[tenancy before] (none)
[l2 scrub] mode='read' -- see gpu_measure for the bias it carries
[card] NVIDIA GeForce RTX 4090 128 SM, L2 75 MB

########## add+norm -> qkv|z  G=2  x48  [5120x10240xK4 5120x6144xK4] ##########
    site tile s98 TILE_M=16 rows_pad=16 groups=40 rot=4 trellis=41.9 MB/pass

  add+norm -> qkv|z M=8 tile s98  reps=7 burst>=30 ms  cell 5.1 s
    legA pins 10240:s2(TM16) 6144:s2(TM16)
    i8F        62.73 us/site  inner=120   burst=  30.1 ms
    i8S        64.41 us/site  inner=117   burst=  30.1 ms
    i8U        63.23 us/site  inner=119   burst=  30.1 ms
    legA       57.97 us/site  inner=129   burst=  29.9 ms
    prepF       5.37 us/site  inner=1382  burst=  29.7 ms
    prepS       7.02 us/site  inner=1066  burst=  29.9 ms
      i8S/legA 1.1109x [1.1105-1.1113]  leg A wins
      i8U/legA 1.0906x [1.0905-1.0912]  leg A wins
      i8F/legA 1.0823x [1.0818-1.0824]  leg A wins
      glue fusion i8F/i8U 0.9922x [0.9918-0.9925]; multi-out i8U/i8S 0.9819x
      prepF   5.37 us  prepS   7.02 us  (prep share of i8F 8.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.086e-02 1.074e-02
    site tile s94 TILE_M=32 rows_pad=32 groups=40 rot=4 trellis=41.9 MB/pass

  add+norm -> qkv|z M=8 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s2(TM16) 6144:s2(TM16)
    i8F        69.72 us/site  inner=108   burst=  30.1 ms
    i8S        71.38 us/site  inner=105   burst=  30.0 ms
    i8U        70.22 us/site  inner=107   burst=  30.1 ms
    legA       57.95 us/site  inner=130   burst=  30.1 ms
    prepF       5.37 us/site  inner=1374  burst=  29.5 ms
    prepS       7.02 us/site  inner=1061  burst=  29.8 ms
      i8S/legA 1.2317x [1.2312-1.2322]  leg A wins
      i8U/legA 1.2114x [1.2102-1.2125]  leg A wins
      i8F/legA 1.2033x [1.2017-1.2038]  leg A wins
      glue fusion i8F/i8U 0.9930x [0.9924-0.9939]; multi-out i8U/i8S 0.9837x
      prepF   5.37 us  prepS   7.02 us  (prep share of i8F 7.7%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.086e-02 1.074e-02
    site tile s85 TILE_M=64 rows_pad=64 groups=40 rot=4 trellis=41.9 MB/pass

  add+norm -> qkv|z M=8 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s2(TM16) 6144:s2(TM16)
    i8F        89.16 us/site  inner=85    burst=  30.3 ms
    i8S        90.42 us/site  inner=83    burst=  30.0 ms
    i8U        89.30 us/site  inner=84    burst=  30.0 ms
    legA       57.97 us/site  inner=130   burst=  30.1 ms
    prepF       5.75 us/site  inner=1289  burst=  29.6 ms
    prepS       7.02 us/site  inner=1061  burst=  29.8 ms
      i8S/legA 1.5599x [1.5589-1.5601]  leg A wins
      i8U/legA 1.5405x [1.5398-1.5411]  leg A wins
      i8F/legA 1.5381x [1.5372-1.5382]  leg A wins
      glue fusion i8F/i8U 0.9983x [0.9981-0.9986]; multi-out i8U/i8S 0.9876x
      prepF   5.75 us  prepS   7.02 us  (prep share of i8F 6.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.086e-02 1.074e-02

  add+norm -> qkv|z M=16 tile s98  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 10240:s2(TM16) 6144:s2(TM16)
    i8F        63.13 us/site  inner=119   burst=  30.1 ms
    i8S        64.84 us/site  inner=116   burst=  30.1 ms
    i8U        63.67 us/site  inner=118   burst=  30.1 ms
    legA       59.41 us/site  inner=127   burst=  30.2 ms
    prepF       5.38 us/site  inner=1372  burst=  29.5 ms
    prepS       7.02 us/site  inner=1056  burst=  29.7 ms
      i8S/legA 1.0912x [1.0908-1.0921]  leg A wins
      i8U/legA 1.0716x [1.0712-1.0720]  leg A wins
      i8F/legA 1.0626x [1.0623-1.0646]  leg A wins
      glue fusion i8F/i8U 0.9917x [0.9913-0.9931]; multi-out i8U/i8S 0.9820x
      prepF   5.38 us  prepS   7.02 us  (prep share of i8F 8.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.126e-02 1.093e-02

  add+norm -> qkv|z M=16 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s2(TM16) 6144:s2(TM16)
    i8F        70.21 us/site  inner=107   burst=  30.0 ms
    i8S        71.91 us/site  inner=105   burst=  30.2 ms
    i8U        70.77 us/site  inner=106   burst=  30.0 ms
    legA       59.40 us/site  inner=127   burst=  30.2 ms
    prepF       5.37 us/site  inner=1379  burst=  29.6 ms
    prepS       7.02 us/site  inner=1061  burst=  29.8 ms
      i8S/legA 1.2106x [1.2102-1.2113]  leg A wins
      i8U/legA 1.1913x [1.1912-1.1921]  leg A wins
      i8F/legA 1.1820x [1.1817-1.1826]  leg A wins
      glue fusion i8F/i8U 0.9921x [0.9916-0.9927]; multi-out i8U/i8S 0.9842x
      prepF   5.37 us  prepS   7.02 us  (prep share of i8F 7.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.126e-02 1.093e-02

  add+norm -> qkv|z M=16 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s2(TM16) 6144:s2(TM16)
    i8F        89.47 us/site  inner=84    burst=  30.1 ms
    i8S        90.78 us/site  inner=83    burst=  30.1 ms
    i8U        89.59 us/site  inner=84    burst=  30.1 ms
    legA       59.39 us/site  inner=127   burst=  30.2 ms
    prepF       5.79 us/site  inner=1274  burst=  29.5 ms
    prepS       7.04 us/site  inner=1056  burst=  29.7 ms
      i8S/legA 1.5287x [1.5277-1.5317]  leg A wins
      i8U/legA 1.5083x [1.5076-1.5103]  leg A wins
      i8F/legA 1.5062x [1.5056-1.5077]  leg A wins
      glue fusion i8F/i8U 0.9984x [0.9983-0.9988]; multi-out i8U/i8S 0.9868x
      prepF   5.79 us  prepS   7.04 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.126e-02 1.093e-02

  add+norm -> qkv|z M=32 tile s98  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        73.45 us/site  inner=103   burst=  30.3 ms
    i8S        75.32 us/site  inner=100   burst=  30.1 ms
    i8U        74.06 us/site  inner=102   burst=  30.2 ms
    legA       70.37 us/site  inner=107   burst=  30.1 ms
    prepF       5.46 us/site  inner=1355  burst=  29.6 ms
    prepS       7.19 us/site  inner=1033  burst=  29.7 ms
      i8S/legA 1.0704x [1.0686-1.0713]  leg A wins
      i8U/legA 1.0525x [1.0522-1.0548]  leg A wins
      i8F/legA 1.0439x [1.0436-1.0455]  leg A wins
      glue fusion i8F/i8U 0.9918x [0.9911-0.9924]; multi-out i8U/i8S 0.9834x
      prepF   5.46 us  prepS   7.19 us  (prep share of i8F 7.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.130e-02 1.087e-02

  add+norm -> qkv|z M=32 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        72.89 us/site  inner=103   burst=  30.0 ms
    i8S        74.73 us/site  inner=101   burst=  30.2 ms
    i8U        73.48 us/site  inner=103   burst=  30.3 ms
    legA       70.36 us/site  inner=107   burst=  30.1 ms
    prepF       5.45 us/site  inner=1363  burst=  29.7 ms
    prepS       7.18 us/site  inner=1035  burst=  29.7 ms
      i8S/legA 1.0621x [1.0618-1.0624]  leg A wins
      i8U/legA 1.0442x [1.0439-1.0447]  leg A wins
      i8F/legA 1.0360x [1.0356-1.0364]  leg A wins
      glue fusion i8F/i8U 0.9920x [0.9915-0.9924]; multi-out i8U/i8S 0.9833x
      prepF   5.45 us  prepS   7.18 us  (prep share of i8F 7.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.130e-02 1.087e-02

  add+norm -> qkv|z M=32 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        90.07 us/site  inner=84    burst=  30.3 ms
    i8S        91.50 us/site  inner=82    burst=  30.0 ms
    i8U        90.26 us/site  inner=83    burst=  30.0 ms
    legA       70.39 us/site  inner=107   burst=  30.1 ms
    prepF       5.91 us/site  inner=1254  burst=  29.7 ms
    prepS       7.22 us/site  inner=1029  burst=  29.7 ms
      i8S/legA 1.3002x [1.2903-1.3011]  leg A wins
      i8U/legA 1.2822x [1.2729-1.2836]  leg A wins
      i8F/legA 1.2796x [1.2699-1.2808]  leg A wins
      glue fusion i8F/i8U 0.9980x [0.9969-0.9984]; multi-out i8U/i8S 0.9864x
      prepF   5.91 us  prepS   7.22 us  (prep share of i8F 6.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.130e-02 1.087e-02

  add+norm -> qkv|z M=36 tile s98  reps=7 burst>=30 ms  cell 2.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        89.59 us/site  inner=84    burst=  30.1 ms
    i8S        90.97 us/site  inner=83    burst=  30.2 ms
    i8U        89.69 us/site  inner=84    burst=  30.1 ms
    legA      130.73 us/site  inner=58    burst=  30.3 ms
    prepF       5.66 us/site  inner=1310  burst=  29.7 ms
    prepS       7.26 us/site  inner=1020  burst=  29.6 ms
      i8S/legA 0.6957x [0.6943-0.6990]  int8 WINS
      i8U/legA 0.6857x [0.6845-0.6893]  int8 WINS
      i8F/legA 0.6840x [0.6823-0.6862]  int8 WINS
      glue fusion i8F/i8U 0.9969x [0.9955-0.9993]; multi-out i8U/i8S 0.9860x
      prepF   5.66 us  prepS   7.26 us  (prep share of i8F 6.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.132e-02 1.087e-02

  add+norm -> qkv|z M=36 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        89.68 us/site  inner=84    burst=  30.1 ms
    i8S        91.02 us/site  inner=83    burst=  30.2 ms
    i8U        89.80 us/site  inner=84    burst=  30.2 ms
    legA      130.68 us/site  inner=58    burst=  30.3 ms
    prepF       5.84 us/site  inner=1267  burst=  29.6 ms
    prepS       7.26 us/site  inner=1020  burst=  29.6 ms
      i8S/legA 0.6964x [0.6960-0.6968]  int8 WINS
      i8U/legA 0.6871x [0.6868-0.6875]  int8 WINS
      i8F/legA 0.6861x [0.6860-0.6867]  int8 WINS
      glue fusion i8F/i8U 0.9989x [0.9983-0.9991]; multi-out i8U/i8S 0.9867x
      prepF   5.84 us  prepS   7.26 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.132e-02 1.087e-02

  add+norm -> qkv|z M=36 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        90.33 us/site  inner=83    burst=  30.0 ms
    i8S        91.73 us/site  inner=82    burst=  30.1 ms
    i8U        90.44 us/site  inner=83    burst=  30.0 ms
    legA      130.63 us/site  inner=58    burst=  30.3 ms
    prepF       5.85 us/site  inner=1269  burst=  29.7 ms
    prepS       7.27 us/site  inner=1026  burst=  29.8 ms
      i8S/legA 0.7021x [0.7014-0.7025]  int8 WINS
      i8U/legA 0.6924x [0.6914-0.6925]  int8 WINS
      i8F/legA 0.6915x [0.6907-0.6917]  int8 WINS
      glue fusion i8F/i8U 0.9988x [0.9986-0.9992]; multi-out i8U/i8S 0.9859x
      prepF   5.85 us  prepS   7.27 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.132e-02 1.087e-02

  add+norm -> qkv|z M=48 tile s98  reps=7 burst>=30 ms  cell 1.8 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        89.79 us/site  inner=84    burst=  30.2 ms
    i8S        91.64 us/site  inner=82    burst=  30.1 ms
    i8U        90.02 us/site  inner=83    burst=  29.9 ms
    legA      132.10 us/site  inner=57    burst=  30.1 ms
    prepF       5.79 us/site  inner=1281  burst=  29.7 ms
    prepS       7.41 us/site  inner=1011  burst=  30.0 ms
      i8S/legA 0.6913x [0.6875-0.6943]  int8 WINS
      i8U/legA 0.6812x [0.6749-0.6832]  int8 WINS
      i8F/legA 0.6784x [0.6718-0.6809]  int8 WINS
      glue fusion i8F/i8U 0.9960x [0.9921-0.9996]; multi-out i8U/i8S 0.9841x
      prepF   5.79 us  prepS   7.41 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.140e-02 1.091e-02

