1505 lines
57 KiB
HTML
1505 lines
57 KiB
HTML
<meta charset="utf-8" lang="en">
|
|
**NVIDIA Vulkan Ray Tracing Tutorial**
|
|
**Trace Rays Indirect**
|
|
|
|
<small>Authors: David Zhao Akeley </small>
|
|
|
|
|
|
|
|
This is an extension of the [Vulkan ray tracing tutorial](vkrt_tutorial.md.html).
|
|
|
|
We will discuss the `vkCmdTraceRaysIndirectKHR` command, which allows the
|
|
`width`, `height`, and `depth` of a trace ray command to be specifed by a
|
|
buffer on the device, rather than directly by the host. As a demonstration,
|
|
this example will add colorful lanterns to the scene that add their own light
|
|
and shadows, with a finite radius of effect. A compute shader will calculate
|
|
scissor rectangles for each lantern, and an indirect trace rays command will
|
|
dispatch rays for lanterns only within those scissor rectangles.
|
|
|
|
# Outline
|
|
|
|
The basic idea is to split up ray tracing into seperate passes. The first pass
|
|
is similar to the original tutorial: it fills in the entire output image,
|
|
calculating lighting from the main light in the scene. Subsequently, one
|
|
pass is run within a scissor rectangle for each lantern to add its light
|
|
contribution to the output image.
|
|
|
|
The steps to accomplish this are:
|
|
|
|
* Add a buffer to store lantern positions, colors, and scissor rectangles.
|
|
These lanterns are separate from the OBJ geometry loaded in the main
|
|
tutorial. Run a compute shader each frame to fill in the scissor rectangles.
|
|
|
|
* Build a BLAS for a lantern, and add lantern instances to the TLAS.
|
|
As lanterns are self-illuminating, the closest hit shader used to shade
|
|
ordinary OBJ geometry is inappropriate, so, we will add a new hit group (new
|
|
closest-hit shader) for lanterns to the SBT, and set `hitGroupId`
|
|
for lantern instances to `1` so this hit group is used.
|
|
|
|
* Modify the ray generation shader so that it emulates additive blending,
|
|
and supports drawing within scissor rectangles (in the main tutorial,
|
|
the raygen shader assumes the ray trace dispatch covers the whole screen).
|
|
|
|
* Add shadow rays in lantern passes, cast towards the lantern whose light
|
|
contribution is being added in the current pass. To detect whether the
|
|
expected lantern was hit, we add a new miss shader and two new closest
|
|
hit shaders (one for OBJ instances, one for lanterns) that return the
|
|
index of the lantern hit (if any).
|
|
|
|
* Add one `vkCmdTraceRaysIndirectKHR` call in `HelloVulkan::raytrace` for
|
|
each lantern in the scene.
|
|
|
|
If everything goes well, we should see something like this (the "lantern debug"
|
|
checkbox enables visualizing the scissor rectangles).
|
|
|
|
|
|
|
|
# Lantern Scissor Rectangles
|
|
|
|
In this step we set up the buffer storing lantern info, and a compute shader
|
|
for calculating the scissor rectangles.
|
|
|
|
## Allocate Host Storage
|
|
|
|
We first need to stage the vector of lanterns on the host in a vector.
|
|
Since this is not an animation example, we won't be concerned with
|
|
keeping track of changes to this vector: changes are forbidden after the
|
|
acceleration structures are built.
|
|
|
|
In `hello_vulkan.h`, declare a struct for holding information about
|
|
the lanterns on the host.
|
|
|
|
```` C
|
|
// Information on each colored lantern illuminating the scene.
|
|
struct Lantern
|
|
{
|
|
nvmath::vec3f position;
|
|
nvmath::vec3f color;
|
|
float brightness;
|
|
float radius; // Max world-space distance that light illuminates.
|
|
};
|
|
````
|
|
|
|
Then declare a vector of `Lantern` and add
|
|
a new function for configuring a new lantern in the scene.
|
|
|
|
```` C
|
|
// Array of lanterns in scene. Not modifiable after acceleration structure build.
|
|
std::vector<Lantern> m_lanterns;
|
|
void addLantern(nvmath::vec3f pos, nvmath::vec3f color, float brightness, float radius);
|
|
````
|
|
|
|
The `addLantern` function is implemented as
|
|
|
|
```` C
|
|
// Add a light-emitting colored lantern to the scene. May only be called before TLAS build.
|
|
void HelloVulkan::addLantern(nvmath::vec3f pos, nvmath::vec3f color, float brightness, float radius)
|
|
{
|
|
assert(m_lanternCount == 0); // Indicates TLAS build has not happened yet.
|
|
|
|
m_lanterns.push_back({pos, color, brightness, radius});
|
|
}
|
|
````
|
|
|
|
In `main.cpp`, we insert calls for adding some lanterns.
|
|
|
|
```` C
|
|
// Creation of the example
|
|
helloVk.loadModel(nvh::findFile("media/scenes/Medieval_building.obj", defaultSearchPaths, true));
|
|
helloVk.loadModel(nvh::findFile("media/scenes/plane.obj", defaultSearchPaths, true));
|
|
helloVk.addLantern({ 8.000f, 1.100f, 3.600f}, {1.0f, 0.0f, 0.0f}, 0.4f, 4.0f);
|
|
helloVk.addLantern({ 8.000f, 0.600f, 3.900f}, {0.0f, 1.0f, 0.0f}, 0.4f, 4.0f);
|
|
helloVk.addLantern({ 8.000f, 1.100f, 4.400f}, {0.0f, 0.0f, 1.0f}, 0.4f, 4.0f);
|
|
helloVk.addLantern({ 1.730f, 1.812f, -1.604f}, {0.0f, 0.4f, 0.4f}, 0.4f, 4.0f);
|
|
helloVk.addLantern({ 1.730f, 1.862f, 1.916f}, {0.0f, 0.2f, 0.4f}, 0.3f, 3.0f);
|
|
helloVk.addLantern({-2.000f, 1.900f, -0.700f}, {0.8f, 0.8f, 0.6f}, 0.4f, 3.9f);
|
|
helloVk.addLantern({ 0.100f, 0.080f, -2.392f}, {1.0f, 0.0f, 1.0f}, 0.5f, 5.0f);
|
|
helloVk.addLantern({ 1.948f, 0.080f, 0.598f}, {1.0f, 1.0f, 1.0f}, 0.6f, 6.0f);
|
|
helloVk.addLantern({-2.300f, 0.080f, 2.100f}, {0.0f, 0.7f, 0.0f}, 0.6f, 6.0f);
|
|
helloVk.addLantern({-1.400f, 4.300f, 0.150f}, {1.0f, 1.0f, 0.0f}, 0.7f, 7.0f);
|
|
````
|
|
|
|
## Lantern Device Storage
|
|
|
|
In `hello_vulkan.h`, declare a struct for storing lanterns on the device.
|
|
This includes the host information, plus a scissor rectangle.
|
|
```` C
|
|
// Information on each colored lantern, plus the info needed for dispatching the
|
|
// indirect ray trace command used to add its brightness effect.
|
|
// The dispatched ray trace covers pixels (offsetX, offsetY) to
|
|
// (offsetX + indirectCommand.width - 1, offsetY + indirectCommand.height - 1).
|
|
struct LanternIndirectEntry
|
|
{
|
|
// Filled in by the device using a compute shader.
|
|
// NOTE: I rely on indirectCommand being the first member.
|
|
VkTraceRaysIndirectCommandKHR indirectCommand;
|
|
int32_t offsetX;
|
|
int32_t offsetY;
|
|
|
|
// Filled in by the host.
|
|
Lantern lantern;
|
|
};
|
|
````
|
|
|
|
!!! NOTE
|
|
`VkTraceRaysIndirectCommandKHR` is just a struct of 3 `int32_t` defining
|
|
the `width`, `height`, `depth` of a trace ray command.
|
|
|
|
|
|
We also declare an equivalent structure for shaders in the file
|
|
`LanternIndirectEntry.glsl`. We avoid using `vec3` due to differences
|
|
in alignment in C++ and GLSL.
|
|
```` C
|
|
struct LanternIndirectEntry
|
|
{
|
|
// VkTraceRaysIndirectCommandKHR
|
|
int indirectWidth;
|
|
int indirectHeight;
|
|
int indirectDepth;
|
|
|
|
// Pixel coordinate of scissor rect upper-left.
|
|
int offsetX;
|
|
int offsetY;
|
|
|
|
// Lantern starts here:
|
|
// Can't use vec3 due to alignment.
|
|
float x, y, z;
|
|
float red, green, blue;
|
|
float brightness;
|
|
float radius;
|
|
};
|
|
````
|
|
|
|
To store the lanterns on the device, declare the Vulkan buffer of `LanternIndirectEntry`
|
|
in `hello_vulkan.h`
|
|
|
|
```` C
|
|
// Buffer to source vkCmdTraceRaysIndirectKHR indirect parameters and lantern color,
|
|
// position, etc. from when doing lantern lighting passes.
|
|
nvvk::Buffer m_lanternIndirectBuffer;
|
|
VkDeviceSize m_lanternCount = 0; // Set to actual lantern count after TLAS build, as
|
|
// that is the point no more lanterns may be added.
|
|
````
|
|
|
|
and fill it with a `createLanternIndirectBuffer` function. For performance,
|
|
we allocate a device-local buffer. We need usage flags for
|
|
|
|
* Storage buffer use, so the compute shader can write to it.
