Untitled

 avatar
unknown
batchfile
a year ago
26 kB
12
Indexable
==============================================================================================================
► 每個『組合實例』(t1 fused vs. t2 separate)

[1] Add         │ MatMul_49
  t1 fused    : 0.023425 ms
  t2 separate : 0.025217 ms  (Δ 0.001792 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002304 ms

[2] Add         │ MatMul_51
  t1 fused    : 0.023424 ms
  t2 separate : 0.025217 ms  (Δ 0.001793 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022849 ms
    Add         : 0.002368 ms

[3] Add         │ MatMul_53
  t1 fused    : 0.023360 ms
  t2 separate : 0.025216 ms  (Δ 0.001856 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022976 ms
    Add         : 0.002240 ms

[4] div+add     │ MatMul_88
  t1 fused    : 0.005568 ms
  t2 separate : 0.007009 ms  (Δ 0.001441 ms, speed-up 1.26×)
  ── t2 breakdown ──
    normal      : 0.004545 ms
    Div         : 0.002464 ms

[5] double_add  │ MatMul_105
  t1 fused    : 0.023713 ms
  t2 separate : 0.026656 ms  (Δ 0.002943 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022976 ms
    Add         : 0.002464 ms
    Add         : 0.001216 ms

[6] roberta_fusion│ MatMul_119
  t1 fused    : 0.034913 ms
  t2 separate : 0.041185 ms  (Δ 0.006272 ms, speed-up 1.18×)
  ── t2 breakdown ──
    normal      : 0.023872 ms
    Add         : 0.003905 ms
    Div         : 0.003808 ms
    Erf         : 0.002080 ms
    Add         : 0.002848 ms
    Mul         : 0.001856 ms
    Mul         : 0.002816 ms

[7] double_add  │ MatMul_129
  t1 fused    : 0.083426 ms
  t2 separate : 0.086275 ms  (Δ 0.002849 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082562 ms
    Add         : 0.002497 ms
    Add         : 0.001216 ms

[8] Add         │ MatMul_143
  t1 fused    : 0.023361 ms
  t2 separate : 0.025248 ms  (Δ 0.001887 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002368 ms

[9] Add         │ MatMul_145
  t1 fused    : 0.023328 ms
  t2 separate : 0.025217 ms  (Δ 0.001889 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022912 ms
    Add         : 0.002305 ms

[10] Add         │ MatMul_147
  t1 fused    : 0.023328 ms
  t2 separate : 0.025344 ms  (Δ 0.002016 ms, speed-up 1.09×)
  ── t2 breakdown ──
    normal      : 0.023072 ms
    Add         : 0.002272 ms

[11] div+add     │ MatMul_182
  t1 fused    : 0.005536 ms
  t2 separate : 0.007040 ms  (Δ 0.001504 ms, speed-up 1.27×)
  ── t2 breakdown ──
    normal      : 0.004544 ms
    Div         : 0.002496 ms

[12] double_add  │ MatMul_199
  t1 fused    : 0.023745 ms
  t2 separate : 0.026625 ms  (Δ 0.002880 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002528 ms
    Add         : 0.001184 ms

[13] roberta_fusion│ MatMul_213
  t1 fused    : 0.035009 ms
  t2 separate : 0.041248 ms  (Δ 0.006239 ms, speed-up 1.18×)
  ── t2 breakdown ──
    normal      : 0.023936 ms
    Add         : 0.003936 ms
    Div         : 0.003808 ms
    Erf         : 0.002080 ms
    Add         : 0.002848 ms
    Mul         : 0.001792 ms
    Mul         : 0.002848 ms

[14] double_add  │ MatMul_223
  t1 fused    : 0.083490 ms
  t2 separate : 0.086209 ms  (Δ 0.002719 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082754 ms
    Add         : 0.002240 ms
    Add         : 0.001215 ms

[15] Add         │ MatMul_237
  t1 fused    : 0.023456 ms
  t2 separate : 0.025120 ms  (Δ 0.001664 ms, speed-up 1.07×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002240 ms

[16] Add         │ MatMul_239
  t1 fused    : 0.023360 ms
  t2 separate : 0.025377 ms  (Δ 0.002017 ms, speed-up 1.09×)
  ── t2 breakdown ──
    normal      : 0.023041 ms
    Add         : 0.002336 ms

