Merge pull request #165 from jlgreathouse/develop

Kent Knox · Kent Knox · commit ee24b26661b0 · 2015-10-16T14:19:15.000-05:00
Fix for #161 + Print GFLOPs :+1:
diff --git a/src/benchmarks/clsparse-bench/functions/clfunc_xSpMdV.hpp b/src/benchmarks/clsparse-bench/functions/clfunc_xSpMdV.hpp
@@ -71,12 +71,12 @@ class xSpMdV: public clsparseFunc
 
     double gflops( )
     {
-        return 0.0;
+        return ((2 * csrMtx.num_nonzeros) / time_in_ns ( ));
     }
 
     std::string gflops_formula( )
     {
-        return "N/A";
+        return "GFLOPs";
     }
 
     double bandwidth( )
@@ -207,12 +207,15 @@ class xSpMdV: public clsparseFunc
         {
           std::cout << "clSPARSE matrix: " << sparseFile << std::endl;
           size_t sparseBytes = sizeof( cl_int )*( csrMtx.num_nonzeros + csrMtx.num_rows ) + sizeof( T ) * ( csrMtx.num_nonzeros + csrMtx.num_cols + csrMtx.num_rows );
+          size_t sparseFlops = 2 * csrMtx.num_nonzeros;
           cpuTimer->pruneOutliers( 3.0 );
           cpuTimer->Print( sparseBytes, "GiB/s" );
+          cpuTimer->Print( sparseFlops, "GFLOPs" );
           cpuTimer->Reset( );
 
           gpuTimer->pruneOutliers( 3.0 );
           gpuTimer->Print( sparseBytes, "GiB/s" );
+          gpuTimer->Print( sparseFlops, "GFLOPs" );
           gpuTimer->Reset( );
         }
 
diff --git a/src/benchmarks/cusparse-bench/functions/cufunc_xSpMdV.hpp b/src/benchmarks/cusparse-bench/functions/cufunc_xSpMdV.hpp
@@ -51,12 +51,12 @@ class xSpMdV : public cusparseFunc
 
     double gflops( )
     {
-        return 0.0;
+        return ((2 * n_vals) / time_in_ns ( ));
     }
 
     std::string gflops_formula( )
     {
-        return "N/A";
+        return "GFLOPs";
     }
 
     double bandwidth( )
diff --git a/src/benchmarks/cusparse-bench/src/main.cpp b/src/benchmarks/cusparse-bench/src/main.cpp
@@ -278,8 +278,10 @@ int main(int argc, char *argv[])
         timer.pruneOutliers( 3.0 );
         std::cout << "cuSPARSE matrix: " << path << std::endl;
         std::cout << "cuSPARSE kernel execution time < ns >: " << my_function->time_in_ns( ) << std::endl;
-        std::cout << "cuSPARSE kernel execution Gflops < " <<
+        std::cout << "cuSPARSE kernel execution < " <<
             my_function->bandwidth_formula( ) << " >: " << my_function->bandwidth( ) << std::endl << std::endl;
+        std::cout << "cuSPARSE kernel execution < " <<
+            my_function->gflops_formula( ) << " >: " << my_function->gflops( ) << std::endl << std::endl;
       }
 
   }
diff --git a/src/library/io/mm-reader.cpp b/src/library/io/mm-reader.cpp
@@ -568,7 +568,6 @@ clsparseSCsrMatrixfromFile(clsparseCsrMatrix* csrMatx, const char* filePath, cls
     MatrixMarketReader< cl_float > mm_reader;
     if( mm_reader.MMReadFormat( filePath ) )
         return clsparseInvalidFile;
-#if 0
     // BUG: We need to check to see if openCL buffers currently exist and deallocate them first!
     // FIX: Below code will check whether the buffers were allocated in the first place;
     {
@@ -590,7 +589,6 @@ clsparseSCsrMatrixfromFile(clsparseCsrMatrix* csrMatx, const char* filePath, cls
         if (validationStatus != clsparseSuccess)
             return validationStatus;
     }
-#endif
 
     // JPA: Shouldn't that just be an assertion check? It seems to me that
     // the user have to call clsparseHeaderfromFile before calling this function,
diff --git a/src/library/kernels/atomic_reduce.cl b/src/library/kernels/atomic_reduce.cl
@@ -35,7 +35,7 @@ R"(
   #endif
 #endif
 
-#if defined(ATOMIC_FLOAT) || defined (ATOMIC_INT)
+#if __OPENCL_VERSION__ <= CL_VERSION_1_0 && (defined(ATOMIC_FLOAT) || defined (ATOMIC_INT))
   #if defined(cl_khr_global_int32_base_atomics) && defined(cl_khr_global_int32_extended_atomics)
     #pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
     #pragma OPENCL_EXTENSION cl_khr_global_int32_extended_atomics : enable
diff --git a/src/library/kernels/csrmm_adaptive.cl b/src/library/kernels/csrmm_adaptive.cl
@@ -42,8 +42,8 @@ R"(
 #if __OPENCL_VERSION__ > CL_VERSION_1_0
   #define ATOM32
 #elif defined(cl_khr_global_int32_base_atomics) && defined(cl_khr_global_int32_extended_atomics)
-  #pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : require
-  #pragma OPENCL_EXTENSION cl_khr_global_int32_extended_atomics : require
+  #pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+  #pragma OPENCL_EXTENSION cl_khr_global_int32_extended_atomics : enable
   #define ATOM32
 #else
   #error "Required integer atomics not supported by this OpenCL implemenation."
diff --git a/src/library/kernels/csrmv_adaptive.cl b/src/library/kernels/csrmv_adaptive.cl
@@ -35,8 +35,8 @@ R"(
 #if __OPENCL_VERSION__ > CL_VERSION_1_0
   #define ATOM32
 #elif defined(cl_khr_global_int32_base_atomics) && defined(cl_khr_global_int32_extended_atomics)
-  #pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : require
-  #pragma OPENCL_EXTENSION cl_khr_global_int32_extended_atomics : require
+  #pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+  #pragma OPENCL_EXTENSION cl_khr_global_int32_extended_atomics : enable
   #define ATOM32
 #else
   #error "Required integer atomics not supported by this OpenCL implemenation."

