Tune: Further file hierarchy improvements

include/isaac/driver/device.h

@@ -58,6 +58,8 @@ public:
   Vendor vendor() const;
   Architecture architecture() const;
 
+  std::string infos() const;
+
   backend_type backend() const;
   size_t clock_rate() const;
   unsigned int address_bits() const;

lib/driver/device.cpp

@@ -1,5 +1,7 @@
-#include "isaac/driver/device.h"
 #include <algorithm>
+#include <sstream>
+
+#include "isaac/driver/device.h"
 #include "helpers/ocl/infos.hpp"
 
 namespace isaac
@@ -19,13 +21,17 @@ int Device::cuGetInfo() const
 }
 
 Device::Device(int ordinal): backend_(CUDA), h_(backend_, true)
-{ cuda::check(cuDeviceGet(h_.cu.get(), ordinal)); }
+{
+  cuda::check(cuDeviceGet(h_.cu.get(), ordinal));
+}
 
 #endif
 
 
 Device::Device(cl_device_id const & device, bool take_ownership) : backend_(OPENCL), h_(backend_, take_ownership)
-{ h_.cl() = device; }
+{
+    h_.cl() = device;
+}
 
 
 bool Device::operator==(Device const & other) const
@@ -205,6 +211,22 @@ bool Device::fp64_support() const
   }
 }
 
+std::string Device::infos() const
+{
+  std::ostringstream oss;
+  std::vector<size_t> max_wi_sizes = max_work_item_sizes();
+
+  oss << "Platform: " << platform().name() << std::endl;
+  oss << "Vendor: " << vendor_str() << std::endl;
+  oss << "Name: " << name() << std::endl;
+  oss << "Maximum total work-group size: " << max_work_group_size() << std::endl;
+  oss << "Maximum individual work-group sizes: " << max_wi_sizes[0] << ", " << max_wi_sizes[1] << ", " << max_wi_sizes[2] << std::endl;
+  oss << "Local memory size: " << local_mem_size() << std::endl;
+
+  return oss.str();
+}
+
+// Properties
 
 #ifdef ISAAC_WITH_CUDA
     #define CUDACASE(CUNAME) case CUDA: return cuGetInfo<CUNAME>();

lib/kernels/templates/base.cpp

@@ -77,13 +77,13 @@ unsigned int base::temporary_workspace(expressions_tuple const &) const
 { return 0; }
 
 base::~base()
-{ }
+{
+}
 
 std::string base::generate(std::string const & suffix, expressions_tuple const & expressions, driver::Device const & device)
 {
   expressions_tuple::data_type::const_iterator sit;
   std::vector<mapping_type>::iterator mit;
-
   int err = is_invalid(expressions, device);
   if(err != 0)
     throw operation_not_supported_exception("The supplied parameters for this template are invalid : err " + tools::to_string(err));

lib/kernels/templates/gemm.cpp

@@ -22,7 +22,9 @@ gemm_parameters::gemm_parameters(unsigned int simd_width
                           , unsigned int local_fetch_0, unsigned int local_fetch_1): base::parameters_type(simd_width, local_size_0, local_size_1, 1),
   kL(KL), depth(D), mS(ms), kS(ks), nS(ns), A_fetching_policy(A_fetching_policy), B_fetching_policy(B_fetching_policy),
   local_fetch_0(local_fetch_0), local_fetch_1(local_fetch_1),
-  mL(ms*local_size_0), nL(ns*local_size_1){}
+  mL(ms*local_size_0), nL(ns*local_size_1)
+{
+}
 
 
   unsigned int gemm::lmem_usage(expressions_tuple const & expressions) const
@@ -59,7 +61,6 @@ gemm_parameters::gemm_parameters(unsigned int simd_width
     if(p_.A_fetching_policy!=FETCH_FROM_LOCAL || p_.B_fetching_policy!=FETCH_FROM_LOCAL)
       return TEMPLATE_INVALID_FETCHING_POLICY_TYPE;
 
-
     if ((p_.mS % p_.simd_width) > 0 || (p_.nS % p_.simd_width) > 0)
       return TEMPLATE_MS_NS_MUST_BE_SIMD_WIDTH_MULTIPLE;
 