[17] Add         │ MatMul_241
  t1 fused    : 0.023393 ms
  t2 separate : 0.025248 ms  (Δ 0.001855 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002368 ms

[18] div+add     │ MatMul_276
  t1 fused    : 0.005537 ms
  t2 separate : 0.007104 ms  (Δ 0.001567 ms, speed-up 1.28×)
  ── t2 breakdown ──
    normal      : 0.004608 ms
    Div         : 0.002496 ms

[19] double_add  │ MatMul_293
  t1 fused    : 0.023744 ms
  t2 separate : 0.026496 ms  (Δ 0.002752 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022912 ms
    Add         : 0.002336 ms
    Add         : 0.001248 ms

[20] roberta_fusion│ MatMul_307
  t1 fused    : 0.034977 ms
  t2 separate : 0.041697 ms  (Δ 0.006720 ms, speed-up 1.19×)
  ── t2 breakdown ──
    normal      : 0.024321 ms
    Add         : 0.003872 ms
    Div         : 0.003872 ms
    Erf         : 0.002080 ms
    Add         : 0.002848 ms
    Mul         : 0.001856 ms
    Mul         : 0.002848 ms

[21] double_add  │ MatMul_317
  t1 fused    : 0.083394 ms
  t2 separate : 0.086147 ms  (Δ 0.002753 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082658 ms
    Add         : 0.002272 ms
    Add         : 0.001217 ms

[22] Add         │ MatMul_331
  t1 fused    : 0.023392 ms
  t2 separate : 0.025153 ms  (Δ 0.001761 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022849 ms
    Add         : 0.002304 ms

[23] Add         │ MatMul_333
  t1 fused    : 0.023617 ms
  t2 separate : 0.025409 ms  (Δ 0.001792 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002528 ms

[24] Add         │ MatMul_335
  t1 fused    : 0.023457 ms
  t2 separate : 0.025121 ms  (Δ 0.001664 ms, speed-up 1.07×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002240 ms

[25] div+add     │ MatMul_370
  t1 fused    : 0.005600 ms
  t2 separate : 0.007073 ms  (Δ 0.001473 ms, speed-up 1.26×)
  ── t2 breakdown ──
    normal      : 0.004545 ms
    Div         : 0.002528 ms

[26] double_add  │ MatMul_387
  t1 fused    : 0.023680 ms
  t2 separate : 0.026625 ms  (Δ 0.002945 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022944 ms
    Add         : 0.002465 ms
    Add         : 0.001216 ms

[27] roberta_fusion│ MatMul_401
  t1 fused    : 0.035008 ms
  t2 separate : 0.040993 ms  (Δ 0.005985 ms, speed-up 1.17×)
  ── t2 breakdown ──
    normal      : 0.023969 ms
    Add         : 0.003648 ms
    Div         : 0.003808 ms
    Erf         : 0.002048 ms
    Add         : 0.002848 ms
    Mul         : 0.001856 ms
    Mul         : 0.002816 ms

[28] double_add  │ MatMul_411
  t1 fused    : 0.083489 ms
  t2 separate : 0.086146 ms  (Δ 0.002657 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082626 ms
    Add         : 0.002304 ms
    Add         : 0.001216 ms

[29] Add         │ MatMul_425
  t1 fused    : 0.023361 ms
  t2 separate : 0.025152 ms  (Δ 0.001791 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002272 ms

[30] Add         │ MatMul_427
  t1 fused    : 0.023456 ms
  t2 separate : 0.025216 ms  (Δ 0.001760 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002336 ms

[31] Add         │ MatMul_429
  t1 fused    : 0.023424 ms
  t2 separate : 0.025153 ms  (Δ 0.001729 ms, speed-up 1.07×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002272 ms

[32] div+add     │ MatMul_464
  t1 fused    : 0.005568 ms
  t2 separate : 0.007072 ms  (Δ 0.001504 ms, speed-up 1.27×)
  ── t2 breakdown ──
    normal      : 0.004576 ms
    Div         : 0.002496 ms

[33] double_add  │ MatMul_481
  t1 fused    : 0.023713 ms
  t2 separate : 0.026656 ms  (Δ 0.002943 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.023072 ms
    Add         : 0.002336 ms
    Add         : 0.001248 ms

