Merge pull request #238 from intel/develop

Develop

README.md

@@ -27,13 +27,13 @@ YASK contains a domain-specific compiler to convert stencil-equation specificati
   for multi-socket and multi-node operation or
   Intel(R) Parallel Studio XE Composer Edition for C++ Linux
   for single-socket only
-  (2018 or later; 2019.3 or later recommended).
-     * There was an issue in Intel(R) MPI versions 2019.1 and 2019.2 that
+  (2018 or later; 2019 update 5 (2019.0.5 or 2019u5) or later recommended).
+     * There was an issue in Intel(R) MPI versions 2019u1 and 2019u2 that
        caused the application to crash when allocating very
        large shared-memory (shm) regions, so those
        versions are not recommended when using the `-use_shm` feature.
-       This issue was resolved in MPI version 2019.3.
-     * If you are using g++ version 8.x or later, Intel(R) C++ version 2019.x or later
+       This issue was resolved in MPI version 2019u3.
+     * If you are using g++ version 8.x or later, Intel(R) C++ version 2019
        is required.
      * Building a YASK kernel with the Gnu C++ compiler is possible.
        Limited testing with g++ 8.2.0 shows the "iso3dfd" kernel

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 = "3.02.01";
+    const string version = "3.03.00";
 
     string yask_get_version_string() {
         return version;

src/compiler/compiler_main.cpp

@@ -95,6 +95,7 @@ void usage(const string& cmd,
         "      avx         YASK stencil classes for CORE AVX ISA (256-bit HW SIMD vectors).\n"
         "      avx2        YASK stencil classes for CORE AVX2 ISA (256-bit HW SIMD vectors).\n"
         "      avx512      YASK stencil classes for CORE AVX-512 ISA (512-bit HW SIMD vectors).\n"
+        "      avx512lo    YASK stencil classes for CORE AVX-512 ISA (256-bit HW SIMD vectors).\n"
         "      knc         YASK stencil classes for Knights-Corner ISA (512-bit HW SIMD vectors).\n"
         "      knl         YASK stencil classes for Knights-Landing (MIC) AVX-512 ISA (512-bit HW SIMD vectors).\n"
         "      intel64     YASK stencil classes for generic C++ (no explicit HW SIMD vectors).\n"

src/compiler/lib/CppIntrin.hpp

@@ -244,6 +244,7 @@ namespace yask {
     // Print 512-bit AVX intrinsic code.
     class YASKAvx512Printer : public YASKCppPrinter {
     protected:
+        bool _is_lo;
         virtual CppVecPrintHelper* newCppVecPrintHelper(VecInfoVisitor& vv,
                                                         CounterVisitor& cv) {
             return new CppAvx512PrintHelper(vv, _settings, _dims, &cv,
@@ -254,10 +255,14 @@ namespace yask {
         YASKAvx512Printer(StencilSolution& stencil,
                           EqBundles& eqBundles,
                           EqBundlePacks& eqBundlePacks,
-                          EqBundles& clusterEqBundles) :
-            YASKCppPrinter(stencil, eqBundles, eqBundlePacks, clusterEqBundles) { }
+                          EqBundles& clusterEqBundles,
+                          bool is_lo = false) :
+            YASKCppPrinter(stencil, eqBundles, eqBundlePacks, clusterEqBundles),
+            _is_lo(is_lo) { }
 
-        virtual int num_vec_elems() const { return 64 / _settings._elem_bytes; }
+        virtual int num_vec_elems() const {
+            return (_is_lo ? 32 : 64) / _settings._elem_bytes;
+        }
 
         // Whether multi-dim folding is efficient.
         virtual bool is_folding_efficient() const { return true; }

src/compiler/lib/Solution.cpp

@@ -239,6 +239,8 @@ namespace yask {
             _printer = new YASKAvx256Printer(*this, *_eqBundles, *_eqBundlePacks, *_clusterEqBundles);
         else if (target == "avx512" || target == "knl")
             _printer = new YASKAvx512Printer(*this, *_eqBundles, *_eqBundlePacks, *_clusterEqBundles);
+        else if (target == "avx512lo")
+            _printer = new YASKAvx512Printer(*this, *_eqBundles, *_eqBundlePacks, *_clusterEqBundles, true);
         else if (target == "dot")
             _printer = new DOTPrinter(*this, *_clusterEqBundles, false);
         else if (target == "dot-lite")