  add+norm -> qkv|z M=48 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        89.89 us/site  inner=84    burst=  30.2 ms
    i8S        91.39 us/site  inner=82    burst=  30.0 ms
    i8U        90.06 us/site  inner=84    burst=  30.3 ms
    legA      132.09 us/site  inner=57    burst=  30.1 ms
    prepF       5.96 us/site  inner=1234  burst=  29.4 ms
    prepS       7.41 us/site  inner=1002  burst=  29.7 ms
      i8S/legA 0.6919x [0.6904-0.6921]  int8 WINS
      i8U/legA 0.6818x [0.6816-0.6821]  int8 WINS
      i8F/legA 0.6804x [0.6803-0.6808]  int8 WINS
      glue fusion i8F/i8U 0.9980x [0.9976-0.9984]; multi-out i8U/i8S 0.9855x
      prepF   5.96 us  prepS   7.41 us  (prep share of i8F 6.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.140e-02 1.091e-02

  add+norm -> qkv|z M=48 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        90.74 us/site  inner=83    burst=  30.1 ms
    i8S        92.21 us/site  inner=82    burst=  30.2 ms
    i8U        90.94 us/site  inner=83    burst=  30.2 ms
    legA      132.00 us/site  inner=57    burst=  30.1 ms
    prepF       5.95 us/site  inner=1247  burst=  29.7 ms
    prepS       7.42 us/site  inner=1003  burst=  29.8 ms
      i8S/legA 0.6987x [0.6984-0.6992]  int8 WINS
      i8U/legA 0.6891x [0.6887-0.6894]  int8 WINS
      i8F/legA 0.6875x [0.6872-0.6878]  int8 WINS
      glue fusion i8F/i8U 0.9978x [0.9974-0.9982]; multi-out i8U/i8S 0.9863x
      prepF   5.95 us  prepS   7.42 us  (prep share of i8F 6.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.140e-02 1.091e-02

  add+norm -> qkv|z M=64 tile s98  reps=7 burst>=30 ms  cell 1.8 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F       107.40 us/site  inner=70    burst=  30.1 ms
    i8S       108.67 us/site  inner=69    burst=  30.0 ms
    i8U       108.29 us/site  inner=70    burst=  30.3 ms
    legA      136.69 us/site  inner=56    burst=  30.6 ms
    prepF       6.11 us/site  inner=1223  burst=  29.9 ms
    prepS       7.59 us/site  inner=990   burst=  30.1 ms
      i8S/legA 0.8033x [0.7899-0.8187]  int8 WINS
      i8U/legA 0.7950x [0.7844-0.8034]  int8 WINS
      i8F/legA 0.7872x [0.7806-0.7942]  int8 WINS
      glue fusion i8F/i8U 0.9924x [0.9825-0.9977]; multi-out i8U/i8S 0.9939x
      prepF   6.11 us  prepS   7.59 us  (prep share of i8F 5.7%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.136e-02 1.087e-02

  add+norm -> qkv|z M=64 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        90.01 us/site  inner=84    burst=  30.2 ms
    i8S        91.54 us/site  inner=82    burst=  30.0 ms
    i8U        90.18 us/site  inner=84    burst=  30.3 ms
    legA      134.94 us/site  inner=56    burst=  30.2 ms
    prepF       6.07 us/site  inner=1228  burst=  29.8 ms
    prepS       7.54 us/site  inner=985   burst=  29.7 ms
      i8S/legA 0.6784x [0.6783-0.6788]  int8 WINS
      i8U/legA 0.6682x [0.6681-0.6686]  int8 WINS
      i8F/legA 0.6670x [0.6668-0.6672]  int8 WINS
      glue fusion i8F/i8U 0.9981x [0.9978-0.9982]; multi-out i8U/i8S 0.9850x
      prepF   6.07 us  prepS   7.54 us  (prep share of i8F 6.7%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.136e-02 1.087e-02

  add+norm -> qkv|z M=64 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 10240:s5(TM32) 6144:s5(TM32)
    i8F        91.64 us/site  inner=82    burst=  30.1 ms
    i8S        93.19 us/site  inner=81    burst=  30.2 ms
    i8U        91.83 us/site  inner=82    burst=  30.1 ms
    legA      134.89 us/site  inner=56    burst=  30.2 ms
    prepF       6.07 us/site  inner=1222  burst=  29.7 ms
    prepS       7.54 us/site  inner=989   burst=  29.8 ms
      i8S/legA 0.6909x [0.6903-0.6913]  int8 WINS
      i8U/legA 0.6807x [0.6803-0.6845]  int8 WINS
      i8F/legA 0.6792x [0.6788-0.6797]  int8 WINS
      glue fusion i8F/i8U 0.9978x [0.9923-0.9982]; multi-out i8U/i8S 0.9854x
      prepF   6.07 us  prepS   7.54 us  (prep share of i8F 6.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.136e-02 1.087e-02

########## add+norm -> q|k|v  G=3  x16  [5120x12288xK4 5120x1024xK4 5120x1024xK4] ##########
    site tile s98 TILE_M=16 rows_pad=16 groups=40 rot=5 trellis=36.7 MB/pass

  add+norm -> q|k|v M=8 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 12288:s3(TM16) 1024:s2(TM16) 1024:s2(TM16)
    i8F        93.91 us/site  inner=64    burst=  30.1 ms
    i8S        96.63 us/site  inner=62    burst=  30.0 ms
    i8U        94.38 us/site  inner=64    burst=  30.2 ms
    legA       68.57 us/site  inner=88    burst=  30.2 ms
    prepF       6.72 us/site  inner=888   burst=  29.8 ms
    prepS       9.38 us/site  inner=639   burst=  30.0 ms
      i8S/legA 1.4089x [1.4084-1.4095]  leg A wins
      i8U/legA 1.3763x [1.3758-1.3768]  leg A wins
      i8F/legA 1.3696x [1.3688-1.3697]  leg A wins
      glue fusion i8F/i8U 0.9949x [0.9948-0.9952]; multi-out i8U/i8S 0.9767x
      prepF   6.72 us  prepS   9.38 us  (prep share of i8F 7.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.096e-02 1.095e-02 1.113e-02
    site tile s94 TILE_M=32 rows_pad=32 groups=40 rot=5 trellis=36.7 MB/pass

  add+norm -> q|k|v M=8 tile s94  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s2(TM16) 1024:s2(TM16)
    i8F       105.71 us/site  inner=57    burst=  30.1 ms
    i8S       108.32 us/site  inner=56    burst=  30.3 ms
    i8U       106.08 us/site  inner=57    burst=  30.2 ms
    legA       68.58 us/site  inner=88    burst=  30.2 ms
    prepF       6.78 us/site  inner=876   burst=  29.7 ms
    prepS       9.39 us/site  inner=636   burst=  29.9 ms
      i8S/legA 1.5796x [1.5784-1.5802]  leg A wins
      i8U/legA 1.5469x [1.5457-1.5471]  leg A wins
      i8F/legA 1.5412x [1.5405-1.5414]  leg A wins
      glue fusion i8F/i8U 0.9963x [0.9961-0.9966]; multi-out i8U/i8S 0.9793x
      prepF   6.78 us  prepS   9.39 us  (prep share of i8F 6.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.096e-02 1.095e-02 1.113e-02
    site tile s85 TILE_M=64 rows_pad=64 groups=40 rot=5 trellis=36.7 MB/pass

  add+norm -> q|k|v M=8 tile s85  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s2(TM16) 1024:s2(TM16)
    i8F       131.70 us/site  inner=46    burst=  30.3 ms
    i8S       133.78 us/site  inner=45    burst=  30.1 ms
    i8U       131.53 us/site  inner=46    burst=  30.3 ms
    legA       68.57 us/site  inner=88    burst=  30.2 ms
    prepF       7.43 us/site  inner=803   burst=  29.8 ms
    prepS       9.44 us/site  inner=632   burst=  29.8 ms
      i8S/legA 1.9510x [1.9506-1.9517]  leg A wins
      i8U/legA 1.9182x [1.9178-1.9191]  leg A wins
      i8F/legA 1.9208x [1.9201-1.9215]  leg A wins
      glue fusion i8F/i8U 1.0011x [1.0008-1.0019]; multi-out i8U/i8S 0.9833x
      prepF   7.43 us  prepS   9.44 us  (prep share of i8F 5.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.096e-02 1.095e-02 1.113e-02

  add+norm -> q|k|v M=16 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 12288:s3(TM16) 1024:s2(TM16) 1024:s2(TM16)
    i8F        94.74 us/site  inner=64    burst=  30.3 ms
    i8S        97.53 us/site  inner=62    burst=  30.2 ms
    i8U        95.28 us/site  inner=63    burst=  30.0 ms
    legA       73.17 us/site  inner=82    burst=  30.0 ms
    prepF       6.71 us/site  inner=885   burst=  29.7 ms
    prepS       9.44 us/site  inner=631   burst=  29.8 ms
      i8S/legA 1.3328x [1.3325-1.3336]  leg A wins
      i8U/legA 1.3022x [1.3018-1.3026]  leg A wins
      i8F/legA 1.2948x [1.2944-1.2954]  leg A wins
      glue fusion i8F/i8U 0.9943x [0.9942-0.9948]; multi-out i8U/i8S 0.9770x
      prepF   6.71 us  prepS   9.44 us  (prep share of i8F 7.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.084e-02 1.081e-02 1.082e-02

  add+norm -> q|k|v M=16 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 12288:s3(TM16) 1024:s2(TM16) 1024:s2(TM16)
    i8F       106.28 us/site  inner=57    burst=  30.3 ms
    i8S       108.96 us/site  inner=55    burst=  30.0 ms
    i8U       106.69 us/site  inner=57    burst=  30.4 ms
    legA       73.18 us/site  inner=82    burst=  30.0 ms
    prepF       6.76 us/site  inner=882   burst=  29.8 ms
    prepS       9.45 us/site  inner=632   burst=  29.9 ms
      i8S/legA 1.4888x [1.4885-1.4967]  leg A wins
      i8U/legA 1.4580x [1.4571-1.4589]  leg A wins
      i8F/legA 1.4523x [1.4518-1.4529]  leg A wins
      glue fusion i8F/i8U 0.9963x [0.9958-0.9965]; multi-out i8U/i8S 0.9791x
      prepF   6.76 us  prepS   9.45 us  (prep share of i8F 6.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.084e-02 1.081e-02 1.082e-02

  add+norm -> q|k|v M=16 tile s85  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s2(TM16) 1024:s2(TM16)
    i8F       132.25 us/site  inner=46    burst=  30.4 ms
    i8S       134.30 us/site  inner=45    burst=  30.2 ms
    i8U       132.05 us/site  inner=46    burst=  30.4 ms
    legA       73.17 us/site  inner=82    burst=  30.0 ms
    prepF       7.43 us/site  inner=801   burst=  29.8 ms
    prepS       9.46 us/site  inner=632   burst=  29.9 ms
      i8S/legA 1.8354x [1.8344-1.8365]  leg A wins
      i8U/legA 1.8048x [1.8040-1.8057]  leg A wins
      i8F/legA 1.8074x [1.8069-1.8083]  leg A wins
      glue fusion i8F/i8U 1.0016x [1.0013-1.0020]; multi-out i8U/i8S 0.9833x
      prepF   7.43 us  prepS   9.46 us  (prep share of i8F 5.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.084e-02 1.081e-02 1.082e-02

  add+norm -> q|k|v M=32 tile s98  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       100.92 us/site  inner=60    burst=  30.3 ms
    i8S       103.89 us/site  inner=58    burst=  30.1 ms
    i8U       101.49 us/site  inner=60    burst=  30.4 ms
    legA      109.61 us/site  inner=55    burst=  30.1 ms
    prepF       6.81 us/site  inner=875   burst=  29.8 ms
    prepS       9.67 us/site  inner=621   burst=  30.0 ms
      i8S/legA 0.9478x [0.9471-0.9482]  int8 WINS
      i8U/legA 0.9260x [0.9259-0.9264]  int8 WINS
      i8F/legA 0.9208x [0.9202-0.9212]  int8 WINS
      glue fusion i8F/i8U 0.9943x [0.9938-0.9946]; multi-out i8U/i8S 0.9771x
      prepF   6.81 us  prepS   9.67 us  (prep share of i8F 6.7%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.082e-02 1.089e-02

  add+norm -> q|k|v M=32 tile s94  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       109.54 us/site  inner=55    burst=  30.1 ms
    i8S       112.46 us/site  inner=54    burst=  30.4 ms
    i8U       110.12 us/site  inner=55    burst=  30.3 ms
    legA      109.65 us/site  inner=55    burst=  30.2 ms
    prepF       6.80 us/site  inner=869   burst=  29.6 ms
    prepS       9.67 us/site  inner=617   burst=  29.8 ms
      i8S/legA 1.0257x [1.0255-1.0263]  leg A wins
      i8U/legA 1.0043x [1.0037-1.0050]  leg A wins
      i8F/legA 0.9990x [0.9986-0.9994]  int8 WINS
      glue fusion i8F/i8U 0.9947x [0.9943-0.9957]; multi-out i8U/i8S 0.9792x
      prepF   6.80 us  prepS   9.67 us  (prep share of i8F 6.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.082e-02 1.089e-02

