Performance Optimisations¶

This page documents every significant optimisation that was applied to RayON, roughly in the order they were introduced, together with the measured or estimated impact of each one. The cumulative effect is a ~1 060× speedup over the single-threaded CPU baseline for a typical 720p scene with 1 024 SPP.

1 — CPU multi-threading¶

What it is: tile-based work dispatch using std::async. The image is divided into blocks; each block is submitted as an independent std::future and picked up by a thread-pool of N−2 hardware threads.

Why it helps: path tracing is embarrassingly parallel — each pixel is independent. Saturating all cores gives a near-linear speedup.

Measured impact: ~15× on a 16-core machine.

// renderer_cpu_parallel.hpp — simplified
for (int ty = 0; ty < num_tiles_y; ++ty) {
    for (int tx = 0; tx < num_tiles_x; ++tx) {
        futures.push_back(std::async(std::launch::async, [=] {
            renderTile(tx, ty, tile_w, tile_h);
        }));
    }
}
for (auto& f : futures) f.get();

2 — CUDA GPU rendering¶

What it is: a CUDA kernel that assigns one thread per pixel. Each thread independently traces rays, with no communication between threads.

Why it helps: a modern GPU has thousands of streaming processors. The GPU also hides memory latency through warp switching — while one warp waits for a memory transaction, another warp executes.

Measured impact: ~400× vs. single-threaded CPU on a 720p scene (without BVH).

3 — Thread-block shape: 32 × 8¶

What it is: the kernel is launched with 2-D thread blocks of 32 columns × 8 rows = 256 threads (16 × 16 on devices that allow fewer threads per block).

Why it helps: 32 threads exactly fill one CUDA warp — the unit of SIMD execution. With 32 threads per row, adjacent threads access adjacent pixel addresses, which coalesces into a single memory transaction. A 16×8 block would split one row across two warps and introduce cross-warp divergence on conditional branches. Register-cap and block-shape experiments (--maxrregcount 48–80, 32×8 / 16×16 / 8×32) gave no consistent gain across resolutions, so 32×8 with 80 registers stayed the default.

Measured impact: ~5–10% throughput gain over 16×8 at the same occupancy.

dim3 block_size(32, 8);   // 256 threads, one warp per row
dim3 grid_size(
    (width  + block_size.x - 1) / block_size.x,
    (height + block_size.y - 1) / block_size.y
);
renderPixelsKernel<<<grid_size, block_size>>>(...);

4 — Cosine-weighted hemisphere sampling¶

What it is: diffuse surfaces scatter rays with probability proportional to \(\cos\theta / \pi\) rather than \(1 / 2\pi\) (uniform).

Why it helps: the Monte Carlo integrand for a Lambertian surface contains a \(\cos\theta\) factor. When the PDF matches that factor, the weights become constant:

Every sample contributes equally — variance drops dramatically near shadow boundaries and at grazing angles.

Measured impact: 4–8× fewer samples needed for equivalent convergence vs. uniform hemisphere sampling (scene-dependent).

5 — Russian roulette path termination¶

What it is: after each bounce, a path is terminated randomly with probability proportional to \(1 - \max(\text{throughput})\). Surviving paths are compensated to maintain an unbiased estimate.

Why it helps: carrying a path to MAX_DEPTH bounces even when throughput is near zero (e.g. after five rough-mirror reflections) wastes GPU cycles. Russian roulette cuts those paths early while redistributing the saved compute to paths that still carry energy.

Measured impact: ~15–20% throughput improvement on typical scenes with MAX_DEPTH=16.

// After each bounce in the GPU kernel
float p = fmaxf(throughput.x, fmaxf(throughput.y, throughput.z));
if (curand_uniform(&rng) > p) break;  // terminate — unbiased
throughput /= p;                       // compensate

6 — Persistent curand RNG states¶

What it is: one curandState per pixel is allocated in device memory at startup and reused across all frames in progressive mode.

