Skip to content
Snippets Groups Projects
Commit cc1974f0 authored by Hashim Sharif's avatar Hashim Sharif
Browse files

Merging

parents 88f0ea83 0b66ed64
No related branches found
No related tags found
No related merge requests found
Expand all Hide whitespace changes
Inline Side-by-side
Showing
with 2160 additions and 0 deletions
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_160.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,344.736,1.29476e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.74231,36242
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.35918,36299
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,835.655,3.42441e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.09224,42331
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.70002,42246
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,24.1042,124515
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,305.48,1.39635e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,2.33278,45027
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,1.73698,45038
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,448.089,2.15419e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.05134,47781
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.79851,47789
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,17.4017,110314
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,201.163,1.05746e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,3.5874,49107
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.32724,49126
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,298.623,1.58674e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.25227,50638
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.11102,50752
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.3346,91373
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.4071,60007
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.79112,50828
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,10.504,60122
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.05558,50600
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.80988,50603
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,352.716,1.87049e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,70.9986,377044
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,29.693,162274
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,502.505,2.64698e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,317.991,1.60963e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,319.62,1.6465e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,48.7611,272981
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,31.3448,186827
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,12.0792,81356
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,49.0611,267293
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_161.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,378.483,1.44377e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,4.16813,37354
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.25211,37202
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,905.01,3.78967e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.41608,43519
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.68306,43573
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,19.4582,105246
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,331.99,1.57104e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,2.31531,46344
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,2.26824,46361
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,485.924,2.40972e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,1.85947,49214
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.37848,49150
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,20.9687,122884
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,197.661,1.06623e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.63787,50453
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.2229,50453
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,296.445,1.62111e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.2027,52149
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.28997,52146
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.8256,104486
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.95045,78259
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,2.4269,52341
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,10.6427,68164
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,3.70788,52034
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.69813,51958
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,346.748,1.8922e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,68.4736,377821
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,28.9813,156405
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,536.616,2.89048e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,344.142,1.82033e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,344.951,1.83101e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,50.2737,285095
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,33.3015,206213
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,12.5197,92863
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,46.8158,273128
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_162.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,376.051,1.43644e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.85018,37259
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.19583,37275
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,903.953,3.79981e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.36888,43749
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,2.17659,43743
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,19.4205,113593
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,333.107,1.57337e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,3.22801,46684
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,2.21371,46628
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,481.782,2.40292e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,1.91195,49343
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.08034,49362
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,19.4,109107
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,197.676,1.04809e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.03656,50768
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.57624,50806
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,298.469,1.64777e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,3.01322,52469
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.26031,52393
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.8955,105111
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.20323,58051
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,3.18967,56831
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,10.956,72962
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.10368,52316
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,6.02162,52242
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,353.033,1.93595e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,67.0839,369030
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,28.8589,156817
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,529.37,2.85811e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,339.888,1.78149e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,338.989,1.80672e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,47.7108,271072
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,31.8336,202963
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,13.1019,88026
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,46.4985,278681
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_163.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,376.623,1.43041e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,4.2712,37335
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.29774,37259
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,906.73,3.79983e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,1.93634,43709
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.93493,43740
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,20.1087,113270
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,340.112,1.62416e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,3.40813,46586
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,2.31909,46626
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,482.654,2.42242e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.32494,49383
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.15611,49380
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,20.2263,123328
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,204.243,1.10377e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.71524,50639
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.19861,50639
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,292.261,1.60557e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.17489,52530
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,1.91954,52530
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.6709,105099
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.89142,72967
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,2.52353,52569
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,9.81087,62950
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.02868,52377
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,4.7711,52300
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,345.339,1.88692e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,65.0372,368276
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,28.1288,151188
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,514.73,2.7732e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,329.902,1.73695e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,330.174,1.75558e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,47.4751,270664