src/kernel/Makefile

@@ -64,7 +64,9 @@ ifneq ($(filter $(arch),avx snb ivb),)
   TARGET		:=	avx
 else ifneq ($(filter $(arch),avx2 hsw bdw),)
   TARGET		:=	avx2
-else ifneq ($(filter $(arch),avx512 avx512f skx skl clx),)
+else ifneq ($(filter $(arch),avx512lo),)
+  TARGET		:=	avx512lo
+else ifneq ($(filter $(arch),avx512 avx512hi avx512f skx skl clx),)
   TARGET		:=	avx512
 else ifneq ($(filter $(arch),knl),)
   TARGET		:=	knl
@@ -73,7 +75,7 @@ else ifneq ($(filter $(arch),knc),)
 else ifneq ($(filter $(arch),intel64 cpp),)
   TARGET		:=	intel64
 else
-  $(error Target not recognized; use arch=avx512, avx2, avx, knl, knc, or intel64)
+  $(error Target not recognized; use arch=avx512, avx512lo, avx2, avx, knl, knc, or intel64)
 endif
 
 # Specify YK_STENCIL and/or YK_ARCH on 'make' cmd-line to name binaries differently.
@@ -108,6 +110,12 @@ else ifeq ($(TARGET),avx512)
  CXX_ISA	?=	-march=skylake-avx512
  MACROS		+=	USE_INTRIN512
 
+else ifeq ($(TARGET),avx512lo)
+
+ ICC_ISA	?=	-xCORE-AVX512
+ CXX_ISA	?=	-march=skylake-avx512
+ MACROS		+=	USE_INTRIN512LO
+
 else ifeq ($(TARGET),avx2)
 
  ICC_ISA	?=	-xCORE-AVX2

src/kernel/lib/alloc.cpp

@@ -145,7 +145,7 @@ namespace yask {
             os << endl;
         }
 #endif
-        
+
         // Base ptrs for all default-alloc'd data.
         // These pointers will be shared by the ones in the var
         // objects, which will take over ownership when these go
@@ -156,7 +156,7 @@ namespace yask {
 #ifdef USE_PMEM
         auto preferredNUMASize = opts->_numa_pref_max * 1024*1024*(size_t)1024;
 #endif
-        
+
         // Pass 0: assign PMEM node when preferred NUMA node is not enough.
         // Pass 1: count required size for each NUMA node, allocate chunk of memory at end.
         // Pass 2: distribute parts of already-allocated memory chunk.
@@ -281,7 +281,7 @@ namespace yask {
 
                     // Get calculated max dist needed for this var.
                     int maxdist = gp->get_halo_exchange_l1_norm();
-                    
+
                     // Always use max dist with WF. Do this because edge
                     // and/or corner values may be needed in WF extensions
                     // even it not needed w/o WFs.
@@ -300,7 +300,7 @@ namespace yask {
                                   "' (max L1-norm = " << maxdist << ")");
                         continue; // to next var.
                     }
-                    
+
                     // Lookup first & last domain indices and calc exchange sizes
                     // for this var.
                     bool found_delta = false;
@@ -640,7 +640,7 @@ namespace yask {
 
         // At this point, we have all the buffers configured.
         // Now we need to allocate space for them.
-        
+
         // Base ptrs for all alloc'd data.
         // These pointers will be shared by the ones in the var
         // objects, which will take over ownership when these go
@@ -656,7 +656,7 @@ namespace yask {
 
