feat: 环境背景管线（IBL 阶段①）——HDR 等距柱状转 cube + Skybox pass
- Cubemap：stbi_loadf 加载 .hdr，CPU 双线性重采样 6 面（GL 规范方向表）， f32→fp16 转换后整块上传 eCubeCompatible image（R16G16B16A16Sfloat） - SkyboxPass：HDRColor 的 first writer，承担清屏职责；无环境贴图自动 退化为纯清屏（Forward 不受影响） - frag 逐像素 NDC 反投影取世界方向（invVP push，投影变换不可线性插值） - Forward color 附件 Clear→Load：几何合成在天空之上 - SceneRenderer 产出 skyInvVP = inverse(proj × mat4(mat3(view)))
This commit is contained in:
@@ -0,0 +1,293 @@
|
||||
//
|
||||
// Created by atdunbg on 2026/9/23.
|
||||
//
|
||||
// Cubemap 实现：HDR 等距柱状 → CPU 重采样 6 面 → fp16 → cube image。
|
||||
//
|
||||
// 一次性加载成本（亚秒级 CPU 重采样），换取运行时零转换开销；
|
||||
// 硬件 Cube 采样（filtering/边缘接缝由 driver 处理）。
|
||||
// 后续 IBL 阶段的 irradiance/prefilter 卷积复用 FaceDir 方向表。
|
||||
//
|
||||
|
||||
#include <Engine/Renderer/Cubemap.h>
|
||||
#include <Engine/Vulkan/Device.h>
|
||||
#include <Engine/Vulkan/MemoryAllocator.h>
|
||||
#include <Engine/Core/Log.h>
|
||||
|
||||
#include <stb/stb_image.h>
|
||||
|
||||
#include <glm/glm.hpp>
|
||||
#include <glm/gtc/constants.hpp>
|
||||
|
||||
#include <algorithm>
|
||||
#include <bit>
|
||||
#include <cmath>
|
||||
#include <cstring>
|
||||
#include <vector>
|
||||
|
||||
namespace Engine
|
||||
{
|
||||
namespace
|
||||
{
|
||||
constexpr u32 kFaceCount = 6;
|
||||
|
||||
// GL/Vulkan 规范的 cube 面方向表（与 GL spec 立方体选择表互逆）。
|
||||
// (u,v)∈[0,1]（u 向右 v 向下）→ s,t∈[-1,1] → 单位化前方向。
|
||||
// Vulkan 面序与 GL 一致：+X -X +Y -Y +Z -Z
|
||||
glm::vec3 FaceDir(u32 face, f32 u, f32 v)
|
||||
{
|
||||
const f32 s = u * 2.0f - 1.0f;
|
||||
const f32 t = v * 2.0f - 1.0f;
|
||||
switch (face) {
|
||||
case 0: return { 1.0f, -t, -s }; // +X（面片顶行 = +Y）
|
||||
case 1: return {-1.0f, -t, s }; // -X
|
||||
case 2: return { s, 1.0f, t }; // +Y
|
||||
case 3: return { s,-1.0f, -t }; // -Y
|
||||
case 4: return { s, -t, 1.0f }; // +Z
|
||||
default: return {-s, -t,-1.0f }; // -Z
|
||||
}
|
||||
}
|
||||
|
||||
// 等距柱状双线性采样（x 环绕跨经度接缝，y 钳制到极点）。
|
||||
// 图像行 0 = 天顶（+Y）：v = acos(y)/π，与 stbi 行序一致，无需翻转
|
||||
glm::vec3 SampleEquirect(const f32* img, u32 w, u32 h, const glm::vec3& dir)
|
||||
{
|
||||
const f32 u = std::atan2(dir.z, dir.x) * (0.5f / glm::pi<f32>()) + 0.5f;
|
||||
const f32 v = std::acos(glm::clamp(dir.y, -1.0f, 1.0f)) / glm::pi<f32>();
|
||||
|
||||
const f32 fx = u * static_cast<f32>(w) - 0.5f;
|
||||
const f32 fy = v * static_cast<f32>(h) - 0.5f;
|
||||
const u32 x0 = static_cast<u32>(std::floor(fx));
|
||||
const u32 y0 = static_cast<u32>(std::floor(fy));
|
||||
const f32 tx = fx - std::floor(fx);
|
||||
const f32 ty = fy - std::floor(fy);
|
||||
|
||||
auto fetch = [&](u32 x, u32 y) {
|
||||
const u32 xc = ((x % w) + w) % w; // 经度环绕
|
||||
const u32 yc = glm::clamp(y, 0u, h - 1u); // 纬度钳制
|
||||
const size_t off = (static_cast<size_t>(yc) * w + xc) * 4;
|
||||
return glm::vec3(img[off + 0], img[off + 1], img[off + 2]);
|
||||
};
|
||||
|
||||
const glm::vec3 c00 = fetch(x0, y0);
|
||||
const glm::vec3 c10 = fetch(x0 + 1, y0);
|
||||
