Elementwise: Bugfix for FETCH_LOCAL_CONTIGUOUS

lib/jit/generation/elementwise_2d.cpp

@@ -47,7 +47,6 @@ expression_type elementwise_2d::type() const
 
 std::string elementwise_2d::generate_impl(std::string const & suffix, expression_tree const  & tree, driver::Device const & device, symbolic::symbols_table const & symbols) const
 {
-  std::string init0, upper_bound0, inc0, init1, upper_bound1, inc1;
   driver::backend_type backend = device.backend();
   kernel_generation_stream stream(backend);
 
@@ -70,35 +69,24 @@ std::string elementwise_2d::generate_impl(std::string const & suffix, expression
   stream << "{" << std::endl;
   stream.inc_tab();
 
+  element_wise_loop_1D(stream, fetch_, 1, "i", "M", "$GLOBAL_IDX_0", "$GLOBAL_SIZE_0", device, [&](unsigned int){
+    element_wise_loop_1D(stream, fetch_, 1, "j", "N", "$GLOBAL_IDX_1", "$GLOBAL_SIZE_1", device, [&](unsigned int){
+      //Declares register to store results
+      for(symbolic::leaf* sym: symbolic::extract<symbolic::leaf>(tree, symbols, assigned_left, false))
+        stream << sym->process("#scalartype #name;") << std::endl;
 
-  fetching_loop_info(fetch_, "M", stream, init0, upper_bound0, inc0,  "$GLOBAL_IDX_0", "$GLOBAL_SIZE_0", device);
+      //Load to registers
-  stream << "for($SIZE_T i = " << init0 << "; i < " << upper_bound0 << "; i += " << inc0 << ")" << std::endl;
+      for(symbolic::leaf* sym: symbolic::extract<symbolic::leaf>(tree, symbols, assigned_right, false))
-  stream << "{" << std::endl;
+        stream << sym->process("#scalartype #name = at(i, j);") << std::endl;
-  stream.inc_tab();
-  fetching_loop_info(fetch_, "N", stream, init1, upper_bound1, inc1, "$GLOBAL_IDX_1", "$GLOBAL_SIZE_1", device);
-  stream << "for($SIZE_T j = " << init1 << "; j < " << upper_bound1 << "; j += " << inc1 << ")" << std::endl;
-  stream << "{" << std::endl;
-  stream.inc_tab();
 
-  //Declares register to store results
+      for(std::size_t idx: assigned)
-  for(symbolic::leaf* sym: symbolic::extract<symbolic::leaf>(tree, symbols, assigned_left, false))
+        stream << symbols.at(idx)->evaluate({{"leaf", "#name"}}) << ";" << std::endl;
-    stream << sym->process("#scalartype #name;") << std::endl;
 
-  //Load to registers
+      //Writes back
-  for(symbolic::leaf* sym: symbolic::extract<symbolic::leaf>(tree, symbols, assigned_right, false))
+      for(symbolic::leaf* sym: symbolic::extract<symbolic::leaf>(tree, symbols, assigned_left, false))
-    stream << sym->process("#scalartype #name = at(i, j);") << std::endl;
+        stream << sym->process("at(i, j) = #name;") << std::endl;
-
+    });
-  for(std::size_t idx: assigned)
+  });
-    stream << symbols.at(idx)->evaluate({{"leaf", "#name"}}) << ";" << std::endl;
-
-  //Writes back
-  for(symbolic::leaf* sym: symbolic::extract<symbolic::leaf>(tree, symbols, assigned_left, false))
-    stream << sym->process("at(i, j) = #name;") << std::endl;
-
-  stream.dec_tab();
-  stream << "}" << std::endl;
-  stream.dec_tab();
-  stream << "}" << std::endl;
 
 
   stream.dec_tab();

