ell_grid_solve: split residual computation into its own file

1 files changed, 31 insertions, 192 deletions

diff --git a/ell_grid_solve.c b/ell_grid_solve.c
index 20de6bc..1893214 100644
--- a/ell_grid_solve.c
+++ b/ell_grid_solve.c
@@ -28,12 +28,12 @@
 #include <lapacke.h>
 
 #include "common.h"
-#include "cpu.h"
 #include "ell_grid_solve.h"
 #include "log.h"
 #include "mg2d_boundary.h"
 #include "mg2d_boundary_internal.h"
 #include "mg2d_constants.h"
+#include "residual_calc.h"
 
 static const double relax_factors[FD_STENCIL_MAX] = {
     [0] = 1.0 / 5,
@@ -63,18 +63,12 @@ struct EGSInternal {
     double *residual_base;
     double *diff_coeffs_base[MG2D_DIFF_COEFF_NB];
 
-    double *residual_max;
-    size_t  residual_max_size;
-
-    void (*residual_calc_line)(size_t linesize, double *dst, double *dst_max,
-                               ptrdiff_t stride, const double *u, const double *rhs,
-                               const double * const diff_coeffs[MG2D_DIFF_COEFF_NB],
-                               const double *fd_factors);
-    size_t calc_blocksize;
     ptrdiff_t residual_calc_offset;
     size_t residual_calc_size[2];
     double fd_factors[MG2D_DIFF_COEFF_NB];
 
+    ResidualCalcContext *rescalc;
+
     union {
         EGSRelaxInternal r;
         EGSExactInternal e;
@@ -124,158 +118,21 @@ static const double fd_denoms[][MG2D_DIFF_COEFF_NB] = {
     },
 };
 
