layers/best_practices_utils.cpp

/* Copyright (c) 2015-2022 The Khronos Group Inc.
 * Copyright (c) 2015-2022 Valve Corporation
 * Copyright (c) 2015-2022 LunarG, Inc.
 * Modifications Copyright (C) 2020 Advanced Micro Devices, Inc. All rights reserved.
 *
 * 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.
 *
 * Author: Camden Stocker <camden@lunarg.com>
 * Author: Nadav Geva <nadav.geva@amd.com>
 */

#include "best_practices_validation.h"
#include "layer_chassis_dispatch.h"
#include "best_practices_error_enums.h"
#include "shader_validation.h"
#include "sync_utils.h"
#include "cmd_buffer_state.h"
#include "device_state.h"
#include "render_pass_state.h"

#include <string>
#include <bitset>
#include <memory>

struct VendorSpecificInfo {
    EnableFlags vendor_id;
    std::string name;
};

const std::map<BPVendorFlagBits, VendorSpecificInfo> kVendorInfo = {{kBPVendorArm, {vendor_specific_arm, "Arm"}},
                                                                    {kBPVendorAMD, {vendor_specific_amd, "AMD"}},
                                                                    {kBPVendorIMG, {vendor_specific_img, "IMG"}},
                                                                    {kBPVendorNVIDIA, {vendor_specific_nvidia, "NVIDIA"}}};

const SpecialUseVUIDs kSpecialUseInstanceVUIDs {
    kVUID_BestPractices_CreateInstance_SpecialUseExtension_CADSupport,
    kVUID_BestPractices_CreateInstance_SpecialUseExtension_D3DEmulation,
    kVUID_BestPractices_CreateInstance_SpecialUseExtension_DevTools,
    kVUID_BestPractices_CreateInstance_SpecialUseExtension_Debugging,
    kVUID_BestPractices_CreateInstance_SpecialUseExtension_GLEmulation,
};

const SpecialUseVUIDs kSpecialUseDeviceVUIDs {
    kVUID_BestPractices_CreateDevice_SpecialUseExtension_CADSupport,
    kVUID_BestPractices_CreateDevice_SpecialUseExtension_D3DEmulation,
    kVUID_BestPractices_CreateDevice_SpecialUseExtension_DevTools,
    kVUID_BestPractices_CreateDevice_SpecialUseExtension_Debugging,
    kVUID_BestPractices_CreateDevice_SpecialUseExtension_GLEmulation,
};

static constexpr std::array<VkFormat, 12> kCustomClearColorCompressedFormatsNVIDIA = {
    VK_FORMAT_R8G8B8A8_UNORM,           VK_FORMAT_B8G8R8A8_UNORM,           VK_FORMAT_A8B8G8R8_UNORM_PACK32,
    VK_FORMAT_A2R10G10B10_UNORM_PACK32, VK_FORMAT_A2B10G10R10_UNORM_PACK32, VK_FORMAT_R16G16B16A16_UNORM,
    VK_FORMAT_R16G16B16A16_SNORM,       VK_FORMAT_R16G16B16A16_UINT,        VK_FORMAT_R16G16B16A16_SINT,
    VK_FORMAT_R16G16B16A16_SFLOAT,      VK_FORMAT_R32G32B32A32_SFLOAT,      VK_FORMAT_B10G11R11_UFLOAT_PACK32,
};

ReadLockGuard BestPractices::ReadLock() {
    if (fine_grained_locking) {
        return ReadLockGuard(validation_object_mutex, std::defer_lock);
    } else {
        return ReadLockGuard(validation_object_mutex);
    }
}

WriteLockGuard BestPractices::WriteLock() {
    if (fine_grained_locking) {
        return WriteLockGuard(validation_object_mutex, std::defer_lock);
    } else {
        return WriteLockGuard(validation_object_mutex);
    }
}

std::shared_ptr<CMD_BUFFER_STATE> BestPractices::CreateCmdBufferState(VkCommandBuffer cb,
                                                                      const VkCommandBufferAllocateInfo* pCreateInfo,
                                                                      const COMMAND_POOL_STATE* pool) {
    return std::static_pointer_cast<CMD_BUFFER_STATE>(std::make_shared<bp_state::CommandBuffer>(this, cb, pCreateInfo, pool));
}

bp_state::CommandBuffer::CommandBuffer(BestPractices* bp, VkCommandBuffer cb, const VkCommandBufferAllocateInfo* pCreateInfo,
                                       const COMMAND_POOL_STATE* pool)
    : CMD_BUFFER_STATE(bp, cb, pCreateInfo, pool) {}

bool BestPractices::VendorCheckEnabled(BPVendorFlags vendors) const {
    for (const auto& vendor : kVendorInfo) {
        if (vendors & vendor.first && enabled[vendor.second.vendor_id]) {
            return true;
        }
    }
    return false;
}

const char* VendorSpecificTag(BPVendorFlags vendors) {
    // Cache built vendor tags in a map
    static layer_data::unordered_map<BPVendorFlags, std::string> tag_map;

    auto res = tag_map.find(vendors);
    if (res == tag_map.end()) {
        // Build the vendor tag string
        std::stringstream vendor_tag;

        vendor_tag << "[";
        bool first_vendor = true;
        for (const auto& vendor : kVendorInfo) {
            if (vendors & vendor.first) {
                if (!first_vendor) {
                    vendor_tag << ", ";
                }
                vendor_tag << vendor.second.name;
                first_vendor = false;
            }
        }
        vendor_tag << "]";

        tag_map[vendors] = vendor_tag.str();
        res = tag_map.find(vendors);
    }

    return res->second.c_str();
}

const char* DepReasonToString(ExtDeprecationReason reason) {
    switch (reason) {
        case kExtPromoted:
            return "promoted to";
            break;
        case kExtObsoleted:
            return "obsoleted by";
            break;
        case kExtDeprecated:
            return "deprecated by";
            break;
        default:
            return "";
            break;
    }
}

bool BestPractices::ValidateDeprecatedExtensions(const char* api_name, const char* extension_name, uint32_t version,
                                                 const char* vuid) const {
    bool skip = false;
    auto dep_info_it = deprecated_extensions.find(extension_name);
    if (dep_info_it != deprecated_extensions.end()) {
        auto dep_info = dep_info_it->second;
        if (((dep_info.target.compare("VK_VERSION_1_1") == 0) && (version >= VK_API_VERSION_1_1)) ||
            ((dep_info.target.compare("VK_VERSION_1_2") == 0) && (version >= VK_API_VERSION_1_2)) ||
            ((dep_info.target.compare("VK_VERSION_1_3") == 0) && (version >= VK_API_VERSION_1_3))) {
            skip |=
                LogWarning(instance, vuid, "%s(): Attempting to enable deprecated extension %s, but this extension has been %s %s.",
                           api_name, extension_name, DepReasonToString(dep_info.reason), (dep_info.target).c_str());
        } else if (dep_info.target.find("VK_VERSION") == std::string::npos) {
            if (dep_info.target.length() == 0) {
                skip |= LogWarning(instance, vuid,
                                   "%s(): Attempting to enable deprecated extension %s, but this extension has been deprecated "
                                   "without replacement.",
                                   api_name, extension_name);
            } else {
                skip |= LogWarning(instance, vuid,
                                   "%s(): Attempting to enable deprecated extension %s, but this extension has been %s %s.",
                                   api_name, extension_name, DepReasonToString(dep_info.reason), (dep_info.target).c_str());
            }
        }
    }
    return skip;
}

bool BestPractices::ValidateSpecialUseExtensions(const char* api_name, const char* extension_name, const SpecialUseVUIDs& special_use_vuids) const
{
    bool skip = false;
    auto dep_info_it = special_use_extensions.find(extension_name);

    if (dep_info_it != special_use_extensions.end()) {
        const char* const format = "%s(): Attempting to enable extension %s, but this extension is intended to support %s "
                                   "and it is strongly recommended that it be otherwise avoided.";
        auto& special_uses = dep_info_it->second;

        if (special_uses.find("cadsupport") != std::string::npos) {
            skip |= LogWarning(instance, special_use_vuids.cadsupport, format, api_name, extension_name,
                               "specialized functionality used by CAD/CAM applications");
        }
        if (special_uses.find("d3demulation") != std::string::npos) {
            skip |= LogWarning(instance, special_use_vuids.d3demulation, format, api_name, extension_name,
                "D3D emulation layers, and applications ported from D3D, by adding functionality specific to D3D");
        }
        if (special_uses.find("devtools") != std::string::npos) {
            skip |= LogWarning(instance, special_use_vuids.devtools, format, api_name, extension_name,
                "developer tools such as capture-replay libraries");
        }
        if (special_uses.find("debugging") != std::string::npos) {
            skip |= LogWarning(instance, special_use_vuids.debugging, format, api_name, extension_name,
                "use by applications when debugging");
        }
        if (special_uses.find("glemulation") != std::string::npos) {
            skip |= LogWarning(instance, special_use_vuids.glemulation, format, api_name, extension_name,
                "OpenGL and/or OpenGL ES emulation layers, and applications ported from those APIs, by adding functionality "
                "specific to those APIs");
        }
    }
    return skip;
}

bool BestPractices::PreCallValidateCreateInstance(const VkInstanceCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator,
                                                  VkInstance* pInstance) const {
    bool skip = false;

    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
        if (white_list(pCreateInfo->ppEnabledExtensionNames[i], kDeviceExtensionNames)) {
            skip |= LogWarning(instance, kVUID_BestPractices_CreateInstance_ExtensionMismatch,
                               "vkCreateInstance(): Attempting to enable Device Extension %s at CreateInstance time.",
                               pCreateInfo->ppEnabledExtensionNames[i]);
        }
        uint32_t specified_version =
            (pCreateInfo->pApplicationInfo ? pCreateInfo->pApplicationInfo->apiVersion : VK_API_VERSION_1_0);
        skip |= ValidateDeprecatedExtensions("CreateInstance", pCreateInfo->ppEnabledExtensionNames[i], specified_version,
                                             kVUID_BestPractices_CreateInstance_DeprecatedExtension);
        skip |= ValidateSpecialUseExtensions("CreateInstance", pCreateInfo->ppEnabledExtensionNames[i], kSpecialUseInstanceVUIDs);
    }

    return skip;
}

bool BestPractices::PreCallValidateCreateDevice(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo* pCreateInfo,
                                                const VkAllocationCallbacks* pAllocator, VkDevice* pDevice) const {
    bool skip = false;

    // get API version of physical device passed when creating device.
    VkPhysicalDeviceProperties physical_device_properties{};
    DispatchGetPhysicalDeviceProperties(physicalDevice, &physical_device_properties);
    auto device_api_version = physical_device_properties.apiVersion;

    // Check api versions and log an info message when instance api Version is higher than version on device.
    if (api_version > device_api_version) {
        std::string inst_api_name = StringAPIVersion(api_version);
        std::string dev_api_name = StringAPIVersion(device_api_version);

        skip |= LogInfo(device, kVUID_BestPractices_CreateDevice_API_Mismatch,
                           "vkCreateDevice(): API Version of current instance, %s is higher than API Version on device, %s",
                           inst_api_name.c_str(), dev_api_name.c_str());
    }

    std::vector<std::string> extensions;
    {
        uint32_t property_count = 0;
        if (DispatchEnumerateDeviceExtensionProperties(physicalDevice, nullptr, &property_count, nullptr) == VK_SUCCESS) {
            std::vector<VkExtensionProperties> property_list(property_count);
            if (DispatchEnumerateDeviceExtensionProperties(physicalDevice, nullptr, &property_count, property_list.data()) == VK_SUCCESS) {
                extensions.reserve(property_list.size());
                for (const VkExtensionProperties& properties : property_list) {
                    extensions.push_back(properties.extensionName);
                }
            }
        }
    }

    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
        const char *extension_name = pCreateInfo->ppEnabledExtensionNames[i];

        uint32_t extension_api_version = std::min(api_version, device_api_version);

        if (white_list(extension_name, kInstanceExtensionNames)) {
            skip |= LogWarning(instance, kVUID_BestPractices_CreateDevice_ExtensionMismatch,
                               "vkCreateDevice(): Attempting to enable Instance Extension %s at CreateDevice time.",
                               extension_name);
            extension_api_version = api_version;
        }

        skip |= ValidateDeprecatedExtensions("CreateDevice", extension_name, extension_api_version,
                                             kVUID_BestPractices_CreateDevice_DeprecatedExtension);
        skip |= ValidateSpecialUseExtensions("CreateDevice", extension_name, kSpecialUseDeviceVUIDs);
    }

    const auto bp_pd_state = Get<bp_state::PhysicalDevice>(physicalDevice);
    if ((bp_pd_state->vkGetPhysicalDeviceFeaturesState == UNCALLED) && (pCreateInfo->pEnabledFeatures != NULL)) {
        skip |= LogWarning(device, kVUID_BestPractices_CreateDevice_PDFeaturesNotCalled,
                           "vkCreateDevice() called before getting physical device features from vkGetPhysicalDeviceFeatures().");
    }

    if ((VendorCheckEnabled(kBPVendorArm) || VendorCheckEnabled(kBPVendorAMD) || VendorCheckEnabled(kBPVendorIMG)) &&
        (pCreateInfo->pEnabledFeatures != nullptr) && (pCreateInfo->pEnabledFeatures->robustBufferAccess == VK_TRUE)) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreateDevice_RobustBufferAccess,
            "%s %s %s: vkCreateDevice() called with enabled robustBufferAccess. Use robustBufferAccess as a debugging tool during "
            "development. Enabling it causes loss in performance for accesses to uniform buffers and shader storage "
            "buffers. Disable robustBufferAccess in release builds. Only leave it enabled if the application use-case "
            "requires the additional level of reliability due to the use of unverified user-supplied draw parameters.",
            VendorSpecificTag(kBPVendorArm), VendorSpecificTag(kBPVendorAMD), VendorSpecificTag(kBPVendorIMG));
    }

    const bool enabled_pageable_device_local_memory = IsExtEnabled(device_extensions.vk_ext_pageable_device_local_memory);
    if (VendorCheckEnabled(kBPVendorNVIDIA) && !enabled_pageable_device_local_memory &&
        std::find(extensions.begin(), extensions.end(), VK_EXT_PAGEABLE_DEVICE_LOCAL_MEMORY_EXTENSION_NAME) != extensions.end()) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreateDevice_PageableDeviceLocalMemory,
            "%s vkCreateDevice() called without pageable device local memory. "
            "Use pageableDeviceLocalMemory from VK_EXT_pageable_device_local_memory when it is available.",
            VendorSpecificTag(kBPVendorNVIDIA));
    }

    return skip;
}

bool BestPractices::PreCallValidateCreateBuffer(VkDevice device, const VkBufferCreateInfo* pCreateInfo,
                                                const VkAllocationCallbacks* pAllocator, VkBuffer* pBuffer) const {
    bool skip = false;

    if ((pCreateInfo->queueFamilyIndexCount > 1) && (pCreateInfo->sharingMode == VK_SHARING_MODE_EXCLUSIVE)) {
        std::stringstream buffer_hex;
        buffer_hex << "0x" << std::hex << HandleToUint64(pBuffer);

        skip |= LogWarning(
            device, kVUID_BestPractices_SharingModeExclusive,
            "Warning: Buffer (%s) specifies a sharing mode of VK_SHARING_MODE_EXCLUSIVE while specifying multiple queues "
            "(queueFamilyIndexCount of %" PRIu32 ").",
            buffer_hex.str().c_str(), pCreateInfo->queueFamilyIndexCount);
    }

