TheRock/rocm-libraries/dnn-providers/integration-tests/tests/TestSynthesizeInputs.cpp

// Copyright © Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT

#include <gtest/gtest.h>

#include <cstdint>
#include <memory>
#include <set>
#include <vector>

#include <hipdnn_data_sdk/utilities/Tensor.hpp>
#include <hipdnn_flatbuffers_sdk/data_objects/graph_generated.h>

#include "harness/input-init/SynthesizeInputs.hpp"
#include <hipdnn_test_sdk/utilities/FlatbufferGraphTestUtils.hpp>
#include <hipdnn_test_sdk/utilities/detail/FlatbufferTensorAttributesUtils.hpp>

// NOLINTBEGIN(readability-identifier-naming)

using namespace hipdnn_flatbuffers_sdk::data_objects;
using namespace hipdnn_integration_tests;

namespace
{

const std::vector<int64_t> kDims = {2, 3};
const std::vector<int64_t> kStrides = {3, 1};

InputTensorMap makeTensors(const std::vector<int64_t>& uids)
{
    InputTensorMap map;
    for(const int64_t uid : uids)
    {
        map[uid] = std::make_unique<hipdnn_data_sdk::utilities::Tensor<float>>(kDims, kStrides);
        map[uid]->fillTensorWithValue(0.f);
    }
    return map;
}

struct GraphResult
{
    flatbuffers::FlatBufferBuilder builder;
    const Graph* graph = nullptr;

    const Node& node(uint32_t i) const
    {
        return *graph->nodes()->Get(i);
    }

    std::vector<int64_t> leafInputUids(const std::set<int64_t>& outputUids) const
    {
        std::vector<int64_t> uids;
        for(const auto* t : *graph->tensors())
        {
            if(!t->virtual_() && outputUids.count(t->uid()) == 0)
            {
                uids.push_back(t->uid());
            }
        }
        return uids;
    }
};

// ── Conv fwd (single node) ──────────────────────────────────────────────────

GraphResult buildConvFwdGraph()
{
    GraphResult r;
    auto& b = r.builder;

    std::vector<flatbuffers::Offset<TensorAttributes>> tensors;
    tensors.push_back(CreateTensorAttributesDirect(b, 1, "x", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 2, "w", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 3, "y", DataType::FLOAT, &kStrides, &kDims));

    auto conv = CreateConvolutionFwdAttributesDirect(b, 1, 2, 3);

    std::vector<flatbuffers::Offset<Node>> nodes;
    nodes.push_back(CreateNodeDirect(
        b, "conv", DataType::FLOAT, NodeAttributes::ConvolutionFwdAttributes, conv.Union()));

    auto graph = CreateGraphDirect(
        b, "test", DataType::FLOAT, DataType::FLOAT, DataType::FLOAT, &tensors, &nodes);
    b.Finish(graph);

    r.graph = GetGraph(b.GetBufferPointer());
    return r;
}

// ── Conv + bias (2-node fused) ──────────────────────────────────────────────
// conv.y (uid 10) is virtual; bias (uid 4) is leaf

GraphResult buildConvBiasGraph()
{
    GraphResult r;
    auto& b = r.builder;

    std::vector<flatbuffers::Offset<TensorAttributes>> tensors;
    tensors.push_back(CreateTensorAttributesDirect(b, 1, "x", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 2, "w", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 10, "conv_y", DataType::FLOAT, &kStrides, &kDims, true));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 4, "bias", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 5, "out", DataType::FLOAT, &kStrides, &kDims));

    auto conv = CreateConvolutionFwdAttributesDirect(b, 1, 2, 10);
    auto add = CreatePointwiseAttributes(b,
                                         PointwiseMode::ADD,
                                         flatbuffers::nullopt,
                                         flatbuffers::nullopt,
                                         flatbuffers::nullopt,
                                         flatbuffers::nullopt,
                                         10,
                                         4,
                                         flatbuffers::nullopt,
                                         5);

    std::vector<flatbuffers::Offset<Node>> nodes;
    nodes.push_back(CreateNodeDirect(
        b, "conv", DataType::FLOAT, NodeAttributes::ConvolutionFwdAttributes, conv.Union()));
    nodes.push_back(CreateNodeDirect(
        b, "bias_add", DataType::FLOAT, NodeAttributes::PointwiseAttributes, add.Union()));

    auto graph = CreateGraphDirect(
        b, "test", DataType::FLOAT, DataType::FLOAT, DataType::FLOAT, &tensors, &nodes);
    b.Finish(graph);

    r.graph = GetGraph(b.GetBufferPointer());
    return r;
}

// ── Conv + bias + relu (3-node fused) ───────────────────────────────────────
// conv.y (uid 10) virtual, bias_add.out (uid 11) virtual, relu.in_0=uid 11, relu.out_0=uid 6