Why it helps: calling curand_init() is expensive (~50 ns per thread). Re-initialising 720p = 921 600 states every frame would cost ~46 ms — longer than the render itself at 60 fps. Persistent states also avoid repeating the same random sequence, which would cause visible banding in accumulated renders.

Measured impact: eliminates a ~46 ms per-frame overhead in interactive mode at 720p.

7 — Accumulation on GPU + uint8 D2H transfer¶

What it is: sample results are accumulated in a float buffer that stays on the GPU. After each batch, a lightweight gamma-correction kernel converts the float sums to uint8. Only the 3-byte-per-pixel uint8 result is copied host ← device.

Why it helps: the previous design transferred the full 3×float accumulation buffer each frame (12 bytes/pixel). The new design reduces D2H bandwidth by 4×. At 1920×1080, that drops the PCIe transfer from ~24 MB/frame to ~6 MB/frame — comfortably below 60 fps budget.

// GPU kernel: accumulate in float4 (one atomic per pixel)
atomicAdd(&d_accum[pixel_idx * 3 + 0], pixel_color.x);

// GPU gamma kernel: convert + pack (no CPU involvement)
display[idx*3+0] = (uint8_t)(clamp(sqrtf(accum[idx*3+0]/spp), 0.f, 1.f) * 255.f);

8 — BVH acceleration (SAH)¶

What it is: a Bounding Volume Hierarchy built on the CPU with Surface Area Heuristic (SAH) splitting. The flat node array is uploaded to the GPU once and traversed iteratively by every kernel thread.

Why it helps: without BVH, every ray tests all \(N\) objects — \(O(N)\). With a SAH-BVH of depth \(\log_2 N\), average traversal cost drops to \(O(\log N)\) with tight bounds.

Measured impact: up to 14.6× on a 300-object scene. See Performance for the full table.

Key implementation details:

• 8 split candidates per axis (3 axes × 8 = 24 candidates per node)
• 64-byte aligned BVHNode — one complete node fits in one L2 cache line
• Iterative stack traversal on the GPU (depth-32 local stack, no recursion)
• Nearer child pushed last — the near child is at the top of the stack and processed first, allowing the "farther than current best" early-out to skip more nodes

struct alignas(64) BVHNode {
    float3 aabb_min, aabb_max;  // 24 bytes
    int    left_child;           //  4 bytes
    int    right_child;          //  4 bytes
    int    prim_start;           //  4 bytes
    int    prim_count;           //  4 bytes  (> 0 → leaf)
    // padding to 64 bytes
};

9 — Inlined material dispatch¶

What it is: GPU material evaluation uses a switch statement over a MaterialType enum rather than virtual functions. The compiler inlines every case at build time.

Why it helps: virtual function calls on the GPU require two indirect memory accesses (vtable pointer + vtable entry) and break warp coherence when threads in the same warp hit different materials. An inlined switch eliminates both costs.

Measured impact: ~5–10% throughput improvement on mixed-material scenes.

10 — Adaptive sampling (converged-pixel skipping)¶

What it is: each pixel tracks a running sample count. After ≥ 32 accumulated samples, the renderer checks whether the relative change in luminance between the previous batch and the new batch is below a threshold (~10⁻⁴·⁵). Converged pixels are flagged (negative sample count) and skipped in all subsequent batches.

Why it helps: in most scenes, large uniform regions (sky, flat walls) converge quickly while complex areas (shadow boundaries, caustics) need many more samples. Skipping converged pixels redirects the GPU to the pixels that still need work.

A heatmap visualisation (purple = few samples, yellow = many) can be toggled in the ImGui panel to show where samples are being spent.

Measured impact: 20–50% effective speedup in mixed-complexity scenes; less useful in uniformly complex scenes.

Disable with --no-adaptive-sampling.

