OSDev.org

I'm writing own GEMM implementation on my AMD platform.
How to determine where to insert prefetch/prefetchnta instructions to make it as fast as AOCL.

000000000040183f <cblas_dgemm.L7>:
40183f: 4d 89 c8 mov %r9,%r8
401842: c5 fc 77 vzeroall
401845: b9 1e 00 00 00 mov $0x1e,%ecx

000000000040184a <cblas_dgemm.L9>:
40184a: c4 c1 7d 28 22 vmovapd (%r10),%ymm4
40184f: c4 c1 7d 28 6a 20 vmovapd 0x20(%r10),%ymm5
401855: c4 42 dd b8 00 vfmadd231pd (%r8),%ymm4,%ymm8
40185a: c4 42 d5 b8 40 20 vfmadd231pd 0x20(%r8),%ymm5,%ymm8
401860: c4 42 dd b8 0c 00 vfmadd231pd (%r8,%rax,1),%ymm4,%ymm9
401866: c4 42 d5 b8 4c 00 20 vfmadd231pd 0x20(%r8,%rax,1),%ymm5,%ymm9
40186d: c4 42 dd b8 14 40 vfmadd231pd (%r8,%rax,2),%ymm4,%ymm10
401873: c4 42 d5 b8 54 40 20 vfmadd231pd 0x20(%r8,%rax,2),%ymm5,%ymm10
40187a: c4 42 dd b8 1c 18 vfmadd231pd (%r8,%rbx,1),%ymm4,%ymm11
401880: c4 42 d5 b8 5c 18 20 vfmadd231pd 0x20(%r8,%rbx,1),%ymm5,%ymm11
401887: 49 83 c2 40 add $0x40,%r10
40188b: 49 83 c0 40 add $0x40,%r8
40188f: 48 83 e9 01 sub $0x1,%rcx
401893: 75 b5 jne 40184a <cblas_dgemm.L9>

AMD and Intel publish optimization guides that (usually) give good advice. There are also independent resources like this that will help.

Measure how fast your code actually runs. Not every optimization will help as much as you expect. Sometimes "worse" code is better.

Every CPU microarchitecture is different. An optimization that makes your code faster on one CPU might be slower on another.

Check the licenses on AOCL. Maybe you can copy some of the existing code instead of writing it from scratch.

Thanks for the advice!

I am working on 2000*1000*240 (M*N*K) data set.
Basically it runs as fast as AOCL, if no SIMD store operation with vmovntpd.
Otherwise it's 15% slower.