| DynamicDispatch Offload - not offloaded | |
| +-------------------+-------+------------------------+------------------------+ | |
| | Op Type | Count | Inputs | Outputs | | |
| +===================+=======+========================+========================+ | |
| | CastAvx | 1 | [1,1024,1024,4] - | [1,1024,1024,4] - | | |
| | | | FLOAT | BFLOAT16 | | |
| | CastAvx | 1 | [1,128,128,8] - | [1,128,128,8] - FLOAT | | |
| | | | BFLOAT16 | | | |
| | Pad | 1 | [1,3,1024,1024] - | [1,4,1024,1024] - | | |
| | | | FLOAT | FLOAT | | |
| | | | [8] - INT64 | | | |
| | | | [1] - FLOAT | | | |
| | Slice | 1 | [1,8,128,128] - FLOAT | [1,4,128,128] - FLOAT | | |
| | | | [1] - INT64 | | | |
| | | | [1] - INT64 | | | |
| | | | [1] - INT64 | | | |
| | Transpose | 1 | [1,4,1024,1024] - | [1,1024,1024,4] - | | |
| | | | FLOAT | FLOAT | | |
| | Transpose | 1 | [1,128,128,8] - FLOAT | [1,8,128,128] - FLOAT | | |
| +-------------------+-------+------------------------+------------------------+ | |
| | Not offloaded sum | 6 | | | | |
| +-------------------+-------+------------------------+------------------------+ | |
| DynamicDispatch Offload - offloaded | |
| +--------------------+-------------+--------------------+---------------------+ | |
| | Op Type | Count | Inputs | Outputs | | |
| +====================+=============+====================+=====================+ | |
| | DynamicDispatch | 1 | [1,1024,1024,4] - | [1,128,128,8] - | | |
| | | | BFLOAT16 | BFLOAT16 | | |
| +--------------------+-------------+--------------------+---------------------+ | |
| | Offloaded sum | 1 | | | | |
| +--------------------+-------------+--------------------+---------------------+ | |
| | Offloaded Op Types | SDAdd | | | | |
| | | SDMHA_VAE | | | | |
| | | SDConv | | | | |
| | | SDGroupNorm | | | | |
| | | SDGemm | | | | |
| +--------------------+-------------+--------------------+---------------------+ | |
| | Offloaded sum (dd | 66 | | | | |
| | fusion) | | | | | |
| | Offload Ratio (dd | 91.67% | | | | |
| | fusion) | | | | | |
| +--------------------+-------------+--------------------+---------------------+ | |