11 — Non-blocking CUDA stream + pinned-memory D2H pipeline¶

What it is: the display path (gamma-correction kernel + device→host copy) runs on a dedicated non-blocking CUDA stream (cudaStreamNonBlocking). The host memory target is page-locked (pinned), allocated with cudaMallocHost.

Why it helps: the old design used the default CUDA stream and cudaDeviceSynchronize(). cudaDeviceSynchronize() is a global barrier — it drains every outstanding GPU operation before returning. In interactive mode this means the next render batch cannot start until the display pipeline has completely finished, stalling the CPU and the GPU at the same time.

The new design:

1. Creates a separate, non-blocking stream (s_display_stream) for display work. Non-blocking means it will never implicitly synchronize with the default stream used by the render kernel.
2. Queues both the gamma-correction kernel and the cudaMemcpyAsync on that stream so the GPU processes them in order with no CPU involvement between the two.
3. Uses a pinned host buffer (cudaMallocHost) as the DMA target. Pinned memory has a fixed physical address the GPU's DMA engine can write to directly over PCIe without an extra kernel- initiated copy — DMA throughput is typically 2–4× higher than to pageable memory.
4. Synchronizes only the display stream (cudaStreamSynchronize(s_display_stream)) rather than every GPU activity.

// renderer_cuda_device.cu — display stream setup
cudaStreamCreateWithFlags(&s_display_stream, cudaStreamNonBlocking);
cudaMallocHost(&s_pinned_display, display_size);   // pinned staging buffer

// Per-frame display update — kernel + async DMA on same stream
cudaStream_t stream = s_display_stream;
gammaCorrectKernel<<<blocks, threads, 0, stream>>>(...);
cudaMemcpyAsync(s_pinned_display, d_display, display_size,
                cudaMemcpyDeviceToHost, stream);
cudaStreamSynchronize(stream);  // wait only for this stream
memcpy(display_image, s_pinned_display, display_size);  // fast pinned→pageable

Measured impact: removes the cudaDeviceSynchronize() bubble between consecutive render batches in interactive mode. On the DGX Spark the display path dropped from ~3 ms (blocked) to ~0.8 ms (async), allowing the render kernel to start sooner each frame.

12 — Adaptive depth¶

Removed. Earlier versions could start interactive rendering at 4 bounces and add one bounce per sample stage (--adaptive-depth). The motion regulator of the "Camera Motion" section replaced it: while the camera moves, the batch size and the render resolution are lowered to hold the target frame rate, and the full bounce depth is kept at all times, so the lighting never changes character between the moving and the still image.

19 — OptiX: hardware triangles, cheap shadow rays, slim payload¶

What it is: every triangle of the scene goes into one geometry acceleration structure with the built-in RT-core intersector; only spheres, rectangles, displaced spheres and SDF shapes remain custom primitives, referenced by a one-level instance acceleration structure. Shadow rays terminate on the first hit with a minimal payload, and the radiance payload was trimmed to what the closest-hit programs actually fill.

Measured impact: the mesh scenes (dragon, statue, bunny) run at hardware speed; shadow rays cost a fraction of a closest-hit query.

Files: optix/optix_renderer.cu (buildGAS), optix/optix_programs.cu

20 — Motion regulator: resolution before samples¶

What it is: while the camera moves, the interactive OptiX renderer holds a target frame rate by lowering the render fraction first (down to a quarter of the width and height, 1/16 of the pixels, upscaled for display) and then the samples per frame, with a cost model of the previous frames, and raises both back when there is headroom. The accumulation restarts on every move, so none of this touches the final image. An optional AI denoiser (the OptiX neural denoiser, guided by albedo and normal) smooths the motion frames and fades out over the first still frames.

Measured impact: 60 fps held on the cloud bunny during orbit on the DGX Spark, against a few frames per second at full resolution.