|
|
* Indirect buffer use, so we can source indirect parameters from it when dispatching a ray trace.
|
|
* Device address use, as `vkCmdTraceRaysIndirectKHR` expects a device address.
|
|
* Transfer dst use, so the buffer can be initialized with the lantern colors and positions from `m_lanterns`.
|
|
|
|
```` C
|
|
// Allocate the buffer used to pass lantern info + ray trace indirect parameters to ray tracer.
|
|
// Fill in the lantern info from m_lanterns (indirect info is filled per-frame on device
|
|
// using a compute shader). Must be called only after TLAS build.
|
|
//
|
|
// The buffer is an array of LanternIndirectEntry, entry i is for m_lanterns[i].
|
|
void HelloVulkan::createLanternIndirectBuffer()
|
|
{
|
|
assert(m_lanternCount > 0);
|
|
assert(m_lanternCount == m_lanterns.size());
|
|
|
|
// m_alloc behind the scenes uses cmdBuf to transfer data to the buffer.
|
|
nvvk::CommandPool cmdBufGet(m_device, m_graphicsQueueIndex);
|
|
vk::CommandBuffer cmdBuf = cmdBufGet.createCommandBuffer();
|
|
|
|
using Usage = vk::BufferUsageFlagBits;
|
|
m_lanternIndirectBuffer =
|
|
m_alloc.createBuffer(sizeof(LanternIndirectEntry) * m_lanternCount,
|
|
Usage::eIndirectBuffer | Usage::eTransferDst
|
|
| Usage::eShaderDeviceAddress | Usage::eStorageBuffer,
|
|
vk::MemoryPropertyFlagBits::eDeviceLocal);
|
|
|
|
std::vector<LanternIndirectEntry> entries(m_lanternCount);
|
|
for (size_t i = 0; i < m_lanternCount; ++i) entries[i].lantern = m_lanterns[i];
|
|
cmdBuf.updateBuffer(m_lanternIndirectBuffer.buffer, 0, entries.size() * sizeof entries[0], entries.data());
|
|
|
|
cmdBufGet.submitAndWait(cmdBuf);
|
|
}
|
|
````
|
|
|
|
Call this function in `main.cpp`, after the AS build (the AS build will be modified later).
|
|
|
|
```` C
|
|
helloVk.initRayTracing();
|
|
helloVk.createBottomLevelAS();
|
|
helloVk.createTopLevelAS();
|
|
helloVk.createLanternIndirectBuffer();
|
|
````
|
|
|
|
## Set up Compute Shader
|
|
|
|
The compute shader will need the view and projection matrices, plus the Z near plane,
|
|
screen dimensions, and length of the `LanternIndirectBuffer` array, in order to
|
|
compute the scissor rectangles. Feed all of this in with a push constant, declared
|
|
in `hello_vulkan.h`
|
|
|
|
```` C
|
|
// Push constant for compute shader filling lantern indirect buffer.
|
|
// Barely fits in 128-byte push constant limit guaranteed by spec.
|
|
struct LanternIndirectPushConstants
|
|
{
|
|
nvmath::vec4 viewRowX; // First 3 rows of view matrix.
|
|
nvmath::vec4 viewRowY; // Set w=1 implicitly in shader.
|
|
nvmath::vec4 viewRowZ;
|
|
|
|
nvmath::mat4 proj; // Perspective matrix
|
|
float nearZ; // Near plane used to create projection matrix.
|
|
|
|
// Pixel dimensions of output image (needed to scale NDC to screen coordinates).
|
|
int32_t screenX;
|
|
int32_t screenY;
|
|
|
|
// Length of the LanternIndirectEntry array.
|
|
int32_t lanternCount;
|
|
} m_lanternIndirectPushConstants;
|
|
````
|
|
|
|
!!! NOTE Push Constant Limit
|
|
The Vulkan spec only guarantees 128 bytes of push constant, so to make everything
|
|
fit, we have to chop off the implicit bottom row of the view matrix.
|
|
|
|
This push constant is consumed in the compute shader `shaders/lanternIndirect.comp`.
|
|
We go through `lanternCount` iterations of a loop that fill in the scissor
|
|
rectangle for each `LanternIndirectEntry` (splitting the work among 128
|
|
invocations of the work group).
|
|
|
|
```` C
|
|
#version 460
|
|
#extension GL_GOOGLE_include_directive : enable
|
|
|
|
// Compute shader for filling in raytrace indirect parameters for each lantern
|
|
// based on the current camera position (passed as view and proj matrix in
|
|
// push constant).
|
|
//
|
|
// Designed to be dispatched with only one work group; it alone fills in
|
|
// the entire lantern array (of length lanternCount, in also push constant).
|
|
|
|
#define LOCAL_SIZE 128
|
|
layout(local_size_x = LOCAL_SIZE, local_size_y = 1, local_size_z = 1) in;
|
|
|
|
#include "LanternIndirectEntry.glsl"
|
|
|
|
layout(binding = 0, set = 0) buffer LanternArray { LanternIndirectEntry lanterns[]; } lanterns;
|
|
|
|
layout(push_constant) uniform Constants
|
|
{
|
|
vec4 viewRowX;
|
|
vec4 viewRowY;
|
|
vec4 viewRowZ;
|
|
mat4 proj;
|
|
float nearZ;
|
|
int screenX;
|
|
int screenY;
|
|
int lanternCount;
|
|
}
|
|
pushC;
|
|
|
|
// Copy the technique of "2D Polyhedral Bounds of a Clipped,
|
|
// Perspective-Projected 3D Sphere" M. Mara M. McGuire
|
|
// http://jcgt.org/published/0002/02/05/paper.pdf
|
|
// to compute a screen-space rectangle covering the given Lantern's
|
|
// light radius-of-effect. Result is in screen (pixel) coordinates.
|
|
void getScreenCoordBox(in LanternIndirectEntry lantern, out ivec2 lower, out ivec2 upper);
|
|
|
|
// Use the xyz and radius of lanterns[i] plus the transformation matrices
|
|
// in pushC to fill in the offset and indirect parameters of lanterns[i]
|
|
// (defines the screen rectangle that this lantern's light is bounded in).
|
|
void fillIndirectEntry(int i)
|
|
{
|
|
LanternIndirectEntry lantern = lanterns.lanterns[i];
|
|
ivec2 lower, upper;
|
|
getScreenCoordBox(lantern, lower, upper);
|
|
|
|
lanterns.lanterns[i].indirectWidth = max(0, upper.x - lower.x);
|
|
lanterns.lanterns[i].indirectHeight = max(0, upper.y - lower.y);
|
|
lanterns.lanterns[i].indirectDepth = 1;
|
|
lanterns.lanterns[i].offsetX = lower.x;
|
|
lanterns.lanterns[i].offsetY = lower.y;
|
|
}
|
|
|
|
void main()
|
|
{
|
|
for (int i = int(gl_LocalInvocationID.x); i < pushC.lanternCount; i += LOCAL_SIZE)
|
|
{
|
|
fillIndirectEntry(i);
|
|
}
|
|
}
|
|
|
|
/** Center is in camera space */
|
|
void getBoundingBox(
|
|
in vec3 center,
|
|
in float radius,
|
|
in float nearZ,
|
|
in mat4 projMatrix,
|
|
out vec2 ndc_low,
|
|
out vec2 ndc_high) {
|
|
````
|
|
!!! TIP
|
|
Omitted code for computing scissor rectangles, taken from "2D Polyhedral Bounds of a Clipped,
|
|
Perspective-Projected 3D Sphere" by Michael Mara and Morgan McGuire.
|
|
http://jcgt.org/published/0002/02/05/paper.pdf
|
|
```` C
|
|
}
|
|
|
|
void getScreenCoordBox(in LanternIndirectEntry lantern, out ivec2 lower, out ivec2 upper)
|
|
{
|
|
vec4 lanternWorldCenter = vec4(lantern.x, lantern.y, lantern.z, 1);
|
|
vec3 center = vec3(
|
|
dot(pushC.viewRowX, lanternWorldCenter),
|
|
dot(pushC.viewRowY, lanternWorldCenter),
|
|
dot(pushC.viewRowZ, lanternWorldCenter));
|
|
vec2 ndc_low, ndc_high;
|
|
float paperNearZ = -abs(pushC.nearZ); // Paper expected negative nearZ, took 2 days to figure out!
|
|
getBoundingBox(center, lantern.radius, paperNearZ, pushC.proj, ndc_low, ndc_high);
|
|
|
|
// Convert NDC [-1,+1]^2 coordinates to screen coordinates, and clamp to stay in bounds.
|
|
|
|
lower.x = clamp(int((ndc_low.x * 0.5 + 0.5) * pushC.screenX), 0, pushC.screenX);
|
|
lower.y = clamp(int((ndc_low.y * 0.5 + 0.5) * pushC.screenY), 0, pushC.screenY);
|
|
upper.x = clamp(int((ndc_high.x * 0.5 + 0.5) * pushC.screenX), 0, pushC.screenX);
|
|
upper.y = clamp(int((ndc_high.y * 0.5 + 0.5) * pushC.screenY), 0, pushC.screenY);
|
|
}
|
|
````
|
|
|
|
Now we just have to fill out the usual boilerplate for setting up the descriptor
|
|
set (passes the `LanternIndirectEntry` array) and compute pipeline. We only have
|
|
to allocate one descriptor as the `LanternIndirectEntry` array never changes.
