MIOpen icon indicating copy to clipboard operation
MIOpen copied to clipboard

Implement ResourcApplyGradientDescent

Open hieule88 opened this issue 10 months ago • 0 comments

  • Added ResourceApplyGradientDescent.

  • Added driver test and gtest for ResourceApplyGradientDescent.

  • New API is guarded by MIOPEN_BETA_API macro.

  • Average over all cases:

  • ResourceApplyGradientDescent

Type Forward
float16 2.08
float32 1.92
bfloat16 2.10

FP16
op_name dtype input_size direction rocm_kernel_avg kernel_duration MIOpen over rocm
ResourceApplyGradientDescent float16 [50 100] fwd 11776 4728 2.490693739
ResourceApplyGradientDescent float16 [100 50] fwd 11280 4889 2.30722029
ResourceApplyGradientDescent float16 [100 100] fwd 11408 4711 2.421566546
ResourceApplyGradientDescent float16 [100 300] fwd 11504 4871 2.361732704
ResourceApplyGradientDescent float16 [300 100] fwd 11584 4640 2.496551724
ResourceApplyGradientDescent float16 [200 300] fwd 12016 4800 2.503333333
ResourceApplyGradientDescent float16 [205 350] fwd 11856 5031 2.356589147
ResourceApplyGradientDescent float16 [350 105] fwd 11488 4871 2.358447957
ResourceApplyGradientDescent float16 [405 200] fwd 11936 5155 2.31542192
ResourceApplyGradientDescent float16 [10 10 10] fwd 10176 3928 2.590631365
ResourceApplyGradientDescent float16 [10 10 30] fwd 10944 4835 2.263495346
ResourceApplyGradientDescent float16 [10 30 10] fwd 10976 4729 2.320998097
ResourceApplyGradientDescent float16 [30 10 10] fwd 11600 4711 2.462322225
ResourceApplyGradientDescent float16 [30 30 30] fwd 11552 4746 2.434049726
ResourceApplyGradientDescent float16 [50 100 50] fwd 14416 6720 2.145238095
ResourceApplyGradientDescent float16 [100 50 100] fwd 14768 8266 1.786595693
ResourceApplyGradientDescent float16 [100 100 100] fwd 20576 11626 1.769826252
ResourceApplyGradientDescent float16 [100 100 300] fwd 38336 25138 1.525021879
ResourceApplyGradientDescent float16 [300 100 100] fwd 37936 25049 1.514471636
ResourceApplyGradientDescent float16 [10 10 10 10] fwd 11152 4906 2.273134937
ResourceApplyGradientDescent float16 [10 10 10 30] fwd 11584 4818 2.404317144
ResourceApplyGradientDescent float16 [30 10 10 10] fwd 11568 4871 2.37487169
ResourceApplyGradientDescent float16 [30 30 30 30] fwd 16432 10382 1.582739357
ResourceApplyGradientDescent float16 [50 100 50 100] fwd 239327 172503 1.38737877
ResourceApplyGradientDescent float16 [100 50 100 50] fwd 239439 172521 1.387883214
ResourceApplyGradientDescent float16 [100 100 100 100] fwd 979692 678811 1.443247089
ResourceApplyGradientDescent float16 [100 100 300 100] fwd 2860371 2021410 1.415037523
ResourceApplyGradientDescent float16 [300 100 100 100] fwd 2864580 2023310 1.415788979