[34] roberta_fusion│ MatMul_495
  t1 fused    : 0.034977 ms
  t2 separate : 0.041186 ms  (Δ 0.006209 ms, speed-up 1.18×)
  ── t2 breakdown ──
    normal      : 0.023808 ms
    Add         : 0.003936 ms
    Div         : 0.003808 ms
    Erf         : 0.002081 ms
    Add         : 0.002848 ms
    Mul         : 0.001856 ms
    Mul         : 0.002849 ms

[35] double_add  │ MatMul_505
  t1 fused    : 0.083490 ms
  t2 separate : 0.086051 ms  (Δ 0.002561 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082594 ms
    Add         : 0.002240 ms
    Add         : 0.001217 ms

[36] Add         │ MatMul_519
  t1 fused    : 0.023392 ms
  t2 separate : 0.025152 ms  (Δ 0.001760 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002272 ms

[37] Add         │ MatMul_521
  t1 fused    : 0.023360 ms
  t2 separate : 0.025313 ms  (Δ 0.001953 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002432 ms

[38] Add         │ MatMul_523
  t1 fused    : 0.023328 ms
  t2 separate : 0.025217 ms  (Δ 0.001889 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002336 ms

[39] div+add     │ MatMul_558
  t1 fused    : 0.005601 ms
  t2 separate : 0.007040 ms  (Δ 0.001439 ms, speed-up 1.26×)
  ── t2 breakdown ──
    normal      : 0.004512 ms
    Div         : 0.002528 ms

[40] double_add  │ MatMul_575
  t1 fused    : 0.023681 ms
  t2 separate : 0.026496 ms  (Δ 0.002815 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002368 ms
    Add         : 0.001215 ms

[41] roberta_fusion│ MatMul_589
  t1 fused    : 0.034656 ms
  t2 separate : 0.041089 ms  (Δ 0.006433 ms, speed-up 1.19×)
  ── t2 breakdown ──
    normal      : 0.024097 ms
    Add         : 0.003712 ms
    Div         : 0.003776 ms
    Erf         : 0.002080 ms
    Add         : 0.002816 ms
    Mul         : 0.001824 ms
    Mul         : 0.002784 ms

[42] double_add  │ MatMul_599
  t1 fused    : 0.083457 ms
  t2 separate : 0.086211 ms  (Δ 0.002754 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082722 ms
    Add         : 0.002273 ms
    Add         : 0.001216 ms

[43] Add         │ MatMul_613
  t1 fused    : 0.023392 ms
  t2 separate : 0.025185 ms  (Δ 0.001793 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002304 ms

[44] Add         │ MatMul_615
  t1 fused    : 0.023424 ms
  t2 separate : 0.025184 ms  (Δ 0.001760 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002304 ms

[45] Add         │ MatMul_617
  t1 fused    : 0.023424 ms
  t2 separate : 0.025185 ms  (Δ 0.001761 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002304 ms

[46] div+add     │ MatMul_652
  t1 fused    : 0.005600 ms
  t2 separate : 0.006976 ms  (Δ 0.001376 ms, speed-up 1.25×)
  ── t2 breakdown ──
    normal      : 0.004512 ms
    Div         : 0.002464 ms

[47] double_add  │ MatMul_669
  t1 fused    : 0.023713 ms
  t2 separate : 0.026464 ms  (Δ 0.002751 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022912 ms
    Add         : 0.002336 ms
    Add         : 0.001216 ms

[48] roberta_fusion│ MatMul_683
  t1 fused    : 0.034945 ms
  t2 separate : 0.040991 ms  (Δ 0.006046 ms, speed-up 1.17×)
  ── t2 breakdown ──
    normal      : 0.024064 ms
    Add         : 0.003615 ms
    Div         : 0.003776 ms
    Erf         : 0.002048 ms
    Add         : 0.002816 ms
    Mul         : 0.001856 ms
    Mul         : 0.002816 ms

[49] double_add  │ MatMul_693
  t1 fused    : 0.083394 ms
  t2 separate : 0.086274 ms  (Δ 0.002880 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082658 ms
    Add         : 0.002368 ms
    Add         : 0.001248 ms

[50] Add         │ MatMul_707
  t1 fused    : 0.023424 ms
  t2 separate : 0.025280 ms  (Δ 0.001856 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.023008 ms
    Add         : 0.002272 ms

[51] Add         │ MatMul_709
  t1 fused    : 0.023360 ms
  t2 separate : 0.025153 ms  (Δ 0.001793 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022849 ms
    Add         : 0.002304 ms