|
|
|
|
`hello_vulkan.h`:
|
|
|
|
```` C
|
|
nvvk::DescriptorSetBindings m_lanternIndirectDescSetLayoutBind;
|
|
vk::DescriptorPool m_lanternIndirectDescPool;
|
|
vk::DescriptorSetLayout m_lanternIndirectDescSetLayout;
|
|
vk::DescriptorSet m_lanternIndirectDescSet;
|
|
vk::PipelineLayout m_lanternIndirectCompPipelineLayout;
|
|
vk::Pipeline m_lanternIndirectCompPipeline;
|
|
````
|
|
|
|
`hello_vulkan.cpp`:
|
|
|
|
```` C
|
|
//--------------------------------------------------------------------------------------------------
|
|
// The compute shader just needs read/write access to the buffer of LanternIndirectEntry.
|
|
void HelloVulkan::createLanternIndirectDescriptorSet()
|
|
{
|
|
using vkDT = vk::DescriptorType;
|
|
using vkSS = vk::ShaderStageFlagBits;
|
|
using vkDSLB = vk::DescriptorSetLayoutBinding;
|
|
|
|
// Lantern buffer (binding = 0)
|
|
m_lanternIndirectDescSetLayoutBind.addBinding( //
|
|
vkDSLB(0, vkDT::eStorageBuffer, 1, vkSS::eCompute));
|
|
|
|
m_lanternIndirectDescPool = m_lanternIndirectDescSetLayoutBind.createPool(m_device);
|
|
m_lanternIndirectDescSetLayout = m_lanternIndirectDescSetLayoutBind.createLayout(m_device);
|
|
m_lanternIndirectDescSet =
|
|
m_device.allocateDescriptorSets({m_lanternIndirectDescPool, 1, &m_lanternIndirectDescSetLayout})[0];
|
|
|
|
assert(m_lanternIndirectBuffer.buffer);
|
|
vk::DescriptorBufferInfo lanternBufferInfo{
|
|
m_lanternIndirectBuffer.buffer, 0, m_lanternCount * sizeof(LanternIndirectEntry)};
|
|
|
|
std::vector<vk::WriteDescriptorSet> writes;
|
|
writes.emplace_back(m_lanternIndirectDescSetLayoutBind.makeWrite(m_lanternIndirectDescSet, 0, &lanternBufferInfo));
|
|
m_device.updateDescriptorSets(static_cast<uint32_t>(writes.size()), writes.data(), 0, nullptr);
|
|
}
|
|
|
|
// Create compute pipeline used to fill m_lanternIndirectBuffer with parameters
|
|
// for dispatching the correct number of ray traces.
|
|
void HelloVulkan::createLanternIndirectCompPipeline()
|
|
{
|
|
// Compile compute shader and package as stage.
|
|
vk::ShaderModule computeShader =
|
|
nvvk::createShaderModule(m_device, //
|
|
nvh::loadFile("shaders/lanternIndirect.comp.spv", true, defaultSearchPaths, true));
|
|
vk::PipelineShaderStageCreateInfo stageInfo;
|
|
stageInfo.setStage(vk::ShaderStageFlagBits::eCompute);
|
|
stageInfo.setModule(computeShader);
|
|
stageInfo.setPName("main");
|
|
|
|
// Set up push constant and pipeline layout.
|
|
constexpr auto pushSize = static_cast<uint32_t>(sizeof(m_lanternIndirectPushConstants));
|
|
vk::PushConstantRange pushCRange = {vk::ShaderStageFlagBits::eCompute, 0, pushSize};
|
|
static_assert(pushSize <= 128, "Spec guarantees only 128 byte push constant");
|
|
vk::PipelineLayoutCreateInfo layoutInfo;
|
|
layoutInfo.setSetLayoutCount(1);
|
|
layoutInfo.setPSetLayouts(&m_lanternIndirectDescSetLayout);
|
|
layoutInfo.setPushConstantRangeCount(1);
|
|
layoutInfo.setPPushConstantRanges(&pushCRange);
|
|
m_lanternIndirectCompPipelineLayout = m_device.createPipelineLayout(layoutInfo);
|
|
|
|
// Create compute pipeline.
|
|
vk::ComputePipelineCreateInfo pipelineInfo;
|
|
pipelineInfo.setStage(stageInfo);
|
|
pipelineInfo.setLayout(m_lanternIndirectCompPipelineLayout);
|
|
m_lanternIndirectCompPipeline = static_cast<const vk::Pipeline&>(m_device.createComputePipeline({}, pipelineInfo));
|
|
|
|
m_device.destroy(computeShader);
|
|
}
|
|
````
|
|
|
|
`main.cpp` (add after indirect buffer initialization).
|
|
|
|
```` C
|
|
// #VKRay
|
|
helloVk.initRayTracing();
|
|
helloVk.createBottomLevelAS();
|
|
helloVk.createTopLevelAS();
|
|
helloVk.createLanternIndirectBuffer();
|
|
helloVk.createRtDescriptorSet();
|
|
helloVk.createRtPipeline();
|
|
helloVk.createLanternIndirectDescriptorSet();
|
|
helloVk.createLanternIndirectCompPipeline();
|
|
````
|
|
|
|
## Call Compute Shader
|
|
|
|
In `HelloVulkan::raytrace`, we have to fill in the earlier push constant and
|
|
dispatch the compute shader before moving
|
|
on to the actual ray tracing. This is rather verbose due to the need for a
|
|
pipeline barrier synchronizing access to the `LanternIndirectEntry` array
|
|
between the compute shader and indirect draw stages.
|
|
|
|
```` C
|
|
void HelloVulkan::raytrace(const vk::CommandBuffer& cmdBuf, const nvmath::vec4f& clearColor)
|
|
{
|
|
// Before tracing rays, we need to dispatch the compute shaders that
|
|
// fill in the ray trace indirect parameters for each lantern pass.
|
|
|
|
// First, barrier before, ensure writes aren't visible to previous frame.
|
|
vk::BufferMemoryBarrier bufferBarrier;
|
|
bufferBarrier.setSrcAccessMask(vk::AccessFlagBits::eIndirectCommandRead);
|
|
bufferBarrier.setDstAccessMask(vk::AccessFlagBits::eShaderWrite);
|
|
bufferBarrier.setSrcQueueFamilyIndex(VK_QUEUE_FAMILY_IGNORED);
|
|
bufferBarrier.setDstQueueFamilyIndex(VK_QUEUE_FAMILY_IGNORED);
|
|
bufferBarrier.setBuffer(m_lanternIndirectBuffer.buffer);
|
|
bufferBarrier.offset = 0;
|
|
bufferBarrier.size = m_lanternCount * sizeof m_lanterns[0];
|
|
cmdBuf.pipelineBarrier( //
|
|
vk::PipelineStageFlagBits::eDrawIndirect, //
|
|
vk::PipelineStageFlagBits::eComputeShader,//
|
|
vk::DependencyFlags(0), //
|
|
{}, {bufferBarrier}, {});
|
|
|
|
// Bind compute shader, update push constant and descriptors, dispatch compute.
|
|
cmdBuf.bindPipeline(vk::PipelineBindPoint::eCompute, m_lanternIndirectCompPipeline);
|
|
nvmath::mat4 view = getViewMatrix();
|
|
m_lanternIndirectPushConstants.viewRowX = view.row(0);
|
|
m_lanternIndirectPushConstants.viewRowY = view.row(1);
|
|
m_lanternIndirectPushConstants.viewRowZ = view.row(2);
|
|
m_lanternIndirectPushConstants.proj = getProjMatrix();
|
|
m_lanternIndirectPushConstants.nearZ = nearZ;
|
|
m_lanternIndirectPushConstants.screenX = m_size.width;
|
|
m_lanternIndirectPushConstants.screenY = m_size.height;
|
|
m_lanternIndirectPushConstants.lanternCount = m_lanternCount;
|
|
cmdBuf.pushConstants<LanternIndirectPushConstants>(
|
|
m_lanternIndirectCompPipelineLayout,
|
|
vk::ShaderStageFlagBits::eCompute,
|
|
0, m_lanternIndirectPushConstants);
|
|
cmdBuf.bindDescriptorSets(
|
|
vk::PipelineBindPoint::eCompute, m_lanternIndirectCompPipelineLayout, 0, {m_lanternIndirectDescSet}, {});
|
|
cmdBuf.dispatch(1, 1, 1);
|
|
|
|
// Ensure compute results are visible when doing indirect ray trace.
|
|
bufferBarrier.setSrcAccessMask(vk::AccessFlagBits::eShaderWrite);
|
|
bufferBarrier.setDstAccessMask(vk::AccessFlagBits::eIndirectCommandRead);
|
|
cmdBuf.pipelineBarrier( //
|
|
vk::PipelineStageFlagBits::eComputeShader, //
|
|
vk::PipelineStageFlagBits::eDrawIndirect, //
|
|
vk::DependencyFlags(0), //
|
|
{}, {bufferBarrier}, {});
|
|
|
|
|
|
// Now move on to the actual ray tracing.
|
|
m_debug.beginLabel(cmdBuf, "Ray trace");
|
|
````
|
|
|
|
!!! TIP `eDrawIndirect`
|
|
`vk::PipelineStageFlagBits::eDrawIndirect` (`VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT`)
|
|
covers the stage that sources indirect paramaters for compute and ray trace
|
|
indirect commands, not just graphics draw indirect commands.