const glm::vec3 c01 = fetch(x0, y0 + 1);
|
||||
const glm::vec3 c11 = fetch(x0 + 1, y0 + 1);
|
||||
|
||||
return glm::mix(glm::mix(c00, c10, tx), glm::mix(c01, c11, tx), ty);
|
||||
}
|
||||
|
||||
// IEEE754 f32 → fp16（round-to-nearest-even；溢出饱和为 Inf）。
|
||||
// HDR 环境值域内常规路径足够，次正规/NaN 路径只为数学完备
|
||||
u16 FloatToHalf(f32 value)
|
||||
{
|
||||
const u32 x = std::bit_cast<u32>(value);
|
||||
const u16 sign = static_cast<u16>((x >> 16) & 0x8000u);
|
||||
const u32 absVal = x & 0x7fffffffu;
|
||||
|
||||
if (absVal > 0x7f800000u) {
|
||||
return static_cast<u16>(sign | 0x7e00u); // NaN
|
||||
}
|
||||
if (absVal >= 0x47800000u) {
|
||||
return static_cast<u16>(sign | 0x7c00u); // Inf / 溢出饱和
|
||||
}
|
||||
if (absVal >= 0x38800000u) {
|
||||
// 常规数：指数重偏（127→15），尾数右移 13 位并舍入到最近偶数
|
||||
const u32 bits = absVal - 0x38000000u;
|
||||
const u32 round = 0x0fffu + ((bits >> 13) & 1u);
|
||||
return static_cast<u16>(sign | ((bits + round) >> 13));
|
||||
}
|
||||
if (absVal < 0x33000000u) {
|
||||
return sign; // < 2^-25：舍入为 0
|
||||
}
|
||||
// half 次正规 [2^-24, 2^-14)：补隐含位后按目标精度舍入
|
||||
const u32 mantissa = (absVal & 0x007fffffu) | 0x00800000u;
|
||||
const u32 shift = 126u - (absVal >> 23u);
|
||||
const u32 m = (mantissa + (1u << (shift - 1))) >> shift;
|
||||
// 舍入进位到 0x400 恰好是最小正规 half 的位型，无需特判
|
||||
return static_cast<u16>(sign | std::min(m, 0x400u));
|
||||
}
|
||||
}
|
||||
|
||||
Scope<Cubemap> Cubemap::CreateFromHdrEquirect(Device& device,
|
||||
MemoryAllocator& allocator,
|
||||
const std::string& path,
|
||||
u32 faceSize)
|
||||
{
|
||||
// ---- 1. CPU 加载 HDR（float RGBA）----
|
||||
int w = 0, h = 0, channels = 0;
|
||||
f32* pixels = stbi_loadf(path.c_str(), &w, &h, &channels, STBI_rgb_alpha);
|
||||
if (!pixels) {
|
||||
ENGINE_ERROR("Cubemap: failed to load HDR: {}", path);
|
||||
return nullptr;
|
||||
}
|
||||
const u32 srcW = static_cast<u32>(w);
|
||||
const u32 srcH = static_cast<u32>(h);
|
||||
|
||||
// ---- 2. 重采样 6 面 + fp16（RGBA16F = 每像素 8B）----
|
||||
const u32 facePixels = faceSize * faceSize;
|
||||
std::vector<u16> faces(kFaceCount * facePixels * 4);
|
||||
for (u32 f = 0; f < kFaceCount; ++f) {
|
||||
u16* dst = faces.data() + f * facePixels * 4;
|
||||
for (u32 y = 0; y < faceSize; ++y) {
|
||||
for (u32 x = 0; x < faceSize; ++x) {
|
||||
const f32 u = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(faceSize);
|
||||
const f32 v = (static_cast<f32>(y) + 0.5f) / static_cast<f32>(faceSize);
|
||||
const glm::vec3 rgb =
|
||||
SampleEquirect(pixels, srcW, srcH, FaceDir(f, u, v));
|
||||
u16* px = dst + (static_cast<size_t>(y) * faceSize + x) * 4;
|
||||
px[0] = FloatToHalf(rgb.r);
|
||||
px[1] = FloatToHalf(rgb.g);
|
||||
px[2] = FloatToHalf(rgb.b);
|
||||
px[3] = 0x3C00; // 1.0
|
||||
}
|
||||
}
|
||||
}
|
||||
stbi_image_free(pixels);
|
||||
|
||||
// ---- 3. staging（HOST_VISIBLE 持久映射，逐面偏移拷入）----
|
||||
const vk::DeviceSize faceBytes =
|
||||
static_cast<vk::DeviceSize>(facePixels) * 8;
|
||||
