Optimizing Matrix Multiply - San Diego ….ppt

Optimizing for the serial processors;Outline;Computational Intensity: Key to algorithm efficiency;Warm up: Matrix-vector multiplication;Warm up: Matrix-vector multiplication; Modeling Matrix-Vector Multiplication;What Simplifying Assumptions;Validating the Model;“Na?ve” Matrix Multiply;Matrix Multiply on RS/6000 ;Note on Matrix Storage;“Na?ve” Matrix Multiply;Matrix Multiply on RS/6000 ;“Na?ve” Matrix Multiply;“Na?ve” Matrix Multiply;Blocked (Tiled) Matrix Multiply;Blocked (Tiled) Matrix Multiply;Blocked (Tiled) Matrix Multiply;Blocked (Tiled) Matrix Multiply;Blocked (Tiled) Matrix Multiply;Blocked (Tiled) Matrix Multiply;Blocked (Tiled) Matrix Multiply;Blocked (Tiled) Matrix Multiply;Using Analysis to Understand Machines;Limits to Optimizing Matrix Multiply;Basic Linear Algebra Subroutines;BLAS speeds on an IBM RS6000/590;Search Over Block Sizes;What the Search Space Looks Like;ATLAS (DGEMM n = 500);Tiling Alone Might Not Be Enough;Optimizing in Practice;Removing False Dependencies;Exploit Multiple Registers;Minimize Pointer Updates;Loop Unrolling;Expose Independent Operations;Copy optimization;Recursive Data Layouts;Strassen’s Matrix Multiply;Strassen (continued);Locality in Other Algorithms;Summary;Reading ;Questions You Should Be Able to Answer;Review from Last Lecture;Membench: What to Expect;Memory Hierarchy on a Sun Ultra-2i;PHiPAC: Portable High Performance ANSI C