-#if HAVE_EXTERNAL_ASM
-void mg2di_residual_calc_line_s1_fma3(size_t linesize, double *dst, double *dst_max,
-                                      ptrdiff_t stride, const double *u, const double *rhs,
-                                      const double * const diff_coeffs[MG2D_DIFF_COEFF_NB],
-                                      const double *fd_factors);
-void mg2di_residual_calc_line_s2_fma3(size_t linesize, double *dst, double *dst_max,
-                                      ptrdiff_t stride, const double *u, const double *rhs,
-                                      const double * const diff_coeffs[MG2D_DIFF_COEFF_NB],
-                                      const double *fd_factors);
-#endif
-
-static void
-derivatives_calc_s1(double *dst, const double *u, const double *fd_factors, ptrdiff_t stride)
-{
-    dst[MG2D_DIFF_COEFF_00] = u[0];
-    dst[MG2D_DIFF_COEFF_10] = (u[1]      - u[-1])      * fd_factors[MG2D_DIFF_COEFF_10];
-    dst[MG2D_DIFF_COEFF_01] = (u[stride] - u[-stride]) * fd_factors[MG2D_DIFF_COEFF_01];
-
-    dst[MG2D_DIFF_COEFF_20] = (u[1]      - 2.0 * u[0] + u[-1])      * fd_factors[MG2D_DIFF_COEFF_20];
-    dst[MG2D_DIFF_COEFF_02] = (u[stride] - 2.0 * u[0] + u[-stride]) * fd_factors[MG2D_DIFF_COEFF_02];
-
-    dst[MG2D_DIFF_COEFF_11] = (u[1 + stride] - u[stride - 1] - u[-stride + 1] + u[-stride - 1]) * fd_factors[MG2D_DIFF_COEFF_11];
-}
-
-static void
-derivatives_calc_s2(double *dst, const double *u, const double *fd_factors, ptrdiff_t stride)
-{
-    const double val    = u[0];
-
-    const double valxp1 = u[ 1];
-    const double valxp2 = u[ 2];
-    const double valxm1 = u[-1];
-    const double valxm2 = u[-2];
-    const double valyp1 = u[ 1 * stride];
-    const double valyp2 = u[ 2 * stride];
-    const double valym1 = u[-1 * stride];
-    const double valym2 = u[-2 * stride];
-
-    const double valxp1yp1 = u[ 1 + 1 * stride];
-    const double valxp1yp2 = u[ 1 + 2 * stride];
-    const double valxp1ym1 = u[ 1 - 1 * stride];
-    const double valxp1ym2 = u[ 1 - 2 * stride];
-
-    const double valxp2yp1 = u[ 2 + 1 * stride];
-    const double valxp2yp2 = u[ 2 + 2 * stride];
-    const double valxp2ym1 = u[ 2 - 1 * stride];
-    const double valxp2ym2 = u[ 2 - 2 * stride];
-
-    const double valxm1yp1 = u[-1 + 1 * stride];
-    const double valxm1yp2 = u[-1 + 2 * stride];
-    const double valxm1ym1 = u[-1 - 1 * stride];
-    const double valxm1ym2 = u[-1 - 2 * stride];
-
-    const double valxm2yp1 = u[-2 + 1 * stride];
-    const double valxm2yp2 = u[-2 + 2 * stride];
-    const double valxm2ym1 = u[-2 - 1 * stride];
-    const double valxm2ym2 = u[-2 - 2 * stride];
-
-    dst[MG2D_DIFF_COEFF_00] = val;
-    dst[MG2D_DIFF_COEFF_10] = (-1.0 * valxp2 + 8.0 * valxp1      - 8.0 * valxm1 + 1.0 * valxm2) * fd_factors[MG2D_DIFF_COEFF_10];
-    dst[MG2D_DIFF_COEFF_01] = (-1.0 * valyp2 + 8.0 * valyp1      - 8.0 * valym1 + 1.0 * valym2) * fd_factors[MG2D_DIFF_COEFF_01];
-
-    dst[MG2D_DIFF_COEFF_20] = (-1.0 * valxp2 + 16.0 * valxp1 - 30.0 * val + 16.0 * valxm1 - 1.0 * valxm2) * fd_factors[MG2D_DIFF_COEFF_20];
-    dst[MG2D_DIFF_COEFF_02] = (-1.0 * valyp2 + 16.0 * valyp1 - 30.0 * val + 16.0 * valym1 - 1.0 * valym2) * fd_factors[MG2D_DIFF_COEFF_02];
-
-    dst[MG2D_DIFF_COEFF_11] = ( 1.0 * valxp2yp2 -  8.0 * valxp2yp1 +  8.0 * valxp2ym1 - 1.0 * valxp2ym2
-                               -8.0 * valxp1yp2 + 64.0 * valxp1yp1 - 64.0 * valxp1ym1 + 8.0 * valxp1ym2
-                               +8.0 * valxm1yp2 - 64.0 * valxm1yp1 + 64.0 * valxm1ym1 - 8.0 * valxm1ym2
-                               -1.0 * valxm2yp2 +  8.0 * valxm2yp1 -  8.0 * valxm2ym1 + 1.0 * valxm2ym2) * fd_factors[MG2D_DIFF_COEFF_11];
-}
-
-static void residual_calc_line_s1_c(size_t linesize, double *dst, double *dst_max,
-                                    ptrdiff_t stride, const double *u, const double *rhs,
-                                    const double * const diff_coeffs[MG2D_DIFF_COEFF_NB],
-                                    const double *fd_factors)
-{
-    double res_max = 0.0, res_abs;
-    for (size_t i = 0; i < linesize; i++) {
-        double u_vals[MG2D_DIFF_COEFF_NB];
-        double res;
-
-        derivatives_calc_s1(u_vals, u + i, fd_factors, stride);
-
-        res = -rhs[i];
-        for (int j = 0; j < ARRAY_ELEMS(u_vals); j++)
-            res += u_vals[j] * diff_coeffs[j][i];
-        dst[i] = res;
-
-        res_abs = fabs(res);
-        res_max = MAX(res_max, res_abs);
-    }
-
-    *dst_max = MAX(*dst_max, res_max);
-}
-
-static void residual_calc_line_s2_c(size_t linesize, double *dst, double *dst_max,
-                                    ptrdiff_t stride, const double *u, const double *rhs,
-                                    const double * const diff_coeffs[MG2D_DIFF_COEFF_NB],
-                                    const double *fd_factors)
-{
-    double res_max = 0.0, res_abs;
-    for (size_t i = 0; i < linesize; i++) {
-        double u_vals[MG2D_DIFF_COEFF_NB];
-        double res;
-
-        derivatives_calc_s2(u_vals, u + i, fd_factors, stride);
-
-        res = -rhs[i];
-        for (int j = 0; j < ARRAY_ELEMS(u_vals); j++)
-            res += u_vals[j] * diff_coeffs[j][i];
-        dst[i] = res;
-
-        res_abs = fabs(res);
-        res_max = MAX(res_max, res_abs);
-    }
-
-    *dst_max = MAX(*dst_max, res_max);
-}
-
-static int residual_calc_task(void *arg, unsigned int job_idx, unsigned int thread_idx)
-{
-    EGSContext  *ctx = arg;
-    EGSInternal *priv = ctx->priv;
-    const ptrdiff_t offset = priv->residual_calc_offset + job_idx * priv->stride;
-    const double *diff_coeffs[MG2D_DIFF_COEFF_NB];
-
-    for (int i = 0; i < ARRAY_ELEMS(diff_coeffs); i++)
-        diff_coeffs[i] = ctx->diff_coeffs[i] + offset;
-
-    priv->residual_calc_line(priv->residual_calc_size[0], ctx->residual + offset,
-                             priv->residual_max + thread_idx * priv->calc_blocksize,
-                             priv->stride, ctx->u + offset, ctx->rhs + offset,
-                             diff_coeffs, priv->fd_factors);
-
-    return 0;
-}
-
 static void residual_calc(EGSContext *ctx)
 {
     EGSInternal *priv = ctx->priv;
-    double res_max = 0.0;
+    const double *diff_coeffs[MG2D_DIFF_COEFF_NB];
     int64_t start;
 
-    memset(priv->residual_max, 0, sizeof(*priv->residual_max) * priv->residual_max_size);
-
     start = gettime();
 