    return skip;
}

bool BestPractices::PreCallValidateCreateImage(VkDevice device, const VkImageCreateInfo* pCreateInfo,
                                               const VkAllocationCallbacks* pAllocator, VkImage* pImage) const {
    bool skip = false;

    if ((pCreateInfo->queueFamilyIndexCount > 1) && (pCreateInfo->sharingMode == VK_SHARING_MODE_EXCLUSIVE)) {
        std::stringstream image_hex;
        image_hex << "0x" << std::hex << HandleToUint64(pImage);

        skip |=
            LogWarning(device, kVUID_BestPractices_SharingModeExclusive,
                       "Warning: Image (%s) specifies a sharing mode of VK_SHARING_MODE_EXCLUSIVE while specifying multiple queues "
                       "(queueFamilyIndexCount of %" PRIu32 ").",
                       image_hex.str().c_str(), pCreateInfo->queueFamilyIndexCount);
    }

    if ((pCreateInfo->flags & VK_IMAGE_CREATE_EXTENDED_USAGE_BIT) && !(pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT)) {
        skip |= LogWarning(device, kVUID_BestPractices_ImageCreateFlags,
                           "vkCreateImage(): pCreateInfo->flags has VK_IMAGE_CREATE_EXTENDED_USAGE_BIT set, but not "
                           "VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT, therefore image views created from this image will have to use the "
                           "same format and VK_IMAGE_CREATE_EXTENDED_USAGE_BIT will not have any effect.");
    }

    if (VendorCheckEnabled(kBPVendorArm) || VendorCheckEnabled(kBPVendorIMG)) {
        if (pCreateInfo->samples > VK_SAMPLE_COUNT_1_BIT && !(pCreateInfo->usage & VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT)) {
            skip |= LogPerformanceWarning(
                device, kVUID_BestPractices_CreateImage_NonTransientMSImage,
                "%s %s vkCreateImage(): Trying to create a multisampled image, but createInfo.usage did not have "
                "VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT set. Multisampled images may be resolved on-chip, "
                "and do not need to be backed by physical storage. "
                "TRANSIENT_ATTACHMENT allows tiled GPUs to not back the multisampled image with physical memory.",
                VendorSpecificTag(kBPVendorArm), VendorSpecificTag(kBPVendorIMG));
        }
    }

    if (VendorCheckEnabled(kBPVendorArm) && pCreateInfo->samples > kMaxEfficientSamplesArm) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreateImage_TooLargeSampleCount,
            "%s vkCreateImage(): Trying to create an image with %u samples. "
            "The hardware revision may not have full throughput for framebuffers with more than %u samples.",
            VendorSpecificTag(kBPVendorArm), static_cast<uint32_t>(pCreateInfo->samples), kMaxEfficientSamplesArm);
    }

    if (VendorCheckEnabled(kBPVendorIMG) && pCreateInfo->samples > kMaxEfficientSamplesImg) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreateImage_TooLargeSampleCount,
            "%s vkCreateImage(): Trying to create an image with %u samples. "
            "The device may not have full support for true multisampling for images with more than %u samples. "
            "XT devices support up to 8 samples, XE up to 4 samples.",
            VendorSpecificTag(kBPVendorIMG), static_cast<uint32_t>(pCreateInfo->samples), kMaxEfficientSamplesImg);
    }

    if (VendorCheckEnabled(kBPVendorIMG) && (pCreateInfo->format == VK_FORMAT_PVRTC1_2BPP_SRGB_BLOCK_IMG ||
                                             pCreateInfo->format == VK_FORMAT_PVRTC1_2BPP_UNORM_BLOCK_IMG ||
                                             pCreateInfo->format == VK_FORMAT_PVRTC1_4BPP_SRGB_BLOCK_IMG ||
                                             pCreateInfo->format == VK_FORMAT_PVRTC1_4BPP_UNORM_BLOCK_IMG ||
                                             pCreateInfo->format == VK_FORMAT_PVRTC2_2BPP_SRGB_BLOCK_IMG ||
                                             pCreateInfo->format == VK_FORMAT_PVRTC2_2BPP_UNORM_BLOCK_IMG ||
                                             pCreateInfo->format == VK_FORMAT_PVRTC2_4BPP_SRGB_BLOCK_IMG ||
                                             pCreateInfo->format == VK_FORMAT_PVRTC2_4BPP_UNORM_BLOCK_IMG)) {
        skip |= LogPerformanceWarning(device, kVUID_BestPractices_Texture_Format_PVRTC_Outdated,
                                      "%s vkCreateImage(): Trying to create an image with a PVRTC format. Both PVRTC1 and PVRTC2 "
                                      "are slower than standard image formats on PowerVR GPUs, prefer ETC, BC, ASTC, etc.",
                                      VendorSpecificTag(kBPVendorIMG));
    }

    if (VendorCheckEnabled(kBPVendorAMD)) {
        std::stringstream image_hex;
        image_hex << "0x" << std::hex << HandleToUint64(pImage);

        if ((pCreateInfo->usage & (VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT)) &&
            (pCreateInfo->sharingMode == VK_SHARING_MODE_CONCURRENT)) {
            skip |= LogPerformanceWarning(device,
                            kVUID_BestPractices_vkImage_AvoidConcurrentRenderTargets,
                            "%s Performance warning: image (%s) is created as a render target with VK_SHARING_MODE_CONCURRENT. "
                            "Using a SHARING_MODE_CONCURRENT "
                            "is not recommended with color and depth targets",
                            VendorSpecificTag(kBPVendorAMD), image_hex.str().c_str());
        }

        if ((pCreateInfo->usage &
             (VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_STORAGE_BIT | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT)) &&
            (pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT)) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_vkImage_DontUseMutableRenderTargets,
                        "%s Performance warning: image (%s) is created as a render target with VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT. "
                        "Using a MUTABLE_FORMAT is not recommended with color, depth, and storage targets",
                        VendorSpecificTag(kBPVendorAMD), image_hex.str().c_str());
        }

        if ((pCreateInfo->usage & (VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT)) &&
                (pCreateInfo->usage & VK_IMAGE_USAGE_STORAGE_BIT)) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_vkImage_DontUseStorageRenderTargets,
                        "%s Performance warning: image (%s) is created as a render target with VK_IMAGE_USAGE_STORAGE_BIT. Using a "
                        "VK_IMAGE_USAGE_STORAGE_BIT is not recommended with color and depth targets",
                        VendorSpecificTag(kBPVendorAMD), image_hex.str().c_str());
        }
    }

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        std::stringstream image_hex;
        image_hex << "0x" << std::hex << HandleToUint64(pImage);

        if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_CreateImage_TilingLinear,
                                          "%s Performance warning: image (%s) is created with tiling VK_IMAGE_TILING_LINEAR. "
                                          "Use VK_IMAGE_TILING_OPTIMAL instead.",
                VendorSpecificTag(kBPVendorNVIDIA), image_hex.str().c_str());
        }

        if (pCreateInfo->format == VK_FORMAT_D32_SFLOAT || pCreateInfo->format == VK_FORMAT_D32_SFLOAT_S8_UINT) {
            skip |= LogPerformanceWarning(
                device, kVUID_BestPractices_CreateImage_Depth32Format,
                "%s Performance warning: image (%s) is created with a 32-bit depth format. Use VK_FORMAT_D24_UNORM_S8_UINT or "
                "VK_FORMAT_D16_UNORM instead, unless the extra precision is needed.",
                VendorSpecificTag(kBPVendorNVIDIA), image_hex.str().c_str());
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateCreateSwapchainKHR(VkDevice device, const VkSwapchainCreateInfoKHR* pCreateInfo,
                                                      const VkAllocationCallbacks* pAllocator, VkSwapchainKHR* pSwapchain) const {
    bool skip = false;

    const auto* bp_pd_state = GetPhysicalDeviceState();
    if (bp_pd_state) {
        if (bp_pd_state->vkGetPhysicalDeviceSurfaceCapabilitiesKHRState == UNCALLED) {
            skip |= LogWarning(device, kVUID_BestPractices_Swapchain_GetSurfaceNotCalled,
                               "vkCreateSwapchainKHR() called before getting surface capabilities from "
                               "vkGetPhysicalDeviceSurfaceCapabilitiesKHR().");
        }

        if ((pCreateInfo->presentMode != VK_PRESENT_MODE_FIFO_KHR) &&
            (bp_pd_state->vkGetPhysicalDeviceSurfacePresentModesKHRState != QUERY_DETAILS)) {
            skip |= LogWarning(device, kVUID_BestPractices_Swapchain_GetSurfaceNotCalled,
                               "vkCreateSwapchainKHR() called before getting surface present mode(s) from "
                               "vkGetPhysicalDeviceSurfacePresentModesKHR().");
        }

        if (bp_pd_state->vkGetPhysicalDeviceSurfaceFormatsKHRState != QUERY_DETAILS) {
            skip |= LogWarning(
                device, kVUID_BestPractices_Swapchain_GetSurfaceNotCalled,
                "vkCreateSwapchainKHR() called before getting surface format(s) from vkGetPhysicalDeviceSurfaceFormatsKHR().");
        }
    }

    if ((pCreateInfo->queueFamilyIndexCount > 1) && (pCreateInfo->imageSharingMode == VK_SHARING_MODE_EXCLUSIVE)) {
        skip |=
            LogWarning(device, kVUID_BestPractices_SharingModeExclusive,
                       "Warning: A Swapchain is being created which specifies a sharing mode of VK_SHARING_MODE_EXCLUSIVE while "
                       "specifying multiple queues (queueFamilyIndexCount of %" PRIu32 ").",
                       pCreateInfo->queueFamilyIndexCount);
    }

    const auto present_mode = pCreateInfo->presentMode;
    if (((present_mode == VK_PRESENT_MODE_MAILBOX_KHR) || (present_mode == VK_PRESENT_MODE_FIFO_KHR)) &&
        (pCreateInfo->minImageCount == 2)) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_SuboptimalSwapchainImageCount,
            "Warning: A Swapchain is being created with minImageCount set to %" PRIu32
            ", which means double buffering is going "
            "to be used. Using double buffering and vsync locks rendering to an integer fraction of the vsync rate. In turn, "
            "reducing the performance of the application if rendering is slower than vsync. Consider setting minImageCount to "
            "3 to use triple buffering to maximize performance in such cases.",
            pCreateInfo->minImageCount);
    }

    if (VendorCheckEnabled(kBPVendorArm) && (pCreateInfo->presentMode != VK_PRESENT_MODE_FIFO_KHR)) {
        skip |= LogWarning(device, kVUID_BestPractices_CreateSwapchain_PresentMode,
                           "%s Warning: Swapchain is not being created with presentation mode \"VK_PRESENT_MODE_FIFO_KHR\". "
                           "Prefer using \"VK_PRESENT_MODE_FIFO_KHR\" to avoid unnecessary CPU and GPU load and save power. "
                           "Presentation modes which are not FIFO will present the latest available frame and discard other "
                           "frame(s) if any.",
                           VendorSpecificTag(kBPVendorArm));
    }

    return skip;
}

bool BestPractices::PreCallValidateCreateSharedSwapchainsKHR(VkDevice device, uint32_t swapchainCount,
                                                             const VkSwapchainCreateInfoKHR* pCreateInfos,
                                                             const VkAllocationCallbacks* pAllocator,
                                                             VkSwapchainKHR* pSwapchains) const {
    bool skip = false;

    for (uint32_t i = 0; i < swapchainCount; i++) {
        if ((pCreateInfos[i].queueFamilyIndexCount > 1) && (pCreateInfos[i].imageSharingMode == VK_SHARING_MODE_EXCLUSIVE)) {
            skip |= LogWarning(
                device, kVUID_BestPractices_SharingModeExclusive,
                "Warning: A shared swapchain (index %" PRIu32
                ") is being created which specifies a sharing mode of VK_SHARING_MODE_EXCLUSIVE while specifying multiple "
                "queues (queueFamilyIndexCount of %" PRIu32 ").",
                i, pCreateInfos[i].queueFamilyIndexCount);
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateCreateRenderPass(VkDevice device, const VkRenderPassCreateInfo* pCreateInfo,
                                                    const VkAllocationCallbacks* pAllocator, VkRenderPass* pRenderPass) const {
    bool skip = false;

    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; ++i) {
        VkFormat format = pCreateInfo->pAttachments[i].format;
        if (pCreateInfo->pAttachments[i].initialLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
            if ((FormatIsColor(format) || FormatHasDepth(format)) &&
                pCreateInfo->pAttachments[i].loadOp == VK_ATTACHMENT_LOAD_OP_LOAD) {
                skip |= LogWarning(device, kVUID_BestPractices_RenderPass_Attatchment,
                                   "Render pass has an attachment with loadOp == VK_ATTACHMENT_LOAD_OP_LOAD and "
                                   "initialLayout == VK_IMAGE_LAYOUT_UNDEFINED.  This is probably not what you "
                                   "intended.  Consider using VK_ATTACHMENT_LOAD_OP_DONT_CARE instead if the "
                                   "image truely is undefined at the start of the render pass.");
            }
            if (FormatHasStencil(format) && pCreateInfo->pAttachments[i].stencilLoadOp == VK_ATTACHMENT_LOAD_OP_LOAD) {
                skip |= LogWarning(device, kVUID_BestPractices_RenderPass_Attatchment,
                                   "Render pass has an attachment with stencilLoadOp == VK_ATTACHMENT_LOAD_OP_LOAD "
                                   "and initialLayout == VK_IMAGE_LAYOUT_UNDEFINED.  This is probably not what you "
                                   "intended.  Consider using VK_ATTACHMENT_LOAD_OP_DONT_CARE instead if the "
                                   "image truely is undefined at the start of the render pass.");
            }
        }

        const auto& attachment = pCreateInfo->pAttachments[i];
        if (attachment.samples > VK_SAMPLE_COUNT_1_BIT) {
            bool access_requires_memory =
                attachment.loadOp == VK_ATTACHMENT_LOAD_OP_LOAD || attachment.storeOp == VK_ATTACHMENT_STORE_OP_STORE;

            if (FormatHasStencil(format)) {
                access_requires_memory |= attachment.stencilLoadOp == VK_ATTACHMENT_LOAD_OP_LOAD ||
                                          attachment.stencilStoreOp == VK_ATTACHMENT_STORE_OP_STORE;
            }

            if (access_requires_memory) {
                skip |= LogPerformanceWarning(
                    device, kVUID_BestPractices_CreateRenderPass_ImageRequiresMemory,
                    "Attachment %u in the VkRenderPass is a multisampled image with %u samples, but it uses loadOp/storeOp "
                    "which requires accessing data from memory. Multisampled images should always be loadOp = CLEAR or DONT_CARE, "
                    "storeOp = DONT_CARE. This allows the implementation to use lazily allocated memory effectively.",
                    i, static_cast<uint32_t>(attachment.samples));
            }
        }
    }

    for (uint32_t dependency = 0; dependency < pCreateInfo->dependencyCount; dependency++) {
        skip |= CheckPipelineStageFlags("vkCreateRenderPass", pCreateInfo->pDependencies[dependency].srcStageMask);
        skip |= CheckPipelineStageFlags("vkCreateRenderPass", pCreateInfo->pDependencies[dependency].dstStageMask);
    }

    return skip;
}

bool BestPractices::ValidateAttachments(const VkRenderPassCreateInfo2* rpci, uint32_t attachmentCount,
                                        const VkImageView* image_views) const {
    bool skip = false;

