From 57d4eec367a6e96323588f74acf5c48974383a45 Mon Sep 17 00:00:00 2001
From: Anton Khirnov <anton@khirnov.net>
Date: Tue, 2 Apr 2019 10:25:08 +0200
Subject: egs: optimize the correction step

---
 ell_grid_solve.c | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

(limited to 'ell_grid_solve.c')

diff --git a/ell_grid_solve.c b/ell_grid_solve.c
index 9aee0e8..dfc48c6 100644
--- a/ell_grid_solve.c
+++ b/ell_grid_solve.c
@@ -29,6 +29,7 @@
 
 #include "bicgstab.h"
 #include "common.h"
+#include "cpu.h"
 #include "ell_grid_solve.h"
 #include "log.h"
 #include "mg2d_boundary.h"
@@ -43,6 +44,7 @@ static const double relax_factors[FD_STENCIL_MAX] = {
 };
 
 typedef struct EGSRelaxInternal {
+    void (*line_add)(ptrdiff_t, double *, const double *, double);
     double relax_factor;
 } EGSRelaxInternal;
 
@@ -281,16 +283,31 @@ static void boundaries_apply(EGSContext *ctx, int init)
     ctx->count_boundaries++;
 }
 
+#if HAVE_EXTERNAL_ASM
+void mg2di_line_madd_fma3(ptrdiff_t linesize, double *dst, const double *src, double c);
+#endif
+
+static void line_madd_c(ptrdiff_t linesize, double *dst, const double *src, double c)
+{
+    for (ptrdiff_t i = 0; i < linesize; i++)
+        dst[i] += c * src[i];
+}
+
 static int residual_add_task(void *arg, unsigned int job_idx, unsigned int thread_idx)
 {
+#if 1
     EGSContext   *ctx = arg;
     EGSInternal *priv = ctx->priv;
 
+    priv->r.line_add(ctx->domain_size[0], ctx->u->data + ctx->u->stride[0] * job_idx,
+                     ctx->residual->data + ctx->residual->stride[0] * job_idx, priv->r.relax_factor);
+#else
     for (int idx0 = 0; idx0 < ctx->domain_size[0]; idx0++) {
         ptrdiff_t idx = job_idx * ctx->u->stride[0] + idx0;
 
         ctx->u->data[idx] += priv->r.relax_factor * ctx->residual->data[idx];
     }
+#endif
 
     return 0;
 }
@@ -683,6 +700,12 @@ int mg2di_egs_init(EGSContext *ctx, int flags)
 
         if (r->relax_multiplier > 0.0)
             priv->r.relax_factor *= r->relax_multiplier;
+
+        priv->r.line_add = line_madd_c;
+#if HAVE_EXTERNAL_ASM
+        if (ctx->cpuflags & MG2DI_CPU_FLAG_FMA3)
+            priv->r.line_add = mg2di_line_madd_fma3;
+#endif
     }
 
     priv->fd_factors[MG2D_DIFF_COEFF_00] = 1.0 /  fd_denoms[ctx->fd_stencil - 1][MG2D_DIFF_COEFF_00];
-- 
cgit v1.2.3