|
|
|
|
Since the near plane and view/projection matrices are used in multiple places now,
|
|
they were factored out to common code in `hello_vulkan.h`.
|
|
|
|
```` C
|
|
nvmath::mat4 getViewMatrix()
|
|
{
|
|
return CameraManip.getMatrix();
|
|
}
|
|
|
|
static constexpr float nearZ = 0.1f;
|
|
nvmath::mat4 getProjMatrix()
|
|
{
|
|
const float aspectRatio = m_size.width / static_cast<float>(m_size.height);
|
|
return nvmath::perspectiveVK(CameraManip.getFov(), aspectRatio, nearZ, 1000.0f);
|
|
}
|
|
````
|
|
|
|
The function for updating the uniform buffer is tweaked to match.
|
|
|
|
```` C
|
|
void HelloVulkan::updateUniformBuffer(const vk::CommandBuffer& cmdBuf)
|
|
{
|
|
const float aspectRatio = m_size.width / static_cast<float>(m_size.height);
|
|
|
|
CameraMatrices ubo = {};
|
|
ubo.view = getViewMatrix();
|
|
ubo.proj = getProjMatrix();
|
|
````
|
|
|
|
# Lantern Acceleration Structures and Closest Hit Shader
|
|
|
|
## Bottom-level Acceleration Structure
|
|
|
|
Lanterns will be drawn as spheres approximated by a triangular mesh. Declare
|
|
in `hello_vulkan.h` functions for generating this mesh, and declare Vulkan
|
|
buffers for storing the mesh's positions and indices, and a `BlasInput`
|
|
for delivering this sphere mesh to the BLAS builder.
|
|
|
|
```` C
|
|
private:
|
|
void fillLanternVerts(std::vector<nvmath::vec3f>& vertices, std::vector<uint32_t>& indices);
|
|
void createLanternModel();
|
|
|
|
// Used to store lantern model, generated at runtime.
|
|
const float m_lanternModelRadius = 0.125;
|
|
nvvk::Buffer m_lanternVertexBuffer;
|
|
nvvk::Buffer m_lanternIndexBuffer;
|
|
nvvk::RaytracingBuilderKHR::BlasInput m_lanternBlasInput{};
|
|
|
|
// Index of lantern's BLAS in the BLAS array stored in m_rtBuilder.
|
|
size_t m_lanternBlasId;
|
|
````
|
|
|
|
In order to focus on the ray tracing, I omit the code for generating those vertex and index
|
|
buffers. The relevent code in `HelloVulkan::createLanternModel` for creating the `BlasInput` is
|
|
|
|
```` C
|
|
// Package vertex and index buffers as BlasInput.
|
|
vk::DeviceAddress vertexAddress = m_device.getBufferAddress({m_lanternVertexBuffer.buffer});
|
|
vk::DeviceAddress indexAddress = m_device.getBufferAddress({m_lanternIndexBuffer.buffer});
|
|
|
|
uint32_t maxPrimitiveCount = uint32_t(indices.size() / 3);
|
|
|
|
// Describe buffer as packed array of float vec3.
|
|
vk::AccelerationStructureGeometryTrianglesDataKHR triangles;
|
|
triangles.setVertexFormat(vk::Format::eR32G32B32Sfloat); // vec3 vertex position data.
|
|
triangles.setVertexData(vertexAddress);
|
|
triangles.setVertexStride(sizeof(nvmath::vec3f));
|
|
// Describe index data (32-bit unsigned int)
|
|
triangles.setIndexType(vk::IndexType::eUint32);
|
|
triangles.setIndexData(indexAddress);
|
|
// Indicate identity transform by setting transformData to null device pointer.
|
|
triangles.setTransformData({});
|
|
triangles.setMaxVertex(vertices.size());
|
|
|
|
// Identify the above data as containing opaque triangles.
|
|
vk::AccelerationStructureGeometryKHR asGeom;
|
|
asGeom.setGeometryType(vk::GeometryTypeKHR::eTriangles);
|
|
asGeom.setFlags(vk::GeometryFlagBitsKHR::eOpaque);
|
|
asGeom.geometry.setTriangles(triangles);
|
|
|
|
// The entire array will be used to build the BLAS.
|
|
vk::AccelerationStructureBuildRangeInfoKHR offset;
|
|
offset.setFirstVertex(0);
|
|
offset.setPrimitiveCount(maxPrimitiveCount);
|
|
offset.setPrimitiveOffset(0);
|
|
offset.setTransformOffset(0);
|
|
|
|
// Our blas is made from only one geometry, but could be made of many geometries
|
|
m_lanternBlasInput.asGeometry.emplace_back(asGeom);
|
|
m_lanternBlasInput.asBuildOffsetInfo.emplace_back(offset);
|
|
````
|
|
|
|
The principle difference from before is that the vertex array is now a packed array of
|
|
float 3-vectors, hence, we call `triangles.setVertexStride(sizeof(nvmath::vec3f));`.
|
|
|
|
Then, we add a call to create a lantern model and add the lantern model to the list of
|
|
BLAS to build in `HelloVulkan::createBottomLevelAS`. Since we'll need the index of
|
|
the lantern BLAS later to add lantern instances in the TLAS build, store the
|
|
BLAS index for the lantern in `m_lanternBlasId`.
|
|
|
|
```` C
|
|
// Build the array of BLAS in m_rtBuilder. There are `m_objModel.size() + 1`-many BLASes.
|
|
// The first `m_objModel.size()` are used for OBJ model BLASes, and the last one
|
|
// is used for the lanterns (model generated at runtime).
|
|
void HelloVulkan::createBottomLevelAS()
|
|
{
|
|
// BLAS - Storing each primitive in a geometry
|
|
std::vector<nvvk::RaytracingBuilderKHR::BlasInput> allBlas;
|
|
allBlas.reserve(m_objModel.size() + 1);
|
|
|
|
// Add OBJ models.
|
|
for(const auto& obj : m_objModel)
|
|
{
|
|
auto blas = objectToVkGeometryKHR(obj);
|
|
|
|
// We could add more geometry in each BLAS, but we add only one for now
|
|
allBlas.emplace_back(blas);
|
|
}
|
|
|
|
// Add lantern model.
|
|
createLanternModel();
|
|
m_lanternBlasId = allBlas.size();
|
|
allBlas.emplace_back(m_lanternBlasInput);
|
|
|
|
m_rtBuilder.buildBlas(allBlas, vk::BuildAccelerationStructureFlagBitsKHR::ePreferFastTrace);
|
|
}
|
|
````
|
|
|
|
## Top-level acceleration structure
|
|
|
|
In the TLAS build function, we add a loop for adding each lantern instance. This is also
|
|
the point that the lanterns are set-in-stone (no more modifications to `m_lanterns`), so
|
|
write `m_lanternCount`.
|
|
|
|
```` C
|
|
// Build the TLAS in m_rtBuilder. Requires that the BLASes were already built and
|
|
// that all ObjInstance and lanterns have been added. One instance with hitGroupId=0
|
|
// is created for every OBJ instance, and one instance with hitGroupId=1 for each lantern.
|
|
//
|
|
// gl_InstanceCustomIndexEXT will be the index of the instance or lantern in m_objInstance or
|
|
// m_lanterns respectively.
|
|
void HelloVulkan::createTopLevelAS()
|
|
{
|
|
assert(m_lanternCount == 0);
|
|
m_lanternCount = m_lanterns.size();
|
|
|
|
std::vector<nvvk::RaytracingBuilderKHR::Instance> tlas;
|
|
tlas.reserve(m_objInstance.size() + m_lanternCount);
|
|
|
|
// Add the OBJ instances.
|
|
for(int i = 0; i < static_cast<int>(m_objInstance.size()); i++)
|
|
{
|
|
nvvk::RaytracingBuilderKHR::Instance rayInst;
|
|
rayInst.transform = m_objInstance[i].transform; // Position of the instance
|
|
rayInst.instanceCustomId = i; // gl_InstanceCustomIndexEXT
|
|
rayInst.blasId = m_objInstance[i].objIndex;
|
|
rayInst.hitGroupId = 0; // We will use the same hit group for all OBJ
|
|
rayInst.flags = VK_GEOMETRY_INSTANCE_TRIANGLE_FACING_CULL_DISABLE_BIT_KHR;
|
|
tlas.emplace_back(rayInst);
|
|
}
|
|
|
|
// Add lantern instances.
|
|
for(int i = 0; i < static_cast<int>(m_lanterns.size()); ++i)
|
|
{
|
|
nvvk::RaytracingBuilderKHR::Instance lanternInstance;
|
|
lanternInstance.transform = nvmath::translation_mat4(m_lanterns[i].position);
|
|
lanternInstance.instanceCustomId = i;
|
|
lanternInstance.blasId = m_lanternBlasId;
|
|
lanternInstance.hitGroupId = 1; // Next hit group is for lanterns.
|
|
lanternInstance.flags = VK_GEOMETRY_INSTANCE_TRIANGLE_FACING_CULL_DISABLE_BIT_KHR;
|
|
tlas.emplace_back(lanternInstance);
|
|
}
|
|
|
|
m_rtBuilder.buildTlas(tlas, vk::BuildAccelerationStructureFlagBitsKHR::ePreferFastTrace);
|
|
}
|
|
````
|
|
|
|
The principle differences are:
|
|
|
|
* `instanceCustomId` is set to the index of the lantern in `m_lanternIndirectBuffer`, so we
|
|
can look up the lantern color in the forthcoming closest hit shader.
