All tutorials · Updated September 21, 2026 · tinygrad 8ad8f73
Kernel names are debugging output. Their dimensions and colors summarize an optimized iteration structure, not necessarily the original Tensor shape or a literal GPU grid.
Current naming lives in Scheduler.get_optimized_ast. It chooses a reduction or elementwise prefix and incorporates special dimensions and optimized ranges. The old explanation based on a Linearizer.linearize method is obsolete.
from tinygrad.uop.ops import axis_colors
for axis, color in axis_colors.items():
print(axis.name, color)
assert axis_colors
The exact mapping comes from uop/ops.py. Scheduler.colors additionally handles weak ranges according to whether they participate in outputs and can become global work, so the dictionary alone does not explain every colored field.
A GLOBAL axis and a LOCAL axis describe different execution roles. UPCAST/UNROLL describe expanded work. REDUCE/GROUP_REDUCE describe reduction organization. Multiplying every visible extent and calling it a thread count conflates these roles.
Run an initialized Tensor computation with DEBUG=2 to find the kernel’s name and execution statistics. Use DEBUG=4 or VIZ to inspect source and launch metadata. Check the target’s mapping of global and local dimensions before interpreting a printed factor as a block or thread count.
Names can change after a compiler optimization without changing the output. Tests should normally assert behavior, not incidental kernel spellings. When investigating performance, preserve the full target/configuration alongside names and timings.
Original chapter by Di Zhu: historical version.