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 spec for LOAD (#16221)

Browse source 
* add load to the spec

* can
This commit is contained in:
George Hotz 2026-05-15 14:46:00 -07:00 • committed by GitHub
commit 2570bded8b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 20 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

BIN
spec/tinyspec.pdf
View file

Binary file not shown.

30
spec/tinyspec.tex
View file

@ -16,6 +16,7 @@
\definecolor{elwyellow}{HTML}{F9A825}
\definecolor{callblue}{HTML}{1565C0}
\definecolor{assignbrown}{HTML}{795548}
\definecolor{loadred}{HTML}{c08080}
\definecolor{multipurple}{HTML}{7B1FA2}
\definecolor{markerorange}{HTML}{E65100}
% AxisType colors (from tinygrad)
@ -48,16 +49,16 @@ All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op
\toprule
\textbf{Op} & \textbf{src} & \textbf{arg} & \textbf{Semantics} \\
\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{BufferView} & (buf,) & size, dtype, offset &
Typed access into a buffer. Zero-copy $(\textit{size},)$ slice at offset; inherits addrspace. \\
\op{Param} & $(\mathbf{s})$ or $(\mathbf{s}, \text{min}, \text{max})$ & slot, dtype, device? &
Placeholder with shape $\mathbf{s}$. Substituted in \op{Function}. \\[4pt]
\op{Const} & () & value, dtype &
A scalar constant with shape $(\ )$. \\
\op{Vconst} & () & values, dtype &
A vector constant with shape $(n,)$. \\
& & & Form vector consts with \op{Stack} \\
\op{Binary} & () & data & Raw binary data, has dtype uint8 and shape len($data$) \\
\bottomrule
\end{tabular}
@ -90,7 +91,7 @@ A \op{Buffer}'s \textbf{addrspace} is \texttt{GLOBAL}, \texttt{LOCAL}, or \textt
\toprule
\textbf{Op} & \textbf{src} & \textbf{arg} & \textbf{Semantics} \\
\midrule
\op{Reduce} & $(T,)$ & op, axes & Reduce $T$ along axes. Op is \op{Add}, \op{Max}, or \op{Mul}. \\
\op{Reduce} & ($T$, $r_0$, $r_1$, \ldots) & op, axes & Reduce $T$ along axes or ranges. Op is \op{Add}, \op{Max}, or \op{Mul}. \\
\bottomrule
\end{tabular}
@ -109,13 +110,25 @@ A \op{Buffer}'s \textbf{addrspace} is \texttt{GLOBAL}, \texttt{LOCAL}, or \textt
\end{tabular}
%% ============================================================
\subsection*{{\color{multipurple}Store Ops} \normalfont\small--- side effects}
\subsection*{{\color{loadred}Load Ops} \normalfont\small--- can change device or addrspace}
\begin{tabular}{@{}l l l l@{}}
\toprule
\textbf{Op} & \textbf{src} & \textbf{arg} & \textbf{Semantics} \\
\midrule
\op{Store} & (buf, val, gate?) & --- & Write val into buf. buf.shape $=$ val.shape. \\
\op{Load} & (buf, alt?, gate?) & device, addrspace & Read (pull) from buffer into a new anonymous buffer. \\
& & & Note: this replaces \op{Copy} and \op{Contiguous}. \\
\bottomrule
\end{tabular}
%% ============================================================
\subsection*{{\color{multipurple}Store Ops} \normalfont\small--- the only op with observable side effects}
\begin{tabular}{@{}l l l l@{}}
\toprule
\textbf{Op} & \textbf{src} & \textbf{arg} & \textbf{Semantics} \\
\midrule
\op{Store} & (buf, val, gate?) & --- & Write (push) val into buf. buf.shape $=$ val.shape. \\
& & & If gate is present, write only when gate is true. Output is void. \\
\bottomrule
\end{tabular}
@ -205,7 +218,6 @@ Ternary & $(P, A, B)$
\op{Contiguous} & $(T,)$ & --- & Force contiguous memory layout. \\
\op{ContiguousBackward} & $(T,)$ & --- & Force contiguous in backward pass. \\
\op{Detach} & $(T,)$ & --- & Stops gradient propagation. \\
\op{Copy} & $(T,)$ & device & Copy to target device. \\
\bottomrule
\end{tabular}
@ -216,8 +228,6 @@ Ternary & $(P, A, B)$
\toprule
\textbf{Op} & \textbf{src} & \textbf{arg} & \textbf{Semantics} \\
\midrule
\op{Load} & (idx,alt?,gate?) & --- & Dereference: read element at index from buffer. \\
& & & All loads will be replaced by \op{Store}. \\
\op{Barrier} & (deps\ldots) & --- & Synchronize threads within a workgroup. \\
\op{Ins} & \ldots & \ldots & A single machine instruction (e.g.\ AMD ISA). \\
\op{Special} & (bound,) & name & GPU thread/workgroup index (e.g.\ \texttt{gidx0}, \texttt{lidx1}). \\