Merge pull request #202 from arrayfire/arrayfire-release-test

Fixing issue with beta == 0 in AutoGemm kernels

Author: guacamoleo

src/library/blas/AutoGemm/KernelOpenCL.py

@@ -162,12 +162,6 @@ def makeOpenCLKernelString(kernel):
         "  REG.s0 = mad( -ALPHA.s1, REG.s1, REG.s0 ); \\\\" + endLine +
         "  REG.s1 *= ALPHA.s0; \\\\" + endLine +
         "  REG.s1 = mad(  ALPHA.s1, type_mad_tmp, REG.s1 ); \\\\" + endLine +
-        "  /* (2) */ \\\\" + endLine +
-        "  REG.s0 = mad(  BETA.s0, DST.s0, REG.s0 ); \\\\" + endLine +
-        "  REG.s0 = mad( -BETA.s1, DST.s1, REG.s0 ); \\\\" + endLine +
-        "  REG.s1 = mad(  BETA.s1, DST.s0, REG.s1 ); \\\\" + endLine +
-        "  REG.s1 = mad(  BETA.s0, DST.s1, REG.s1 ); \\\\" + endLine +
-        "  /* (3) */ \\\\" + endLine +
         "  DST = REG;" + endLine )
 
   ####################################
@@ -357,11 +351,11 @@ def makeOpenCLKernelString(kernel):
   kStr += endLine
   kStr += "    /* load global -> local */" + endLine
   numALoads  = (kernel.workGroupNumRows*kernel.microTileNumRows*kernel.unroll) \
-      / (kernel.workGroupNumRows*kernel.workGroupNumCols)
+      // (kernel.workGroupNumRows*kernel.workGroupNumCols) # // -- integer divide
   numALoadsR = (kernel.workGroupNumRows*kernel.microTileNumRows*kernel.unroll) \
       % (kernel.workGroupNumRows*kernel.workGroupNumCols)
   numBLoads  = (kernel.workGroupNumCols*kernel.microTileNumCols*kernel.unroll) \
-      / (kernel.workGroupNumRows*kernel.workGroupNumCols)
+      // (kernel.workGroupNumRows*kernel.workGroupNumCols) # // - integer divide
   numBLoadsR = (kernel.workGroupNumCols*kernel.microTileNumCols*kernel.unroll) \
       % (kernel.workGroupNumRows*kernel.workGroupNumCols)
 

src/library/blas/AutoGemm/KernelParameters.py

@@ -89,11 +89,11 @@ def isValid(self):
     return True
     """
     numALoads = (self.workGroupNumRows*self.microTileNumRows*self.unroll) \
-        / (self.workGroupNumRows*self.workGroupNumCols)
+        // (self.workGroupNumRows*self.workGroupNumCols)
     numALoadsR = (self.workGroupNumRows*self.microTileNumRows*self.unroll) \
         % (self.workGroupNumRows*self.workGroupNumCols)
     numBLoads = (self.workGroupNumCols*self.microTileNumCols*self.unroll) \
-        / (self.workGroupNumRows*self.workGroupNumCols)
+        // (self.workGroupNumRows*self.workGroupNumCols)
     numBLoadsR = (self.workGroupNumCols*self.microTileNumCols*self.unroll) \
         % (self.workGroupNumRows*self.workGroupNumCols)
     if (numALoads>0 and numALoadsR>0):

src/library/blas/trtri/TrtriKernelSourceIncludes.h

@@ -6,11 +6,11 @@
 
 //**** compiler flags
 //**** online compilation flags
-const char * const TrtriBuildOptions = "-cl-std=CL2.0";
-const char * const TrtribinBuildOptions = "-cl-std=CL2.0";
+const char * const TrtriBuildOptions = "-cl-std=CL" OPENCL_VERSION;
+const char * const TrtribinBuildOptions = "-cl-std=CL" OPENCL_VERSION;
 
 /*mod 192 dtrsm*/
-extern const char * const diag_dtrtri_upper_192_12_src; 
+extern const char * const diag_dtrtri_upper_192_12_src;
 extern unsigned char *diag_dtrtri_upper_192_12_bin;
 extern size_t diag_dtrtri_upper_192_12_binSize;