| 53 | } |
| 54 | |
| 55 | void ConvBiasForwardImpl::AlgoChanwiseSmall::exec(const ExecArgs& args) const { |
| 56 | WorkspaceBundle bundle{args.workspace.raw_ptr, {get_workspace_in_bytes(args)}}; |
| 57 | TensorND conv_dst_tensor = *args.dst_tensor; |
| 58 | if (args.dst_layout->dtype.enumv() != args.bias_layout->dtype.enumv()) { |
| 59 | conv_dst_tensor = TensorND{bundle.get(0), conv_dst_tensor.layout}; |
| 60 | conv_dst_tensor.layout.dtype = DType(); |
| 61 | args.opr->check_or_deduce_dtype_fwd( |
| 62 | args.src_layout->dtype, args.filter_layout->dtype, |
| 63 | conv_dst_tensor.layout.dtype); |
| 64 | } |
| 65 | { |
| 66 | auto kparam = chanwise::Param::from_fwd_args(args); |
| 67 | auto stream = cuda_stream(args.handle); |
| 68 | switch (args.src_layout->dtype.enumv()) { |
| 69 | case DTypeEnum::Float32: |
| 70 | chanwise::run_fwd_small( |
| 71 | conv_dst_tensor.ptr<float>(), args.src_tensor->ptr<float>(), |
| 72 | args.filter_tensor->ptr<float>(), kparam, stream); |
| 73 | break; |
| 74 | #if CUDA_VERSION >= 9000 |
| 75 | case DTypeEnum::Float16: |
| 76 | chanwise::run_fwd_small( |
| 77 | static_cast<half*>(conv_dst_tensor.raw_ptr()), |
| 78 | static_cast<half*>(args.src_tensor->raw_ptr()), |
| 79 | static_cast<half*>(args.filter_tensor->raw_ptr()), kparam, |
| 80 | stream); |
| 81 | break; |
| 82 | #endif |
| 83 | default: |
| 84 | megdnn_assert_internal(0); |
| 85 | } |
| 86 | } |
| 87 | handle_bias_and_nonlinear( |
| 88 | args.handle, args.nonlinear_mode, &conv_dst_tensor, args.dst_tensor, |
| 89 | args.bias_tensor); |
| 90 | } |
| 91 | |
| 92 | // vim: syntax=cpp.doxygen |
nothing calls this directly
no test coverage detected