Get disassembled code to verify SIMD optimizations

[shubhamvishu]

This uses perf tool to generate the disassembled code for the run. It generates a perf<X>.data file (where X=1,2,3...) for each corresponding benchmark run/configuration. One could then simply look into the disassembled code like below :

Currently it records the executed instructions from the user space. We can add other events also if we want, but this could be a good start and keep a decent file size

sudo perf record -e instructions:u -g java <...>

To look into the disassembled code (disassembling is super slow) :

• Directly open the view or file

sudo perf annotate -i perf.data -f

• Print the file content

PAGER=cat perf annotate -i perf1.data -f --stdio 

[shubhamvishu]

I used this to verify the SDOT instruction was getting used correctly in the 7-bit case for #13572.

// 4-bit; No matches

>> PAGER=cat perf annotate -i perf1.data -f --stdio | head -100000 | grep "sdot"         

// 7 bit; found matches  

>> PAGER=cat perf annotate -i perf2.data -f --stdio | head -100000 | grep "sdot"        
    0.00 :   79c:    sdot    v0.4s, v2.16b, v4.16b
    0.00 :   7a8:    sdot    v7.4s, v6.16b, v5.16b
    0.00 :   7b0:    sdot    v22.4s, v3.16b, v16.16b
    0.00 :   7bc:    sdot    v21.4s, v18.16b, v17.16b
    0.28 :   7e4:    sdot    v0.4s, v19.16b, v20.16b
    0.55 :   7ec:    sdot    v7.4s, v23.16b, v30.16b
    0.44 :   7f4:    sdot    v22.4s, v24.16b, v31.16b
    1.81 :   7fc:    sdot    v21.4s, v25.16b, v1.16b
    0.34 :   804:    sdot    v0.4s, v26.16b, v2.16b
    0.58 :   80c:    sdot    v7.4s, v27.16b, v4.16b
    0.61 :   814:    sdot    v22.4s, v28.16b, v5.16b
    1.80 :   81c:    sdot    v21.4s, v29.16b, v3.16b
    0.00 :   8b4:    sdot    v0.4s, v22.16b, v16.16b
    0.00 :   8c0:    sdot    v0.4s, v17.16b, v18.16b
    0.00 :   8cc:    sdot    v0.4s, v19.16b, v23.16b
    0.00 :   8d8:    sdot    v0.4s, v24.16b, v25.16b
    0.00 :   8e4:    sdot    v0.4s, v26.16b, v27.16b
    0.00 :   8f0:    sdot    v0.4s, v28.16b, v29.16b
    0.00 :   914:    sdot    v0.4s, v20.16b, v30.16b
    0.00 :   928:    sdot    v0.4s, v31.16b, v6.16b
    0.00 :   94c:    sdot    v0.4s, v16.16b, v22.16b
    0.00 :   950:    sdot    v0.4s, v21.16b, v17.16b
    0.00 :   954:    sdot    v0.4s, v5.16b, v18.16b
    0.00 :   958:    sdot    v0.4s, v4.16b, v19.16b
    0.00 :   95c:    sdot    v0.4s, v3.16b, v23.16b
    0.00 :   960:    sdot    v0.4s, v2.16b, v24.16b
    0.00 :   a04:    sdot    v25.4s, v28.16b, v29.16b
    0.00 :   a14:    sdot    v25.4s, v20.16b, v30.16b
    0.00 :   a24:    sdot    v25.4s, v31.16b, v21.16b
    0.00 :   a34:    sdot    v25.4s, v5.16b, v4.16b
    0.00 :   a44:    sdot    v25.4s, v3.16b, v2.16b
    0.00 :   a54:    sdot    v25.4s, v6.16b, v22.16b
    0.00 :   a88:    sdot    v25.4s, v16.16b, v17.16b
    0.00 :   aa0:    sdot    v25.4s, v23.16b, v24.16b
    0.00 :   aa4:    sdot    v25.4s, v18.16b, v19.16b
    0.00 :   ab0:    sdot    v25.4s, v0.16b, v7.16b
    0.00 :   abc:    sdot    v25.4s, v1.16b, v26.16b
    0.00 :   ac8:    sdot    v25.4s, v27.16b, v28.16b
    0.00 :   ad4:    sdot    v25.4s, v29.16b, v20.16b
    0.00 :   ae0:    sdot    v25.4s, v30.16b, v31.16b
    0.00 :   bf8:    sdot    v0.4s, v21.16b, v6.16b
    0.00 :   c0c:    sdot    v25.4s, v27.16b, v26.16b

// No quantization

>> PAGER=cat perf annotate -i perf3.data -f --stdio | head -100000 | grep "sdot"

[github-actions]

This PR has not had activity in the past 2 weeks, labeling it as stale. If the PR is waiting for review, notify the [email protected] list. Thank you for your contribution!