scs 0.4.0 → 0.4.1

data/vendor/scs/linsys/mkl/direct/private.h

@@ -0,0 +1,38 @@
+#ifndef PRIV_H_GUARD
+#define PRIV_H_GUARD
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "csparse.h"
+#include "linsys.h"
+
+struct SCS_LIN_SYS_WORK {
+  ScsMatrix *kkt; /* Upper triangular KKT matrix (in CSR format) */
+  scs_float *sol; /* solution to the KKT system */
+  scs_int n;      /* number of QP variables */
+  scs_int m;      /* number of QP constraints */
+
+  /* Pardiso variables */
+  void *pt[64];      /* internal solver memory pointer pt */
+  scs_int iparm[64]; /* Pardiso control parameters */
+  scs_int n_plus_m;  /* dimension of the linear system */
+  scs_int mtype;     /* matrix type (-2 for real and symmetric indefinite) */
+  scs_int nrhs;      /* number of right-hand sides (1) */
+  scs_int maxfct;    /* maximum number of factors (1) */
+  scs_int mnum;      /* indicates matrix for the solution phase (1) */
+  scs_int phase;     /* control the execution phases of the solver */
+  scs_int error;     /* the error indicator (0 for no error) */
+  scs_int msglvl;    /* Message level information (0 for no output) */
+
+  /* These are required for matrix updates */
+  scs_int *diag_r_idxs; /* indices where R appears */
+  scs_float *diag_p;    /* Diagonal values of P */
+};
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

data/vendor/scs/linsys/scs_matrix.c

@@ -117,6 +117,7 @@ static void compute_ruiz_mats(ScsMatrix *P, ScsMatrix *A, scs_float *Dt,
                               scs_float *Et, ScsConeWork *cone) {
   scs_int i, j, kk;
   scs_float wrk;
+  scs_float nm_a_col;
 
   /****************************  D  ****************************/
 
@@ -138,7 +139,8 @@ static void compute_ruiz_mats(ScsMatrix *P, ScsMatrix *A, scs_float *Dt,
 
   /* invert temporary vec to form D */
   for (i = 0; i < A->m; ++i) {
-    Dt[i] = SAFEDIV_POS(1.0, SQRTF(apply_limit(Dt[i])));
+    Dt[i] = SQRTF(apply_limit(Dt[i]));
+    Dt[i] = SAFEDIV_POS(1.0, Dt[i]);
   }
 
   /****************************  E  ****************************/
@@ -169,8 +171,10 @@ static void compute_ruiz_mats(ScsMatrix *P, ScsMatrix *A, scs_float *Dt,
 
   /* calculate col norms, E */
   for (i = 0; i < A->n; ++i) {
-    Et[i] = MAX(Et[i], SCS(norm_inf)(&(A->x[A->p[i]]), A->p[i + 1] - A->p[i]));
-    Et[i] = SAFEDIV_POS(1.0, SQRTF(apply_limit(Et[i])));
+    nm_a_col = SCS(norm_inf)(&(A->x[A->p[i]]), A->p[i + 1] - A->p[i]);
+    Et[i] = MAX(Et[i], nm_a_col);
+    Et[i] = SQRTF(apply_limit(Et[i]));
+    Et[i] = SAFEDIV_POS(1.0, Et[i]);
   }
 }
 
@@ -201,7 +205,8 @@ static void compute_l2_mats(ScsMatrix *P, ScsMatrix *A, scs_float *Dt,
   SCS(enforce_cone_boundaries)(cone, Dt, &SCS(mean));
 
   for (i = 0; i < A->m; ++i) {
-    Dt[i] = SAFEDIV_POS(1.0, SQRTF(apply_limit(Dt[i])));
+    Dt[i] = SQRTF(apply_limit(Dt[i]));
+    Dt[i] = SAFEDIV_POS(1.0, Dt[i]);
   }
 