  add+norm -> q|k|v M=32 tile s85  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       133.24 us/site  inner=46    burst=  30.6 ms
    i8S       135.68 us/site  inner=45    burst=  30.5 ms
    i8U       133.27 us/site  inner=46    burst=  30.7 ms
    legA      109.67 us/site  inner=55    burst=  30.2 ms
    prepF       7.42 us/site  inner=801   burst=  29.7 ms
    prepS       9.72 us/site  inner=616   burst=  29.9 ms
      i8S/legA 1.2370x [1.2364-1.2377]  leg A wins
      i8U/legA 1.2153x [1.2147-1.2157]  leg A wins
      i8F/legA 1.2152x [1.2140-1.2155]  leg A wins
      glue fusion i8F/i8U 0.9998x [0.9994-1.0000]; multi-out i8U/i8S 0.9823x
      prepF   7.42 us  prepS   9.72 us  (prep share of i8F 5.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.082e-02 1.089e-02

  add+norm -> q|k|v M=36 tile s98  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       119.58 us/site  inner=51    burst=  30.5 ms
    i8S       122.28 us/site  inner=50    burst=  30.6 ms
    i8U       119.85 us/site  inner=51    burst=  30.6 ms
    legA      164.40 us/site  inner=37    burst=  30.4 ms
    prepF       7.12 us/site  inner=838   burst=  29.8 ms
    prepS       9.73 us/site  inner=614   burst=  29.9 ms
      i8S/legA 0.7439x [0.7380-0.7442]  int8 WINS
      i8U/legA 0.7290x [0.7231-0.7294]  int8 WINS
      i8F/legA 0.7274x [0.7218-0.7276]  int8 WINS
      glue fusion i8F/i8U 0.9978x [0.9967-0.9981]; multi-out i8U/i8S 0.9802x
      prepF   7.12 us  prepS   9.73 us  (prep share of i8F 6.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.080e-02 1.083e-02 1.086e-02

  add+norm -> q|k|v M=36 tile s94  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       124.29 us/site  inner=49    burst=  30.5 ms
    i8S       126.66 us/site  inner=48    burst=  30.4 ms
    i8U       124.20 us/site  inner=49    burst=  30.4 ms
    legA      164.39 us/site  inner=37    burst=  30.4 ms
    prepF       7.41 us/site  inner=803   burst=  29.8 ms
    prepS       9.76 us/site  inner=612   burst=  29.9 ms
      i8S/legA 0.7703x [0.7701-0.7707]  int8 WINS
      i8U/legA 0.7556x [0.7546-0.7562]  int8 WINS
      i8F/legA 0.7559x [0.7557-0.7562]  int8 WINS
      glue fusion i8F/i8U 1.0006x [0.9993-1.0017]; multi-out i8U/i8S 0.9807x
      prepF   7.41 us  prepS   9.76 us  (prep share of i8F 6.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.080e-02 1.083e-02 1.086e-02

  add+norm -> q|k|v M=36 tile s85  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       133.40 us/site  inner=45    burst=  30.0 ms
    i8S       135.76 us/site  inner=45    burst=  30.5 ms
    i8U       133.36 us/site  inner=46    burst=  30.7 ms
    legA      164.20 us/site  inner=37    burst=  30.4 ms
    prepF       7.41 us/site  inner=806   burst=  29.9 ms
    prepS       9.75 us/site  inner=614   burst=  29.9 ms
      i8S/legA 0.8268x [0.8264-0.8270]  int8 WINS
      i8U/legA 0.8121x [0.8118-0.8124]  int8 WINS
      i8F/legA 0.8121x [0.8120-0.8126]  int8 WINS
      glue fusion i8F/i8U 1.0002x [0.9997-1.0007]; multi-out i8U/i8S 0.9824x
      prepF   7.41 us  prepS   9.75 us  (prep share of i8F 5.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.080e-02 1.083e-02 1.086e-02

  add+norm -> q|k|v M=48 tile s98  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       119.49 us/site  inner=51    burst=  30.5 ms
    i8S       122.33 us/site  inner=50    burst=  30.6 ms
    i8U       119.86 us/site  inner=51    burst=  30.6 ms
    legA      169.04 us/site  inner=36    burst=  30.4 ms
    prepF       7.15 us/site  inner=835   burst=  29.9 ms
    prepS       9.86 us/site  inner=608   burst=  30.0 ms
      i8S/legA 0.7237x [0.7233-0.7241]  int8 WINS
      i8U/legA 0.7090x [0.7087-0.7096]  int8 WINS
      i8F/legA 0.7068x [0.7066-0.7070]  int8 WINS
      glue fusion i8F/i8U 0.9968x [0.9958-0.9975]; multi-out i8U/i8S 0.9797x
      prepF   7.15 us  prepS   9.86 us  (prep share of i8F 6.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.080e-02 1.076e-02 1.085e-02

  add+norm -> q|k|v M=48 tile s94  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       124.17 us/site  inner=49    burst=  30.4 ms
    i8S       126.73 us/site  inner=48    burst=  30.4 ms
    i8U       124.25 us/site  inner=49    burst=  30.4 ms
    legA      169.09 us/site  inner=36    burst=  30.4 ms
    prepF       7.39 us/site  inner=806   burst=  29.8 ms
    prepS       9.89 us/site  inner=605   burst=  29.9 ms
      i8S/legA 0.7496x [0.7478-0.7503]  int8 WINS
      i8U/legA 0.7350x [0.7330-0.7356]  int8 WINS
      i8F/legA 0.7342x [0.7332-0.7347]  int8 WINS
      glue fusion i8F/i8U 0.9990x [0.9981-1.0002]; multi-out i8U/i8S 0.9807x
      prepF   7.39 us  prepS   9.89 us  (prep share of i8F 6.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.080e-02 1.076e-02 1.085e-02

  add+norm -> q|k|v M=48 tile s85  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       133.91 us/site  inner=45    burst=  30.1 ms
    i8S       136.48 us/site  inner=44    burst=  30.0 ms
    i8U       134.02 us/site  inner=45    burst=  30.2 ms
    legA      169.03 us/site  inner=36    burst=  30.4 ms
    prepF       7.38 us/site  inner=809   burst=  29.8 ms
    prepS       9.89 us/site  inner=605   burst=  29.9 ms
      i8S/legA 0.8071x [0.8068-0.8118]  int8 WINS
      i8U/legA 0.7928x [0.7920-0.7933]  int8 WINS
      i8F/legA 0.7922x [0.7914-0.7928]  int8 WINS
      glue fusion i8F/i8U 0.9994x [0.9990-0.9997]; multi-out i8U/i8S 0.9819x
      prepF   7.38 us  prepS   9.89 us  (prep share of i8F 5.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.080e-02 1.076e-02 1.085e-02

  add+norm -> q|k|v M=64 tile s98  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       120.09 us/site  inner=51    burst=  30.6 ms
    i8S       122.79 us/site  inner=49    burst=  30.1 ms
    i8U       120.23 us/site  inner=50    burst=  30.1 ms
    legA      203.32 us/site  inner=30    burst=  30.5 ms
    prepF       7.43 us/site  inner=802   burst=  29.8 ms
    prepS      10.07 us/site  inner=594   burst=  29.9 ms
      i8S/legA 0.6039x [0.6029-0.6043]  int8 WINS
      i8U/legA 0.5913x [0.5907-0.5916]  int8 WINS
      i8F/legA 0.5906x [0.5900-0.5908]  int8 WINS
      glue fusion i8F/i8U 0.9987x [0.9981-0.9993]; multi-out i8U/i8S 0.9791x
      prepF   7.43 us  prepS  10.07 us  (prep share of i8F 6.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.081e-02 1.073e-02 1.085e-02

  add+norm -> q|k|v M=64 tile s94  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       124.33 us/site  inner=49    burst=  30.5 ms
    i8S       127.04 us/site  inner=48    burst=  30.5 ms
    i8U       124.51 us/site  inner=49    burst=  30.5 ms
    legA      203.22 us/site  inner=30    burst=  30.5 ms
    prepF       7.44 us/site  inner=801   burst=  29.8 ms
    prepS      10.08 us/site  inner=594   burst=  29.9 ms
      i8S/legA 0.6252x [0.6243-0.6253]  int8 WINS
      i8U/legA 0.6127x [0.6117-0.6128]  int8 WINS
      i8F/legA 0.6118x [0.6113-0.6119]  int8 WINS
      glue fusion i8F/i8U 0.9986x [0.9984-0.9996]; multi-out i8U/i8S 0.9800x
      prepF   7.44 us  prepS  10.08 us  (prep share of i8F 6.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.081e-02 1.073e-02 1.085e-02

  add+norm -> q|k|v M=64 tile s85  reps=7 burst>=30 ms  cell 1.4 s
    legA pins 12288:s3(TM16) 1024:s5(TM32) 1024:s5(TM32)
    i8F       135.01 us/site  inner=45    burst=  30.4 ms
    i8S       137.71 us/site  inner=44    burst=  30.3 ms
    i8U       135.17 us/site  inner=45    burst=  30.4 ms
    legA      203.25 us/site  inner=30    burst=  30.5 ms
    prepF       7.42 us/site  inner=798   burst=  29.6 ms
    prepS      10.04 us/site  inner=596   burst=  29.9 ms
      i8S/legA 0.6775x [0.6771-0.6781]  int8 WINS
      i8U/legA 0.6650x [0.6646-0.6654]  int8 WINS
      i8F/legA 0.6642x [0.6639-0.6647]  int8 WINS
      glue fusion i8F/i8U 0.9989x [0.9987-0.9990]; multi-out i8U/i8S 0.9815x
      prepF   7.42 us  prepS  10.04 us  (prep share of i8F 5.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.081e-02 1.073e-02 1.085e-02

########## add+norm -> gate|up  G=2  x64  [5120x17408xK4 5120x17408xK4] ##########
    site tile s98 TILE_M=16 rows_pad=16 groups=40 rot=2 trellis=89.1 MB/pass

  add+norm -> gate|up M=8 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s4(TM16) 17408:s4(TM16)
    i8F       107.44 us/site  inner=139   burst=  29.9 ms
    i8S       108.01 us/site  inner=139   burst=  30.0 ms
    i8U       107.71 us/site  inner=140   burst=  30.2 ms
    legA      112.96 us/site  inner=134   burst=  30.3 ms
    prepF       5.49 us/site  inner=2694  burst=  29.6 ms
    prepS       7.19 us/site  inner=2056  burst=  29.5 ms
      i8S/legA 0.9558x [0.9556-0.9565]  int8 WINS
      i8U/legA 0.9533x [0.9529-0.9537]  int8 WINS
      i8F/legA 0.9510x [0.9446-0.9595]  int8 WINS
      glue fusion i8F/i8U 0.9974x [0.9904-1.0065]; multi-out i8U/i8S 0.9972x
      prepF   5.49 us  prepS   7.19 us  (prep share of i8F 5.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.079e-02 1.072e-02
    site tile s94 TILE_M=32 rows_pad=32 groups=40 rot=2 trellis=89.1 MB/pass

  add+norm -> gate|up M=8 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s4(TM16) 17408:s4(TM16)
    i8F       111.04 us/site  inner=135   burst=  30.0 ms
    i8S       111.51 us/site  inner=135   burst=  30.1 ms
    i8U       111.26 us/site  inner=135   burst=  30.0 ms
    legA      113.03 us/site  inner=133   burst=  30.1 ms
    prepF       5.50 us/site  inner=2665  burst=  29.3 ms
    prepS       7.20 us/site  inner=2043  burst=  29.4 ms
      i8S/legA 0.9867x [0.9858-0.9870]  int8 WINS
      i8U/legA 0.9844x [0.9837-0.9849]  int8 WINS
      i8F/legA 0.9826x [0.9818-0.9827]  int8 WINS
      glue fusion i8F/i8U 0.9980x [0.9977-0.9983]; multi-out i8U/i8S 0.9977x
      prepF   5.50 us  prepS   7.20 us  (prep share of i8F 5.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.079e-02 1.072e-02
    site tile s85 TILE_M=64 rows_pad=64 groups=40 rot=2 trellis=89.1 MB/pass

  add+norm -> gate|up M=8 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s4(TM16) 17408:s4(TM16)
    i8F       135.83 us/site  inner=111   burst=  30.2 ms
    i8S       136.92 us/site  inner=110   burst=  30.1 ms
    i8U       135.84 us/site  inner=111   burst=  30.2 ms
    legA      113.01 us/site  inner=133   burst=  30.1 ms
    prepF       5.90 us/site  inner=2494  burst=  29.4 ms
    prepS       7.24 us/site  inner=2039  burst=  29.5 ms
      i8S/legA 1.2114x [1.2107-1.2130]  leg A wins
      i8U/legA 1.2019x [1.2008-1.2025]  leg A wins
      i8F/legA 1.2021x [1.2003-1.2025]  leg A wins
      glue fusion i8F/i8U 0.9998x [0.9990-1.0013]; multi-out i8U/i8S 0.9921x
      prepF   5.90 us  prepS   7.24 us  (prep share of i8F 4.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.079e-02 1.072e-02