const vk::DeviceSize total = faceBytes * kFaceCount;
|
||||
|
||||
auto stagingAlloc = allocator.CreateBuffer(
|
||||
total,
|
||||
vk::BufferUsageFlagBits::eTransferSrc,
|
||||
vk::MemoryPropertyFlagBits::eHostVisible |
|
||||
vk::MemoryPropertyFlagBits::eHostCoherent);
|
||||
std::memcpy(stagingAlloc.MappedPtr, faces.data(),
|
||||
static_cast<size_t>(total));
|
||||
faces.clear();
|
||||
faces.shrink_to_fit(); // CPU 侧临时数据尽早释放
|
||||
|
||||
// ---- 4. cube image（eCubeCompatible，6 层，fp16）----
|
||||
vk::ImageCreateInfo imageInfo;
|
||||
imageInfo.setImageType(vk::ImageType::e2D)
|
||||
.setFormat(vk::Format::eR16G16B16A16Sfloat)
|
||||
.setExtent({faceSize, faceSize, 1})
|
||||
.setMipLevels(1)
|
||||
.setArrayLayers(kFaceCount)
|
||||
.setSamples(vk::SampleCountFlagBits::e1)
|
||||
.setTiling(vk::ImageTiling::eOptimal)
|
||||
.setFlags(vk::ImageCreateFlagBits::eCubeCompatible)
|
||||
.setUsage(vk::ImageUsageFlagBits::eTransferDst |
|
||||
vk::ImageUsageFlagBits::eSampled)
|
||||
.setSharingMode(vk::SharingMode::eExclusive)
|
||||
.setInitialLayout(vk::ImageLayout::eUndefined);
|
||||
|
||||
auto imageAlloc = allocator.CreateImage(imageInfo,
|
||||
vk::MemoryPropertyFlagBits::eDeviceLocal);
|
||||
const vk::Image image = imageAlloc.Image;
|
||||
|
||||
// ---- 5. cube view + 环境采样器（clamp-to-edge：cube 采样寻址由面自动处理）----
|
||||
vk::ImageViewCreateInfo viewInfo;
|
||||
viewInfo.setImage(image)
|
||||
.setViewType(vk::ImageViewType::eCube)
|
||||
.setFormat(vk::Format::eR16G16B16A16Sfloat)
|
||||
.setSubresourceRange({
|
||||
vk::ImageAspectFlagBits::eColor, 0, 1, 0, kFaceCount});
|
||||
vk::ImageView view = device.GetDevice().createImageView(viewInfo);
|
||||
|
||||
vk::SamplerCreateInfo samplerInfo;
|
||||
samplerInfo.setMagFilter(vk::Filter::eLinear)
|
||||
.setMinFilter(vk::Filter::eLinear)
|
||||
.setMipmapMode(vk::SamplerMipmapMode::eNearest)
|
||||
.setAddressModeU(vk::SamplerAddressMode::eClampToEdge)
|
||||
.setAddressModeV(vk::SamplerAddressMode::eClampToEdge)
|
||||
.setAddressModeW(vk::SamplerAddressMode::eClampToEdge)
|
||||
.setAnisotropyEnable(VK_FALSE)
|
||||
.setBorderColor(vk::BorderColor::eIntOpaqueBlack)
|
||||
.setUnnormalizedCoordinates(VK_FALSE)
|
||||
.setCompareEnable(VK_FALSE)
|
||||
.setMinLod(0.0f)
|
||||
.setMaxLod(0.0f);
|
||||
vk::Sampler sampler = device.GetDevice().createSampler(samplerInfo);
|
||||
|
||||
// ---- 6. 一次性命令缓冲：整 cube → TransferDst，逐面拷贝 → ShaderRead ----
|
||||
vk::CommandPoolCreateInfo poolInfo;
|
||||
poolInfo.setQueueFamilyIndex(device.GetGraphicsFamily())
|
||||
.setFlags(vk::CommandPoolCreateFlagBits::eTransient);
|
||||
vk::CommandPool pool = device.GetDevice().createCommandPool(poolInfo);
|
||||
|
||||
vk::CommandBuffer cmd = device.GetDevice().allocateCommandBuffers(
|
||||
vk::CommandBufferAllocateInfo{}
|
||||
.setCommandPool(pool)
|
||||
.setLevel(vk::CommandBufferLevel::ePrimary)
|
||||
.setCommandBufferCount(1))[0];
|
||||
|
||||
cmd.begin(vk::CommandBufferBeginInfo{
|
||||
