Track and check step-var indices. Closes #194.

README.md

@@ -58,6 +58,7 @@ YASK contains a domain-specific compiler to convert scalar stencil code to SIMD-
       for functional testing if you don't have native support for any given instruction set.
 
 ### Backward-compatibility notices, including changes in default behavior:
+* Version 2.20.00 added checking of the step-dimension index value in the `yk_grid::get_element()` and similar APIs. Previously, invalid values silently "wrapped" around to valid ones. Now, the step index must be valid when reading, and the valid step indices are updated when writing.
 * Version 2.19.01 turned off multi-pass tuning by default. Enable with `-auto_tune_each_pass`.
 * Version 2.18.03 allowed the default radius to be stencil-specific and changed the names of example stencil "9axis" to "3axis_with_diags".
 * Version 2.18.00 added the ability to specify the global-domain size, and it will calculate the local-domain sizes from it.

include/yask_compiler_api.hpp

@@ -182,7 +182,7 @@ namespace yask {
         /**
            C++ initializer-list version with same semantics as
            new_grid(const std::string& name, const std::vector<yc_index_node_ptr>& dims).
-           @note This version is not available (or needed) in SWIG-based APIs, e.g., Python.
+           @note This version is not available (or needed) in the Python API.
            @returns Pointer to the new \ref yc_grid object.
         */
         virtual yc_grid_ptr
@@ -223,7 +223,7 @@ namespace yask {
         /**
            C++ initializer-list version with same semantics as
            new_scratch_grid(const std::string& name, const std::vector<yc_index_node_ptr>& dims).
-           @note This version is not available (or needed) in SWIG-based APIs, e.g., Python.
+           @note This version is not available (or needed) in the Python API.
            @returns Pointer to the new \ref yc_grid object.
         */
         virtual yc_grid_ptr
@@ -492,7 +492,7 @@ namespace yask {
         /**
            C++ initializer-list version with same semantics as
            new_grid_point(std::vector<yc_index_node_ptr> index_exprs).
-           @note This version is not available (or needed) in SWIG-based APIs, e.g., Python.
+           @note This version is not available (or needed) in the Python API.
            @returns Pointer to AST node used to read or write from point in grid. */
         virtual yc_grid_point_node_ptr
         new_grid_point(const std::initializer_list<yc_number_node_ptr>& index_exprs) = 0;
@@ -523,7 +523,7 @@ namespace yask {
         /**
            C++ initializer-list version with same semantics as
            new_relative_grid_point(std::vector<int> dim_offsets).
-           @note This version is not available (or needed) in SWIG-based APIs, e.g., Python.
+           @note This version is not available (or needed) in the Python API.
            @returns Pointer to AST node used to read or write from point in grid. */
         virtual yc_grid_point_node_ptr
         new_relative_grid_point(const std::initializer_list<int>& dim_offsets) = 0;

include/yk_solution_api.hpp

@@ -709,7 +709,7 @@ namespace yask {
         /**
            See documentation for the version of new_grid() with a vector of dimension names
            as a parameter.
-           @note This version is not available (or needed) in SWIG-based APIs, e.g., Python.
+           @note This version is not available (or needed) in the Python API.
            @returns Pointer to the new grid.
         */
         virtual yk_grid_ptr
@@ -786,7 +786,7 @@ namespace yask {
         /**
            See documentation for the version of new_fixed_size_grid() with a vector of dimension names
            as a parameter.
-           @note This version is not available (or needed) in SWIG-based APIs, e.g., Python.
+           @note This version is not available (or needed) in the Python API.
            @returns Pointer to the new grid.
         */
         virtual yk_grid_ptr

src/common/common_utils.cpp

@@ -43,7 +43,7 @@ namespace yask {
     // for numbers above 9 (at least up to 99).
 
     // Format: "major.minor.patch".
-    const string version = "2.19.04";
+    const string version = "2.20.00";
 
     string yask_get_version_string() {
         return version;

src/kernel/Makefile

@@ -781,7 +781,7 @@ py-api-no-yc:
 	$(MAKE) $(NO_YC_MAKE_FLAGS) py-api
 
 # Validation runs for each binary.
-DEF_TEST_ARGS := -stencil $(YK_STENCIL) -arch $(YK_ARCH) -v \
+DEF_TEST_ARGS := -ranks $(ranks) -stencil $(YK_STENCIL) -arch $(YK_ARCH) -v \
 	-trial_steps 2 -max_threads 6 -block_threads 2 -no-use_shm -overlap_comms
 
 val0	:=	$(DEF_TEST_ARGS) -l 48 -r 32 -b 24 -rt 0 $(EXTRA_TEST_ARGS)
@@ -800,24 +800,18 @@ yk-tests:
 	$(YK_SCRIPT) $(val2)
 	$(YK_SCRIPT) $(val3)
 	$(YK_SCRIPT) $(val4a)
-
-yk-mp-tests:
-	$(YK_SCRIPT) -ranks $(ranks) $(val0)
-	$(YK_SCRIPT) -ranks $(ranks) $(val1)
-	$(YK_SCRIPT) -ranks $(ranks) $(val2)
-	$(YK_SCRIPT) -ranks $(ranks) $(val3)
-	$(YK_SCRIPT) -ranks $(ranks) $(val4b)
-	$(YK_SCRIPT) -ranks $(ranks) $(val4c)
+	$(YK_SCRIPT) $(val4b)
+	$(YK_SCRIPT) $(val4c)
 
