# rocBLAS solution indices for the prompt path's FP16-out GEMMs; valid for this architecture and rocBLAS build only
# rows: type N K ldy T solution   (type f16o: FP16 in, FP16 out, FP32 compute; Gemm::f16_inplace)
# solution: rocBLAS's solution index, or `default` where its own choice measured best at that bucket
STRATA_ROCBLAS_TUNING_V1 gfx1030 5.2.0.dabb6df2b98-dirty
f16o 10240 2560 10240 256 -1086328261   # 2.54294 ms -> 0.345287 ms
f16o 10240 2560 10240 512 -1086328277   # 5.16948 ms -> 0.669216 ms
f16o 10240 2560 10240 768 -1086328245   # 7.56462 ms -> 1.02345 ms
f16o 10240 2560 10240 1024 -1086328245   # 10.2347 ms -> 1.3804 ms
f16o 10240 2560 10240 1536 default   # 1.96263 ms
f16o 10240 2560 10240 2048 default   # 2.54689 ms
f16o 10240 2560 10240 4096 default   # 5.17027 ms
f16o 10240 2560 10240 8192 default   # 11.689 ms
f16o 12288 2560 12288 256 -1086328276   # 2.95806 ms -> 0.496324 ms
f16o 12288 2560 12288 512 -1086328277   # 0.996887 ms -> 0.820375 ms
f16o 12288 2560 12288 768 -1086328277   # 1.56971 ms -> 1.28913 ms
f16o 12288 2560 12288 1024 default   # 1.82413 ms
f16o 12288 2560 12288 1536 -1086328245   # 2.63703 ms -> 2.39408 ms
f16o 12288 2560 12288 2048 -1086328245   # 3.41317 ms -> 3.1995 ms
f16o 12288 2560 12288 4096 default   # 6.82793 ms
f16o 12288 2560 12288 8192 default   # 13.75 ms
f16o 6144 2560 6144 256 -1086328259   # 1.01279 ms -> 0.27815 ms
f16o 6144 2560 6144 512 -1086328261   # 1.8456 ms -> 0.501155 ms
f16o 6144 2560 6144 768 default   # 0.663882 ms
f16o 6144 2560 6144 1024 default   # 0.82505 ms
f16o 6144 2560 6144 1536 default   # 1.28901 ms
f16o 6144 2560 6144 2048 -1086328245   # 2.35832 ms -> 1.74556 ms
f16o 6144 2560 6144 4096 default   # 3.17068 ms
f16o 6144 2560 6144 8192 default   # 6.35074 ms
f16o 2560 6144 2560 256 -1086328322   # 1.37524 ms -> 0.286257 ms
f16o 2560 6144 2560 512 -1086328261   # 1.28905 ms -> 0.466465 ms
f16o 2560 6144 2560 768 default   # 0.758482 ms
f16o 2560 6144 2560 1024 -1086328277   # 0.970742 ms -> 0.898318 ms
f16o 2560 6144 2560 1536 -1086328261   # 1.56328 ms -> 1.40264 ms
f16o 2560 6144 2560 2048 default   # 1.80918 ms
f16o 2560 6144 2560 4096 default   # 3.65266 ms
f16o 2560 6144 2560 8192 default   # 7.53115 ms
f16o 2560 640 2560 256 -1086328464   # 0.0431323 ms -> 0.038199 ms
f16o 2560 640 2560 512 -1086328271   # 0.0721713 ms -> 0.0576917 ms
f16o 2560 640 2560 768 -1086328493   # 0.0951573 ms -> 0.0785843 ms
f16o 2560 640 2560 1024 -1086328250   # 0.10769 ms -> 0.0937707 ms
f16o 2560 640 2560 1536 -1086328271   # 0.155556 ms -> 0.136023 ms
f16o 2560 640 2560 2048 default   # 0.180861 ms
f16o 2560 640 2560 4096 default   # 0.34475 ms
f16o 2560 640 2560 8192 -1086328238   # 0.693967 ms -> 0.651795 ms
f16o 2560 2560 2560 256 default   # 0.128383 ms
f16o 2560 2560 2560 512 -1086328261   # 0.217194 ms -> 0.200714 ms
f16o 2560 2560 2560 768 default   # 0.296738 ms
f16o 2560 2560 2560 1024 -1086328245   # 0.412855 ms -> 0.351483 ms
f16o 2560 2560 2560 1536 default   # 0.513452 ms
f16o 2560 2560 2560 2048 default   # 0.688994 ms
f16o 2560 2560 2560 4096 -1086328261   # 1.50943 ms -> 1.33379 ms
f16o 2560 2560 2560 8192 -1086328261   # 2.80126 ms -> 2.6054 ms
f16o 640 2560 640 256 -1086328281   # 0.162249 ms -> 0.0554783 ms
f16o 640 2560 640 512 -1086328282   # 0.195902 ms -> 0.083451 ms
f16o 640 2560 640 768 -1086328283   # 0.152316 ms -> 0.114144 ms
f16o 640 2560 640 1024 -1086328276   # 0.213034 ms -> 0.125423 ms
f16o 640 2560 640 1536 -1086328275   # 0.260087 ms -> 0.197715 ms
