Merge pull request #100 from intel/develop

Develop

README.md

@@ -9,7 +9,7 @@ YASK--Yet Another Stencil Kernel: A framework to facilitate exploration of the H
 
 YASK contains a domain-specific compiler to convert scalar C++ stencil code to SIMD-optimized code for Intel(R) Xeon Phi(TM) and Intel(R) Xeon(R) processors.
 
-Supported Platforms
+Supported Platforms and Processors:
 * 64-bit Linux
 * Intel(R) Xeon Phi(TM) processor supporting the MIC_AVX512 instruction set.
 * Intel(R) Xeon(R) processor supporting the AVX, AVX2, or CORE_AVX512 instruction sets.
@@ -24,9 +24,11 @@ Pre-requisites:
   (optional: for functional testing if you don't have native ISA support).
 * Intel(R) MPI Library, https://software.intel.com/en-us/intel-mpi-library,
   or equivalent (optional: for multi-core and multi-node operation).
+* Linux libraries 'librt' and 'libnuma'.
 * Perl (5.010 or later).
 * Awk.
 * Gnu make.
+* Bash shell.
 * The 'indent' or 'gindent' utility (optional: to make the generated code easier for humans to read).
 * SWIG (3.0.12 or later),
   http://www.swig.org (optional: for creating the Python interface).

bin/gen_layouts.pl

@@ -277,7 +277,7 @@ END
 
       # Creation.
       print " else if (ndims == $n && do_wrap == $wrap)\n",
-        "  gp = make_shared<YkElemGrid<$layout, $wrap>>(_dims, name, dims, &_ostr);\n";
+        "  gp = make_shared<YkElemGrid<$layout, $wrap>>(_dims, name, dims, _opts, &_ostr);\n";
     }
   }
 

bin/gen_loops.pl

@@ -477,7 +477,7 @@ ($$$$)
         push @$code,
             " // This value of $divar covers ".dimStr($dim)." from $stvar to (but not including) $spvar.",
             " idx_t $stvar = std::max($abvar + ($divar * $svar), $bvar);",
-            " idx_t $spvar = std::min($stvar + $svar, $evar);";
+            " idx_t $spvar = std::min($abvar + (($divar+1) * $svar), $evar);";
     }
 }
 

bin/yask.sh

@@ -69,8 +69,6 @@ while true; do
         echo "  -host <hostname>|-mic <N>"
         echo "     Specify host to run executable on."
         echo "     'ssh <hostname>' will be pre-pended to the sh_prefix command."
-        echo "     If -arch 'knl' is given, it implies the following (which can be overridden):"
-        echo "       -exe_prefix 'numactl --preferred=1'"
         echo "     If -mic <N> is given, it implies the following (which can be overridden):"
         echo "       -arch 'knc'"
         echo "       -host "`hostname`"-mic<N>"
@@ -188,12 +186,6 @@ if [[ -z ${arch:+ok} ]]; then
     exit 1
 fi
 
-# Set defaults for KNL.
-# TODO: run numactl [on host] to determine if in flat mode.
-if [[ "$arch" == "knl" ]]; then
-    true ${exe_prefix='numactl --preferred=1'}
-fi
-
 # Simplified MPI in x-dim only.
 if [[ -n "$nranks" ]]; then
     true ${mpi_cmd="mpirun -np $nranks"}
@@ -276,3 +268,7 @@ else
 fi
 
 echo "Log saved in '$logfile'."
+
+if [[ `grep -c FAILED $logfile` > 0 ]]; then
+    exit 1;
+fi

include/yask_kernel_api.hpp

@@ -46,6 +46,15 @@ namespace yask {
     typedef std::int64_t idx_t;
 #endif
 
+    /// Allocate grids on local NUMA node.
+    const int yask_numa_local = -1;
+
+    /// Allocate grids across all available NUMA nodes.
+    const int yask_numa_interleave = -2;
+
+    /// Do not specify any NUMA binding.
+    const int yask_numa_none = -9;
+
     // Forward declarations of classes and pointers.
 
     class yk_env;
@@ -96,7 +105,7 @@ namespace yask {
         virtual yk_solution_ptr
         new_solution(yk_env_ptr env /**< [in] Pointer to env info. */) const;
 
