Well, I as many of you, have made my own multmatrix function using the vfpu trough asm instructions, an I have executed the simplest benchmark, multiply the same 4x4 matrices 10.000.000 times, having the next results:
SW multiplication: 61.585938
HW multiplication 1: 54.400757 Speedup1: 1.132054
HW multiplication 2: 50.423547 Speedup2: 1.221372
HW multiplication 3: 39.609375 Speedup3: 1.554846
The differences between the three HW multiplications are the matrices, in first case both are null matrices (all 0's), in second case are two non null matrices, and the result was discarded, in last case multiplications where accumulative, so I think the HW has any means to view if any component is NaN and in such case the multiplication is faster. Relevant case is the second, where we have an 1.22 speedup.
Talking about code size:
ASM function is coded in nearly 20 OP, and SW function in 26-30, so we have an improvement of ~1.2 in space and in time.
Timing includes the overhead of the for loop calling the multiplication, but 2 integer operations and a jump are far leser than de fp op's