 # Run the default YASK compiler and kernel.
 yc-and-yk-test: $(YK_EXEC) $(YK_SCRIPT)
-	$(MAKE) yk-tests
-	if (( $(ranks) > 1 )); then $(MAKE) yk-mp-tests; fi
+	$(MAKE) ranks=1 yk-tests
+	if (( $(ranks) > 1 )); then $(MAKE) yk-tests; fi
 
 # Run the YASK kernel test without implicity using the YASK compiler.
 yk-test-no-yc: kernel-no-yc $(YK_SCRIPT)
-	$(MAKE) yk-tests
-	if (( $(ranks) > 1 )); then $(MAKE) yk-mp-tests; fi
+	$(MAKE) ranks=1 yk-tests
+	if (( $(ranks) > 1 )); then $(MAKE) yk-tests; fi
 
 # Run the kernel API tests for C++ and Python with and w/o expected exceptions.
 api-tests:

src/kernel/lib/context.cpp

@@ -172,7 +172,7 @@ namespace yask {
                 // rank. This is needed because neighbors will not know what
                 // grids are actually dirty, and all ranks must have the same
                 // information about which grids are possibly dirty.
-                mark_grids_dirty(nullptr, start_t, stop_t);
+                update_grids(nullptr, start_t, stop_t, true);
 
             } // all bundles.
 
@@ -183,7 +183,7 @@ namespace yask {
         exchange_halos();
 
         run_time.stop();
-    }
+    } // run_ref.
 
     // Eval stencil bundle pack(s) over grid(s) using optimized code.
     void StencilContext::run_solution(idx_t first_step_index,
@@ -405,7 +405,7 @@ namespace yask {
                         // same information about which grids are possibly
                         // dirty.  TODO: make this smarter to save unneeded
                         // MPI exchanges.
-                        mark_grids_dirty(bp, start_t, stop_t);
+                        update_grids(bp, start_t, stop_t, true);
 
                         // Do the appropriate steps for halo exchange of exterior.
                         // TODO: exchange halo for each dim as soon as it's done.
@@ -426,8 +426,8 @@ namespace yask {
 #include "yask_rank_loops.hpp"
 
                     // Mark as dirty only if we did exterior.
-                    if (do_mpi_left || do_mpi_right)
-                        mark_grids_dirty(bp, start_t, stop_t);
+                    bool mark_dirty = do_mpi_left || do_mpi_right;
+                    update_grids(bp, start_t, stop_t, mark_dirty);
 
                     // Do the appropriate steps for halo exchange depending
                     // on 'do_mpi_*' flags.
@@ -483,7 +483,7 @@ namespace yask {
                     } // domain dims.
 
                     // Mark grids dirty for all packs.
-                    mark_grids_dirty(bp, start_t, stop_t);
+                    update_grids(bp, start_t, stop_t, true);
 
                     // Do the appropriate steps for halo exchange of exterior.
                     // TODO: exchange halo for each dim as soon as it's done.
@@ -504,8 +504,8 @@ namespace yask {
 #include "yask_rank_loops.hpp"
 
                 // Mark as dirty only if we did exterior.
-                if (do_mpi_left || do_mpi_right)
-                    mark_grids_dirty(bp, start_t, stop_t);
+                bool mark_dirty = do_mpi_left || do_mpi_right;
+                update_grids(bp, start_t, stop_t, mark_dirty);
 
                 // Do the appropriate steps for halo exchange depending
                 // on 'do_mpi_*' flags.
@@ -1639,10 +1639,13 @@ namespace yask {
                 continue;
 
             // Check all allocated step indices.
-            idx_t stop_t = 1;
-            if (gp->is_dim_used(step_dim))
-                stop_t = gp->get_alloc_size(step_dim);
-            for (idx_t t = 0; t < stop_t; t++) {
+            // Use '0' for grids that don't use the step dim.
+            idx_t start_t = 0, stop_t = 1;
+            if (gp->is_dim_used(step_dim)) {
+                start_t = gp->get_first_valid_step_index();
+                stop_t = gp->get_last_valid_step_index() + 1;
+            }
+            for (idx_t t = start_t; t < stop_t; t++) {
 
                 // Only need to swap grids whose halos are not up-to-date
                 // for this step.
@@ -1652,13 +1655,14 @@ namespace yask {
                 // Swap this grid.
                 gridsToSwap[gname] = gp;
 
