AKTIVE

OpenACS and Tcl/Tk Conference 2026, July 17-18

Andreas Kupries
AKTIVE

AK Tcl Image Vector Extension
MIT/BSD Licensed
Inspired by VIPS
Mixes C and Tcl via Critcl

AK Tcl Image Vector Extension

Last year: Introduced by use,
light on internals

See [PDF] [ZIP] [WU Lecturecast]
(Entire conference)
https://core.tcl-lang.org/akupries/aktive/doc/trunk/doc/presentations/eurotcl-2025-bologna.pdf

https://core.tcl-lang.org/akupries/aktive/doc/trunk/doc/presentations/eurotcl-2025-bologna.zip

https://learn.wu.ac.at/eurotcl2025/lecturecasts/753988258?m=delivery

https://learn.wu.ac.at/eurotcl2025/lecturecasts/?page=1

AK Tcl Image Vector Extension

Last year: Introduction by use,
light on internals
Today: A deeper look at the internals
- Demand-driven: Lazy construction
  - Data structures (definition, run)
- Horizontally Threaded Execution
  - Thread coordination, ordered and not
Benchmarking & Performance
- Loop unrolling
- (SIMD) Vectorization

AK Tcl Image Vector Extension

Lazy construction

Image creation builds definition structures
- Linear pipe
- Tree
- Directed acyclic graph

AK Tcl Image Vector Extension

Lazy construction

Image materialization builds execution
structures
- ... computes things
- ... and tears the execution structures down again
- The definition structures stay

AK Tcl Image Vector Extension

Lazy construction

Examples of materializators
- Saving pixels to file, channel, memory, ...
- Computing whole-image statistics

AK Tcl Image Vector Extension

Preface: Inspection

aktive format as d2        $image
aktive format as markdown  $image
aktive format as tclscript $image

set file1 [aktive read from netpbm file path ...];# FO(2): tmp2,tmp4
set tmp2 [aktive op view $file1 port {-1 -1 382 252}]
set tmp3 [aktive op tile max $tmp2 radius 1]
set tmp4 [aktive op view $file1 port {-1 -1 382 252}]
set tmp5 [aktive op tile min $tmp4 radius 1]
set tmp6 [aktive op math sub $tmp3 $tmp5]
set tmp7 [aktive op math1 invert $tmp6]
set tmp8 [aktive op math1 gamma expand $tmp7]
set tmp9 [aktive op color scRGB to Grey $tmp8]
set result [aktive op math1 scale $tmp9 factor 0.01]

AK Tcl Image Vector Extension

Preface: Inspection

aktive format as d2        $image
aktive format as markdown  $image
aktive format as tclscript $image

Charcoal effect

full size (assets/charcoal.svg)

AK Tcl Image Vector Extension

Definition Structures

Shoutout: Pikchr @ https://pikchr.org

type (function vectors, parameter defs)
parameters
srcs
domain (geometry)
state

AK Tcl Image Vector Extension

Definition Structures

type (function vectors, parameter defs)
content (parameters, srcs, domain, state)
metadata
refcount

AK Tcl Image Vector Extension

Definition Structures

Types and Parameters

AK Tcl Image Vector Extension

Definition Structures

Content and Meta

AK Tcl Image Vector Extension

Definition Structures

C Types

AK Tcl Image Vector Extension

Runtime Structures

(runtime) state
(runtime) srcs
result buffer
image
- param
- type
- domain
- state

AK Tcl Image Vector Extension

Runtime Structures

region: result buffer, image, type
region_info: runtime state, runtime srcs, param, domain, def state
context

AK Tcl Image Vector Extension

Runtime Structures

AK Tcl Image Vector Extension

Runtime Structures

AK Tcl Image Vector Extension

Horizontally Threaded Execution

Generate tasks
Process tasks
Collect results

AK Tcl Image Vector Extension

Horizontally Threaded Execution

Generate tasks --- Maker
Process tasks
Collect results

AK Tcl Image Vector Extension

Horizontally Threaded Execution

Generate tasks --- Maker
Process tasks --- Workers
Collect results

AK Tcl Image Vector Extension

Horizontally Threaded Execution

Generate tasks --- Maker
Process tasks --- Workers
Collect results --- Completer

AK Tcl Image Vector Extension

Horizontally Threaded Execution

Generate tasks --- Maker
Process tasks --- Workers

Here are the runtime structures

Collect results --- Completer

Coordination ?

