My presentation in 2026 was about Accelerating machine vision and computation tasks with Sparsity, Vulkan and SPIR-V.
SPIR-V Wave Intrinsics (subgroup operations)
• A powerful feature for SIMD (Single Instruction, Multiple Data) processing on GPUs, allowing threads
within a “wave” to exchange data without explicit barriers
• Waves/Subgroups act as a group of threads executed in lockstep (typically 32 or 64 lanes)
• The most critical part of wave organization is how the hardware handles Sparsity (if-statements)
• The registers are “sliced” across the wave. When you call a wave intrinsic like WaveActiveSum(), the
hardware doesn’t send data to/from RAM, it uses an internal Crossbar Switch or Shuffle Network
Then I talked about the imaging approach for 3D micrscopy using sparsity.
Sparsity of Illumination
• Standard wide-field microscopy use Köhler bathing the subject in uniform light. This intentionally reduces
the phase contrast of the system as it isolates the focal plane. The Light rays arriving from opposing angles
mutually cancel the shadow casting and phase gradients that define high-frequency surface details
• By activating a single, incoherent point-light source (LED) at a time, we ensure that the photon ‘origin’ is
largely spatially distinct. This sparsity results in maximum phase conversion at all depths of field. This allows
the system to recover sharp edge responses even in regions that would be considered “out of focus” or lowcontrast in a traditional bright-field arrangement
• This also results in an arrangement which is not light starved as the LED is able to illuminate the camera
sensor and is only blocked where light is absorbed or refracted away allowing very high-speed capture with
minimal shot noise
Sparsity of the Subject
• The second governing principle is the sparsity of the subject relative to the detector. The system is designed
to significantly oversample the optical resolution limit (100 nm pixel pitch vs. ≈ 980 nm Rayleigh limit)
• In many biological contexts and areas like photonics there are large homogeneous areas that are punctuated
with “sparse” small features separated from neighbouring structures by distances exceeding the Rayleigh
criterion (separated enough they do not blur together). When an isolated sub-resolution feature is imaged
with this sparse illumination strategy, it behaves as a discrete signal source rather than part of a convoluted
texture
• This capability is particularly evident in low-contrast biological samples. For example, visualizing the endpiece of a sperm flagella is typically too small and low contrast. However, because the feature is sparse even
the weak gradient of the photometric response of the flagella tip can be isolated and reconstructed. The
inverse rendering algorithm detects this sparse signal by its consistent orientation response across the
different off axis lighting angles