[PATCH] D33099: [X86] Model 256-bit AVX instructions in the AMD Jaguar scheduler (PR28573)

[Andrew V. Tischenko via Phabricator via llvm-commits]

avt77 added a comment.

Simon, finaly I'm able to create ClothAvx test executable with clang. And I created it with this patch and without it. And I got the following results on AMD laptop (CPU AMD A10-8700P Radeon R6, 10 Compute Cores 4C+6G 1.80 GHz):

C:\Users\andre\Downloads\working\ClothExe>type avxcloth-patch.log
0.00 57.67 60.91 60.28 27.26 62.62 62.56 67.97
SIMD Width = 1
28.43 23.37 22.71 22.93 22.65 23.00 23.07 23.14 22.71 22.89 22.94 22.81 22.79 23.32 23.14
SIMD Width = 4
36.39 57.45 57.61 56.97 57.05 57.82 57.12 57.38 57.08 57.57 57.28 57.88 57.43 56.96 57.07 57.33
SIMD Width = 8
68.71 71.78 71.31 71.78 71.79 71.67 71.97 71.79 71.25 72.55 71.96 71.52 72.04 70.67 71.78 70.39 70.94
C:\Users\andre\Downloads\working\ClothExe>type avxcloth-trunk.log
0.00 55.19 59.88 58.43 19.96 60.22 58.58 57.37 59.34 60.11
SIMD Width = 1
24.51 21.64 21.29 21.42 21.37 21.38 21.43 21.19 22.37 23.09 24.14 23.77 23.23 23.12 22.43 22.30
SIMD Width = 4
35.74 58.77 56.04 55.87 56.56 55.44 55.24 55.26 55.44 54.91 56.47 57.75 56.36 56.72 55.56 56.11 56.59
SIMD Width = 8
65.76 70.74 70.04 70.26 70.95 72.26 73.50 70.77 69.86 69.93 70.76 70.25 70.41 71.99

As you see the patched version is slightly faster than the trunk one. Are you sure you saw any degradation with this patch?
BTW, the number are Flps per second (they are being calculated every one second).


https://reviews.llvm.org/D33099