-    tp_execute(ctx->tp, priv->residual_calc_size[1], residual_calc_task, ctx);
+    for (int i = 0; i < ARRAY_ELEMS(diff_coeffs); i++)
+        diff_coeffs[i] = ctx->diff_coeffs[i] + priv->residual_calc_offset;
 
-    for (size_t i = 0; i < priv->residual_max_size; i++)
-        res_max = MAX(res_max, priv->residual_max[i]);
-    ctx->residual_max = res_max;
+    mg2di_residual_calc(priv->rescalc, priv->residual_calc_size, priv->stride,
+                        &ctx->residual_max, ctx->residual + priv->residual_calc_offset,
+                        ctx->u + priv->residual_calc_offset, ctx->rhs + priv->residual_calc_offset,
+                        diff_coeffs, priv->fd_factors);
 
     ctx->time_res_calc += gettime() - start;
     ctx->count_res++;
@@ -704,39 +561,17 @@ int mg2di_egs_solve(EGSContext *ctx)
 int mg2di_egs_init(EGSContext *ctx)
 {
     EGSInternal *priv = ctx->priv;
-    double *tmp;
     int ret;
 
-    priv->calc_blocksize = 1;
-    switch (ctx->fd_stencil) {
-    case 1:
-        if (ctx->solver_type == EGS_SOLVER_EXACT)
-            priv->e.fill_mat = fill_mat_s1;
-
-        priv->residual_calc_line = residual_calc_line_s1_c;
-#if HAVE_EXTERNAL_ASM
-        if (ctx->cpuflags & MG2DI_CPU_FLAG_FMA3) {
-            priv->residual_calc_line = mg2di_residual_calc_line_s1_fma3;
-            priv->calc_blocksize     = 4;
-        }
-#endif
-        break;
-    case 2:
-        if (ctx->solver_type == EGS_SOLVER_EXACT)
-            priv->e.fill_mat = fill_mat_s2;
-
-        priv->residual_calc_line = residual_calc_line_s2_c;
-#if HAVE_EXTERNAL_ASM
-        if (ctx->cpuflags & MG2DI_CPU_FLAG_FMA3) {
-            priv->residual_calc_line = mg2di_residual_calc_line_s2_fma3;
-            priv->calc_blocksize     = 4;
+    if (ctx->solver_type == EGS_SOLVER_EXACT) {
+        switch (ctx->fd_stencil) {
+        case 1: priv->e.fill_mat = fill_mat_s1; break;
+        case 2: priv->e.fill_mat = fill_mat_s2; break;
+        default:
+            mg2di_log(&ctx->logger, 0, "Invalid finite difference stencil: %zd\n",
+                      ctx->fd_stencil);
+            return -EINVAL;
         }
-#endif
-        break;
-    default:
-        mg2di_log(&ctx->logger, 0, "Invalid finite difference stencil: %zd\n",
-                  ctx->fd_stencil);
-        return -EINVAL;
     }
 
     if (ctx->step[0] <= DBL_EPSILON || ctx->step[1] <= DBL_EPSILON) {
@@ -791,14 +626,13 @@ int mg2di_egs_init(EGSContext *ctx)
         }
     }
 
-    priv->residual_max_size = tp_get_nb_threads(ctx->tp) * priv->calc_blocksize;
-    tmp = realloc(priv->residual_max,
-                  sizeof(*priv->residual_max) * priv->residual_max_size);
-    if (!tmp) {
-        priv->residual_max_size = 0;
-        return -ENOMEM;
-    }
-    priv->residual_max = tmp;
+    priv->rescalc->tp         = ctx->tp;
+    priv->rescalc->fd_stencil = ctx->fd_stencil;
+    priv->rescalc->cpuflags   = ctx->cpuflags;
+
+    ret = mg2di_residual_calc_init(priv->rescalc);
+    if (ret < 0)
+        return ret;
 
     boundaries_apply(ctx);
     residual_calc(ctx);
@@ -909,6 +743,10 @@ EGSContext *mg2di_egs_alloc(enum EGSType type, size_t domain_size[2])
     ctx->domain_size[0] = domain_size[0];
     ctx->domain_size[1] = domain_size[1];
 
+    ctx->priv->rescalc = mg2di_residual_calc_alloc();
+    if (!ctx->priv->rescalc)
+        goto fail;
+
     return ctx;
 fail:
     mg2di_egs_free(&ctx);
@@ -922,6 +760,8 @@ void mg2di_egs_free(EGSContext **pctx)
     if (!ctx)
         return;
 
+    mg2di_residual_calc_free(&ctx->priv->rescalc);
+
     free(ctx->solver_data);
 
     if (ctx->solver_type == EGS_SOLVER_EXACT) {
@@ -936,7 +776,6 @@ void mg2di_egs_free(EGSContext **pctx)
     free(ctx->priv->u_base);
     free(ctx->priv->rhs_base);
     free(ctx->priv->residual_base);
-    free(ctx->priv->residual_max);
     for (int i = 0; i < ARRAY_ELEMS(ctx->priv->diff_coeffs_base); i++)
         free(ctx->priv->diff_coeffs_base[i]);
     for (int i = 0; i < ARRAY_ELEMS(ctx->boundaries); i++)