GraphResult buildConvBiasReluGraph()
{
    GraphResult r;
    auto& b = r.builder;

    std::vector<flatbuffers::Offset<TensorAttributes>> tensors;
    tensors.push_back(CreateTensorAttributesDirect(b, 1, "x", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 2, "w", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 10, "conv_y", DataType::FLOAT, &kStrides, &kDims, true));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 4, "bias", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 11, "bias_out", DataType::FLOAT, &kStrides, &kDims, true));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 6, "out", DataType::FLOAT, &kStrides, &kDims));

    auto conv = CreateConvolutionFwdAttributesDirect(b, 1, 2, 10);
    auto add = CreatePointwiseAttributes(b,
                                         PointwiseMode::ADD,
                                         flatbuffers::nullopt,
                                         flatbuffers::nullopt,
                                         flatbuffers::nullopt,
                                         flatbuffers::nullopt,
                                         10,
                                         4,
                                         flatbuffers::nullopt,
                                         11);
    auto relu = CreatePointwiseAttributes(b,
                                          PointwiseMode::RELU_FWD,
                                          flatbuffers::nullopt,
                                          flatbuffers::nullopt,
                                          flatbuffers::nullopt,
                                          flatbuffers::nullopt,
                                          11,
                                          flatbuffers::nullopt,
                                          flatbuffers::nullopt,
                                          6);

    std::vector<flatbuffers::Offset<Node>> nodes;
    nodes.push_back(CreateNodeDirect(
        b, "conv", DataType::FLOAT, NodeAttributes::ConvolutionFwdAttributes, conv.Union()));
    nodes.push_back(CreateNodeDirect(
        b, "bias_add", DataType::FLOAT, NodeAttributes::PointwiseAttributes, add.Union()));
    nodes.push_back(CreateNodeDirect(
        b, "relu", DataType::FLOAT, NodeAttributes::PointwiseAttributes, relu.Union()));

    auto graph = CreateGraphDirect(
        b, "test", DataType::FLOAT, DataType::FLOAT, DataType::FLOAT, &tensors, &nodes);
    b.Finish(graph);

    r.graph = GetGraph(b.GetBufferPointer());
    return r;
}

// ── Batchnorm training with runtime PBV scalars ─────────────────────────────
// uids: x=1, y=2, scale=3, bias=4, epsilon=5, prev_mean=8, prev_variance=9, momentum=10
GraphResult buildBatchnormTrainingRuntimePbvGraph()
{
    GraphResult result;
    result.builder = hipdnn_test_sdk::utilities::createValidBatchnormFwdTrainingGraph(
        kStrides,
        kDims,
        /*withMeanVariance=*/false,
        /*overrideShapeEnabled=*/false,
        /*runtimeEpsilon=*/true,
        /*withRunningStatsAndMomentum=*/true,
        /*runtimeMomentum=*/true);
    result.graph = GetGraph(result.builder.GetBufferPointer());
    return result;
}

InputTensorMap makeTensorsFromGraph(const GraphResult& gr, const std::vector<int64_t>& uids)
{
    std::unordered_map<int64_t, const TensorAttributes*> attributes;
    for(const auto* tensor : *gr.graph->tensors())
    {
        attributes.emplace(tensor->uid(), tensor);
    }

    InputTensorMap inputs;
    for(const int64_t uid : uids)
    {
        inputs.emplace(uid,
                       hipdnn_test_sdk::detail::createTensorFromAttribute(*attributes.at(uid)));
    }
    return inputs;
}

float scalarValue(const InputTensorMap& inputs, int64_t uid)
{
    return *static_cast<const float*>(inputs.at(uid)->rawHostData());
}

// ── SDPA forward (no structured optionals) ──────────────────────────────────

GraphResult buildSdpaFwdGraph()
{
    GraphResult r;
    auto& b = r.builder;

    std::vector<flatbuffers::Offset<TensorAttributes>> tensors;
    tensors.push_back(CreateTensorAttributesDirect(b, 1, "q", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 2, "k", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 3, "v", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 4, "o", DataType::FLOAT, &kStrides, &kDims));

    auto sdpa = CreateSdpaAttributes(b, 1, 2, 3, 4);

    std::vector<flatbuffers::Offset<Node>> nodes;
    nodes.push_back(CreateNodeDirect(
        b, "sdpa_fwd", DataType::FLOAT, NodeAttributes::SdpaAttributes, sdpa.Union()));