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,30.4123,187537
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,10.2215,83300
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,46.3845,273875
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_164.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,379.306,1.42757e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.64497,37278
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.88347,37297
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,900.541,3.74501e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.1469,43521
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.54092,43559
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,18.5551,100862
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,329.675,1.53371e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,1.86776,46220
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,1.98069,46501
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,479.022,2.36315e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.10373,49269
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,1.92033,49278
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,18.7268,119116
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,204.671,1.0882e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.35297,50563
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,1.81918,50620
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,293.578,1.61435e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.22158,52268
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.11317,52301
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.4222,88604
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.56134,68623
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,3.0971,52511
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,8.69792,57728
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.90547,52226
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.07817,52155
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,353.537,1.93273e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,68.0307,372313
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,28.517,156425
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,547.356,2.93903e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,352.325,1.86501e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,352.467,1.90256e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,51.5187,301084
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,30.5184,185959
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,10.2667,66566
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,43.6303,263236
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_165.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,353.773,1.32322e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.4627,36510
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,1.8332,36567
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,878.627,3.57603e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.05928,42794
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.75336,42851
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,17.1158,93353
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,316.068,1.43626e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,2.90983,45593
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,2.22248,45720
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,475.687,2.29564e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.32846,48366
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,1.68287,48382
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,19.1992,110644
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,199.562,1.04256e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.51739,49706
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,3.51719,53685
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,295.002,1.59147e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.27073,51350
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.21199,51350
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,12.6138,102966
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.33603,60989
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.8852,51521
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,10.8473,56777
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.30992,51375
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.90882,51280
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,344.155,1.85001e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,62.0464,346209
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,26.5284,148623
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,485.063,2.57151e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,317.001,1.59544e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,313.651,1.6023e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,45.7252,259752
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,27.9466,162611
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,9.16157,70558
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,40.4661,232799
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_166.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,352.972,1.3171e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,4.25133,36683
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.45931,36702
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,885.443,3.58108e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.38741,42659
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,2.3844,42771
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,19.8063,110988
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,328.712,1.51501e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,2.41333,45436
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,2.27317,45524
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,477.967,2.32662e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.2292,48092
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,1.84405,48203
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,19.5186,111310
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,194.861,1.01432e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,3.30804,53532
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.18683,49610
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,296.149,1.59912e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.56193,51248
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.39044,51305
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.6294,97584
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.35232,61017
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.95976,51534
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,10.3614,66179
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.53209,51195
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,6.12908,51175
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,357.974,1.90505e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,70.0006,380612
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,29.5305,157813
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,544.245,2.8828e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,349.204,1.81432e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,346.818,1.81672e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,51.9549,306831
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,32.7948,203765
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,13.1344,91337
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,46.9918,268839
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_167.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,350.949,1.31074e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,4.10317,36453
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.50366,36453
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,878.849,3.60669e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,1.85388,42579
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,2.69125,42598
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,19.9735,106383
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,325.095,1.50842e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,2.37345,45411
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,2.45422,45408
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,472.051,2.28945e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.13227,48127
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.41281,48104
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,19.9196,120940
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,194.514,1.01048e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.62532,49456
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.35806,49396
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,296.671,1.5972e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.58171,51020
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.92363,51134
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.0122,102384
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,10.6179,76199
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,2.00952,51154
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,10.6804,66734
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.01533,50910
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,6.19199,50682
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,349.427,1.86051e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,70.2114,389789
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,29.1434,157984
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,550.625,2.90433e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,354.709,1.85793e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,352.117,1.8427e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,48.5219,270571