|
|
|
|
* `hitGroupId` is set to `1`, so that lanterns will use a new closest hit shader instead
|
|
of the old one for OBJs.
|
|
|
|
!!! TIP Helper Reminders
|
|
`instanceCustomId` corresponds to `VkAccelerationStructureInstanceKHR::instanceCustomIndex` in host
|
|
code and `gl_InstanceCustomIndexEXT` in shader code.
|
|
|
|
`hitGroupId` corresponds to `VkAccelerationStructureInstanceKHR::instanceShaderBindingTableRecordOffset`.
|
|
|
|
`blasId` has no Vulkan equivalent; it is translated to a BLAS device address in the `m_rtBuilder` helper.
|
|
|
|
## Lantern Primary Ray Closest Hit Shader
|
|
|
|
We now implement the closest hit shader for lanterns hit by primary rays (rays
|
|
cast starting from the eye). First, we need to do a bit of preparation:
|
|
|
|
* Add a bool to `hitPayload` to control whether additive blending is enabled or
|
|
not. The lanterns will be drawn at a constant brightness, so additive blending
|
|
is enabled for rays hitting OBJ instances and disabled for rays hitting lanterns.
|
|
The raygen shader will be updated later to take this bool into account.
|
|
|
|
* Access the GLSL definition of `LanternIndirectEntry` so we can look up the lantern color.
|
|
|
|
* Add a descriptor set to the raytrace pipeline to deliver the
|
|
|
|
We do the first two tasks in `raycommon.glsl`.
|
|
|
|
```` C
|
|
#include "LanternIndirectEntry.glsl"
|
|
|
|
struct hitPayload
|
|
{
|
|
vec3 hitValue;
|
|
bool additiveBlending;
|
|
};
|
|
````
|
|
|
|
The last task is done in `HelloVulkan::createRtDescriptorSet`
|
|
|
|
```` C
|
|
// This descriptor set holds the Acceleration structure, output image, and lanterns array buffer.
|
|
//
|
|
void HelloVulkan::createRtDescriptorSet()
|
|
{
|
|
using vkDT = vk::DescriptorType;
|
|
using vkSS = vk::ShaderStageFlagBits;
|
|
using vkDSLB = vk::DescriptorSetLayoutBinding;
|
|
|
|
// ...
|
|
|
|
// Lantern buffer (binding = 2)
|
|
m_rtDescSetLayoutBind.addBinding( //
|
|
vkDSLB(2, vkDT::eStorageBuffer, 1, vkSS::eRaygenKHR | vkSS::eClosestHitKHR));
|
|
assert(m_lanternCount > 0);
|
|
|
|
// ...
|
|
|
|
std::vector<vk::WriteDescriptorSet> writes;
|
|
|
|
// ...
|
|
|
|
writes.emplace_back(m_rtDescSetLayoutBind.makeWrite(m_rtDescSet, 2, &lanternBufferInfo));
|
|
m_device.updateDescriptorSets(static_cast<uint32_t>(writes.size()), writes.data(), 0, nullptr);
|
|
}
|
|
````
|
|
|
|
Now we can implement the new closest hit shader. Name this shader `lantern.rchit`.
|
|
|
|
```` C
|
|
#version 460
|
|
#extension GL_EXT_ray_tracing : require
|
|
#extension GL_EXT_nonuniform_qualifier : enable
|
|
#extension GL_EXT_scalar_block_layout : enable
|
|
#extension GL_GOOGLE_include_directive : enable
|
|
#include "raycommon.glsl"
|
|
|
|
// Closest hit shader invoked when a primary ray hits a lantern.
|
|
|
|
// clang-format off
|
|
layout(location = 0) rayPayloadInEXT hitPayload prd;
|
|
|
|
layout(binding = 2, set = 0) buffer LanternArray { LanternIndirectEntry lanterns[]; } lanterns;
|
|
|
|
// clang-format on
|
|
|
|
void main()
|
|
{
|
|
// Just look up this lantern's color. Self-illuminating, so no lighting calculations.
|
|
LanternIndirectEntry lantern = lanterns.lanterns[nonuniformEXT(gl_InstanceCustomIndexEXT)];
|
|
prd.hitValue = vec3(lantern.red, lantern.green, lantern.blue);
|
|
prd.additiveBlending = false;
|
|
}
|
|
````
|
|
|
|
This shader is fairly simple, we just had to look up the lantern color and return it in the
|
|
payload. Here, we used the fact that in the TLAS instances setup, we set a lantern instance's
|
|
`gl_InstanceCustomIndexEXT` to its position in the lanterns array.
|
|
|
|
Now we just have to add the new hit group to the pipeline. This is more of the same,
|
|
in the `HelloVulkan::createRtPipeline` function, we add the lantern closest hit
|
|
group after the OBJ hit group, to match the `hitGroupId`s assigned earlier in the
|
|
TLAS build.
|
|
|
|
```` C
|
|
// OBJ Primary Ray Hit Group - Closest Hit + AnyHit (not used)
|
|
vk::ShaderModule chitSM =
|
|
nvvk::createShaderModule(m_device, //
|
|
nvh::loadFile("shaders/raytrace.rchit.spv", true, paths, true));
|
|
|
|
vk::RayTracingShaderGroupCreateInfoKHR hg{vk::RayTracingShaderGroupTypeKHR::eTrianglesHitGroup,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR};
|
|
stages.push_back({{}, vk::ShaderStageFlagBits::eClosestHitKHR, chitSM, "main"});
|
|
hg.setClosestHitShader(static_cast<uint32_t>(stages.size() - 1));
|
|
m_rtShaderGroups.push_back(hg);
|
|
|
|
// Lantern Primary Ray Hit Group
|
|
vk::ShaderModule lanternChitSM =
|
|
nvvk::createShaderModule(m_device, //
|
|
nvh::loadFile("shaders/lantern.rchit.spv", true, paths, true));
|
|
|
|
vk::RayTracingShaderGroupCreateInfoKHR lanternHg{
|
|
vk::RayTracingShaderGroupTypeKHR::eTrianglesHitGroup,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR};
|
|
stages.push_back({{}, vk::ShaderStageFlagBits::eClosestHitKHR, lanternChitSM, "main"});
|
|
lanternHg.setClosestHitShader(static_cast<uint32_t>(stages.size() - 1));
|
|
m_rtShaderGroups.push_back(lanternHg);
|
|
|
|
// ...
|
|
|
|
m_device.destroy(lanternChitSM);
|
|
````
|
|
|
|
We don't have to modify `HelloVulkan::createRtShaderBindingTable`. Changes to the number of
|
|
group handles to copy into the SBT are picked up automatically from `m_rtShaderGroups.size()`.
|
|
|
|
# Ray Generation Shader
|
|
|
|
## Draw Within Scissor Rectangle
|
|
|
|
The original ray generation shader assumed that `gl_LaunchSizeEXT` is the size of the entire
|
|
screen. As this is no longer the case for scissor rectangles, we communicate the screen
|
|
size through push constant instead. In addition, we also add to the push constants a number
|
|
indicating which lantern pass is currently being drawn (-1 for the original full screen pass).
|
|
|
|
Modify `m_rtPushConstants` in `hello_vulkan.h`.
|
|
|
|
```` C
|
|
// Push constant for ray trace pipeline.
|
|
struct RtPushConstant
|
|
{
|
|
// Background color
|
|
nvmath::vec4f clearColor;
|
|
|
|
// Information on the light in the sky used when lanternPassNumber = -1.
|
|
nvmath::vec3f lightPosition;
|
|
float lightIntensity;
|
|
int32_t lightType;
|
|
|
|
// -1 if this is the full-screen pass. Otherwise, this pass is to add light
|
|
// from lantern number lanternPassNumber. We use this to lookup trace indirect
|
|
// parameters in m_lanternIndirectBuffer.
|
|
int32_t lanternPassNumber;
|
|
|
|
// Pixel dimensions of the output image.
|
|
int32_t screenX;
|
|
int32_t screenY;
|
|
|
|
// See m_lanternDebug.
|
|
int32_t lanternDebug;
|
|
} m_rtPushConstants;
|
|
````
|
|
|
|
We also update the GLSL push constant to match. Since the raygen shader now needs
|
|
access to the push constant, move the push constant definition from `raytrace.rchit`
|
|
to `raycommon.glsl`.
|
|
|
|
```` C
|
|
layout(push_constant) uniform Constants
|
|
{
|
|
vec4 clearColor;
|
|
vec3 lightPosition;
|
|
float lightIntensity;
|
|
int lightType; // 0: point, 1: infinite
|
|
int lanternPassNumber; // -1 if this is the full-screen pass. Otherwise, used to lookup trace indirect parameters.
|
|
int screenX;
|
|
int screenY;
|
|
int lanternDebug;
|
|
}
|
|
pushC;
|
|
````
|
|
|
|
(`lanternDebug` will be used later to toggle visualising the scissor rectangles)
|
|
|
|
|
|
This move also requires us to tweak `raytrace.rmiss`.
