tmptp

1 files changed, 43 insertions, 30 deletions

diff --git a/residual_calc.c b/residual_calc.c
index 3b83c63..8e6f66d 100644
--- a/residual_calc.c
+++ b/residual_calc.c
@@ -30,7 +30,11 @@
 #include "mg2d_constants.h"
 #include "residual_calc.h"
 
+#define MIN_ELEM_PER_TASK 1
+
 typedef struct ResidualCalcTask {
+    size_t    lines_total;
+    size_t    lines_per_task;
     size_t    line_size;
 
     double               *dst;
@@ -248,35 +252,41 @@ static int residual_calc_task(void *arg, unsigned int job_idx, unsigned int thre
     ResidualCalcInternal *priv = arg;
     ResidualCalcTask     *task = &priv->task;
 
-    const double *diff_coeffs[MG2D_DIFF_COEFF_NB];
-    double *dst = task->dst + job_idx * task->dst_stride;
-
-    for (int i = 0; i < ARRAY_ELEMS(diff_coeffs); i++)
-        diff_coeffs[i] = task->diff_coeffs[i] + job_idx * task->diff_coeffs_stride;
-
-    if (task->u_mult == 0.0) {
-        priv->residual_calc_line(task->line_size, dst,
-                                 priv->residual_max + thread_idx * priv->calc_blocksize,
-                                 task->u_stride, task->u + job_idx * task->u_stride,
-                                 task->rhs + job_idx * task->rhs_stride,
-                                 diff_coeffs, task->res_mult);
-    } else {
-        priv->residual_add_line(task->line_size, dst,
-                                priv->residual_max + thread_idx * priv->calc_blocksize,
-                                task->u_stride, task->u + job_idx * task->u_stride,
-                                task->rhs + job_idx * task->rhs_stride,
-                                diff_coeffs, task->res_mult, task->u_mult);
-    }
-    if (task->mirror_line & (1 << 0)) {
-        for (int i = 1; i <= FD_STENCIL_MAX; i++)
-            dst[-i] = dst[i];
-    }
-    if (task->mirror_line & (1 << 1)) {
-        for (int i = 1; i <= FD_STENCIL_MAX; i++)
-            dst[task->line_size - 1 + i] = dst[task->line_size - 1 - i];
-    }
-    if ((task->mirror_line & (1 << 2)) && job_idx > 0 && job_idx <= FD_STENCIL_MAX) {
-        memcpy(task->dst - job_idx * task->dst_stride, dst, sizeof(*dst) * task->line_size);
+    const size_t line_min = job_idx * task->lines_per_task;
+    const size_t line_max = MIN((job_idx + 1) * task->lines_per_task, task->lines_total);
+
+    for (size_t line = line_min; line < line_max; line++) {
+        double *dst = task->dst + line * task->dst_stride;
+        const double *diff_coeffs[MG2D_DIFF_COEFF_NB];
+
+        for (int i = 0; i < ARRAY_ELEMS(diff_coeffs); i++)
+            diff_coeffs[i] = task->diff_coeffs[i] + line * task->diff_coeffs_stride;
+
+        if (task->u_mult == 0.0) {
+            priv->residual_calc_line(task->line_size, dst,
+                                     priv->residual_max + thread_idx * priv->calc_blocksize,
+                                     task->u_stride, task->u + line * task->u_stride,
+                                     task->rhs + line * task->rhs_stride,
+                                     diff_coeffs, task->res_mult);
+        } else {
+            priv->residual_add_line(task->line_size, dst,
+                                    priv->residual_max + thread_idx * priv->calc_blocksize,
+                                    task->u_stride, task->u + line * task->u_stride,
+                                    task->rhs + line * task->rhs_stride,
+                                    diff_coeffs, task->res_mult, task->u_mult);
+        }
+
+        if (task->mirror_line & (1 << 0)) {
+            for (int i = 1; i <= FD_STENCIL_MAX; i++)
+                dst[-i] = dst[i];
+        }
+        if (task->mirror_line & (1 << 1)) {
+            for (int i = 1; i <= FD_STENCIL_MAX; i++)
+                dst[task->line_size - 1 + i] = dst[task->line_size - 1 - i];
+        }
+        if ((task->mirror_line & (1 << 2)) && line > 0 && line <= FD_STENCIL_MAX) {
+            memcpy(task->dst - line * task->dst_stride, dst, sizeof(*dst) * task->line_size);
+        }
     }
 
     return 0;
@@ -309,8 +319,11 @@ int mg2di_residual_calc(ResidualCalcContext *ctx, size_t size[2],
     task->u_mult             = u_mult;
     task->res_mult           = res_mult;
     task->mirror_line        = mirror_line;
+    task->lines_total        = size[1];
+    task->lines_per_task     = MAX(1, MIN_ELEM_PER_TASK / size[0]);
 
-    tp_execute(ctx->tp, size[1], residual_calc_task, priv);
+    tp_execute(ctx->tp, (size[1] + task->lines_per_task - 1) / task->lines_per_task,
+               residual_calc_task, priv);
 
     for (size_t i = 0; i < priv->residual_max_size; i++)
         res_max = MAX(res_max, priv->residual_max[i]);