// Copyright © Advanced Micro Devices, Inc., or its affiliates. // SPDX-License-Identifier: MIT #include #include #include #include #include "../utils.hpp" int main() { using InputType = float; const int64_t n = 4; const int64_t c = 64; const int64_t h = 16; const int64_t w = 16; const int64_t k = 32; const int64_t r = 3; const int64_t s = 3; auto buildSubMulMulAddConvbwdRelubwdBnwrwGraph = [=](hipdnnHandle_t handle) { auto graph = std::make_shared(); graph->set_name("sub_mul_mul_add_convbwd_relubwd_bnwrw_graph") .set_io_data_type(hipdnn_frontend::getDataTypeEnumFromType()) .set_intermediate_data_type(hipdnn_frontend::getDataTypeEnumFromType()) .set_compute_data_type(hipdnn_frontend::DataType::FLOAT); auto xBn = std::make_shared( hipdnn_frontend::graph::Tensor_attributes() .set_name("x_bn") .set_dim({n, c, h, w}) .set_stride({c * h * w, 1, c * w, c})); auto meanBn = std::make_shared( hipdnn_frontend::graph::Tensor_attributes() .set_name("mean_bn") .set_dim({1, c, 1, 1}) .set_stride({c, 1, c, c})); auto invStdBn = std::make_shared( hipdnn_frontend::graph::Tensor_attributes() .set_name("invstd_bn") .set_dim({1, c, 1, 1}) .set_stride({c, 1, c, c})); auto scaleBn = std::make_shared( hipdnn_frontend::graph::Tensor_attributes() .set_name("scale_bn") .set_dim({1, c, 1, 1}) .set_stride({c, 1, c, c})); auto biasBn = std::make_shared( hipdnn_frontend::graph::Tensor_attributes() .set_name("bias_bn") .set_dim({1, c, 1, 1}) .set_stride({c, 1, c, c})); auto dy = std::make_shared( hipdnn_frontend::graph::Tensor_attributes() .set_name("dy") .set_dim({n, k, h, w}) .set_stride({k * h * w, 1, k * w, k})); auto filter = std::make_shared( hipdnn_frontend::graph::Tensor_attributes() .set_name("filter") .set_dim({k, c, r, s}) .set_stride({c * r * s, 1, c * s, c})); auto subAttrs = hipdnn_frontend::graph::PointwiseAttributes() .set_name("sub_node") .set_mode(hipdnn_frontend::PointwiseMode::SUB); auto subOut = graph->pointwise(xBn, meanBn, subAttrs); auto mulAttrs0 = hipdnn_frontend::graph::PointwiseAttributes() .set_name("mul0_node") .set_mode(hipdnn_frontend::PointwiseMode::MUL); auto mulOut0 = graph->pointwise(subOut, invStdBn, mulAttrs0); auto mulAttrs1 = hipdnn_frontend::graph::PointwiseAttributes() .set_name("mul1_node") .set_mode(hipdnn_frontend::PointwiseMode::MUL); auto mulOut1 = graph->pointwise(mulOut0, scaleBn, mulAttrs1); auto addAttrs = hipdnn_frontend::graph::PointwiseAttributes() .set_name("add_node") .set_mode(hipdnn_frontend::PointwiseMode::ADD); auto addOut = graph->pointwise(mulOut1, biasBn, addAttrs); auto convAttrs = hipdnn_frontend::graph::ConvDgradAttributes() .set_name("conv_dgrad_node") .set_padding({1, 1}) .set_stride({1, 1}) .set_dilation({1, 1}); auto dx = graph->conv_dgrad(dy, filter, convAttrs); auto reluBwdAttrs = hipdnn_frontend::graph::PointwiseAttributes() .set_name("relu_bwd_node") .set_mode(hipdnn_frontend::PointwiseMode::RELU_BWD); auto dRelu = graph->pointwise(dx, addOut, reluBwdAttrs); dRelu->set_output(true); auto bnWgradAttrs = hipdnn_frontend::graph::BatchnormBackwardWeightAttributes().set_name( "bn_backward_weight_node"); auto [dscale, dbias, eqScaleDy, eqScaleX, eqBias] = graph->dbn_weight(dRelu, xBn, meanBn, invStdBn, scaleBn, bnWgradAttrs); dscale->set_output(true); dbias->set_output(true); eqScaleDy->set_output(true); eqScaleX->set_output(true); eqBias->set_output(true); HIPDNN_FE_CHECK(graph->build(handle)); return std::make_tuple(graph, xBn, meanBn, invStdBn, scaleBn, biasBn, dy, filter, dRelu, dscale, dbias, eqScaleDy, eqScaleX, eqBias); }; auto backend = hipdnn_frontend::detail::hipdnnBackend(); if(!backend) { std::cout << "Creat backend failed. \n"; return 1; } hipdnnHandle_t handle; HIPDNN_CHECK(backend->create(&handle)); auto [graph, xBn, meanBn, invStdBn, scaleBn, biasBn, dy, filter, dRelu, dscale, dbias, eqScaleDy, eqScaleX, eqBias] = buildSubMulMulAddConvbwdRelubwdBnwrwGraph(handle); hipdnn_data_sdk::utilities::Tensor xBnTensor(xBn->get_dim(), xBn->get_stride()); hipdnn_data_sdk::utilities::Tensor meanBnTensor(meanBn->get_dim(), meanBn->get_stride()); hipdnn_data_sdk::utilities::Tensor invStdBnTensor(invStdBn->get_dim(), invStdBn->get_stride()); hipdnn_data_sdk::utilities::Tensor scaleBnTensor(scaleBn->get_dim(), scaleBn->get_stride()); hipdnn_data_sdk::utilities::Tensor biasBnTensor(biasBn->get_dim(), biasBn->get_stride()); hipdnn_data_sdk::utilities::Tensor dyTensor(dy->get_dim(), dy->get_stride()); hipdnn_data_sdk::utilities::Tensor filterTensor(filter->get_dim(), filter->get_stride()); hipdnn_data_sdk::utilities::Tensor dReluTensor(dRelu->get_dim(), dRelu->get_stride()); hipdnn_data_sdk::utilities::Tensor dscaleTensor(dscale->get_dim(), dscale->get_stride()); hipdnn_data_sdk::utilities::Tensor dbiasTensor(dbias->get_dim(), dbias->get_stride()); hipdnn_data_sdk::utilities::Tensor eqScaleDyTensor(eqScaleDy->get_dim(), eqScaleDy->get_stride()); hipdnn_data_sdk::utilities::Tensor eqScaleXTensor(eqScaleX->get_dim(), eqScaleX->get_stride()); hipdnn_data_sdk::utilities::Tensor eqBiasTensor(eqBias->get_dim(), eqBias->get_stride()); std::unordered_map variantPack; variantPack[xBn->get_uid()] = xBnTensor.memory().deviceData(); variantPack[meanBn->get_uid()] = meanBnTensor.memory().deviceData(); variantPack[invStdBn->get_uid()] = invStdBnTensor.memory().deviceData(); variantPack[scaleBn->get_uid()] = scaleBnTensor.memory().deviceData(); variantPack[biasBn->get_uid()] = biasBnTensor.memory().deviceData(); variantPack[dy->get_uid()] = dyTensor.memory().deviceData(); variantPack[filter->get_uid()] = filterTensor.memory().deviceData(); variantPack[dRelu->get_uid()] = dReluTensor.memory().deviceData(); variantPack[dscale->get_uid()] = dscaleTensor.memory().deviceData(); variantPack[dbias->get_uid()] = dbiasTensor.memory().deviceData(); variantPack[eqScaleDy->get_uid()] = eqScaleDyTensor.memory().deviceData(); variantPack[eqScaleX->get_uid()] = eqScaleXTensor.memory().deviceData(); variantPack[eqBias->get_uid()] = eqBiasTensor.memory().deviceData(); int64_t workspaceSize = 0; HIPDNN_FE_CHECK(graph->get_workspace_size(workspaceSize)); const hipdnn_data_sdk::utilities::Workspace workspace(static_cast(workspaceSize)); HIPDNN_FE_CHECK(graph->execute(handle, variantPack, workspace.get())); std::cout << "SubMulMulAddConvbwdRelubwdBnwrw graph execution complete. \n"; HIPDNN_CHECK(backend->destroy(handle)); return 0; }