   /****************************  E  ****************************/
@@ -233,7 +238,8 @@ static void compute_l2_mats(ScsMatrix *P, ScsMatrix *A, scs_float *Dt,
   /* calculate col norms, E */
   for (i = 0; i < A->n; ++i) {
     Et[i] += SCS(norm_sq)(&(A->x[A->p[i]]), A->p[i + 1] - A->p[i]);
-    Et[i] = SAFEDIV_POS(1.0, SQRTF(apply_limit(SQRTF(Et[i]))));
+    Et[i] = SQRTF(apply_limit(SQRTF(Et[i])));
+    Et[i] = SAFEDIV_POS(1.0, Et[i]);
   }
 }
 

data/vendor/scs/scs.mk

@@ -69,6 +69,7 @@ DIRSRC = $(LINSYS)/cpu/direct
 INDIRSRC = $(LINSYS)/cpu/indirect
 GPUDIR = $(LINSYS)/gpu/direct
 GPUINDIR = $(LINSYS)/gpu/indirect
+MKLSRC = $(LINSYS)/mkl/direct
 
 EXTSRC = $(LINSYS)/external
 
@@ -83,44 +84,56 @@ ifeq ($(PREFIX),)
   PREFIX = /usr/local
 endif
 
-OPT_FLAGS =
-########### OPTIONAL FLAGS ##########
+########### CUSTOM FLAGS ##########
 # these can all be override from the command line
+CUSTOM_FLAGS =
 # e.g. make DLONG=1 will override the setting below
 DLONG = 0
 ifneq ($(DLONG), 0)
-OPT_FLAGS += -DDLONG=$(DLONG) # use longs rather than ints
+CUSTOM_FLAGS += -DDLONG=$(DLONG) # use longs rather than ints
 endif
 CTRLC = 1
 ifneq ($(CTRLC), 0)
-OPT_FLAGS += -DCTRLC=$(CTRLC) # graceful interrupts with ctrl-c
+CUSTOM_FLAGS += -DCTRLC=$(CTRLC) # graceful interrupts with ctrl-c
 endif
 SFLOAT = 0
 ifneq ($(SFLOAT), 0)
-OPT_FLAGS += -DSFLOAT=$(SFLOAT) # use floats rather than doubles
-endif
-NOTIMER = 0
-ifneq ($(NOTIMER), 0)
-OPT_FLAGS += -DNOTIMER=$(NOTIMER) # no timing, times reported as nan
+CUSTOM_FLAGS += -DSFLOAT=$(SFLOAT) # use floats rather than doubles
 endif
 GPU_TRANSPOSE_MAT = 1
 ifneq ($(GPU_TRANSPOSE_MAT), 0)
-OPT_FLAGS += -DGPU_TRANSPOSE_MAT=$(GPU_TRANSPOSE_MAT) # tranpose A mat in GPU memory
+CUSTOM_FLAGS += -DGPU_TRANSPOSE_MAT=$(GPU_TRANSPOSE_MAT) # transpose A mat in GPU memory
+endif
+NO_TIMER = 0
+ifneq ($(NO_TIMER), 0)
+CUSTOM_FLAGS += -DNO_TIMER=$(NO_TIMER) # no timing, times reported as nan
+endif
+NO_VALIDATE = 0
+ifneq ($(NO_VALIDATE), 0)
+CUSTOM_FLAGS += -DNO_VALIDATE=$(NO_VALIDATE) # perform problem validation or skip
+endif
+NO_PRINTING = 0
+ifneq ($(NO_PRINTING), 0)
+CUSTOM_FLAGS += -DNO_PRINTING=$(NO_PRINTING) # disable printing
 endif
-NOVALIDATE = 0
-ifneq ($(NOVALIDATE), 0)
-OPT_FLAGS += -DNOVALIDATE=$(NOVALIDATE) # perform problem validation or skip
+NO_READ_WRITE = 0
+ifneq ($(NO_READ_WRITE), 0)
+CUSTOM_FLAGS += -DNO_READ_WRITE=$(NO_READ_WRITE) # disable printing
 endif
 ### VERBOSITY LEVELS: 0,1,2,...
 VERBOSITY = 0
 ifneq ($(VERBOSITY), 0)
-OPT_FLAGS += -DVERBOSITY=$(VERBOSITY) # verbosity level
+CUSTOM_FLAGS += -DVERBOSITY=$(VERBOSITY) # verbosity level
 endif
 COVERAGE = 0
 ifneq ($(COVERAGE), 0)
-override CFLAGS += --coverage # generate test coverage data
+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
 
