Merge pull request #278 from intel/develop

Develop

Makefile

@@ -266,18 +266,22 @@ combo-test: $(COMBO_TEST_EXEC)
 	$(RUN_PREFIX) $<
 
 api-tests: compiler-api
-	$(MAKE) combo-api-tests
 	$(YC_MAKE) $@
 	$(YK_MAKE) $@
+	$(MAKE) combo-api-tests
 
-unit-tests:
+common-unit-tests:
 	$(MAKE) tuple-test
 	$(MAKE) combo-test
 
+unit-tests: common-unit-tests
+	$(YC_MAKE) $@
+	$(YK_MAKE) $@
+
 example-tests:
 	$(EX_MAKE) all-tests
 
-all-tests: compiler-api unit-tests
+all-tests: compiler-api common-unit-tests
 	$(YC_MAKE) $@
 	$(YK_MAKE) $@
 	$(MAKE) combo-api-tests

README.md

@@ -49,9 +49,9 @@ and Intel(R) graphics processors.
       You'll get a warning when running `make` if one of these doesn't exist.
       Everything will still work, but the generated code will be difficult to read.
       Reading the generated code is only necessary for debug or curiosity.
-    * SWIG (3.0.12 or later; 4.0.0 or later recommended),
+    * SWIG (4.0.0 or later):
       http://www.swig.org, for creating the Python interface.
-    * Python 3 (3.6.1 or later, recommended):
+    * Python 3 (3.6.1 or later):
       https://www.python.org/downloads, for creating and using the Python interface.
     * Doxygen (1.9.0 or later):
       https://www.doxygen.nl, for creating updated API documentation.

src/common/common.mk

@@ -70,7 +70,6 @@ YC_EXEC		:=	$(BIN_OUT_DIR)/$(YC_BASE).exe
 YC_SRC_DIR	:=	$(SRC_DIR)/compiler
 
 # Tools.
-CXX		:=	icpx
 SWIG		:=	swig
 PERL		:=	perl
 MKDIR		:=	mkdir -p -v
@@ -178,6 +177,7 @@ YK_EXT_BASE	:=	$(YK_BASE).$(YK_TAG)
 YK_LIB		:=	$(LIB_OUT_DIR)/lib$(YK_EXT_BASE)$(SO_SUFFIX)
 
 # Compiler for building kernel lib and apps.
+CXX		:=	icpx
 YK_CXX		:=	$(CXX)
 MPI_CXX 	:=	mpiicpc
 ifeq ($(mpi),1)
@@ -190,8 +190,12 @@ ifeq ($(offload),1)
 endif
 
 # Base compiler flags for building kernel lib and apps.
+ifeq ($(offload),1)
+ YK_CXXDBG	:=	-gline-tables-only
+else
+ YK_CXXDBG	:=	-g
+endif
 YK_CXXOPT	:=	-O3
-YK_CXXDBG	:=	-g
 YK_CXXWARN	:=	-Wall
 YK_CXXFLAGS	:=	-std=c++17 $(YK_CXXDBG) $(YK_CXXOPT) $(YK_CXXWARN) -I$(INC_DIR) $(EXTRA_YK_CXXFLAGS)
 ifeq ($(mpi),1)
@@ -202,9 +206,9 @@ endif
 YK_LIBS		:=	-lrt
 YK_LFLAGS	:=	-Wl,-rpath=$(LIB_OUT_DIR) -L$(LIB_OUT_DIR) -l$(YK_EXT_BASE)
 
-# Default number of ranks for running tests.
-# 4 is good because it tests in-plane diagonal exchanges for 2D and 3D tests.
-# 8 would test all exchanges for 3D tests.
+# Default number of ranks for running MPI tests.
+# 4 tests in-plane diagonal exchanges for 2D and 3D tests.
+# 8 tests all exchanges for 3D tests.
 ifneq ($(mpi),1)
 ranks	:=	1
 else

src/common/common_utils.cpp

@@ -44,7 +44,7 @@ namespace yask {
     // for numbers above 9 (at least up to 99).
 