Original file line number	Diff line number	Diff line change
`@@ -71,12 +71,12 @@ class xSpMdV: public clsparseFunc`
`71`	`71`
`72`	`72`	`double gflops( )`
`73`	`73`	`{`
`74`		`- return 0.0;`
	`74`	`+ return ((2 * csrMtx.num_nonzeros) / time_in_ns ( ));`
`75`	`75`	`}`
`76`	`76`
`77`	`77`	`std::string gflops_formula( )`
`78`	`78`	`{`
`79`		`- return "N/A";`
	`79`	`+ return "GFLOPs";`
`80`	`80`	`}`
`81`	`81`
`82`	`82`	`double bandwidth( )`
`@@ -207,12 +207,15 @@ class xSpMdV: public clsparseFunc`
`207`	`207`	`{`
`208`	`208`	`std::cout << "clSPARSE matrix: " << sparseFile << std::endl;`
`209`	`209`	`size_t sparseBytes = sizeof( cl_int )( csrMtx.num_nonzeros + csrMtx.num_rows ) + sizeof( T ) ( csrMtx.num_nonzeros + csrMtx.num_cols + csrMtx.num_rows );`
	`210`	`+ size_t sparseFlops = 2 * csrMtx.num_nonzeros;`
`210`	`211`	`cpuTimer->pruneOutliers( 3.0 );`
`211`	`212`	`cpuTimer->Print( sparseBytes, "GiB/s" );`
	`213`	`+ cpuTimer->Print( sparseFlops, "GFLOPs" );`
`212`	`214`	`cpuTimer->Reset( );`
`213`	`215`
`214`	`216`	`gpuTimer->pruneOutliers( 3.0 );`
`215`	`217`	`gpuTimer->Print( sparseBytes, "GiB/s" );`
	`218`	`+ gpuTimer->Print( sparseFlops, "GFLOPs" );`
`216`	`219`	`gpuTimer->Reset( );`
`217`	`220`	`}`
`218`	`221`
Original file line number	Diff line number	Diff line change
`@@ -51,12 +51,12 @@ class xSpMdV : public cusparseFunc`
`51`	`51`
`52`	`52`	`double gflops( )`
`53`	`53`	`{`
`54`		`- return 0.0;`
	`54`	`+ return ((2 * n_vals) / time_in_ns ( ));`
`55`	`55`	`}`
`56`	`56`
`57`	`57`	`std::string gflops_formula( )`
`58`	`58`	`{`
`59`		`- return "N/A";`
	`59`	`+ return "GFLOPs";`
`60`	`60`	`}`
`61`	`61`
`62`	`62`	`double bandwidth( )`
Original file line number	Diff line number	Diff line change
`@@ -278,8 +278,10 @@ int main(int argc, char *argv[])`
`278`	`278`	`timer.pruneOutliers( 3.0 );`
`279`	`279`	`std::cout << "cuSPARSE matrix: " << path << std::endl;`
`280`	`280`	`std::cout << "cuSPARSE kernel execution time < ns >: " << my_function->time_in_ns( ) << std::endl;`
`281`		`- std::cout << "cuSPARSE kernel execution Gflops < " <<`
	`281`	`+ std::cout << "cuSPARSE kernel execution < " <<`
`282`	`282`	`my_function->bandwidth_formula( ) << " >: " << my_function->bandwidth( ) << std::endl << std::endl;`
	`283`	`+ std::cout << "cuSPARSE kernel execution < " <<`
	`284`	`+ my_function->gflops_formula( ) << " >: " << my_function->gflops( ) << std::endl << std::endl;`
`283`	`285`	`}`
`284`	`286`
`285`	`287`	`}`
Original file line number	Diff line number	Diff line change
`@@ -568,7 +568,6 @@ clsparseSCsrMatrixfromFile(clsparseCsrMatrix* csrMatx, const char* filePath, cls`
`568`	`568`	`MatrixMarketReader< cl_float > mm_reader;`
`569`	`569`	`if( mm_reader.MMReadFormat( filePath ) )`
`570`	`570`	`return clsparseInvalidFile;`
`571`		`-#if 0`
`572`	`571`	`// BUG: We need to check to see if openCL buffers currently exist and deallocate them first!`
`573`	`572`	`// FIX: Below code will check whether the buffers were allocated in the first place;`
`574`	`573`	`{`
`@@ -590,7 +589,6 @@ clsparseSCsrMatrixfromFile(clsparseCsrMatrix* csrMatx, const char* filePath, cls`
`590`	`589`	`if (validationStatus != clsparseSuccess)`
`591`	`590`	`return validationStatus;`
`592`	`591`	`}`
`593`		`-#endif`
`594`	`592`
`595`	`593`	`// JPA: Shouldn't that just be an assertion check? It seems to me that`
`596`	`594`	`// the user have to call clsparseHeaderfromFile before calling this function,`