 ############ OPENMP: ############
 # set USE_OPENMP = 1 to allow openmp (multi-threaded matrix multiplies):
@@ -138,44 +151,44 @@ endif
 # NB: point the libraries to the locations where
 # you have blas and lapack installed
 
+BLASLDFLAGS =
 USE_LAPACK = 1
 ifneq ($(USE_LAPACK), 0)
   # edit these for your setup:
-  BLASLDFLAGS = -llapack -lblas # -lgfortran
-  LDFLAGS += $(BLASLDFLAGS)
-  OPT_FLAGS += -DUSE_LAPACK
+  BLASLDFLAGS += -llapack -lblas # -lgfortran
+  CUSTOM_FLAGS += -DUSE_LAPACK
 
   BLAS64 = 0
   ifneq ($(BLAS64), 0)
-  OPT_FLAGS += -DBLAS64=$(BLAS64) # if blas/lapack lib uses 64 bit ints
+  CUSTOM_FLAGS += -DBLAS64=$(BLAS64) # if blas/lapack lib uses 64 bit ints
   endif
 
   NOBLASSUFFIX = 0
   ifneq ($(NOBLASSUFFIX), 0)
-  OPT_FLAGS += -DNOBLASSUFFIX=$(NOBLASSUFFIX) # hack to strip blas suffix
+  CUSTOM_FLAGS += -DNOBLASSUFFIX=$(NOBLASSUFFIX) # hack to strip blas suffix
   endif
 
   BLASSUFFIX = "_"
   ifneq ($(BLASSUFFIX), "_")
-  OPT_FLAGS += -DBLASSUFFIX=$(BLASSUFFIX) # blas suffix (underscore usually)
+  CUSTOM_FLAGS += -DBLASSUFFIX=$(BLASSUFFIX) # blas suffix (underscore usually)
   endif
 endif
 
 MATLAB_MEX_FILE = 0
 ifneq ($(MATLAB_MEX_FILE), 0)
-OPT_FLAGS += -DMATLAB_MEX_FILE=$(MATLAB_MEX_FILE) # matlab mex
+CUSTOM_FLAGS += -DMATLAB_MEX_FILE=$(MATLAB_MEX_FILE) # matlab mex
 endif
 PYTHON = 0
 ifneq ($(PYTHON), 0)
-OPT_FLAGS += -DPYTHON=$(PYTHON) # python extension
+CUSTOM_FLAGS += -DPYTHON=$(PYTHON) # python extension
 endif
 USING_R = 0
 ifneq ($(USING_R), 0)
-OPT_FLAGS += -DUSING_R=$(USING_R) # R extension
+CUSTOM_FLAGS += -DUSING_R=$(USING_R) # R extension
 endif
 
 # debug to see var values, e.g. 'make print-OBJECTS' shows OBJECTS value
 print-%: ; @echo $*=$($*)
 
-override CFLAGS += $(OPT_FLAGS)
-CUDAFLAGS += $(OPT_FLAGS)
+override CFLAGS += $(CUSTOM_FLAGS)
+CUDAFLAGS += $(CUSTOM_FLAGS)

data/vendor/scs/src/cones.c

@@ -4,10 +4,8 @@
 #include "scs_blas.h" /* contains BLAS(X) macros and type info */
 #include "util.h"
 
-#define CONE_TOL (1e-9)
-#define CONE_THRESH (1e-8)
-#define EXP_CONE_MAX_ITERS (100)
 #define BOX_CONE_MAX_ITERS (25)
+#define POW_CONE_TOL (1e-9)
 #define POW_CONE_MAX_ITERS (20)
 
 /* Box cone limits (+ or -) taken to be INF */
@@ -35,6 +33,11 @@ void BLAS(scal)(const blas_int *n, const scs_float *sa, scs_float *sx,
 
 #endif
 
+/* Forward declare exponential cone projection routine.
+ * Implemented in exp_cone.c.
+ */
+scs_float SCS(proj_pd_exp_cone)(scs_float *v0, scs_int primal);
+
 void SCS(free_cone)(ScsCone *k) {
   if (k) {
     if (k->bu)
@@ -285,7 +288,7 @@ void SCS(finish_cone)(ScsConeWork *c) {
 }
 
 char *SCS(get_cone_header)(const ScsCone *k) {
-  char *tmp = (char *)scs_malloc(sizeof(char) * 512);
+  char *tmp = (char *)scs_malloc(512); /* sizeof(char) = 1 */
   scs_int i, soc_vars, sd_vars;
   sprintf(tmp, "cones: ");
   if (k->z) {
@@ -325,118 +328,6 @@ char *SCS(get_cone_header)(const ScsCone *k) {
   return tmp;
 }
 
