mirrors/tinygrad
2
0
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-06-24 02:14:17 +00:00
Code Issues Projects Releases Packages Wiki Activity

update buffer in spec (#16507)

Browse source
This commit is contained in:
George Hotz 2026-06-04 10:12:31 -07:00 • committed by GitHub
commit 4fb8ce1831
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 7 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

BIN
spec/tinyspec.pdf
View file

Binary file not shown.

16
spec/tinyspec.tex
View file

@ -40,7 +40,7 @@
\section*{UOps}
All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op},\;\mathrm{src},\;\mathrm{arg},\;\mathrm{tag})$ where $\mathrm{op}$ is from the set below, $\mathrm{src}$ is a tuple of input UOps, $\mathrm{arg}$ is op-dependent, and $\mathrm{tag}$ is for temporary processing. The full program is a DAG of UOps. Each UOp has five derived properties --- \textbf{dtype}, \textbf{shape}, \textbf{device}, \textbf{min\_max}, and \textbf{axis} --- determined by the rules at the end of this document.
All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op},\;\mathrm{src},\;\mathrm{arg},\;\mathrm{tag})$ where $\mathrm{op}$ is from the set below, $\mathrm{src}$ is a tuple of input UOps, $\mathrm{arg}$ is op-dependent, and $\mathrm{tag}$ is for temporary processing. The full program is a DAG of UOps. Each UOp has five derived properties --- \textbf{dtype}, \textbf{shape}, \textbf{device}, \textbf{addrspace}, and \textbf{min\_max} --- determined by the rules at the end of this document.
%% ============================================================
\subsection*{Source Ops \normalfont\small--- leaf nodes}
@ -51,8 +51,8 @@ All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op
\midrule
\op{Param} & $(\mathbf{s})$ & slot, dtype, device?, addrspace? &
Placeholder with shape $\mathbf{s}$. Substituted in \op{Function}. \\[4pt]
\op{Buffer} & () & size, dtype, device, addrspace &
Shape $(n \cdot \textit{size},)$ if device is $n$-tuple, else $(\textit{size},)$. \\
\op{Buffer} & $(\mathbf{s})$ & slot, dtype, device, addrspace &
Concrete buffer slot with shape $\mathbf{s}$. If device is a tuple, it creates the fully sized buffer across multiple devices. \\
\op{Const} & () & value, dtype &
A scalar constant with shape $(\ )$. \\
& & & Form vector consts with \op{Stack} \\
@ -61,7 +61,7 @@ All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op
\end{tabular}
\smallskip
A \op{Buffer}'s \textbf{addrspace} is \texttt{GLOBAL}, \texttt{LOCAL}, or \texttt{REG}.
\textbf{addrspace} is \texttt{GLOBAL}, \texttt{LOCAL}, or \texttt{REG}.
%% ============================================================
\subsection*{{\color{movgreen}Movement Ops} \normalfont\small--- no arithmetic; view, indexing, and reinterpretation only}
@ -74,12 +74,10 @@ A \op{Buffer}'s \textbf{addrspace} is \texttt{GLOBAL}, \texttt{LOCAL}, or \textt
\op{Flip} & $(T,)$ & bools $\mathbf{f}$ & Reverse along flagged axes. \\
\op{Reshape} & $(T, \mathbf{s'})$ & --- & Reinterpret in row-major order. $\prod s_k = \prod s'_k$. \\
\op{Expand} & $(T, \mathbf{s'})$ & --- & Broadcast size-1 axes. $s_k \in \{1, s'_k\}$. \\
\op{Pad} & $(T, \mathbf{o}, \mathbf{s'})$ & --- & Place $T$ at offset $o_k$ in a zero-filled output of shape $s'_k$. \\
\op{Pad} & $(T, \mathbf{o}, \mathbf{s'})$ & --- & Place $T$ at offset $o_k$ in an invalid-filled output of shape $s'_k$. \\
\op{Shrink} & $(T, \mathbf{o}, \mathbf{s'})$ & --- & Keep $s'_k$ elements starting at offset $o_k$ per axis. Inverse of \op{Pad}. \\
\op{Index} & $(T, i_0, i_1, \ldots)$ & --- & Index from left. $()$-shaped $i$ removes dim; $(k,)$-shaped makes it $k$. \\
\op{Stack} & $(T_0, T_1, \ldots)$ & --- & Join along a newly created leading axis. All shapes must match. \\
\op{Replicated} & $(T,)$ & axes & Mark $T$ as replicated along axes. Collapse axes to $1$. \\
\op{Slice} & $(T, \mathrm{offset})$ & size, dtype & Zero-copy \textit{size} elems of dtype; offset is elems of $T$ dtype. \\
\op{Bitcast} & $(T,)$ & dtype & Reinterpret storage as target dtype; preserve total bytes. \\
\bottomrule
\end{tabular}
@ -248,14 +246,14 @@ These ops are not part of the core specification and are subject to change.
%% ============================================================
\subsection*{Derived Properties}
Every UOp has a \textbf{dtype}, \textbf{shape}, \textbf{device}, \textbf{min\_max}, and \textbf{axis}, derived from its op, src, and arg:
Every UOp has a \textbf{dtype}, \textbf{shape}, \textbf{device}, \textbf{addrspace}, and \textbf{min\_max}, derived from its op, src, and arg:
\medskip
\begin{tabular}{@{}l l l l l@{}}
\toprule
\textbf{Op} & \textbf{dtype} & \textbf{shape} & \textbf{device} & \textbf{min\_max} \\
\midrule
\op{Buffer} & from arg & $(\text{size},)$ from arg & from arg & dtype range \\
\op{Buffer} & from arg & from $\mathrm{src}[0]$ & from arg & dtype range \\
\op{Const} & from arg & $()$ & \textsc{null} & $[v, v]$ \\
\op{Param} & from arg & from $\mathrm{src}[0]$ & from arg & from src or dtype range \\[3pt]
Movement ops & $\mathrm{src}[0].\mathrm{dtype}$ & (see op) & $\mathrm{src}[0].\mathrm{device}$ & $\mathrm{src}[0]$ \\