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,29.9688,172978
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,10.4235,71395
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,46.5118,258348
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_perf_fp16_168.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,350.029,1.32438e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,2.79413,36356
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.08987,36432
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,876.486,3.58641e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.02325,42661
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.5181,42756
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,18.5093,101688
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,316.926,1.45832e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,2.64824,45388
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,1.77112,45462
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,470.364,2.27227e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,1.93378,48096
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,1.78824,48147
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,21.2716,134999
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,190.053,999012
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.09608,49460
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.13989,49533
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,295.297,1.59056e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,3.03483,51195
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.25576,51135
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.7841,97299
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,10.2158,76672
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,2.06478,51401
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,10.7958,72279
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.4984,56242
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,6.26824,51008
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,344.484,1.85136e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,67.526,370185
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,28.0443,147920
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,525.1,2.77217e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,338.872,1.77491e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,338.019,1.80635e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,47.0941,254613
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,27.9775,173259
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,9.21542,60441
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,41.7087,227963
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_red_samp_fp16_42.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,373.814,1.50342e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.2666,39127
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,1.80386,39069
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,1320.04,5.89533e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.02645,47995
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,2.67969,48105
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,19.2226,119613
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,414.702,2.13905e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,2.43845,51060
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,1.92418,50984
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,679.246,3.67719e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.11397,54118
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.50782,54156
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,20.3157,136255
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,249.499,1.46504e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.95217,55715
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.73291,55734
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,403.468,2.41794e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.74872,57540
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.27275,57578
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.5194,115175
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.8221,86147
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,2.35854,57635
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,11.5996,103494
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.49869,57213
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.70418,57021
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,351.842,2.1004e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,54.4989,328936
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,28.7613,164526
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,794.891,4.65583e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,408.553,2.31782e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,409.279,2.35626e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,50.8358,322330
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,26.4958,168790
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,14.8298,101643
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,51.1146,335945
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_red_samp_fp16_44.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,373.04,1.49763e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,2.97739,38877
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,1.75301,38877
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,1318.17,5.84816e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,1.5918,47981
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.52329,48175
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,17.0323,106123
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,405.754,2.06704e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,1.85906,50977
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,1.29455,51003
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,678.918,3.62642e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,1.84783,54143
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.59723,54251
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,21.4198,151819
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,253.501,1.49005e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.88187,55678
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.6939,55620
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,412.319,2.46737e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.74961,57593
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,1.98414,57593
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.5037,109358
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.03222,68299
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.92216,57574
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,11.4418,97628
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,3.94646,57346
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.55647,57120
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,359.484,2.12355e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,54.3003,329059
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,29.5784,175790
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,796.238,4.63702e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,407.145,2.32402e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,409.181,2.36025e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,50.0483,326751
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,30.6403,207453
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,14.9928,106593
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,52.6524,329883
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_red_samp_fp16_46.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,372.942,1.50283e+06
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.42906,39184
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,1.94965,39241
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,1324.67,5.88102e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,2.11742,47981
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.79925,48021
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,18.3743,119504
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,410.906,2.10147e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,1.97375,51067
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,2.66037,51086
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,676.427,3.65834e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.06719,54117
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.34689,54269
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,20.9467,147949
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,251.52,1.48384e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.26901,55681
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.01198,55678
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,408.643,2.46041e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.6836,57634
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.81573,57558
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.2832,109364
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.36828,68324
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.89762,57691
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,11.4947,103479
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.2993,57538
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,6.14501,57462
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,353.979,2.13926e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,56.9928,346984
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,29.1811,175386
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,797.413,4.68361e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,407.74,2.34011e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,408.232,2.37399e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,49.4458,318489