[52] Add         │ MatMul_711
  t1 fused    : 0.023360 ms
  t2 separate : 0.025377 ms  (Δ 0.002017 ms, speed-up 1.09×)
  ── t2 breakdown ──
    normal      : 0.023105 ms
    Add         : 0.002272 ms

[53] div+add     │ MatMul_746
  t1 fused    : 0.005568 ms
  t2 separate : 0.007072 ms  (Δ 0.001504 ms, speed-up 1.27×)
  ── t2 breakdown ──
    normal      : 0.004576 ms
    Div         : 0.002496 ms

[54] double_add  │ MatMul_763
  t1 fused    : 0.023808 ms
  t2 separate : 0.026464 ms  (Δ 0.002656 ms, speed-up 1.11×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002336 ms
    Add         : 0.001248 ms

[55] roberta_fusion│ MatMul_777
  t1 fused    : 0.034753 ms
  t2 separate : 0.040961 ms  (Δ 0.006208 ms, speed-up 1.18×)
  ── t2 breakdown ──
    normal      : 0.023936 ms
    Add         : 0.003617 ms
    Div         : 0.003808 ms
    Erf         : 0.002048 ms
    Add         : 0.002848 ms
    Mul         : 0.001856 ms
    Mul         : 0.002848 ms

[56] double_add  │ MatMul_787
  t1 fused    : 0.083617 ms
  t2 separate : 0.086339 ms  (Δ 0.002722 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082754 ms
    Add         : 0.002336 ms
    Add         : 0.001249 ms

[57] Add         │ MatMul_801
  t1 fused    : 0.023329 ms
  t2 separate : 0.025345 ms  (Δ 0.002016 ms, speed-up 1.09×)
  ── t2 breakdown ──
    normal      : 0.022977 ms
    Add         : 0.002368 ms

[58] Add         │ MatMul_803
  t1 fused    : 0.023361 ms
  t2 separate : 0.025217 ms  (Δ 0.001856 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002304 ms

[59] Add         │ MatMul_805
  t1 fused    : 0.023393 ms
  t2 separate : 0.025280 ms  (Δ 0.001887 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022912 ms
    Add         : 0.002368 ms

[60] div+add     │ MatMul_840
  t1 fused    : 0.005568 ms
  t2 separate : 0.007040 ms  (Δ 0.001472 ms, speed-up 1.26×)
  ── t2 breakdown ──
    normal      : 0.004544 ms
    Div         : 0.002496 ms

[61] double_add  │ MatMul_857
  t1 fused    : 0.023681 ms
  t2 separate : 0.026465 ms  (Δ 0.002784 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002304 ms
    Add         : 0.001248 ms

[62] roberta_fusion│ MatMul_871
  t1 fused    : 0.034657 ms
  t2 separate : 0.041087 ms  (Δ 0.006430 ms, speed-up 1.19×)
  ── t2 breakdown ──
    normal      : 0.023840 ms
    Add         : 0.003904 ms
    Div         : 0.003776 ms
    Erf         : 0.002080 ms
    Add         : 0.002848 ms
    Mul         : 0.001824 ms
    Mul         : 0.002815 ms

[63] double_add  │ MatMul_881
  t1 fused    : 0.083522 ms
  t2 separate : 0.086082 ms  (Δ 0.002560 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082594 ms
    Add         : 0.002240 ms
    Add         : 0.001248 ms

[64] Add         │ MatMul_895
  t1 fused    : 0.023360 ms
  t2 separate : 0.025249 ms  (Δ 0.001889 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002336 ms

[65] Add         │ MatMul_897
  t1 fused    : 0.023584 ms
  t2 separate : 0.025056 ms  (Δ 0.001472 ms, speed-up 1.06×)
  ── t2 breakdown ──
    normal      : 0.022848 ms
    Add         : 0.002208 ms

[66] Add         │ MatMul_899
  t1 fused    : 0.023424 ms
  t2 separate : 0.025312 ms  (Δ 0.001888 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022944 ms
    Add         : 0.002368 ms

[67] div+add     │ MatMul_934
  t1 fused    : 0.005568 ms
  t2 separate : 0.007072 ms  (Δ 0.001504 ms, speed-up 1.27×)
  ── t2 breakdown ──
    normal      : 0.004544 ms
    Div         : 0.002528 ms

[68] double_add  │ MatMul_951
  t1 fused    : 0.023745 ms
  t2 separate : 0.026528 ms  (Δ 0.002783 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002400 ms
    Add         : 0.001248 ms

