Optimize AVX2 SGEMM & STRMM

[wjc404]

Replace KERNEL_16x6 with KERNEL_8x12 to slow down reading on packed matrix A(in L3 cache), as mentioned in issue #2210 . The performance catchs up with MKL2019 after the change.

[wjc404]

1).1-thread SGEMM test with m=n=k=7023, transa=N and transb=T, on i9 9900K at 4.4 GHz, theoretical 140.8 GFLOPS:

old kernel:

new kernel:

2).4-thread SGEMM test with m=n=k=10000, transa=T and transb=N, on the same CPU at 4.2 GHz, theoretical 538 GFLOPS:

old kernel:

new kernel:

3).8-thread test with m=n=k=20000, transa=transb=N on the same CPU at 4.1 GHz, theoretical 1050 GFLOPS:

old kernel:

new kernel:

4).1-thread STRMM test with m=n=6971, side=L, uplo=U, transa=N and diag=N, on the same CPU at 4.4 GHz, theoretical 140.8 GFLOPS:

old kernel:

new kernel:

[wjc404]

1-thread SGEMM test with m=n=k=5999, transa=N and transb=N, on r7 3700x at 3.6 GHz, theoretical 115.2 GFLOPS:

old kernel:

new kernel:

4-thread test with m=n=k=10000, transa=transb=T, on the same CPU at 3.6 GHz, theoretical 461 GFLOPS:

old kernel:

new kernel:

[wjc404]

The STRMM kernel passed 1-thread reliability test.

Test code:
strmm_compare_test.zip

[wjc404]

The SGEMM kernel passed 1-thread reliability test. Ready for merging.

[martin-frbg]

Great, thanks a lot. Still fascinating to see how much performance can be improved by making things slower.