Files: renderer_optix_progressive_host.hpp, sdl_gui_handler.hpp (MotionQualitySettings)

21 — Volumes: baked density field and empty-space skipping¶

What it is: the procedural density of a cloud (distance to the mesh, shape noise, interior noise) is baked once into a 3D texture of 160 cells, re-baked in milliseconds when a slider moves, so the delta-tracking loop pays one texture fetch per step. The signed distance grid of the mesh bounds how far a point outside the cloud is from any density, and the tracking loop jumps that distance instead of taking null-collision steps.

Measured impact: the cloud bunny renders 512 spp at 720p in 11 s; the same transport on a density grid without the distance skip takes 40 s.

Files: volume_sdf.cu, volume_medium.cuh

GPU Implementation Techniques¶

The sections above cover algorithmic and system-level decisions. This section documents the lower-level CUDA and OptiX implementation details that address four recurring hardware bottlenecks:

13 — Precomputed inverse ray direction for BVH traversal¶

What it is: the slab-method AABB test computes 1/dir.x, 1/dir.y, 1/dir.z for every bounding-box test during BVH traversal. Since the ray direction is constant across the entire traversal, these three reciprocal divisions are redundant. The inverse is precomputed once per ray and passed as a parameter to hit_aabb().

// hit_scene() — computed once per ray
const f3 inv_dir(1.0f / r.dir.x, 1.0f / r.dir.y, 1.0f / r.dir.z);

// hit_aabb() — uses precomputed inverse, no divisions
__device__ __forceinline__ bool hit_aabb(
    const ray_simple &r, const f3 &inv_dir,
    const f3 &box_min, const f3 &box_max,
    float t_min, float t_max)
{
    float t0_x = (box_min.x - r.orig.x) * inv_dir.x;  // multiply, not divide
    // ...
}

Measured impact: eliminates 3 fdiv instructions per AABB test. For a BVH of depth 12 with 300+ objects, each ray saves ~36 divisions per bounce.

Files: cuda_raytracer.cuh — hit_aabb(), hit_scene()

14 — __launch_bounds__ on the path-tracing kernel¶

What it is: the __launch_bounds__(256) annotation tells the CUDA compiler that the path-tracing kernel is always launched with at most 256 threads per block (our 32 × 8 configuration). Without it the compiler must assume a generic thread count and may over-allocate registers or spill to slow local memory.

__global__ void __launch_bounds__(256)
renderAccKernel(float4 *accum_buffer, ...)
{
    // ... path tracing logic ...
}

Why 256? The kernel is register-heavy (ray state, hit records, material data, RNG state). With 256 threads per block, the compiler can allocate up to 256 registers per thread on modern GPUs without spilling — giving better occupancy than if it had to assume a higher thread count.

Measured impact: ~5–10% throughput improvement from better register allocation; avoids spills to slow local memory.

Files: shaders/render_acc_kernel.cu, shaders/render_acc_kernel.cuh

15 — GPU-side converged pixel counting (warp-shuffle reduction)¶

What it is: adaptive sampling tracks per-pixel convergence via a device-side int array (negative values mark converged pixels). The original code copied the entire array (~3.5 MB at 720p) to the host, then counted on the CPU. A replacement single-pass GPU reduction kernel uses warp-shuffle instructions; only one int (4 bytes) is transferred back to the host.

__global__ void countConvergedKernel(
    const int *pixel_sample_counts, int num_pixels, int *d_converged_count)
{
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int converged = (idx < num_pixels && pixel_sample_counts[idx] < 0) ? 1 : 0;

    // Warp-level reduction — no shared memory needed
    for (int offset = 16; offset > 0; offset >>= 1)
        converged += __shfl_down_sync(0xFFFFFFFF, converged, offset);

    if ((threadIdx.x & 31) == 0)
        atomicAdd(d_converged_count, converged);
}

Measured impact: eliminates the per-frame 3.5 MB D2H transfer. The GPU kernel runs in < 0.1 ms; the old copy + CPU loop took ~1–2 ms per frame.

