Fixing a bug in the ExecutionKernel

pc2 · Aug 8, 2021 · 370b794 · 370b794
1 parent 0265d8e
commit 370b794
Show file tree

Hide file tree

Showing 3 changed files with 34 additions and 8 deletions.
diff --git a/StencilStream/ExecutionKernel.hpp b/StencilStream/ExecutionKernel.hpp
@@ -163,7 +163,7 @@ class ExecutionKernel
                     stencil_buffer[stage],
                     UID(grid_width, grid_height));
 
-                if (stage < n_generations)
+                if (i_generation + stage < n_generations)
                 {
                     value = trans_func(stencil);
                 }

diff --git a/StencilStream/StencilExecutor.hpp b/StencilStream/StencilExecutor.hpp
@@ -113,11 +113,6 @@ class StencilExecutor
         for (uindex_t i = 0; i < n_passes; i++)
         {
             Grid output_grid = input_grid.make_output_grid();
-            uindex_t n_generations_per_pass = pipeline_length;
-            if (i == n_passes - 1 && n_generations % pipeline_length != 0)
-            {
-                n_generations_per_pass = n_generations % pipeline_length;
-            }
 
             for (uindex_t c = 0; c < input_grid.get_tile_range().c; c++)
             {
@@ -129,7 +124,7 @@ class StencilExecutor
                                                                      { cgh.single_task(ExecutionKernelImpl(
                                                                            trans_func,
                                                                            i_generation,
-                                                                           n_generations_per_pass,
+                                                                           n_generations,
                                                                            c * tile_width,
                                                                            r * tile_height,
                                                                            grid_width,
@@ -145,7 +140,7 @@ class StencilExecutor
                 }
             }
             input_grid = output_grid;
-            i_generation += n_generations_per_pass;
+            i_generation += std::min(n_generations - i_generation, pipeline_length);
         }
     }
 

diff --git a/tests/src/units/ExecutionKernel.cpp b/tests/src/units/ExecutionKernel.cpp
@@ -137,5 +137,36 @@ TEST_CASE("Halo values inside the pipeline are handled correctly", "[ExecutionKe
     }
 
     REQUIRE(in_pipe::empty());
+    REQUIRE(out_pipe::empty());
+}
+
+TEST_CASE("Incomplete Pipeline with i_generation != 0", "[ExecutionKernel]")
+{
+    using Cell = uint8_t;
+    auto trans_func = [](Stencil<Cell, 1> const &stencil) {
+        return stencil[ID(0,0)] + 1;
+    };
+
+    using in_pipe = HostPipe<class IncompletePipelineInPipeID, Cell>;
+    using out_pipe = HostPipe<class IncompletePipelineOutPipeID, Cell>;
+    using TestExecutionKernel = ExecutionKernel<decltype(trans_func), Cell, 1, 16, 64, 64, in_pipe, out_pipe>;
+
+    for (int c = -16; c < 16+64; c++) {
+        for (int r = -16; r < 16+64; r++) {
+            in_pipe::write(0);
+        }
+    }
+
+    TestExecutionKernel kernel(trans_func, 16, 20, 0, 0, 64, 64, 0);
+    kernel.operator()();
+
+    REQUIRE(in_pipe::empty());
+
+    for (int c = 0; c < 64; c++) {
+        for (int r = 0; r < 64; r++) {
+            REQUIRE(out_pipe::read() == 4);
+        }
+    }
+
     REQUIRE(out_pipe::empty());
 }