ITXGEMM performance

• John A. Gunnels, Greg M. Henry, and Robert A. van de Geijn, " A Family of High-Performance Matrix Algorithms, " in Computational Science - 2001, Part I, Lecture Notes in Computer Science 2073, pp. 51-60, Springer, 2001

• John Gunnels and Robert van de Geijn, " Developing Linear Algebra Algorithms: A Collection of Class Projects. " FLAME Working Note #3, The University of Texas at Austin, Department of Computer Sciences. Technical Report TR-2001-??. May 2001.

Platform (Robert's laptop):

• Intel Pentium (TR) III processor (650 MHz) with 256K L2 cache and 16K L1 (data) cache.
• Linux Redhat 6.2
• double precision (64-bit) arithmetic
• ATLAS R3.1 for PIII with a 256K Cache.
  This particular version was specifically pre-compiled for us by the Programmer-Formerly-Known-As-Richard.

Performance graphs that demonstrates the benefits of a family of algorithms

• Best ITXGEMM: For each set of matrix dimensions we report the best performance (in MFLOPS) attained over all members of the matrix-matrix multiplication algorithm family.

• Best ITXGEMM vs. MPM-MMP-MPM: This graphs shows the performance penalty incurred (in MFLOPS) if the matrix multiply always uses the MPM-MMP-MPM member of the family, rather than the one that attains the best performance.

• Best ITXGEMM vs. MMP-MPM-MMP: This graphs shows the performance penalty incurred (in MFLOPS) if the matrix multiply always uses the MMP-MPM-MMP member of the family, rather than the one that attains the best performance.

• Best ITXGEMM vs. MPM-MPP-MPM: This graphs shows the performance penalty incurred (in MFLOPS) if the matrix multiply always uses the MPM-MPP-MPM member of the family, rather than the one that attains the best performance.

• Best ITXGEMM vs. MPM-MPP-MMP: This graphs shows the performance penalty incurred (in MFLOPS) if the matrix multiply always uses the MPM-MPP-MPM member of the family, rather than the one that attains the best performance.

Comparison with ATLAS

• Best ITXGEMM vs. ATLAS For each set of matrix dimensions we report the best performance (in MFLOPS) attained over all members of the matrix-matrix multiplication algorithm family. We also show the performance of ATLAS.

• Best ITXGEMM > ATLAS For each set of matrix dimensions we report the best performance (in MFLOPS) attained over all members of the matrix-matrix multiplication algorithm family. We show all matrix sizes where ITXGEMM beats the performance of ATLAS.

• Best ITXGEMM < ATLAS For each set of matrix dimensions we report the best performance (in MFLOPS) attained over all members of the matrix-matrix multiplication algorithm family. We show all matrix sizes where ATLAS beats the performance of ITXGEMM.