Fixed pipeline code. Allows multiple runs within the code. Fixed visc.mk to...

Fixed pipeline code. Allows multiple runs within the code. Fixed visc.mk to support SPIR and PTX both

llvm/test/VISC/parboil/benchmarks/pipeline/src/visc/main.cc

@@ -87,12 +87,13 @@ B    : structural element for dilation - erosion ([0 1 0; 1 1 1; 0 1 0])
 L    : output (laplacian of the image)
 Need 2D grid, a thread per pixel
 */
-OutStruct laplacianEstimate(float *I, size_t bytesI,
+void laplacianEstimate(float *I, size_t bytesI,
                             float *B, size_t bytesB,
                             float *L, size_t bytesL,
                             int m, int n) {
 
-  __visc__hint(visc::GPU_TARGET);
+  __visc__hint(visc::DEVICE);
+  //__visc__hint(visc::GPU_TARGET);
   __visc__attributes(2, I, B, 1, L);
   // 3x3 image area
   float imageArea[SZB][SZB];
@@ -163,16 +164,17 @@ OutStruct laplacianEstimate(float *I, size_t bytesI,
     float laplacian = dilatedPixel + erodedPixel - 2 * imageArea[1][1];
     L[gy*n+gx] = laplacian;
   }
-  OutStruct output = {bytesB, bytesL};
+  //OutStruct output = {bytesB, bytesL};
   //if(gx == m-1 && gy == n-1)
     //std::cout << "Exit laplacian\n";
-  return output;
+  __visc__return(bytesB, bytesL);
 }
 
 OutStruct WrapperlaplacianEstimate(float *I, size_t bytesI,
                           float *B, size_t bytesB,
                           float *L, size_t bytesL,
                           int m, int n) {
+  __visc__hint(visc::CPU_TARGET);
   __visc__attributes(2, I, B, 1, L);
   void* LNode = __visc__createNode2D(laplacianEstimate, m, n);
   __visc__bindIn(LNode, 0, 0, 0); // Bind I
@@ -198,11 +200,12 @@ B    : structural element for dilation - erosion ([0 1 0; 1 1 1; 0 1 0])
 S    : output (sign of the image)
 Need 2D grid, a thread per pixel
  */
-OutStruct computeZeroCrossings(float *L, size_t bytesL,
+void computeZeroCrossings(float *L, size_t bytesL,
                           float *B, size_t bytesB,
                           float *S, size_t bytesS,
                           int m, int n) {
-  __visc__hint(visc::GPU_TARGET);
+  __visc__hint(visc::DEVICE);
+  //__visc__hint(visc::SPIR_TARGET);
   __visc__attributes(2, L, B, 1, S);
 
   // 3x3 image area
@@ -273,16 +276,17 @@ OutStruct computeZeroCrossings(float *L, size_t bytesL,
     float pixelSign = dilatedPixel - erodedPixel;
     S[gy*n+gx] = pixelSign;
   }
-  OutStruct output = {bytesB, bytesS};
+  //OutStruct output = {bytesB, bytesS};
   //if(gx == n-1 && gy == n-1)
     //std::cout << "Exit ZC\n";
-  return output;
+  __visc__return(bytesB, bytesS);
 }
 
 OutStruct WrapperComputeZeroCrossings(float *L, size_t bytesL,
                           float *B, size_t bytesB,
                           float *S, size_t bytesS,
                           int m, int n) {
+  __visc__hint(visc::CPU_TARGET);
   __visc__attributes(2, L, B, 1, S);
   void* ZCNode = __visc__createNode2D(computeZeroCrossings, m, n);
   __visc__bindIn(ZCNode, 0, 0, 0); // Bind L
@@ -340,6 +344,9 @@ OutStruct edgeDetection(float *I, size_t bytesI,
 }
 
 }
+#define NUM_RUNS 5
+#define NUM_FRAMES 20
+
 int main (int argc, char *argv[]) {
 
     struct pb_Parameters *params;
@@ -397,8 +404,7 @@ int main (int argc, char *argv[]) {
         matS[i] = 0.0f;
     }
 
-    pb_SwitchToTimer( &timers, pb_TimerID_NONE );
-    
+    pb_SwitchToTimer( &timers, visc_TimerID_COMPUTATION );
     struct InStruct* args = (struct InStruct*)malloc (sizeof(InStruct));
     
     packData(args, &matI.front(), I_sz,
@@ -408,7 +414,9 @@ int main (int argc, char *argv[]) {
                    matIrow, matIcol);
 
     assert(I_sz % BlockSize == 0);
-    void* DFG = __visc__launch(1, edgeDetection, (void*)args);
+    for(unsigned j=0; j<NUM_RUNS; j++) {
+      std::cout << "Run: " << j << "\n";
+      void* DFG = __visc__launch(1, edgeDetection, (void*)args);
 
       llvm_visc_track_mem(&matI[0], I_sz);
       llvm_visc_track_mem(&matL[0], L_sz);
@@ -417,8 +425,10 @@ int main (int argc, char *argv[]) {
       //packData(args, &matA[0], BlockSize, &matB[i], BlockSize, &matC[i], BlockSize, BlockElements);
       __visc__push(DFG, args);
       __visc__push(DFG, args);
-      __visc__push(DFG, args);
-      __visc__pop(DFG);
+      for(unsigned i=0; i<NUM_FRAMES-2; i++) {
+        __visc__push(DFG, args);
+        __visc__pop(DFG);
+      }
       __visc__pop(DFG);
       __visc__pop(DFG);
       llvm_visc_request_mem(&matS[0], S_sz);
@@ -440,8 +450,8 @@ int main (int argc, char *argv[]) {
       //llvm_visc_untrack_mem(&matC[i]);
     //}
 
-    __visc__wait(DFG);
-
+      __visc__wait(DFG);
+    }
     pb_SwitchToTimer( &timers, pb_TimerID_COPY );
     //llvm_visc_request_mem(&matC.front(), C_sz);
 

llvm/test/VISC/parboil/common/mk/visc.mk

@@ -103,18 +103,17 @@ TEST_OBJS = $(call INBUILDDIR,$(VISC_OBJS))
 PARBOIL_OBJS = $(call INBUILDDIR,parboil.ll)
 KERNEL = $(TEST_OBJS).kernels.ll
 KERNEL_OPT = $(BUILDDIR)/$(APP).kernels.opt.ll
-ifneq ($(TARGET),x86)
+ifeq ($(TARGET),x86)
+  SPIR_ASSEMBLY = $(TEST_OBJS).kernels.bc
+else ifeq ($(TARGET),seq)
+else
   KERNEL_LINKED = $(BUILDDIR)/$(APP).kernels.linked.ll
   #KERNEL = $(TEST_OBJS).kernels.ll
   PTX_ASSEMBLY = $(TEST_OBJS).nvptx.s
-else
-  SPIR_ASSEMBLY = $(TEST_OBJS).kernels.bc
 endif
 HOST_LINKED = $(BUILDDIR)/$(APP).linked.ll
 HOST = $(BUILDDIR)/$(APP).host.ll
 
-APP_BINS = $(PTX_ASSEMBLY) $(BIN) $(KERNEL)
-
 ifeq ($(DEBUGGER),)
 DEBUGGER=gdb
 endif