-static scs_float exp_newton_one_d(scs_float rho, scs_float y_hat,
-                                  scs_float z_hat, scs_float w) {
-  scs_float t_prev, t = MAX(w - z_hat, MAX(-z_hat, 1e-9));
-  scs_float f = 1., fp = 1.;
-  scs_int i;
-  for (i = 0; i < EXP_CONE_MAX_ITERS; ++i) {
-    t_prev = t;
-    f = t * (t + z_hat) / rho / rho - y_hat / rho + log(t / rho) + 1;
-    fp = (2 * t + z_hat) / rho / rho + 1 / t;
-
-    t = t - f / fp;
-
-    if (t <= -z_hat) {
-      t = -z_hat;
-      break;
-    } else if (t <= 0) {
-      t = 0;
-      break;
-    } else if (ABS(t - t_prev) < CONE_TOL) {
-      break;
-    } else if (SQRTF(f * f / fp) < CONE_TOL) {
-      break;
-    }
-  }
-  if (i == EXP_CONE_MAX_ITERS) {
-    scs_printf("warning: exp cone newton step hit maximum %i iters\n", (int)i);
-    scs_printf("rho=%1.5e; y_hat=%1.5e; z_hat=%1.5e; w=%1.5e; f=%1.5e, "
-               "fp=%1.5e, t=%1.5e, t_prev= %1.5e\n",
-               rho, y_hat, z_hat, w, f, fp, t, t_prev);
-  }
-  return t + z_hat;
-}
-
-static void exp_solve_for_x_with_rho(const scs_float *v, scs_float *x,
-                                     scs_float rho, scs_float w) {
-  x[2] = exp_newton_one_d(rho, v[1], v[2], w);
-  x[1] = (x[2] - v[2]) * x[2] / rho;
-  x[0] = v[0] - rho;
-}
-
-static scs_float exp_calc_grad(const scs_float *v, scs_float *x, scs_float rho,
-                               scs_float w) {
-  exp_solve_for_x_with_rho(v, x, rho, w);
-  if (x[1] <= 1e-12) {
-    return x[0];
-  }
-  return x[0] + x[1] * log(x[1] / x[2]);
-}
-
-static void exp_get_rho_ub(const scs_float *v, scs_float *x, scs_float *ub,
-                           scs_float *lb) {
-  *lb = 0;
-  *ub = 0.125;
-  while (exp_calc_grad(v, x, *ub, v[1]) > 0) {
-    *lb = *ub;
-    (*ub) *= 2;
-  }
-}
-
-/* project onto the exponential cone, v has dimension *exactly* 3 */
-static scs_int proj_exp_cone(scs_float *v) {
-  scs_int i;
-  scs_float ub, lb, rho, g, x[3];
-  scs_float r = v[0], s = v[1], t = v[2];
-
-  /* v in cl(Kexp) */
-  if ((s * exp(r / s) - t <= CONE_THRESH && s > 0) ||
-      (r <= 0 && s == 0 && t >= 0)) {
-    return 0;
-  }
-
-  /* -v in Kexp^* */
-  if ((r > 0 && r * exp(s / r) + exp(1) * t <= CONE_THRESH) ||
-      (r == 0 && s <= 0 && t <= 0)) {
-    memset(v, 0, 3 * sizeof(scs_float));
-    return 0;
-  }
-
-  /* special case with analytical solution */
-  if (r < 0 && s < 0) {
-    v[1] = 0.0;
-    v[2] = MAX(v[2], 0);
-    return 0;
-  }
-
-  /* iterative procedure to find projection, bisects on dual variable: */
-  exp_get_rho_ub(v, x, &ub, &lb); /* get starting upper and lower bounds */
-  for (i = 0; i < EXP_CONE_MAX_ITERS; ++i) {
-    rho = (ub + lb) / 2; /* halfway between upper and lower bounds */
-    g = exp_calc_grad(v, x, rho, x[1]); /* calculates gradient wrt dual var */
-    if (g > 0) {
-      lb = rho;
-    } else {
-      ub = rho;
-    }
-    if (ub - lb < CONE_TOL) {
-      break;
-    }
-  }
-#if VERBOSITY > 10
-  scs_printf("exponential cone proj iters %i\n", (int)i);
-#endif
-  if (i == EXP_CONE_MAX_ITERS) {
-    scs_printf("warning: exp cone outer step hit maximum %i iters\n", (int)i);
-    scs_printf("r=%1.5e; s=%1.5e; t=%1.5e\n", r, s, t);
-  }
-  v[0] = x[0];
-  v[1] = x[1];
-  v[2] = x[2];
-  return 0;
-}
-
 static scs_int set_up_sd_cone_work_space(ScsConeWork *c, const ScsCone *k) {
   scs_int i;
 #ifdef USE_LAPACK
@@ -741,13 +632,13 @@ static void proj_power_cone(scs_float *v, scs_float a) {
   scs_int i;
   /* v in K_a */
   if (xh >= 0 && yh >= 0 &&
-      CONE_THRESH + POWF(xh, a) * POWF(yh, (1 - a)) >= rh) {
+      POW_CONE_TOL + POWF(xh, a) * POWF(yh, (1 - a)) >= rh) {
     return;
   }
 