    // Check for non-transient attachments that should be transient and vice versa
    for (uint32_t i = 0; i < attachmentCount; ++i) {
        const auto& attachment = rpci->pAttachments[i];
        bool attachment_should_be_transient =
            (attachment.loadOp != VK_ATTACHMENT_LOAD_OP_LOAD && attachment.storeOp != VK_ATTACHMENT_STORE_OP_STORE);

        if (FormatHasStencil(attachment.format)) {
            attachment_should_be_transient &= (attachment.stencilLoadOp != VK_ATTACHMENT_LOAD_OP_LOAD &&
                                               attachment.stencilStoreOp != VK_ATTACHMENT_STORE_OP_STORE);
        }

        auto view_state = Get<IMAGE_VIEW_STATE>(image_views[i]);
        if (view_state) {
            const auto& ici = view_state->image_state->createInfo;

            bool image_is_transient = (ici.usage & VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT) != 0;

            // The check for an image that should not be transient applies to all GPUs
            if (!attachment_should_be_transient && image_is_transient) {
                skip |= LogPerformanceWarning(
                    device, kVUID_BestPractices_CreateFramebuffer_AttachmentShouldNotBeTransient,
                    "Attachment %u in VkFramebuffer uses loadOp/storeOps which need to access physical memory, "
                    "but the image backing the image view has VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT set. "
                    "Physical memory will need to be backed lazily to this image, potentially causing stalls.",
                    i);
            }

            bool supports_lazy = false;
            for (uint32_t j = 0; j < phys_dev_mem_props.memoryTypeCount; j++) {
                if (phys_dev_mem_props.memoryTypes[j].propertyFlags & VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT) {
                    supports_lazy = true;
                }
            }

            // The check for an image that should be transient only applies to GPUs supporting
            // lazily allocated memory
            if (supports_lazy && attachment_should_be_transient && !image_is_transient) {
                skip |= LogPerformanceWarning(
                    device, kVUID_BestPractices_CreateFramebuffer_AttachmentShouldBeTransient,
                    "Attachment %u in VkFramebuffer uses loadOp/storeOps which never have to be backed by physical memory, "
                    "but the image backing the image view does not have VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT set. "
                    "You can save physical memory by using transient attachment backed by lazily allocated memory here.",
                    i);
            }
        }
    }
    return skip;
}

bool BestPractices::PreCallValidateCreateFramebuffer(VkDevice device, const VkFramebufferCreateInfo* pCreateInfo,
                                                     const VkAllocationCallbacks* pAllocator, VkFramebuffer* pFramebuffer) const {
    bool skip = false;

    auto rp_state = Get<RENDER_PASS_STATE>(pCreateInfo->renderPass);
    if (rp_state && !(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT)) {
        skip = ValidateAttachments(rp_state->createInfo.ptr(), pCreateInfo->attachmentCount, pCreateInfo->pAttachments);
    }

    return skip;
}

bool BestPractices::PreCallValidateAllocateDescriptorSets(VkDevice device, const VkDescriptorSetAllocateInfo* pAllocateInfo,
                                                          VkDescriptorSet* pDescriptorSets, void* ads_state_data) const {
    bool skip = false;
    skip |= ValidationStateTracker::PreCallValidateAllocateDescriptorSets(device, pAllocateInfo, pDescriptorSets, ads_state_data);

    if (!skip) {
        const auto pool_state = Get<bp_state::DescriptorPool>(pAllocateInfo->descriptorPool);
        // if the number of freed sets > 0, it implies they could be recycled instead if desirable
        // this warning is specific to Arm
        if (VendorCheckEnabled(kBPVendorArm) && pool_state && (pool_state->freed_count > 0)) {
            skip |= LogPerformanceWarning(
                device, kVUID_BestPractices_AllocateDescriptorSets_SuboptimalReuse,
                "%s Descriptor set memory was allocated via vkAllocateDescriptorSets() for sets which were previously freed in the "
                "same logical device. On some drivers or architectures it may be most optimal to re-use existing descriptor sets.",
                VendorSpecificTag(kBPVendorArm));
        }

        if (IsExtEnabled(device_extensions.vk_khr_maintenance1)) {
            // Track number of descriptorSets allowable in this pool
            if (pool_state->GetAvailableSets() < pAllocateInfo->descriptorSetCount) {
                skip |= LogWarning(pool_state->Handle(), kVUID_BestPractices_EmptyDescriptorPool,
                                 "vkAllocateDescriptorSets(): Unable to allocate %" PRIu32 " descriptorSets from %s"
                                 ". This pool only has %" PRIu32 " descriptorSets remaining.",
                                 pAllocateInfo->descriptorSetCount, report_data->FormatHandle(pool_state->Handle()).c_str(),
                                 pool_state->GetAvailableSets());
            }
        }
    }

    return skip;
}

void BestPractices::ManualPostCallRecordAllocateDescriptorSets(VkDevice device, const VkDescriptorSetAllocateInfo* pAllocateInfo,
                                                               VkDescriptorSet* pDescriptorSets, VkResult result, void* ads_state) {
    if (result == VK_SUCCESS) {
        auto pool_state = Get<bp_state::DescriptorPool>(pAllocateInfo->descriptorPool);
        if (pool_state) {
            // we record successful allocations by subtracting the allocation count from the last recorded free count
            const auto alloc_count = pAllocateInfo->descriptorSetCount;
            // clamp the unsigned subtraction to the range [0, last_free_count]
            if (pool_state->freed_count > alloc_count) {
                pool_state->freed_count -= alloc_count;
            } else {
                pool_state->freed_count = 0;
            }
        }
    }
}

void BestPractices::PostCallRecordFreeDescriptorSets(VkDevice device, VkDescriptorPool descriptorPool, uint32_t descriptorSetCount,
                                                     const VkDescriptorSet* pDescriptorSets, VkResult result) {
    ValidationStateTracker::PostCallRecordFreeDescriptorSets(device, descriptorPool, descriptorSetCount, pDescriptorSets, result);
    if (result == VK_SUCCESS) {
        auto pool_state = Get<bp_state::DescriptorPool>(descriptorPool);
        // we want to track frees because we're interested in suggesting re-use
        if (pool_state) {
            pool_state->freed_count += descriptorSetCount;
        }
    }
}

void BestPractices::PreCallRecordAllocateMemory(VkDevice device, const VkMemoryAllocateInfo* pAllocateInfo,
                                                const VkAllocationCallbacks* pAllocator, VkDeviceMemory* pMemory) {
    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        WriteLockGuard guard{memory_free_events_lock_};

        // Release old allocations to avoid overpopulating the container
        const auto now = std::chrono::high_resolution_clock::now();
        const auto last_old = std::find_if(memory_free_events_.rbegin(), memory_free_events_.rend(), [now](const MemoryFreeEvent& event) {
            return now - event.time > kAllocateMemoryReuseTimeThresholdNVIDIA;
        });
        memory_free_events_.erase(memory_free_events_.begin(), last_old.base());
    }
}

bool BestPractices::PreCallValidateAllocateMemory(VkDevice device, const VkMemoryAllocateInfo* pAllocateInfo,
                                                  const VkAllocationCallbacks* pAllocator, VkDeviceMemory* pMemory) const {
    bool skip = false;

    if ((Count<DEVICE_MEMORY_STATE>() + 1) > kMemoryObjectWarningLimit) {
        skip |= LogPerformanceWarning(device, kVUID_BestPractices_AllocateMemory_TooManyObjects,
                                      "Performance Warning: This app has > %" PRIu32 " memory objects.", kMemoryObjectWarningLimit);
    }

    if (pAllocateInfo->allocationSize < kMinDeviceAllocationSize) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_AllocateMemory_SmallAllocation,
            "vkAllocateMemory(): Allocating a VkDeviceMemory of size %" PRIu64 ". This is a very small allocation (current "
            "threshold is %" PRIu64 " bytes). "
            "You should make large allocations and sub-allocate from one large VkDeviceMemory.",
            pAllocateInfo->allocationSize, kMinDeviceAllocationSize);
    }

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        if (!device_extensions.vk_ext_pageable_device_local_memory &&
            !LvlFindInChain<VkMemoryPriorityAllocateInfoEXT>(pAllocateInfo->pNext)) {
            skip |= LogPerformanceWarning(
                device, kVUID_BestPractices_AllocateMemory_SetPriority,
                "%s Use VkMemoryPriorityAllocateInfoEXT to provide the operating system information on the allocations that "
                "should stay in video memory and which should be demoted first when video memory is limited. "
                "The highest priority should be given to GPU-written resources like color attachments, depth attachments, "
                "storage images, and buffers written from the GPU.",
                VendorSpecificTag(kBPVendorNVIDIA));
        }

        {
            // Size in bytes for an allocation to be considered "compatible"
            static constexpr VkDeviceSize size_threshold = VkDeviceSize{1} << 20;

            ReadLockGuard guard{memory_free_events_lock_};

            const auto now = std::chrono::high_resolution_clock::now();
            const VkDeviceSize alloc_size = pAllocateInfo->allocationSize;
            const uint32_t memory_type_index = pAllocateInfo->memoryTypeIndex;
            const auto latest_event = std::find_if(memory_free_events_.rbegin(), memory_free_events_.rend(), [&](const MemoryFreeEvent& event) {
                return (memory_type_index == event.memory_type_index) && (alloc_size <= event.allocation_size) &&
                       (alloc_size - event.allocation_size <= size_threshold) && (now - event.time < kAllocateMemoryReuseTimeThresholdNVIDIA);
            });

            if (latest_event != memory_free_events_.rend()) {
                const auto time_delta = std::chrono::duration_cast<std::chrono::milliseconds>(now - latest_event->time);
                if (time_delta < std::chrono::milliseconds{5}) {
                    skip |=
                        LogPerformanceWarning(device, kVUID_BestPractices_AllocateMemory_ReuseAllocations,
                                              "%s Reuse memory allocations instead of releasing and reallocating. A memory allocation "
                                              "has just been released, and it could have been reused in place of this allocation.",
                                              VendorSpecificTag(kBPVendorNVIDIA));
                } else {
                    const uint32_t seconds = static_cast<uint32_t>(time_delta.count() / 1000);
                    const uint32_t milliseconds = static_cast<uint32_t>(time_delta.count() % 1000);

                    skip |= LogPerformanceWarning(
                        device, kVUID_BestPractices_AllocateMemory_ReuseAllocations,
                        "%s Reuse memory allocations instead of releasing and reallocating. A memory allocation has been released "
                        "%" PRIu32 ".%03" PRIu32 " seconds ago, and it could have been reused in place of this allocation.",
                        VendorSpecificTag(kBPVendorNVIDIA), seconds, milliseconds);
                }
            }
        }
    }

    // TODO: Insert get check for GetPhysicalDeviceMemoryProperties once the state is tracked in the StateTracker

    return skip;
}

void BestPractices::ManualPostCallRecordAllocateMemory(VkDevice device, const VkMemoryAllocateInfo* pAllocateInfo,
                                                       const VkAllocationCallbacks* pAllocator, VkDeviceMemory* pMemory,
                                                       VkResult result) {
    if (result != VK_SUCCESS) {
        static std::vector<VkResult> error_codes = {VK_ERROR_OUT_OF_HOST_MEMORY, VK_ERROR_OUT_OF_DEVICE_MEMORY,
                                                    VK_ERROR_TOO_MANY_OBJECTS, VK_ERROR_INVALID_EXTERNAL_HANDLE,
                                                    VK_ERROR_INVALID_OPAQUE_CAPTURE_ADDRESS};
        static std::vector<VkResult> success_codes = {};
        ValidateReturnCodes("vkAllocateMemory", result, error_codes, success_codes);
        return;
    }
}

void BestPractices::ValidateReturnCodes(const char* api_name, VkResult result, const std::vector<VkResult>& error_codes,
                                        const std::vector<VkResult>& success_codes) const {
    auto error = std::find(error_codes.begin(), error_codes.end(), result);
    if (error != error_codes.end()) {
        static const std::vector<VkResult> common_failure_codes = {VK_ERROR_OUT_OF_DATE_KHR,
                                                                   VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT};

        auto common_failure = std::find(common_failure_codes.begin(), common_failure_codes.end(), result);
        if (common_failure != common_failure_codes.end()) {
            LogInfo(instance, kVUID_BestPractices_Failure_Result, "%s(): Returned error %s.", api_name, string_VkResult(result));
        } else {
            LogWarning(instance, kVUID_BestPractices_Error_Result, "%s(): Returned error %s.", api_name, string_VkResult(result));
        }
        return;
    }
    auto success = std::find(success_codes.begin(), success_codes.end(), result);
    if (success != success_codes.end()) {
        LogInfo(instance, kVUID_BestPractices_NonSuccess_Result, "%s(): Returned non-success return code %s.", api_name,
                string_VkResult(result));
    }
}

void BestPractices::PreCallRecordFreeMemory(VkDevice device, VkDeviceMemory memory, const VkAllocationCallbacks* pAllocator) {
    if (memory != VK_NULL_HANDLE && VendorCheckEnabled(kBPVendorNVIDIA)) {
        auto mem_info = Get<DEVICE_MEMORY_STATE>(memory);

        // Exclude memory free events on dedicated allocations, or imported/exported allocations.
        if (!mem_info->IsDedicatedBuffer() && !mem_info->IsDedicatedImage() && !mem_info->IsExport() && !mem_info->IsImport()) {
            MemoryFreeEvent event;
            event.time = std::chrono::high_resolution_clock::now();
            event.memory_type_index = mem_info->alloc_info.memoryTypeIndex;
            event.allocation_size = mem_info->alloc_info.allocationSize;

            WriteLockGuard guard{memory_free_events_lock_};
            memory_free_events_.push_back(event);
        }
    }

    ValidationStateTracker::PreCallRecordFreeMemory(device, memory, pAllocator);
}

bool BestPractices::PreCallValidateFreeMemory(VkDevice device, VkDeviceMemory memory,
                                              const VkAllocationCallbacks* pAllocator) const {
    if (memory == VK_NULL_HANDLE) return false;
    bool skip = false;

    auto mem_info = Get<DEVICE_MEMORY_STATE>(memory);

    for (const auto& item : mem_info->ObjectBindings()) {
        const auto& obj = item.first;
        LogObjectList objlist(device);
        objlist.add(obj);
        objlist.add(mem_info->mem());
        skip |= LogWarning(objlist, layer_name.c_str(), "VK Object %s still has a reference to mem obj %s.",
                           report_data->FormatHandle(obj).c_str(), report_data->FormatHandle(mem_info->mem()).c_str());
    }

    return skip;
}

bool BestPractices::ValidateBindBufferMemory(VkBuffer buffer, VkDeviceMemory memory, const char* api_name) const {
    bool skip = false;
    auto buffer_state = Get<BUFFER_STATE>(buffer);

    if (!buffer_state->memory_requirements_checked && !buffer_state->external_memory_handle) {
        skip |= LogWarning(device, kVUID_BestPractices_BufferMemReqNotCalled,
                           "%s: Binding memory to %s but vkGetBufferMemoryRequirements() has not been called on that buffer.",
                           api_name, report_data->FormatHandle(buffer).c_str());
    }

    auto mem_state = Get<DEVICE_MEMORY_STATE>(memory);

    if (mem_state && mem_state->alloc_info.allocationSize == buffer_state->createInfo.size &&
        mem_state->alloc_info.allocationSize < kMinDedicatedAllocationSize) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_SmallDedicatedAllocation,
            "%s: Trying to bind %s to a memory block which is fully consumed by the buffer. "
            "The required size of the allocation is %" PRIu64 ", but smaller buffers like this should be sub-allocated from "
            "larger memory blocks. (Current threshold is %" PRIu64 " bytes.)",
            api_name, report_data->FormatHandle(buffer).c_str(), mem_state->alloc_info.allocationSize, kMinDedicatedAllocationSize);
    }

    skip |= ValidateBindMemory(device, memory);

    return skip;
}

bool BestPractices::PreCallValidateBindBufferMemory(VkDevice device, VkBuffer buffer, VkDeviceMemory memory,
                                                    VkDeviceSize memoryOffset) const {
    bool skip = false;
    const char* api_name = "BindBufferMemory()";

    skip |= ValidateBindBufferMemory(buffer, memory, api_name);

    return skip;
}

bool BestPractices::PreCallValidateBindBufferMemory2(VkDevice device, uint32_t bindInfoCount,
                                                     const VkBindBufferMemoryInfo* pBindInfos) const {
    char api_name[64];
    bool skip = false;

    for (uint32_t i = 0; i < bindInfoCount; i++) {
        snprintf(api_name, sizeof(api_name), "vkBindBufferMemory2() pBindInfos[%u]", i);
        skip |= ValidateBindBufferMemory(pBindInfos[i].buffer, pBindInfos[i].memory, api_name);
    }

    return skip;
}

bool BestPractices::PreCallValidateBindBufferMemory2KHR(VkDevice device, uint32_t bindInfoCount,
                                                        const VkBindBufferMemoryInfo* pBindInfos) const {
    char api_name[64];
    bool skip = false;

    for (uint32_t i = 0; i < bindInfoCount; i++) {
        snprintf(api_name, sizeof(api_name), "vkBindBufferMemory2KHR() pBindInfos[%u]", i);
        skip |= ValidateBindBufferMemory(pBindInfos[i].buffer, pBindInfos[i].memory, api_name);
    }

    return skip;
}

bool BestPractices::ValidateBindImageMemory(VkImage image, VkDeviceMemory memory, const char* api_name) const {
    bool skip = false;
    auto image_state = Get<IMAGE_STATE>(image);

    if (image_state->disjoint == false) {
        if (!image_state->memory_requirements_checked[0] && !image_state->external_memory_handle) {
            skip |= LogWarning(device, kVUID_BestPractices_ImageMemReqNotCalled,
                               "%s: Binding memory to %s but vkGetImageMemoryRequirements() has not been called on that image.",
                               api_name, report_data->FormatHandle(image).c_str());
        }
    } else {
        // TODO If binding disjoint image then this needs to check that VkImagePlaneMemoryRequirementsInfo was called for each
        // plane.
    }

    auto mem_state = Get<DEVICE_MEMORY_STATE>(memory);

    if (mem_state->alloc_info.allocationSize == image_state->requirements[0].size &&
        mem_state->alloc_info.allocationSize < kMinDedicatedAllocationSize) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_SmallDedicatedAllocation,
            "%s: Trying to bind %s to a memory block which is fully consumed by the image. "
            "The required size of the allocation is %" PRIu64 ", but smaller images like this should be sub-allocated from "
            "larger memory blocks. (Current threshold is %" PRIu64 " bytes.)",
            api_name, report_data->FormatHandle(image).c_str(), mem_state->alloc_info.allocationSize, kMinDedicatedAllocationSize);
    }

    // If we're binding memory to a image which was created as TRANSIENT and the image supports LAZY allocation,
    // make sure this type is actually used.
    // This warning will only trigger if this layer is run on a platform that supports LAZILY_ALLOCATED_BIT
    // (i.e.most tile - based renderers)
    if (image_state->createInfo.usage & VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT) {
        bool supports_lazy = false;
        uint32_t suggested_type = 0;

        for (uint32_t i = 0; i < phys_dev_mem_props.memoryTypeCount; i++) {
            if ((1u << i) & image_state->requirements[0].memoryTypeBits) {
                if (phys_dev_mem_props.memoryTypes[i].propertyFlags & VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT) {
                    supports_lazy = true;
                    suggested_type = i;
                    break;
                }
            }
        }

        uint32_t allocated_properties = phys_dev_mem_props.memoryTypes[mem_state->alloc_info.memoryTypeIndex].propertyFlags;

        if (supports_lazy && (allocated_properties & VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT) == 0) {
            skip |= LogPerformanceWarning(
                device, kVUID_BestPractices_NonLazyTransientImage,
                "%s: Attempting to bind memory type %u to VkImage which was created with TRANSIENT_ATTACHMENT_BIT,"
                "but this memory type is not LAZILY_ALLOCATED_BIT. You should use memory type %u here instead to save "
                "%" PRIu64 " bytes of physical memory.",
                api_name, mem_state->alloc_info.memoryTypeIndex, suggested_type, image_state->requirements[0].size);
        }
    }

    skip |= ValidateBindMemory(device, memory);

    return skip;
}

bool BestPractices::PreCallValidateBindImageMemory(VkDevice device, VkImage image, VkDeviceMemory memory,
                                                   VkDeviceSize memoryOffset) const {
    bool skip = false;
    const char* api_name = "vkBindImageMemory()";

    skip |= ValidateBindImageMemory(image, memory, api_name);

    return skip;
}

bool BestPractices::PreCallValidateBindImageMemory2(VkDevice device, uint32_t bindInfoCount,
                                                    const VkBindImageMemoryInfo* pBindInfos) const {
    char api_name[64];
    bool skip = false;

    for (uint32_t i = 0; i < bindInfoCount; i++) {
        snprintf(api_name, sizeof(api_name), "vkBindImageMemory2() pBindInfos[%u]", i);
        if (!LvlFindInChain<VkBindImageMemorySwapchainInfoKHR>(pBindInfos[i].pNext)) {
            skip |= ValidateBindImageMemory(pBindInfos[i].image, pBindInfos[i].memory, api_name);
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateBindImageMemory2KHR(VkDevice device, uint32_t bindInfoCount,
                                                       const VkBindImageMemoryInfo* pBindInfos) const {
    char api_name[64];
    bool skip = false;

    for (uint32_t i = 0; i < bindInfoCount; i++) {
        snprintf(api_name, sizeof(api_name), "vkBindImageMemory2KHR() pBindInfos[%u]", i);
        skip |= ValidateBindImageMemory(pBindInfos[i].image, pBindInfos[i].memory, api_name);
    }

    return skip;
}

void BestPractices::PreCallRecordSetDeviceMemoryPriorityEXT(VkDevice device, VkDeviceMemory memory, float priority) {
    auto mem_info = std::static_pointer_cast<bp_state::DeviceMemory>(Get<DEVICE_MEMORY_STATE>(memory));
    mem_info->dynamic_priority.emplace(priority);
}

static inline bool FormatHasFullThroughputBlendingArm(VkFormat format) {
    switch (format) {
        case VK_FORMAT_B10G11R11_UFLOAT_PACK32:
        case VK_FORMAT_R16_SFLOAT:
        case VK_FORMAT_R16G16_SFLOAT:
        case VK_FORMAT_R16G16B16_SFLOAT:
        case VK_FORMAT_R16G16B16A16_SFLOAT:
        case VK_FORMAT_R32_SFLOAT:
        case VK_FORMAT_R32G32_SFLOAT:
        case VK_FORMAT_R32G32B32_SFLOAT:
        case VK_FORMAT_R32G32B32A32_SFLOAT:
            return false;

        default:
            return true;
    }
}

bool BestPractices::ValidateMultisampledBlendingArm(uint32_t createInfoCount,
                                                    const VkGraphicsPipelineCreateInfo* pCreateInfos) const {
    bool skip = false;

    for (uint32_t i = 0; i < createInfoCount; i++) {
        auto create_info = &pCreateInfos[i];

        if (!create_info->pColorBlendState || !create_info->pMultisampleState ||
            create_info->pMultisampleState->rasterizationSamples == VK_SAMPLE_COUNT_1_BIT ||
            create_info->pMultisampleState->sampleShadingEnable) {
            return skip;
        }

        auto rp_state = Get<RENDER_PASS_STATE>(create_info->renderPass);
        const auto& subpass = rp_state->createInfo.pSubpasses[create_info->subpass];

        // According to spec, pColorBlendState must be ignored if subpass does not have color attachments.
        uint32_t num_color_attachments = std::min(subpass.colorAttachmentCount, create_info->pColorBlendState->attachmentCount);

        for (uint32_t j = 0; j < num_color_attachments; j++) {
            const auto& blend_att = create_info->pColorBlendState->pAttachments[j];
            uint32_t att = subpass.pColorAttachments[j].attachment;

            if (att != VK_ATTACHMENT_UNUSED && blend_att.blendEnable && blend_att.colorWriteMask) {
                if (!FormatHasFullThroughputBlendingArm(rp_state->createInfo.pAttachments[att].format)) {
                    skip |= LogPerformanceWarning(device, kVUID_BestPractices_CreatePipelines_MultisampledBlending,
                                                  "%s vkCreateGraphicsPipelines() - createInfo #%u: Pipeline is multisampled and "
                                                  "color attachment #%u makes use "
                                                  "of a format which cannot be blended at full throughput when using MSAA.",
                                                  VendorSpecificTag(kBPVendorArm), i, j);
                }
            }
        }
    }

    return skip;
}

void BestPractices::ManualPostCallRecordCreateComputePipelines(VkDevice device, VkPipelineCache pipelineCache, uint32_t createInfoCount,
                                                         const VkComputePipelineCreateInfo* pCreateInfos,
                                                         const VkAllocationCallbacks* pAllocator, VkPipeline* pPipelines,
                                                         VkResult result, void* pipe_state) {
    // AMD best practice
    pipeline_cache_ = pipelineCache;
}

bool BestPractices::PreCallValidateCreateGraphicsPipelines(VkDevice device, VkPipelineCache pipelineCache, uint32_t createInfoCount,
                                                           const VkGraphicsPipelineCreateInfo* pCreateInfos,
                                                           const VkAllocationCallbacks* pAllocator, VkPipeline* pPipelines,
                                                           void* cgpl_state_data) const {
    bool skip = StateTracker::PreCallValidateCreateGraphicsPipelines(device, pipelineCache, createInfoCount, pCreateInfos,
                                                                     pAllocator, pPipelines, cgpl_state_data);
    if (skip) {
        return skip;
    }
    create_graphics_pipeline_api_state* cgpl_state = reinterpret_cast<create_graphics_pipeline_api_state*>(cgpl_state_data);

    if ((createInfoCount > 1) && (!pipelineCache)) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreatePipelines_MultiplePipelines,
            "Performance Warning: This vkCreateGraphicsPipelines call is creating multiple pipelines but is not using a "
            "pipeline cache, which may help with performance");
    }

    for (uint32_t i = 0; i < createInfoCount; i++) {
        const auto& create_info = pCreateInfos[i];

        if (!(cgpl_state->pipe_state[i]->active_shaders & VK_SHADER_STAGE_MESH_BIT_NV) && create_info.pVertexInputState) {
            const auto& vertex_input = *create_info.pVertexInputState;
            uint32_t count = 0;
            for (uint32_t j = 0; j < vertex_input.vertexBindingDescriptionCount; j++) {
                if (vertex_input.pVertexBindingDescriptions[j].inputRate == VK_VERTEX_INPUT_RATE_INSTANCE) {
                    count++;
                }
            }
            if (count > kMaxInstancedVertexBuffers) {
                skip |= LogPerformanceWarning(
                    device, kVUID_BestPractices_CreatePipelines_TooManyInstancedVertexBuffers,
                    "The pipeline is using %u instanced vertex buffers (current limit: %u), but this can be inefficient on the "
                    "GPU. If using instanced vertex attributes prefer interleaving them in a single buffer.",
                    count, kMaxInstancedVertexBuffers);
            }
        }

        if ((pCreateInfos[i].pRasterizationState) && (pCreateInfos[i].pRasterizationState->depthBiasEnable) &&
            (pCreateInfos[i].pRasterizationState->depthBiasConstantFactor == 0.0f) &&
            (pCreateInfos[i].pRasterizationState->depthBiasSlopeFactor == 0.0f) && VendorCheckEnabled(kBPVendorArm)) {
            skip |= LogPerformanceWarning(
                device, kVUID_BestPractices_CreatePipelines_DepthBias_Zero,
                "%s Performance Warning: This vkCreateGraphicsPipelines call is created with depthBiasEnable set to true "
                "and both depthBiasConstantFactor and depthBiasSlopeFactor are set to 0. This can cause reduced "
                "efficiency during rasterization. Consider disabling depthBias or increasing either "
                "depthBiasConstantFactor or depthBiasSlopeFactor.",
                VendorSpecificTag(kBPVendorArm));
        }

        skip |= VendorCheckEnabled(kBPVendorArm) && ValidateMultisampledBlendingArm(createInfoCount, pCreateInfos);
    }
    if (VendorCheckEnabled(kBPVendorAMD) || VendorCheckEnabled(kBPVendorNVIDIA)) {
        auto prev_pipeline = pipeline_cache_.load();
        if (pipelineCache && prev_pipeline && pipelineCache != prev_pipeline) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_CreatePipelines_MultiplePipelineCaches,
                                          "%s %s Performance Warning: A second pipeline cache is in use. "
                                          "Consider using only one pipeline cache to improve cache hit rate.",
                                          VendorSpecificTag(kBPVendorAMD), VendorSpecificTag(kBPVendorNVIDIA));
        }
    }
    if (VendorCheckEnabled(kBPVendorAMD)) {
        if (num_pso_ > kMaxRecommendedNumberOfPSOAMD) {
            skip |=
                LogPerformanceWarning(device, kVUID_BestPractices_CreatePipelines_TooManyPipelines,
                                          "%s Performance warning: Too many pipelines created, consider consolidation",
                                          VendorSpecificTag(kBPVendorAMD));
        }

        if (pCreateInfos->pInputAssemblyState && pCreateInfos->pInputAssemblyState->primitiveRestartEnable) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_CreatePipelines_AvoidPrimitiveRestart,
                                          "%s Performance warning: Use of primitive restart is not recommended",
                                          VendorSpecificTag(kBPVendorAMD));
        }

        // TODO: this might be too aggressive of a check
        if (pCreateInfos->pDynamicState && pCreateInfos->pDynamicState->dynamicStateCount > kDynamicStatesWarningLimitAMD) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_CreatePipelines_MinimizeNumDynamicStates,
                "%s Performance warning: Dynamic States usage incurs a performance cost. Ensure that they are truly needed",
                VendorSpecificTag(kBPVendorAMD));
        }
    }

    return skip;
}

static std::vector<bp_state::AttachmentInfo> GetAttachmentAccess(const safe_VkGraphicsPipelineCreateInfo& create_info,
                                                                 std::shared_ptr<const RENDER_PASS_STATE>& rp) {
    std::vector<bp_state::AttachmentInfo> result;
    if (!rp || rp->UsesDynamicRendering()) {
        return result;
    }

    const auto& subpass = rp->createInfo.pSubpasses[create_info.subpass];

    // NOTE: see PIPELINE_LAYOUT and safe_VkGraphicsPipelineCreateInfo constructors. pColorBlendState and pDepthStencilState
    // are only non-null if they are enabled.
    if (create_info.pColorBlendState) {
        // According to spec, pColorBlendState must be ignored if subpass does not have color attachments.
        uint32_t num_color_attachments = std::min(subpass.colorAttachmentCount, create_info.pColorBlendState->attachmentCount);
        for (uint32_t j = 0; j < num_color_attachments; j++) {
            if (create_info.pColorBlendState->pAttachments[j].colorWriteMask != 0) {
                uint32_t attachment = subpass.pColorAttachments[j].attachment;
                if (attachment != VK_ATTACHMENT_UNUSED) {
                    result.push_back({attachment, VK_IMAGE_ASPECT_COLOR_BIT});
                }
            }
        }
    }

    if (create_info.pDepthStencilState &&
        (create_info.pDepthStencilState->depthTestEnable || create_info.pDepthStencilState->depthBoundsTestEnable ||
         create_info.pDepthStencilState->stencilTestEnable)) {
        uint32_t attachment = subpass.pDepthStencilAttachment ? subpass.pDepthStencilAttachment->attachment : VK_ATTACHMENT_UNUSED;
        if (attachment != VK_ATTACHMENT_UNUSED) {
            VkImageAspectFlags aspects = 0;
            if (create_info.pDepthStencilState->depthTestEnable || create_info.pDepthStencilState->depthBoundsTestEnable) {
                aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
            }
            if (create_info.pDepthStencilState->stencilTestEnable) {
                aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
            }
            result.push_back({attachment, aspects});
        }
    }
    return result;
}

bp_state::Pipeline::Pipeline(const ValidationStateTracker* state_data, const VkGraphicsPipelineCreateInfo* pCreateInfo,
                             std::shared_ptr<const RENDER_PASS_STATE>&& rpstate,
                             std::shared_ptr<const PIPELINE_LAYOUT_STATE>&& layout)
    : PIPELINE_STATE(state_data, pCreateInfo, std::move(rpstate), std::move(layout)),
      access_framebuffer_attachments(GetAttachmentAccess(create_info.graphics, rp_state)) {}

std::shared_ptr<PIPELINE_STATE> BestPractices::CreateGraphicsPipelineState(
    const VkGraphicsPipelineCreateInfo* pCreateInfo, std::shared_ptr<const RENDER_PASS_STATE>&& render_pass,
    std::shared_ptr<const PIPELINE_LAYOUT_STATE>&& layout) const {
    return std::static_pointer_cast<PIPELINE_STATE>(
        std::make_shared<bp_state::Pipeline>(this, pCreateInfo, std::move(render_pass), std::move(layout)));
}

void BestPractices::ManualPostCallRecordCreateGraphicsPipelines(VkDevice device, VkPipelineCache pipelineCache, uint32_t count,
                                                                const VkGraphicsPipelineCreateInfo* pCreateInfos,
                                                                const VkAllocationCallbacks* pAllocator, VkPipeline* pPipelines,
                                                                VkResult result, void* cgpl_state_data) {
    // AMD best practice
    pipeline_cache_ = pipelineCache;
}

bool BestPractices::PreCallValidateCreateComputePipelines(VkDevice device, VkPipelineCache pipelineCache, uint32_t createInfoCount,
                                                          const VkComputePipelineCreateInfo* pCreateInfos,
                                                          const VkAllocationCallbacks* pAllocator, VkPipeline* pPipelines,
                                                          void* ccpl_state_data) const {
    bool skip = StateTracker::PreCallValidateCreateComputePipelines(device, pipelineCache, createInfoCount, pCreateInfos,
                                                                    pAllocator, pPipelines, ccpl_state_data);

    if ((createInfoCount > 1) && (!pipelineCache)) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreatePipelines_MultiplePipelines,
            "Performance Warning: This vkCreateComputePipelines call is creating multiple pipelines but is not using a "
            "pipeline cache, which may help with performance");
    }

    if (VendorCheckEnabled(kBPVendorAMD)) {
        auto prev_pipeline = pipeline_cache_.load();
        if (pipelineCache && prev_pipeline && pipelineCache != prev_pipeline) {
            skip |= LogPerformanceWarning(
                device, kVUID_BestPractices_CreatePipelines_MultiplePipelines,
                "%s Performance Warning: A second pipeline cache is in use. Consider using only one pipeline cache to "
                "improve cache hit rate",
                VendorSpecificTag(kBPVendorAMD));
        }
    }

    for (uint32_t i = 0; i < createInfoCount; i++) {
        const VkComputePipelineCreateInfo& createInfo = pCreateInfos[i];
        if (VendorCheckEnabled(kBPVendorArm)) {
            skip |= ValidateCreateComputePipelineArm(createInfo);
        }

        if (IsExtEnabled(device_extensions.vk_khr_maintenance4)) {
            auto module_state = Get<SHADER_MODULE_STATE>(createInfo.stage.module);
            for (const Instruction* inst : module_state->GetBuiltinDecorationList()) {
                if (inst->GetBuiltIn() == spv::BuiltInWorkgroupSize) {
                    skip |= LogWarning(device, kVUID_BestPractices_SpirvDeprecated_WorkgroupSize,
                                       "vkCreateComputePipelines(): pCreateInfos[ %" PRIu32
                                       "] is using the Workgroup built-in which SPIR-V 1.6 deprecated. The VK_KHR_maintenance4 "
                                       "extension exposes a new LocalSizeId execution mode that should be used instead.",
                                       i);
                }
            }
        }
    }

    return skip;
}

bool BestPractices::ValidateCreateComputePipelineArm(const VkComputePipelineCreateInfo& createInfo) const {
    bool skip = false;
    auto module_state = Get<SHADER_MODULE_STATE>(createInfo.stage.module);
    // Generate warnings about work group sizes based on active resources.
    auto entrypoint_optional = module_state->FindEntrypoint(createInfo.stage.pName, createInfo.stage.stage);
    if (!entrypoint_optional) return false;

    const Instruction& entrypoint = *entrypoint_optional;
    uint32_t x = 1, y = 1, z = 1;
    module_state->FindLocalSize(entrypoint, x, y, z);

    uint32_t thread_count = x * y * z;

    // Generate a priori warnings about work group sizes.
    if (thread_count > kMaxEfficientWorkGroupThreadCountArm) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreateComputePipelines_ComputeWorkGroupSize,
            "%s vkCreateComputePipelines(): compute shader with work group dimensions (%u, %u, "
            "%u) (%u threads total), has more threads than advised in a single work group. It is advised to use work "
            "groups with less than %u threads, especially when using barrier() or shared memory.",
            VendorSpecificTag(kBPVendorArm), x, y, z, thread_count, kMaxEfficientWorkGroupThreadCountArm);
    }

    if (thread_count == 1 || ((x > 1) && (x & (kThreadGroupDispatchCountAlignmentArm - 1))) ||
        ((y > 1) && (y & (kThreadGroupDispatchCountAlignmentArm - 1))) ||
        ((z > 1) && (z & (kThreadGroupDispatchCountAlignmentArm - 1)))) {
        skip |= LogPerformanceWarning(device, kVUID_BestPractices_CreateComputePipelines_ComputeThreadGroupAlignment,
                                      "%s vkCreateComputePipelines(): compute shader with work group dimensions (%u, "
                                      "%u, %u) is not aligned to %u "
                                      "threads. On Arm Mali architectures, not aligning work group sizes to %u may "
                                      "leave threads idle on the shader "
                                      "core.",
                                      VendorSpecificTag(kBPVendorArm), x, y, z, kThreadGroupDispatchCountAlignmentArm,
                                      kThreadGroupDispatchCountAlignmentArm);
    }

    auto accessible_ids = module_state->MarkAccessibleIds(entrypoint_optional);
    auto descriptor_uses = module_state->CollectInterfaceByDescriptorSlot(accessible_ids);

    unsigned dimensions = 0;
    if (x > 1) dimensions++;
    if (y > 1) dimensions++;
    if (z > 1) dimensions++;
    // Here the dimension will really depend on the dispatch grid, but assume it's 1D.
    dimensions = std::max(dimensions, 1u);

    // If we're accessing images, we almost certainly want to have a 2D workgroup for cache reasons.
    // There are some false positives here. We could simply have a shader that does this within a 1D grid,
    // or we may have a linearly tiled image, but these cases are quite unlikely in practice.
    bool accesses_2d = false;
    for (const auto& usage : descriptor_uses) {
        auto dim = module_state->GetShaderResourceDimensionality(usage.second);
        if (dim < 0) continue;
        auto spvdim = spv::Dim(dim);
        if (spvdim != spv::Dim1D && spvdim != spv::DimBuffer) accesses_2d = true;
    }

    if (accesses_2d && dimensions < 2) {
        LogPerformanceWarning(device, kVUID_BestPractices_CreateComputePipelines_ComputeSpatialLocality,
                              "%s vkCreateComputePipelines(): compute shader has work group dimensions (%u, %u, %u), which "
                              "suggests a 1D dispatch, but the shader is accessing 2D or 3D images. The shader may be "
                              "exhibiting poor spatial locality with respect to one or more shader resources.",
                              VendorSpecificTag(kBPVendorArm), x, y, z);
    }

    return skip;
}

bool BestPractices::CheckPipelineStageFlags(const std::string& api_name, VkPipelineStageFlags flags) const {
    bool skip = false;

    if (flags & VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT) {
        skip |= LogWarning(device, kVUID_BestPractices_PipelineStageFlags,
                           "You are using VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT when %s is called\n", api_name.c_str());
    } else if (flags & VK_PIPELINE_STAGE_ALL_COMMANDS_BIT) {
        skip |= LogWarning(device, kVUID_BestPractices_PipelineStageFlags,
                           "You are using VK_PIPELINE_STAGE_ALL_COMMANDS_BIT when %s is called\n", api_name.c_str());
    }

    return skip;
}

bool BestPractices::CheckPipelineStageFlags(const std::string& api_name, VkPipelineStageFlags2KHR flags) const {
    bool skip = false;

    if (flags & VK_PIPELINE_STAGE_2_ALL_GRAPHICS_BIT_KHR) {
        skip |= LogWarning(device, kVUID_BestPractices_PipelineStageFlags,
                           "You are using VK_PIPELINE_STAGE_2_ALL_GRAPHICS_BIT_KHR when %s is called\n", api_name.c_str());
    } else if (flags & VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT_KHR) {
        skip |= LogWarning(device, kVUID_BestPractices_PipelineStageFlags,
                           "You are using VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT_KHR when %s is called\n", api_name.c_str());
    }

    return skip;
}

bool BestPractices::CheckDependencyInfo(const std::string& api_name, const VkDependencyInfoKHR& dep_info) const {
    bool skip = false;
    auto stage_masks = sync_utils::GetGlobalStageMasks(dep_info);

    skip |= CheckPipelineStageFlags(api_name, stage_masks.src);
    skip |= CheckPipelineStageFlags(api_name, stage_masks.dst);
    for (uint32_t i = 0; i < dep_info.imageMemoryBarrierCount; ++i) {
        skip |= ValidateImageMemoryBarrier(
            api_name, dep_info.pImageMemoryBarriers[i].oldLayout, dep_info.pImageMemoryBarriers[i].newLayout,
            dep_info.pImageMemoryBarriers[i].srcAccessMask, dep_info.pImageMemoryBarriers[i].dstAccessMask,
            dep_info.pImageMemoryBarriers[i].subresourceRange.aspectMask);
    }

    return skip;
}

void BestPractices::ManualPostCallRecordQueuePresentKHR(VkQueue queue, const VkPresentInfoKHR* pPresentInfo, VkResult result) {
    for (uint32_t i = 0; i < pPresentInfo->swapchainCount; ++i) {
        auto swapchains_result = pPresentInfo->pResults ? pPresentInfo->pResults[i] : result;
        if (swapchains_result == VK_SUBOPTIMAL_KHR) {
            LogPerformanceWarning(
                pPresentInfo->pSwapchains[i], kVUID_BestPractices_SuboptimalSwapchain,
                "vkQueuePresentKHR: %s :VK_SUBOPTIMAL_KHR was returned. VK_SUBOPTIMAL_KHR - Presentation will still succeed, "
                "subject to the window resize behavior, but the swapchain is no longer configured optimally for the surface it "
                "targets. Applications should query updated surface information and recreate their swapchain at the next "
                "convenient opportunity.",
                report_data->FormatHandle(pPresentInfo->pSwapchains[i]).c_str());
        }
    }

    // AMD best practice
    // end-of-frame cleanup
    num_queue_submissions_ = 0;
    num_barriers_objects_ = 0;
    ClearPipelinesUsedInFrame();
}

bool BestPractices::PreCallValidateQueueSubmit(VkQueue queue, uint32_t submitCount, const VkSubmitInfo* pSubmits,
                                               VkFence fence) const {
    bool skip = false;

    for (uint32_t submit = 0; submit < submitCount; submit++) {
        for (uint32_t semaphore = 0; semaphore < pSubmits[submit].waitSemaphoreCount; semaphore++) {
            skip |= CheckPipelineStageFlags("vkQueueSubmit", pSubmits[submit].pWaitDstStageMask[semaphore]);
        }
        if (pSubmits[submit].signalSemaphoreCount == 0 && pSubmits[submit].pSignalSemaphores != nullptr) {
            skip |=
                LogWarning(device, kVUID_BestPractices_SemaphoreCount,
                           "pSubmits[%" PRIu32 "].pSignalSemaphores is set, but pSubmits[%" PRIu32 "].signalSemaphoreCount is 0.",
                           submit, submit);
        }
        if (pSubmits[submit].waitSemaphoreCount == 0 && pSubmits[submit].pWaitSemaphores != nullptr) {
            skip |= LogWarning(device, kVUID_BestPractices_SemaphoreCount,
                               "pSubmits[%" PRIu32 "].pWaitSemaphores is set, but pSubmits[%" PRIu32 "].waitSemaphoreCount is 0.",
                               submit, submit);
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateQueueSubmit2KHR(VkQueue queue, uint32_t submitCount, const VkSubmitInfo2KHR* pSubmits,
                                                   VkFence fence) const {
    bool skip = false;

    for (uint32_t submit = 0; submit < submitCount; submit++) {
        for (uint32_t semaphore = 0; semaphore < pSubmits[submit].waitSemaphoreInfoCount; semaphore++) {
            skip |= CheckPipelineStageFlags("vkQueueSubmit2KHR", pSubmits[submit].pWaitSemaphoreInfos[semaphore].stageMask);
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateQueueSubmit2(VkQueue queue, uint32_t submitCount, const VkSubmitInfo2* pSubmits,
                                                VkFence fence) const {
    bool skip = false;

    for (uint32_t submit = 0; submit < submitCount; submit++) {
        for (uint32_t semaphore = 0; semaphore < pSubmits[submit].waitSemaphoreInfoCount; semaphore++) {
            skip |= CheckPipelineStageFlags("vkQueueSubmit2", pSubmits[submit].pWaitSemaphoreInfos[semaphore].stageMask);
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateCreateCommandPool(VkDevice device, const VkCommandPoolCreateInfo* pCreateInfo,
                                                     const VkAllocationCallbacks* pAllocator, VkCommandPool* pCommandPool) const {
    bool skip = false;

    if (pCreateInfo->flags & VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT) {
        skip |= LogPerformanceWarning(
            device, kVUID_BestPractices_CreateCommandPool_CommandBufferReset,
            "vkCreateCommandPool(): VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT is set. Consider resetting entire "
            "pool instead.");
    }

    return skip;
}

void BestPractices::PreCallRecordBeginCommandBuffer(VkCommandBuffer commandBuffer,
                                                    const VkCommandBufferBeginInfo* pBeginInfo) {
    StateTracker::PreCallRecordBeginCommandBuffer(commandBuffer, pBeginInfo);

    auto cb = GetWrite<bp_state::CommandBuffer>(commandBuffer);
    if (!cb) return;

    cb->num_submits = 0;
    cb->is_one_time_submit = (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT) != 0;
}

bool BestPractices::PreCallValidateBeginCommandBuffer(VkCommandBuffer commandBuffer,
                                                      const VkCommandBufferBeginInfo* pBeginInfo) const {
    bool skip = false;

    if (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT) {
        skip |= LogPerformanceWarning(device, kVUID_BestPractices_BeginCommandBuffer_SimultaneousUse,
                                      "vkBeginCommandBuffer(): VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT is set.");
    }

    if (VendorCheckEnabled(kBPVendorArm)) {
        if (!(pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT)) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_BeginCommandBuffer_OneTimeSubmit,
                                          "%s vkBeginCommandBuffer(): VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT is not set. "
                                          "For best performance on Mali GPUs, consider setting ONE_TIME_SUBMIT by default.",
                                          VendorSpecificTag(kBPVendorArm));
        }
    }
    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        auto cb = GetRead<bp_state::CommandBuffer>(commandBuffer);
        if (cb->num_submits == 1 && !cb->is_one_time_submit) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_BeginCommandBuffer_OneTimeSubmit,
                                          "%s vkBeginCommandBuffer(): VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT was not set "
                                          "and the command buffer has only been submitted once. "
                                          "For best performance on NVIDIA GPUs, use ONE_TIME_SUBMIT.",
                                          VendorSpecificTag(kBPVendorNVIDIA));
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateCmdSetEvent(VkCommandBuffer commandBuffer, VkEvent event, VkPipelineStageFlags stageMask) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdSetEvent", stageMask);

    return skip;
}

bool BestPractices::PreCallValidateCmdSetEvent2KHR(VkCommandBuffer commandBuffer, VkEvent event,
                                                   const VkDependencyInfoKHR* pDependencyInfo) const {
    return CheckDependencyInfo("vkCmdSetEvent2KHR", *pDependencyInfo);
}

bool BestPractices::PreCallValidateCmdSetEvent2(VkCommandBuffer commandBuffer, VkEvent event,
                                                const VkDependencyInfo* pDependencyInfo) const {
    return CheckDependencyInfo("vkCmdSetEvent2", *pDependencyInfo);
}

bool BestPractices::PreCallValidateCmdResetEvent(VkCommandBuffer commandBuffer, VkEvent event,
                                                 VkPipelineStageFlags stageMask) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdResetEvent", stageMask);

    return skip;
}

bool BestPractices::PreCallValidateCmdResetEvent2KHR(VkCommandBuffer commandBuffer, VkEvent event,
                                                     VkPipelineStageFlags2KHR stageMask) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdResetEvent2KHR", stageMask);

    return skip;
}

bool BestPractices::PreCallValidateCmdResetEvent2(VkCommandBuffer commandBuffer, VkEvent event,
                                                  VkPipelineStageFlags2 stageMask) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdResetEvent2", stageMask);

    return skip;
}

bool BestPractices::PreCallValidateCmdWaitEvents(VkCommandBuffer commandBuffer, uint32_t eventCount, const VkEvent* pEvents,
                                                 VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask,
                                                 uint32_t memoryBarrierCount, const VkMemoryBarrier* pMemoryBarriers,
                                                 uint32_t bufferMemoryBarrierCount,
                                                 const VkBufferMemoryBarrier* pBufferMemoryBarriers,
                                                 uint32_t imageMemoryBarrierCount,
                                                 const VkImageMemoryBarrier* pImageMemoryBarriers) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdWaitEvents", srcStageMask);
    skip |= CheckPipelineStageFlags("vkCmdWaitEvents", dstStageMask);

    return skip;
}

bool BestPractices::PreCallValidateCmdWaitEvents2KHR(VkCommandBuffer commandBuffer, uint32_t eventCount, const VkEvent* pEvents,
                                                     const VkDependencyInfoKHR* pDependencyInfos) const {
    bool skip = false;
    for (uint32_t i = 0; i < eventCount; i++) {
        skip = CheckDependencyInfo("vkCmdWaitEvents2KHR", pDependencyInfos[i]);
    }

    return skip;
}

bool BestPractices::PreCallValidateCmdWaitEvents2(VkCommandBuffer commandBuffer, uint32_t eventCount, const VkEvent* pEvents,
                                                     const VkDependencyInfo* pDependencyInfos) const {
    bool skip = false;
    for (uint32_t i = 0; i < eventCount; i++) {
        skip = CheckDependencyInfo("vkCmdWaitEvents2", pDependencyInfos[i]);
    }

    return skip;
}

bool BestPractices::ValidateAccessLayoutCombination(const std::string& api_name, VkAccessFlags2 access, VkImageLayout layout,
                                                    VkImageAspectFlags aspect) const {
    bool skip = false;

    const VkAccessFlags all = VK_ACCESS_FLAG_BITS_MAX_ENUM;
    bool none_allowed = false;
    VkAccessFlags allowed = 0;

    // Combinations taken from https://github.com/KhronosGroup/Vulkan-ValidationLayers/issues/2918
    switch (layout) {
        case VK_IMAGE_LAYOUT_UNDEFINED:
            allowed = all;
            none_allowed = true;
            break;
        case VK_IMAGE_LAYOUT_GENERAL:
            allowed = all;
            break;
        case VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL:
            allowed = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
                      VK_ACCESS_COLOR_ATTACHMENT_READ_NONCOHERENT_BIT_EXT;
            break;
        case VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL:
            allowed = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
            break;
        case VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL:
            allowed = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
            break;
        case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL:
            allowed = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT | VK_ACCESS_SHADER_READ_BIT;
            break;
        case VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL:
            allowed = VK_ACCESS_TRANSFER_READ_BIT;
            break;
        case VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL:
            allowed = VK_ACCESS_TRANSFER_WRITE_BIT;
            break;
        case VK_IMAGE_LAYOUT_PREINITIALIZED:
            allowed = VK_ACCESS_HOST_WRITE_BIT;
            break;
        case VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL:
            if (aspect & VK_IMAGE_ASPECT_DEPTH_BIT) {
                allowed |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
            }
            if (aspect & VK_IMAGE_ASPECT_STENCIL_BIT) {
                allowed |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
            }
            break;
        case VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL:
            if (aspect & VK_IMAGE_ASPECT_DEPTH_BIT) {
                allowed |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
            }
            if (aspect & VK_IMAGE_ASPECT_STENCIL_BIT) {
                allowed |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
            }
            break;
        case VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL:
            allowed = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
            break;
        case VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_OPTIMAL:
            allowed = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
            break;
        case VK_IMAGE_LAYOUT_STENCIL_ATTACHMENT_OPTIMAL:
            allowed = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
            break;
        case VK_IMAGE_LAYOUT_STENCIL_READ_ONLY_OPTIMAL:
            allowed = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
            break;
        case VK_IMAGE_LAYOUT_PRESENT_SRC_KHR:
            allowed = VK_ACCESS_NONE;
            break;
        case VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR:
            allowed = all;
            none_allowed = true;
            break;
        case VK_IMAGE_LAYOUT_SHADING_RATE_OPTIMAL_NV:
            allowed = VK_ACCESS_SHADING_RATE_IMAGE_READ_BIT_NV;
            break;
        case VK_IMAGE_LAYOUT_FRAGMENT_DENSITY_MAP_OPTIMAL_EXT:
            allowed = VK_ACCESS_FRAGMENT_DENSITY_MAP_READ_BIT_EXT;
            break;
        default:
            // If a new layout is added, will need to manually add it
            return false;
    }

    if (access == 0 && !none_allowed) {
        skip |= LogWarning(device, kVUID_BestPractices_ImageBarrierAccessLayout,
                           "%s: accessMask is VK_ACCESS_NONE, but for layout %s expected accessMask are %s.", api_name.c_str(),
                           string_VkImageLayout(layout), string_VkAccessFlags2(allowed).c_str());
    } else if ((allowed | access) != allowed) {
        skip |=
            LogWarning(device, kVUID_BestPractices_ImageBarrierAccessLayout,
                       "%s: accessMask is %s, but for layout %s expected accessMask are %s.", string_VkAccessFlags2(access).c_str(),
                       api_name.c_str(), string_VkImageLayout(layout), string_VkAccessFlags2(allowed).c_str());
    }

    return skip;
}

bool BestPractices::ValidateImageMemoryBarrier(const std::string& api_name, VkImageLayout oldLayout, VkImageLayout newLayout,
                                               VkAccessFlags2 srcAccessMask, VkAccessFlags2 dstAccessMask,
                                               VkImageAspectFlags aspectMask) const {
    bool skip = false;

    if (oldLayout == VK_IMAGE_LAYOUT_UNDEFINED && IsImageLayoutReadOnly(newLayout)) {
        skip |= LogWarning(device, kVUID_BestPractices_TransitionUndefinedToReadOnly,
                           "VkImageMemoryBarrier is being submitted with oldLayout VK_IMAGE_LAYOUT_UNDEFINED and the contents "
                           "may be discarded, but the newLayout is %s, which is read only.",
                           string_VkImageLayout(newLayout));
    }

    skip |= ValidateAccessLayoutCombination(api_name, srcAccessMask, oldLayout, aspectMask);
    skip |= ValidateAccessLayoutCombination(api_name, dstAccessMask, newLayout, aspectMask);

    return skip;
}

bool BestPractices::PreCallValidateCmdPipelineBarrier(VkCommandBuffer commandBuffer, VkPipelineStageFlags srcStageMask,
                                                      VkPipelineStageFlags dstStageMask, VkDependencyFlags dependencyFlags,
                                                      uint32_t memoryBarrierCount, const VkMemoryBarrier* pMemoryBarriers,
                                                      uint32_t bufferMemoryBarrierCount,
                                                      const VkBufferMemoryBarrier* pBufferMemoryBarriers,
                                                      uint32_t imageMemoryBarrierCount,
                                                      const VkImageMemoryBarrier* pImageMemoryBarriers) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdPipelineBarrier", srcStageMask);
    skip |= CheckPipelineStageFlags("vkCmdPipelineBarrier", dstStageMask);

    for (uint32_t i = 0; i < imageMemoryBarrierCount; ++i) {
        skip |=
            ValidateImageMemoryBarrier("vkCmdPipelineBarrier", pImageMemoryBarriers[i].oldLayout, pImageMemoryBarriers[i].newLayout,
                                       pImageMemoryBarriers[i].srcAccessMask, pImageMemoryBarriers[i].dstAccessMask,
                                       pImageMemoryBarriers[i].subresourceRange.aspectMask);
    }

    if (VendorCheckEnabled(kBPVendorAMD)) {
        auto num = num_barriers_objects_.load();
        if (num + imageMemoryBarrierCount + bufferMemoryBarrierCount > kMaxRecommendedBarriersSizeAMD) {
            skip |= LogPerformanceWarning(device, kVUID_BestPractices_CmdBuffer_highBarrierCount,
                                          "%s Performance warning: In this frame, %" PRIu32
                                          " barriers were already submitted. Barriers have a high cost and can "
                                          "stall the GPU. "
                                          "Consider consolidating and re-organizing the frame to use fewer barriers.",
                                          VendorSpecificTag(kBPVendorAMD), num);
        }
    }
    if (VendorCheckEnabled(kBPVendorAMD) || VendorCheckEnabled(kBPVendorNVIDIA)) {
        static constexpr std::array<VkImageLayout, 3> read_layouts = {
            VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL,
            VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
            VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
        };

        for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
            // read to read barriers
            const auto &image_barrier = pImageMemoryBarriers[i];
            bool old_is_read_layout = std::find(read_layouts.begin(), read_layouts.end(), image_barrier.oldLayout) != read_layouts.end();
            bool new_is_read_layout = std::find(read_layouts.begin(), read_layouts.end(), image_barrier.newLayout) != read_layouts.end();

            if (old_is_read_layout && new_is_read_layout) {
                skip |= LogPerformanceWarning(device, kVUID_BestPractices_PipelineBarrier_readToReadBarrier,
                                              "%s %s Performance warning: Don't issue read-to-read barriers. "
                                              "Get the resource in the right state the first time you use it.",
                                              VendorSpecificTag(kBPVendorAMD), VendorSpecificTag(kBPVendorNVIDIA));
            }

            // general with no storage
            if (VendorCheckEnabled(kBPVendorAMD) && image_barrier.newLayout == VK_IMAGE_LAYOUT_GENERAL) {
                auto image_state = Get<IMAGE_STATE>(pImageMemoryBarriers[i].image);
                if (!(image_state->createInfo.usage & VK_IMAGE_USAGE_STORAGE_BIT)) {
                    skip |= LogPerformanceWarning(device, kVUID_BestPractices_vkImage_AvoidGeneral,
                                                  "%s Performance warning: VK_IMAGE_LAYOUT_GENERAL should only be used with "
                                                  "VK_IMAGE_USAGE_STORAGE_BIT images.",
                                                  VendorSpecificTag(kBPVendorAMD));
                }
            }
        }
    }

    for (uint32_t i = 0; i < imageMemoryBarrierCount; ++i) {
        skip |= ValidateCmdPipelineBarrierImageBarrier(commandBuffer, pImageMemoryBarriers[i]);
    }

    return skip;
}

bool BestPractices::PreCallValidateCmdPipelineBarrier2KHR(VkCommandBuffer commandBuffer,
                                                          const VkDependencyInfoKHR* pDependencyInfo) const {
    bool skip = false;

    skip |= CheckDependencyInfo("vkCmdPipelineBarrier2KHR", *pDependencyInfo);

    for (uint32_t i = 0; i < pDependencyInfo->imageMemoryBarrierCount; ++i) {
        skip |= ValidateCmdPipelineBarrierImageBarrier(commandBuffer, pDependencyInfo->pImageMemoryBarriers[i]);
    }

    return skip;
}

bool BestPractices::PreCallValidateCmdPipelineBarrier2(VkCommandBuffer commandBuffer,
                                                       const VkDependencyInfo* pDependencyInfo) const {
    bool skip = false;

    skip |= CheckDependencyInfo("vkCmdPipelineBarrier2", *pDependencyInfo);

    for (uint32_t i = 0; i < pDependencyInfo->imageMemoryBarrierCount; ++i) {
        skip |= ValidateCmdPipelineBarrierImageBarrier(commandBuffer, pDependencyInfo->pImageMemoryBarriers[i]);
    }

    return skip;
}

template <typename ImageMemoryBarrier>
bool BestPractices::ValidateCmdPipelineBarrierImageBarrier(VkCommandBuffer commandBuffer,
                                                           const ImageMemoryBarrier& barrier) const {

    bool skip = false;

    const auto cmd_state = GetRead<bp_state::CommandBuffer>(commandBuffer);
    assert(cmd_state);

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        if (barrier.oldLayout == VK_IMAGE_LAYOUT_UNDEFINED && barrier.newLayout != VK_IMAGE_LAYOUT_UNDEFINED) {
            skip |= ValidateZcull(*cmd_state, barrier.image, barrier.subresourceRange);
        }
    }

    return skip;
}

bool BestPractices::PreCallValidateCmdWriteTimestamp(VkCommandBuffer commandBuffer, VkPipelineStageFlagBits pipelineStage,
                                                     VkQueryPool queryPool, uint32_t query) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdWriteTimestamp", static_cast<VkPipelineStageFlags>(pipelineStage));

    return skip;
}

bool BestPractices::PreCallValidateCmdWriteTimestamp2KHR(VkCommandBuffer commandBuffer, VkPipelineStageFlags2KHR pipelineStage,
                                                         VkQueryPool queryPool, uint32_t query) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdWriteTimestamp2KHR", pipelineStage);

    return skip;
}

bool BestPractices::PreCallValidateCmdWriteTimestamp2(VkCommandBuffer commandBuffer, VkPipelineStageFlags2 pipelineStage,
                                                      VkQueryPool queryPool, uint32_t query) const {
    bool skip = false;

    skip |= CheckPipelineStageFlags("vkCmdWriteTimestamp2", pipelineStage);

    return skip;
}

void BestPractices::PreCallRecordCmdBindPipeline(VkCommandBuffer commandBuffer, VkPipelineBindPoint pipelineBindPoint,
                                                 VkPipeline pipeline) {
    StateTracker::PreCallRecordCmdBindPipeline(commandBuffer, pipelineBindPoint, pipeline);

    auto pipeline_info = Get<PIPELINE_STATE>(pipeline);
    auto cb = GetWrite<bp_state::CommandBuffer>(commandBuffer);

    assert(pipeline_info);
    assert(cb);

    if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS && VendorCheckEnabled(kBPVendorNVIDIA)) {
        using TessGeometryMeshState = bp_state::CommandBufferStateNV::TessGeometryMesh::State;
        auto& tgm = cb->nv.tess_geometry_mesh;