         // Make sure pad is big enough for shm locks.
         assert(_data_buf_pad >= sizeof(SimpleLock));
- 
+
         // Allocate MPI buffers.
         // Pass 0: count required size, allocate chunk of memory at end.
         // Pass 1: distribute parts of already-allocated memory chunk.
@@ -685,7 +685,7 @@ namespace yask {
                     for (auto& gtab : sb_ofs[gname])
                         gtab.resize(env->num_shm_ranks, 0);
                 }
-                
+
                 // Visit buffers for each neighbor for this var.
                 var_mpi_data.visitNeighbors
                     ([&](const IdxTuple& roffsets,
@@ -704,7 +704,7 @@ namespace yask {
                             numa_pref = _shmem_key;
                             assert(nshm_rank < env->num_shm_ranks);
                         }
-                            
+
                         // Send and recv.
                         for (int bd = 0; bd < MPIBufs::nBufDirs; bd++) {
                             auto& buf = var_mpi_data.getBuf(MPIBufs::BufDir(bd), roffsets);
@@ -758,7 +758,7 @@ namespace yask {
                                 nbufs[numa_pref]++;
                                 if (pass == 0)
                                     TRACE_MSG("  MPI buf '" << buf.name << "' needs " <<
-                                              makeByteStr(sbytes) << 
+                                              makeByteStr(sbytes) <<
                                               " (mem-key = " << numa_pref << ")");
                             }
                         } // snd/rcv.
@@ -828,7 +828,7 @@ namespace yask {
         }
         TRACE_MSG("allocScratchData: max mini-block size across pack(s) is " <<
                   mblksize.makeDimValStr(" * "));
-        
+
         // Pass 0: count required size, allocate chunk of memory at end.
         // Pass 1: distribute parts of already-allocated memory chunk.
         for (int pass = 0; pass < 2; pass++) {
@@ -900,5 +900,5 @@ namespace yask {
 
         } // scratch-var passes.
     }
-    
+
 } // namespace yask.

src/kernel/lib/auto_tuner.cpp

@@ -43,7 +43,7 @@ namespace yask {
             _name += "(" + name + ")";
         clear(settings->_do_auto_tune);
     }
-    
+
     // Eval auto-tuner for given number of steps.
     void StencilContext::eval_auto_tuner(idx_t num_steps) {
         STATE_VARS(this);
@@ -57,7 +57,7 @@ namespace yask {
         else
             _at.eval();
     }
-    
+
     // Reset auto-tuners.
     void StencilContext::reset_auto_tuner(bool enable, bool verbose) {
         for (auto& sp : stPacks)
@@ -77,7 +77,7 @@ namespace yask {
             done = _at.is_done();
         return !done;
     }
-    
+
     // Apply auto-tuning immediately, i.e., not as part of normal processing.
     // Will alter data in vars.
     void StencilContext::run_auto_tuner_now(bool verbose) {
@@ -159,7 +159,7 @@ namespace yask {
         DEBUG_MSG(_name << ": sub-block-size: " <<
                   _settings->_sub_block_sizes.removeDim(step_posn).makeDimValStr(" * "));
     }
-    
+
     // Access settings.
     bool AutoTuner::tune_mini_blks() const {
         return _context->get_settings()->_tune_mini_blks;
@@ -210,7 +210,7 @@ namespace yask {
     // Check whether sizes within search limits.
     bool AutoTuner::checkSizes(const IdxTuple& bsize) {
         bool ok = true;
-        
+
         // Too small?
         if (ok && get_num_domain_points(bsize) < min_pts) {
             n2small++;
@@ -228,7 +228,7 @@ namespace yask {
         }
         return ok;
     }
-    
+
     // Evaluate the previous run and take next auto-tuner step.
     void AutoTuner::eval() {
         STATE_VARS(this);
@@ -257,7 +257,7 @@ namespace yask {
                   ctime << " secs (" << rate <<
                   " steps/sec) cumulative; best-rate = " << best_rate <<
                   "; min-secs = " << min_secs);
-        
+
         // Still in warmup?
         if (in_warmup) {
 
@@ -413,7 +413,7 @@ namespace yask {
                 // Check sizes.
                 if (ok && !checkSizes(bsize))
                     ok = false;
-                
+
 
                 // Valid size and not already checked?
                 if (ok && results.count(bsize) == 0) {
@@ -475,7 +475,7 @@ namespace yask {
 
         // Restore step-dim value for block.
         target_sizes()[step_posn] = target_steps;
-        
+
         // Change derived sizes to 0 so adjustSettings()
         // will set them to the default.
         if (!tune_mini_blks()) {

src/kernel/lib/auto_tuner.hpp

@@ -43,7 +43,7 @@ namespace yask {
         // Null stream to throw away debug info.
         yask_output_factory yof;
         yask_output_ptr nullop = yof.new_null_output();
-        
+
         // Whether to print progress.
         bool verbose = false;
 
@@ -112,12 +112,12 @@ namespace yask {
             return tune_mini_blks() ?
                 _settings->_block_sizes : _settings->_region_sizes;
         }
-        
+
         // Change settings pointers.
         void set_settings(KernelSettings* p) {
             _settings = p;
         }
-        
+
         // Reset all state to beginning.
         void clear(bool mark_done, bool verbose = false);