  add+norm -> gate|up M=16 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s4(TM16) 17408:s4(TM16)
    i8F       108.61 us/site  inner=139   burst=  30.2 ms
    i8S       109.27 us/site  inner=137   burst=  29.9 ms
    i8U       108.83 us/site  inner=138   burst=  30.0 ms
    legA      121.23 us/site  inner=124   burst=  30.1 ms
    prepF       5.51 us/site  inner=2640  burst=  29.1 ms
    prepS       7.27 us/site  inner=2039  burst=  29.6 ms
      i8S/legA 0.9014x [0.9009-0.9022]  int8 WINS
      i8U/legA 0.8977x [0.8974-0.8984]  int8 WINS
      i8F/legA 0.8960x [0.8954-0.8965]  int8 WINS
      glue fusion i8F/i8U 0.9982x [0.9973-0.9985]; multi-out i8U/i8S 0.9959x
      prepF   5.51 us  prepS   7.27 us  (prep share of i8F 5.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.079e-02 1.079e-02

  add+norm -> gate|up M=16 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s4(TM16) 17408:s4(TM16)
    i8F       112.27 us/site  inner=134   burst=  30.1 ms
    i8S       112.80 us/site  inner=133   burst=  30.0 ms
    i8U       112.51 us/site  inner=134   burst=  30.2 ms
    legA      121.18 us/site  inner=124   burst=  30.1 ms
    prepF       5.49 us/site  inner=2664  burst=  29.3 ms
    prepS       7.28 us/site  inner=2039  burst=  29.7 ms
      i8S/legA 0.9308x [0.9304-0.9317]  int8 WINS
      i8U/legA 0.9285x [0.9279-0.9289]  int8 WINS
      i8F/legA 0.9263x [0.9261-0.9267]  int8 WINS
      glue fusion i8F/i8U 0.9979x [0.9973-0.9983]; multi-out i8U/i8S 0.9973x
      prepF   5.49 us  prepS   7.28 us  (prep share of i8F 4.9%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.079e-02 1.079e-02

  add+norm -> gate|up M=16 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s4(TM16) 17408:s4(TM16)
    i8F       136.40 us/site  inner=110   burst=  30.0 ms
    i8S       137.63 us/site  inner=110   burst=  30.3 ms
    i8U       136.48 us/site  inner=111   burst=  30.3 ms
    legA      121.20 us/site  inner=124   burst=  30.1 ms
    prepF       5.93 us/site  inner=2468  burst=  29.3 ms
    prepS       7.29 us/site  inner=2024  burst=  29.5 ms
      i8S/legA 1.1355x [1.1348-1.1362]  leg A wins
      i8U/legA 1.1258x [1.1254-1.1266]  leg A wins
      i8F/legA 1.1251x [1.1240-1.1255]  leg A wins
      glue fusion i8F/i8U 0.9992x [0.9981-0.9997]; multi-out i8U/i8S 0.9915x
      prepF   5.93 us  prepS   7.29 us  (prep share of i8F 4.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.079e-02 1.079e-02

  add+norm -> gate|up M=32 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       126.82 us/site  inner=120   burst=  30.4 ms
    i8S       127.11 us/site  inner=118   burst=  30.0 ms
    i8U       126.03 us/site  inner=119   burst=  30.0 ms
    legA      140.95 us/site  inner=107   burst=  30.2 ms
    prepF       5.58 us/site  inner=2634  burst=  29.4 ms
    prepS       7.39 us/site  inner=1987  burst=  29.4 ms
      i8S/legA 0.9022x [0.8929-0.9193]  int8 WINS
      i8U/legA 0.8943x [0.8852-0.9013]  int8 WINS
      i8F/legA 0.8950x [0.8840-0.9072]  int8 WINS
      glue fusion i8F/i8U 1.0068x [0.9902-1.0150]; multi-out i8U/i8S 0.9904x
      prepF   5.58 us  prepS   7.39 us  (prep share of i8F 4.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.084e-02 1.085e-02

  add+norm -> gate|up M=32 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       114.24 us/site  inner=132   burst=  30.2 ms
    i8S       115.11 us/site  inner=130   burst=  29.9 ms
    i8U       114.74 us/site  inner=131   burst=  30.1 ms
    legA      140.73 us/site  inner=107   burst=  30.1 ms
    prepF       5.56 us/site  inner=2634  burst=  29.3 ms
    prepS       7.40 us/site  inner=2004  burst=  29.7 ms
      i8S/legA 0.8178x [0.8174-0.8183]  int8 WINS
      i8U/legA 0.8151x [0.8149-0.8156]  int8 WINS
      i8F/legA 0.8117x [0.8111-0.8123]  int8 WINS
      glue fusion i8F/i8U 0.9959x [0.9954-0.9961]; multi-out i8U/i8S 0.9967x
      prepF   5.56 us  prepS   7.40 us  (prep share of i8F 4.9%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.084e-02 1.085e-02

  add+norm -> gate|up M=32 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       136.96 us/site  inner=110   burst=  30.1 ms
    i8S       138.26 us/site  inner=109   burst=  30.1 ms
    i8U       137.07 us/site  inner=110   burst=  30.2 ms
    legA      140.78 us/site  inner=107   burst=  30.1 ms
    prepF       5.98 us/site  inner=2451  burst=  29.3 ms
    prepS       7.44 us/site  inner=1987  burst=  29.6 ms
      i8S/legA 0.9821x [0.9819-0.9824]  int8 WINS
      i8U/legA 0.9737x [0.9733-0.9740]  int8 WINS
      i8F/legA 0.9727x [0.9725-0.9735]  int8 WINS
      glue fusion i8F/i8U 0.9993x [0.9987-0.9998]; multi-out i8U/i8S 0.9914x
      prepF   5.98 us  prepS   7.44 us  (prep share of i8F 4.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.084e-02 1.085e-02

  add+norm -> gate|up M=36 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       165.93 us/site  inner=91    burst=  30.2 ms
    i8S       166.49 us/site  inner=91    burst=  30.3 ms
    i8U       165.67 us/site  inner=91    burst=  30.2 ms
    legA      252.30 us/site  inner=61    burst=  30.8 ms
    prepF       5.95 us/site  inner=2548  burst=  30.3 ms
    prepS       7.46 us/site  inner=1974  burst=  29.5 ms
      i8S/legA 0.6629x [0.6562-0.6759]  int8 WINS
      i8U/legA 0.6569x [0.6526-0.6906]  int8 WINS
      i8F/legA 0.6582x [0.6524-0.6906]  int8 WINS
      glue fusion i8F/i8U 0.9998x [0.9752-1.0192]; multi-out i8U/i8S 0.9945x
      prepF   5.95 us  prepS   7.46 us  (prep share of i8F 3.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.082e-02

  add+norm -> gate|up M=36 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       167.55 us/site  inner=90    burst=  30.2 ms
    i8S       167.17 us/site  inner=90    burst=  30.1 ms
    i8U       167.67 us/site  inner=90    burst=  30.2 ms
    legA      248.19 us/site  inner=61    burst=  30.3 ms
    prepF       6.11 us/site  inner=2417  burst=  29.5 ms
    prepS       7.46 us/site  inner=1960  burst=  29.2 ms
      i8S/legA 0.6735x [0.6734-0.6738]  int8 WINS
      i8U/legA 0.6756x [0.6754-0.6761]  int8 WINS
      i8F/legA 0.6752x [0.6748-0.6754]  int8 WINS
      glue fusion i8F/i8U 0.9992x [0.9988-0.9996]; multi-out i8U/i8S 1.0031x
      prepF   6.11 us  prepS   7.46 us  (prep share of i8F 3.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.082e-02

  add+norm -> gate|up M=36 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       137.36 us/site  inner=110   burst=  30.2 ms
    i8S       138.62 us/site  inner=109   burst=  30.2 ms
    i8U       137.36 us/site  inner=110   burst=  30.2 ms
    legA      248.19 us/site  inner=61    burst=  30.3 ms
    prepF       6.10 us/site  inner=2417  burst=  29.5 ms
    prepS       7.44 us/site  inner=1970  burst=  29.3 ms
      i8S/legA 0.5585x [0.5584-0.5590]  int8 WINS
      i8U/legA 0.5536x [0.5534-0.5566]  int8 WINS
      i8F/legA 0.5534x [0.5528-0.5540]  int8 WINS
      glue fusion i8F/i8U 0.9995x [0.9944-1.0005]; multi-out i8U/i8S 0.9912x
      prepF   6.10 us  prepS   7.44 us  (prep share of i8F 4.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.082e-02

  add+norm -> gate|up M=48 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       166.61 us/site  inner=91    burst=  30.3 ms
    i8S       167.17 us/site  inner=90    burst=  30.1 ms
    i8U       167.01 us/site  inner=90    burst=  30.1 ms
    legA      265.35 us/site  inner=58    burst=  30.8 ms
    prepF       5.96 us/site  inner=2468  burst=  29.4 ms
    prepS       7.77 us/site  inner=1954  burst=  30.4 ms
      i8S/legA 0.6340x [0.6275-0.6448]  int8 WINS
      i8U/legA 0.6414x [0.6268-0.6544]  int8 WINS
      i8F/legA 0.6406x [0.6256-0.6561]  int8 WINS
      glue fusion i8F/i8U 0.9966x [0.9754-1.0293]; multi-out i8U/i8S 0.9991x
      prepF   5.96 us  prepS   7.77 us  (prep share of i8F 3.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.080e-02

  add+norm -> gate|up M=48 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       169.51 us/site  inner=89    burst=  30.2 ms
    i8S       168.88 us/site  inner=89    burst=  30.1 ms
    i8U       169.53 us/site  inner=89    burst=  30.2 ms
    legA      259.80 us/site  inner=58    burst=  30.1 ms
    prepF       6.08 us/site  inner=2393  burst=  29.1 ms
    prepS       7.58 us/site  inner=1954  burst=  29.6 ms
      i8S/legA 0.6499x [0.6498-0.6502]  int8 WINS
      i8U/legA 0.6525x [0.6523-0.6526]  int8 WINS
      i8F/legA 0.6524x [0.6522-0.6527]  int8 WINS
      glue fusion i8F/i8U 0.9998x [0.9996-1.0002]; multi-out i8U/i8S 1.0038x
      prepF   6.08 us  prepS   7.58 us  (prep share of i8F 3.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.080e-02

  add+norm -> gate|up M=48 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       138.37 us/site  inner=109   burst=  30.2 ms
    i8S       139.50 us/site  inner=108   burst=  30.1 ms
    i8U       138.34 us/site  inner=109   burst=  30.2 ms
    legA      259.80 us/site  inner=58    burst=  30.1 ms
    prepF       6.08 us/site  inner=2425  burst=  29.5 ms
    prepS       7.58 us/site  inner=1956  burst=  29.7 ms
      i8S/legA 0.5370x [0.5368-0.5371]  int8 WINS
      i8U/legA 0.5325x [0.5323-0.5327]  int8 WINS
      i8F/legA 0.5327x [0.5323-0.5329]  int8 WINS
      glue fusion i8F/i8U 1.0001x [0.9996-1.0008]; multi-out i8U/i8S 0.9918x
      prepF   6.08 us  prepS   7.58 us  (prep share of i8F 4.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.082e-02 1.080e-02

  add+norm -> gate|up M=64 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       203.87 us/site  inner=74    burst=  30.2 ms
    i8S       205.13 us/site  inner=74    burst=  30.4 ms
    i8U       204.28 us/site  inner=74    burst=  30.2 ms
    legA      282.34 us/site  inner=55    burst=  31.1 ms
    prepF       6.23 us/site  inner=2368  burst=  29.5 ms
    prepS       7.82 us/site  inner=1911  burst=  29.9 ms
      i8S/legA 0.7353x [0.7062-0.7627]  int8 WINS
      i8U/legA 0.7265x [0.7047-0.7494]  int8 WINS
      i8F/legA 0.7257x [0.7017-0.7715]  int8 WINS
      glue fusion i8F/i8U 0.9981x [0.9422-1.0508]; multi-out i8U/i8S 0.9979x
      prepF   6.23 us  prepS   7.82 us  (prep share of i8F 3.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.081e-02 1.080e-02

  add+norm -> gate|up M=64 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       172.29 us/site  inner=88    burst=  30.3 ms
    i8S       171.10 us/site  inner=88    burst=  30.1 ms
    i8U       172.06 us/site  inner=87    burst=  29.9 ms
    legA      277.23 us/site  inner=55    burst=  30.5 ms
    prepF       6.22 us/site  inner=2351  burst=  29.2 ms
    prepS       7.75 us/site  inner=1906  burst=  29.5 ms
      i8S/legA 0.6172x [0.6170-0.6175]  int8 WINS
      i8U/legA 0.6207x [0.6206-0.6210]  int8 WINS
      i8F/legA 0.6215x [0.6214-0.6220]  int8 WINS
      glue fusion i8F/i8U 1.0014x [1.0011-1.0016]; multi-out i8U/i8S 1.0057x
      prepF   6.22 us  prepS   7.75 us  (prep share of i8F 3.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.081e-02 1.080e-02

  add+norm -> gate|up M=64 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 17408:s10(TM32) 17408:s10(TM32)
    i8F       140.18 us/site  inner=108   burst=  30.3 ms
    i8S       140.67 us/site  inner=107   burst=  30.1 ms
    i8U       139.65 us/site  inner=108   burst=  30.2 ms
    legA      277.09 us/site  inner=55    burst=  30.5 ms
    prepF       6.21 us/site  inner=2368  burst=  29.4 ms
    prepS       7.75 us/site  inner=1908  burst=  29.6 ms
      i8S/legA 0.5075x [0.5071-0.5080]  int8 WINS
      i8U/legA 0.5038x [0.5036-0.5043]  int8 WINS
      i8F/legA 0.5057x [0.5056-0.5059]  int8 WINS
      glue fusion i8F/i8U 1.0036x [1.0026-1.0041]; multi-out i8U/i8S 0.9929x
      prepF   6.21 us  prepS   7.75 us  (prep share of i8F 4.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.081e-02 1.080e-02