     // Format: "major.minor.patch[-alpha|-beta]".
-    const string version = "4.04.01";
+    const string version = "4.04.02";
 
     string yask_get_version_string() {
         return version;

src/compiler/Makefile

@@ -195,7 +195,10 @@ api-tests:
 	$(MAKE) cxx-yc-api-test-with-exception
 	$(MAKE) py-yc-api-test-with-exception
 
-all-tests: api-tests
+# no unit tests yet.
+unit-tests:
+
+all-tests: api-tests unit-tests
 
 all:
 	$(MAKE) compiler

src/compiler/lib/Cpp.cpp

@@ -1269,7 +1269,6 @@ namespace yask {
     }
 
     // Print loop-invariant data values for each VarPoint.
-    // TODO: fix warning from loading invariant real_vec_t outside of OMP device region.
     string CppPreLoopPrintDataVisitor::visit(VarPoint* gp) {
         assert(gp);
 
@@ -1281,9 +1280,9 @@ namespace yask {
 
             // Not already loaded?
             if (!_cvph.lookup_point_var(*gp)) {
-                string expr = _ph.read_from_point(_os, *gp);
+                string expr = string("get_copyable(") + _ph.read_from_point(_os, *gp) + ")";
                 string res;
-                make_next_temp_var(res, gp, "expr", "") << expr << _ph.get_line_suffix();
+                make_next_temp_var(res, gp, "expr", "", true) << expr << _ph.get_line_suffix();
 
                 // Save for future use.
                 _cvph.save_point_var(*gp, res);

src/compiler/lib/Print.cpp

@@ -39,7 +39,8 @@ namespace yask {
     // If 'comment' is set, use it for the comment.
     // Return stream to continue w/RHS.
     ostream& PrintVisitorBase::make_next_temp_var(string& res, Expr* ex,
-                                                  string prefix, string comment) {
+                                                  string prefix, string comment,
+                                                  bool use_auto_type) {
         res = _ph.make_var_name(prefix);
         if (ex) {
             _temp_vars[ex] = res;
@@ -48,7 +49,8 @@ namespace yask {
         }
         if (comment.length())
             _os << endl << " // " << res << " = " << comment << "." << endl;
-        _os << _ph.get_line_prefix() << _ph.get_var_type() << " " << res << " = ";
+        string vtype = use_auto_type ? "auto" : _ph.get_var_type();
+        _os << _ph.get_line_prefix() << vtype << " " << res << " = ";
         return _os;
     }
 

src/compiler/lib/Print.hpp

@@ -184,7 +184,8 @@ namespace yask {
         // If 'comment' is set, use it for the comment.
         // Return stream to continue w/RHS.
         virtual ostream& make_next_temp_var(string& res, Expr* ex,
-                                            string prefix, string comment);
+                                            string prefix, string comment,
+                                            bool use_auto_type = false);
 
     public:
         // os is used for printing intermediate results as needed.

src/examples/swe_main.cpp

@@ -572,7 +572,7 @@ int main(int argc, char** argv) {
         auto ydur = ystats->get_elapsed_secs();
         os << fixed << setprecision(2) <<
             "Duration (s): " << dur << endl <<
-            "Rate (ms/step): " << (1e3 * dur / (nt+1)) << endl <<
+            "Rate (ms/step): " << (1e3 * dur / nt) << endl <<
             "Time in YASK kernel (s): " << ydur <<
             " (" << (100. * ydur / dur) << "%)" << endl <<
             flush;

src/examples/wave_eq_main.cpp

@@ -355,7 +355,7 @@ int main(int argc, char** argv) {
         auto ydur = ystats->get_elapsed_secs();
         os << fixed << setprecision(2) <<
             "Duration (s): " << dur << endl <<
-            "Rate (ms/step): " << (1e3 * dur / (nt+1)) << endl <<
+            "Rate (ms/step): " << (1e3 * dur / nt) << endl <<
             "Time in YASK kernel (s): " << ydur <<
             " (" << (100. * ydur / dur) << "%)" << endl <<
             flush;

src/kernel/Makefile

@@ -551,7 +551,12 @@ $(YK_EXT_OBJ_DIR)/%.o: %.cpp $(YK_GEN_HEADERS) $(YK_INC_GLOB)
 # C++-preprocessor rule.
 %.i: %.cpp $(YK_GEN_HEADERS) $(YK_INC_GLOB)
 	g++ -E -CC $(YK_CXX_INCFLAGS) $< > $@
-	@ls -l $@
+	@ls -l `realpath $@`
+
+# Asm rule.
+%.s: %.cpp $(YK_GEN_HEADERS) $(YK_INC_GLOB)
+	$(CXX_PREFIX) $(YK_CXXCMD) $(YK_CXXFLAGS2) -x c++ -S -o $@ $<
+	@ls -l `realpath $@`
 