-        /// Create a stencil solution by copying the settings from another.
+        /// **[Advanced]** Create a stencil solution by copying the settings from another.
         /**
            All the settings that were specified via the `yk_solution::set_*()`
            functions in the source solution will be copied to the new solution.
@@ -284,12 +293,17 @@ namespace yask {
            also be smaller than the specified size when the block is at the
            edge of the domain. The block size cannot be set in the
            solution-step dimension (because temporal blocking is not yet enabled).
+
+           Unless auto-tuning is disabled, the block size will be used as
+           a starting point for an automated search for a higher-performing
+           block size.
         */
         virtual void
         set_block_size(const std::string& dim
                        /**< [in] Name of dimension to set.  Must be one of
                           the names from get_domain_dim_names(). */,
-                       idx_t size /**< [in] Elements in a block in this `dim`. */ ) =0;
+                       idx_t size
+                       /**< [in] Elements in a block in this `dim`. */ ) =0;
 
         /// Get the block size.
         /**
@@ -302,21 +316,6 @@ namespace yask {
                         /**< [in] Name of dimension to get.  Must be one of
                            the names from get_domain_dim_names(). */) const =0;
 
-        /// Set performance parameters from an option string.
-        /**
-           Parses the string for options as if from a command-line.
-           Example: "-bx 64 -block_threads 4" sets the block-size in the *x*
-           dimension to 64 and the number of threads used to process each
-           block to 4.
-           See the help message from the YASK kernel binary for documentation
-           on the command-line options.
-
-           @returns Any strings that were not recognized by the parser as options.
-        */
-        virtual std::string
-        apply_command_line_options(const std::string& args
-                                   /**< [in] String of arguments to parse. */ ) =0;
-
         /// Set the number of MPI ranks in the given dimension.
         /**
            The *product* of the number of ranks across all dimensions must
@@ -703,6 +702,49 @@ namespace yask {
                           Must be exatly one size for each dimension. */ ) =0;
 #endif
 
+        /// **[Advanced]** Set the default preferred NUMA node on which to allocate data.
+        /**
+           This value is used when allocating grids and MPI buffers.
+           The NUMA "preferred node allocation" policy is used, meaning that
+           memory will be allocated in an alternative node if the preferred one
+           doesn't have enough space available or is otherwise restricted.
+           Instead of specifying a NUMA node, a special value may be used
+           to specify another policy as listed.
+           This setting may be overridden for any specific grid.
+        */
+        virtual void
+        set_default_numa_preferred(int numa_node
+                                   /**< [in] Preferred NUMA node for data
+                                      allocation.  Alternatively, use
+                                      `yask_numa_local` for explicit
+                                      local-node allocation,
+                                      `yask_numa_interleave` for
+                                      interleaving pages across all nodes,
+                                      or `yask_numa_none` for no NUMA
+                                      policy. */) =0;
+
+        /// **[Advanced]** Get the default preferred NUMA node on which to allocate data.
+        /**
+           @returns Current setting of preferred NUMA node.
+        */
+        virtual int
+        get_default_numa_preferred() const =0;
+
+        /// **[Advanced]** Set performance parameters from an option string.
+        /**
+           Parses the string for options as if from a command-line.
+           Example: "-bx 64 -block_threads 4" sets the block-size in the *x*
+           dimension to 64 and the number of threads used to process each
+           block to 4.
+           See the help message from the YASK kernel binary for documentation
+           on the command-line options.
+
+           @returns Any strings that were not recognized by the parser as options.
+        */
+        virtual std::string
+        apply_command_line_options(const std::string& args
+                                   /**< [in] String of arguments to parse. */ ) =0;
+
         /// **[Advanced]** Use data-storage from existing grids in specified solution.
         /**
            Calls yk_grid::share_storage() for each pair of grids that have the same name
@@ -1310,7 +1352,23 @@ namespace yask {
         get_num_storage_elements() const =0;
 
         /* Advanced APIs for yk_grid found below are not needed for most applications. */
-
+
+        /// **[Advanced]** Set the default preferred NUMA node on which to allocate data.
+        /**
+           This value is used when allocating data for this grid.
+        */
+        virtual void
+        set_numa_preferred(int numa_node
+                           /**< [in] Preferred NUMA node.
+                              See set_default_numa_preferred() for other options. */) =0;
+
+        /// **[Advanced]** Get the default preferred NUMA node on which to allocate data.
+        /**
+           @returns Current setting of preferred NUMA node for this grid.
+        */
+        virtual int
+        get_numa_preferred() const =0;
+
         /// **[Advanced]** Set the left halo size in the specified dimension.
         /**
            This value is typically set by the stencil compiler, but

src/common/common_utils.cpp

@@ -41,7 +41,7 @@ namespace yask {
     // for numbers above 9 (at least up to 99).
 