FP32
op_name dtype input_size direction rocm_kernel_avg kernel_duration MIOpen over rocm
ResourceApplyGradientDescent float32 [50 100] fwd 9984 4764 2.095717884
ResourceApplyGradientDescent float32 [100 50] fwd 9856 4711 2.092124814
ResourceApplyGradientDescent float32 [100 100] fwd 10048 4729 2.124762106
ResourceApplyGradientDescent float32 [100 300] fwd 10352 4942 2.094698503
ResourceApplyGradientDescent float32 [300 100] fwd 10624 4871 2.181071649
ResourceApplyGradientDescent float32 [200 300] fwd 11040 5173 2.134158129
ResourceApplyGradientDescent float32 [205 350] fwd 11248 5155 2.181959263
ResourceApplyGradientDescent float32 [350 105] fwd 11008 4871 2.259905564
ResourceApplyGradientDescent float32 [405 200] fwd 11120 5386 2.064611957
ResourceApplyGradientDescent float32 [10 10 10] fwd 8512 4889 1.74105134
ResourceApplyGradientDescent float32 [10 10 30] fwd 9936 4746 2.093552465
ResourceApplyGradientDescent float32 [10 30 10] fwd 10608 4782 2.218318695
ResourceApplyGradientDescent float32 [30 10 10] fwd 10192 4711 2.163447251
ResourceApplyGradientDescent float32 [30 30 30] fwd 10512 4942 2.127074059
ResourceApplyGradientDescent float32 [50 100 50] fwd 13488 7484 1.802244789
ResourceApplyGradientDescent float32 [100 50 100] fwd 14768 9280 1.59137931
ResourceApplyGradientDescent float32 [100 100 100] fwd 24544 14987 1.637685995
ResourceApplyGradientDescent float32 [100 100 300] fwd 52736 32658 1.614795762
ResourceApplyGradientDescent float32 [300 100 100] fwd 52736 32338 1.63077494
ResourceApplyGradientDescent float32 [10 10 10 10] fwd 10192 4728 2.155668359
ResourceApplyGradientDescent float32 [10 10 10 30] fwd 10816 4924 2.19658814
ResourceApplyGradientDescent float32 [30 10 10 10] fwd 10528 4924 2.138099106
ResourceApplyGradientDescent float32 [30 30 30 30] fwd 21472 13191 1.627776514
ResourceApplyGradientDescent float32 [50 100 50 100] fwd 375838 243829 1.541399916
ResourceApplyGradientDescent float32 [100 50 100 50] fwd 376158 243794 1.542933788
ResourceApplyGradientDescent float32 [100 100 100 100] fwd 1493306 945557 1.579287129
ResourceApplyGradientDescent float32 [100 100 300 100] fwd 4483885 2815180 1.592752506
ResourceApplyGradientDescent float32 [300 100 100 100] fwd 4464925 2816790 1.585111066

BFP16
op_name dtype input_size direction rocm_kernel_avg kernel_duration MIOpen over rocm
ResourceApplyGradientDescent bfloat16 [50 100] fwd 15696 4818 3.257783313
ResourceApplyGradientDescent bfloat16 [100 50] fwd 14480 4764 3.039462636
ResourceApplyGradientDescent bfloat16 [100 100] fwd 13264 4835 2.743329886
ResourceApplyGradientDescent bfloat16 [100 300] fwd 10624 4853 2.189161343
ResourceApplyGradientDescent bfloat16 [300 100] fwd 10704 4764 2.246851385
ResourceApplyGradientDescent bfloat16 [200 300] fwd 11056 4942 2.237150951
ResourceApplyGradientDescent bfloat16 [205 350] fwd 9952 4924 2.02112104
ResourceApplyGradientDescent bfloat16 [350 105] fwd 13104 4835 2.710237849
ResourceApplyGradientDescent bfloat16 [405 200] fwd 10368 5262 1.970353478
ResourceApplyGradientDescent bfloat16 [10 10 10] fwd 14768 3857 3.828882551
ResourceApplyGradientDescent bfloat16 [10 10 30] fwd 13600 4711 2.886860539
ResourceApplyGradientDescent bfloat16 [10 30 10] fwd 13360 4657 2.868799656
ResourceApplyGradientDescent bfloat16 [30 10 10] fwd 13344 4711 2.832519635
ResourceApplyGradientDescent bfloat16 [30 30 30] fwd 10512 4853 2.166082835
ResourceApplyGradientDescent bfloat16 [50 100 50] fwd 10576 6755 1.56565507
ResourceApplyGradientDescent bfloat16 [100 50 100] fwd 12448 8302 1.499397735
ResourceApplyGradientDescent bfloat16 [100 100 100] fwd 19936 11893 1.676280165
ResourceApplyGradientDescent bfloat16 [100 100 300] fwd 32688 25991 1.257666115
ResourceApplyGradientDescent bfloat16 [300 100 100] fwd 33920 25760 1.316770186
ResourceApplyGradientDescent bfloat16 [10 10 10 10] fwd 14752 4818 3.061851391
ResourceApplyGradientDescent bfloat16 [10 10 10 30] fwd 11792 4835 2.438883144
ResourceApplyGradientDescent bfloat16 [30 10 10 10] fwd 10528 4782 2.201589293
ResourceApplyGradientDescent bfloat16 [30 30 30 30] fwd 14400 10417 1.382355765
ResourceApplyGradientDescent bfloat16 [50 100 50 100] fwd 199342 180130 1.106656304
ResourceApplyGradientDescent bfloat16 [100 50 100 50] fwd 197758 180183 1.097539724
ResourceApplyGradientDescent bfloat16 [100 100 100 100] fwd 761881 721869 1.055428339
ResourceApplyGradientDescent bfloat16 [100 100 300 100] fwd 2282092 2114300 1.079360545
ResourceApplyGradientDescent bfloat16 [300 100 100 100] fwd 2282588 2123540 1.074897577

hieule88 avatar Feb 20 '25 04:02 hieule88