Fix invalid memory access on Nvidia GPUs in CSR-Adaptive SpMV kernel (#…

…192)

* corrected condition to prevent buffer overflow (fixes bug on Nvidia K20m)

* removed unneeded line

src/library/kernels/csrmv_adaptive.cl

@@ -422,8 +422,7 @@ csrmv_adaptive(__global const VALUE_TYPE * restrict const vals,
           // However, this may change in the future (e.g. with shared virtual memory.)
           // This causes a minor performance loss because this is the last workgroup
           // to be launched, and this loop can't be unrolled.
-          const unsigned int max_to_load = rowPtrs[stop_row] - rowPtrs[row];
-          for(int i = 0; i < ((int)max_to_load-(int)lid); i += WG_SIZE)
+          for(int i = 0; col+i < rowPtrs[stop_row]; i += WG_SIZE)
               partialSums[lid + i] = alpha * vals[col + i] * vec[cols[col + i]];
       }
       barrier(CLK_LOCAL_MEM_FENCE);