    auto graph = CreateGraphDirect(
        b, "test", DataType::FLOAT, DataType::FLOAT, DataType::FLOAT, &tensors, &nodes);
    b.Finish(graph);

    r.graph = GetGraph(b.GetBufferPointer());
    return r;
}

// ── SDPA forward with structured seq_len_q ──────────────────────────────────

GraphResult buildSdpaFwdWithStructuredGraph()
{
    GraphResult r;
    auto& b = r.builder;

    std::vector<flatbuffers::Offset<TensorAttributes>> tensors;
    tensors.push_back(CreateTensorAttributesDirect(b, 1, "q", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 2, "k", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 3, "v", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 4, "o", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 5, "seq_len_q", DataType::FLOAT, &kStrides, &kDims));

    auto sdpa = CreateSdpaAttributes(b,
                                     1,
                                     2,
                                     3,
                                     4,
                                     flatbuffers::nullopt, // attn_mask
                                     flatbuffers::nullopt, // scale
                                     5); // seq_len_q

    std::vector<flatbuffers::Offset<Node>> nodes;
    nodes.push_back(CreateNodeDirect(
        b, "sdpa_fwd", DataType::FLOAT, NodeAttributes::SdpaAttributes, sdpa.Union()));

    auto graph = CreateGraphDirect(
        b, "test", DataType::FLOAT, DataType::FLOAT, DataType::FLOAT, &tensors, &nodes);
    b.Finish(graph);

    r.graph = GetGraph(b.GetBufferPointer());
    return r;
}

// ── SDPA backward standalone ────────────────────────────────────────────────
// O and stats are leaf inputs (not virtual) → DERIVED → refuses

GraphResult buildSdpaBwdStandaloneGraph()
{
    GraphResult r;
    auto& b = r.builder;

    std::vector<flatbuffers::Offset<TensorAttributes>> tensors;
    tensors.push_back(CreateTensorAttributesDirect(b, 1, "q", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 2, "k", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 3, "v", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 4, "o", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 5, "do", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 6, "stats", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 7, "dq", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 8, "dk", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 9, "dv", DataType::FLOAT, &kStrides, &kDims));

    auto bwd = CreateSdpaBackwardAttributes(b, 1, 2, 3, 4, 5, 6, 7, 8, 9);

    std::vector<flatbuffers::Offset<Node>> nodes;
    nodes.push_back(CreateNodeDirect(
        b, "sdpa_bwd", DataType::FLOAT, NodeAttributes::SdpaBackwardAttributes, bwd.Union()));

    auto graph = CreateGraphDirect(
        b, "test", DataType::FLOAT, DataType::FLOAT, DataType::FLOAT, &tensors, &nodes);
    b.Finish(graph);

    r.graph = GetGraph(b.GetBufferPointer());
    return r;
}

// ── SDPA fwd+bwd fused ─────────────────────────────────────────────────────
// O (uid 10) and stats (uid 11) are virtual inter-node tensors.
// Leaf inputs: Q(1), K(2), V(3) from fwd + dO(5) from bwd.
// Outputs: dQ(7), dK(8), dV(9).

GraphResult buildSdpaFwdBwdFusedGraph()
{
    GraphResult r;
    auto& b = r.builder;

    std::vector<flatbuffers::Offset<TensorAttributes>> tensors;
    tensors.push_back(CreateTensorAttributesDirect(b, 1, "q", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 2, "k", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 3, "v", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 10, "o", DataType::FLOAT, &kStrides, &kDims, true));
    tensors.push_back(
        CreateTensorAttributesDirect(b, 11, "stats", DataType::FLOAT, &kStrides, &kDims, true));
    tensors.push_back(CreateTensorAttributesDirect(b, 5, "do", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 7, "dq", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 8, "dk", DataType::FLOAT, &kStrides, &kDims));
    tensors.push_back(CreateTensorAttributesDirect(b, 9, "dv", DataType::FLOAT, &kStrides, &kDims));

    auto fwd = CreateSdpaAttributes(b,
                                    1,
                                    2,
                                    3,
                                    10,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    flatbuffers::nullopt,
                                    11); // stats_tensor_uid

    auto bwd = CreateSdpaBackwardAttributes(b, 1, 2, 3, 10, 5, 11, 7, 8, 9);

    std::vector<flatbuffers::Offset<Node>> nodes;
    nodes.push_back(CreateNodeDirect(
        b, "sdpa_fwd", DataType::FLOAT, NodeAttributes::SdpaAttributes, fwd.Union()));
    nodes.push_back(CreateNodeDirect(
        b, "sdpa_bwd", DataType::FLOAT, NodeAttributes::SdpaBackwardAttributes, bwd.Union()));

    auto graph = CreateGraphDirect(
        b, "test", DataType::FLOAT, DataType::FLOAT, DataType::FLOAT, &tensors, &nodes);
    b.Finish(graph);

    r.graph = GetGraph(b.GetBufferPointer());
    return r;
}

SynthesisResult runSynthesis(const GraphResult& gr, const std::set<int64_t>& outputUids)
{
    const auto leafUids = gr.leafInputUids(outputUids);
    auto inputs = makeTensors(leafUids);
    SynthesisConfig config;

    synthesizeInputs(*gr.graph, inputs, leafUids, config);

    auto missing = config.unfilled(leafUids);
    if(!missing.empty())
    {
        std::string msg = "cannot synthesize:";
        for(const int64_t uid : missing)
        {
            const auto init = config.fill(uid);
            const char* kind = init.kind == FillSpec::Kind::STRUCTURED ? "structured" : "derived";
            msg += " uid=" + std::to_string(uid) + " (" + kind + ")";
        }
        return SynthesisResult::unsupported(msg);
    }

    return SynthesisResult::ok();
}

} // namespace

// ── Test cases ──────────────────────────────────────────────────────────────

TEST(TestSynthesizeInputs, SingleConvFwd)
{
    const auto gr = buildConvFwdGraph();
    const auto result = runSynthesis(gr, {3});

    EXPECT_TRUE(result.filled) << result.reason;
}

TEST(TestSynthesizeInputs, ConvPlusBiasFused)
{
    const auto gr = buildConvBiasGraph();
    const auto result = runSynthesis(gr, {5});

    EXPECT_TRUE(result.filled) << result.reason;
}

TEST(TestSynthesizeInputs, ConvPlusBiasPlusReluFused)
{
    const auto gr = buildConvBiasReluGraph();
    const auto result = runSynthesis(gr, {6});

    EXPECT_TRUE(result.filled) << result.reason;
}

// Unit seam: verify the synthesis policy itself. Epsilon is the fixed-value
// path; momentum is the seeded random path and must reproduce across runs.
TEST(TestSynthesizeInputs, RuntimePbvScalarsUseFixedAndDeterministicRandomFills)
{
    const auto graph = buildBatchnormTrainingRuntimePbvGraph();
    const std::vector<int64_t> leafUids = {1, 3, 4, 5, 8, 9, 10};

    auto firstInputs = makeTensorsFromGraph(graph, leafUids);
    SynthesisConfig firstConfig;
    const auto firstResult = synthesizeInputs(*graph.graph, firstInputs, leafUids, firstConfig);
    ASSERT_TRUE(firstResult.filled) << firstResult.reason;

    EXPECT_FLOAT_EQ(scalarValue(firstInputs, 5), 1e-5f);
    const float firstMomentum = scalarValue(firstInputs, 10);
    EXPECT_GE(firstMomentum, 0.0f);
    EXPECT_LE(firstMomentum, 1.0f);

    auto secondInputs = makeTensorsFromGraph(graph, leafUids);
    SynthesisConfig secondConfig;
    const auto secondResult = synthesizeInputs(*graph.graph, secondInputs, leafUids, secondConfig);
    ASSERT_TRUE(secondResult.filled) << secondResult.reason;
    EXPECT_FLOAT_EQ(scalarValue(secondInputs, 10), firstMomentum);
}

TEST(TestSynthesizeInputs, SdpaFwdNoStructuredOptionals)
{
    const auto gr = buildSdpaFwdGraph();
    const auto result = runSynthesis(gr, {4});

    EXPECT_TRUE(result.filled) << result.reason;
}

TEST(TestSynthesizeInputs, SdpaFwdWithStructuredInputRefuses)
{
    const auto gr = buildSdpaFwdWithStructuredGraph();
    const auto result = runSynthesis(gr, {4});

    EXPECT_FALSE(result.filled);
    EXPECT_NE(result.reason.find("uid=5"), std::string::npos);
    EXPECT_NE(result.reason.find("structured"), std::string::npos);
}

TEST(TestSynthesizeInputs, SdpaBwdStandaloneRefusesDerived)
{
    const auto gr = buildSdpaBwdStandaloneGraph();
    const auto result = runSynthesis(gr, {7, 8, 9});

    EXPECT_FALSE(result.filled);
    EXPECT_NE(result.reason.find("derived"), std::string::npos);
}

TEST(TestSynthesizeInputs, SdpaFwdBwdFusedSucceeds)
{
    const auto gr = buildSdpaFwdBwdFusedGraph();
    const auto result = runSynthesis(gr, {7, 8, 9});

    EXPECT_TRUE(result.filled) << result.reason;
}

// NOLINTEND(readability-identifier-naming)