|
|
|
|
````
|
|
#version 460
|
|
#extension GL_EXT_ray_tracing : require
|
|
#extension GL_GOOGLE_include_directive : enable
|
|
#include "raycommon.glsl"
|
|
|
|
layout(location = 0) rayPayloadInEXT hitPayload prd;
|
|
|
|
void main()
|
|
{
|
|
prd.hitValue = pushC.clearColor.xyz * 0.8;
|
|
prd.additiveBlending = false;
|
|
}
|
|
````
|
|
|
|
We will cover initializing the new push constants later, when we look at `vkCmdTraceRaysIndirectKHR`.
|
|
|
|
In `raytrace.rgen`, we have to replace the old code for calculating the pixel center.
|
|
|
|
```` C
|
|
void main()
|
|
{
|
|
const vec2 pixelCenter = vec2(gl_LaunchIDEXT.xy) + vec2(0.5);
|
|
const vec2 inUV = pixelCenter / vec2(gl_LaunchSizeEXT.xy);
|
|
````
|
|
|
|
with
|
|
|
|
```` C
|
|
layout(binding = 2, set = 0) buffer LanternArray { LanternIndirectEntry lanterns[]; } lanterns;
|
|
|
|
void main()
|
|
{
|
|
// Global light pass is a full screen rectangle (lower corner 0,0), but
|
|
// lantern passes are only run within rectangles that may be offset.
|
|
ivec2 pixelOffset = ivec2(0);
|
|
if (pushC.lanternPassNumber >= 0)
|
|
{
|
|
pixelOffset.x = lanterns.lanterns[pushC.lanternPassNumber].offsetX;
|
|
pixelOffset.y = lanterns.lanterns[pushC.lanternPassNumber].offsetY;
|
|
}
|
|
|
|
const ivec2 pixelIntCoord = ivec2(gl_LaunchIDEXT.xy) + pixelOffset;
|
|
const vec2 pixelCenter = vec2(pixelIntCoord) + vec2(0.5);
|
|
const vec2 inUV = pixelCenter / vec2(pushC.screenX, pushC.screenY);
|
|
vec2 d = inUV * 2.0 - 1.0;
|
|
````
|
|
|
|
Let's recap why this works. If `pushC.lanternPassNumber` is negative, we're drawing
|
|
the first, full-screen pass, and this code behaves identically as before, except
|
|
that `inUV` performs division by `(pushC.screenX, pushC.screenY)` instead of
|
|
relying on `gl_LaunchSizeEXT` to be the screen size.
|
|
|
|
Otherwise (`pushC.lanternPassNumber >= 0`), we're drawing a scissor rectangle for
|
|
the given lantern number. Look up that lantern's `LanternIndirectEntry` in the
|
|
array (notice that the descriptor binding for it is added). Its scissor rectangle
|
|
is defined by:
|
|
|
|
* `LanternIndirectEntry::offsetX`,`offsetY`: the pixel coordinate of the scissor box's
|
|
upper-left.
|
|
|
|
* `LanternIndirectEntry::width`,`height`: the dimensions of the scissor box (not
|
|
directly used here; consumed by `vkCmdTraceRaysIndirectKHR`).
|
|
|
|
The `gl_LaunchIDEXT` variable ranges from `(0,0)` to `(width-1, height-1)`, so to
|
|
cover the correct pixels within the scissor, we just have to reposition
|
|
`gl_LaunchIDEXT` by the offset `(offsetX, offsetY)`.
|
|
|
|
## Additive Blending
|
|
|
|
We also have to emulate additive blending. Instead of always writing to the output
|
|
image:
|
|
|
|
```` C
|
|
imageStore(image, ivec2(gl_LaunchIDEXT.xy), vec4(prd.hitValue, 1.0));
|
|
````
|
|
|
|
we do
|
|
|
|
```` C
|
|
// Either add to or replace output image color based on prd.additiveBlending.
|
|
// Global pass always replaces color as it is the first pass.
|
|
vec3 oldColor = vec3(0);
|
|
if (prd.additiveBlending && pushC.lanternPassNumber >= 0) {
|
|
oldColor = imageLoad(image, pixelIntCoord).rgb;
|
|
}
|
|
imageStore(image, pixelIntCoord, vec4(prd.hitValue + oldColor, 1.0));
|
|
````
|
|
|
|
thus adding the ray payload's color to the old image color if `prd.additiveBlending`
|
|
is true and this is not the first, full-screen pass (the first pass must replace the
|
|
output image color as its existing contents are garbage).
|
|
|
|
# Lantern Shadow Rays
|
|
|
|
We now have to set up a system for casting shadow rays from the OBJ closest hit
|
|
shader to the lanterns. This requires us to
|
|
|
|
* Detect in `raycast.rchit` whether we are in a lantern pass, and use this
|
|
to decide between casting shadow rays to the main light (as in the base
|
|
tutorial) or casting shadow rays to a lantern.
|
|
|
|
* Declare a payload for which lantern (if any) was hit, and add a new miss shader
|
|
and two new closest hit shaders for filling that payload.
|
|
|
|
* Use the `sbtRecordOffset` parameter of `traceRayEXT` to skip over the earlier
|
|
hit groups.
|
|
|
|
## New payload
|
|
|
|
In `raytrace.rchit` (called when an OBJ instance is hit by a primary ray), declare
|
|
the new payload and the array of lanterns.
|
|
|
|
```` C
|
|
layout(location = 2) rayPayloadEXT int hitLanternInstance;
|
|
|
|
layout(binding = 0, set = 0) uniform accelerationStructureEXT topLevelAS;
|
|
layout(binding = 2, set = 0) buffer LanternArray { LanternIndirectEntry lanterns[]; } lanterns;
|
|
````
|
|
|
|
## New shaders
|
|
|
|
We need a few simple shaders to report the number of the lantern hit (if any) by the shadow ray.
|
|
First is the miss shader, `lanternShadow.rmiss`.
|
|
|
|
```` C
|
|
// Miss shader invoked when tracing shadow rays (rays towards lantern)
|
|
// in lantern passes. Misses shouldn't really happen, but if they do,
|
|
// report we did not hit any lantern by setting hitLanternInstance = -1.
|
|
layout(location = 2) rayPayloadInEXT int hitLanternInstance;
|
|
|
|
void main()
|
|
{
|
|
hitLanternInstance = -1;
|
|
}
|
|
````
|
|
|
|
Then a closest hit shader for OBJ instances hit by a lantern shadow ray.
|
|
This also returns `-1` for "no lantern". Call this `lanternShadowObj.rchit`.
|
|
|
|
```` C
|
|
#version 460
|
|
#extension GL_EXT_ray_tracing : require
|
|
#extension GL_GOOGLE_include_directive : enable
|
|
#include "raycommon.glsl"
|
|
|
|
// During a lantern pass, this closest hit shader is invoked when
|
|
// shadow rays (rays towards lantern) hit a regular OBJ. Report back
|
|
// that no lantern was hit (-1).
|
|
|
|
// clang-format off
|
|
layout(location = 2) rayPayloadInEXT int hitLanternInstance;
|
|
|
|
// clang-format on
|
|
|
|
void main()
|
|
{
|
|
hitLanternInstance = -1;
|
|
}
|
|
````
|
|
|
|
Finally, a closest hit shader for lantern instances, named `lanternShadowLantern.rchit`.
|
|
|
|
```` C
|
|
#version 460
|
|
#extension GL_EXT_ray_tracing : require
|
|
#extension GL_GOOGLE_include_directive : enable
|
|
#include "raycommon.glsl"
|
|
|
|
// During a lantern pass, this closest hit shader is invoked when
|
|
// shadow rays (rays towards lantern) hit a lantern. Report back
|
|
// which lantern was hit.
|
|
|
|
// clang-format off
|
|
layout(location = 2) rayPayloadInEXT int hitLanternInstance;
|
|
|
|
// clang-format on
|
|
|
|
void main()
|
|
{
|
|
hitLanternInstance = gl_InstanceCustomIndexEXT;
|
|
}
|
|
````
|
|
|
|
Note that we really need to report back the lantern number, and
|
|
not just a boolean "lantern hit" flag. In order to have lanterns cast
|
|
shadows on each other, we must be able to detect that the shadow ray
|
|
hit the "wrong" lantern.
|
|
|
|
|
|
|
|
## Add Shaders to Pipeline
|
|
|
|
We add the new miss shader as miss shader 2 in the SBT, and the closest hit
|
|
shaders as hit groups 2 and 3 in the SBT, following the earlier 2 hit
|
|
groups for primary rays. Add the following code to `HelloVulkan::createRtPipeline`
|
|
after loading `raytraceShadow.rmiss.spv`.
|
|
|
|
```` C
|
|
// Miss shader 2 is invoked when a shadow ray for lantern lighting misses the
|
|
// lantern. It shouldn't be invoked, but I include it just in case.