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,28.7016,197598
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,12.5925,96118
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,48.8592,302847
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_samp_fp16_261.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,199.096,768777
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.87226,36317
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.31064,36355
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,862.768,3.5928e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,1.47999,42966
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,0.916697,42889
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,14.7617,85988
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,243.939,1.10794e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,1.35692,45410
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,0.885912,45430
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,432.939,2.0384e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,1.30457,48006
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,0.967639,48083
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,12.2797,91598
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,157.17,786742
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,1.45318,49513
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,1.8852,49551
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,297.318,1.59494e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.36708,51385
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.21256,51366
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.5806,97816
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,8.09633,57082
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,0.937241,51595
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,7.82347,51595
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,2.95617,51424
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.27187,51290
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,402.601,2.16753e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,100.073,540576
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,38.2652,213693
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,940.594,5.07676e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,529.699,2.81573e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,538.929,2.93758e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,68.0023,405177
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,40.2808,250158
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,11.5333,69055
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,53.7619,322533
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_samp_fp16_262.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,201.491,782079
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,3.64602,36755
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.25925,36812
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,863.481,3.60566e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,1.323,43292
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.23471,43349
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,15.2389,86831
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,244.387,1.08745e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,1.66831,45676
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,0.88134,45676
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,434.161,2.03201e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,1.44482,48634
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,1.05884,48634
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,11.3014,82389
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,155.029,784704
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,1.46854,49989
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,1.37976,50063
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,294.808,1.5918e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,2.66305,51777
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.31285,51929
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.7435,93346
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,8.66416,62669
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.54901,52005
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,7.01051,52005
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,2.76116,46791
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.01021,51763
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,401.791,2.17637e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,101.101,554195
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,38.6083,214821
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,980.097,5.33162e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,567.423,3.03884e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,582.842,3.20481e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,70.175,419035
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,40.4284,263401
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,13.3167,91066
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,53.7532,315081
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_samp_fp16_263.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,239.067,970071
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,4.33674,38093
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.34824,38131
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,1255.39,5.6059e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,1.37852,47241
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,0.771448,47298
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,14.7832,94710
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,341.94,1.68639e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,1.16271,50082
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,0.818232,50099
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,635.421,3.26296e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.23858,53688
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,1.90917,53652
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,17.0536,113283
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,233.522,1.32983e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,2.39134,55217
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.274,55138
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,411.418,2.44319e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,1.74261,57437
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.19678,57456
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.5902,109239
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.51897,73973
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.93151,57415
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,8.79818,63113
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,4.12541,57053
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,5.56009,56904
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,370.749,2.16788e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,73.0082,435852
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,29.2538,196008
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,719.422,4.17407e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,477.161,2.70832e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,482.106,2.80035e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,53.2209,328026
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,31.3862,203294
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,9.51689,73278
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,42.6197,273975
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
llvm/projects/hpvm-tensor-rt/PPoPP_results/profiling_results/alexnet_canny_fp16/alexnet_canny_fp16_samp_fp16_264.txt 0 → 100644
+ 108
− 0
View file @ cc1974f0
Conv1,239.767,971021
Conv1_f2h,0,0
Conv1_h2f,0,0
Add1,4.04048,38189
Add1_f2h,0,0
Add1_h2f,0,0
Tanh1,2.52667,38189
Tanh1_f2h,0,0
Tanh1_h2f,0,0
Conv2,1257.19,5.61616e+06
Conv2_f2h,0,0
Conv2_h2f,0,0
Add2,1.40312,47337
Add2_f2h,0,0
Add2_h2f,0,0
Tanh2,1.07093,47356
Tanh2_f2h,0,0
Tanh2_h2f,0,0
Pool1,15.0031,94899
Pool1_f2h,0,0
Pool1_h2f,0,0
Conv3,342.292,1.69339e+06
Conv3_f2h,0,0
Conv3_h2f,0,0
Add3,1.14918,50061
Add3_f2h,0,0
Add3_h2f,0,0
Tanh3,1.01116,50119
Tanh3_f2h,0,0
Tanh3_h2f,0,0
Conv4,635.865,3.30913e+06
Conv4_f2h,0,0
Conv4_h2f,0,0
Add4,2.45573,53609
Add4_f2h,0,0
Add4_h2f,0,0
Tanh4,2.24347,53686
Tanh4_f2h,0,0
Tanh4_h2f,0,0
Pool2,18.5907,135377
Pool2_f2h,0,0
Pool2_h2f,0,0
Conv5,233.494,1.36403e+06
Conv5_f2h,0,0
Conv5_h2f,0,0
Add5,3.06065,55385
Add5_f2h,0,0
Add5_h2f,0,0
Tanh5,2.60712,55398
Tanh5_f2h,0,0
Tanh5_h2f,0,0
Conv6,412.356,2.47544e+06
Conv6_f2h,0,0
Conv6_h2f,0,0
Add6,1.91035,57419
Add6_f2h,0,0
Add6_h2f,0,0
Tanh6,2.64814,57473
Tanh6_f2h,0,0
Tanh6_h2f,0,0
Pool3,11.7297,115070
Pool3_f2h,0,0
Pool3_h2f,0,0
Mul1,9.47385,68360
Mul1_f2h,0,0
Mul1_h2f,0,0
Add7,1.72888,57565
Add7_f2h,0,0
Add7_h2f,0,0
Softmax1,8.59251,57568
Softmax1_f2h,0,0
Softmax1_h2f,0,0
ArgMax1,2.73992,57114
ArgMax1_f2h,0,0
ArgMax1_h2f,0,0
Select1,4.57165,57041
Select1_f2h,0,0
Select1_h2f,0,0
Contract1,352.018,2.10195e+06
Contract1_f2h,0,0
Contract1_h2f,0,0
tensorReduce1,69.4555,427759
tensorReduce1_f2h,0,0
tensorReduce1_h2f,0,0
tensorMap11,28.3901,175566
tensorMap11_f2h,0,0
tensorMap11_h2f,0,0
Conv7,762.599,4.49019e+06
Conv7_f2h,0,0
Conv7_h2f,0,0
Conv8,458.711,2.59714e+06
Conv8_f2h,0,0
Conv8_h2f,0,0
Conv9,464.862,2.68914e+06
Conv9_f2h,0,0
Conv9_h2f,0,0
tensorMap21,51.7606,322724
tensorMap21_f2h,0,0
tensorMap21_h2f,0,0
tensorReduce2,29.6339,202986
tensorReduce2_f2h,0,0
tensorReduce2_h2f,0,0
tensorReduce3,11.0134,84635
tensorReduce3_f2h,0,0
tensorReduce3_h2f,0,0
tensorMap22,41.902,274721
tensorMap22_f2h,0,0
tensorMap22_h2f,0,0
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment