render_delegate.cpp

//
// SPDX-License-Identifier: Apache-2.0
//

// Copyright 2019 Luma Pictures
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// Modifications Copyright 2022 Autodesk, Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "render_delegate.h"
#include "reader.h"

#include <pxr/base/tf/getenv.h>
#include <pxr/base/tf/envSetting.h>

#include <pxr/imaging/hd/bprim.h>
#include <pxr/imaging/hd/camera.h>
#include <pxr/imaging/hd/extComputation.h>
#include <pxr/imaging/hd/instancer.h>
#include <pxr/imaging/hd/resourceRegistry.h>
#include <pxr/imaging/hd/rprim.h>
#include <pxr/imaging/hd/tokens.h>
#ifdef ENABLE_SCENE_INDEX
#include <pxr/imaging/hd/dirtyBitsTranslator.h>
#include <pxr/imaging/hd/retainedDataSource.h>
#endif
#include <common_utils.h>
#include <constant_strings.h>
#include <shape_utils.h>

#include "basis_curves.h"
#include "camera.h"
#include "config.h"
#include "gaussian_splat.h"
#include "instancer.h"
#include "light.h"
#include "mesh.h"
#include "native_rprim.h"
#include "node_graph.h"
#include "nodes/nodes.h"
#include "openvdb_asset.h"
#include "options.h"
#include "points.h"
#include "procedural_custom.h"
#include "render_buffer.h"
#include "render_pass.h"
#include "volume.h"
#include <cctype>

#ifdef ENABLE_HYDRA2_RENDERSETTINGS
#include "render_settings.h"
#endif

PXR_NAMESPACE_OPEN_SCOPE

// clang-format off
TF_DEFINE_PRIVATE_TOKENS(_tokens,
    (arnold)
    ((aovDriverFormat, "driver:parameters:aov:format"))
    ((aovFormat, "arnold:format"))
    (ArnoldOptions)
    (openvdbAsset)
    ((arnoldGlobal, "arnold:global:"))
    ((arnoldDriver, "arnold:driver"))
    ((arnoldNamespace, "arnold:"))
    ((colorManagerNamespace, "color_manager:"))
    (batchCommandLine)
    (percentDone)
    (totalClockTime)
    (renderProgressAnnotation)
    (delegateRenderProducts)
    (orderedVars)
    ((aovSettings, "aovDescriptor.aovSettings"))
    (productType)
    (productName)
    (pixelAspectRatio)
    (driver_exr)
    (sourceType)
    (sourceName)
    (dataType)
    (huskErrorStatus)
    ((format, "aovDescriptor.format"))
    ((clearValue, "aovDescriptor.clearValue"))
    ((multiSampled, "aovDescriptor.multiSampled"))
    ((aovName, "driver:parameters:aov:name"))
    (deep)
    (raw)
    (instantaneousShutter)
    ((aovShadersArray, "aov_shaders:i"))
    (GeometryLight)
    (dataWindowNDC)
    (resolution)
    // The following tokens are also defined in read_options.cpp, we need them
    // here for the conversion from TfToken to HdFormat, while in read_options they
    // are used for the conversion of HdFormat to TfToken.
    ((_float, "float"))
    ((_int, "int"))
    (i8) (int8)
    (ui8) (uint8)
    (half) (float16)
    (float2) (float3) (float4)
    (half2) (half3) (half4)
    (color2f) (color3f) (color4f)
    (color2h) (color3h) (color4h)
    (color2u8) (color3u8) (color4u8)
    (color2i8) (color3i8) (color4i8)
    (int2) (int3) (int4)
    (uint2) (uint3) (uint4)
);
// clang-format on

#define PXR_VERSION_STR \
    ARNOLD_XSTR(PXR_MAJOR_VERSION) "." ARNOLD_XSTR(PXR_MINOR_VERSION) "." ARNOLD_XSTR(PXR_PATCH_VERSION)

TF_DEFINE_ENV_SETTING(HDARNOLD_SHAPE_INSTANCING, "", "Set to 0 to disable inner shape instancing");

namespace {

const HdFormat _GetHdFormatFromToken(const TfToken& token)
{
    if (token == _tokens->uint8) {
        return HdFormatUNorm8;
    } else if (token == _tokens->color2u8) {
        return HdFormatUNorm8Vec2;
    } else if (token == _tokens->color3u8) {
        return HdFormatUNorm8Vec3;
    } else if (token == _tokens->color4u8) {
        return HdFormatUNorm8Vec4;
    } else if (token == _tokens->int8) {
        return HdFormatSNorm8;
    } else if (token == _tokens->color2i8) {
        return HdFormatSNorm8Vec2;
    } else if (token == _tokens->color3i8) {
        return HdFormatSNorm8Vec3;
    } else if (token == _tokens->color4i8) {
        return HdFormatSNorm8Vec4;
    } else if (token == _tokens->half) {
        return HdFormatFloat16;
    } else if (token == _tokens->half2 || token == _tokens->color2h) {
        return HdFormatFloat16Vec2;
    } else if (token == _tokens->half3 || token == _tokens->color3h) {
        return HdFormatFloat16Vec3;
    } else if (token == _tokens->half4 || token == _tokens->color4h) {
        return HdFormatFloat16Vec4;
    } else if (token == _tokens->_float) {
        return HdFormatFloat32;
    } else if (token == _tokens->float2 || token == _tokens->color2f) {
        return HdFormatFloat32Vec2;
    } else if (token == _tokens->float3 || token == _tokens->color3f) {
        return HdFormatFloat32Vec3;
    } else if (token == _tokens->float4 || token == _tokens->color4f) {
        return HdFormatFloat32Vec4;
    } else if (token == _tokens->_int) {
        return HdFormatInt32;
    } else if (token == _tokens->int2) {
        return HdFormatInt32Vec2;
    } else if (token == _tokens->int3) {
        return HdFormatInt32Vec3;
    } else if (token == _tokens->int4) {
        return HdFormatInt32Vec4;
    } else {
        return HdFormatInvalid;
    }
}

VtValue _GetNodeParamValue(AtNode* node, const AtParamEntry* pentry)
{
    if (Ai_unlikely(pentry == nullptr)) {
        return {};
    }
    const auto ptype = AiParamGetType(pentry);
    if (ptype == AI_TYPE_INT) {
        return VtValue(AiNodeGetInt(node, AiParamGetName(pentry)));
    } else if (ptype == AI_TYPE_FLOAT) {
        return VtValue(AiNodeGetFlt(node, AiParamGetName(pentry)));
    } else if (ptype == AI_TYPE_BOOLEAN) {
        return VtValue(AiNodeGetBool(node, AiParamGetName(pentry)));
    } else if (ptype == AI_TYPE_STRING || ptype == AI_TYPE_ENUM) {
        return VtValue(std::string(AiNodeGetStr(node, AiParamGetName(pentry))));
    }
    return {};
}

void _SetNodeParam(AtNode* node, const TfToken& key, const VtValue& value)
{
    // Some applications might send integers instead of booleans.
    if (value.IsHolding<int>()) {
        const auto* nodeEntry = AiNodeGetNodeEntry(node);
        auto* paramEntry = AiNodeEntryLookUpParameter(nodeEntry, AtString(key.GetText()));
        if (paramEntry != nullptr) {
            const auto paramType = AiParamGetType(paramEntry);
            if (paramType == AI_TYPE_INT) {
                AiNodeSetInt(node, AtString(key.GetText()), value.UncheckedGet<int>());
            } else if (paramType == AI_TYPE_BOOLEAN) {
                AiNodeSetBool(node, AtString(key.GetText()), value.UncheckedGet<int>() != 0);
            }
        }
        // Or longs.
    } else if (value.IsHolding<long>()) {
        const auto* nodeEntry = AiNodeGetNodeEntry(node);
        auto* paramEntry = AiNodeEntryLookUpParameter(nodeEntry, AtString(key.GetText()));
        if (paramEntry != nullptr) {
            const auto paramType = AiParamGetType(paramEntry);
            if (paramType == AI_TYPE_INT) {
                AiNodeSetInt(node, AtString(key.GetText()), static_cast<int>(value.UncheckedGet<long>()));
            } else if (paramType == AI_TYPE_BOOLEAN) {
                AiNodeSetBool(node, AtString(key.GetText()), value.UncheckedGet<long>() != 0);
            }
        }
        // Or long longs.
    } else if (value.IsHolding<long long>()) {
        const auto* nodeEntry = AiNodeGetNodeEntry(node);
        auto* paramEntry = AiNodeEntryLookUpParameter(nodeEntry, AtString(key.GetText()));
        if (paramEntry != nullptr) {
            const auto paramType = AiParamGetType(paramEntry);
            if (paramType == AI_TYPE_INT) {
                AiNodeSetInt(node, AtString(key.GetText()), static_cast<int>(value.UncheckedGet<long long>()));
            } else if (paramType == AI_TYPE_BOOLEAN) {
                AiNodeSetBool(node, AtString(key.GetText()), value.UncheckedGet<long long>() != 0);
            }
        }
    } else if (value.IsHolding<float>()) {
        AiNodeSetFlt(node, AtString(key.GetText()), value.UncheckedGet<float>());
    } else if (value.IsHolding<double>()) {
        AiNodeSetFlt(node, AtString(key.GetText()), static_cast<float>(value.UncheckedGet<double>()));
    } else if (value.IsHolding<bool>()) {
        AiNodeSetBool(node, AtString(key.GetText()), value.UncheckedGet<bool>());
    } else if (value.IsHolding<std::string>()) {
        AiNodeSetStr(node, AtString(key.GetText()), AtString(value.UncheckedGet<std::string>().c_str()));
    } else if (value.IsHolding<TfToken>()) {
        AiNodeSetStr(node, AtString(key.GetText()), AtString(value.UncheckedGet<TfToken>().GetText()));
    }
}


inline const TfTokenVector& _SupportedSprimTypes()
{
    // Hd_PrimTypeIndex::SyncPrims walks types in this order; every prim of type N
    // is fully synced before any prim of type N+1. Light shaders (and filters) can
    // target ArnoldNodeGraph prims, which must therefore appear before light types.
    // Scene-index dependency forwarding dirties lights when graphs change but does
    // not reorder this pass.
    static const TfTokenVector r{HdPrimTypeTokens->camera,
                                 HdPrimTypeTokens->material,
                                 str::t_ArnoldNodeGraph,
                                 HdPrimTypeTokens->distantLight,
                                 HdPrimTypeTokens->sphereLight,
                                 HdPrimTypeTokens->diskLight,
                                 HdPrimTypeTokens->rectLight,
                                 HdPrimTypeTokens->cylinderLight,
                                 HdPrimTypeTokens->domeLight,
#ifdef ENABLE_SCENE_INDEX
                                 HdPrimTypeTokens->meshLight,
#endif
                                 _tokens->GeometryLight,
                                 _tokens->ArnoldOptions,
                                 HdPrimTypeTokens->extComputation
                                 /*HdPrimTypeTokens->simpleLight*/};
    return r;
}


inline const TfTokenVector& _SupportedBprimTypes(bool ownsUniverse)
{
    // For the hydra render delegate plugin, when we own the arnold universe, we don't want 
    // to support the render settings primitives as Bprims since it will be passed through SetRenderSettings

#if PXR_VERSION >= 2208
    if (!ownsUniverse) {
        static const TfTokenVector r{HdPrimTypeTokens->renderBuffer, _tokens->openvdbAsset, HdPrimTypeTokens->renderSettings};
        return r;
    } else
#endif
    {
#ifdef ENABLE_HYDRA2_RENDERSETTINGS
        static const TfTokenVector r{HdPrimTypeTokens->renderBuffer, _tokens->openvdbAsset, HdPrimTypeTokens->renderSettings};
#else
        static const TfTokenVector r{HdPrimTypeTokens->renderBuffer, _tokens->openvdbAsset};
#endif
        return r;
    }
}

struct SupportedRenderSetting {
    /// Constructor with no default value.
    SupportedRenderSetting(const char* _label) : label(_label) {}

    /// Constructor with a default value.
    template <typename T>
    SupportedRenderSetting(const char* _label, const T& _defaultValue) : label(_label), defaultValue(_defaultValue)
    {
    }

    TfToken label;
    VtValue defaultValue;
};

using SupportedRenderSettings = std::vector<std::pair<TfToken, SupportedRenderSetting>>;
using VtStringArray = VtArray<std::string>;
const SupportedRenderSettings& _GetSupportedRenderSettings()
{
    static const auto& config = HdArnoldConfig::GetInstance();
    static const SupportedRenderSettings data{
        // Global settings to control rendering
        {str::t_enable_progressive_render, {"Enable Progressive Render", config.enable_progressive_render}},
        {str::t_progressive_min_AA_samples,
         {"Progressive Render Minimum AA Samples", config.progressive_min_AA_samples}},
        {str::t_enable_adaptive_sampling, {"Enable Adaptive Sampling", config.enable_adaptive_sampling}},
#ifndef __APPLE__
        {str::t_enable_gpu_rendering, {"Enable GPU Rendering", config.enable_gpu_rendering}},
#endif
        {str::t_interactive_target_fps, {"Target FPS for Interactive Rendering", config.interactive_target_fps}},
        {str::t_interactive_target_fps_min,
         {"Minimum Target FPS for Interactive Rendering", config.interactive_target_fps_min}},
        {str::t_interactive_fps_min, {"Minimum FPS for Interactive Rendering", config.interactive_fps_min}},
        // Threading settings
        {str::t_threads, {"Number of Threads", config.threads}},
        // Sampling settings
        {str::t_AA_samples, {"AA Samples", config.AA_samples}},
        {str::t_AA_samples_max, {"AA Samples Max"}},
        {str::t_GI_diffuse_samples, {"Diffuse Samples", config.GI_diffuse_samples}},
        {str::t_GI_specular_samples, {"Specular Samples", config.GI_specular_samples}},
        {str::t_GI_transmission_samples, {"Transmission Samples", config.GI_transmission_samples}},
        {str::t_GI_sss_samples, {"SubSurface Scattering Samples", config.GI_sss_samples}},
        {str::t_GI_volume_samples, {"Volume Samples", config.GI_volume_samples}},
        // Depth settings
        {str::t_auto_transparency_depth, {"Auto Transparency Depth"}},
        {str::t_GI_diffuse_depth, {"Diffuse Depth", config.GI_diffuse_depth}},
        {str::t_GI_specular_depth, {"Specular Depth", config.GI_specular_depth}},
        {str::t_GI_transmission_depth, {"Transmission Depth", config.GI_transmission_depth}},
        {str::t_GI_volume_depth, {"Volume Depth"}},
        {str::t_GI_total_depth, {"Total Depth"}},
        // Ignore settings
        {str::t_abort_on_error, {"Abort On Error", config.abort_on_error}},
        {str::t_ignore_textures, {"Ignore Textures"}},
        {str::t_ignore_shaders, {"Ignore Shaders"}},
        {str::t_ignore_atmosphere, {"Ignore Atmosphere"}},
        {str::t_ignore_lights, {"Ignore Lights"}},
        {str::t_ignore_shadows, {"Ignore Shadows"}},
        {str::t_ignore_subdivision, {"Ignore Subdivision"}},
        {str::t_ignore_displacement, {"Ignore Displacement"}},
        {str::t_ignore_bump, {"Ignore Bump"}},
        {str::t_ignore_motion, {"Ignore Motion"}},
        {str::t_ignore_motion_blur, {"Ignore Motion Blur"}},
        {str::t_ignore_dof, {"Ignore Depth of Field"}},
        {str::t_ignore_smoothing, {"Ignore Smoothing"}},
        {str::t_ignore_sss, {"Ignore SubSurface Scattering"}},
        {str::t_ignore_operators, {"Ignore Operators"}},
        // HTML Report Settings
        {str::t_report_file, {"HTML Report Path", config.report_file}},
        // Log Settings
        {str::t_log_verbosity, {"Log Verbosity (0-5)", config.log_verbosity}},
        {str::t_log_file, {"Log File Path", config.log_file}},
        // Profiling Settings
        {str::t_profile_file, {"File Output for Profiling", config.profile_file}},
        // Stats Settings
        {str::t_stats_file, {"File Output for Stats", config.stats_file}},
        // Search paths
        {str::t_plugin_searchpath, {"Plugin search path.", config.plugin_searchpath}},
#if ARNOLD_VERSION_NUM <= 70403
        {str::t_plugin_searchpath, {"Plugin search path.", config.plugin_searchpath}},
        {str::t_procedural_searchpath, {"Procedural search path.", config.procedural_searchpath}},
#else
        {str::t_asset_searchpath, {"Asset search path.", config.asset_searchpath}},
#endif
        {str::t_osl_includepath, {"OSL include path.", config.osl_includepath}},

        {str::t_subdiv_dicing_camera, {"Subdiv Dicing Camera", std::string{}}},
        {str::t_subdiv_frustum_culling, {"Subdiv Frustum Culling"}},
        {str::t_subdiv_frustum_padding, {"Subdiv Frustum Padding"}},

        {str::t_shader_override, {"Path to the shader_override node graph", std::string{}}},
        {str::t_background, {"Path to the background node graph.", std::string{}}},
        {str::t_atmosphere, {"Path to the atmosphere node graph.", std::string{}}},
        {str::t_aov_shaders, {"Path to the aov_shaders node graph.", std::string{}}},
        {str::t_imager, {"Path to the imagers node graph.", std::string{}}},
        {str::t_texture_auto_generate_tx, {"Auto-generate Textures to TX", config.auto_generate_tx}},
    };
    return data;
}

int _GetLogFlagsFromVerbosity(int verbosity)
{
    if (verbosity <= 0) {
        return 0;
    }
    if (verbosity >= 5) {
        return AI_LOG_ALL & ~AI_LOG_COLOR;
    }

    int flags = AI_LOG_ERRORS | AI_LOG_TIMESTAMP | AI_LOG_MEMORY | AI_LOG_BACKTRACE;

    if (verbosity >= 2) {
        flags |= AI_LOG_WARNINGS;
        if (verbosity >= 3) {
            // Don't want progress without info, as otherwise it never prints a
            // "render done" message!
            flags |= AI_LOG_INFO | AI_LOG_PROGRESS;
            if (verbosity >= 4) {
                flags |= AI_LOG_STATS | AI_LOG_PLUGINS;
            }
        }
    }
    return flags;
}

template <typename F>
void _CheckForBoolValue(const VtValue& value, F&& f)
{
    if (value.IsHolding<bool>()) {
        f(value.UncheckedGet<bool>());
    } else if (value.IsHolding<int>()) {
        f(value.UncheckedGet<int>() != 0);
    } else if (value.IsHolding<long>()) {
        f(value.UncheckedGet<long>() != 0);
    } else if (value.IsHolding<long long>()) {
        f(value.UncheckedGet<long long>() != 0);
    }
}

template <typename F>
void _CheckForIntValue(const VtValue& value, F&& f)
{
    if (value.IsHolding<int>()) {
        f(value.UncheckedGet<int>());
    } else if (value.IsHolding<long>()) {
        f(static_cast<int>(value.UncheckedGet<long>()));
    } else if (value.IsHolding<long long>()) {
        f(static_cast<int>(value.UncheckedGet<long long>()));
    }
}

template <typename F>
void _CheckForFloatValue(const VtValue& value, F&& f)
{
    if (value.IsHolding<float>()) {
        f(value.UncheckedGet<float>());
    } else if (value.IsHolding<double>()) {
        f(static_cast<float>(value.UncheckedGet<double>()));
    } else if (value.IsHolding<GfHalf>()) {
        f(value.UncheckedGet<GfHalf>());
    }
}

void _RemoveArnoldGlobalPrefix(const TfToken& key, TfToken& key_new)
{
    if (TfStringStartsWith(key, _tokens->arnoldGlobal))
        key_new = TfToken{key.GetText() + _tokens->arnoldGlobal.size()};
    else if (TfStringStartsWith(key, _tokens->arnoldNamespace))
        key_new = TfToken{key.GetText() + _tokens->arnoldNamespace.size()};
    else 
        key_new = key;
}

} // namespace

std::mutex HdArnoldRenderDelegate::_mutexResourceRegistry;
std::atomic_int HdArnoldRenderDelegate::_counterResourceRegistry;
HdResourceRegistrySharedPtr HdArnoldRenderDelegate::_resourceRegistry;

AtNode* HydraArnoldAPI::CreateArnoldNode(const char* type, const char* name)
{
    return _renderDelegate->CreateArnoldNode(AtString(type), AtString(name));
}
const AtNode* HydraArnoldAPI::GetProceduralParent() const
{
    return _renderDelegate->GetProceduralParent();
}

void HydraArnoldAPI::AddNodeName(const std::string &name, AtNode *node)
{
    _renderDelegate->AddNodeName(name, node);
}
AtNode* HydraArnoldAPI::LookupTargetNode(const char* targetName, const AtNode* source, ConnectionType c)
{
    return _renderDelegate->LookupNode(targetName, true);
}
const AtString& HydraArnoldAPI::GetPxrMtlxPath() 
{
    return _renderDelegate->GetPxrMtlxPath();
}

HdArnoldRenderDelegate::HdArnoldRenderDelegate(bool isBatch, const TfToken &context, AtUniverse *universe, AtSessionMode renderSessionType, AtNode* procParent) : 
    _apiAdapter(this),
    _universe(universe),
    _procParent(procParent),
    _renderSessionType(renderSessionType),
    _context(context),
    _isBatch(isBatch), 
    _renderDelegateOwnsUniverse(universe==nullptr)
{     

    _lightLinkingChanged.store(false, std::memory_order_release);
    _meshLightsChanged.store(false, std::memory_order_release);
    _id = SdfPath(TfToken(TfStringPrintf("/HdArnoldRenderDelegate_%p", this)));
    // use the "render" tag by default
    _renderTags.push_back(UsdGeomTokens->render);
    // We first need to check if arnold has already been initialized.
    // If not, we need to call AiBegin, and the destructor on we'll call AiEnd
    bool isArnoldActive = 
#if ARNOLD_VERSION_NUM >= 70100
        AiArnoldIsActive();
#else
        AiUniverseIsActive();
#endif

    if (_isBatch && _renderDelegateOwnsUniverse) {
#if ARNOLD_VERSION_NUM >= 70104
        // Ensure that the ADP dialog box will not pop up and hang the application
        // We only want to do this when we own the universe (e.g. with husk), 
        // otherwise this would prevent CER from showing up when we're rendering it
        // through a procedural, or scene format plugin
        AiADPDisableDialogWindow();
        AiErrorReportingSetEnabled(false);
#endif
    }
    if (!isArnoldActive) {
        AiADPAddProductMetadata(AI_ADP_PLUGINNAME, AtString{"arnold-usd"});
        AiADPAddProductMetadata(AI_ADP_PLUGINVERSION, AtString{AI_VERSION});
        AiADPAddProductMetadata(AI_ADP_HOSTNAME, AtString{"Hydra"});
        AiADPAddProductMetadata(AI_ADP_HOSTVERSION, AtString{PXR_VERSION_STR});
        AiBegin(_renderSessionType);
    }
    _supportedRprimTypes = {HdPrimTypeTokens->mesh, HdPrimTypeTokens->volume, HdPrimTypeTokens->points,
                            HdPrimTypeTokens->basisCurves, str::t_procedural_custom};
#if PXR_VERSION >= 2603
    _supportedRprimTypes.push_back(HdPrimTypeTokens->particleField);
#endif
    if (_mask & AI_NODE_SHAPE) {
        auto* shapeIter = AiUniverseGetNodeEntryIterator(AI_NODE_SHAPE);
        while (!AiNodeEntryIteratorFinished(shapeIter)) {
            const auto* nodeEntry = AiNodeEntryIteratorGetNext(shapeIter);
            TfToken rprimType{ArnoldUsdMakeCamelCase(TfStringPrintf("Arnold_%s", AiNodeEntryGetName(nodeEntry)))};
            _supportedRprimTypes.push_back(rprimType);
            _nativeRprimTypes.insert({rprimType, AiNodeEntryGetNameAtString(nodeEntry)});

            NativeRprimParamList paramList;
            auto* paramIter = AiNodeEntryGetParamIterator(nodeEntry);
            while (!AiParamIteratorFinished(paramIter)) {
                const auto* param = AiParamIteratorGetNext(paramIter);
                const auto paramName = AiParamGetName(param);
                if (ArnoldUsdIgnoreParameter(paramName)) {
                    continue;
                }
                paramList.emplace(TfToken{TfStringPrintf("arnold:%s", paramName.c_str())}, param);
            }

            _nativeRprimParams.emplace(AiNodeEntryGetNameAtString(nodeEntry), std::move(paramList));
            AiParamIteratorDestroy(paramIter);
        }
        AiNodeEntryIteratorDestroy(shapeIter);
    }
    std::lock_guard<std::mutex> guard(_mutexResourceRegistry);
    if (_counterResourceRegistry.fetch_add(1) == 0) {
        _resourceRegistry.reset(new HdResourceRegistry());
    }

    const auto& config = HdArnoldConfig::GetInstance();
    if (_renderDelegateOwnsUniverse) {
        // Msg & log settings should be skipped if we have a procedural parent. 
        // In this case, a procedural shouldn't affect the global scene settings
        AiMsgSetConsoleFlags(
            #if ARNOLD_VERSION_NUM < 70100
                GetRenderSession(),
            #else
                _universe,
            #endif
            (config.log_flags_console >= 0) ? 
                config.log_flags_console : _verbosityLogFlags);

        AiMsgSetLogFileFlags(
            #if ARNOLD_VERSION_NUM < 70100
                GetRenderSession(),
            #else
                _universe,
            #endif
            (config.log_flags_file >= 0) ? 
                config.log_flags_file : _verbosityLogFlags);

        if (!config.log_file.empty())
        {
            AiMsgSetLogFileName(config.log_file.c_str());
        }
        if (!config.stats_file.empty())
        {
            AiStatsSetFileName(config.stats_file.c_str());
        }
        if (!config.profile_file.empty())
        {
            AiProfileSetFileName(config.profile_file.c_str());
        }
    }

    hdArnoldInstallNodes();
    // Check the USD environment variable for custom Materialx node definitions.
    // We need to use this to pass it on to Arnold's MaterialX
    const char *pxrMtlxPath = std::getenv("PXR_MTLX_STDLIB_SEARCH_PATHS");
    if (pxrMtlxPath) {
        _pxrMtlxPath = AtString(pxrMtlxPath);
    }

    if (_renderDelegateOwnsUniverse) {
        _universe = AiUniverse();
        _renderSession = AiRenderSession(_universe, _renderSessionType);
    }

    _renderParam.reset(new HdArnoldRenderParam(this));
    // To set the default value.
    _fps = _renderParam->GetFPS();
    _options = AiUniverseGetOptions(_universe);
    if (_renderDelegateOwnsUniverse) {
        for (const auto& o : _GetSupportedRenderSettings()) {
            _SetRenderSetting(o.first, o.second.defaultValue);
        }
        AiRenderSetHintStr(
            GetRenderSession(), str::render_context, AtString(_context.GetText()));

        // We need access to both beauty and P at the same time.
        if (_isBatch) {
            AiRenderSetHintBool(GetRenderSession(), str::progressive, false);
            AiNodeSetBool(_options, str::enable_progressive_render, false);
        } else {
            AiRenderSetHintBool(GetRenderSession(), str::progressive_show_all_outputs, true);
        }
    }

    // Check if shape instancing is supported
#if ARNOLD_VERSION_NUM >= 70405
    std::string envShapeInstancing = TfGetEnvSetting(HDARNOLD_SHAPE_INSTANCING);
    _supportShapeInstancing = envShapeInstancing != std::string("0");
#else
    _supportShapeInstancing = false;
#endif    
    
}

HdArnoldRenderDelegate::~HdArnoldRenderDelegate()
{
    std::lock_guard<std::mutex> guard(_mutexResourceRegistry);
    if (_counterResourceRegistry.fetch_sub(1) == 1) {
        _resourceRegistry.reset();
    }
    _renderParam->Interrupt();
    if (_renderDelegateOwnsUniverse) {
        AiRenderSessionDestroy(GetRenderSession());
        AiUniverseDestroy(_universe);
    }

}

HdRenderParam* HdArnoldRenderDelegate::GetRenderParam() const { return _renderParam.get(); }

void HdArnoldRenderDelegate::CommitResources(HdChangeTracker* tracker) {}

const TfTokenVector& HdArnoldRenderDelegate::GetSupportedRprimTypes() const { return _supportedRprimTypes; }

const TfTokenVector& HdArnoldRenderDelegate::GetSupportedSprimTypes() const { return _SupportedSprimTypes(); }

const TfTokenVector& HdArnoldRenderDelegate::GetSupportedBprimTypes() const { return _SupportedBprimTypes(_renderDelegateOwnsUniverse); }

void HdArnoldRenderDelegate::_SetRenderSetting(const TfToken& _key, const VtValue& _value)
{    
    // function to get or create the color manager and set it on the options node
    auto getOrCreateColorManager = [](HdArnoldRenderDelegate *renderDelegate, AtNode* options) -> AtNode* {
        AtNode* colorManager = static_cast<AtNode*>(AiNodeGetPtr(options, str::color_manager));
        if (colorManager == nullptr) {
            const char *ocio_path = std::getenv("OCIO");
            if (ocio_path) {
                colorManager = renderDelegate->CreateArnoldNode(str::color_manager_ocio, 
                    str::color_manager_ocio);
                AiNodeSetPtr(options, str::color_manager, colorManager);
                AiNodeSetStr(colorManager, str::config, AtString(ocio_path));
            }
            else
                // use the default color manager
                colorManager = renderDelegate->LookupNode("ai_default_color_manager_ocio");
        }
        return colorManager;
    };

    // When husk/houdini changes frame, they set the new frame number via the render settings.
    if (_key == str::t_houdiniFrame) {
        if (_value.IsHolding<double>()) {
            const float frame = static_cast<float>(_value.UncheckedGet<double>());
            AiNodeSetFlt(_options, str::frame, frame);
        }

        // We want to restart a new render in that case.
        _renderParam->Restart();
    }
    
    // See the HDK hydra docs for more details
    //   https://www.sidefx.com/docs/hdk/_h_d_k__u_s_d_hydra.html#HDK_USDHydraCopTextures
    if (_key == str::t_houdiniCopTextureChanged) {
        // COP textures need updating, flush the texture cache to trigger a refresh
        // of all the image_cop nodes
        _renderParam->Pause();
        AiUniverseCacheFlush(_universe, AI_CACHE_TEXTURE);
        _renderParam->Restart();
    }

    // Special setting that describes custom output, like deep AOVs or other arnold drivers #1422.
    if (_key == _tokens->delegateRenderProducts) {
        _ParseDelegateRenderProducts(_value);
        return;
    }
    TfToken key;
    _RemoveArnoldGlobalPrefix(_key, key);

    // Currently usdview can return double for floats, so until it's fixed
    // we have to convert doubles to float.
    auto value = _value.IsHolding<double>() ? VtValue(static_cast<float>(_value.UncheckedGet<double>())) : _value;
    // Certain applications might pass boolean values via ints or longs.
    if (key == str::t_enable_gpu_rendering) {
        _CheckForBoolValue(value, [&](const bool b) {
            AiNodeSetStr(_options, str::render_device, b ? str::GPU : str::CPU);
            AiDeviceAutoSelect(GetRenderSession());
        });
    } else if (key == str::t_log_verbosity) {
        if (value.IsHolding<int>()) {
            _verbosityLogFlags = _GetLogFlagsFromVerbosity(value.UncheckedGet<int>());
            static const auto& config = HdArnoldConfig::GetInstance();
            // Do not set the console and file flags, if the corresponding
            // environment variable was set in the config
            if (config.log_flags_console < 0) {
                AiMsgSetConsoleFlags(
                    #if ARNOLD_VERSION_NUM < 70100                    
                        GetRenderSession(), 
                    #else
                        _universe,
                    #endif
                    _verbosityLogFlags);
            }
            if (config.log_flags_file < 0) {
                AiMsgSetLogFileFlags(
                    #if ARNOLD_VERSION_NUM < 70100                    
                        GetRenderSession(), 
                    #else
                        _universe,
                    #endif
                    _verbosityLogFlags);

            }
        }
    } else if (key == str::t_log_file) {
        if (value.IsHolding<std::string>()) {
            _logFile = value.UncheckedGet<std::string>();
            AiMsgSetLogFileName(_logFile.c_str());
        }
#if ARNOLD_VERSION_NUM >= 70401
    } else if (key == str::t_report_file) {
        if (value.IsHolding<std::string>()) {
            _reportFile = value.UncheckedGet<std::string>();
            AiReportSetFileName(_reportFile.c_str());
        }
#endif
    } else if (key == str::t_stats_file) {
        if (value.IsHolding<std::string>()) {
            _statsFile = value.UncheckedGet<std::string>();
            AiStatsSetFileName(_statsFile.c_str());
        }
    } else if (key == str::t_profile_file) {
        if (value.IsHolding<std::string>()) {
            _profileFile = value.UncheckedGet<std::string>();
            AiProfileSetFileName(_profileFile.c_str());
        }
    } else if (key == str::t_enable_progressive_render) {
        if (!_isBatch) {
            _CheckForBoolValue(value, [&](const bool b) {
                AiRenderSetHintBool(GetRenderSession(), str::progressive, b);
                AiNodeSetBool(_options, str::enable_progressive_render, b);
            });
        }
    } else if (key == str::t_progressive_min_AA_samples) {
        if (!_isBatch) {
            _CheckForIntValue(value, [&](const int i) {
                AiRenderSetHintInt(GetRenderSession(), str::progressive_min_AA_samples, i);
            });
        }
    } else if (key == str::t_interactive_target_fps) {
        if (!_isBatch) {
            if (value.IsHolding<float>()) {
                AiRenderSetHintFlt(GetRenderSession(), str::interactive_target_fps, value.UncheckedGet<float>());
            }
        }
    } else if (key == str::t_interactive_target_fps_min) {
        if (!_isBatch) {
            if (value.IsHolding<float>()) {
                AiRenderSetHintFlt(GetRenderSession(), str::interactive_target_fps_min, value.UncheckedGet<float>());
            }
        }
    } else if (key == str::t_interactive_fps_min) {
        if (!_isBatch) {
            if (value.IsHolding<float>()) {
                AiRenderSetHintFlt(GetRenderSession(), str::interactive_fps_min, value.UncheckedGet<float>());
            }
        }
    } else if (key == _tokens->instantaneousShutter) {
        // If the arnold-specific attribute "ignore_motion_blur" is set, it should take 
        // precedence over the usd builtin instantaneousShutter
        _CheckForBoolValue(value, [&](const bool b) { 
            if (!_forceIgnoreMotionBlur)
                AiNodeSetBool(_options, str::ignore_motion_blur, b); 
        });
    } else if (key == str::t_ignore_motion_blur) {
        _CheckForBoolValue(value, [&](const bool b) { 
            AiNodeSetBool(_options, str::ignore_motion_blur, b); 
            _forceIgnoreMotionBlur = true; 
        });
    } else if (key == str::t_houdiniFps) {
        _CheckForFloatValue(value, [&](const float f) {
            _fps = f;
            AiNodeSetFlt(_options, str::fps, _fps);
        });
    } else if (key == str::t_background) {
        ArnoldUsdCheckForSdfPathValue(value, [&](const SdfPath& p) { _background = p; });
    } else if (key == str::t_atmosphere) {
        ArnoldUsdCheckForSdfPathValue(value, [&](const SdfPath& p) { _atmosphere = p; });
    } else if (key == str::t_aov_shaders) {
        ArnoldUsdCheckForSdfPathVectorValue(value, [&](const SdfPathVector& p) { _aov_shaders = p; });
    } else if (key == str::t_imager) {
        ArnoldUsdCheckForSdfPathValue(value, [&](const SdfPath& p) { _imager = p; });
    } else if (key == str::t_shader_override) {
        ArnoldUsdCheckForSdfPathValue(value, [&](const SdfPath& p) { _shader_override = p; });
    } else if (key == str::t_subdiv_dicing_camera) {
        ArnoldUsdCheckForSdfPathValue(value, [&](const SdfPath& p) {
            _subdiv_dicing_camera = p; 
            AiNodeSetPtr(_options, str::subdiv_dicing_camera, LookupNode(_subdiv_dicing_camera.GetText()));
        });
    } else if (key == str::color_space_linear) {
        if (value.IsHolding<std::string>()) {
            AtNode* colorManager = getOrCreateColorManager(this, _options);
            AiNodeSetStr(colorManager, str::color_space_linear, AtString(value.UncheckedGet<std::string>().c_str()));
        }
    } else if (key == str::color_space_narrow) {
        if (value.IsHolding<std::string>()) {
            AtNode* colorManager = getOrCreateColorManager(this, _options);
            AiNodeSetStr(colorManager, str::color_space_narrow, AtString(value.UncheckedGet<std::string>().c_str()));
        }
    } else if (key == _tokens->dataWindowNDC) {
        if (value.IsHolding<GfVec4f>()) {
            _windowNDC = value.UncheckedGet<GfVec4f>();
        }
    } else if (key == _tokens->pixelAspectRatio) {
        if (value.IsHolding<float>()) {
            _pixelAspectRatio = value.UncheckedGet<float>();
        }
    } 
    else if (key == _tokens->resolution) {
        if (value.IsHolding<GfVec2i>()) {
            _resolution = value.UncheckedGet<GfVec2i>();
        }
    } else if (key == _tokens->batchCommandLine) {
        // Solaris-specific command line, it can have an argument "-o output.exr" to override
        // the output image. We might end up using this for arnold drivers
        if (value.IsHolding<VtStringArray>()) {
            const VtStringArray &commandLine = value.UncheckedGet<VtArray<std::string>>();
            for (unsigned int i = 0; i < commandLine.size(); ++i) {
                // husk argument for output image
                if (commandLine[i] == "-o" && i < commandLine.size() - 2) {
                    _outputOverride = commandLine[++i];
                    continue;
                }
                // husk argument for thread count (#1077)
                if ((commandLine[i] == "-j" || commandLine[i] == "--threads") 
                        && i < commandLine.size() - 2) {
                    // if for some reason the argument value is not a number, atoi should return 0
                    // which is also the default arnold value. 
                    AiNodeSetInt(_options, str::threads, std::atoi(commandLine[++i].c_str()));
                }
            }
        }
    } else if (TfStringStartsWith(key.GetString(), _tokens->colorManagerNamespace)) {
        std::string cmParam = key.GetString().substr(_tokens->colorManagerNamespace.GetString().length());
        AtNode* colorManager = getOrCreateColorManager(this, _options);
        AtString cmParamStr(cmParam.c_str());
        if (AiNodeEntryLookUpParameter(AiNodeGetNodeEntry(colorManager), 
            AtString(cmParam.c_str())) != nullptr) {
            _SetNodeParam(colorManager, TfToken(cmParam), value);
        }
    } 
    else {
        auto* optionsEntry = AiNodeGetNodeEntry(_options);
        // Sometimes the Render Delegate receives parameters that don't exist
        // on the options node. For example, if the host application ignores the
        // render setting descriptor list.
        if (AiNodeEntryLookUpParameter(optionsEntry, AtString(key.GetText())) != nullptr) {
            _SetNodeParam(_options, key, value);
        }
    }
}

void HdArnoldRenderDelegate::_ParseDelegateRenderProducts(const VtValue& value)
{
    // Details about the data layout can be found here:
    // https://www.sidefx.com/docs/hdk/_h_d_k__u_s_d_hydra.html#HDK_USDHydraHuskDRP
    ClearDelegateRenderProducts();
    
    using DataType = VtArray<HdAovSettingsMap>;
    if (!value.IsHolding<DataType>()) {
        return;
    }
    auto products = value.UncheckedGet<DataType>();
    // For Render Delegate products, we want to eventually create arnold drivers
    // during batch rendering #1422
    for (auto& productIter : products) {
        HdArnoldDelegateRenderProduct product;
        const auto* productType = TfMapLookupPtr(productIter, _tokens->productType);

        // check the product type, and see if we support it
        if (productType == nullptr || !productType->IsHolding<TfToken>())
            continue;

        TfToken renderProductType = productType->UncheckedGet<TfToken>();
        // We only consider render products with type set to "arnold",
        // as well as "deep" for backwards compatibility #1422
        if (renderProductType != str::t_arnold && renderProductType != _tokens->deep)
            continue;

        // default driver is exr
        TfToken driverType = _tokens->driver_exr;
        // Special case for "deep" for backwards compatibility, we want a deepexr driver
        if (renderProductType == _tokens->deep)
            driverType = str::t_driver_deepexr;
        else {
            const auto* arnoldDriver = TfMapLookupPtr(productIter, _tokens->arnoldDriver);
            if (arnoldDriver != nullptr ) { 
                // arnold:driver is set in this render product, we use that for the driver type
                if (arnoldDriver->IsHolding<TfToken>()) {
                    driverType = arnoldDriver->UncheckedGet<TfToken>();
                } else if (arnoldDriver->IsHolding<std::string>()) {
                    driverType = TfToken(arnoldDriver->UncheckedGet<std::string>());
                }
            }
        }

        // Let's check if a driver type exists as this render product type #1422
        if (AiNodeEntryLookUp(AtString(driverType.GetText())) == nullptr) {
            // Arnold doesn't know how to render with this driver, let's skip it
            AiMsgWarning("Unknown Arnold Driver Type %s", driverType.GetText());
            continue; 
        }

        // Ignoring cases where productName is not set.
        const auto* productName = TfMapLookupPtr(productIter, _tokens->productName);
        if (productName == nullptr || !productName->IsHolding<TfToken>()) {
            continue;
        }
        product.productName = productName->UncheckedGet<TfToken>();
        product.productType = driverType;
        productIter.erase(_tokens->productType);
        productIter.erase(_tokens->productName);
        // Elements of the HdAovSettingsMap in the product are either a list of RenderVars or generic attributes
        // of the render product.
        for (const auto& productElem : productIter) {
            // If the key is "aovDescriptor.aovSettings" then we got the list of RenderVars.
            if (productElem.first == _tokens->orderedVars) {
                if (!productElem.second.IsHolding<DataType>()) {
                    continue;
                }
                const auto& renderVars = productElem.second.UncheckedGet<DataType>();
                for (const auto& renderVarIter : renderVars) {
                    HdArnoldRenderVar renderVar;
                    renderVar.sourceType = _tokens->raw;
                    // Each element either contains a setting, or "aovDescriptor.aovSettings" which will hold
                    // extra settings for the RenderVar including metadata.
                    for (const auto& renderVarElem : renderVarIter) {
                        if (renderVarElem.first == _tokens->aovSettings) {
                            if (!renderVarElem.second.IsHolding<HdAovSettingsMap>()) {
                                continue;
                            }
                            renderVar.settings = renderVarElem.second.UncheckedGet<HdAovSettingsMap>();
                            // name is not coming through as a top parameter.
                            const auto* aovName = TfMapLookupPtr(renderVar.settings, _tokens->aovName);
                            if (aovName != nullptr) {
                                if (aovName->IsHolding<std::string>()) {
                                    renderVar.name = aovName->UncheckedGet<std::string>();
                                } else if (aovName->IsHolding<TfToken>()) {
                                    renderVar.name = aovName->UncheckedGet<TfToken>().GetString();
                                }
                            }
                        } else if (
                            renderVarElem.first == _tokens->sourceName &&
                            renderVarElem.second.IsHolding<std::string>()) {
                            renderVar.sourceName = renderVarElem.second.UncheckedGet<std::string>();
                        } else if (
                            renderVarElem.first == _tokens->sourceType && renderVarElem.second.IsHolding<TfToken>()) {
                            renderVar.sourceType = renderVarElem.second.UncheckedGet<TfToken>();
                        } else if (
                            renderVarElem.first == _tokens->dataType && renderVarElem.second.IsHolding<TfToken>()) {
                            renderVar.dataType = renderVarElem.second.UncheckedGet<TfToken>();
                        } else if (
                            renderVarElem.first == _tokens->format && renderVarElem.second.IsHolding<HdFormat>()) {
                            renderVar.format = renderVarElem.second.UncheckedGet<HdFormat>();
                        } else if (renderVarElem.first == _tokens->clearValue) {
                            renderVar.clearValue = renderVarElem.second;
                        } else if (
                            renderVarElem.first == _tokens->multiSampled && renderVarElem.second.IsHolding<bool>()) {
                            renderVar.multiSampled = renderVarElem.second.UncheckedGet<bool>();
                        }
                    }

                    // Look for driver:parameters:aov:format and arnold:format overrides
                    const auto* aovDriverFormat = TfMapLookupPtr(renderVar.settings, _tokens->aovDriverFormat);
                    if (aovDriverFormat != nullptr && aovDriverFormat->CanCast<TfToken>()) {
                        TfToken aovDriverFormatToken = VtValue::Cast<TfToken>(*aovDriverFormat).UncheckedGet<TfToken>();
                        renderVar.format = _GetHdFormatFromToken(aovDriverFormatToken);
                    }
                    const auto* arnoldFormat = TfMapLookupPtr(renderVar.settings, _tokens->aovFormat);
                    if (arnoldFormat != nullptr && arnoldFormat->CanCast<TfToken>()) {
                        TfToken arnoldFormatToken = VtValue::Cast<TfToken>(*arnoldFormat).UncheckedGet<TfToken>();
                        renderVar.format = _GetHdFormatFromToken(arnoldFormatToken);
                    }

                    if (!renderVar.sourceName.empty()) {
                        // if drivers:parameters:aov:name is not defined, use sourceName instead #2572
                        if (renderVar.name.empty())
                            renderVar.name = renderVar.sourceName;
                        product.renderVars.emplace_back(std::move(renderVar));
                    }
                }
            } else {
                // It's a setting describing the RenderProduct.
                product.settings.insert({productElem.first, productElem.second});
            }
        }
        _delegateRenderProducts.emplace_back(std::move(product));
    }
}

void HdArnoldRenderDelegate::SetRenderSetting(const TfToken& key, const VtValue& value)
{
    _renderParam->Interrupt();
    _SetRenderSetting(key, value);
}

VtValue HdArnoldRenderDelegate::GetRenderSetting(const TfToken& _key) const
{
    TfToken key;
    _RemoveArnoldGlobalPrefix(_key, key);
    if (key == str::t_enable_gpu_rendering) {
        return VtValue(AiNodeGetStr(_options, str::render_device) == str::GPU);
    } else if (key == str::t_enable_progressive_render) {
        bool v = true;
        AiRenderGetHintBool(GetRenderSession(), str::progressive, v);
        return VtValue(v);
    } else if (key == str::t_progressive_min_AA_samples) {
        int v = -4;
        AiRenderGetHintInt(GetRenderSession(), str::progressive_min_AA_samples, v);
        return VtValue(v);
    } else if (key == str::t_log_verbosity) {
        return VtValue(ArnoldUsdGetLogVerbosityFromFlags(_verbosityLogFlags));
    } else if (key == str::t_log_file) {
        return VtValue(_logFile);
    } else if (key == str::t_report_file) {
        return VtValue(_reportFile);
    } else if (key == str::t_stats_file) {
        return VtValue(_statsFile);
    } else if (key == str::t_profile_file) {
        return VtValue(_profileFile);
    } else if (key == str::t_interactive_target_fps) {
        float v = 1.0f;
        AiRenderGetHintFlt(GetRenderSession(), str::interactive_target_fps, v);
        return VtValue(v);
    } else if (key == str::t_interactive_target_fps_min) {
        float v = 1.0f;
        AiRenderGetHintFlt(GetRenderSession(), str::interactive_target_fps_min, v);
        return VtValue(v);
    } else if (key == str::t_interactive_fps_min) {
        float v = 1.0f;
        AiRenderGetHintFlt(GetRenderSession(), str::interactive_fps_min, v);
        return VtValue(v);
    } else if (key == str::t_profile_file) {
        return VtValue(std::string(AiProfileGetFileName().c_str()));
    } else if (key == str::t_background) {
        return VtValue(_background.GetString());
    } else if (key == str::t_atmosphere) {
        return VtValue(_atmosphere.GetString());
    } else if (key == str::t_aov_shaders) {
        // There should be a function in common_utils.cpp
        std::vector<std::string> pathsAsString;
        std::transform(_aov_shaders.begin(), _aov_shaders.end(), std::back_inserter(pathsAsString), [](const auto &p){return p.GetString();});
        return VtValue(TfStringJoin(pathsAsString));
    } else if (key == str::t_imager) {
        return VtValue(_imager.GetString());
    }  else if (key == str::t_subdiv_dicing_camera) {
        return VtValue(_subdiv_dicing_camera.GetString());
    } else if (key == str::t_shader_override) {
        return VtValue(_shader_override.GetString());
    }
    const auto* nentry = AiNodeGetNodeEntry(_options);
    const auto* pentry = AiNodeEntryLookUpParameter(nentry, AtString(key.GetText()));
    return _GetNodeParamValue(_options, pentry);
}

// For now we only support a few parameter types, that are expected to have
// UI code in usdview / Maya to Hydra.
HdRenderSettingDescriptorList HdArnoldRenderDelegate::GetRenderSettingDescriptors() const
{
    const auto* nentry = AiNodeGetNodeEntry(_options);
    HdRenderSettingDescriptorList ret;
    for (const auto& it : _GetSupportedRenderSettings()) {
        HdRenderSettingDescriptor desc;
        desc.name = it.second.label;
        desc.key = it.first;
        if (it.second.defaultValue.IsEmpty()) {
            const auto* pentry = AiNodeEntryLookUpParameter(nentry, AtString(it.first.GetText()));
            desc.defaultValue = _GetNodeParamValue(_options, pentry);
        } else {
            desc.defaultValue = it.second.defaultValue;
        }
        ret.emplace_back(std::move(desc));
    }
    return ret;
}

// for testing in batch mode. TODO: correctly check the if we can and want to use the hydra render settings
bool HdArnoldRenderDelegate::IsUsingHydraRenderSettings() const { 
#ifdef ENABLE_HYDRA2_RENDERSETTINGS
    return true; 
#else
    return false;
#endif
}

VtDictionary HdArnoldRenderDelegate::GetRenderStats() const
{
    VtDictionary stats;

    float total_progress = 100.0f;
    AiRenderGetHintFlt(GetRenderSession(), str::total_progress, total_progress);
    stats[_tokens->percentDone] = total_progress;

    const double elapsed = _renderParam->GetElapsedRenderTime() / 1000.0;
    stats[_tokens->totalClockTime] = VtValue(elapsed);

    std::string renderStatus = _renderParam->GetRenderStatusString();
    if(!renderStatus.empty())
    {
        // Beautify the log - 'Rendering' looks nicer than 'rendering'
        // in the viewport annotation
        renderStatus[0] = std::toupper(renderStatus[0]);
    }
    const int width = AiNodeGetInt(_options, str::xres);
    const int height = AiNodeGetInt(_options, str::yres);
    constexpr std::size_t maxResChars{256};
    char resolutionBuffer[maxResChars];
    snprintf(&resolutionBuffer[0], maxResChars, "%s %i x %i", renderStatus.c_str(), width, height);
    stats[_tokens->renderProgressAnnotation] = VtValue(resolutionBuffer);

    if (total_progress >= 100) {
        // If there are cryptomatte drivers, we look for the metadata that is stored in each of them.
        // In theory, we could just look for the first driver, but for safety we're doing it for all of them.
        // This doesn't need to be done while the render is in progress as it's only required
        // when the exr file is authored on disk #2334
        for (const auto& cryptoDriver : _cryptomatteDrivers) {
            const AtNode *driver = LookupNode(cryptoDriver.c_str());
            if (!driver)
                continue;
            if (AiNodeLookUpUserParameter(driver, str::custom_attributes) == nullptr)
                continue;
            const AtArray *customAttrsArray = AiNodeGetArray(driver, str::custom_attributes);
            if (customAttrsArray == nullptr)
                continue;
            unsigned int customAttrsCount = AiArrayGetNumElements(customAttrsArray);
            for (unsigned int i = 0; i < customAttrsCount; ++i) {
                AtString customAttr = AiArrayGetStr(customAttrsArray, i);
                std::string customAttrStr(customAttr.c_str());
                // the custom_attributes attribute will be an array of strings, where each  
                // element is set like:
                // "STRING cryptomatte/f834d0a/conversion uint32_to_float32"
                // where the second element is the metadata name and the last one
                // is the metadata value
                size_t pos = customAttrStr.find_first_of(' ');
                if (pos == std::string::npos)
                    continue;
                std::string customAttrType = customAttrStr.substr(0, pos);
                customAttrStr = customAttrStr.substr(pos + 1);

                pos = customAttrStr.find_first_of(' ');
                if (pos == std::string::npos)
                    continue;
                std::string metadataName = customAttrStr.substr(0, pos);
                std::string metadataVal = customAttrStr.substr(pos + 1);
                // TODO do we want to check if the metadata is not a string ?
                stats[TfToken(metadataName)] = TfToken(metadataVal);
            }
        }
    }

    // If we have a render error, we want to send it to husk which will exit with this error code.
    // Husk will also display a message like "[15:16:41] Delegate fatal error (1) detected" 
    auto errorCode = _renderParam->GetErrorCode();
    if (errorCode != AI_SUCCESS) {
        stats[_tokens->huskErrorStatus] = static_cast<int>(errorCode);
    }
    return stats;
}

HdResourceRegistrySharedPtr HdArnoldRenderDelegate::GetResourceRegistry() const { return _resourceRegistry; }

HdRenderPassSharedPtr HdArnoldRenderDelegate::CreateRenderPass(
    HdRenderIndex* index, const HdRprimCollection& collection)
{
    return HdRenderPassSharedPtr(new HdArnoldRenderPass(this, index, collection));
}

HdInstancer* HdArnoldRenderDelegate::CreateInstancer(HdSceneDelegate* delegate, const SdfPath& id)
{
    return new HdArnoldInstancer(this, delegate, id);
}

void HdArnoldRenderDelegate::DestroyInstancer(HdInstancer* instancer) { delete instancer; }

HdRprim* HdArnoldRenderDelegate::CreateRprim(const TfToken& typeId, const SdfPath& rprimId)
{
    if (!(_mask & AI_NODE_SHAPE))
        return nullptr;

    _renderParam->Interrupt();
    if (typeId == HdPrimTypeTokens->mesh) {
        return new HdArnoldMesh(this, rprimId);
    }
    if (typeId == HdPrimTypeTokens->volume) {
        return new HdArnoldVolume(this, rprimId);
    }
    if (typeId == HdPrimTypeTokens->points) {
        return new HdArnoldPoints(this, rprimId);
    }
#if PXR_VERSION >= 2603
    if (typeId == HdPrimTypeTokens->particleField) {
        return new HdArnoldGaussianSplat(this, rprimId);
    }
#endif
    if (typeId == HdPrimTypeTokens->basisCurves) {
        return new HdArnoldBasisCurves(this, rprimId);
    }
    if (typeId == str::t_procedural_custom) {
        return new HdArnoldProceduralCustom(this, rprimId);
    }
    auto typeIt = _nativeRprimTypes.find(typeId);
    if (typeIt != _nativeRprimTypes.end()) {
        return new HdArnoldNativeRprim(this, typeIt->second, rprimId);
    }
    TF_CODING_ERROR("Unknown Rprim Type %s", typeId.GetText());
    return nullptr;
}

void HdArnoldRenderDelegate::DestroyRprim(HdRprim* rPrim)
{
    _renderParam->Interrupt();
    delete rPrim;
}

HdSprim* HdArnoldRenderDelegate::CreateSprim(const TfToken& typeId, const SdfPath& sprimId)
{
    _renderParam->Interrupt();
    // We're creating a new Sprim. It's possible that it is already referenced
    // by another prim (which can happen when shaders are disconnected/reconnected).
    // In this case we need to dirty it so that all source prims are properly updated.
    // Note : for now we're only tracking dependencies for Sprim targets, but 
    // this could be extended
    const auto &it = _targetToSourcesMap.find(sprimId);
    if (it != _targetToSourcesMap.end())
        DirtyDependency(sprimId);

    if (typeId == HdPrimTypeTokens->camera) {
        return (_mask & AI_NODE_CAMERA) ? 
            new HdArnoldCamera(this, sprimId) : nullptr;
    }
    if (typeId == HdPrimTypeTokens->material) {
        return (_mask & AI_NODE_SHADER) ? 
            new HdArnoldNodeGraph(this, sprimId) : nullptr;
    }
    if (typeId == str::t_ArnoldNodeGraph) {
        return (_mask & AI_NODE_SHADER) ? 
            new HdArnoldNodeGraph(this, sprimId) : nullptr;
    }
    if (typeId == _tokens->ArnoldOptions) {
        return (_mask & AI_NODE_OPTIONS) ? 
            new HdArnoldOptions(this, sprimId) : nullptr;
    }

    if (typeId == HdPrimTypeTokens->sphereLight) {
        return (_mask & AI_NODE_LIGHT) ? 
            HdArnoldLight::CreatePointLight(this, sprimId) : nullptr;
    }
    if (typeId == HdPrimTypeTokens->distantLight) {
        return (_mask & AI_NODE_LIGHT) ? 
            HdArnoldLight::CreateDistantLight(this, sprimId) : nullptr;
    }
    if (typeId == HdPrimTypeTokens->diskLight) {
        return (_mask & AI_NODE_LIGHT) ? 
            HdArnoldLight::CreateDiskLight(this, sprimId) : nullptr;
    }
    if (typeId == HdPrimTypeTokens->rectLight) {
        return (_mask & AI_NODE_LIGHT) ? 
            HdArnoldLight::CreateRectLight(this, sprimId) : nullptr;
    }
    if (typeId == HdPrimTypeTokens->cylinderLight) {
        return (_mask & AI_NODE_LIGHT) ? 
            HdArnoldLight::CreateCylinderLight(this, sprimId) : nullptr;
    }
    if (typeId == HdPrimTypeTokens->domeLight) {
        return (_mask & AI_NODE_LIGHT) ? 
            HdArnoldLight::CreateDomeLight(this, sprimId) : nullptr;
    }
    if (typeId == _tokens->GeometryLight
#ifdef ENABLE_SCENE_INDEX
        || typeId == HdPrimTypeTokens->meshLight
#endif
    ) {
        return (_mask & AI_NODE_LIGHT) ? HdArnoldLight::CreateGeometryLight(this, sprimId) : nullptr;
    }
    if (typeId == HdPrimTypeTokens->simpleLight) {
        return nullptr;
    }
    if (typeId == HdPrimTypeTokens->extComputation) {
        return new HdExtComputation(sprimId);
    }
    TF_CODING_ERROR("Unknown Sprim Type %s", typeId.GetText());
    return nullptr;
}

HdSprim* HdArnoldRenderDelegate::CreateFallbackSprim(const TfToken& typeId)
{
    return nullptr;
}

void HdArnoldRenderDelegate::DestroySprim(HdSprim* sPrim)
{
    if (sPrim == nullptr)
        return;

    _renderParam->Interrupt();
    const auto &id = sPrim->GetId();
    // We could be destroying a Sprim that is being referenced by 
    // another source. We need to keep track of this, so that
    // all the references are properly updated
    if (_targetToSourcesMap.find(id) != _targetToSourcesMap.end())
        RemoveDependency(id);
    delete sPrim;
}

HdBprim* HdArnoldRenderDelegate::CreateBprim(const TfToken& typeId, const SdfPath& bprimId)
{
    // Neither of these will create Arnold nodes.
    if (typeId == HdPrimTypeTokens->renderBuffer) {
        return new HdArnoldRenderBuffer(bprimId);
    }
    if (typeId == _tokens->openvdbAsset) {
        return new HdArnoldOpenvdbAsset(this, bprimId);
    }

#if PXR_VERSION >= 2208
    // Only support render settings when we don't own the universe (procedural context).
    // When we own the universe (batch context), settings come through SetRenderSettings.
#ifdef ENABLE_HYDRA2_RENDERSETTINGS
    if (typeId == HdPrimTypeTokens->renderSettings /*&& !_renderDelegateOwnsUniverse*/) {
        return new HdArnoldRenderSettings(bprimId);
    }
#else
    if (typeId == HdPrimTypeTokens->renderSettings)
        return nullptr;
#endif // ENABLE_HYDRA2_RENDERSETTINGS
#endif // PXR_VERSION >= 2208

    TF_CODING_ERROR("Unknown Bprim Type %s", typeId.GetText());
    return nullptr;
}

HdBprim* HdArnoldRenderDelegate::CreateFallbackBprim(const TfToken& typeId)
{
    return CreateBprim(typeId, SdfPath());
}

void HdArnoldRenderDelegate::DestroyBprim(HdBprim* bPrim)
{
    // RenderBuffers can be in use in drivers.
    _renderParam->Interrupt();
    delete bPrim;
}

TfToken HdArnoldRenderDelegate::GetMaterialBindingPurpose() const { return HdTokens->full; }

TfTokenVector HdArnoldRenderDelegate::GetMaterialRenderContexts() const
{
    return {_tokens->arnold, str::t_mtlx};
}

AtString HdArnoldRenderDelegate::GetLocalNodeName(const AtString& name) const
{
    return AtString(_id.AppendChild(TfToken(name.c_str())).GetText());
}

AtUniverse* HdArnoldRenderDelegate::GetUniverse() const { return _universe; }

AtRenderSession* HdArnoldRenderDelegate::GetRenderSession() const
{
    if (_renderDelegateOwnsUniverse) {
        return _renderSession;
    } else {
        return AiUniverseGetRenderSession(GetUniverse());
    }
}

AtNode* HdArnoldRenderDelegate::GetOptions() const { return _options; }

AtNode* HdArnoldRenderDelegate::GetFallbackSurfaceShader()
{
    if (_fallbackShader)
        return _fallbackShader;

    std::lock_guard<std::mutex> guard(_defaultShadersMutex);
    if (_fallbackShader == nullptr) {
        _fallbackShader = CreateArnoldNode(str::standard_surface, 
            AtString("_fallbackShader"));
    
        AtNode *userDataReader = CreateArnoldNode(str::user_data_rgb,
            AtString("_fallbackShader_userDataReader"));
        
        AiNodeSetStr(userDataReader, str::attribute, str::displayColor);
        AiNodeSetRGB(userDataReader, str::_default, 1.0f, 1.0f, 1.0f);
        AiNodeLink(userDataReader, str::base_color, _fallbackShader);
    }
    return _fallbackShader; 
}

AtNode* HdArnoldRenderDelegate::GetFallbackVolumeShader()
{
    if (_fallbackVolumeShader)
        return _fallbackVolumeShader;

    std::lock_guard<std::mutex> guard(_defaultShadersMutex);
    if (_fallbackVolumeShader == nullptr) {
        _fallbackVolumeShader = CreateArnoldNode(str::standard_volume,
            AtString("_fallbackVolume"));
    }
    return _fallbackVolumeShader;
}

HdAovDescriptor HdArnoldRenderDelegate::GetDefaultAovDescriptor(const TfToken& name) const
{
    if (name == HdAovTokens->color) {
#if 1
        return HdAovDescriptor(HdFormatFloat32Vec4, false, VtValue(GfVec4f(0.0f)));
#else
        return HdAovDescriptor(HdFormatUNorm8Vec4, false, VtValue(GfVec4f(0.0f, 0.0f, 0.0f, 0.0f)));
#endif
    } else if (name == HdAovTokens->depth) {
        return HdAovDescriptor(HdFormatFloat32, false, VtValue(1.0f));
    } else if (name == HdAovTokens->primId) {
        return HdAovDescriptor(HdFormatInt32, false, VtValue(-1));
    } else if (name == HdAovTokens->instanceId || name == HdAovTokens->elementId || name == HdAovTokens->pointId) {
        // We are only supporting the prim id buffer for now.
        return HdAovDescriptor(HdFormatInt32, false, VtValue(-1));
    } else if (
        name == HdAovTokens->normal || name == HdAovTokens->Neye ||
        name == "linearDepth" || // This was changed to cameraDepth after 0.19.11.
        name == "cameraDepth") {
        // More built-in aovs.
        return HdAovDescriptor();
    } else if (TfStringStartsWith(name.GetString(), HdAovTokens->primvars)) {
        // Primvars.
        return HdAovDescriptor(HdFormatFloat32Vec3, false, VtValue(GfVec3f(0.0f)));
    } else if (TfStringStartsWith(name.GetString(), HdAovTokens->lpe)) {
        // LPEs.
        return HdAovDescriptor(HdFormatFloat32Vec3, false, VtValue(GfVec3f(0.0f)));
    } else {
        // Anything else. We can't decide what the AOV might contain based on the name, so we are just returning a
        // default value.
        return HdAovDescriptor(HdFormatFloat32Vec3, false, VtValue(GfVec3f(0.0f)));
    }
}

void HdArnoldRenderDelegate::RegisterLightLinking(const TfToken& name, HdLight* light, bool isShadow)
{
    std::lock_guard<std::mutex> guard(_lightLinkingMutex);
    auto& links = isShadow ? _shadowLinks : _lightLinks;
    auto it = links.find(name);
    if (it == links.end()) {
        if (!name.IsEmpty() || !links.empty()) {
            _lightLinkingChanged.store(true, std::memory_order_release);
        }
        links.emplace(name, std::vector<HdLight*>{light});
    } else {
        if (std::find(it->second.begin(), it->second.end(), light) == it->second.end()) {
            // We only trigger the change if we are registering a non-empty collection, or there are more than one
            // collections.
            if (!name.IsEmpty() || links.size() > 1) {
                _lightLinkingChanged.store(true, std::memory_order_release);
            }
            it->second.push_back(light);
        }
    }
}

void HdArnoldRenderDelegate::DeregisterLightLinking(const TfToken& name, HdLight* light, bool isShadow)
{
    std::lock_guard<std::mutex> guard(_lightLinkingMutex);
    auto& links = isShadow ? _shadowLinks : _lightLinks;
    auto it = links.find(name);
    if (it != links.end()) {
        // We only trigger updates if either deregistering a named collection, or deregistering the empty
        // collection and there are other collection.
        if (!name.IsEmpty() || links.size() > 1) {
            _lightLinkingChanged.store(true, std::memory_order_release);
        }
        it->second.erase(std::remove(it->second.begin(), it->second.end(), light), it->second.end());
        if (it->second.empty()) {
            links.erase(name);
        }
    }
}

void HdArnoldRenderDelegate::_ApplyLightLinking(AtNode* shape, const VtArray<TfToken>& categories)
{
    std::lock_guard<std::mutex> guard(_lightLinkingMutex);
    // We need to reset the parameter if either there are no light links, or the only light link is the default
    // group.
    const auto lightEmpty = _lightLinks.empty() || (_lightLinks.size() == 1 && _lightLinks.count(TfToken{}) == 1);
    const auto shadowEmpty = _shadowLinks.empty() || (_shadowLinks.size() == 1 && _shadowLinks.count(TfToken{}) == 1);
    if (lightEmpty) {
        AiNodeResetParameter(shape, str::use_light_group);
        AiNodeResetParameter(shape, str::light_group);
    }
    if (shadowEmpty) {
        AiNodeResetParameter(shape, str::use_shadow_group);
        AiNodeResetParameter(shape, str::shadow_group);
    }
    if (lightEmpty && shadowEmpty) {
        return;
    }
    auto applyGroups = [=](const AtString& group, const AtString& useGroup, const LightLinkingMap& links) {
        std::vector<AtNode*> lights;
        for (const auto& category : categories) {
            auto it = links.find(category);
            if (it != links.end()) {
                for (auto* light : it->second) {
                    auto* arnoldLight = HdArnoldLight::GetLightNode(light);
                    if (arnoldLight != nullptr) {
                        lights.push_back(arnoldLight);
                    }
                }
            }
        }
        // Add the lights with an empty collection to the list.
        auto it = links.find(TfToken{});
        if (it != links.end()) {
            for (auto* light : it->second) {
                auto* arnoldLight = HdArnoldLight::GetLightNode(light);
                if (arnoldLight != nullptr) {
                    lights.push_back(arnoldLight);
                }
            }
        }

        // Add the mesh lights as well, they are not registered as light in hydra unfortunatelly
        {
            std::lock_guard<std::mutex> guard(_meshLightsMutex);
            for (AtNode * meshLight:_meshLights) {
                lights.push_back(meshLight); // TODO except if they have a collection yeah
            }
        }

        // If lights is empty, then no lights affect the shape, and we still have to set useGroup to true.
        if (lights.empty()) {
            AiNodeResetParameter(shape, group);
        } else {
            AiNodeSetArray(
                shape, group, AiArrayConvert(static_cast<uint32_t>(lights.size()), 1, AI_TYPE_NODE, lights.data()));
        }
        AiNodeSetBool(shape, useGroup, true);
    };

    if (!lightEmpty) {
        std::lock_guard<std::mutex> guard(_deferredFunctionCallsMutex);
        _deferredFunctionCalls.push_back(std::bind(applyGroups, str::light_group, str::use_light_group, _lightLinks));
    }
    if (!shadowEmpty) {
        std::lock_guard<std::mutex> guard(_deferredFunctionCallsMutex);
        _deferredFunctionCalls.push_back(std::bind(applyGroups, str::shadow_group, str::use_shadow_group, _shadowLinks));
    }
}

void HdArnoldRenderDelegate::ApplyLightLinking(HdSceneDelegate *sceneDelegate, AtNode* node, SdfPath const& id) {
    const SdfPath &instancerId = sceneDelegate->GetInstancerId(id);
    VtArray<TfToken> instancerCategories;
    // If this shape is instanced, we store the list of "categories"
    // (aka collections) associated with it.
    if (!instancerId.IsEmpty()) {
        instancerCategories = sceneDelegate->GetCategories(instancerId);
    }
    if (instancerCategories.empty()) {
        // If there are no collections associated with eventual instancers,
        // we just pass the reference to the categories array to avoid useless copies
        _ApplyLightLinking(node, sceneDelegate->GetCategories(id));
    } else {
        // We want to concatenate the shape's categories with the
        // instancer's categories, and call ApplyLightLinking with the full list
        VtArray<TfToken> categories = sceneDelegate->GetCategories(id);
        categories.reserve(categories.size() + instancerCategories.size());
        for (const auto &instanceCategory : instancerCategories)
            categories.push_back(instanceCategory);
        _ApplyLightLinking(node, categories);
    }
}


void HdArnoldRenderDelegate::ProcessConnections()
{
    _apiAdapter.ProcessConnections();
}

bool HdArnoldRenderDelegate::CanUpdateScene()
{
    // For interactive renders, it is always possible to update the scene
    if (_renderSessionType == AI_SESSION_INTERACTIVE)
        return true;
    // For batch renders, only update the scene if the render hasn't started yet,
    // or if it's finished
    const int status = AiRenderGetStatus(_renderSession);
    return status != AI_RENDER_STATUS_RESTARTING && status != AI_RENDER_STATUS_RENDERING;
}

bool HdArnoldRenderDelegate::HasPendingChanges(HdRenderIndex* renderIndex, const SdfPath& cameraId, const GfVec2f& shutter)
{
    if (!_deferredFunctionCalls.empty()) {
        // TODO: depending on the size, we could do the update in parallel.
        for (const auto& func : _deferredFunctionCalls) {
            func();
        }
        _deferredFunctionCalls.clear();
    }

    HdDirtyBits bits = HdChangeTracker::Clean;
    // If Light Linking have changed, we have to dirty the categories on all rprims to force updating the
    // the light linking information.
    if (_lightLinkingChanged.exchange(false, std::memory_order_acq_rel)) {
        bits |= HdChangeTracker::DirtyCategories;
    }

    // MeshLight changes
    if (_meshLightsChanged.exchange(false, std::memory_order_acq_rel)) {
        bits |= HdChangeTracker::DirtyCategories;
    }

    // When shutter open and shutter close significantly changes, we might not have enough samples for transformation
    // and deformation, so we need to force re-syncing all the prims.
    if (_renderParam->UpdateShutter(shutter)) {
        bits |= HdChangeTracker::DirtyPoints | HdChangeTracker::DirtyTransform | HdChangeTracker::DirtyInstancer |
                HdChangeTracker::DirtyPrimvar;
        // If the shutter has changed, we also need to update the render camera
        if (!cameraId.IsEmpty()) {
            renderIndex->GetChangeTracker().MarkSprimDirty(cameraId, HdCamera::AllDirty);
        }

    }
    /// TODO(pal): Investigate if this is needed.
    /// When FPS changes we have to dirty points and primvars.
    if (_renderParam->UpdateFPS(_fps)) {
        bits |= HdChangeTracker::DirtyPoints | HdChangeTracker::DirtyPrimvar;
    }
    auto& changeTracker = renderIndex->GetChangeTracker();
    
    bool changes = false;
    if (bits != HdChangeTracker::Clean) {
#ifdef ENABLE_SCENE_INDEX
        // Unfortunately the MarkAllRprimsDirty doesn't work as we would expect in USD 25.05 with hydra 2, it marks dirty the prims of the legacy scene index which doesn't contain rprims, so all
        // the dirty notifications get discarded. We have to use a workaround to get the same behaviour as before by propagating the dirtyness to a dedicated scene index filter.
        if (HdRenderIndex::IsSceneIndexEmulationEnabled()) {
            if (HdSceneIndexBaseRefPtr sceneIndex = renderIndex->GetTerminalSceneIndex()) {
                // Regarding the used of TfHashMap to carry the dirtyness, we unfortunately can't pass
                // directly HdSceneIndexObserver::DirtiedPrimEntries due to template functions implementation
                // missing for this type. Vectors of SdfPath and HdDataSourceLocatorSet don't work as well. So
                // we revert to an hashmap which it not ideal, but it allows us to decouple the libraries using
                // only available functions and supported types known by each libs.
                TfHashMap<SdfPath, HdDataSourceLocatorSet, SdfPath::Hash> dirtyEntries;
                for (const SdfPath& id : renderIndex->GetRprimIds()) {
                    HdSceneIndexPrim prim = sceneIndex->GetPrim(id);
                    HdDataSourceLocatorSet locators;
                    // TODO: We might have to also translate bespoke arnold nodes parameters if they are not supported.
                    HdDirtyBitsTranslator::RprimDirtyBitsToLocatorSet(prim.primType, bits, &locators);
                    if (!locators.IsEmpty()) {
                        dirtyEntries.insert({id, locators});
                    }
                }
                // if the shutter was modified, we also want to update the camera
                if (bits & HdChangeTracker::DirtyTransform && !cameraId.IsEmpty()) {
                    HdSceneIndexPrim prim = sceneIndex->GetPrim(cameraId);
                    HdDataSourceLocatorSet locators;
                    HdDirtyBitsTranslator::SprimDirtyBitsToLocatorSet(prim.primType, HdCamera::AllDirty, &locators);
                    if (!locators.IsEmpty()) {
                        dirtyEntries.insert({cameraId, locators});
                    }
                }

                if (!dirtyEntries.empty()) {
                    auto dataSource = HdRetainedTypedSampledDataSource<
                        TfHashMap<SdfPath, HdDataSourceLocatorSet, SdfPath::Hash>>::New(dirtyEntries);
                    sceneIndex->SystemMessage(str::t_ArnoldMarkPrimsDirty, dataSource);
                }
            }
        } else {
            renderIndex->GetChangeTracker().MarkAllRprimsDirty(bits);
        }
#else
        renderIndex->GetChangeTracker().MarkAllRprimsDirty(bits);
#endif // ENABLE_SCENE_INDEX
        changes = true;
    }
    SdfPath id;
    auto markPrimDirty = [&](const SdfPath& source, HdDirtyBits bits) {
        // Marking a primitive as being dirty. But the function to invoke
        // depends on the prim type. For now we're checking first if a Rprim
        // exists with this name, to choose between Rprims and Sprims.
        if (renderIndex->HasRprim(source)) {
            changeTracker.MarkRprimDirty(source, bits);
        }
        else {
            // Depending on the Sprim type, the dirty bits must be different. See .//pxr/imaging/hd/dirtyBitsTranslator.cpp
            changeTracker.MarkSprimDirty(source, bits);
        }
    };
    // First let's process all the dependencies that were removed.
    // We need to remove it from all our maps, and mark all the 
    // sources as being dirty, so that they can update their 
    // new reference properly
    while (_dependencyRemovalQueue.try_pop(id)) {        
        auto targetIt = _targetToSourcesMap.find(id);
        if (targetIt != _targetToSourcesMap.end()) {
            changes = true; // this requires a render update
            for (const auto& source : targetIt->second) {
                // for each source referencing the current target
                // we need to remove the target from its list
                auto sourceIt = _sourceToTargetsMap.find(source);
                if (sourceIt != _sourceToTargetsMap.end()) {
                    sourceIt->second.erase(id);
                    if (sourceIt->second.empty()) {
                        _sourceToTargetsMap.erase(sourceIt);
                    }
                }
                // This source primitive needs to be updated
                auto bits = _dependencyToDirtyBitsMap[{id, source}];
                markPrimDirty(source, bits);
                _dependencyToDirtyBitsMap.erase({id, source});
            }

            // Erase the map from this target to all its sources
            _targetToSourcesMap.erase(id);
        }        
    }