[69] roberta_fusion│ MatMul_965
  t1 fused    : 0.035041 ms
  t2 separate : 0.041569 ms  (Δ 0.006528 ms, speed-up 1.19×)
  ── t2 breakdown ──
    normal      : 0.024321 ms
    Add         : 0.003904 ms
    Div         : 0.003776 ms
    Erf         : 0.002048 ms
    Add         : 0.002848 ms
    Mul         : 0.001824 ms
    Mul         : 0.002848 ms

[70] double_add  │ MatMul_975
  t1 fused    : 0.083490 ms
  t2 separate : 0.086082 ms  (Δ 0.002592 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082594 ms
    Add         : 0.002272 ms
    Add         : 0.001216 ms

[71] Add         │ MatMul_989
  t1 fused    : 0.023457 ms
  t2 separate : 0.025440 ms  (Δ 0.001983 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.023040 ms
    Add         : 0.002400 ms

[72] Add         │ MatMul_991
  t1 fused    : 0.023425 ms
  t2 separate : 0.025153 ms  (Δ 0.001728 ms, speed-up 1.07×)
  ── t2 breakdown ──
    normal      : 0.022881 ms
    Add         : 0.002272 ms

[73] Add         │ MatMul_993
  t1 fused    : 0.023393 ms
  t2 separate : 0.025217 ms  (Δ 0.001824 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002304 ms

[74] div+add     │ MatMul_1028
  t1 fused    : 0.005600 ms
  t2 separate : 0.007008 ms  (Δ 0.001408 ms, speed-up 1.25×)
  ── t2 breakdown ──
    normal      : 0.004512 ms
    Div         : 0.002496 ms

[75] double_add  │ MatMul_1045
  t1 fused    : 0.023776 ms
  t2 separate : 0.026528 ms  (Δ 0.002752 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022913 ms
    Add         : 0.002431 ms
    Add         : 0.001184 ms

[76] roberta_fusion│ MatMul_1059
  t1 fused    : 0.034849 ms
  t2 separate : 0.041026 ms  (Δ 0.006177 ms, speed-up 1.18×)
  ── t2 breakdown ──
    normal      : 0.023840 ms
    Add         : 0.003712 ms
    Div         : 0.003840 ms
    Erf         : 0.002048 ms
    Add         : 0.002817 ms
    Mul         : 0.001824 ms
    Mul         : 0.002945 ms

[77] double_add  │ MatMul_1069
  t1 fused    : 0.083426 ms
  t2 separate : 0.086146 ms  (Δ 0.002720 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082626 ms
    Add         : 0.002304 ms
    Add         : 0.001216 ms

[78] Add         │ MatMul_1083
  t1 fused    : 0.023361 ms
  t2 separate : 0.025184 ms  (Δ 0.001823 ms, speed-up 1.08×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002304 ms

[79] Add         │ MatMul_1085
  t1 fused    : 0.023584 ms
  t2 separate : 0.025121 ms  (Δ 0.001537 ms, speed-up 1.07×)
  ── t2 breakdown ──
    normal      : 0.022849 ms
    Add         : 0.002272 ms

[80] Add         │ MatMul_1087
  t1 fused    : 0.023360 ms
  t2 separate : 0.025088 ms  (Δ 0.001728 ms, speed-up 1.07×)
  ── t2 breakdown ──
    normal      : 0.022848 ms
    Add         : 0.002240 ms

[81] div+add     │ MatMul_1122
  t1 fused    : 0.005600 ms
  t2 separate : 0.006976 ms  (Δ 0.001376 ms, speed-up 1.25×)
  ── t2 breakdown ──
    normal      : 0.004480 ms
    Div         : 0.002496 ms

[82] double_add  │ MatMul_1139
  t1 fused    : 0.023808 ms
  t2 separate : 0.026624 ms  (Δ 0.002816 ms, speed-up 1.12×)
  ── t2 breakdown ──
    normal      : 0.022880 ms
    Add         : 0.002528 ms
    Add         : 0.001216 ms

[83] roberta_fusion│ MatMul_1153
  t1 fused    : 0.034880 ms
  t2 separate : 0.041440 ms  (Δ 0.006560 ms, speed-up 1.19×)
  ── t2 breakdown ──
    normal      : 0.024160 ms
    Add         : 0.003936 ms
    Div         : 0.003808 ms
    Erf         : 0.002048 ms
    Add         : 0.002848 ms
    Mul         : 0.001824 ms
    Mul         : 0.002816 ms

