Overview

Welcome to Caten/APIs

WIP

proceed is a function that evaluates a computational graph.

lisp

CATEN-USER> (proceed (make-tensor `(3 3) :initial-element 1.0))

Result

{Tensor{LISPBUFFER}[float32] :shape (3 3) :id STC28090
   ((1.0 1.0 1.0)
    (1.0 1.0 1.0)
    (1.0 1.0 1.0))
  :op #<PROCEEDNODE {100328CD63}>
  :requires-grad NIL
  :variables (TID28027)
  :tracker #<TRACKER :order={row(0 1)} :shape=(3 3) :contiguous-p=T>}

ctx:with-contextvar is a macro that sets the context variable. BACKEND=CLANG to use JIT, JIT_DEBUG=4 to see the generated code. DOT=1 to debug the pattern matcher on your browser. (need graphviz)

lisp

CATEN-USER> (ctx:with-contextvar (:BACKEND "CLANG" :jit_debug 4 :dot 0)
  (caten (forward (Embedding 10 10) (make-tensor `(b c)))))

Result

#<CLANGRUNTIME {(c b) -> ((val_35), nil)}
    val_2 = allocate(shape=(10, 10), stride=(10, 1), from=<Realized Buffer>);
    val_9 = allocate(shape=(), stride=());
    val_10 = load(val_9); // :value = C
    val_15 = allocate(shape=(), stride=());
    val_16 = load(val_15); // :value = C
    val_17 = allocate(shape=(), stride=());
    val_18 = load(val_17); // :value = B
    val_22 = allocate(shape=(val_18, val_16), stride=(val_10, 1));
    val_12 = allocate(shape=(), stride=());
    val_13 = load(val_12); // :value = 10
    val_14 = mul(val_13, val_10);
    val_34 = allocate(shape=(val_18, val_16, 1, 10), stride=(val_14, 10, 10, 1));
    val_34 = fused_sumnode_embedding73130(val_34, c, b, val_2, val_22); // JIT: <CLANG[FUSED_SUMNODE_EMBEDDING73130] : #<EXPR (3*(100*(b*c)))>FLOP>
    val_35 = pause/backward(val_34);
>

Overview

Share via

Overview