local changes to sgemm visc_sh. Removed some comments etc.

llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/main.cc

@@ -91,25 +91,17 @@ void packData(RootIn* args,
     args->grid_y = grid_y;
 }
 
-// TODO: decide between dynamic vs static allocation. Merely a convension - will
-// be translated.
 void Allocation(int block_x, int block_y) {
-    // Memory shared between threadblocks
-    //int bl_x = TILE_N;
-    //int bl_y = TILE_TB_HEIGHT;
-    //void* shB = __visc__malloc(bl_x*bl_y*sizeof(float));
     void* shB = __visc__malloc(block_x*block_y*sizeof(float));
-
-    //__visc__return(shB, bl_x*bl_y*sizeof(float));
     __visc__return(shB, block_x*block_y*sizeof(float));
 }
 
 
-void SgemmLeaf( float* A, size_t bytesA, int lda, float* B, size_t bytesB, int ldb, float* C, size_t bytesC, int ldc, int k, float alpha, float beta, float* shB, size_t bytesshB )
-{
+void SgemmLeaf( float* A, size_t bytesA, int lda, float* B, size_t bytesB, int ldb, float* C, size_t bytesC, int ldc, int k, float alpha, float beta, float* shB, size_t bytesshB) {
     __visc__hint(visc::DEVICE);
     //__visc__hint(visc::SPIR_TARGET);
-    // TODO: shB is not an in or out attribute
+    //__visc__hint(visc::GPU_TARGET);
+
     __visc__attributes(3, A, B, C, 1, C);
 
     void* thisNode = __visc__getNode();
@@ -153,7 +145,7 @@ void SgemmLeaf( float* A, size_t bytesA, int lda, float* B, size_t bytesB, int l
     }
 }
 
-// Thread block node for sgemm - Creates allocation node and leaf (thread) node
+// Work group node for sgemm - Creates allocation node and leaf (work item) node
 void SgemmTB(float *A, size_t bytesA,
              int lda,
              float *B, size_t bytesB,
@@ -192,10 +184,9 @@ void SgemmTB(float *A, size_t bytesA,
     __visc__edge(AllocationNode, SgemmLeafNode, 1, 0, 12, 0); // Edge local_B
     __visc__edge(AllocationNode, SgemmLeafNode, 1, 1, 13, 0); // Edge bytes_local_B
 
-    //TODO: bindOut : for now with out attribute
 }
 
-// Root node for sgemm - Creates thread block node
+// Root node for sgemm - Creates work group node
 void SgemmRoot(float *A, size_t bytesA,
                int lda,
                float *B, size_t bytesB,
@@ -229,7 +220,6 @@ void SgemmRoot(float *A, size_t bytesA,
     __visc__bindIn(SgemmTBNode, 12, 12, 0); // Bind block_x
     __visc__bindIn(SgemmTBNode, 13, 13, 0); // Bind block_y
 
-    //TODO: bindOut : for now with out attribute
 }
 
 // Creates root node for sgemm
@@ -254,14 +244,12 @@ __attribute__((noinline)) void basicSgemm(struct pb_TimerSet* timers, char trans
 
 //    unsigned db[2] = {TILE_N,TILE_TB_HEIGHT};
 //    unsigned dg[2] = {m*TILE_N/TILE_M,n*TILE_TB_HEIGHT/TILE_N};
-//    unsigned dg[2] = {m*db[0]/TILE_M,n*db[1]/TILE_N};
-
-//    unsigned sgemmDFG = __visc__node(mysgemmNT, 2, 2, db[0], db[1], dg[0]/db[0], dg[1]/db[1], 12, A, bytesA, lda, B, bytesB, ldb, C, bytesC, ldc, k, alpha, beta, 0);
 
     int block_x = TILE_N;
     int block_y = TILE_TB_HEIGHT;
     int grid_x = m/TILE_M;
     int grid_y = n/TILE_N;
+
     // Pack data in struct
     RootIn* args = (RootIn*) malloc(sizeof(RootIn));
     packData(args,
@@ -344,7 +332,8 @@ int main (int argc, char *argv[]) {
 
     // Use standard sgemm interface
     basicSgemm(&timers, 'N', 'T', matArow, matBcol, matAcol, 1.0f, \
-               &matA.front(), A_sz, matArow, &matBT.front(), B_sz, matBcol, 0.0f, &matC.front(), C_sz, matArow);
+               &matA.front(), A_sz, matArow, &matBT.front(), B_sz, matBcol,
+               0.0f, &matC.front(), C_sz, matArow);
 
     pb_SwitchToTimer( &timers, pb_TimerID_COPY );
     llvm_visc_request_mem(&matC.front(), C_sz);
@@ -360,11 +349,9 @@ int main (int argc, char *argv[]) {
     __visc__cleanup();
 
     if (params->outFile) {
-
-        /* Write C to file */
-        //pb_SwitchToTimer(&timers, pb_TimerID_IO);
-        writeColMajorMatrixFile(params->outFile,
-                                matArow, matBcol, matC);
+      /* Write C to file */
+      //pb_SwitchToTimer(&timers, pb_TimerID_IO);
+      writeColMajorMatrixFile(params->outFile, matArow, matBcol, matC);
     }
 
     double GPUtime = pb_GetElapsedTime(&(timers.timers[pb_TimerID_KERNEL]));