[84] double_add  │ MatMul_1163
  t1 fused    : 0.083490 ms
  t2 separate : 0.086241 ms  (Δ 0.002751 ms, speed-up 1.03×)
  ── t2 breakdown ──
    normal      : 0.082721 ms
    Add         : 0.002304 ms
    Add         : 0.001216 ms

==============================================================================================================
► 每個『組合類型』總和
  Add         : fused 0.842637 ms   sep 0.908116 ms   Δ 0.065479 ms   speed-up 1.08×
  div+add     : fused 0.066914 ms   sep 0.084482 ms   Δ 0.017568 ms   speed-up 1.26×
  double_add  : fused 1.286492 ms   sep 1.352830 ms   Δ 0.066338 ms   speed-up 1.05×
  roberta_fusion: fused 0.418665 ms   sep 0.494472 ms   Δ 0.075807 ms   speed-up 1.18×

==============================================================================================================
► 四組合累計 Σ
  Σ fused     : 2.614708 ms
  Σ separate  : 2.839900 ms
  Δ (sep-fuse): 0.225192 ms   speed-up 1.09×

==============================================================================================================
► 其餘 Run_with_type 彙總 (未納入組合)
  Add         : t1     0.091556 ms │ t2     0.133474 ms │ Δ     0.041918 ms
  Cast        : t1     0.001248 ms │ t2     0.001248 ms │ Δ     0.000000 ms
  Concat      : t1     0.726835 ms │ t2     0.727441 ms │ Δ     0.000606 ms
  CumSum      : t1     0.021504 ms │ t2     0.021441 ms │ Δ    -0.000063 ms
  Div         : t1     0.061023 ms │ t2     0.059298 ms │ Δ    -0.001725 ms
  Equal       : t1     0.001376 ms │ t2     0.001376 ms │ Δ     0.000000 ms
  Gather      : t1     1.236570 ms │ t2     1.165435 ms │ Δ    -0.071135 ms
  Gemm        : t1     0.052865 ms │ t2     0.031521 ms │ Δ    -0.021344 ms
  Mul         : t1     0.061124 ms │ t2     0.061184 ms │ Δ     0.000060 ms
  Not         : t1     0.000960 ms │ t2     0.000928 ms │ Δ    -0.000032 ms
  Pow         : t1     0.035649 ms │ t2     0.035548 ms │ Δ    -0.000101 ms
  ReduceMean  : t1     0.139874 ms │ t2     0.136323 ms │ Δ    -0.003551 ms
  Shape       : t1     0.015360 ms │ t2     0.015361 ms │ Δ     0.000001 ms
  Softmax     : t1     0.031298 ms │ t2     0.033025 ms │ Δ     0.001727 ms
  Sqrt        : t1     0.022780 ms │ t2     0.022783 ms │ Δ     0.000003 ms
  Sub         : t1     0.052162 ms │ t2     0.050335 ms │ Δ    -0.001827 ms
  Tanh        : t1     0.000991 ms │ t2     0.000960 ms │ Δ    -0.000031 ms
  Transpose   : t1     0.099648 ms │ t2     0.100450 ms │ Δ     0.000802 ms

==============================================================================================================
► 子-op (含 normal) 跨全部組合累積時間
  normal      : 2.434452 ms
  Add         : 0.248934 ms
  Div         : 0.075648 ms
  Mul         : 0.056097 ms
  Erf         : 0.024769 ms

==============================================================================================================
► 整體檔案總執行時間
  t1 total : 5.340941 ms
  t2 total : 5.511313 ms
  Δ        : 0.170372 ms   speed-up 1.03×

► CPU
t1:
Unsqueeze              99.0        0.316840        0.003200
Reshape                48.0        0.253456        0.005280
ConstantOfShape         2.0        0.031961        0.015981
0.602257

t2:
Unsqueeze              99.0        0.312523        0.003157
Reshape                48.0        0.247475        0.005156
ConstantOfShape         2.0        0.038691        0.019346
0.598689

► 整體檔案總執行時間 + CPU
  t1 total : 5.943198 ms
  t2 total : 6.110002 ms
  Δ        : 0.166804 ms   speed-up 1.03×
Editor is loading...
Leave a Comment