Unordered
Sequential

AK Tcl Image Vector Extension

Horizontally Threaded Execution

Unordered

AK Tcl Image Vector Extension

Horizontally Threaded Execution

Sequential

AK Tcl Image Vector Extension

Benchmarking & Performance



op-bench {
    aktive image from value // width %W height %H depth 1 value 0.5
}

AK Tcl Image Vector Extension

Benchmarking & Performance

    foreach sink {
	null
	null-s
    } {
	# varying concurrency ... unthreaded, ...
	#                     ... threads up to num cpu cores, ...
	#                     ... then overcommitted cpu ...
	foreach cores {
	    -1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
	} {
	    # varying sizes ... small to large
	    foreach size {
		1 10 100 1000 10000 100000 1000000 10000000
	    } { ;#       1K   10K   100K   1M      10M
		# varying shapes of the same size ... tall to wide
		for { set h $size ; set w 1 } { $h > 0 } {
		    set w [expr {$w*10}] ; set h [expr {$h/10}]
		} { ...

AK Tcl Image Vector Extension

Benchmarking & Performance

Unordered vs. Sequential

AK Tcl Image Vector Extension

Benchmarking & Performance

Unordered vs. Sequential

AK Tcl Image Vector Extension

Benchmarking & Performance

Currently: Performance through threading
Can we gain more ? How ?
Make the workers themselves faster!

AK Tcl Image Vector Extension

Benchmarking & Performance

Make the workers themselves faster!

Idea 1: Loop unrolling

Out of Order Superscalar

Idea 2: Vectorization (SIMD)

x86 SSE*, AVX*, Arm NEON, SVE, RISC-V "V"
Google Highway (Portable, Length-Agnostic)
@ https://github.com/google/highway

AK Tcl Image Vector Extension

Function specification

set unary0 {
    acos           { scalar { v = acos  (v);     } }
    [...]
}

set unary1 {
    [...]
    shift  {
	scalar  { v = v + a;      }
	highway { v = Add (v, a); }
    }
    [...]
}

AK Tcl Image Vector Extension

Templating

set u4_unary0_vdecl {
    extern void aktive_vector4_unary_@name@ (
        double* d, double* s, aktive_uint n);
}
set u4_unary0_vdef {
    void aktive_vector4_unary_@name@ (double* d, double* s, aktive_uint n) {
	// 4-unroll
 	for (; n > 3; n -= 4, d += 4, s += 4) {
	    double v0 = s[0]; // v1, v2, v3 = s[1,2,3]
#define v v0 // ... v1 v2 v3
	    @opcode@
#undef  v
	    d[0] = v0; // d[1,2,3] = v1, v2, v3
	}
	// 2-unroll,
 	for (; n > 1; n -= 2, d += 2, s += 2) { ... }
	// remainder
 	for (; n > 0; n--, d++, s++) { ... }
    }
}

AK Tcl Image Vector Extension

Templating

set u1_unary0_hdecl {
    void aktive_highway1_unary_@name@ (
        double* d, double* s, aktive_uint n);}
set u1_unary0_hdef  {
    HWY_ATTR void aktive_highway1_unary_@name@ (
        double* d, double* s, aktive_uint n) {
	const int32_t N = Lanes(f64);
	@decls@
	if (n < N) {	// not a full block - do a masked partial op
	    auto mask = FirstN(f64, n);
	    auto v    = MaskedLoad(mask, f64, s);
	    @opcode@
	    BlendedStore (v, mask, f64, d);	    return;
	}
	aktive_uint k, border = n - N; 	// lane-sized blocks
	for (k = 0; k < n; k += N) {
	    aktive_uint at = HWY_MIN (k, border);
	    auto v = LoadU (f64, s + at);
	    @opcode@
	    StoreU (v, f64, d + at);
	}
    }
}

AK Tcl Image Vector Extension

Templating

set u1_unary0_hdecl {
    void aktive_highway1_unary_@name@ (
        double* d, double* s, aktive_uint n);}
set u1_unary0_hdef  {
    HWY_ATTR void aktive_highway1_unary_@name@ (
        double* d, double* s, aktive_uint n) {
	const int32_t N = Lanes(f64);
	@decls@
	if (n < N) {	// not a full block - do a masked partial op
	    auto mask = FirstN(f64, n);
	    auto v    = MaskedLoad(mask, f64, s);
	    @opcode@
	    BlendedStore (v, mask, f64, d);	    return;
	}
	aktive_uint k, border = n - N; 	// lane-sized blocks
	for (k = 0; k < n; k += N) {
	    aktive_uint at = HWY_MIN (k, border);
	    auto v = LoadU (f64, s + at);
	    @opcode@
	    StoreU (v, f64, d + at);
	}
    }
}