Bridging the Semantic Gaps of GPU Acceleration for Scaleout CNN-based Big Data Processing: Think Big, See Small