########## silu*mul -> down_proj  G=1  x64  [17408x5120xK4] ##########
    site tile s98 TILE_M=16 rows_pad=16 groups=136 rot=4 trellis=44.6 MB/pass

  silu*mul -> down_proj M=8 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s2(TM16)
    i8F        91.96 us/site  inner=82    burst=  30.2 ms
    i8S        91.11 us/site  inner=83    burst=  30.2 ms
    i8U        91.13 us/site  inner=83    burst=  30.3 ms
    legA       55.48 us/site  inner=135   burst=  30.0 ms
    prepF       7.45 us/site  inner=998   burst=  29.7 ms
    prepS       6.56 us/site  inner=1128  burst=  29.6 ms
      i8S/legA 1.6432x [1.6420-1.6442]  leg A wins
      i8U/legA 1.6436x [1.6415-1.6446]  leg A wins
      i8F/legA 1.6592x [1.6565-1.6599]  leg A wins
      glue fusion i8F/i8U 1.0093x [1.0091-1.0097]; multi-out i8U/i8S 1.0001x
      prepF   7.45 us  prepS   6.56 us  (prep share of i8F 8.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.052e-02
    site tile s94 TILE_M=32 rows_pad=32 groups=136 rot=4 trellis=44.6 MB/pass

  silu*mul -> down_proj M=8 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F       110.46 us/site  inner=68    burst=  30.0 ms
    i8S       109.58 us/site  inner=69    burst=  30.2 ms
    i8U       109.59 us/site  inner=69    burst=  30.2 ms
    legA       55.44 us/site  inner=136   burst=  30.2 ms
    prepF       7.46 us/site  inner=994   burst=  29.7 ms
    prepS       6.57 us/site  inner=1128  burst=  29.7 ms
      i8S/legA 1.9766x [1.9748-1.9782]  leg A wins
      i8U/legA 1.9771x [1.9750-1.9784]  leg A wins
      i8F/legA 1.9927x [1.9904-1.9937]  leg A wins
      glue fusion i8F/i8U 1.0080x [1.0074-1.0083]; multi-out i8U/i8S 1.0001x
      prepF   7.46 us  prepS   6.57 us  (prep share of i8F 6.8%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.052e-02
    site tile s85 TILE_M=64 rows_pad=64 groups=136 rot=4 trellis=44.6 MB/pass

  silu*mul -> down_proj M=8 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F       143.18 us/site  inner=53    burst=  30.4 ms
    i8S       142.17 us/site  inner=53    burst=  30.1 ms
    i8U       142.18 us/site  inner=53    burst=  30.1 ms
    legA       55.43 us/site  inner=136   burst=  30.2 ms
    prepF       7.69 us/site  inner=965   burst=  29.7 ms
    prepS       6.64 us/site  inner=1117  burst=  29.7 ms
      i8S/legA 2.5644x [2.5632-2.5656]  leg A wins
      i8U/legA 2.5650x [2.5635-2.5670]  leg A wins
      i8F/legA 2.5832x [2.5819-2.5835]  leg A wins
      glue fusion i8F/i8U 1.0070x [1.0064-1.0074]; multi-out i8U/i8S 1.0001x
      prepF   7.69 us  prepS   6.64 us  (prep share of i8F 5.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.052e-02

  silu*mul -> down_proj M=16 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s2(TM16)
    i8F        92.39 us/site  inner=82    burst=  30.3 ms
    i8S        91.82 us/site  inner=82    burst=  30.1 ms
    i8U        91.79 us/site  inner=82    burst=  30.1 ms
    legA       56.79 us/site  inner=132   burst=  30.0 ms
    prepF       7.44 us/site  inner=998   burst=  29.7 ms
    prepS       6.80 us/site  inner=1091  burst=  29.7 ms
      i8S/legA 1.6167x [1.6163-1.6171]  leg A wins
      i8U/legA 1.6162x [1.6157-1.6169]  leg A wins
      i8F/legA 1.6270x [1.6264-1.6273]  leg A wins
      glue fusion i8F/i8U 1.0066x [1.0063-1.0069]; multi-out i8U/i8S 0.9998x
      prepF   7.44 us  prepS   6.80 us  (prep share of i8F 8.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.056e-02

  silu*mul -> down_proj M=16 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F       110.65 us/site  inner=68    burst=  30.1 ms
    i8S       110.04 us/site  inner=69    burst=  30.4 ms
    i8U       110.04 us/site  inner=69    burst=  30.4 ms
    legA       56.83 us/site  inner=132   burst=  30.0 ms
    prepF       7.45 us/site  inner=995   burst=  29.6 ms
    prepS       6.82 us/site  inner=1086  burst=  29.6 ms
      i8S/legA 1.9364x [1.9357-1.9375]  leg A wins
      i8U/legA 1.9366x [1.9355-1.9375]  leg A wins
      i8F/legA 1.9473x [1.9465-1.9480]  leg A wins
      glue fusion i8F/i8U 1.0055x [1.0052-1.0059]; multi-out i8U/i8S 1.0000x
      prepF   7.45 us  prepS   6.82 us  (prep share of i8F 6.7%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.056e-02

  silu*mul -> down_proj M=16 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F       143.35 us/site  inner=53    burst=  30.4 ms
    i8S       142.60 us/site  inner=53    burst=  30.2 ms
    i8U       142.62 us/site  inner=53    burst=  30.2 ms
    legA       56.81 us/site  inner=132   burst=  30.0 ms
    prepF       7.68 us/site  inner=965   burst=  29.6 ms
    prepS       6.97 us/site  inner=1061  burst=  29.6 ms
      i8S/legA 2.5104x [2.5089-2.5115]  leg A wins
      i8U/legA 2.5104x [2.5095-2.5117]  leg A wins
      i8F/legA 2.5234x [2.5223-2.5254]  leg A wins
      glue fusion i8F/i8U 1.0051x [1.0048-1.0056]; multi-out i8U/i8S 1.0001x
      prepF   7.68 us  prepS   6.97 us  (prep share of i8F 5.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.056e-02

  silu*mul -> down_proj M=32 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F        92.43 us/site  inner=81    burst=  29.9 ms
    i8S        92.05 us/site  inner=82    burst=  30.2 ms
    i8U        92.07 us/site  inner=82    burst=  30.2 ms
    legA       68.34 us/site  inner=110   burst=  30.1 ms
    prepF       7.48 us/site  inner=990   burst=  29.6 ms
    prepS       7.08 us/site  inner=1042  burst=  29.5 ms
      i8S/legA 1.3472x [1.3378-1.3480]  leg A wins
      i8U/legA 1.3473x [1.3377-1.3525]  leg A wins
      i8F/legA 1.3524x [1.3449-1.3535]  leg A wins
      glue fusion i8F/i8U 1.0044x [1.0002-1.0054]; multi-out i8U/i8S 1.0000x
      prepF   7.48 us  prepS   7.08 us  (prep share of i8F 8.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.054e-02

  silu*mul -> down_proj M=32 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       115.12 us/site  inner=66    burst=  30.4 ms
    i8S       114.58 us/site  inner=66    burst=  30.2 ms
    i8U       114.57 us/site  inner=66    burst=  30.2 ms
    legA       68.49 us/site  inner=110   burst=  30.1 ms
    prepF       7.48 us/site  inner=987   burst=  29.5 ms
    prepS       7.08 us/site  inner=1042  burst=  29.5 ms
      i8S/legA 1.6728x [1.6674-1.6768]  leg A wins
      i8U/legA 1.6727x [1.6673-1.6743]  leg A wins
      i8F/legA 1.6784x [1.6753-1.6886]  leg A wins
      glue fusion i8F/i8U 1.0048x [1.0028-1.0085]; multi-out i8U/i8S 1.0000x
      prepF   7.48 us  prepS   7.08 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.054e-02

  silu*mul -> down_proj M=32 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       143.58 us/site  inner=53    burst=  30.4 ms
    i8S       143.05 us/site  inner=53    burst=  30.3 ms
    i8U       143.04 us/site  inner=53    burst=  30.3 ms
    legA       68.54 us/site  inner=110   burst=  30.2 ms
    prepF       7.56 us/site  inner=977   burst=  29.6 ms
    prepS       7.14 us/site  inner=1034  burst=  29.5 ms
      i8S/legA 2.0870x [2.0864-2.0956]  leg A wins
      i8U/legA 2.0867x [2.0860-2.0878]  leg A wins
      i8F/legA 2.0950x [2.0939-2.0960]  leg A wins
      glue fusion i8F/i8U 1.0038x [1.0037-1.0043]; multi-out i8U/i8S 0.9999x
      prepF   7.56 us  prepS   7.14 us  (prep share of i8F 5.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.054e-02

  silu*mul -> down_proj M=36 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F        92.82 us/site  inner=81    burst=  30.1 ms
    i8S        92.89 us/site  inner=81    burst=  30.1 ms
    i8U        92.68 us/site  inner=81    burst=  30.0 ms
    legA      130.64 us/site  inner=58    burst=  30.3 ms
    prepF       7.51 us/site  inner=981   burst=  29.5 ms
    prepS       7.37 us/site  inner=999   burst=  29.4 ms
      i8S/legA 0.7104x [0.7018-0.7232]  int8 WINS
      i8U/legA 0.7096x [0.7042-0.7152]  int8 WINS
      i8F/legA 0.7105x [0.7058-0.7131]  int8 WINS
      glue fusion i8F/i8U 1.0012x [0.9970-1.0034]; multi-out i8U/i8S 0.9980x
      prepF   7.51 us  prepS   7.37 us  (prep share of i8F 8.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.049e-02

  silu*mul -> down_proj M=36 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       115.16 us/site  inner=66    burst=  30.4 ms
    i8S       114.96 us/site  inner=66    burst=  30.4 ms
    i8U       114.99 us/site  inner=66    burst=  30.4 ms
    legA      130.34 us/site  inner=58    burst=  30.2 ms
    prepF       7.56 us/site  inner=977   burst=  29.6 ms
    prepS       7.38 us/site  inner=1002  burst=  29.6 ms
      i8S/legA 0.8821x [0.8774-0.8826]  int8 WINS
      i8U/legA 0.8823x [0.8768-0.8829]  int8 WINS
      i8F/legA 0.8836x [0.8807-0.8845]  int8 WINS
      glue fusion i8F/i8U 1.0015x [1.0012-1.0045]; multi-out i8U/i8S 1.0003x
      prepF   7.56 us  prepS   7.38 us  (prep share of i8F 6.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.049e-02

  silu*mul -> down_proj M=36 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       143.67 us/site  inner=53    burst=  30.5 ms
    i8S       143.44 us/site  inner=53    burst=  30.4 ms
    i8U       143.48 us/site  inner=53    burst=  30.4 ms
    legA      130.46 us/site  inner=58    burst=  30.3 ms
    prepF       7.56 us/site  inner=977   burst=  29.6 ms
    prepS       7.38 us/site  inner=998   burst=  29.5 ms
      i8S/legA 1.0996x [1.0991-1.1001]  leg A wins
      i8U/legA 1.0998x [1.0991-1.1006]  leg A wins
      i8F/legA 1.1015x [1.1005-1.1021]  leg A wins
      glue fusion i8F/i8U 1.0014x [1.0011-1.0015]; multi-out i8U/i8S 1.0002x
      prepF   7.56 us  prepS   7.38 us  (prep share of i8F 5.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.049e-02

  silu*mul -> down_proj M=48 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F        93.10 us/site  inner=81    burst=  30.2 ms
    i8S        92.96 us/site  inner=81    burst=  30.1 ms
    i8U        93.00 us/site  inner=81    burst=  30.1 ms
    legA      132.51 us/site  inner=57    burst=  30.2 ms
    prepF       7.55 us/site  inner=977   burst=  29.5 ms
    prepS       7.65 us/site  inner=973   burst=  29.8 ms
      i8S/legA 0.7001x [0.6958-0.7055]  int8 WINS
      i8U/legA 0.7031x [0.6949-0.7054]  int8 WINS
      i8F/legA 0.7008x [0.6983-0.7067]  int8 WINS
      glue fusion i8F/i8U 0.9995x [0.9965-1.0054]; multi-out i8U/i8S 0.9999x
      prepF   7.55 us  prepS   7.65 us  (prep share of i8F 8.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.059e-02

