Changeset 3946 for trunk


Ignore:
Timestamp:
Aug 1, 2014, 5:33:13 PM (5 years ago)
Author:
linmengl
Message:

add perf results. 4 inserts version saves 100+ lines of asm, but hurt the whole performance

Location:
trunk/lib_ir
Files:
2 edited

Legend:

Unmodified
Added
Removed
  • trunk/lib_ir/s2p.h

    r3945 r3946  
    4343//s2p_bytepack_inline, written in pure IR in order to use immediate constants in shifting.
    4444//s2p_ideal: use 3 stage packh2,4,8 (only available on Haswell avx2)
    45 #define S2P_ALGORITHM s2p_bytepack_inline
     45#define S2P_ALGORITHM s2p_ideal
    4646
    4747#define s2p_ideal(s0, s1, s2, s3, s4, s5, s6, s7, p0, p1, p2, p3, p4, p5, p6, p7) \
  • trunk/lib_ir/xmlwf/perf.txt

    r3945 r3946  
    5151use PEXT for packh_2/4/8
    5252xmlwf_perf   &  4.349   &   4.81   &    4.979   &   5.342   &   5.661 \\ \hline
     53
     54===========================================
     55use PEXT, no shl/or, use truncate and build_vector of v4i32
     56xmlwf_perf   &  4.804   &   5.262   &   5.431   &   5.779   &   6.112 \\ \hline
Note: See TracChangeset for help on using the changeset viewer.