Skip to content

Technical Analysis: Post-Process Optimizations (March 2026)

This document details GPU optimization paths for the post-processing pipeline of suckless-ogl, based on migrating computations to Compute Shaders.

General Objectives

  1. Reduce CPU overhead: Eliminate Framebuffer (FBO) switches and multiple draw calls.

  2. Maximize GPU occupancy: Leverage the massive parallelism of compute units (EUs/CUs) via Compute Shaders.

  3. Reduce synchronization barriers: Minimize waits between passes.

Part 1: Auto-Exposure (Luminance Calculation)

Concept

Replace the rasterization pass (Fragment Shader on a 64x64 quad) with a Compute Shader processing the scene texture.

Critical Points (Lessons Learned)

To maintain ISO parity with master, the Compute Shader must strictly replicate the physical logic:

  • 4x4 Sampling: Do not settle for a single texture() at the center. Average a pixel block (box filter) to capture light peaks.

  • Exclusion Threshold (0.05): Ignore pixels with luminance below 0.05. Without this, black sky or deep shadows pull the average down, causing massive overexposure.

  • Sentinel Value (-100.0): If a block is entirely black, it must be marked for the adaptation step to ignore it.

Implementation Steps

  1. Shader: Create shaders/lum_downsample.comp with a 4x4 sampling loop.

  2. Texture: Switch downsample_tex to R32F format for image storage (image2D).

  3. C Code: Remove downsample_fbo. Replace glDrawArrays with glDispatchCompute(8, 8, 1).

  4. Barrier: Add glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT) before the adaptation step.

Part 2: Bloom (Single-Pass Downsampling)

Concept

Replace the 5–6 successive downsampling passes with a single Compute Shader dispatch (similar to AMD's Single Pass Downsampler).

Technical Details

  • Mip Hierarchy: Use glBindImageTexture to bind multiple mipmap levels (1 to 4) simultaneously.

  • Parallelism: Each work group (8x8) processes a region and writes to the corresponding mips.

  • Format: Use R11F_G11F_B10F for compact, performant HDR storage.

Implementation Steps

  1. Shader: Create shaders/bloom_downsample.comp.

  2. C Code: Modify fx_bloom_init to configure mip textures with image access.

  3. Render: Replace the fragment render loop with a single dispatch. Keep Raster mode for Upsampling (which benefits from hardware blending GL_ONE, GL_ONE).

Part 3: Resource Management & RAII

Concept

Ensure reliable resource deletion during engine restarts or resolution changes.

Recommendations

  • SHADER_SAFE_DESTROY Macro: Always use a macro that checks for null before calling shader_destroy.

  • FBO Cleanup: Ensure that textures attached to FBOs are freed after the FBOs to avoid dangling pointers in the driver.

Part 4: Validation and Metrics

Test Methodology

  1. Visual Parity: Use tests/test_visual_fx.c to compare Raster and Compute output pixel-by-pixel. Any mean luminance difference > 1% must be treated as a bug.

  2. Benchmarking: Use apitrace to verify the elimination of "bubbles" in the pipeline (zones where the GPU waits for the CPU).

  3. Watchdog: For tests under Wine, implement an exit timeout if the application ignores the Escape signal due to X11 desynchronization.

Analysis performed on March 13, 2026.