    ArnoldDependencyChange dependencyChange;
    while (_dependencyTrackQueue.try_pop(dependencyChange)) {
        // We have a new list of dependencies for a given source.
        // We need to ensure that the previous dependencies were properly cleared
        const auto &newTargetsWithBits = dependencyChange.targets;
        const auto &source = dependencyChange.source;
        auto prevTargets = _sourceToTargetsMap[source];
        PathSet newTargets;
        for (const auto &pathAndBits : newTargetsWithBits) {
            newTargets.insert(pathAndBits.first);
            _dependencyToDirtyBitsMap.insert({{pathAndBits.first, source}, pathAndBits.second});
        }
        // Set the new targets for this source
        _sourceToTargetsMap[source] = newTargets;

        // Now check, for all targets that were set previously to this source,
        // if they're still present in the new list. If they're not, then we need
        // to remove the source from the target map
        for (const auto &prevTarget : prevTargets) {
            if (newTargets.find(prevTarget) == newTargets.end()) {
                _targetToSourcesMap[prevTarget].erase(source);
                _dependencyToDirtyBitsMap.erase({prevTarget, source});
            }
        }
        
        for (const auto& target : newTargets) {
            // for each target, we want to add all the source to its map
            _targetToSourcesMap[target].insert(source);
        }
    }
    
    // Finally, we're processing all the dependencies that were marked as dirty.
    // For each of them, we need to update all the sources pointing at it
    while (_dependencyDirtyQueue.try_pop(id)) {
        auto it = _targetToSourcesMap.find(id);
        if (it != _targetToSourcesMap.end()) {
            changes = true;
            // mark each source as being dirty
            for (const auto &source: it->second) {
                auto bits = _dependencyToDirtyBitsMap[{id, source}];
                markPrimDirty(source, bits);
            }
        }
    }

    // If we have connections in our stack, it means that some nodes were re-exported, 
    // and therefore that the render was already interrupted
    ProcessConnections();
    return changes;
}

bool HdArnoldRenderDelegate::IsPauseSupported() const { return false; }

bool HdArnoldRenderDelegate::IsStopSupported() const { return true; }

#if PXR_VERSION >= 2203
bool HdArnoldRenderDelegate::Stop(bool blocking)
#else
bool HdArnoldRenderDelegate::Stop()
#endif
{
    _renderParam->Pause();
    return true;
}

bool HdArnoldRenderDelegate::Resume()
{
    _renderParam->Resume();
    return true;
}

bool HdArnoldRenderDelegate::Restart()
{    
    _renderParam->Restart();
    return true;
}

#if PXR_VERSION >= 2203
bool HdArnoldRenderDelegate::IsStopped() const
{   
    int status = AiRenderGetStatus(GetRenderSession());
    return (status != AI_RENDER_STATUS_RENDERING && status != AI_RENDER_STATUS_RESTARTING);
}
#endif

const HdArnoldRenderDelegate::NativeRprimParamList* HdArnoldRenderDelegate::GetNativeRprimParamList(
    const AtString& arnoldNodeType) const
{
    const auto it = _nativeRprimParams.find(arnoldNodeType);
    return it == _nativeRprimParams.end() ? nullptr : &it->second;
}

void HdArnoldRenderDelegate::DirtyDependency(const SdfPath& id) 
{
    _dependencyDirtyQueue.emplace(id);
}

void HdArnoldRenderDelegate::RemoveDependency(const SdfPath& id) 
{
    _dependencyRemovalQueue.emplace(id); 
}

void HdArnoldRenderDelegate::TrackDependencies(const SdfPath& source, const PathSetWithDirtyBits& targets)
{
    _dependencyTrackQueue.emplace(source, targets);
}

void HdArnoldRenderDelegate::ClearDependencies(const SdfPath& source)
{
    // Originaly TrackDependencies(source, {});
    auto it = _sourceToTargetsMap.find(source);
    if (it != _sourceToTargetsMap.end()) {
        for(const auto &target: it->second) {
            _dependencyRemovalQueue.emplace(target);
        }
    }
}

void HdArnoldRenderDelegate::TrackRenderTag(AtNode* node, const TfToken& tag)
{
    if (node == nullptr)
        return;

    AiNodeSetDisabled(node, !IsVisibleRenderTag(tag));
    // If a specific render tag (i.e. node default nor geometry) is set on this 
    // primitive, we create a user data to track it later on
    if (tag != str::t__default && tag != str::t_geometry) {
        if (!AiNodeLookUpUserParameter(node, str::usd_purpose)) 
            AiNodeDeclare(node, str::usd_purpose, str::constantString);
        AiNodeSetStr(node, str::usd_purpose, AtString(tag.GetText()));
    }
}

bool HdArnoldRenderDelegate::SetRenderTags(const TfTokenVector& renderTags)
{
    // In this function we store the provided render tags, and we want to return
    // whether they have changed since the previous iteration
    if (renderTags == _renderTags)
        return false;

    // if the amount of elements has not changed, we also want to check if we're receiving 
    // the render tags in a different order
    bool renderTagsChanged = renderTags.size() != _renderTags.size();
    if (!renderTagsChanged) {
        for (auto t : renderTags) {
            if (std::find(_renderTags.begin(), _renderTags.end(), t) == _renderTags.end()) {
                renderTagsChanged = true;
                break;
            }
        }
    }
    _renderTags = renderTags;
    return renderTagsChanged;
}

AtNode* HdArnoldRenderDelegate::GetBackground(HdRenderIndex* renderIndex)
{
    const HdArnoldNodeGraph *nodeGraph = HdArnoldNodeGraph::GetNodeGraph(renderIndex, _background, this);
    if (nodeGraph)
        return nodeGraph->GetCachedTerminal(str::t_background);
    return nullptr;
}

AtNode* HdArnoldRenderDelegate::GetAtmosphere(HdRenderIndex* renderIndex)
{
    const HdArnoldNodeGraph *nodeGraph = HdArnoldNodeGraph::GetNodeGraph(renderIndex, _atmosphere, this);
    if (nodeGraph)
        return nodeGraph->GetCachedTerminal(str::t_atmosphere);
    return nullptr;
}

std::vector<AtNode*> HdArnoldRenderDelegate::GetAovShaders(HdRenderIndex* renderIndex)
{
    std::vector<AtNode *> nodes;
    for (const auto &materialPath: _aov_shaders) {
        HdArnoldNodeGraph *nodeGraph = HdArnoldNodeGraph::GetNodeGraph(renderIndex, materialPath, this);
        if (nodeGraph) {
            const auto &terminals = nodeGraph->GetCachedTerminals(_tokens->aovShadersArray);
            std::copy(terminals.begin(), terminals.end(), std::back_inserter(nodes));
        }
    }
    return nodes;
}

AtNode* HdArnoldRenderDelegate::GetImager(HdRenderIndex* renderIndex)
{
    const HdArnoldNodeGraph *nodeGraph = HdArnoldNodeGraph::GetNodeGraph(renderIndex, _imager, this);
    if (nodeGraph) {
        // Indicate to this node graph that it is an imager graph, so that 
        // it doesn't interrupt / restart the render when a node changes
        const_cast<HdArnoldNodeGraph*>(nodeGraph)->SetImagerGraph(true);
        return nodeGraph->GetCachedTerminal(str::t_input);
    }
    return nullptr;
}

AtNode* HdArnoldRenderDelegate::GetSubdivDicingCamera(HdRenderIndex* renderIndex)
{
    if (_subdiv_dicing_camera.IsEmpty())
        return nullptr;

    return LookupNode(_subdiv_dicing_camera.GetText());
}

AtNode* HdArnoldRenderDelegate::GetShaderOverride(HdRenderIndex* renderIndex)
{
    const HdArnoldNodeGraph *nodeGraph = HdArnoldNodeGraph::GetNodeGraph(renderIndex, _shader_override, this);
    if (nodeGraph)
        return nodeGraph->GetCachedTerminal(str::t_shader_override);
    return nullptr;
}

void HdArnoldRenderDelegate::RegisterCryptomatteDriver(const AtString& driver)
{
    _cryptomatteDrivers.insert(driver);
    SetHasCryptomatte(true);

}
void HdArnoldRenderDelegate::ClearCryptomatteDrivers()
{
   _cryptomatteDrivers.clear();
}
void HdArnoldRenderDelegate::SetInstancerCryptoOffset(AtNode *node, size_t numInstances)
{   
    // For cryptomatte to give consistent results with instances in hydra2, we need to use the 
    // user data "crypto_object" and "crypto_object_offset" that will be used by cryptomatte 
    // instead of the node names (prototype names aren't consistent in hydra2). 
    if (AiNodeLookUpUserParameter(node, str::instance_crypto_object_offset)) {
        AtArray *array = AiNodeGetArray(node, str::instance_crypto_object_offset);
        // if the crypto offsets were already set with the same amount of instances,
        // we don't need to do it again
        if (array && AiArrayGetNumElements(array) == numInstances)
            return;        
    }
    else
        AiNodeDeclare(node, str::instance_crypto_object_offset, "constant array INT");        
        
    std::vector<int> crypto_object_offsets(numInstances);
    for (size_t i = 0; i < numInstances; ++i)
        crypto_object_offsets[i] = i;
    AiNodeSetArray(node, str::instance_crypto_object_offset, AiArrayConvert(numInstances, 1, AI_TYPE_INT, &crypto_object_offsets[0]));
}

void HdArnoldRenderDelegate::SetHasCryptomatte(bool b)
{    
    if (b == _hasCryptomatte)
        return;
    _hasCryptomatte = b;
    if (_hasCryptomatte) {
        // We set this flag for the first time, let's iterate through all the 
        // instancers in the scene in order to set their user data for crypto_object_offset.
        // This is needed to ensure instancers have a consistent name for cryptomatte.
        AtNodeIterator* nodeIter = AiUniverseGetNodeIterator(_universe, AI_NODE_SHAPE);
        while (!AiNodeIteratorFinished(nodeIter))
        {
            AtNode *node = AiNodeIteratorGetNext(nodeIter);
            if (!AiNodeIs(node, str::instancer))
                continue;
            
            const AtArray *instanceMatrix = AiNodeGetArray(node, str::instance_matrix);
            size_t numInstances = instanceMatrix ? AiArrayGetNumElements(instanceMatrix) : 0;
            SetInstancerCryptoOffset(node, numInstances);
        }        
        AiNodeIteratorDestroy(nodeIter);
    }

}

#if PXR_VERSION >= 2108
HdCommandDescriptors HdArnoldRenderDelegate::GetCommandDescriptors() const
{
    HdCommandDescriptors descriptors;
    descriptors.emplace_back(TfToken("flush_texture"), "Flush textures");
    return descriptors;
}

bool HdArnoldRenderDelegate::InvokeCommand(const TfToken& command, const HdCommandArgs& args)
{
    if (command == TfToken("flush_texture")) {
        // Stop render
        _renderParam->Pause();
        // Flush texture
        AiUniverseCacheFlush(_universe, AI_CACHE_TEXTURE);
        // Restart the render
        _renderParam->Restart();
    }
    return false;
}
#endif // PXR_VERSION

PXR_NAMESPACE_CLOSE_SCOPE