        // Make sure the message is only signaled once per command buffer
        tgm.threshold_signaled = tgm.num_switches >= kNumBindPipelineTessGeometryMeshSwitchesThresholdNVIDIA;

        // Track pipeline switches with tessellation, geometry, and/or mesh shaders enabled, and disabled
        auto tgm_stages = VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT | VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT |
                          VK_SHADER_STAGE_GEOMETRY_BIT | VK_SHADER_STAGE_TASK_BIT_NV | VK_SHADER_STAGE_MESH_BIT_NV;
        auto new_tgm_state = (pipeline_info->active_shaders & tgm_stages) != 0
                                                  ? TessGeometryMeshState::Enabled
                                                  : TessGeometryMeshState::Disabled;
        if (tgm.state != new_tgm_state && tgm.state != TessGeometryMeshState::Unknown) {
            tgm.num_switches++;
        }
        tgm.state = new_tgm_state;

        // Track depthTestEnable and depthCompareOp
        auto &pipeline_create_info = pipeline_info->GetCreateInfo<VkGraphicsPipelineCreateInfo>();
        auto depth_stencil_state = pipeline_create_info.pDepthStencilState;
        auto dynamic_state = pipeline_create_info.pDynamicState;
        if (depth_stencil_state && dynamic_state) {
            auto dynamic_state_begin = dynamic_state->pDynamicStates;
            auto dynamic_state_end = dynamic_state->pDynamicStates + dynamic_state->dynamicStateCount;

            bool dynamic_depth_test_enable = std::find(dynamic_state_begin, dynamic_state_end, VK_DYNAMIC_STATE_DEPTH_TEST_ENABLE) != dynamic_state_end;
            bool dynamic_depth_func = std::find(dynamic_state_begin, dynamic_state_end, VK_DYNAMIC_STATE_DEPTH_COMPARE_OP) != dynamic_state_end;

            if (!dynamic_depth_test_enable) {
                RecordSetDepthTestState(*cb, cb->nv.depth_compare_op, depth_stencil_state->depthTestEnable != VK_FALSE);
            }
            if (!dynamic_depth_func) {
                RecordSetDepthTestState(*cb, depth_stencil_state->depthCompareOp, cb->nv.depth_test_enable);
            }
        }
    }
}

void BestPractices::PostCallRecordCmdBindPipeline(VkCommandBuffer commandBuffer, VkPipelineBindPoint pipelineBindPoint,
                                                  VkPipeline pipeline) {
    StateTracker::PostCallRecordCmdBindPipeline(commandBuffer, pipelineBindPoint, pipeline);

    // AMD best practice
    PipelineUsedInFrame(pipeline);

    if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS) {
        auto pipeline_state = Get<bp_state::Pipeline>(pipeline);
        // check for depth/blend state tracking
        if (pipeline_state) {
            auto cb_node = GetWrite<bp_state::CommandBuffer>(commandBuffer);
            assert(cb_node);
            auto& render_pass_state = cb_node->render_pass_state;

            render_pass_state.nextDrawTouchesAttachments = pipeline_state->access_framebuffer_attachments;
            render_pass_state.drawTouchAttachments = true;

            const auto* blend_state = pipeline_state->ColorBlendState();
            const auto* stencil_state = pipeline_state->DepthStencilState();

            if (blend_state) {
                // assume the pipeline is depth-only unless any of the attachments have color writes enabled
                render_pass_state.depthOnly = true;
                for (size_t i = 0; i < blend_state->attachmentCount; i++) {
                    if (blend_state->pAttachments[i].colorWriteMask != 0) {
                        render_pass_state.depthOnly = false;
                    }
                }
            }

            // check for depth value usage
            render_pass_state.depthEqualComparison = false;

            if (stencil_state && stencil_state->depthTestEnable) {
                switch (stencil_state->depthCompareOp) {
                    case VK_COMPARE_OP_EQUAL:
                    case VK_COMPARE_OP_GREATER_OR_EQUAL:
                    case VK_COMPARE_OP_LESS_OR_EQUAL:
                        render_pass_state.depthEqualComparison = true;
                        break;
                    default:
                        break;
                }
            }
        }
    }
}

void BestPractices::PreCallRecordCmdSetDepthCompareOp(VkCommandBuffer commandBuffer, VkCompareOp depthCompareOp) {
    StateTracker::PreCallRecordCmdSetDepthCompareOp(commandBuffer, depthCompareOp);

    auto cb = GetWrite<bp_state::CommandBuffer>(commandBuffer);
    assert(cb);

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        RecordSetDepthTestState(*cb, depthCompareOp, cb->nv.depth_test_enable);
    }
}

void BestPractices::PreCallRecordCmdSetDepthCompareOpEXT(VkCommandBuffer commandBuffer, VkCompareOp depthCompareOp) {
    StateTracker::PreCallRecordCmdSetDepthCompareOpEXT(commandBuffer, depthCompareOp);

    auto cb = GetWrite<bp_state::CommandBuffer>(commandBuffer);
    assert(cb);

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        RecordSetDepthTestState(*cb, depthCompareOp, cb->nv.depth_test_enable);
    }
}

void BestPractices::PreCallRecordCmdSetDepthTestEnable(VkCommandBuffer commandBuffer, VkBool32 depthTestEnable) {
    StateTracker::PreCallRecordCmdSetDepthTestEnable(commandBuffer, depthTestEnable);

    auto cb = GetWrite<bp_state::CommandBuffer>(commandBuffer);
    assert(cb);

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        RecordSetDepthTestState(*cb, cb->nv.depth_compare_op, depthTestEnable != VK_FALSE);
    }
}

void BestPractices::PreCallRecordCmdSetDepthTestEnableEXT(VkCommandBuffer commandBuffer, VkBool32 depthTestEnable) {
    StateTracker::PreCallRecordCmdSetDepthTestEnableEXT(commandBuffer, depthTestEnable);

    auto cb = GetWrite<bp_state::CommandBuffer>(commandBuffer);
    assert(cb);

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        RecordSetDepthTestState(*cb, cb->nv.depth_compare_op, depthTestEnable != VK_FALSE);
    }
}

void BestPractices::RecordSetDepthTestState(bp_state::CommandBuffer& cmd_state, VkCompareOp new_depth_compare_op, bool new_depth_test_enable) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    if (cmd_state.nv.depth_compare_op != new_depth_compare_op) {
        switch (new_depth_compare_op) {
        case VK_COMPARE_OP_LESS:
        case VK_COMPARE_OP_LESS_OR_EQUAL:
            cmd_state.nv.zcull_direction = bp_state::CommandBufferStateNV::ZcullDirection::Less;
            break;
        case VK_COMPARE_OP_GREATER:
        case VK_COMPARE_OP_GREATER_OR_EQUAL:
            cmd_state.nv.zcull_direction = bp_state::CommandBufferStateNV::ZcullDirection::Greater;
            break;
        default:
            // The other ops carry over the previous state.
            break;
        }
    }
    cmd_state.nv.depth_compare_op = new_depth_compare_op;
    cmd_state.nv.depth_test_enable = new_depth_test_enable;
}

void BestPractices::RecordCmdBeginRenderingCommon(VkCommandBuffer commandBuffer) {
    auto cmd_state = GetWrite<bp_state::CommandBuffer>(commandBuffer);
    assert(cmd_state);

    auto rp = cmd_state->activeRenderPass.get();
    assert(rp);

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        std::shared_ptr<IMAGE_VIEW_STATE> depth_image_view_shared_ptr;
        IMAGE_VIEW_STATE* depth_image_view = nullptr;
        layer_data::optional<VkAttachmentLoadOp> load_op;

        if (rp->use_dynamic_rendering || rp->use_dynamic_rendering_inherited) {
            const auto depth_attachment = rp->dynamic_rendering_begin_rendering_info.pDepthAttachment;
            if (depth_attachment) {
                load_op.emplace(depth_attachment->loadOp);
                depth_image_view_shared_ptr = Get<IMAGE_VIEW_STATE>(depth_attachment->imageView);
                depth_image_view = depth_image_view_shared_ptr.get();
            }

            for (uint32_t i = 0; i < rp->dynamic_rendering_begin_rendering_info.colorAttachmentCount; ++i) {
                const auto& color_attachment = rp->dynamic_rendering_begin_rendering_info.pColorAttachments[i];
                if (color_attachment.loadOp == VK_ATTACHMENT_LOAD_OP_CLEAR) {
                    const VkFormat format = Get<IMAGE_VIEW_STATE>(color_attachment.imageView)->create_info.format;
                    RecordClearColor(format, color_attachment.clearValue.color);
                }
            }

        } else {
            if (rp->createInfo.pAttachments) {
                if (rp->createInfo.subpassCount > 0) {
                    const auto depth_attachment = rp->createInfo.pSubpasses[0].pDepthStencilAttachment;
                    if (depth_attachment) {
                        const uint32_t attachment_index = depth_attachment->attachment;
                        if (attachment_index != VK_ATTACHMENT_UNUSED) {
                            load_op.emplace(rp->createInfo.pAttachments[attachment_index].loadOp);
                            depth_image_view = (*cmd_state->active_attachments)[attachment_index];
                        }
                    }
                }
                for (uint32_t i = 0; i < cmd_state->activeRenderPassBeginInfo.clearValueCount; ++i) {
                    const auto& attachment = rp->createInfo.pAttachments[i];
                    if (attachment.loadOp == VK_ATTACHMENT_LOAD_OP_CLEAR) {
                        const auto& clear_color = cmd_state->activeRenderPassBeginInfo.pClearValues[i].color;
                        RecordClearColor(attachment.format, clear_color);
                    }
                }
            }
        }
        if (depth_image_view && (depth_image_view->create_info.subresourceRange.aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) != 0U) {
            const VkImage depth_image = depth_image_view->image_state->image();
            const VkImageSubresourceRange& subresource_range = depth_image_view->create_info.subresourceRange;
            RecordBindZcullScope(*cmd_state, depth_image, subresource_range);
        } else {
            RecordUnbindZcullScope(*cmd_state);
        }
        if (load_op) {
            if (*load_op == VK_ATTACHMENT_LOAD_OP_CLEAR || *load_op == VK_ATTACHMENT_LOAD_OP_DONT_CARE) {
                RecordResetScopeZcullDirection(*cmd_state);
            }
        }
    }
}

void BestPractices::RecordCmdEndRenderingCommon(VkCommandBuffer commandBuffer) {
    auto cmd_state = GetWrite<bp_state::CommandBuffer>(commandBuffer);
    assert(cmd_state);

    auto rp = cmd_state->activeRenderPass.get();
    assert(rp);

    if (VendorCheckEnabled(kBPVendorNVIDIA)) {
        layer_data::optional<VkAttachmentStoreOp> store_op;

        if (rp->use_dynamic_rendering || rp->use_dynamic_rendering_inherited) {
            const auto depth_attachment = rp->dynamic_rendering_begin_rendering_info.pDepthAttachment;
            if (depth_attachment) {
                store_op.emplace(depth_attachment->storeOp);
            }
        } else {
            if (rp->createInfo.subpassCount > 0) {
                const uint32_t last_subpass = rp->createInfo.subpassCount - 1;
                const auto depth_attachment = rp->createInfo.pSubpasses[last_subpass].pDepthStencilAttachment;
                if (depth_attachment) {
                    const uint32_t attachment = depth_attachment->attachment;
                    if (attachment != VK_ATTACHMENT_UNUSED) {
                        store_op.emplace(rp->createInfo.pAttachments[attachment].storeOp);
                    }
                }
            }
        }

        if (store_op) {
            if (*store_op == VK_ATTACHMENT_STORE_OP_DONT_CARE || *store_op == VK_ATTACHMENT_STORE_OP_NONE) {
                RecordResetScopeZcullDirection(*cmd_state);
            }
        }

        RecordUnbindZcullScope(*cmd_state);
    }
}

void BestPractices::RecordBindZcullScope(bp_state::CommandBuffer& cmd_state, VkImage depth_attachment, const VkImageSubresourceRange& subresource_range) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    if (depth_attachment == VK_NULL_HANDLE) {
        cmd_state.nv.zcull_scope = {};
        return;
    }

    assert((subresource_range.aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) != 0U);

    auto image_state = Get<IMAGE_STATE>(depth_attachment);
    assert(image_state);

    const uint32_t mip_levels = image_state->createInfo.mipLevels;
    const uint32_t array_layers = image_state->createInfo.arrayLayers;

    auto& tree = cmd_state.nv.zcull_per_image[depth_attachment];
    if (tree.states.empty()) {
        tree.mip_levels = mip_levels;
        tree.array_layers = array_layers;
        tree.states.resize(array_layers * mip_levels);
    }

    cmd_state.nv.zcull_scope.image = depth_attachment;
    cmd_state.nv.zcull_scope.range = subresource_range;
    cmd_state.nv.zcull_scope.tree = &tree;
}

void BestPractices::RecordUnbindZcullScope(bp_state::CommandBuffer& cmd_state) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    RecordBindZcullScope(cmd_state, VK_NULL_HANDLE, VkImageSubresourceRange{});
}

void BestPractices::RecordResetScopeZcullDirection(bp_state::CommandBuffer& cmd_state) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    auto& scope = cmd_state.nv.zcull_scope;
    RecordResetZcullDirection(cmd_state, scope.image, scope.range);
}

template <typename Func>
static void ForEachSubresource(const IMAGE_STATE& image, const VkImageSubresourceRange& range, Func&& func)
{
    const uint32_t layerCount =
        (range.layerCount == VK_REMAINING_ARRAY_LAYERS) ? (image.full_range.layerCount - range.baseArrayLayer) : range.layerCount;
    const uint32_t levelCount =
        (range.levelCount == VK_REMAINING_MIP_LEVELS) ? (image.full_range.levelCount - range.baseMipLevel) : range.levelCount;

    for (uint32_t i = 0; i < layerCount; ++i) {
        const uint32_t layer = range.baseArrayLayer + i;
        for (uint32_t j = 0; j < levelCount; ++j) {
            const uint32_t level = range.baseMipLevel + j;
            func(layer, level);
        }
    }
}

void BestPractices::RecordResetZcullDirection(bp_state::CommandBuffer& cmd_state, VkImage depth_image,
                                              const VkImageSubresourceRange& subresource_range) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    RecordSetZcullDirection(cmd_state, depth_image, subresource_range, bp_state::CommandBufferStateNV::ZcullDirection::Unknown);

    const auto image_it = cmd_state.nv.zcull_per_image.find(depth_image);
    if (image_it == cmd_state.nv.zcull_per_image.end()) {
        return;
    }
    auto& tree = image_it->second;

    auto image = Get<IMAGE_STATE>(depth_image);
    if (!image) return;

