diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-07-02 19:44:08 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-07-02 19:44:08 +0200 |
| commit | 6a10d3577232f1fd646a20f22a4a1ab4ec29f28a (patch) | |
| tree | e4a077b064876ddb39144935fa256a4ab0692488 /deps/cudnn-frontend/include/cudnn_frontend/node/batchnorm_inference.h | |
| parent | 98a713e86c23bc6098bf7a5558507a9f27a7a1c1 (diff) | |
| download | ffwd-6a10d3577232f1fd646a20f22a4a1ab4ec29f28a.tar.gz ffwd-6a10d3577232f1fd646a20f22a4a1ab4ec29f28a.zip | |
cuda: add cuDNN fused attention path with runtime fallback
Diffstat (limited to 'deps/cudnn-frontend/include/cudnn_frontend/node/batchnorm_inference.h')
| -rw-r--r-- | deps/cudnn-frontend/include/cudnn_frontend/node/batchnorm_inference.h | 156 |
1 files changed, 156 insertions, 0 deletions
diff --git a/deps/cudnn-frontend/include/cudnn_frontend/node/batchnorm_inference.h b/deps/cudnn-frontend/include/cudnn_frontend/node/batchnorm_inference.h new file mode 100644 index 0000000..9fb433f --- /dev/null +++ b/deps/cudnn-frontend/include/cudnn_frontend/node/batchnorm_inference.h @@ -0,0 +1,156 @@ +#pragma once + +#include "../graph_helpers.h" +#include "../node_interface.h" + +namespace cudnn_frontend { + +namespace graph { +class BatchnormInferenceNode : public NodeCRTP<BatchnormInferenceNode> { + public: + Batchnorm_inference_attributes attributes; + + BatchnormInferenceNode(Batchnorm_inference_attributes&& attributes_, detail::Context const& context) + : NodeCRTP(context), attributes(std::move(attributes_)) {} + + Type + getType() override final { + return Type::BATCHNORM_INFERENCE; + } + + error_t + infer_properties_node() override final { + CUDNN_FE_LOG_LABEL_ENDL("INFO: Inferencing properties for batchnorm inference node " << attributes.name); + + attributes.fill_from_context(context); + + auto X = attributes.inputs[Batchnorm_inference_attributes::input_names::X]; + auto Y = attributes.outputs[Batchnorm_inference_attributes::output_names::Y]; + // Only infer dims and strides if user did not set them + if (Y->get_dim().empty()) { + Y->set_dim(X->get_dim()); + } + if (Y->get_stride().empty()) { + Y->set_stride(X->get_stride()); + } + + return {error_code_t::OK, ""}; + } + + error_t + create_cudnn_operations( + std::unordered_set<Tensor_attributes::uid_t>& uids_involved_in_operations, + std::vector<std::shared_ptr<cudnn_frontend::Operation>>& operations, + managed_backend_descriptor_t& raw_operations, + std::unordered_map<int64_t, std::shared_ptr<cudnn_frontend::Tensor>>& tensors) const override final { + CUDNN_FRONTEND_UNUSED(raw_operations); + CUDNN_FE_LOG_LABEL("INFO: Building BatchnormInferenceNode operations " << attributes.name << " "); + + // Create operation by directly calling cuDNN backend API + Operation_v8 batchnorm_operation; + + _CUDNN_CHECK_CUDNN_ERROR( + batchnorm_operation.initialize_managed_backend_pointer(CUDNN_BACKEND_OPERATION_NORM_FORWARD_DESCRIPTOR)); + + // Set norm mode to BATCH_NORM + cudnnBackendNormMode_t cudnn_norm_mode; + + _CUDNN_CHECK_CUDNN_ERROR(detail::convert_to_cudnn_type(NormMode_t::BATCH_NORM, cudnn_norm_mode)); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_MODE, + CUDNN_TYPE_NORM_MODE, + 1, + &cudnn_norm_mode)); + + // Set forward phase to INFERENCE + cudnnBackendNormFwdPhase_t cudnn_norm_fwd_phase; + + _CUDNN_CHECK_CUDNN_ERROR(detail::convert_to_cudnn_type(NormFwdPhase_t::INFERENCE, cudnn_norm_fwd_phase)); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_PHASE, + CUDNN_TYPE_NORM_FWD_PHASE, + 1, + &cudnn_norm_fwd_phase)); + + // Set input tensor X + CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(X, Batchnorm_inference_attributes::input_names::X); + auto x_desc = tensors.at(X->second->get_uid())->get_raw_desc(); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_XDESC, + CUDNN_TYPE_BACKEND_DESCRIPTOR, + 1, + &x_desc)); + + // Set mean and inv_variance (as inputs for inference) + CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(MEAN, Batchnorm_inference_attributes::input_names::MEAN); + auto mean_desc = tensors.at(MEAN->second->get_uid())->get_raw_desc(); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_MEAN_DESC, + CUDNN_TYPE_BACKEND_DESCRIPTOR, + 1, + &mean_desc)); + + CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(INV_VARIANCE, + Batchnorm_inference_attributes::input_names::INV_VARIANCE); + auto inv_var_desc = tensors.at(INV_VARIANCE->second->get_uid())->get_raw_desc(); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_INV_VARIANCE_DESC, + CUDNN_TYPE_BACKEND_DESCRIPTOR, + 1, + &inv_var_desc)); + + // Set scale and bias tensors + CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(SCALE, Batchnorm_inference_attributes::input_names::SCALE); + auto scale_desc = tensors.at(SCALE->second->get_uid())->get_raw_desc(); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_SCALE_DESC, + CUDNN_TYPE_BACKEND_DESCRIPTOR, + 1, + &scale_desc)); + + CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(BIAS, Batchnorm_inference_attributes::input_names::BIAS); + auto bias_desc = tensors.at(BIAS->second->get_uid())->get_raw_desc(); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_BIAS_DESC, + CUDNN_TYPE_BACKEND_DESCRIPTOR, + 1, + &bias_desc)); + + // Set output tensor Y + CUDNN_FE_VALIDATE_AND_ASSIGN_OUTPUT_TENSOR(Y, Batchnorm_inference_attributes::output_names::Y); + auto y_desc = tensors.at(Y->second->get_uid())->get_raw_desc(); + + _CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(), + CUDNN_ATTR_OPERATION_NORM_FWD_YDESC, + CUDNN_TYPE_BACKEND_DESCRIPTOR, + 1, + &y_desc)); + + _CUDNN_CHECK_CUDNN_ERROR(detail::finalize(batchnorm_operation.get_raw_desc())); + + operations.push_back(std::make_shared<Operation_v8>(std::move(batchnorm_operation))); + + auto const& non_virtual_uids = attributes.get_non_virtual_uids(); + uids_involved_in_operations.insert(non_virtual_uids.begin(), non_virtual_uids.end()); + return {error_code_t::OK, ""}; + } + +#ifndef CUDNN_FRONTEND_SKIP_JSON_LIB + virtual void + serialize(json& j) const override final { + j = attributes; + j.update(R"( {"tag": "BATCHNORM_INFERENCE"})"_json); + } +#endif +}; + +} // namespace graph + +} // namespace cudnn_frontend
\ No newline at end of file |