@@ -459,23 +460,23 @@ gemm_parameters::gemm_parameters(unsigned int simd_width
     stream << "}" << std::endl;
 
 
-    if(A_trans_=='N' || B_trans_=='T')
-        stream << "int Ky = K - idT.y;" << std::endl;
-    if(A_trans_=='T' || B_trans_=='N')
-        stream << "int Kx = K - idT.x;" << std::endl;
+//    if(A_trans_=='N' || B_trans_=='T')
+//        stream << "int Ky = K - idT.y;" << std::endl;
+//    if(A_trans_=='T' || B_trans_=='N')
+//        stream << "int Kx = K - idT.x;" << std::endl;
 
-    if(A_trans_=='N' || B_trans_=='T')
-        for(unsigned int k = 0; k < p_.kL; k += p_.local_fetch_1)
-            stream << "int condy" << k << " = " << k << " < Ky;" << std::endl;
+//    if(A_trans_=='N' || B_trans_=='T')
+//        for(unsigned int k = 0; k < p_.kL; k += p_.local_fetch_1)
+//            stream << "int condy" << k << " = " << k << " < Ky;" << std::endl;
 
-    if(A_trans_=='T' || B_trans_=='N')
-    {
-        for(unsigned int k = 0 ; k < p_.kL ; k += p_.local_fetch_0*p_.simd_width)
-            for(unsigned int s = 0 ; s < p_.simd_width ; ++s)
-            stream << "int condx" << k + s << " = " << k + s << " < Kx;" << std::endl;
-    }
+//    if(A_trans_=='T' || B_trans_=='N')
+//    {
+//        for(unsigned int k = 0 ; k < p_.kL ; k += p_.local_fetch_0*p_.simd_width)
+//            for(unsigned int s = 0 ; s < p_.simd_width ; ++s)
+//            stream << "int condx" << k + s << " = " << k + s << " < Kx;" << std::endl;
+//    }
 
-    fetch_to_lds(true);
+//    fetch_to_lds(true);
 
     stream << "//Write back C" << std::endl;
     stream << "M += ids.x;" << std::endl;
@@ -567,9 +568,6 @@ gemm_parameters::gemm_parameters(unsigned int simd_width
       stream << "}" << std::endl;
     }
 
-
-//    if(p_.simd_width>1)
-//        std::cout << stream.str() << std::endl;
     return stream.str();
 
 #undef VLOAD
@@ -746,7 +744,8 @@ gemm_parameters::gemm_parameters(unsigned int simd_width
                            , fetching_policy_type Afetch , fetching_policy_type Bfetch
                            , int_t lfetch0, int_t lfetch1, bool check_bound) :
     gemm(gemm_parameters(simd, ls0, KL, ls1, D, ms, ks, ns, Afetch, Bfetch, lfetch0, lfetch1), check_bound, 'N', 'N')
-  { }
+  {
+  }
 
   //
   gemm_tn::gemm_tn(unsigned int simd

lib/profiles/profiles.cpp

@@ -217,6 +217,7 @@ profiles::map_type& profiles::init(driver::CommandQueue const & queue)
   driver::Device const & device = queue.device();
   presets_type::const_iterator it = presets_.find(std::make_tuple(device.type(), device.vendor(), device.architecture()));
   if(it==presets_.end()){
+        //FIXME: Hadle this case
 //      import(presets_.at(std::make_tuple(device.type(), device.vendor(), driver::Device::Architecture::UNKNOWN)), queue);
   }
   else

python/setup.py

@@ -143,7 +143,7 @@ def main():
                     libraries=libraries)]
     
     #External
-    extensions += [Extension('autotuning.external.sklearn._tree',
+    extensions += [Extension('external.sklearn._tree',
                              ['external/sklearn/_tree.c'],
                              include_dirs = [numpy_include])]
 