    ForEachSubresource(*image, subresource_range, [&tree](uint32_t layer, uint32_t level) {
        auto& subresource = tree.GetState(layer, level);
        subresource.num_less_draws = 0;
        subresource.num_greater_draws = 0;
    });
}

void BestPractices::RecordSetScopeZcullDirection(bp_state::CommandBuffer& cmd_state, bp_state::CommandBufferStateNV::ZcullDirection mode) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    auto& scope = cmd_state.nv.zcull_scope;
    RecordSetZcullDirection(cmd_state, scope.image, scope.range, mode);
}

void BestPractices::RecordSetZcullDirection(bp_state::CommandBuffer& cmd_state, VkImage depth_image,
                                            const VkImageSubresourceRange& subresource_range,
                                            bp_state::CommandBufferStateNV::ZcullDirection mode) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    const auto image_it = cmd_state.nv.zcull_per_image.find(depth_image);
    if (image_it == cmd_state.nv.zcull_per_image.end()) {
        return;
    }
    auto& tree = image_it->second;

    auto image = Get<IMAGE_STATE>(depth_image);
    if (!image) return;

    ForEachSubresource(*image, subresource_range, [&tree, &cmd_state](uint32_t layer, uint32_t level) {
        tree.GetState(layer, level).direction = cmd_state.nv.zcull_direction;
    });
}

void BestPractices::RecordZcullDraw(bp_state::CommandBuffer& cmd_state) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    // Add one draw to each subresource depending on the current Z-cull direction
    auto& scope = cmd_state.nv.zcull_scope;

    auto image = Get<IMAGE_STATE>(scope.image);
    if (!image) return;

    ForEachSubresource(*image, scope.range, [&scope](uint32_t layer, uint32_t level) {
        auto& subresource = scope.tree->GetState(layer, level);

        switch (subresource.direction) {
        case bp_state::CommandBufferStateNV::ZcullDirection::Unknown:
            // Unreachable
            assert(0);
            break;
        case bp_state::CommandBufferStateNV::ZcullDirection::Less:
            ++subresource.num_less_draws;
            break;
        case bp_state::CommandBufferStateNV::ZcullDirection::Greater:
            ++subresource.num_greater_draws;
            break;
        }
    });
}

bool BestPractices::ValidateZcullScope(const bp_state::CommandBuffer& cmd_state) const {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    bool skip = false;

    if (cmd_state.nv.depth_test_enable) {
        auto& scope = cmd_state.nv.zcull_scope;
        skip |= ValidateZcull(cmd_state, scope.image, scope.range);
    }

    return skip;
}

bool BestPractices::ValidateZcull(const bp_state::CommandBuffer& cmd_state, VkImage image,
                                  const VkImageSubresourceRange& subresource_range) const {
    bool skip = false;

    const char* good_mode = nullptr;
    const char* bad_mode = nullptr;
    bool is_balanced = false;

    const auto image_it = cmd_state.nv.zcull_per_image.find(image);
    if (image_it == cmd_state.nv.zcull_per_image.end()) {
        return skip;
    }
    const auto& tree = image_it->second;

    auto image_state = Get<IMAGE_STATE>(image);
    if (!image_state) {
        return skip;
    }

    ForEachSubresource(*image_state, subresource_range, [&](uint32_t layer, uint32_t level) {
        if (is_balanced) {
            return;
        }
        const auto& resource = tree.GetState(layer, level);
        const uint64_t num_draws = resource.num_less_draws + resource.num_greater_draws;

        if (num_draws == 0) {
            return;
        }
        const uint64_t less_ratio = (resource.num_less_draws * 100) / num_draws;
        const uint64_t greater_ratio = (resource.num_greater_draws * 100) / num_draws;

        if ((less_ratio > kZcullDirectionBalanceRatioNVIDIA) && (greater_ratio > kZcullDirectionBalanceRatioNVIDIA)) {
            is_balanced = true;

            if (greater_ratio > less_ratio) {
                good_mode = "GREATER";
                bad_mode = "LESS";
            } else {
                good_mode = "LESS";
                bad_mode = "GREATER";
            }
        }
    });

    if (is_balanced) {
        skip |= LogPerformanceWarning(
            cmd_state.commandBuffer(), kVUID_BestPractices_Zcull_LessGreaterRatio,
            "%s Depth attachment %s is primarily rendered with depth compare op %s, but some draws use %s. "
            "Z-cull is disabled for the least used direction, which harms depth testing performance. "
            "The Z-cull direction can be reset by clearing the depth attachment, transitioning from VK_IMAGE_LAYOUT_UNDEFINED, "
            "using VK_ATTACHMENT_LOAD_OP_DONT_CARE, or using VK_ATTACHMENT_STORE_OP_DONT_CARE.",
            VendorSpecificTag(kBPVendorNVIDIA), report_data->FormatHandle(cmd_state.nv.zcull_scope.image).c_str(), good_mode,
            bad_mode);
    }

    return skip;
}

static std::array<uint32_t, 4> GetRawClearColor(VkFormat format, const VkClearColorValue& clear_value) {
    std::array<uint32_t, 4> raw_color{};
    std::copy_n(clear_value.uint32, raw_color.size(), raw_color.data());

    // Zero out unused components to avoid polluting the cache with garbage
    if (!FormatHasRed(format))   raw_color[0] = 0;
    if (!FormatHasGreen(format)) raw_color[1] = 0;
    if (!FormatHasBlue(format))  raw_color[2] = 0;
    if (!FormatHasAlpha(format)) raw_color[3] = 0;

    return raw_color;
}

static bool IsClearColorZeroOrOne(VkFormat format, const std::array<uint32_t, 4> clear_color) {
    static_assert(sizeof(float) == sizeof(uint32_t), "Mismatching float <-> uint32 sizes");
    const float one = 1.0f;
    const float zero = 0.0f;
    uint32_t raw_one{};
    uint32_t raw_zero{};
    memcpy(&raw_one, &one, sizeof(one));
    memcpy(&raw_zero, &zero, sizeof(zero));

    const bool is_one = (!FormatHasRed(format)   || (clear_color[0] == raw_one)) &&
                        (!FormatHasGreen(format) || (clear_color[1] == raw_one)) &&
                        (!FormatHasBlue(format)  || (clear_color[2] == raw_one)) &&
                        (!FormatHasAlpha(format) || (clear_color[3] == raw_one));
    const bool is_zero = (!FormatHasRed(format)   || (clear_color[0] == raw_zero)) &&
                         (!FormatHasGreen(format) || (clear_color[1] == raw_zero)) &&
                         (!FormatHasBlue(format)  || (clear_color[2] == raw_zero)) &&
                         (!FormatHasAlpha(format) || (clear_color[3] == raw_zero));
    return is_one || is_zero;
}

static std::string MakeCompressedFormatListNVIDIA() {
    std::string format_list;
    for (VkFormat compressed_format : kCustomClearColorCompressedFormatsNVIDIA) {
        if (compressed_format == kCustomClearColorCompressedFormatsNVIDIA.back()) {
            format_list += "or ";
        }
        format_list += string_VkFormat(compressed_format);
        if (compressed_format != kCustomClearColorCompressedFormatsNVIDIA.back()) {
            format_list += ", ";
        }
    }
    return format_list;
}

void BestPractices::RecordClearColor(VkFormat format, const VkClearColorValue& clear_value) {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    const std::array<uint32_t, 4> raw_color = GetRawClearColor(format, clear_value);
    if (IsClearColorZeroOrOne(format, raw_color)) {
        // These colors are always compressed
        return;
    }

    const auto it = std::find(kCustomClearColorCompressedFormatsNVIDIA.begin(), kCustomClearColorCompressedFormatsNVIDIA.end(), format);
    if (it == kCustomClearColorCompressedFormatsNVIDIA.end()) {
        // The format cannot be compressed with a custom color
        return;
    }

    // Record custom clear color
    WriteLockGuard guard{clear_colors_lock_};
    if (clear_colors_.size() < kMaxRecommendedNumberOfClearColorsNVIDIA) {
        clear_colors_.insert(raw_color);
    }
}

bool BestPractices::ValidateClearColor(VkCommandBuffer commandBuffer, VkFormat format, const VkClearColorValue& clear_value) const {
    assert(VendorCheckEnabled(kBPVendorNVIDIA));

    bool skip = false;

    const std::array<uint32_t, 4> raw_color = GetRawClearColor(format, clear_value);
    if (IsClearColorZeroOrOne(format, raw_color)) {
        return skip;
    }

    const auto it = std::find(kCustomClearColorCompressedFormatsNVIDIA.begin(), kCustomClearColorCompressedFormatsNVIDIA.end(), format);
    if (it == kCustomClearColorCompressedFormatsNVIDIA.end()) {
        // The format is not compressible
        static const std::string format_list = MakeCompressedFormatListNVIDIA();

        skip |= LogPerformanceWarning(commandBuffer, kVUID_BestPractices_ClearColor_NotCompressed,
                                      "%s Clearing image with format %s without a 1.0f or 0.0f clear color. "
                                      "The clear will not get compressed in the GPU, harming performance. "
                                      "This can be fixed using a clear color of VkClearColorValue{0.0f, 0.0f, 0.0f, 0.0f}, or "
                                      "VkClearColorValue{1.0f, 1.0f, 1.0f, 1.0f}. Alternatively, use %s.",
                                      VendorSpecificTag(kBPVendorNVIDIA), string_VkFormat(format), format_list.c_str());
    } else {
        // The format is compressible
        bool registered = false;
        {
            ReadLockGuard guard{clear_colors_lock_};
            registered = clear_colors_.find(raw_color) != clear_colors_.end();

            if (!registered) {
                // If it's not in the list, it might be new. Check if there's still space for new entries.
                registered = clear_colors_.size() < kMaxRecommendedNumberOfClearColorsNVIDIA;
            }
        }
        if (!registered) {
            std::string clear_color_str;

            if (FormatIsUINT(format)) {
                clear_color_str = std::to_string(clear_value.uint32[0]) + ", " + std::to_string(clear_value.uint32[1]) + ", " +
                                  std::to_string(clear_value.uint32[2]) + ", " + std::to_string(clear_value.uint32[3]);
            } else if (FormatIsSINT(format)) {
                clear_color_str = std::to_string(clear_value.int32[0]) + ", " + std::to_string(clear_value.int32[1]) + ", " +
                                  std::to_string(clear_value.int32[2]) + ", " + std::to_string(clear_value.int32[3]);
            } else {
                clear_color_str = std::to_string(clear_value.float32[0]) + ", " + std::to_string(clear_value.float32[1]) + ", " +
                                  std::to_string(clear_value.float32[2]) + ", " + std::to_string(clear_value.float32[3]);
            }

            skip |= LogPerformanceWarning(
                commandBuffer, kVUID_BestPractices_ClearColor_NotCompressed,
                "%s Clearing image with unregistered VkClearColorValue{%s}. "
                "This clear will not get compressed in the GPU, harming performance. "
                "The clear color is not registered because too many unique colors have been used. "
                "Select a discrete set of clear colors and stick to those. "
                "VkClearColorValue{0, 0, 0, 0} and VkClearColorValue{1.0f, 1.0f, 1.0f, 1.0f} are always registered.",
                VendorSpecificTag(kBPVendorNVIDIA), clear_color_str.c_str());
        }
    }

    return skip;
}

static inline bool RenderPassUsesAttachmentAsResolve(const safe_VkRenderPassCreateInfo2& createInfo, uint32_t attachment) {
    for (uint32_t subpass = 0; subpass < createInfo.subpassCount; subpass++) {
        const auto& subpass_info = createInfo.pSubpasses[subpass];
        if (subpass_info.pResolveAttachments) {
            for (uint32_t i = 0; i < subpass_info.colorAttachmentCount; i++) {
                if (subpass_info.pResolveAttachments[i].attachment == attachment) return true;
            }
        }
    }

    return false;
}

static inline bool RenderPassUsesAttachmentOnTile(const safe_VkRenderPassCreateInfo2& createInfo, uint32_t attachment) {
    for (uint32_t subpass = 0; subpass < createInfo.subpassCount; subpass++) {
        const auto& subpass_info = createInfo.pSubpasses[subpass];

        // If an attachment is ever used as a color attachment,
        // resolve attachment or depth stencil attachment,
        // it needs to exist on tile at some point.

        for (uint32_t i = 0; i < subpass_info.colorAttachmentCount; i++) {
            if (subpass_info.pColorAttachments[i].attachment == attachment) return true;
        }

        if (subpass_info.pResolveAttachments) {
            for (uint32_t i = 0; i < subpass_info.colorAttachmentCount; i++) {
                if (subpass_info.pResolveAttachments[i].attachment == attachment) return true;
            }
        }

        if (subpass_info.pDepthStencilAttachment && subpass_info.pDepthStencilAttachment->attachment == attachment) return true;
    }

    return false;
}

static inline bool RenderPassUsesAttachmentAsImageOnly(const safe_VkRenderPassCreateInfo2& createInfo, uint32_t attachment) {
    if (RenderPassUsesAttachmentOnTile(createInfo, attachment)) {
        return false;
    }

    for (uint32_t subpass = 0; subpass < createInfo.subpassCount; subpass++) {
        const auto& subpassInfo = createInfo.pSubpasses[subpass];

        for (uint32_t i = 0; i < subpassInfo.inputAttachmentCount; i++) {
            if (subpassInfo.pInputAttachments[i].attachment == attachment) {
                return true;
            }
        }
    }

    return false;
}

bool BestPractices::ValidateCmdBeginRenderPass(VkCommandBuffer commandBuffer, RenderPassCreateVersion rp_version,
                                               const VkRenderPassBeginInfo* pRenderPassBegin) const {
    bool skip = false;

    if (!pRenderPassBegin) {
        return skip;
    }

    if (pRenderPassBegin->renderArea.extent.width == 0 || pRenderPassBegin->renderArea.extent.height == 0) {
        skip |= LogWarning(device, kVUID_BestPractices_BeginRenderPass_ZeroSizeRenderArea,
                           "This render pass has a zero-size render area. It cannot write to any attachments, "
                           "and can only be used for side effects such as layout transitions.");
    }

    auto rp_state = Get<RENDER_PASS_STATE>(pRenderPassBegin->renderPass);
    if (rp_state) {
        if (rp_state->createInfo.flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT) {
            const VkRenderPassAttachmentBeginInfo* rpabi = LvlFindInChain<VkRenderPassAttachmentBeginInfo>(pRenderPassBegin->pNext);
            if (rpabi) {
                skip = ValidateAttachments(rp_state->createInfo.ptr(), rpabi->attachmentCount, rpabi->pAttachments);
            }
        }
        // Check if any attachments have LOAD operation on them
        for (uint32_t att = 0; att < rp_state->createInfo.attachmentCount; att++) {
            const auto& attachment = rp_state->createInfo.pAttachments[att];

            bool attachment_has_readback = false;
            if (!FormatIsStencilOnly(attachment.format) && attachment.loadOp == VK_ATTACHMENT_LOAD_OP_LOAD) {
                attachment_has_readback = true;
            }

            if (FormatHasStencil(attachment.format) && attachment.stencilLoadOp == VK_ATTACHMENT_LOAD_OP_LOAD) {
                attachment_has_readback = true;
            }

            bool attachment_needs_readback = false;

            // Check if the attachment is actually used in any subpass on-tile
            if (attachment_has_readback && RenderPassUsesAttachmentOnTile(rp_state->createInfo, att)) {
                attachment_needs_readback = true;
            }

            // Using LOAD_OP_LOAD is expensive on tiled GPUs, so flag it as a potential improvement
            if (attachment_needs_readback && (VendorCheckEnabled(kBPVendorArm) || VendorCheckEnabled(kBPVendorIMG))) {
                skip |=
                    LogPerformanceWarning(device, kVUID_BestPractices_BeginRenderPass_AttachmentNeedsReadback,
                                          "%s %s: Attachment #%u in render pass has begun with VK_ATTACHMENT_LOAD_OP_LOAD.\n"