 ######## Primary targets.
 # NB: must set stencil and arch make vars to generate the desired YASK kernel.
@@ -712,19 +717,19 @@ py-api-no-yc:
 # Make this target before rebuilding YASK with any new parameters.
 clean:
 	rm -fr $(YK_OUT_DIR)
-	rm -fv *.s
+	- find . $(SRC_DIR) -name '*.s' -print -delete
 	- find . $(COMM_DIR) $(COEFF_DIR) -name '*.*.o' -print -delete
 	- find . $(COMM_DIR) $(COEFF_DIR) -name '*.optrpt' -print -delete
 
-# Remove executables, libs, etc.
-# Also remove logs from kernel dir, which are most likely from testing.
+# Remove executables, libs, reports, etc.
 realclean: clean
 	rm -fv $(YK_LIB) $(YK_EXEC) $(MAKE_REPORT_FILE)
 	rm -fv $(YK_PY_MOD)* $(YK_PY_LIB)
 	rm -fv $(YK_API_TEST_EXEC) $(YK_API_TEST_EXEC_WITH_EXCEPTION)
 	rm -fv $(BUILD_OUT_DIR)/*report.txt
-	- find . -name '*.pyc' -print -delete
-	- find . -name '*~' -print -delete
+	rm -fv ./*~
+	- find $(YASK_DIR) -name '*.pyc' -print -delete
+	- find $(SRC_DIR) $(INC_DIR) $(YASK_DIR) $(UTILS_DIR) -name '*~' -print -delete
 
 echo-settings:
 	@echo "Build environment, `date`"; \
@@ -800,11 +805,9 @@ echo-settings:
 	uname -a
 
 # Print stats on inner SIMD loops from asm file.
-code-stats: $(YK_LIB)
-	$(call MK_DIR,$(YK_OBJ_DIR))
-	$(CXX_PREFIX) $(YK_CXXCMD) $(YK_CXXFLAGS2) -x c++ -S -o $(YK_OBJ_DIR)/factory.s $(YK_LIB_SRC_DIR)/factory.cpp
+code-stats: $(YK_LIB_SRC_DIR)/factory.s
 	@echo "Code stats for stencil computation:"
-	$(PERL) $(VIEW_ASM) -p -l -f='calc_' $(YK_OBJ_DIR)/factory.s
+	$(PERL) $(VIEW_ASM) -p -l -f='calc_vectors' $(YK_LIB_SRC_DIR)/factory.s
 
 # Print some usage info.
 help:

src/kernel/lib/generic_var.cpp

@@ -154,10 +154,13 @@ namespace yask {
             // Also update the version on the device.
             #ifdef USE_OFFLOAD_NO_USM
             auto devn = KernelEnv::_omp_devn;
-
+
+            auto cval = get_copyable(val);
             _Pragma("omp target teams distribute parallel for device(devn)")
-                for (idx_t i = 0; i < ne; i++)
-                    elems[i] = val;
+                for (idx_t i = 0; i < ne; i++) {
+                    T kval = cval;
+                    elems[i] = kval;
+                }
             #endif
         }
     }
@@ -184,10 +187,13 @@ namespace yask {
             // Also update the version on the device to the same sequence.
             #ifdef USE_OFFLOAD_NO_USM
             auto devn = KernelEnv::_omp_devn;
-
+
+            auto cseed = get_copyable(seed);
             _Pragma("omp target teams distribute parallel for device(devn)")
-                for (idx_t i = 0; i < ne; i++)
-                    elems[i] = seed * T(imod_flr(i, wrap) + 1);
+                for (idx_t i = 0; i < ne; i++) {
+                    T kseed = cseed;
+                    elems[i] = kseed * T(imod_flr(i, wrap) + 1);
+                }
             #endif
         }
     }

src/kernel/lib/halo.cpp

@@ -288,7 +288,7 @@ namespace yask {
                                          if (send_vec_ok)
                                              nelems = gb.get_vecs_in_slice(bufp, first, last, use_offload);
                                          else
-                                             nelems = gb.get_elements_in_slice(bufp, first, last, use_offload);
+                                             nelems = gb.get_elements_in_slice_void(bufp, first, last, use_offload);
                                          auto nb = nelems * get_element_bytes();
                                          bufp += nb;
                                          npbytes += nb;
@@ -418,9 +418,9 @@ namespace yask {
                                                    (use_offload ? " on device" : " on host"));
                                          idx_t nelems = 0;
                                          if (recv_vec_ok)
-                                             nelems = gp->set_vecs_in_slice(bufp, first, last, use_offload);
+                                             nelems = gb.set_vecs_in_slice(bufp, first, last, use_offload);
                                          else
-                                             nelems = gp->set_elements_in_slice(bufp, first, last, use_offload);
+                                             nelems = gb.set_elements_in_slice_void(bufp, first, last, use_offload);
                                          auto nb = nelems * get_element_bytes();
                                          bufp += nb;
                                          npbytes += nb;