     // Format: "major.minor.patch".
-    const string version = "2.04.00";
+    const string version = "2.05.04";
 
     string yask_get_version_string() {
         return version;

src/compiler/lib/YaskKernel.cpp

@@ -286,7 +286,7 @@ namespace yask {
             ctorCode += " " + grid + "_dim_names = {" +
                 gdims.makeDimStr(", ", "\"", "\"") + "};\n";
             string initCode = " " + grid + "_ptr = std::make_shared<" + typeDef +
-                ">(_dims, \"" + grid + "\", " + grid + "_dim_names, &_ostr);\n"
+                ">(_dims, \"" + grid + "\", " + grid + "_dim_names, _opts, &_ostr);\n"
                 " assert(" + grid + "_ptr);\n";
 
             // Grid vars.
@@ -391,7 +391,7 @@ namespace yask {
                 if (!firstGrid)
                     newGridCode += " else";
                 newGridCode += " if (dims == " + grid + "_dim_names) gp = std::make_shared<" +
-                    typeDef + ">(_dims, name, dims, &_ostr);\n";
+                    typeDef + ">(_dims, name, dims, _opts, &_ostr);\n";
             }
 
         } // grids.
@@ -416,7 +416,7 @@ namespace yask {
         os << "\n // Make a new grid iff its dims match any in the stencil.\n"
             " // Returns pointer to the new grid or nullptr if no match.\n"
             " virtual YkGridPtr newStencilGrid(const std::string& name,"
-            " const GridDimNames& dims) {\n"
+            " const GridDimNames& dims, KernelSettingsPtr settings) {\n"
             " YkGridPtr gp;\n" <<
             newGridCode <<
             " return gp;\n"

src/kernel/Makefile

@@ -157,7 +157,7 @@ else ifeq ($(arch),knl)
 
  ISA		?=	-xMIC-AVX512
  GCXX_ISA	?=	-march=knl
- MACROS		+=	USE_INTRIN512 USE_RCP28
+ MACROS		+=	USE_INTRIN512 USE_RCP28 NUMA_PREF=1
  YC_TARGET  	?=	avx512
  def_block_args	?=	-b 96
  def_block_threads ?=	8
@@ -427,7 +427,25 @@ YK_CXXOPT	?=	-O3
 
 # Linker.
 YK_LD		:=	$(YK_CXX)
-YK_LFLAGS	:=	-lrt -Wl,-rpath=$(LIB_DIR) -L$(LIB_DIR) -l$(YK_BASE2)
+YK_LIBS		:=	-lrt
+YK_LFLAGS	:=	-Wl,-rpath=$(LIB_DIR) -L$(LIB_DIR) -l$(YK_BASE2)
+
+# Work-around missing numaif.h:
+# IF numaif.h is found in /usr/include,
+# THEN enable the macro to use it.
+ifneq ($(shell find /usr/include -name 'numaif.h' | wc -l),0)
+ MACROS		+=	USE_NUMAIF_H
+endif
+
+# Work-around missing libnuma.so:
+# IF libnuma.so.1 exists AND "normal" libnuma can't be found,
+# THEN use hard-coded libnuma.so.1,
+# ELSE use "normal" libnuma.
+ifeq ($(and $(wildcard /usr/lib64/libnuma.so.1),$(shell whereis libnuma |wc -w)),1)
+ YK_LIBS	+=	/usr/lib64/libnuma.so.1
+else
+ YK_LIBS	+=	-lnuma
+endif
 