   /* -v in K_a^* */
   if (xh <= 0 && yh <= 0 &&
-      CONE_THRESH + POWF(-xh, a) * POWF(-yh, 1 - a) >=
+      POW_CONE_TOL + POWF(-xh, a) * POWF(-yh, 1 - a) >=
           rh * POWF(a, a) * POWF(1 - a, 1 - a)) {
     v[0] = v[1] = v[2] = 0;
     return;
@@ -760,7 +651,7 @@ static void proj_power_cone(scs_float *v, scs_float a) {
     y = pow_calc_x(r, yh, rh, 1 - a);
 
     f = pow_calc_f(x, y, r, a);
-    if (ABS(f) < CONE_TOL) {
+    if (ABS(f) < POW_CONE_TOL) {
       break;
     }
 
@@ -829,51 +720,16 @@ static scs_int proj_cone(scs_float *x, const ScsCone *k, ScsConeWork *c,
     }
   }
 
-  if (k->ep) { /* doesn't use r_y */
-               /*
-                * exponential cone is not self dual, if s \in K
-                * then y \in K^* and so if K is the primal cone
-                * here we project onto K^*, via Moreau
-                * \Pi_C^*(y) = y + \Pi_C(-y)
-                */
+  if (k->ep || k->ed) { /* doesn't use r_y */
 #ifdef _OPENMP
 #pragma omp parallel for
 #endif
-    for (i = 0; i < k->ep; ++i) {
-      proj_exp_cone(&(x[count + 3 * i]));
+    for (i = 0; i < k->ep + k->ed; ++i) {
+      /* provided in exp_cone.c */
+      SCS(proj_pd_exp_cone)(&(x[count + 3 * i]), i < k->ep);
     }
-    count += 3 * k->ep;
+    count += 3 * (k->ep + k->ed);
   }
-
-  /* dual exponential cone */
-  if (k->ed) { /* doesn't use r_y */
-    /*
-     * exponential cone is not self dual, if s \in K
-     * then y \in K^* and so if K is the primal cone
-     * here we project onto K^*, via Moreau
-     * \Pi_C^*(y) = y + \Pi_C(-y)
-     */
-    scs_int idx;
-    scs_float r, s, t;
-    SCS(scale_array)(&(x[count]), -1, 3 * k->ed); /* x = -x; */
-#ifdef _OPENMP
-#pragma omp parallel for private(r, s, t, idx)
-#endif
-    for (i = 0; i < k->ed; ++i) {
-      idx = count + 3 * i;
-      r = x[idx];
-      s = x[idx + 1];
-      t = x[idx + 2];
-
-      proj_exp_cone(&(x[idx]));
-
-      x[idx] -= r;
-      x[idx + 1] -= s;
-      x[idx + 2] -= t;
-    }
-    count += 3 * k->ed;
-  }
-
   if (k->psize && k->p) { /* doesn't use r_y */
     scs_float v[3];
     scs_int idx;