Kernels: more generic temporary workspace checks

bench/blas.cpp

@@ -265,7 +265,6 @@ void bench(isc::numeric_type dtype, std::string operation)
   #endif
       std::cout << std::endl;
     }
-    std::cout << "\n\n" << std::flush;
   }
 
   if(operation.substr(0, 4)=="gemv")
@@ -334,6 +333,8 @@ void bench(isc::numeric_type dtype, std::string operation)
     MNKs.push_back(std::make_tuple('N','N',169,384,2304));
     MNKs.push_back(std::make_tuple('N','N',169,192,1728));
     MNKs.push_back(std::make_tuple('N','N',169,128,1728));
+    MNKs.push_back(std::make_tuple('N','T',256,4096,9216));
+
     //AlexNet (Backward)
     MNKs.push_back(std::make_tuple('T','N',1728,128,169));
     MNKs.push_back(std::make_tuple('T','N',1728,192,169));

include/isaac/kernels/templates/base.h

@@ -67,6 +67,7 @@ private:
   virtual std::string generate_impl(std::string const & suffix, expressions_tuple const & expressions, driver::Device const & device, std::vector<mapping_type> const & mapping) const = 0;
 public:
   base(binding_policy_t binding_policy);
+  virtual unsigned int temporary_workspace(expressions_tuple const &) const;
   virtual unsigned int lmem_usage(expressions_tuple const &) const;
   virtual unsigned int registers_usage(expressions_tuple const &) const;
   virtual std::vector<int_t> input_sizes(expressions_tuple const & expressions) const = 0;

include/isaac/kernels/templates/gemm.h

@@ -42,6 +42,7 @@ struct gemm_parameters : public base::parameters_type
 class gemm : public base_impl<gemm, gemm_parameters>
 {
 private:
+  unsigned int temporary_workspace(expressions_tuple const & expressions) const;
   unsigned int lmem_usage(expressions_tuple const & expressions) const;
   unsigned int registers_usage(expressions_tuple const & expressions) const;
   int is_invalid_impl(driver::Device const &, expressions_tuple const &) const;

lib/kernels/templates/base.cpp

@@ -73,6 +73,9 @@ unsigned int base::lmem_usage(expressions_tuple const &) const
 unsigned int base::registers_usage(expressions_tuple const &) const
 { return 0; }
 
+unsigned int base::temporary_workspace(expressions_tuple const &) const
+{ return 0; }
+
 base::~base()
 { }
 
@@ -81,7 +84,8 @@ std::string base::generate(std::string const & suffix, expressions_tuple const &
   expressions_tuple::data_type::const_iterator sit;
   std::vector<mapping_type>::iterator mit;
 
-  if(int err = is_invalid(expressions, device))
+  int err = is_invalid(expressions, device);
+  if(err != 0 && err != TEMPLATE_TEMPORARY_TOO_LARGE)
     throw operation_not_supported_exception("The supplied parameters for this template are invalid : err " + tools::to_string(err));
 
   //Create mapping
@@ -142,6 +146,10 @@ int base_impl<TType, PType>::is_invalid(expressions_tuple const & expressions, d
   if (p_.simd_width!=1 && p_.simd_width!=2 && p_.simd_width!=3 && p_.simd_width!=4)
     return TEMPLATE_INVALID_SIMD_WIDTH;
 
+  //Temporary workspace
+  if(temporary_workspace(expressions) > 2e6)
+      return TEMPLATE_TEMPORARY_TOO_LARGE;
+
   return is_invalid_impl(device, expressions);
 }
 

lib/kernels/templates/gemm.cpp

@@ -46,16 +46,20 @@ gemm_parameters::gemm_parameters(unsigned int simd_width
     return N*size_of(numeric_t);
   }
 
-  int gemm::is_invalid_impl(driver::Device const &, expressions_tuple const & expressions) const
+  unsigned int gemm::temporary_workspace(expressions_tuple const & expressions) const
   {
-    std::vector<int_t> MNK = input_sizes(expressions);
-    int_t M = MNK[0]; int_t N = MNK[1];
+      std::vector<int_t> MNK = input_sizes(expressions);
+      int_t M = MNK[0]; int_t N = MNK[1];
+      if(p_.depth > 1)
+        return M*N*p_.depth;
+      return 0;
+  }
 
+  int gemm::is_invalid_impl(driver::Device const &, expressions_tuple const &) const
+  {
     if(p_.A_fetching_policy!=FETCH_FROM_LOCAL || p_.B_fetching_policy!=FETCH_FROM_LOCAL)
       return TEMPLATE_INVALID_FETCHING_POLICY_TYPE;
 
-    if(p_.depth > 1 && M*N*p_.depth > 2e6)
-      return TEMPLATE_TEMPORARY_TOO_LARGE;
 
     if ((p_.mS % p_.simd_width) > 0 || (p_.nS % p_.simd_width) > 0)
       return TEMPLATE_MS_NS_MUST_BE_SIMD_WIDTH_MULTIPLE;
@@ -66,8 +70,7 @@ gemm_parameters::gemm_parameters(unsigned int simd_width
     if ( p_.kS % p_.kL == 0)
       return TEMPLATE_KS_MUST_BE_SMALLER_THAN_KL;
 
-    if (p_.A_fetching_policy==FETCH_FROM_LOCAL || p_.B_fetching_policy==FETCH_FROM_LOCAL)
-    {
+    if (p_.A_fetching_policy==FETCH_FROM_LOCAL || p_.B_fetching_policy==FETCH_FROM_LOCAL){
       if ((p_.local_fetch_0*p_.local_fetch_1) !=(p_.local_size_0*p_.local_size_1))
         return TEMPLATE_LOCAL_FETCH_PRODUCT_MUST_MATCH_LOCAL_SIZE_PRODUCT;
     }

lib/model/model.cpp

@@ -124,6 +124,7 @@ void model::execute(controller<expressions_tuple> const & expr)
   }
 
   //Execution
+//  std::cout << std::endl << "Label: " << label << std::endl;
   return templates_[label]->enqueue(queue_, program, tools::to_string(label), *fallback_, expr);
 }
 

lib/wrap/clBLAS.cpp

@@ -176,6 +176,7 @@ extern "C"
                             cl_uint numCommandQueues, cl_command_queue *commandQueues,\
                             cl_uint numEventsInWaitList, const cl_event *eventWaitList, cl_event *events)\
     {\
+        /*std::cout << transA << " " <<transB << " " << M << " " << N << " " << K << std::endl;*/\
         cl_mem mA = cmA;\
         cl_mem mB = cmB;\
         if(order==clblasRowMajor){\