|
|
vk::ShaderModule lanternmissSM = nvvk::createShaderModule(
|
|
m_device, nvh::loadFile("shaders/lanternShadow.rmiss.spv", true, paths, true));
|
|
````
|
|
|
|
and add this code for loading the last 2 closest hit shaders after loading
|
|
`lantern.rchit.spv`:
|
|
|
|
```` C
|
|
// OBJ Lantern Shadow Ray Hit Group
|
|
vk::ShaderModule lanternShadowObjChitSM =
|
|
nvvk::createShaderModule(m_device, //
|
|
nvh::loadFile("shaders/lanternShadowObj.rchit.spv", true, paths, true));
|
|
|
|
vk::RayTracingShaderGroupCreateInfoKHR lanternShadowObjHg{
|
|
vk::RayTracingShaderGroupTypeKHR::eTrianglesHitGroup,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR};
|
|
stages.push_back({{}, vk::ShaderStageFlagBits::eClosestHitKHR, lanternShadowObjChitSM, "main"});
|
|
lanternShadowObjHg.setClosestHitShader(static_cast<uint32_t>(stages.size() - 1));
|
|
m_rtShaderGroups.push_back(lanternShadowObjHg);
|
|
|
|
// Lantern Lantern Shadow Ray Hit Group
|
|
vk::ShaderModule lanternShadowLanternChitSM =
|
|
nvvk::createShaderModule(m_device, //
|
|
nvh::loadFile("shaders/lanternShadowLantern.rchit.spv", true, paths, true));
|
|
|
|
vk::RayTracingShaderGroupCreateInfoKHR lanternShadowLanternHg{
|
|
vk::RayTracingShaderGroupTypeKHR::eTrianglesHitGroup,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR,
|
|
VK_SHADER_UNUSED_KHR, VK_SHADER_UNUSED_KHR};
|
|
stages.push_back({{}, vk::ShaderStageFlagBits::eClosestHitKHR, lanternShadowLanternChitSM, "main"});
|
|
lanternShadowLanternHg.setClosestHitShader(static_cast<uint32_t>(stages.size() - 1));
|
|
m_rtShaderGroups.push_back(lanternShadowLanternHg);
|
|
````
|
|
|
|
We need to destroy the added shader modules at the end of the function.
|
|
|
|
```` C
|
|
m_device.destroy(shadowmissSM);
|
|
// ...
|
|
m_device.destroy(lanternShadowObjChitSM);
|
|
m_device.destroy(lanternShadowLanternChitSM);
|
|
````
|
|
|
|
Through all this, we still load shader stages in the same order as they will appear
|
|
in the SBT in order to keep things simple (note `stages.size() - 1`). Add a comment
|
|
at the top of this function to help us keep track of all the new shaders.
|
|
|
|
```` C
|
|
// Shader list:
|
|
//
|
|
// 0 ====== Ray Generation Shaders =====================================================
|
|
//
|
|
// Raygen shader: Ray generation shader. Casts primary rays from camera to scene.
|
|
//
|
|
// 1 ====== Miss Shaders ===============================================================
|
|
//
|
|
// Miss shader 0: Miss shader when casting primary rays. Fill in clear color.
|
|
//
|
|
// 2 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
|
|
//
|
|
// Miss shader 1: Miss shader when casting shadow rays towards main light.
|
|
// Reports no shadow.
|
|
//
|
|
// 3 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
|
|
//
|
|
// Miss shader 2: Miss shader when casting shadow rays towards a lantern.
|
|
// Reports no lantern hit (-1).
|
|
//
|
|
// 4 ====== Hit Groups for Primary Rays (sbtRecordOffset=0) ============================
|
|
//
|
|
// chit shader 0: Closest hit shader for primary rays hitting OBJ instances
|
|
// (hitGroupId=0). Casts shadow ray (to sky light or to lantern,
|
|
// depending on pass number) and returns specular
|
|
// and diffuse light to add to output image.
|
|
//
|
|
// 5 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
|
|
//
|
|
// chit shader 1: Closest hit shader for primary rays hitting lantern instances
|
|
// (hitGroupId=1). Returns color value to replace the current
|
|
// image pixel color with (lanterns are self-illuminating).
|
|
//
|
|
// 6 - - - - Hit Groups for Lantern Shadow Rays (sbtRecordOffset=2) - - - - - - - - - - -
|
|
//
|
|
// chit shader 2: Closest hit shader for OBJ instances hit when casting shadow
|
|
// rays to a lantern. Returns -1 to report that the shadow ray
|
|
// failed to reach the targetted lantern.
|
|
//
|
|
// 7 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
|
|
//
|
|
// chit shader 3: Closest hit shader for lantern instances hit when casting shadow
|
|
// rays to a lantern. Returns the gl_CustomInstanceIndexEXT [lantern
|
|
// number] of the lantern hit.
|
|
//
|
|
// 8 =====================================================================================
|
|
````
|
|
|
|
## Compute Lighting Intensity
|
|
|
|
Because the lanterns have color, we have to replace the scalar `lightIntensity`
|
|
in `raytrace.rchit` with an RGB `colorIntensity`.
|
|
|
|
```` C
|
|
// Vector toward the light
|
|
vec3 L;
|
|
vec3 colorIntensity = vec3(pushC.lightIntensity);
|
|
float lightDistance = 100000.0;
|
|
````
|
|
|
|
Then, we have to check if we're in a lantern pass (`lanternPassNumber >= 0`).
|
|
If so, look up the lantern location in the `LanternIndirectEntry` array,
|
|
and compute the light direction and intensity based on that position.
|
|
|
|
```` C
|
|
// ray direction is towards lantern, if in lantern pass.
|
|
if (pushC.lanternPassNumber >= 0)
|
|
{
|
|
LanternIndirectEntry lantern = lanterns.lanterns[pushC.lanternPassNumber];
|
|
vec3 lDir = vec3(lantern.x, lantern.y, lantern.z) - worldPos;
|
|
lightDistance = length(lDir);
|
|
vec3 color = vec3(lantern.red, lantern.green, lantern.blue);
|
|
// Lantern light decreases linearly. Not physically accurate, but looks good
|
|
// and avoids a hard "edge" at the radius limit. Use a constant value
|
|
// if lantern debug is enabled to clearly see the covered screen rectangle.
|
|
float distanceFade =
|
|
pushC.lanternDebug != 0
|
|
? 0.3
|
|
: max(0, (lantern.radius - lightDistance) / lantern.radius);
|
|
colorIntensity = color * lantern.brightness * distanceFade;
|
|
L = normalize(lDir);
|
|
}
|
|
````
|
|
|
|
otherwise, do the old lighting calculations, except we again have to
|
|
replace `float lightIntensity` with `vec3 colorIntensity`.
|
|
|
|
```` C
|
|
// Non-lantern pass may have point light...
|
|
else if(pushC.lightType == 0)
|
|
{
|
|
vec3 lDir = pushC.lightPosition - worldPos;
|
|
lightDistance = length(lDir);
|
|
colorIntensity = vec3(pushC.lightIntensity / (lightDistance * lightDistance));
|
|
L = normalize(lDir);
|
|
}
|
|
else // or directional light.
|
|
{
|
|
L = normalize(pushC.lightPosition - vec3(0));
|
|
}
|
|
````
|
|
|
|
!!! NOTE `lanternDebug`
|
|
When `lanternDebug` is on, I disable diminishing lighting with distance, so
|
|
that the light will reach the edge of the scissor box, making the scissor
|
|
box easy to see. To toggle this variable, I declare `bool m_lanternDebug`
|
|
in `hello_vulkan.h`, and allow ImGui to control it:
|
|
|
|
```` C
|
|
void renderUI(HelloVulkan& helloVk)
|
|
{
|
|
ImGuiH::CameraWidget();
|
|
if(ImGui::CollapsingHeader("Light"))
|
|
{
|
|
// ...
|
|
ImGui::Checkbox("Lantern Debug", &helloVk.m_lanternDebug);
|
|
}
|
|
}
|
|
````
|
|
|
|
Then, every frame I copy `m_lanternDebug` to the push constant. The reason
|
|
I cannot directly modify the push constant through ImGui is that ImGui expects
|
|
a `bool` (usually 8-bits) while Vulkan expects a 32-bit boolean.
|
|
|
|
## Casting Lantern Shadow Rays
|
|
|
|
Use an `if` to ensure the original shadow rays are cast only in the non-lantern pass.
|
|
|
|
```` C
|
|
// Ordinary shadow from the simple tutorial.
|
|
if (pushC.lanternPassNumber < 0) {
|
|
isShadowed = true;
|
|
uint flags = gl_RayFlagsTerminateOnFirstHitEXT | gl_RayFlagsOpaqueEXT
|
|
| gl_RayFlagsSkipClosestHitShaderEXT;
|
|
traceRayEXT(topLevelAS, // acceleration structure
|
|
flags, // rayFlags
|
|
0xFF, // cullMask
|
|
0, // sbtRecordOffset
|
|
0, // sbtRecordStride
|
|
1, // missIndex
|
|
origin, // ray origin
|
|
tMin, // ray min range
|
|
rayDir, // ray direction
|
|
tMax, // ray max range
|
|
1 // payload (location = 1)
|
|
);
|
|
}
|
|
````
|
|
|
|
Otherwise, we cast a ray towards a lantern. This ray is different in that
|
|
|
|
* We actually need the closest hit shaders to run to return `hitLanternInstance`,
|
|
so do not provide the flags
|
|
` gl_RayFlagsTerminateOnFirstHitEXT | gl_RayFlagsSkipClosestHitShaderEXT`.
|
|
|
|
* Use miss shader 2, which we added earlier.
|
|
|
|
* Pass 2 as `sbtRecordOffset`, so that the closest hit shaders we just added (number 2 and 3)
|
|
are used when hitting OBJ instances (`hitGroupId=0`) and lantern instances (`hitGroupId=1`)
|
|
respectively.