Files: shaders/render_acc_kernel.cu, renderer_cuda_device.cu

16 — Accumulation buffer reset ordering¶

What it is: when a camera move invalidates accumulated samples, the accumulation buffer must be zeroed before the next render kernel reads it. Using cudaMemset() (which enqueues on the default stream 0) while the render kernel runs on a non-blocking custom stream creates a race condition: the kernel can start reading the buffer while the memset is still running.

The fix is to use cudaMemsetAsync on the same non-blocking stream as the kernel. This race affected both backends:

// WRONG — stream 0 races with s_compute_stream / render_stream
cudaMemset(d_accum_buffer, 0, size);

// CORRECT — guaranteed to complete before the next kernel launch on the same stream
cudaMemsetAsync(d_accum_buffer, 0, size, s_compute_stream);   // CUDA
cudaMemsetAsync(g_state.d_accum_buffer, 0, size, getOptiXStream());  // OptiX

Measured impact: eliminates a white-frame artifact in the CUDA renderer (visible on every camera move when adaptive sampling was enabled) and black-streak artifacts in the OptiX renderer. No throughput change — this is a correctness fix.

Files: renderer_cuda_device.cu, optix/optix_renderer.cu

17 — OptiX: GPU-side gamma correction with pinned memory¶

What it is: the original OptiX pipeline downloaded the full float4 accumulation buffer to the host (~14 MB at 720p) before performing gamma correction and format conversion on the CPU. A GPU gamma-correction kernel now converts float4 → uint8 directly on the device; only the compact display buffer (~2.7 MB) is transferred via an async copy to pinned host memory.

GPU (float4 accum) → gammaCorrectKernel → uint8 d_display
                                            ↓
                                    cudaMemcpyAsync (2.7 MB, pinned)
                                            ↓
                                        Host display buffer

Measured impact:

• 5× smaller D2H transfer: 2.7 MB (uint8 RGB) vs. 14 MB (float4 RGBA)
• GPU-parallel gamma correction: no CPU involvement between render and display
• Async DMA transfer: pinned memory allows the GPU's DMA engine to write directly over PCIe

Files: optix/optix_renderer.cu, renderer_optix_host.hpp, renderer_optix_progressive_host.hpp

18 — Firefly rejection (per-sample luminance clamp)¶

What it is: HDR environment map texels (e.g. the sun disk in an outdoor sky image) can have linear luminance > 50 000. A single such sample early in accumulation snaps the pixel to white and takes many subsequent samples to average down. A luminance-preserving clamp caps each sample's contribution before it is added to the accumulation buffer; hue is preserved by scaling all three channels uniformly.

// In renderAccKernel (CUDA) and __raygen__rg (OptiX):
constexpr float FIREFLY_CLAMP = 20.0f;
float sample_lum = 0.2126f * color.x + 0.7152f * color.y + 0.0722f * color.z;
if (sample_lum > FIREFLY_CLAMP)
    color = color * (FIREFLY_CLAMP / sample_lum);  // scale, don't clip per-channel

The threshold of 20.0 (linear) covers the full visible sky (3–15) while rejecting only the extreme sun-disk texels. Per-channel clamping (fminf(r, C)) is avoided because it shifts hue — a luminance scale keeps the colour balanced.

Measured impact: eliminates white-dot flickering during camera motion with HDR environment maps. Introduces a slight bias in extremely bright regions — the standard trade-off in production renderers (Blender Cycles exposes equivalent "Clamp Direct / Indirect" settings).

Files: shaders/render_acc_kernel.cu, optix/optix_programs.cu

Summary¶

The combined CUDA + BVH speedup reaches ~1 660× over single-threaded CPU on the default scene at 720p, 1 024 SPP — measured on an NVIDIA DGX Spark (GB10 GPU).

Techniques 1–16, 19 and 21 do not change the converged image. Techniques 17 and 18 introduce minor biases (OptiX gamma rounding; luminance clamping) that are invisible at normal viewing conditions but eliminate distracting artifacts, and technique 20 only affects the preview while the camera moves.