  silu*mul -> down_proj M=48 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       115.12 us/site  inner=66    burst=  30.4 ms
    i8S       115.17 us/site  inner=66    burst=  30.4 ms
    i8U       115.17 us/site  inner=65    burst=  29.9 ms
    legA      131.61 us/site  inner=57    burst=  30.0 ms
    prepF       7.56 us/site  inner=978   burst=  29.6 ms
    prepS       7.60 us/site  inner=969   burst=  29.4 ms
      i8S/legA 0.8751x [0.8734-0.8776]  int8 WINS
      i8U/legA 0.8751x [0.8744-0.8779]  int8 WINS
      i8F/legA 0.8748x [0.8739-0.8760]  int8 WINS
      glue fusion i8F/i8U 0.9996x [0.9978-0.9997]; multi-out i8U/i8S 1.0000x
      prepF   7.56 us  prepS   7.60 us  (prep share of i8F 6.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.059e-02

  silu*mul -> down_proj M=48 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       143.86 us/site  inner=53    burst=  30.5 ms
    i8S       143.91 us/site  inner=53    burst=  30.5 ms
    i8U       143.90 us/site  inner=53    burst=  30.5 ms
    legA      131.84 us/site  inner=57    burst=  30.1 ms
    prepF       7.57 us/site  inner=977   burst=  29.6 ms
    prepS       7.61 us/site  inner=969   burst=  29.5 ms
      i8S/legA 1.0919x [1.0913-1.0922]  leg A wins
      i8U/legA 1.0916x [1.0913-1.0921]  leg A wins
      i8F/legA 1.0914x [1.0909-1.0918]  leg A wins
      glue fusion i8F/i8U 0.9997x [0.9996-0.9999]; multi-out i8U/i8S 1.0000x
      prepF   7.57 us  prepS   7.61 us  (prep share of i8F 5.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.059e-02

  silu*mul -> down_proj M=64 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F       141.86 us/site  inner=53    burst=  30.1 ms
    i8S       142.23 us/site  inner=53    burst=  30.2 ms
    i8U       142.18 us/site  inner=53    burst=  30.1 ms
    legA      134.09 us/site  inner=56    burst=  30.0 ms
    prepF       7.64 us/site  inner=965   burst=  29.5 ms
    prepS       7.95 us/site  inner=923   burst=  29.3 ms
      i8S/legA 1.0605x [1.0591-1.0622]  leg A wins
      i8U/legA 1.0603x [1.0602-1.0614]  leg A wins
      i8F/legA 1.0575x [1.0569-1.0612]  leg A wins
      glue fusion i8F/i8U 0.9974x [0.9962-0.9999]; multi-out i8U/i8S 0.9998x
      prepF   7.64 us  prepS   7.95 us  (prep share of i8F 5.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.060e-02

  silu*mul -> down_proj M=64 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       115.17 us/site  inner=66    burst=  30.4 ms
    i8S       115.55 us/site  inner=65    burst=  30.0 ms
    i8U       115.56 us/site  inner=65    burst=  30.0 ms
    legA      134.03 us/site  inner=56    burst=  30.0 ms
    prepF       7.64 us/site  inner=965   burst=  29.5 ms
    prepS       7.95 us/site  inner=923   burst=  29.3 ms
      i8S/legA 0.8621x [0.8620-0.8624]  int8 WINS
      i8U/legA 0.8622x [0.8619-0.8625]  int8 WINS
      i8F/legA 0.8593x [0.8588-0.8595]  int8 WINS
      glue fusion i8F/i8U 0.9965x [0.9964-0.9967]; multi-out i8U/i8S 1.0001x
      prepF   7.64 us  prepS   7.95 us  (prep share of i8F 6.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.060e-02

  silu*mul -> down_proj M=64 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F       144.08 us/site  inner=53    burst=  30.5 ms
    i8S       144.43 us/site  inner=52    burst=  30.0 ms
    i8U       144.43 us/site  inner=52    burst=  30.0 ms
    legA      134.21 us/site  inner=56    burst=  30.1 ms
    prepF       7.65 us/site  inner=965   burst=  29.5 ms
    prepS       7.96 us/site  inner=927   burst=  29.5 ms
      i8S/legA 1.0763x [1.0714-1.0778]  leg A wins
      i8U/legA 1.0761x [1.0733-1.0790]  leg A wins
      i8F/legA 1.0736x [1.0688-1.0773]  leg A wins
      glue fusion i8F/i8U 0.9976x [0.9958-0.9984]; multi-out i8U/i8S 1.0000x
      prepF   7.65 us  prepS   7.96 us  (prep share of i8F 5.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.060e-02

########## gate*sig -> o_proj  G=1  x16  [6144x5120xK4] ##########
    site tile s98 TILE_M=16 rows_pad=16 groups=48 rot=10 trellis=15.7 MB/pass

  gate*sig -> o_proj M=8 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s2(TM16)
    i8F        34.39 us/site  inner=88    burst=  30.3 ms
    i8S        34.88 us/site  inner=86    burst=  30.0 ms
    i8U        34.88 us/site  inner=86    burst=  30.0 ms
    legA       23.61 us/site  inner=127   burst=  30.0 ms
    prepF       3.09 us/site  inner=965   burst=  29.8 ms
    prepS       3.60 us/site  inner=829   burst=  29.8 ms
      i8S/legA 1.4773x [1.4770-1.4775]  leg A wins
      i8U/legA 1.4774x [1.4770-1.4775]  leg A wins
      i8F/legA 1.4563x [1.4560-1.4564]  leg A wins
      glue fusion i8F/i8U 0.9858x [0.9856-0.9859]; multi-out i8U/i8S 1.0000x
      prepF   3.09 us  prepS   3.60 us  (prep share of i8F 9.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.088e-02
    site tile s94 TILE_M=32 rows_pad=32 groups=48 rot=10 trellis=15.7 MB/pass

  gate*sig -> o_proj M=8 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        40.81 us/site  inner=74    burst=  30.2 ms
    i8S        41.30 us/site  inner=73    burst=  30.1 ms
    i8U        41.28 us/site  inner=73    burst=  30.1 ms
    legA       23.61 us/site  inner=127   burst=  30.0 ms
    prepF       3.09 us/site  inner=957   burst=  29.5 ms
    prepS       3.60 us/site  inner=829   burst=  29.9 ms
      i8S/legA 1.7490x [1.7479-1.7495]  leg A wins
      i8U/legA 1.7481x [1.7479-1.7487]  leg A wins
      i8F/legA 1.7282x [1.7271-1.7285]  leg A wins
      glue fusion i8F/i8U 0.9885x [0.9881-0.9886]; multi-out i8U/i8S 0.9996x
      prepF   3.09 us  prepS   3.60 us  (prep share of i8F 7.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.088e-02
    site tile s85 TILE_M=64 rows_pad=64 groups=48 rot=10 trellis=15.7 MB/pass

  gate*sig -> o_proj M=8 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        52.42 us/site  inner=58    burst=  30.4 ms
    i8S        52.80 us/site  inner=57    burst=  30.1 ms
    i8U        52.79 us/site  inner=57    burst=  30.1 ms
    legA       23.62 us/site  inner=128   burst=  30.2 ms
    prepF       3.18 us/site  inner=938   burst=  29.8 ms
    prepS       3.61 us/site  inner=826   burst=  29.8 ms
      i8S/legA 2.2353x [2.2335-2.2368]  leg A wins
      i8U/legA 2.2351x [2.2340-2.2362]  leg A wins
      i8F/legA 2.2195x [2.2182-2.2214]  leg A wins
      glue fusion i8F/i8U 0.9930x [0.9928-0.9934]; multi-out i8U/i8S 0.9998x
      prepF   3.18 us  prepS   3.61 us  (prep share of i8F 6.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.088e-02

  gate*sig -> o_proj M=16 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s2(TM16)
    i8F        34.67 us/site  inner=87    burst=  30.2 ms
    i8S        35.13 us/site  inner=86    burst=  30.2 ms
    i8U        35.13 us/site  inner=86    burst=  30.2 ms
    legA       24.73 us/site  inner=122   burst=  30.2 ms
    prepF       3.15 us/site  inner=942   burst=  29.7 ms
    prepS       3.67 us/site  inner=812   burst=  29.8 ms
      i8S/legA 1.4204x [1.4199-1.4209]  leg A wins
      i8U/legA 1.4202x [1.4198-1.4207]  leg A wins
      i8F/legA 1.4015x [1.4008-1.4021]  leg A wins
      glue fusion i8F/i8U 0.9868x [0.9864-0.9871]; multi-out i8U/i8S 0.9999x
      prepF   3.15 us  prepS   3.67 us  (prep share of i8F 9.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.098e-02

  gate*sig -> o_proj M=16 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        41.01 us/site  inner=74    burst=  30.3 ms
    i8S        41.47 us/site  inner=73    burst=  30.3 ms
    i8U        41.47 us/site  inner=73    burst=  30.3 ms
    legA       24.74 us/site  inner=122   burst=  30.2 ms
    prepF       3.15 us/site  inner=942   burst=  29.7 ms
    prepS       3.68 us/site  inner=811   burst=  29.8 ms
      i8S/legA 1.6762x [1.6760-1.6768]  leg A wins
      i8U/legA 1.6761x [1.6754-1.6764]  leg A wins
      i8F/legA 1.6576x [1.6569-1.6580]  leg A wins
      glue fusion i8F/i8U 0.9890x [0.9888-0.9892]; multi-out i8U/i8S 0.9998x
      prepF   3.15 us  prepS   3.68 us  (prep share of i8F 7.7%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.098e-02

  gate*sig -> o_proj M=16 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        52.56 us/site  inner=58    burst=  30.5 ms
    i8S        53.00 us/site  inner=57    burst=  30.2 ms
    i8U        53.00 us/site  inner=57    burst=  30.2 ms
    legA       24.76 us/site  inner=122   burst=  30.2 ms
    prepF       3.21 us/site  inner=924   burst=  29.6 ms
    prepS       3.69 us/site  inner=806   burst=  29.8 ms
      i8S/legA 2.1404x [2.1235-2.1427]  leg A wins
      i8U/legA 2.1402x [2.1239-2.1428]  leg A wins
      i8F/legA 2.1227x [2.1066-2.1251]  leg A wins
      glue fusion i8F/i8U 0.9918x [0.9914-0.9920]; multi-out i8U/i8S 1.0000x
      prepF   3.21 us  prepS   3.69 us  (prep share of i8F 6.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.098e-02

  gate*sig -> o_proj M=32 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F        34.76 us/site  inner=87    burst=  30.2 ms
    i8S        35.35 us/site  inner=85    burst=  30.0 ms
    i8U        35.36 us/site  inner=85    burst=  30.1 ms
    legA       30.05 us/site  inner=100   burst=  30.1 ms
    prepF       3.19 us/site  inner=927   burst=  29.5 ms
    prepS       3.77 us/site  inner=787   burst=  29.7 ms
      i8S/legA 1.1765x [1.1758-1.1767]  leg A wins
      i8U/legA 1.1768x [1.1764-1.1770]  leg A wins
      i8F/legA 1.1568x [1.1562-1.1578]  leg A wins
      glue fusion i8F/i8U 0.9831x [0.9825-0.9838]; multi-out i8U/i8S 1.0002x
      prepF   3.19 us  prepS   3.77 us  (prep share of i8F 9.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.105e-02

  gate*sig -> o_proj M=32 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.86 us/site  inner=70    burst=  30.0 ms
    i8S        43.41 us/site  inner=69    burst=  30.0 ms
    i8U        43.41 us/site  inner=70    burst=  30.4 ms
    legA       30.04 us/site  inner=100   burst=  30.0 ms
    prepF       3.18 us/site  inner=934   burst=  29.7 ms
    prepS       3.76 us/site  inner=792   burst=  29.8 ms
      i8S/legA 1.4451x [1.4434-1.4465]  leg A wins
      i8U/legA 1.4449x [1.4436-1.4456]  leg A wins
      i8F/legA 1.4267x [1.4252-1.4273]  leg A wins
      glue fusion i8F/i8U 0.9874x [0.9873-0.9876]; multi-out i8U/i8S 1.0000x
      prepF   3.18 us  prepS   3.76 us  (prep share of i8F 7.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.105e-02

  gate*sig -> o_proj M=32 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        52.85 us/site  inner=57    burst=  30.1 ms
    i8S        53.43 us/site  inner=57    burst=  30.5 ms
    i8U        53.43 us/site  inner=57    burst=  30.5 ms
    legA       30.05 us/site  inner=100   burst=  30.0 ms
    prepF       3.20 us/site  inner=931   burst=  29.8 ms
    prepS       3.80 us/site  inner=784   burst=  29.8 ms
      i8S/legA 1.7780x [1.7757-1.7798]  leg A wins
      i8U/legA 1.7785x [1.7753-1.7794]  leg A wins
      i8F/legA 1.7587x [1.7561-1.7596]  leg A wins
      glue fusion i8F/i8U 0.9892x [0.9887-0.9897]; multi-out i8U/i8S 1.0001x
      prepF   3.20 us  prepS   3.80 us  (prep share of i8F 6.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.105e-02