@@ -155,7 +155,7 @@ def main():
                 author='Philippe Tillet',
                 author_email='ptillet@g.harvard.edu',
                 license='MPL 2.0',
-                packages=['isaac','isaac.autotuning', 'isaac.autotuning.external', 'isaac.autotuning.external.deap', 'isaac.autotuning.external.deap.tools', 'isaac.autotuning.external.sklearn'],
+                packages=['isaac','isaac.external','isaac.external.sklearn'],
                 ext_package="isaac",
                 ext_modules=extensions,
                 cmdclass={'build_py': build_py, 'build_ext': build_ext_subclass},

python/src/bind/driver.cpp

@@ -139,6 +139,7 @@ void export_driver()
       .add_property("platform", &sc::driver::Device::platform)
       .add_property("vendor", &sc::driver::Device::vendor)
       .add_property("nv_compute_capability", &detail::nv_compute_capability)
+      .add_property("infos", &sc::driver::Device::infos)
       ;
 
   bp::class_<sc::driver::Context, boost::noncopyable>("context", bp::no_init)

python/src/bind/exceptions.cpp

@@ -92,4 +92,16 @@ void export_exceptions()
     wrap::exception<isaac::driver::ocl::exception::mem_object_allocation_failure>("MemObjectAllocationFailure")
         .def("__str__", &isaac::driver::ocl::exception::mem_object_allocation_failure::what)
         ;
+
+    wrap::exception<isaac::driver::ocl::exception::out_of_host_memory>("OutOfHostMemory")
+        .def("__str__", &isaac::driver::ocl::exception::out_of_host_memory::what)
+        ;
+
+    wrap::exception<isaac::driver::ocl::exception::invalid_work_group_size>("InvalidWorkGroupSize")
+        .def("__str__", &isaac::driver::ocl::exception::invalid_work_group_size::what)
+        ;
+
+    wrap::exception<isaac::driver::ocl::exception::invalid_value>("InvalidValue")
+        .def("__str__", &isaac::driver::ocl::exception::invalid_value::what)
+        ;
 }

python/src/bind/kernels.cpp

@@ -29,11 +29,11 @@ void export_templates()
 
 
   bp::enum_<tpt::fetching_policy_type>
-      ("fetching_policy_type");
+      ("fetching_policy_type")
+     .value("FETCH_FROM_LOCAL", tpt::FETCH_FROM_LOCAL)
+     .value("FETCH_FROM_GLOBAL_STRIDED", tpt::FETCH_FROM_GLOBAL_STRIDED)
+     .value("FETCH_FROM_GLOBAL_CONTIGUOUS", tpt::FETCH_FROM_GLOBAL_CONTIGUOUS);
 
-  bp::scope().attr("FETCH_FROM_LOCAL") = tpt::FETCH_FROM_LOCAL;
-  bp::scope().attr("FETCH_FROM_GLOBAL_STRIDED") = tpt::FETCH_FROM_GLOBAL_CONTIGUOUS;
-  bp::scope().attr("FETCH_FROM_GLOBAL_CONTIGUOUS") = tpt::FETCH_FROM_GLOBAL_STRIDED;
 
   //Base
   {

tune/android/main.py

@@ -13,7 +13,7 @@ from kivy.uix.settings import SettingsWithNoMenu
 import isaac as sc
 import json
 
-from isaac.autotuning.tune import tune
+from tune.tune import do_tuning
 
 __version__ = '1.0'
 
@@ -62,7 +62,8 @@ class IsaacApp(App):
         json_path = ''
         #FIXME: Move profiling logics into tuning
         sc.driver.default.queue_properties = sc.driver.PROFILING_ENABLE
-        tune(device, operation, json_path)
+        print device.infos
+        do_tuning(device, operation, json_path)
         
     def show_benchmark(self):
         pass

tune/android/tune/model.py

@@ -1,4 +1,4 @@
-from external.sklearn.forest import RandomForestRegressor
+from isaac.external.sklearn.forest import RandomForestRegressor
 import numpy as np
 
 def gmean(a, axis=0, dtype=None):

tune/android/tune/optimize.py

@@ -13,10 +13,12 @@ from numpy import cumsum
 
 import tools
 
-fetch_types = [sc.templates.FETCH_FROM_GLOBAL_CONTIGUOUS,
-               sc.templates.FETCH_FROM_GLOBAL_STRIDED,
-               sc.templates.FETCH_FROM_LOCAL,
-               sc.templates.FETCH_FROM_LOCAL]
+fetch_types = [sc.templates.fetching_policy_type.FETCH_FROM_GLOBAL_CONTIGUOUS,
+               sc.templates.fetching_policy_type.FETCH_FROM_GLOBAL_STRIDED,
+               sc.templates.fetching_policy_type.FETCH_FROM_LOCAL,
+               sc.templates.fetching_policy_type.FETCH_FROM_LOCAL]
+
+to_catch = (sc.OperationNotSupported, sc.LaunchOutOfResources, sc.MemObjectAllocationFailure, sc.InvalidWorkGroupSize, sc.OutOfHostMemory, sc.InvalidValue)
 
 def exhaustive(template, sizes, context):
     tree, _ = tools.tree_of(template, sizes, context)
@@ -33,7 +35,7 @@ def exhaustive(template, sizes, context):
             time = tools.benchmark(template, parameters, tree)
             if not best or time < best[1]:
                 best = parameters, time
-        except (sc.OperationNotSupported, sc.LaunchOutOfResources, sc.MemObjectAllocationFailure):
+        except to_catch:
             pass
         if best:
             stdout.write('%.2f %% | Best %.2f [ for %s ]\r'%(float(idx*100)/len(ranges),metric(sizes, best[1]), best[0]))
@@ -73,7 +75,6 @@ def genetic(template, sizes, context, naccept=200, niter = 1000, cxpb=0.4, mutpb
     def evaluate(genome):
         idx = tuple(genome)
         if idx not in cache:
-            print decode(genome)
             cache[idx] = tools.benchmark(template, decode(genome), tree)
         return cache[idx],
         
@@ -100,7 +101,7 @@ def genetic(template, sizes, context, naccept=200, niter = 1000, cxpb=0.4, mutpb
         try:
             individual.fitness.values = toolbox.evaluate(genome)
             population += [individual]
-        except (sc.OperationNotSupported, sc.LaunchOutOfResources, sc.MemObjectAllocationFailure ):
+        except to_catch:
             pass
         genome = encode(list(initializer.next()))
     hof.update(population)
@@ -134,7 +135,7 @@ def genetic(template, sizes, context, naccept=200, niter = 1000, cxpb=0.4, mutpb
                 #Reproduction
                 else: 
                     offspring += [random.choice(population)]
-            except (sc.OperationNotSupported, sc.LaunchOutOfResources, sc.MemObjectAllocationFailure):
+            except to_catch:
                 pass
 
 
@@ -173,7 +174,7 @@ def is_local_optimum(parameters, template, sizes, context):
     #Evaluate the provided parameters guess
     try:
         reference = tools.benchmark(template, parameters, tree)
-    except (sc.OperationNotSupported, sc.LaunchOutOfResources, sc.MemObjectAllocationFailure):
+    except to_catch:
         return False
         
     #Latency bound -- ignore
@@ -190,7 +191,7 @@ def is_local_optimum(parameters, template, sizes, context):
             time = tools.benchmark(template, x, tree)
             if time/reference < .97:
                 return False
-        except (sc.OperationNotSupported, sc.LaunchOutOfResources, sc.MemObjectAllocationFailure):
+        except to_catch:
             pass
     return True
     

tune/android/tune/tune.py

@@ -4,7 +4,7 @@ from itertools import chain, product
 from numpy import argsort, argmax
 from operator import mul
 import isaac as sc
-from external.sklearn.forest import RandomForestRegressor
+from isaac.external.sklearn.forest import RandomForestRegressor
 import optimize, tools, model
 from json import encoder
 import json
@@ -21,7 +21,7 @@ def pow2range(a, b):
     return [2**x for x in range(a, b)]
 
 
-def tune(device, operation, json_path): 
+def do_tuning(device, operation, json_path): 
     #Context
     context = sc.driver.context(device)
     

tune/main.py

@@ -1,6 +1,6 @@
 import argparse
 import isaac as sc
-from isaac.autotuning.tune import tune
+from tune.tune import do_tuning
 
 def parse_arguments():
     platforms = sc.driver.get_platforms()
@@ -32,4 +32,4 @@ def parse_arguments():
 if __name__ == "__main__":
     sc.driver.default.queue_properties = sc.driver.PROFILING_ENABLE
     args = parse_arguments()
-    tune(*args)
+    do_tuning(*args)