|
|
|
|
The code is
|
|
|
|
```` C
|
|
// Lantern shadow ray. Cast a ray towards the lantern whose lighting is being
|
|
// added this pass. Only the closest hit shader for lanterns will set
|
|
// hitLanternInstance (payload 2) to non-negative value.
|
|
else {
|
|
// Skip ray if no light would be added anyway.
|
|
if (colorIntensity == vec3(0)) {
|
|
isShadowed = true;
|
|
}
|
|
else {
|
|
uint flags = gl_RayFlagsOpaqueEXT;
|
|
hitLanternInstance = -1;
|
|
traceRayEXT(topLevelAS, // acceleration structure
|
|
flags, // rayFlags
|
|
0xFF, // cullMask
|
|
2, // sbtRecordOffset : lantern shadow hit groups start at index 2.
|
|
0, // sbtRecordStride
|
|
2, // missIndex : lantern shadow miss shader is number 2.
|
|
origin, // ray origin
|
|
tMin, // ray min range
|
|
rayDir, // ray direction
|
|
tMax, // ray max range
|
|
2 // payload (location = 2)
|
|
);
|
|
// Did we hit the lantern we expected?
|
|
isShadowed = (hitLanternInstance != pushC.lanternPassNumber);
|
|
}
|
|
}
|
|
````
|
|
|
|
Notice that we determine whether this lantern is shadowed at this pixel by
|
|
checking if the hit lantern number matches the lantern whose light is being
|
|
added this pass; again, this ensures lanterns correctly shadow each others' light.
|
|
|
|
## Write Payload
|
|
|
|
Replace `lightIntensity` with `colorIntensity` and ask the raygen shader
|
|
for additive blending.
|
|
|
|
```` C
|
|
prd.hitValue = colorIntensity * (attenuation * (diffuse + specular));
|
|
prd.additiveBlending = true;
|
|
````
|
|
|
|
# Trace Rays Indirect
|
|
|
|
Everything is finally set up to actually run the extra lantern passes
|
|
in `HelloVulkan::raytrace`. We've already dispatched the compute
|
|
shader in an earlier section. After that, we can run the first raytrace
|
|
pass. There are minimal changes from before, we just have to
|
|
|
|
* Initialize the new push constant values (especially setting
|
|
`lanternPassNumber=-1` to indicate this is not a lantern pass).
|
|
|
|
```` C
|
|
// Initialize push constant values
|
|
m_rtPushConstants.clearColor = clearColor;
|
|
m_rtPushConstants.lightPosition = m_pushConstant.lightPosition;
|
|
m_rtPushConstants.lightIntensity = m_pushConstant.lightIntensity;
|
|
m_rtPushConstants.lightType = m_pushConstant.lightType;
|
|
m_rtPushConstants.lanternPassNumber = -1; // Global non-lantern pass
|
|
m_rtPushConstants.screenX = m_size.width;
|
|
m_rtPushConstants.screenY = m_size.height;
|
|
m_rtPushConstants.lanternDebug = m_lanternDebug;
|
|
````
|
|
|
|
* Update the addresses of the raygen, miss, and hit group sections of the SBT
|
|
to account for the added shaders.
|
|
|
|
```` C
|
|
using Stride = vk::StridedDeviceAddressRegionKHR;
|
|
std::array<Stride, 4> strideAddresses{
|
|
Stride{sbtAddress + 0u * groupSize, groupStride, groupSize * 1}, // raygen
|
|
Stride{sbtAddress + 1u * groupSize, groupStride, groupSize * 3}, // miss
|
|
Stride{sbtAddress + 4u * groupSize, groupStride, groupSize * 4}, // hit
|
|
Stride{0u, 0u, 0u}}; // callable
|
|
|
|
// First pass, illuminate scene with global light.
|
|
cmdBuf.traceRaysKHR(
|
|
&strideAddresses[0], &strideAddresses[1], //
|
|
&strideAddresses[2], &strideAddresses[3], //
|
|
m_size.width, m_size.height, 1);
|
|
````
|
|
|
|
After that, we can open a loop for performing all lantern passes.
|
|
|
|
```` C
|
|
// Lantern passes, ensure previous pass completed, then add light contribution from each lantern.
|
|
for (int i = 0; i < static_cast<int>(m_lanternCount); ++i)
|
|
{
|
|
````
|
|
|
|
Because the additive blending in the shader requires read-modify-write operations,
|
|
we need a barrier between every pass.
|
|
|
|
```` C
|
|
// Barrier to ensure previous pass finished.
|
|
vk::Image offscreenImage{m_offscreenColor.image};
|
|
vk::ImageSubresourceRange colorRange(
|
|
vk::ImageAspectFlagBits::eColor, 0, VK_REMAINING_MIP_LEVELS, 0, VK_REMAINING_ARRAY_LAYERS
|
|
);
|
|
vk::ImageMemoryBarrier imageBarrier;
|
|
imageBarrier.setOldLayout(vk::ImageLayout::eGeneral);
|
|
imageBarrier.setNewLayout(vk::ImageLayout::eGeneral);
|
|
imageBarrier.setSrcQueueFamilyIndex(VK_QUEUE_FAMILY_IGNORED);
|
|
imageBarrier.setDstQueueFamilyIndex(VK_QUEUE_FAMILY_IGNORED);
|
|
imageBarrier.setImage(offscreenImage);
|
|
imageBarrier.setSubresourceRange(colorRange);
|
|
imageBarrier.setSrcAccessMask(vk::AccessFlagBits::eShaderWrite);
|
|
imageBarrier.setDstAccessMask(vk::AccessFlagBits::eShaderRead);
|
|
cmdBuf.pipelineBarrier(
|
|
vk::PipelineStageFlagBits::eRayTracingShaderKHR, //
|
|
vk::PipelineStageFlagBits::eRayTracingShaderKHR, //
|
|
vk::DependencyFlags(0), //
|
|
{}, {}, {imageBarrier});
|
|
````
|
|
|
|
Then, we can pass the number of the lantern pass being performed (`i`), and look
|
|
up the indirect parameters for that entry. Unlike draw and dispatch indirect, the
|
|
indirect paramater location is passed as a raw device address, so we need to
|
|
perform manual device pointer arithmetic to look up the $i^{th}$ entry of the
|
|
`LanternIndirectEntry` array. Take advantage of the fact that `VkTraceRaysIndirectCommandKHR`
|
|
is the first member of `LanternIndirectEntry`.
|
|
|
|
```` C
|
|
// Set lantern pass number.
|
|
m_rtPushConstants.lanternPassNumber = i;
|
|
cmdBuf.pushConstants<RtPushConstant>(m_rtPipelineLayout,
|
|
vk::ShaderStageFlagBits::eRaygenKHR
|
|
| vk::ShaderStageFlagBits::eClosestHitKHR
|
|
| vk::ShaderStageFlagBits::eMissKHR,
|
|
0, m_rtPushConstants);
|
|
|
|
// Execute lantern pass.
|
|
cmdBuf.traceRaysIndirectKHR(
|
|
&strideAddresses[0], &strideAddresses[1], //
|
|
&strideAddresses[2], &strideAddresses[3], //
|
|
m_device.getBufferAddress({m_lanternIndirectBuffer.buffer}) + i * sizeof(LanternIndirectEntry));
|
|
}
|
|
````
|
|
|
|
Everything should be in order now. We can see in this image that the cyan and purple lanterns
|
|
are both shadowed by the doodad hanging off the side of the building, and the spikes on the
|
|
roof cut shadows in the yellow lantern's light.
|
|
|
|
|
|
|
|
Zoom out and enable the lantern debug checkbox to see the scissor rectangles.
|
|
|
|
|
|
|
|
## Cleanup
|
|
|
|
One last loose end, we have to clean up all the new resources is `HelloVulkan::destroyResources`.
|
|
|
|
```` C
|
|
// Destroying all allocations
|
|
//
|
|
void HelloVulkan::destroyResources()
|
|
{
|
|
// ...
|
|
|
|
// #VKRay
|
|
// ...
|
|
m_device.destroy(m_lanternIndirectDescPool);
|
|
m_device.destroy(m_lanternIndirectDescSetLayout);
|
|
m_device.destroy(m_lanternIndirectCompPipeline);
|
|
m_device.destroy(m_lanternIndirectCompPipelineLayout);
|
|
m_alloc.destroy(m_lanternIndirectBuffer);
|
|
m_alloc.destroy(m_lanternVertexBuffer);
|
|
m_alloc.destroy(m_lanternIndexBuffer);
|
|
}
|
|
````
|
|
|
|
# Final Code
|
|
|
|
You can find the final code in the folder [ray_tracing_indirect_scissor](https://github.com/nvpro-samples/vk_raytracing_tutorial_KHR/tree/master/ray_tracing_indirect_scissor)
|
|
|
|
|
|
<!-- Markdeep: -->
|
|
<link rel="stylesheet" href="vkrt_tutorial.css?">
|
|
<script> window.markdeepOptions = { tocStyle: "medium" };</script>
|
|
<script src="markdeep.min.js" charset="utf-8"></script>
|
|
<script src="https://developer.nvidia.com/sites/default/files/akamai/gameworks/whitepapers/markdeep.min.js" charset="utf-8"></script>
|
|
<script>
|
|
window.alreadyProcessedMarkdeep || (document.body.style.visibility = "visible")
|
|
</script>
|