CUDA Programming in 21 Days

CUDA Programming in 21 Days
A Hands-On Course in C++ and Python
by M. Saqib

================================================================
Go from "what even is a GPU?" to writing, debugging, and tuning
your own CUDA kernels - in three focused weeks.
================================================================

Most GPU books are either a wall of reference material or a thin
layer of copy-paste recipes. This one is a course. Each day is a
single sitting that builds on the last, teaches one idea
properly, and ends with a workshop so the knowledge lands in your
hands - not just your eyes.

You write REAL CUDA C++ that compiles with nvcc, and you see the
Python equivalent (CuPy and Numba) alongside every step, so you
can run and experiment even before your C++ is fluent. Every
speedup in the book is one you measure yourself.

WHAT YOU GET
- 21 chapters (3 weeks x 7 days), ~1.2 million words of careful,
worked teaching - no filler.
- 229 figures, diagrams, and plots, every one generated from a
real computation or a clean schematic.
- Hundreds of runnable listings in CUDA C++, CuPy, and Numba.
- Four formats in one purchase: PDF, EPUB, MOBI, and HTML.

THE THREE WEEKS
Week 1 - Get onto the GPU: why GPUs win, the toolkit, your
first kernel, thread indexing, moving data, and debugging.
Week 2 - Make it fast: the memory hierarchy, coalescing,
shared memory and tiling, synchronization, warps and
divergence, occupancy, and honest profiling with a roofline.
Week 3 - Patterns, libraries, and a real project: reduction,
atomics, scan, streams and overlap, Thrust/cuBLAS/cuFFT/CuPy,
and a complete, profiled image-convolution application.

WHO IT'S FOR
You know a little C or C++ and a little Python. You do NOT need
any GPU experience. You do not even need an expensive GPU - any
recent NVIDIA card works, and Day 2 shows you how to run every
example for free in the cloud if you have none.

BY DAY 21 YOU CAN
- decide whether a problem suits a GPU, and why;
- write, launch, and debug your own kernels;
- lay out memory and choose a launch configuration for speed;
- use reductions, scans, atomics, and streams with confidence;
- reach for the right library - and verify its result;
- build and profile a real GPU application end to end.

The GPU stops being a black box. Go build something that needed
all those threads.