src/kernel/lib/realv.hpp

@@ -575,6 +575,20 @@ namespace yask {
 
     }; // real_vec_t.
 
+    // Function for getting copyable data for OMP offload.
+    #if 1
+
+    // For most types, just return the value.
+    template<typename T>
+    inline auto get_copyable(T v) { return v; }
+
+    // Specialize for 'real_vec_t' to return data union only.
+    template<>
+    inline auto get_copyable(real_vec_t v) { return v.u; }
+    #else
+    #define get_copyable(v) v
+    #endif
+
     // Output using '<<'.
     inline std::ostream& operator<<(std::ostream& os, const real_vec_t& rn) {
         rn.print_reals(os, false);

src/kernel/lib/settings.cpp

@@ -397,14 +397,13 @@ namespace yask {
                            use_shm));
         parser.add_option(make_shared<command_line_parser::bool_option>
                           ("force_scalar_exchange",
-                           "[Debug] Do not allow vectorized halo exchanges.",
+                           "[Debug] Do not allow vectorized halo packing and unpacking.",
                            force_scalar_exchange));
         #endif
         #endif
         parser.add_option(make_shared<command_line_parser::bool_option>
                           ("force_scalar",
-                           "[Debug] Evaluate every var point with scalar stencil operations "
-                           "and exchange halos using only scalar packing and unpacking.",
+                           "[Debug] Do not allow vector stencil computation.",
                            force_scalar));
         parser.add_option(make_shared<command_line_parser::int_option>
                           ("max_threads",
@@ -941,7 +940,7 @@ namespace yask {
             }
             os << " Note: only the nano-block size in the '" <<
                 _dims->_stencil_dims.get_dim_name(_bind_posn) << "' dimension may be used at run-time\n"
-                "  because block-thread binding is enabled on " << num_inner_threads << " block threads.\n";
+                "  because inner-thread data-binding is enabled on " << num_inner_threads << " inner threads.\n";
         }
 
         #ifdef USE_TILING

src/kernel/lib/settings.hpp

@@ -720,7 +720,7 @@ namespace yask {
         int get_num_comp_threads(int& outer_threads, int& inner_threads) const;
 
         // Set number of threads to use for a mega-block.
-        // Enable nested OMP if there are >1 block threads,
+        // Enable nested OMP if there are >1 inner threads,
         // disable otherwise.
         // Return number of threads.
         // Do nothing and return 0 if not properly initialized.

src/kernel/lib/stencil_calc.cpp

@@ -200,7 +200,7 @@ namespace yask {
                     TRACE_MSG("reqd part '" << rp->get_name() << "': " <<
                               mb_idxs3.make_range_str(true) <<
                               " via outer thread " << outer_thread_idx <<
-                              " with " << nbt << " block thread(s) bound to data...");
+                              " with " << nbt << " inner thread(s) bound to data...");
 
                     // Start threads within a block.  Each of these threads
                     // will eventually work on a separate nano-block.  This
@@ -210,7 +210,7 @@ namespace yask {
                         assert(omp_get_num_threads() == nbt);
                         int inner_thread_idx = omp_get_thread_num();
 
-                        // Run the micro-block loops on all block threads and
+                        // Run the micro-block loops on all inner threads and
                         // call calc_nano_block() only by the designated
                         // thread for the given slab index in the binding
                         // dim. This is an explicit replacement for "normal"
@@ -249,7 +249,7 @@ namespace yask {
                     TRACE_MSG("reqd part '" << rp->get_name() << "': " <<
                               mb_idxs3.make_range_str(true) << 
                               " via outer thread " << outer_thread_idx <<
-                              " with " << nbt << " block thread(s) NOT bound to data...");
+                              " with " << nbt << " inner thread(s) NOT bound to data...");
 
                     // Call calc_nano_block() with a different thread for
                     // each nano-block using standard OpenMP scheduling.