  gate*sig -> o_proj M=36 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F        34.84 us/site  inner=87    burst=  30.3 ms
    i8S        35.43 us/site  inner=85    burst=  30.1 ms
    i8U        35.41 us/site  inner=85    burst=  30.1 ms
    legA       54.42 us/site  inner=56    burst=  30.5 ms
    prepF       3.23 us/site  inner=923   burst=  29.8 ms
    prepS       3.82 us/site  inner=779   burst=  29.8 ms
      i8S/legA 0.6508x [0.6505-0.6514]  int8 WINS
      i8U/legA 0.6507x [0.6501-0.6518]  int8 WINS
      i8F/legA 0.6402x [0.6397-0.6413]  int8 WINS
      glue fusion i8F/i8U 0.9839x [0.9831-0.9844]; multi-out i8U/i8S 0.9992x
      prepF   3.23 us  prepS   3.82 us  (prep share of i8F 9.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.097e-02

  gate*sig -> o_proj M=36 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.89 us/site  inner=70    burst=  30.0 ms
    i8S        43.49 us/site  inner=69    burst=  30.0 ms
    i8U        43.46 us/site  inner=69    burst=  30.0 ms
    legA       54.43 us/site  inner=55    burst=  29.9 ms
    prepF       3.23 us/site  inner=923   burst=  29.8 ms
    prepS       3.84 us/site  inner=779   burst=  29.9 ms
      i8S/legA 0.7987x [0.7982-0.8036]  int8 WINS
      i8U/legA 0.7985x [0.7979-0.8011]  int8 WINS
      i8F/legA 0.7881x [0.7874-0.7911]  int8 WINS
      glue fusion i8F/i8U 0.9872x [0.9864-0.9876]; multi-out i8U/i8S 0.9995x
      prepF   3.23 us  prepS   3.84 us  (prep share of i8F 7.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.097e-02

  gate*sig -> o_proj M=36 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        53.04 us/site  inner=57    burst=  30.2 ms
    i8S        53.61 us/site  inner=56    burst=  30.0 ms
    i8U        53.59 us/site  inner=56    burst=  30.0 ms
    legA       54.41 us/site  inner=56    burst=  30.5 ms
    prepF       3.24 us/site  inner=916   burst=  29.7 ms
    prepS       3.84 us/site  inner=774   burst=  29.7 ms
      i8S/legA 0.9854x [0.9846-0.9873]  int8 WINS
      i8U/legA 0.9852x [0.9842-0.9868]  int8 WINS
      i8F/legA 0.9750x [0.9739-0.9766]  int8 WINS
      glue fusion i8F/i8U 0.9896x [0.9893-0.9898]; multi-out i8U/i8S 0.9997x
      prepF   3.24 us  prepS   3.84 us  (prep share of i8F 6.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.097e-02

  gate*sig -> o_proj M=48 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F        34.90 us/site  inner=87    burst=  30.4 ms
    i8S        35.55 us/site  inner=85    burst=  30.2 ms
    i8U        35.54 us/site  inner=85    burst=  30.2 ms
    legA       55.61 us/site  inner=55    burst=  30.6 ms
    prepF       3.28 us/site  inner=909   burst=  29.8 ms
    prepS       3.95 us/site  inner=749   burst=  29.6 ms
      i8S/legA 0.6390x [0.6380-0.6402]  int8 WINS
      i8U/legA 0.6391x [0.6373-0.6404]  int8 WINS
      i8F/legA 0.6272x [0.6260-0.6282]  int8 WINS
      glue fusion i8F/i8U 0.9813x [0.9796-0.9825]; multi-out i8U/i8S 1.0002x
      prepF   3.28 us  prepS   3.95 us  (prep share of i8F 9.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.102e-02

  gate*sig -> o_proj M=48 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.97 us/site  inner=70    burst=  30.1 ms
    i8S        43.62 us/site  inner=69    burst=  30.1 ms
    i8U        43.64 us/site  inner=69    burst=  30.1 ms
    legA       55.56 us/site  inner=55    burst=  30.6 ms
    prepF       3.28 us/site  inner=909   burst=  29.8 ms
    prepS       3.97 us/site  inner=752   burst=  29.9 ms
      i8S/legA 0.7852x [0.7848-0.7857]  int8 WINS
      i8U/legA 0.7853x [0.7851-0.7858]  int8 WINS
      i8F/legA 0.7734x [0.7731-0.7738]  int8 WINS
      glue fusion i8F/i8U 0.9848x [0.9846-0.9849]; multi-out i8U/i8S 1.0002x
      prepF   3.28 us  prepS   3.97 us  (prep share of i8F 7.6%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.102e-02

  gate*sig -> o_proj M=48 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        53.26 us/site  inner=57    burst=  30.4 ms
    i8S        53.91 us/site  inner=56    burst=  30.2 ms
    i8U        53.90 us/site  inner=56    burst=  30.2 ms
    legA       55.64 us/site  inner=55    burst=  30.6 ms
    prepF       3.28 us/site  inner=906   burst=  29.7 ms
    prepS       3.97 us/site  inner=750   burst=  29.8 ms
      i8S/legA 0.9688x [0.9647-0.9714]  int8 WINS
      i8U/legA 0.9688x [0.9650-0.9715]  int8 WINS
      i8F/legA 0.9572x [0.9562-0.9600]  int8 WINS
      glue fusion i8F/i8U 0.9881x [0.9878-0.9909]; multi-out i8U/i8S 1.0001x
      prepF   3.28 us  prepS   3.97 us  (prep share of i8F 6.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.102e-02

  gate*sig -> o_proj M=64 tile s98  reps=7 burst>=30 ms  cell 1.6 s
    legA pins 5120:s5(TM32)
    i8F        52.31 us/site  inner=58    burst=  30.3 ms
    i8S        53.19 us/site  inner=57    burst=  30.3 ms
    i8U        53.19 us/site  inner=57    burst=  30.3 ms
    legA       57.50 us/site  inner=53    burst=  30.5 ms
    prepF       3.31 us/site  inner=895   burst=  29.6 ms
    prepS       4.18 us/site  inner=709   burst=  29.6 ms
      i8S/legA 0.9250x [0.9245-0.9253]  int8 WINS
      i8U/legA 0.9250x [0.9245-0.9329]  int8 WINS
      i8F/legA 0.9097x [0.9089-0.9100]  int8 WINS
      glue fusion i8F/i8U 0.9833x [0.9748-0.9836]; multi-out i8U/i8S 1.0002x
      prepF   3.31 us  prepS   4.18 us  (prep share of i8F 6.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.097e-02

  gate*sig -> o_proj M=64 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.99 us/site  inner=70    burst=  30.1 ms
    i8S        43.86 us/site  inner=69    burst=  30.3 ms
    i8U        43.86 us/site  inner=69    burst=  30.3 ms
    legA       57.51 us/site  inner=53    burst=  30.5 ms
    prepF       3.30 us/site  inner=898   burst=  29.7 ms
    prepS       4.18 us/site  inner=711   burst=  29.7 ms
      i8S/legA 0.7626x [0.7624-0.7628]  int8 WINS
      i8U/legA 0.7626x [0.7625-0.7630]  int8 WINS
      i8F/legA 0.7476x [0.7473-0.7477]  int8 WINS
      glue fusion i8F/i8U 0.9801x [0.9796-0.9805]; multi-out i8U/i8S 1.0001x
      prepF   3.30 us  prepS   4.18 us  (prep share of i8F 7.7%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.097e-02

  gate*sig -> o_proj M=64 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        53.54 us/site  inner=57    burst=  30.5 ms
    i8S        54.39 us/site  inner=56    burst=  30.5 ms
    i8U        54.39 us/site  inner=56    burst=  30.5 ms
    legA       57.53 us/site  inner=53    burst=  30.5 ms
    prepF       3.30 us/site  inner=898   burst=  29.7 ms
    prepS       4.18 us/site  inner=711   burst=  29.7 ms
      i8S/legA 0.9455x [0.9453-0.9460]  int8 WINS
      i8U/legA 0.9455x [0.9452-0.9460]  int8 WINS
      i8F/legA 0.9309x [0.9302-0.9312]  int8 WINS
      glue fusion i8F/i8U 0.9843x [0.9841-0.9846]; multi-out i8U/i8S 1.0000x
      prepF   3.30 us  prepS   4.18 us  (prep share of i8F 6.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.097e-02

########## (none) -> out_proj  G=1  x48  [6144x5120xK4] ##########
    site tile s98 TILE_M=16 rows_pad=16 groups=48 rot=10 trellis=15.7 MB/pass

  (none) -> out_proj M=8 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        33.91 us/site  inner=89    burst=  30.2 ms
    i8S        33.92 us/site  inner=89    burst=  30.2 ms
    i8U        33.92 us/site  inner=89    burst=  30.2 ms
    legA       22.65 us/site  inner=133   burst=  30.1 ms
    prepF       2.63 us/site  inner=1129  burst=  29.7 ms
    prepS       2.63 us/site  inner=1127  burst=  29.7 ms
      i8S/legA 1.4975x [1.4969-1.4981]  leg A wins
      i8U/legA 1.4973x [1.4968-1.4980]  leg A wins
      i8F/legA 1.4973x [1.4967-1.4978]  leg A wins
      glue fusion i8F/i8U 0.9999x [0.9997-1.0001]; multi-out i8U/i8S 1.0000x
      prepF   2.63 us  prepS   2.63 us  (prep share of i8F 7.8%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.120e-02
    site tile s94 TILE_M=32 rows_pad=32 groups=48 rot=10 trellis=15.7 MB/pass

  (none) -> out_proj M=8 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        40.32 us/site  inner=75    burst=  30.2 ms
    i8S        40.32 us/site  inner=75    burst=  30.2 ms
    i8U        40.31 us/site  inner=75    burst=  30.2 ms
    legA       22.65 us/site  inner=133   burst=  30.1 ms
    prepF       2.63 us/site  inner=1133  burst=  29.8 ms
    prepS       2.63 us/site  inner=1133  burst=  29.8 ms
      i8S/legA 1.7802x [1.7797-1.7807]  leg A wins
      i8U/legA 1.7799x [1.7797-1.7804]  leg A wins
      i8F/legA 1.7801x [1.7796-1.7807]  leg A wins
      glue fusion i8F/i8U 1.0001x [0.9999-1.0002]; multi-out i8U/i8S 0.9998x
      prepF   2.63 us  prepS   2.63 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.120e-02
    site tile s85 TILE_M=64 rows_pad=64 groups=48 rot=10 trellis=15.7 MB/pass

  (none) -> out_proj M=8 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        51.83 us/site  inner=58    burst=  30.1 ms
    i8S        51.82 us/site  inner=58    burst=  30.1 ms
    i8U        51.82 us/site  inner=58    burst=  30.1 ms
    legA       22.65 us/site  inner=133   burst=  30.1 ms
    prepF       2.62 us/site  inner=1138  burst=  29.8 ms
    prepS       2.61 us/site  inner=1138  burst=  29.7 ms
      i8S/legA 2.2875x [2.2866-2.3010]  leg A wins
      i8U/legA 2.2873x [2.2867-2.2889]  leg A wins
      i8F/legA 2.2880x [2.2869-2.2896]  leg A wins
      glue fusion i8F/i8U 1.0002x [0.9999-1.0003]; multi-out i8U/i8S 1.0000x
      prepF   2.62 us  prepS   2.61 us  (prep share of i8F 5.0%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.120e-02

  (none) -> out_proj M=16 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        33.96 us/site  inner=89    burst=  30.2 ms
    i8S        33.96 us/site  inner=89    burst=  30.2 ms
    i8U        33.96 us/site  inner=89    burst=  30.2 ms
    legA       23.62 us/site  inner=127   burst=  30.0 ms
    prepF       2.64 us/site  inner=1127  burst=  29.7 ms
    prepS       2.64 us/site  inner=1130  burst=  29.8 ms
      i8S/legA 1.4376x [1.4369-1.4381]  leg A wins
      i8U/legA 1.4374x [1.4367-1.4379]  leg A wins
      i8F/legA 1.4376x [1.4372-1.4380]  leg A wins
      glue fusion i8F/i8U 1.0000x [0.9999-1.0004]; multi-out i8U/i8S 0.9999x
      prepF   2.64 us  prepS   2.64 us  (prep share of i8F 7.8%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.115e-02

  (none) -> out_proj M=16 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        40.46 us/site  inner=75    burst=  30.3 ms
    i8S        40.45 us/site  inner=75    burst=  30.3 ms
    i8U        40.45 us/site  inner=75    burst=  30.3 ms
    legA       23.63 us/site  inner=127   burst=  30.0 ms
    prepF       2.64 us/site  inner=1128  burst=  29.8 ms
    prepS       2.64 us/site  inner=1127  burst=  29.7 ms
      i8S/legA 1.7123x [1.7116-1.7129]  leg A wins
      i8U/legA 1.7124x [1.7115-1.7130]  leg A wins
      i8F/legA 1.7122x [1.7119-1.7129]  leg A wins
      glue fusion i8F/i8U 1.0000x [0.9997-1.0002]; multi-out i8U/i8S 1.0000x
      prepF   2.64 us  prepS   2.64 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.115e-02