-                // Update last step.
-                lastStepsToSwap[gp] = t;
-
-                // First?
-                if (firstStepsToSwap.count(gp) == 0)
+                // Update first step.
+                if (firstStepsToSwap.count(gp) == 0 || t < firstStepsToSwap[gp])
                     firstStepsToSwap[gp] = t;
 
+                // Update last step.
+                if (lastStepsToSwap.count(gp) == 0 || t > lastStepsToSwap[gp])
+                    lastStepsToSwap[gp] = t;
+
             } // steps.
         } // grids.
         TRACE_MSG("exchange_halos: need to exchange halos for " <<
@@ -1781,9 +1785,7 @@ namespace yask {
                                           (send_vec_ok ? "with" : "without") <<
                                           " vector copy into " << buf);
                                 if (send_vec_ok)
-                                    nelems = gp->get_vecs_in_slice(buf,
-                                                                   first, firstStepsToSwap[gp],
-                                                                   last, lastStepsToSwap[gp]);
+                                    nelems = gp->get_vecs_in_slice(buf, first, last);
                                 else
                                     nelems = gp->get_elements_in_slice(buf, first, last);
                                 idx_t nbytes = nelems * get_element_bytes();
@@ -1853,9 +1855,7 @@ namespace yask {
                                           (recv_vec_ok ? "with" : "without") <<
                                           " vector copy from " << buf);
                                 if (recv_vec_ok)
-                                    nelems = gp->set_vecs_in_slice(buf,
-                                                                   first, firstStepsToSwap[gp],
-                                                                   last, lastStepsToSwap[gp]);
+                                    nelems = gp->set_vecs_in_slice(buf, first, last);
                                 else
                                     nelems = gp->set_elements_in_slice(buf, first, last);
                                 assert(nelems <= recvBuf.get_size());
@@ -1918,12 +1918,10 @@ namespace yask {
 #endif
     }
 
-    // Mark grids that have been written to by bundle pack 'sel_bp'.
-    // TODO: only mark grids that are written to in their halo-read area.
-    // TODO: add index for misc dim(s).
-    // TODO: track sub-domain of grid that is dirty.
-    void StencilContext::mark_grids_dirty(const BundlePackPtr& sel_bp,
-                                          idx_t start, idx_t stop) {
+    // Update data in grids that have been written to by bundle pack 'sel_bp'.
+    void StencilContext::update_grids(const BundlePackPtr& sel_bp,
+                                      idx_t start, idx_t stop,
+                                      bool mark_dirty) {
         STATE_VARS(this);
         idx_t stride = (start > stop) ? -1 : 1;
         map<YkGridPtr, set<idx_t>> grids_done;
@@ -1952,18 +1950,20 @@ namespace yask {
                     // scratch grids as dirty because they are never exchanged.
                     for (auto gp : sb->outputGridPtrs) {
 
-                        // Mark output step as dirty if not already done.
+                        // Update if not already done.
                         if (grids_done[gp].count(t_out) == 0) {
-                            gp->set_dirty(true, t_out);
+                            gp->update_valid_step(t_out);
+                            if (mark_dirty)
+                                gp->set_dirty(true, t_out);
                             TRACE_MSG("grid '" << gp->get_name() <<
-                                      "' marked as dirty at step " << t_out);
+                                      "' updated at step " << t_out);
                             grids_done[gp].insert(t_out);
                         }
                     }
                 } // bundles.
             } // steps.
         } // packs.
-    } // mark_grids_dirty().
+    } // update_grids().
 
     // Reset any locks, etc.
     void StencilContext::reset_locks() {

src/kernel/lib/context.hpp

@@ -445,10 +445,12 @@ namespace yask {
         // Call MPI_Test() on all unfinished requests to promote MPI progress.
         void poke_halo_exchange();
 
-        // Mark grids that have been written to by bundle pack 'sel_bp'.
+        // Update valid steps in grids that have been written to by bundle pack 'sel_bp'.
         // If sel_bp==null, use all bundles.
-        void mark_grids_dirty(const BundlePackPtr& sel_bp,
-                              idx_t start, idx_t stop);
+        // If 'mark_dirty', also mark as needing halo exchange.
+        void update_grids(const BundlePackPtr& sel_bp,
+                          idx_t start, idx_t stop,
+                          bool mark_dirty);
 
         // Set various limits in 'idxs' based on current step in region.
         bool shift_region(const Indices& base_start, const Indices& base_stop,

src/kernel/lib/generic_grids.hpp

@@ -51,7 +51,7 @@ namespace yask {
         const static int _numa_unset = -999;
         int _numa_pref = _numa_unset; // use default from _opts.
 
-        // Note that both _dims and *_layout_base hold dimensions unless this
+        // Note that both _dims and *_layout_base hold sizes unless this
         // is a scalar. For a scalar, _dims is empty and _layout_base = 0.
         IdxTuple _grid_dims;         // names and lengths of grid dimensions.
         Layout* _layout_base = 0; // memory layout.