 # Tools.
 SWIG		:=	swig
@@ -462,15 +480,6 @@ ifeq ($(mpi),1)
  MACROS		+=	USE_MPI
 endif
 
-# HBW settings.
-# TODO: change this to use OS-default location.
-ifeq ($(hbw),1)
- MACROS		+=	USE_HBW
- HBW_DIR 	=	$(HOME)/memkind_build
- YK_CXXFLAGS	+=	-I$(HBW_DIR)/include
- YK_LFLAGS	+= 	-lnuma $(HBW_DIR)/lib/libmemkind.a
-endif
-
 # VTUNE settings.
 ifeq ($(vtune),1)
  MACROS		+=	USE_VTUNE
@@ -482,7 +491,7 @@ else
  VTUNE_DIR	=	$(VTUNE_AMPLIFIER_XE_2016_DIR)
 endif
 YK_CXXFLAGS	+=	-I$(VTUNE_DIR)/include
-YK_LFLAGS	+=	$(VTUNE_DIR)/lib64/libittnotify.a
+YK_LIBS		+=	$(VTUNE_DIR)/lib64/libittnotify.a
 endif
 
 # compiler-specific settings
@@ -541,11 +550,11 @@ kernel:	$(YK_EXEC) $(MAKE_REPORT_FILE)
 	@ls -l $@
 
 $(YK_LIB): $(YK_OBJS)
-	$(CXX_PREFIX) $(YK_CXX) $(YK_CXXFLAGS) -shared -o $@ $^
+	$(CXX_PREFIX) $(YK_CXX) $(YK_CXXFLAGS) $(YK_LIBS) -shared -o $@ $^
 	@ls -l $@
 
 $(YK_EXEC): yask_main.cpp $(YK_LIB)
-	$(CXX_PREFIX) $(YK_LD) $(YK_CXXFLAGS) $< $(YK_LFLAGS) -o $@
+	$(CXX_PREFIX) $(YK_LD) $(YK_CXXFLAGS) $< $(YK_LIBS) $(YK_LFLAGS) -o $@
 	@ls -l $@
 
 $(MAKE_REPORT_FILE): $(YK_LIB)
@@ -633,7 +642,7 @@ $(YK_SWIG_DIR)/yask_kernel_api_wrap.o: $(YK_SWIG_DIR)/yask_kernel_api_wrap.cpp
 	$(YK_CXX) $(YK_CXXFLAGS) $(PYINC) -fPIC -c -o $@ $<
 
 $(YK_PY_LIB): $(YK_OBJS) $(YK_SWIG_DIR)/yask_kernel_api_wrap.o
-	$(YK_CXX) $(YK_CXXFLAGS) -shared -o $@ $^
+	$(YK_CXX) $(YK_CXXFLAGS) $(YK_LIBS) -shared -o $@ $^
 
 # Simple tests
 
@@ -736,8 +745,8 @@ all-tests:
 	$(MAKE) clean; $(MAKE) stencil=3plane fold=y=2,z=4 yc-and-yk-test
 	$(MAKE) clean; $(MAKE) stencil=cube fold=x=2,y=2,z=2 yc-and-yk-test
 	$(MAKE) clean; $(MAKE) stencil=test_4d fold=w=2,x=2,y=2,z=2 yc-and-yk-test
-	$(MAKE) clean; $(MAKE) stencil=test_scratch1 fold=x=4 yc-and-yk-test
-	$(MAKE) clean; $(MAKE) stencil=test_scratch2 fold=x=2,y=2,z=2 yc-and-yk-test
+	$(MAKE) clean; $(MAKE) stencil=test_scratch1 real_bytes=8 fold=x=4 yc-and-yk-test
+	$(MAKE) clean; $(MAKE) stencil=test_scratch2 real_bytes=8 fold=x=2,y=1,z=2 yc-and-yk-test
 	$(MAKE) clean; $(MAKE) stencil=iso3dfd fold=x=4,y=2 yc-and-yk-test
 	$(MAKE) clean; $(MAKE) stencil=awp_elastic real_bytes=8 yc-and-yk-test
 	$(MAKE) clean; $(MAKE) stencil=ssg real_bytes=8 yc-and-yk-test
@@ -862,4 +871,4 @@ help:
 	@echo " $(MAKE) -j all"
 	@echo " $(MAKE) -j all ranks=2"
 	@echo " $(MAKE) -j all YK_CXX=g++ YK_CXXOPT=-O2 mpi=0"
-	@echo " $(MAKE) -j all YK_CXX=mpigxx YK_CXXOPT=-O2 ranks=3"
+	@echo " $(MAKE) -j all YK_CXX=mpigxx YK_CXXOPT=-O2 ranks=3 EXTRA_MACROS='DEBUG'"