lib/jit/generation/tools/loop.hpp

@@ -29,13 +29,13 @@ namespace isaac
 namespace templates
 {
 
-inline void fetching_loop_info(fetch_type policy, std::string const & bound, kernel_generation_stream & stream, std::string & init, std::string & upper_bound, std::string & inc, std::string const & domain_id, std::string const & domain_size, driver::Device const &)
+inline void fetching_loop_info(fetch_type policy, std::string const & bound, kernel_generation_stream & stream, std::string & init, std::string & upper_bound, std::string & inc, std::string const & domain_id, std::string const & domain_size, driver::Device const &, std::string const & vwidth)
 {
   if (policy==FETCH_FROM_GLOBAL_STRIDED)
   {
-    init = domain_id;
+    init = domain_id + "*" + vwidth;
     upper_bound = bound;
-    inc = domain_size;
+    inc = domain_size + "*" + vwidth;
   }
   else if (policy==FETCH_FROM_GLOBAL_CONTIGUOUS)
   {
@@ -43,12 +43,12 @@ inline void fetching_loop_info(fetch_type policy, std::string const & bound, ker
     std::string chunk_start = "chunk_start";
     std::string chunk_end = "chunk_end";
 
-    stream << "$SIZE_T " << chunk_size << " = (" << bound << "+" << domain_size << "-1)/" << domain_size << ";" << std::endl;
+    stream << "$SIZE_T " << chunk_size << " = " << vwidth << "*(" << bound << "+" << domain_size << "-1)/(" << vwidth << ");" << std::endl;
     stream << "$SIZE_T " << chunk_start << " =" << domain_id << "*" << chunk_size << ";" << std::endl;
     stream << "$SIZE_T " << chunk_end << " = min(" << chunk_start << "+" << chunk_size << ", " << bound << ");" << std::endl;
     init = chunk_start;
     upper_bound = chunk_end;
-    inc = "1";
+    inc = vwidth;
   }
 }
 
@@ -60,9 +60,9 @@ inline void element_wise_loop_1D(kernel_generation_stream & stream, fetch_type f
   std::string strwidth = tools::to_string(vwidth);
 
   std::string init, upper_bound, inc;
-  fetching_loop_info(fetch, bound, stream, init, upper_bound, inc, domain_id, domain_size, device);
+  fetching_loop_info(fetch, bound, stream, init, upper_bound, inc, domain_id, domain_size, device, strwidth);
   std::string boundround = upper_bound + "/" + strwidth + "*" + strwidth;
-  stream << "for(unsigned int " << i << " = " << init << "*" << strwidth << "; " << i << " < " << boundround << "; " << i << " += " << inc << "*" << strwidth << ")" << std::endl;
+  stream << "for(unsigned int " << i << " = " << init << "; " << i << " < " << boundround << "; " << i << " += " << inc << ")" << std::endl;
   stream << "{" << std::endl;
   stream.inc_tab();
   generate_body(vwidth);

python/setup.py

@@ -73,7 +73,7 @@ def main():
       libraries += ['gnustl_shared']
 
