Rico贾若童的博客

[ML] Data Formats

Pickle, pth, onnx

Pickle A Python serialization format for saving objects to disk and loading them back later. Common use cases include models, dictionaries, lists, pandas DataFrames, preprocessing objects, and int...

Posted by Rico's Nerd Cluster on January 23, 2026

[CUDA - 6] CUDA Functions

attomicAdd, pragma unroll

atomicAdd There’s no such thing as “atomic memory” in CUDA—atomicity is a property of an operation, not a memory type. atomicAdd is an instruction you apply to a memory location (typically in glob...

Posted by Rico's Nerd Cluster on January 21, 2026

[CUDA - 5] SIMT in CUDA

SIMD SIMT

SIMT and SIMD CUDA is a very good embodiment of SIMD (Single-Instruction-Multiple Data). SIMD is great for addressing embarassingly parallel problems, problems that are so “embarassingly” simple, ...

Posted by Rico's Nerd Cluster on January 19, 2026

[CUDA - 4] My First CUDA Kernel - Chamfer Distance

First CUDA Program, JIT Compile

On Chamfer Distance Chamfer distance measures how close two point clouds are by averaging nearest-neighbor distances in both directions. Given two point sets $P_1$ and $P_2$, For each point ...

Posted by Rico's Nerd Cluster on January 18, 2026

[CUDA - 3] - Mixed Precision Training

AT_DISPATCH_FLOATING_TYPES_AND_HALF, gpuAtomicAdd

Run Time Type Dispatch AT_DISPATCH_FLOATING_TYPES_AND_HALF(dtype, name, lambda) expands into something like: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 switch (points_tensor.scalar_type()) { case...

Posted by Rico's Nerd Cluster on January 16, 2026

[CUDA - 2] Introduction to CUDA Coding

CUDA Programming Hierarchy

What is Cuda CUDA is fundamentally C++ language with extensions. kernels, __global__, __device__ etc are defined in the C++ space. If you want to expose C to CUDA, just follow the standard 1 exte...

Posted by Rico's Nerd Cluster on January 14, 2026

[CUDA - 2] CUDA Introduction

First CUDA Program, SIMD SIMT,

What is Cuda Cuda is fundamentally C++ language with extensiions. kernels, __global__, __device__ etc are defined in the C++ space. If you want to expose C to CUDA, just follow the standard 1 ext...

Posted by Rico's Nerd Cluster on January 14, 2026

[CUDA - 1] GPU Architecture

GPU Architecture, Tensor Cores, Pinned Memory

A Great introduction video can be found here GPU (GA102) Architecture A graphics card’s brain is its GPU. NVidia’s Ampere GPU architecture family has GA102 and GA104. GA102 is shared across NVidi...

Posted by Rico's Nerd Cluster on January 11, 2026

[Point Cloud Compression] - Sampling

Furthest Point Sampling Idea: given a set of points 1 P = {P1, P2, ...} Select 1 S = {s1, s2 ...} such that S is maximally spread out. How? Add an intial point in P to S. Then iterativ...

Posted by Rico's Nerd Cluster on January 7, 2026

[ML] -Point-cloud-compression-1-Basic-Concepts

PSNR, Chamfer Distance

[Part 1] PSNR vs SNR In signal processing, the Signal-to-Noise Ratio (SNR) is defined as: \[\mathrm{SNR} = 10 \log_{10} \left( \frac{\mathrm{Var}(\text{signal})}{\mathrm{Var}(\text{noise})} \righ...

Posted by Rico's Nerd Cluster on January 5, 2026

Rico's Nerd Cluster