[KMCompiler][Ascend] ascend masked_scatter_backward

[YangLong114514]

PR Category

Operator

Type of Change

Performance Optimization

Description

add operator masked_scatter_backward for ascend backend.

Performance

test_masked_scatter_backward.py
Operator: masked_scatter_backward  Performance Test (dtype=torch.float16, mode=operator,level=comprehensive)
Status       Torch Latency (ms)    Gems Latency (ms)         Gems Speedup          Torch GBPS            Gems GBPS           Size Detail
-----------------------------------------------------------------------------------------------------------------------------------------
SUCCESS               0.186865            0.222445               0.840               0.110               0.092          [torch.Size([64, 64]), torch.Size([64, 64]), [64, 64]]
SUCCESS               0.191116            0.213633               0.895               1.715               1.534          [torch.Size([256, 256]), torch.Size([256, 256]), [256, 256]]
SUCCESS               0.230341            0.404634               0.569              22.761              12.957          [torch.Size([1024, 1024]), torch.Size([1024, 1024]), [1024, 1024]]
SUCCESS               0.299732            2.089534               0.143             279.871              40.146          [torch.Size([4096, 4096]), torch.Size([4096, 4096]), [4096, 4096]]
SUCCESS               0.834866            8.304328               0.101             401.914              40.406          [torch.Size([1024, 65536]), torch.Size([1024, 65536]), [1024, 65536]]
SUCCESS               0.207613            0.373913               0.555              61.653              34.233          [torch.Size([10000, 256]), torch.Size([10000, 256]), [10000, 256]]
SUCCESS               6.694811           80.605745               0.083             489.454              40.652          [torch.Size([10000, 65536]), torch.Size([10000, 65536]), [10000, 65536]]
SUCCESS               0.214244            0.150338               1.425               0.007               0.010          [torch.Size([17, 17]), torch.Size([17, 17]), [17, 17]]
SUCCESS               0.238289            0.150170               1.587               0.086               0.136          [torch.Size([32, 128]), torch.Size([32, 128]), [32, 128]]
SUCCESS               0.208093            0.151333               1.375               0.197               0.271          [torch.Size([128, 64]), torch.Size([128, 64]), [128, 64]]
SUCCESS               0.192496            0.150580               1.278               0.426               0.544          [torch.Size([256, 64]), torch.Size([256, 64]), [256, 64]]
SUCCESS               0.221345            0.419613               0.527              47.373              24.989          [torch.Size([1024, 2048]), torch.Size([1024, 2048]), [1024, 2048]]


Operator: masked_scatter_backward  Performance Test (dtype=torch.float32, mode=operator,level=comprehensive)
Status       Torch Latency (ms)    Gems Latency (ms)         Gems Speedup          Torch GBPS            Gems GBPS           Size Detail
-----------------------------------------------------------------------------------------------------------------------------------------
SUCCESS               0.190768            0.148341               1.286               0.193               0.249          [torch.Size([64, 64]), torch.Size([64, 64]), [64, 64]]
SUCCESS               0.290923            0.213058               1.365               2.027               2.768          [torch.Size([256, 256]), torch.Size([256, 256]), [256, 256]]
SUCCESS               0.257021            0.437299               0.588              36.718              21.581          [torch.Size([1024, 1024]), torch.Size([1024, 1024]), [1024, 1024]]
SUCCESS               0.432519            2.230018               0.194             349.106              67.710          [torch.Size([4096, 4096]), torch.Size([4096, 4096]), [4096, 4096]]
SUCCESS               1.420952            8.811322               0.161             425.053              68.546          [torch.Size([1024, 65536]), torch.Size([1024, 65536]), [1024, 65536]]
SUCCESS               0.221188            0.406568               0.544             104.165              56.670          [torch.Size([10000, 256]), torch.Size([10000, 256]), [10000, 256]]
SUCCESS              12.251547           85.894108               0.143             481.428              68.669          [torch.Size([10000, 65536]), torch.Size([10000, 65536]), [10000, 65536]]
SUCCESS               0.180288            0.144298               1.249               0.014               0.018          [torch.Size([17, 17]), torch.Size([17, 17]), [17, 17]]
SUCCESS               0.186995            0.151316               1.236               0.197               0.244          [torch.Size([32, 128]), torch.Size([32, 128]), [32, 128]]
SUCCESS               0.221251            0.153422               1.442               0.333               0.481          [torch.Size([128, 64]), torch.Size([128, 64]), [128, 64]]
SUCCESS               0.227187            0.156054               1.456               0.649               0.945          [torch.Size([256, 64]), torch.Size([256, 64]), [256, 64]]
SUCCESS               0.241353            0.420698               0.574              78.202              44.864          [torch.Size([1024, 2048]), torch.Size([1024, 2048]), [1024, 2048]]


Operator: masked_scatter_backward  Performance Test (dtype=torch.bfloat16, mode=operator,level=comprehensive)
Status       Torch Latency (ms)    Gems Latency (ms)         Gems Speedup          Torch GBPS            Gems GBPS           Size Detail
-----------------------------------------------------------------------------------------------------------------------------------------
SUCCESS               0.195788            0.151861               1.289               0.105               0.135          [torch.Size([64, 64]), torch.Size([64, 64]), [64, 64]]
SUCCESS               0.203880            0.215265               0.947               1.607               1.522          [torch.Size([256, 256]), torch.Size([256, 256]), [256, 256]]
SUCCESS               0.228684            0.413876               0.553              22.926              12.668          [torch.Size([1024, 1024]), torch.Size([1024, 1024]), [1024, 1024]]
SUCCESS               0.304732            2.085660               0.146             275.278              40.220          [torch.Size([4096, 4096]), torch.Size([4096, 4096]), [4096, 4096]]
SUCCESS               0.852247            8.287827               0.103             393.717              40.486          [torch.Size([1024, 65536]), torch.Size([1024, 65536]), [1024, 65536]]
SUCCESS               0.223813            0.417465               0.536              57.191              30.661          [torch.Size([10000, 256]), torch.Size([10000, 256]), [10000, 256]]
SUCCESS               6.575993           80.737352               0.081             498.297              40.586          [torch.Size([10000, 65536]), torch.Size([10000, 65536]), [10000, 65536]]
SUCCESS               0.181642            0.145472               1.249               0.008               0.010          [torch.Size([17, 17]), torch.Size([17, 17]), [17, 17]]
SUCCESS               0.199590            0.145099               1.376               0.103               0.141          [torch.Size([32, 128]), torch.Size([32, 128]), [32, 128]]
SUCCESS               0.204784            0.145400               1.408               0.200               0.282          [torch.Size([128, 64]), torch.Size([128, 64]), [128, 64]]
SUCCESS               0.189378            0.147345               1.285               0.433               0.556          [torch.Size([256, 64]), torch.Size([256, 64]), [256, 64]]
SUCCESS               0.214521            0.409541               0.524              48.880              25.604          [torch.Size([1024, 2048]), torch.Size([1024, 2048]), [1024, 2048]]