f16o 640 2560 640 2048 -1086328277   # 0.218461 ms -> 0.201288 ms
f16o 640 2560 640 4096 default   # 0.344818 ms
f16o 640 2560 640 8192 default   # 0.681609 ms
f16o 512 2560 512 256 -1086328332   # 0.150156 ms -> 0.046799 ms
f16o 512 2560 512 512 -1086328280   # 0.198168 ms -> 0.071905 ms
f16o 512 2560 512 768 -1086328331   # 0.145263 ms -> 0.100797 ms
f16o 512 2560 512 1024 -1086328280   # 0.152569 ms -> 0.12405 ms
f16o 512 2560 512 1536 -1086328280   # 0.218594 ms -> 0.169409 ms
f16o 512 2560 512 2048 -1086328261   # 0.209875 ms -> 0.198968 ms
f16o 512 2560 512 4096 default   # 0.342058 ms
f16o 512 2560 512 8192 -1086328245   # 0.698289 ms -> 0.658783 ms
f16o 128 2560 128 256 -1086328327   # 0.0575983 ms -> 0.0334523 ms
f16o 128 2560 128 512 -1086328327   # 0.192688 ms -> 0.0367193 ms
f16o 128 2560 128 768 -1086328327   # 0.109797 ms -> 0.0451057 ms
f16o 128 2560 128 1024 -1086328330   # 0.112637 ms -> 0.049492 ms
f16o 128 2560 128 1536 -1086328330   # 0.153369 ms -> 0.065745 ms
f16o 128 2560 128 2048 -1086328280   # 0.12381 ms -> 0.076158 ms
f16o 128 2560 128 4096 -1086328308   # 0.15169 ms -> 0.12401 ms
f16o 128 2560 128 8192 default   # 0.202315 ms
f16o 48 2560 48 256 -1086328398   # 0.0449723 ms -> 0.020266 ms
f16o 48 2560 48 512 -1086328339   # 0.196942 ms -> 0.0364923 ms
f16o 48 2560 48 768 -1086328334   # 0.102744 ms -> 0.040892 ms
f16o 48 2560 48 1024 -1086328334   # 0.104411 ms -> 0.0430787 ms
f16o 48 2560 48 1536 -1086328334   # 0.111064 ms -> 0.049972 ms
f16o 48 2560 48 2048 -1086328337   # 0.120264 ms -> 0.056292 ms
f16o 48 2560 48 4096 -1086328337   # 0.181915 ms -> 0.0996777 ms
f16o 48 2560 48 8192 -1086328294   # 0.203128 ms -> 0.149476 ms
f16o 320 10240 320 256 -1086328339   # 0.397617 ms -> 0.13093 ms
f16o 320 10240 320 512 -1086328338   # 0.581545 ms -> 0.194875 ms
f16o 320 10240 320 768 -1086328280   # 1.02177 ms -> 0.232008 ms
f16o 320 10240 320 1024 -1086328331   # 1.28118 ms -> 0.352498 ms
f16o 320 10240 320 1536 -1086328276   # 2.00815 ms -> 0.460815 ms
f16o 320 10240 320 2048 -1086328282   # 2.68941 ms -> 0.599478 ms
f16o 320 10240 320 4096 -1086328259   # 8.24766 ms -> 1.09359 ms
f16o 320 10240 320 8192 default   # 2.18572 ms
f16o 10240 320 10240 256 default   # 0.0574917 ms
f16o 10240 320 10240 512 default   # 0.097331 ms
f16o 10240 320 10240 768 default   # 0.140477 ms
f16o 10240 320 10240 1024 default   # 0.181956 ms
f16o 10240 320 10240 1536 -1086328238   # 0.274793 ms -> 0.261327 ms
f16o 10240 320 10240 2048 default   # 0.342885 ms
f16o 10240 320 10240 4096 default   # 0.673544 ms
f16o 10240 320 10240 8192 default   # 1.34294 ms
f16o 4 10240 4 256 -1086328399   # 0.097331 ms -> 0.0496253 ms
f16o 4 10240 4 512 -1086328398   # 0.253181 ms -> 0.0543853 ms
f16o 4 10240 4 768 -1086328398   # 0.251141 ms -> 0.061092 ms
f16o 4 10240 4 1024 -1086328408   # 0.247781 ms -> 0.079345 ms
f16o 4 10240 4 1536 -1086328402   # 0.266167 ms -> 0.110837 ms
f16o 4 10240 4 2048 -1086328337   # 0.293807 ms -> 0.130957 ms
f16o 4 10240 4 4096 -1086328482   # 0.578413 ms -> 0.251874 ms
f16o 4 10240 4 8192 -1086328355   # 1.35875 ms -> 0.482975 ms
f16o 1 2560 1 256 -9   # 0.0427057 ms -> 0.00797333 ms
f16o 1 2560 1 512 -9   # 0.193929 ms -> 0.0102663 ms
f16o 1 2560 1 768 -9   # 0.0855447 ms -> 0.0120797 ms
f16o 1 2560 1 1024 -9   # 0.0907977 ms -> 0.016093 ms
f16o 1 2560 1 1536 -9   # 0.0921313 ms -> 0.018053 ms
f16o 1 2560 1 2048 -9   # 0.119091 ms -> 0.0275593 ms
f16o 1 2560 1 4096 -9   # 0.0724783 ms -> 0.0414923 ms
f16o 1 2560 1 8192 -9   # 0.197969 ms -> 0.0730517 ms