vk::CommandBufferUsageFlagBits::eOneTimeSubmit});
|
||||
|
||||
// 6 层一起转 layout（subresource range layerCount = 6）
|
||||
vk::ImageMemoryBarrier2 toTransfer;
|
||||
toTransfer.setSrcStageMask(vk::PipelineStageFlagBits2::eNone)
|
||||
.setSrcAccessMask(vk::AccessFlagBits2::eNone)
|
||||
.setDstStageMask(vk::PipelineStageFlagBits2::eTransfer)
|
||||
.setDstAccessMask(vk::AccessFlagBits2::eTransferWrite)
|
||||
.setOldLayout(vk::ImageLayout::eUndefined)
|
||||
.setNewLayout(vk::ImageLayout::eTransferDstOptimal)
|
||||
.setImage(image)
|
||||
.setSubresourceRange({vk::ImageAspectFlagBits::eColor,
|
||||
0, 1, 0, kFaceCount});
|
||||
cmd.pipelineBarrier2(
|
||||
vk::DependencyInfo{}.setImageMemoryBarriers(toTransfer));
|
||||
|
||||
for (u32 f = 0; f < kFaceCount; ++f) {
|
||||
vk::BufferImageCopy region;
|
||||
region.setBufferOffset(faceBytes * f)
|
||||
.setBufferRowLength(0)
|
||||
.setBufferImageHeight(0)
|
||||
.setImageSubresource({vk::ImageAspectFlagBits::eColor,
|
||||
0, f, 1})
|
||||
.setImageOffset({0, 0, 0})
|
||||
.setImageExtent({faceSize, faceSize, 1});
|
||||
cmd.copyBufferToImage(stagingAlloc.Buffer, image,
|
||||
vk::ImageLayout::eTransferDstOptimal, region);
|
||||
}
|
||||
|
||||
vk::ImageMemoryBarrier2 toShader;
|
||||
toShader.setSrcStageMask(vk::PipelineStageFlagBits2::eTransfer)
|
||||
.setSrcAccessMask(vk::AccessFlagBits2::eTransferWrite)
|
||||
.setDstStageMask(vk::PipelineStageFlagBits2::eFragmentShader)
|
||||
.setDstAccessMask(vk::AccessFlagBits2::eShaderRead)
|
||||
.setOldLayout(vk::ImageLayout::eTransferDstOptimal)
|
||||
.setNewLayout(vk::ImageLayout::eShaderReadOnlyOptimal)
|
||||
.setImage(image)
|
||||
.setSubresourceRange({vk::ImageAspectFlagBits::eColor,
|
||||
0, 1, 0, kFaceCount});
|
||||
cmd.pipelineBarrier2(
|
||||
vk::DependencyInfo{}.setImageMemoryBarriers(toShader));
|
||||
|
||||
cmd.end();
|
||||
|
||||
// ---- 7. 提交并等待，清理 staging ----
|
||||
vk::Fence fence = device.GetDevice().createFence({});
|
||||
device.GetGraphicsQueue().submit(
|
||||
vk::SubmitInfo{}.setCommandBuffers(cmd), fence);
|
||||
if (device.GetDevice().waitForFences(fence, VK_TRUE, UINT64_MAX) !=
|
||||
vk::Result::eSuccess) {
|
||||
ENGINE_ERROR("Cubemap upload: waitForFences failed");
|
||||
}
|
||||
|
||||
device.GetDevice().destroyFence(fence);
|
||||
device.GetDevice().destroyCommandPool(pool);
|
||||
allocator.Destroy(stagingAlloc);
|
||||
|
||||
ENGINE_INFO("Environment cubemap: {} ({}x6 {} faces, fp16)",
|
||||
path, faceSize, kFaceCount);
|
||||
|
||||
return Scope<Cubemap>(new Cubemap(device, allocator, image, view,
|
||||
sampler, imageAlloc, faceSize));
|
||||
}
|
||||
|
||||
Cubemap::~Cubemap()
|
||||
{
|
||||
const auto& vkDevice = m_Device.GetDevice();
|
||||
|
||||
vkDevice.destroySampler(m_Sampler);
|
||||
vkDevice.destroyImageView(m_View);
|
||||
|
||||
// image 资源与显存经 VMA 成对释放
|
||||
m_Allocator.Destroy(m_Allocation);
|
||||
}
|
||||
|
||||
} // namespace Engine
|
||||
Reference in New Issue
Block a user