  (none) -> out_proj M=16 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s2(TM16)
    i8F        52.01 us/site  inner=58    burst=  30.2 ms
    i8S        52.01 us/site  inner=58    burst=  30.2 ms
    i8U        52.00 us/site  inner=58    burst=  30.2 ms
    legA       23.64 us/site  inner=127   burst=  30.0 ms
    prepF       2.66 us/site  inner=1117  burst=  29.7 ms
    prepS       2.66 us/site  inner=1117  burst=  29.7 ms
      i8S/legA 2.2003x [2.1988-2.2020]  leg A wins
      i8U/legA 2.1997x [2.1988-2.2017]  leg A wins
      i8F/legA 2.2001x [2.1988-2.2021]  leg A wins
      glue fusion i8F/i8U 1.0002x [0.9999-1.0004]; multi-out i8U/i8S 0.9999x
      prepF   2.66 us  prepS   2.66 us  (prep share of i8F 5.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.115e-02

  (none) -> out_proj M=32 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        34.22 us/site  inner=88    burst=  30.1 ms
    i8S        34.23 us/site  inner=88    burst=  30.1 ms
    i8U        34.22 us/site  inner=88    burst=  30.1 ms
    legA       28.91 us/site  inner=104   burst=  30.1 ms
    prepF       2.68 us/site  inner=1111  burst=  29.7 ms
    prepS       2.68 us/site  inner=1111  burst=  29.7 ms
      i8S/legA 1.1841x [1.1829-1.1843]  leg A wins
      i8U/legA 1.1836x [1.1825-1.1843]  leg A wins
      i8F/legA 1.1839x [1.1829-1.1840]  leg A wins
      glue fusion i8F/i8U 1.0001x [0.9997-1.0003]; multi-out i8U/i8S 0.9998x
      prepF   2.68 us  prepS   2.68 us  (prep share of i8F 7.8%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.106e-02

  (none) -> out_proj M=32 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.34 us/site  inner=71    burst=  30.1 ms
    i8S        42.34 us/site  inner=71    burst=  30.1 ms
    i8U        42.34 us/site  inner=71    burst=  30.1 ms
    legA       28.89 us/site  inner=104   burst=  30.0 ms
    prepF       2.68 us/site  inner=1111  burst=  29.7 ms
    prepS       2.68 us/site  inner=1112  burst=  29.8 ms
      i8S/legA 1.4656x [1.4650-1.4731]  leg A wins
      i8U/legA 1.4656x [1.4652-1.4663]  leg A wins
      i8F/legA 1.4654x [1.4650-1.4662]  leg A wins
      glue fusion i8F/i8U 0.9999x [0.9997-0.9999]; multi-out i8U/i8S 1.0001x
      prepF   2.68 us  prepS   2.68 us  (prep share of i8F 6.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.106e-02

  (none) -> out_proj M=32 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        52.37 us/site  inner=58    burst=  30.4 ms
    i8S        52.36 us/site  inner=58    burst=  30.4 ms
    i8U        52.36 us/site  inner=58    burst=  30.4 ms
    legA       28.91 us/site  inner=104   burst=  30.1 ms
    prepF       2.69 us/site  inner=1106  burst=  29.8 ms
    prepS       2.68 us/site  inner=1111  burst=  29.8 ms
      i8S/legA 1.8110x [1.8073-1.8127]  leg A wins
      i8U/legA 1.8109x [1.8077-1.8122]  leg A wins
      i8F/legA 1.8110x [1.8087-1.8120]  leg A wins
      glue fusion i8F/i8U 1.0002x [0.9997-1.0006]; multi-out i8U/i8S 1.0000x
      prepF   2.69 us  prepS   2.68 us  (prep share of i8F 5.1%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.106e-02

  (none) -> out_proj M=36 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        34.26 us/site  inner=88    burst=  30.1 ms
    i8S        34.21 us/site  inner=88    burst=  30.1 ms
    i8U        34.20 us/site  inner=88    burst=  30.1 ms
    legA       53.17 us/site  inner=57    burst=  30.3 ms
    prepF       2.69 us/site  inner=1101  burst=  29.7 ms
    prepS       2.69 us/site  inner=1106  burst=  29.8 ms
      i8S/legA 0.6439x [0.6415-0.6447]  int8 WINS
      i8U/legA 0.6441x [0.6403-0.6449]  int8 WINS
      i8F/legA 0.6442x [0.6439-0.6455]  int8 WINS
      glue fusion i8F/i8U 1.0006x [0.9995-1.0057]; multi-out i8U/i8S 0.9999x
      prepF   2.69 us  prepS   2.69 us  (prep share of i8F 7.9%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.105e-02

  (none) -> out_proj M=36 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.35 us/site  inner=71    burst=  30.1 ms
    i8S        42.35 us/site  inner=71    burst=  30.1 ms
    i8U        42.35 us/site  inner=71    burst=  30.1 ms
    legA       53.31 us/site  inner=57    burst=  30.4 ms
    prepF       2.71 us/site  inner=1101  burst=  29.8 ms
    prepS       2.71 us/site  inner=1098  burst=  29.7 ms
      i8S/legA 0.7944x [0.7937-0.7950]  int8 WINS
      i8U/legA 0.7945x [0.7944-0.7950]  int8 WINS
      i8F/legA 0.7945x [0.7937-0.7948]  int8 WINS
      glue fusion i8F/i8U 0.9998x [0.9987-1.0003]; multi-out i8U/i8S 1.0001x
      prepF   2.71 us  prepS   2.71 us  (prep share of i8F 6.4%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.105e-02

  (none) -> out_proj M=36 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        52.47 us/site  inner=58    burst=  30.4 ms
    i8S        52.47 us/site  inner=58    burst=  30.4 ms
    i8U        52.47 us/site  inner=58    burst=  30.4 ms
    legA       53.17 us/site  inner=57    burst=  30.3 ms
    prepF       2.71 us/site  inner=1101  burst=  29.8 ms
    prepS       2.70 us/site  inner=1106  burst=  29.9 ms
      i8S/legA 0.9867x [0.9841-0.9877]  int8 WINS
      i8U/legA 0.9868x [0.9839-0.9877]  int8 WINS
      i8F/legA 0.9870x [0.9840-0.9876]  int8 WINS
      glue fusion i8F/i8U 1.0001x [0.9999-1.0004]; multi-out i8U/i8S 1.0000x
      prepF   2.71 us  prepS   2.70 us  (prep share of i8F 5.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.105e-02

  (none) -> out_proj M=48 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        34.33 us/site  inner=88    burst=  30.2 ms
    i8S        34.33 us/site  inner=88    burst=  30.2 ms
    i8U        34.31 us/site  inner=88    burst=  30.2 ms
    legA       54.51 us/site  inner=56    burst=  30.5 ms
    prepF       2.73 us/site  inner=1086  burst=  29.6 ms
    prepS       2.73 us/site  inner=1091  burst=  29.8 ms
      i8S/legA 0.6288x [0.6279-0.6309]  int8 WINS
      i8U/legA 0.6299x [0.6266-0.6308]  int8 WINS
      i8F/legA 0.6298x [0.6278-0.6330]  int8 WINS
      glue fusion i8F/i8U 1.0006x [0.9995-1.0041]; multi-out i8U/i8S 0.9998x
      prepF   2.73 us  prepS   2.73 us  (prep share of i8F 7.9%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.102e-02

  (none) -> out_proj M=48 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.50 us/site  inner=72    burst=  30.6 ms
    i8S        42.51 us/site  inner=72    burst=  30.6 ms
    i8U        42.51 us/site  inner=72    burst=  30.6 ms
    legA       54.58 us/site  inner=56    burst=  30.6 ms
    prepF       2.74 us/site  inner=1091  burst=  29.9 ms
    prepS       2.74 us/site  inner=1096  burst=  30.1 ms
      i8S/legA 0.7790x [0.7784-0.7820]  int8 WINS
      i8U/legA 0.7788x [0.7741-0.7792]  int8 WINS
      i8F/legA 0.7788x [0.7768-0.7792]  int8 WINS
      glue fusion i8F/i8U 1.0000x [0.9998-1.0034]; multi-out i8U/i8S 1.0000x
      prepF   2.74 us  prepS   2.74 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.102e-02

  (none) -> out_proj M=48 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        52.71 us/site  inner=57    burst=  30.0 ms
    i8S        52.71 us/site  inner=57    burst=  30.0 ms
    i8U        52.71 us/site  inner=57    burst=  30.0 ms
    legA       54.59 us/site  inner=55    burst=  30.0 ms
    prepF       2.74 us/site  inner=1086  burst=  29.8 ms
    prepS       2.75 us/site  inner=1082  burst=  29.7 ms
      i8S/legA 0.9656x [0.9651-0.9664]  int8 WINS
      i8U/legA 0.9657x [0.9653-0.9664]  int8 WINS
      i8F/legA 0.9657x [0.9651-0.9665]  int8 WINS
      glue fusion i8F/i8U 1.0000x [0.9997-1.0001]; multi-out i8U/i8S 1.0001x
      prepF   2.74 us  prepS   2.75 us  (prep share of i8F 5.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.102e-02

  (none) -> out_proj M=64 tile s98  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        51.88 us/site  inner=58    burst=  30.1 ms
    i8S        51.89 us/site  inner=58    burst=  30.1 ms
    i8U        51.88 us/site  inner=58    burst=  30.1 ms
    legA       56.24 us/site  inner=54    burst=  30.4 ms
    prepF       2.77 us/site  inner=1066  burst=  29.6 ms
    prepS       2.77 us/site  inner=1076  burst=  29.8 ms
      i8S/legA 0.9228x [0.9223-0.9233]  int8 WINS
      i8U/legA 0.9227x [0.9222-0.9231]  int8 WINS
      i8F/legA 0.9226x [0.9221-0.9233]  int8 WINS
      glue fusion i8F/i8U 0.9999x [0.9993-1.0002]; multi-out i8U/i8S 0.9999x
      prepF   2.77 us  prepS   2.77 us  (prep share of i8F 5.3%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.109e-02

  (none) -> out_proj M=64 tile s94  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        42.51 us/site  inner=71    burst=  30.2 ms
    i8S        42.51 us/site  inner=71    burst=  30.2 ms
    i8U        42.52 us/site  inner=71    burst=  30.2 ms
    legA       56.25 us/site  inner=54    burst=  30.4 ms
    prepF       2.77 us/site  inner=1071  burst=  29.7 ms
    prepS       2.77 us/site  inner=1071  burst=  29.7 ms
      i8S/legA 0.7557x [0.7554-0.7566]  int8 WINS
      i8U/legA 0.7560x [0.7554-0.7566]  int8 WINS
      i8F/legA 0.7555x [0.7553-0.7565]  int8 WINS
      glue fusion i8F/i8U 0.9998x [0.9991-0.9999]; multi-out i8U/i8S 1.0003x
      prepF   2.77 us  prepS   2.77 us  (prep share of i8F 6.5%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.109e-02

  (none) -> out_proj M=64 tile s85  reps=7 burst>=30 ms  cell 1.3 s
    legA pins 5120:s5(TM32)
    i8F        53.02 us/site  inner=57    burst=  30.2 ms
    i8S        53.02 us/site  inner=57    burst=  30.2 ms
    i8U        53.02 us/site  inner=57    burst=  30.2 ms
    legA       56.25 us/site  inner=54    burst=  30.4 ms
    prepF       2.77 us/site  inner=1071  burst=  29.7 ms
    prepS       2.77 us/site  inner=1071  burst=  29.7 ms
      i8S/legA 0.9425x [0.9421-0.9428]  int8 WINS
      i8U/legA 0.9425x [0.9421-0.9429]  int8 WINS
      i8F/legA 0.9425x [0.9420-0.9428]  int8 WINS
      glue fusion i8F/i8U 0.9999x [0.9998-1.0001]; multi-out i8U/i8S 1.0000x
      prepF   2.77 us  prepS   2.77 us  (prep share of i8F 5.2%)
      witness: fused prep >1 int8 step on 0.0000% of elems; rel|i8 - legA| 1.109e-02

[tenancy after] 1, 776 MiB
[under load] sm clock median 2700 MHz (max 2700), power median 274 W, sw_power_cap on 114 of 2758, hw_slowdown 0

WHOLE-MODEL verify forward, ms of weight GEMM + glue (400 linears)
     M      legA       i8S       i8U       i8F  i8F/legA  i8S/legA
     8    16.125    19.568    19.457    19.453    1.2064    1.2135
    16    16.945    19.734    19.612    19.595    1.1564    1.1646
    32    20.395    20.716    20.595    20.539    1.0071    1.0158
    36    36.837    23.349    23.154    23.147    0.6284    0.6338
    48    38.015    23.438    23.260    23.242    0.6114    0.6166
    64    40.086    25.499    25.329    25.312    0.6314    0.6361
