unversioned
+# name becomes a symlink to it (libscsdir.so.3.3, libscsdir.3.3.dylib).
+# Windows DLLs stay unversioned.
 ifeq ($(UNAME), Darwin)
 # we're on apple, no need to link rt library
 LDFLAGS += -lm
 SHARED = dylib
 SONAME = -install_name
+versioned = $(1:%.dylib=%.$(SCS_SOVERSION).dylib)
 else
 ifeq ($(ISWINDOWS), 1)
 # we're on windows (cygwin or msys)
 LDFLAGS += -lm
 SHARED = dll
 SONAME = -soname
+versioned = $(1)
 else
 # we're on a linux system, use accurate timer provided by clock_gettime()
-LDFLAGS += -lm -lrt
+LDFLAGS += -lm -lrt -lpthread
 SHARED = so
 SONAME = -soname
+versioned = $(1:%.so=%.so.$(SCS_SOVERSION))
 endif
 endif
 
-#TODO: check if this works for all platforms:
+# Default CUDA path; override with CUDA_PATH=/path/to/cuda
 ifeq ($(CUDA_PATH), )
 CUDA_PATH=/usr/local/cuda
 CUCC = $(CUDA_PATH)/bin/nvcc
@@ -61,8 +70,9 @@ CUDAFLAGS = $(CFLAGS) -I$(CUDA_PATH)/include -Ilinsys/gpu -Wno-c++11-long-long #
 CUDSS_FLAGS = -I$(CUDSS_PATH)/include -I$(CUDA_PATH)/include
 CUDSS_LDFLAGS = $(CULDFLAGS) -L$(CUDSS_PATH)/lib -lcudss
 
-# Add on default CFLAGS
-OPT = -O3
+# Default optimization flags. Override with OPT="-O3 -g" for valgrind, which
+# may not support instruction sets enabled by -march=native.
+OPT ?= -O3 -march=native -fno-math-errno
 INCLUDE = -I. -Iinclude -Ilinsys
 override CFLAGS += -g -Wall -Wwrite-strings -pedantic -funroll-loops -Wstrict-prototypes $(INCLUDE) $(OPT) -Werror=incompatible-pointer-types
 ifneq ($(ISWINDOWS), 1)
@@ -72,10 +82,12 @@ endif
 LINSYS = linsys
 DIRSRC = $(LINSYS)/cpu/direct
 INDIRSRC = $(LINSYS)/cpu/indirect
+DENSESRC = $(LINSYS)/cpu/dense
 GPUDIR = $(LINSYS)/gpu/direct
 GPUINDIR = $(LINSYS)/gpu/indirect
 MKLSRC = $(LINSYS)/mkl/direct
 CUDSSSRC = $(LINSYS)/cudss/direct
+ACCELSRC = $(LINSYS)/accelerate/direct
 
 EXTSRC = $(LINSYS)/external
 
@@ -124,7 +136,7 @@ CUSTOM_FLAGS += -DNO_PRINTING=$(NO_PRINTING) # disable printing
 endif
 NO_READ_WRITE = 0
 ifneq ($(NO_READ_WRITE), 0)
-CUSTOM_FLAGS += -DNO_READ_WRITE=$(NO_READ_WRITE) # disable printing
+CUSTOM_FLAGS += -DNO_READ_WRITE=$(NO_READ_WRITE) # disable read/write (file I/O)
 endif
 ### VERBOSITY LEVELS: 0,1,2,...
 VERBOSITY = 0
@@ -136,10 +148,20 @@ ifneq ($(COVERAGE), 0)
 CUSTOM_FLAGS += --coverage # generate test coverage data
 endif
 
-# See: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-link-line-advisor.html
-# This is probably not correct for other systems. TODO: update this
-# to work for all combinations of platform / compiler / threading options.
-MKLFLAGS = -L$(MKLROOT) -L$(MKLROOT)/lib -Wl,--no-as-needed -lmkl_rt -lmkl_gnu_thread -lmkl_core -lgomp -lpthread -ldl
+# MKL linker flags for Linux + GCC + GNU threading. For other platforms, see:
+# https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-link-line-advisor.html
+#
+# We link against mkl_rt (the single dynamic library) which dispatches to the
+# correct interface layer at runtime. The BLAS integer width (LP64 vs ILP64)
+# is set at startup by the MKL-specific path in src/scs.c, controlled by the
+# BLAS64 compile flag.
+#
+# Note: The PARDISO integer width (pardiso vs pardiso_64) is controlled by DLONG,
+# not by the MKL interface layer. See linsys/mkl/direct/private.c for details.
+# OMPROOT should point to the Intel OpenMP (iomp5) installation, typically
+# /opt/intel/oneapi/compiler/latest. If unset, fall back to MKLROOT.
+OMPROOT ?= $(MKLROOT)
+MKLFLAGS = -L$(MKLROOT)/lib -L$(MKLROOT)/lib/intel64 -L$(OMPROOT)/lib -Wl,--no-as-needed -lmkl_rt -liomp5 -lpthread -ldl
 
 ############ OPENMP: ############
 # set USE_OPENMP = 1 to allow openmp (multi-threaded matrix multiplies):