     #Source files
-    src =  'src/lib/runtime/predictors/random_forest.cpp src/lib/runtime/profiles.cpp src/lib/runtime/database.cpp src/lib/runtime/execute.cpp src/lib/exception/driver.cpp src/lib/exception/api.cpp src/lib/random/rand.cpp src/lib/jit/generation/elementwise_1d.cpp src/lib/jit/generation/reduce_2d.cpp src/lib/jit/generation/reduce_1d.cpp src/lib/jit/generation/base.cpp src/lib/jit/generation/gemm.cpp src/lib/jit/generation/engine/keywords.cpp src/lib/jit/generation/engine/stream.cpp src/lib/jit/generation/elementwise_2d.cpp src/lib/jit/syntax/expression/expression.cpp src/lib/jit/syntax/expression/preset.cpp src/lib/jit/syntax/expression/operations.cpp src/lib/jit/syntax/engine/binder.cpp src/lib/jit/syntax/engine/macro.cpp src/lib/jit/syntax/engine/process.cpp src/lib/jit/syntax/engine/object.cpp src/lib/value_scalar.cpp src/lib/array.cpp src/lib/api/blas/cublas.cpp src/lib/api/blas/clBLAS.cpp src/lib/driver/dispatch.cpp src/lib/driver/kernel.cpp src/lib/driver/backend.cpp src/lib/driver/platform.cpp src/lib/driver/buffer.cpp src/lib/driver/event.cpp src/lib/driver/ndrange.cpp src/lib/driver/device.cpp src/lib/driver/program_cache.cpp src/lib/driver/check.cpp src/lib/driver/command_queue.cpp src/lib/driver/handle.cpp src/lib/driver/context.cpp src/lib/driver/program.cpp '.split() + [os.path.join('src', 'bind', sf)  for sf in ['_isaac.cpp', 'core.cpp', 'driver.cpp', 'kernels.cpp', 'exceptions.cpp']]
+    src =  'src/lib/exception/api.cpp src/lib/exception/driver.cpp src/lib/value_scalar.cpp src/lib/random/rand.cpp src/lib/driver/check.cpp src/lib/driver/ndrange.cpp src/lib/driver/platform.cpp src/lib/driver/backend.cpp src/lib/driver/program.cpp src/lib/driver/command_queue.cpp src/lib/driver/event.cpp src/lib/driver/kernel.cpp src/lib/driver/handle.cpp src/lib/driver/device.cpp src/lib/driver/program_cache.cpp src/lib/driver/buffer.cpp src/lib/driver/context.cpp src/lib/driver/dispatch.cpp src/lib/jit/generation/engine/stream.cpp src/lib/jit/generation/engine/keywords.cpp src/lib/jit/generation/reduce_1d.cpp src/lib/jit/generation/elementwise_1d.cpp src/lib/jit/generation/base.cpp src/lib/jit/generation/elementwise_2d.cpp src/lib/jit/generation/reduce_2d.cpp src/lib/jit/generation/gemm.cpp src/lib/jit/syntax/engine/object.cpp src/lib/jit/syntax/engine/macro.cpp src/lib/jit/syntax/engine/process.cpp src/lib/jit/syntax/engine/binder.cpp src/lib/jit/syntax/expression/operations.cpp src/lib/jit/syntax/expression/expression.cpp src/lib/jit/syntax/expression/preset.cpp src/lib/api/blas/clBLAS.cpp src/lib/api/blas/cublas.cpp src/lib/runtime/execute.cpp src/lib/runtime/predictors/random_forest.cpp src/lib/runtime/profiles.cpp src/lib/runtime/database.cpp src/lib/array.cpp '.split() + [os.path.join('src', 'bind', sf)  for sf in ['_isaac.cpp', 'core.cpp', 'driver.cpp', 'kernels.cpp', 'exceptions.cpp']]
     boostsrc = 'external/boost/libs/'
     for s in ['numpy','python','smart_ptr','system','thread']:
         src = src + [x for x in recursive_glob('external/boost/libs/' + s + '/src/','.cpp') if 'win32' not in x and 'pthread' not in x]

tune/android/tune/tune.py

@@ -128,6 +128,9 @@ class Tuner:
                 profiles = [map(mmap,row) for v in row for row in csv.reader(f, delimiter=',')]
             with open(os.path.join(savepath, 'Y.csv')) as f:
                 Y = [map(float, row) for row in csv.reader(f, delimiter=',')]
+            #for x in X:
+            #    tree, _ = tools.tree_of(operation, x, context)
+            #    Y.append([performance(x, tools.benchmark(operation(*best), tree)) for best in profiles])
         except:
             pass
         
@@ -181,7 +184,6 @@ class Tuner:
                 row = Y[X.index(x)]
                 self.progress_bar.update(1, 1, profiles[argmax(row)], max(row))
         self.progress_bar.set_finished()
-        
         #Adding external profiles
         for prof in tools.external_profiles(operation):
 			for x, y in zip(X, Y):