nandi/vflutter_ffi · README.md

⑂ 59a731a766d2baec2d522605425d2fce9c6c7209

README.md · 120 lines · 4.2 KBmarkdown    Blame HistoryRaw
Add a Mandelbrot example app, and a V kernel worth demonstrating  nandithebull 1# vflutter_ffi example
3A Mandelbrot explorer whose pixels are computed in V. Scroll to zoom smoothly,
4tap to zoom in, right-click to zoom out — all anchored at the cursor. Sliders
5control the iteration cap and how many isolates the frame is split across;
6"Reset view" restores the framing without disturbing either.
8It exists to exercise the parts of the bridge the string demo cannot:
10* **Bulk data across the boundary.** Each frame is a full RGBA buffer, not a
11  string. Dart allocates it and `vf_mandelbrot` fills it in place, so no V
12  memory crosses the boundary and there is nothing to `vf_free`.
13* **Band parallelism.** The image splits into horizontal strips rendered on
14  separate isolates. Bands are independent by construction, and `-gc none`
15  means no stop-the-world phase to serialise them.
16* **An honest benchmark.** The same algorithm is written twice — in V
17  (`src/vflutter.v`) and in Dart (`lib/mandelbrot_dart.dart`) — and the
18  Benchmark button times both. The tests assert the two agree pixel-for-pixel,
19  which is what makes the comparison meaningful.
21## What the numbers actually say
23Measured here (800x600, maxIter 500, release build, 8-core Linux):
25
26
27| V, one isolate | ~152 ms |
28| V, 4 isolates | ~65 ms |
29| Dart AOT, one isolate | ~165 ms |
31**V is not meaningfully faster than Dart at this.** For a scalar
32floating-point loop, V-through-C and Dart AOT land within ~10% of each other.
33The 2.5x comes from the parallel split, which Dart could do on its own without
34any FFI. Treat this app as a demonstration of the *bridge* — bulk buffers,
35caller-owned memory, safe multi-isolate calls — not as a speed claim for V.
37Reach for V here because you want to write the logic in V, or already have it
38in V, not because C-via-V is expected to outrun Dart AOT on arithmetic.
Raise the isolate ceiling to 32, and add the sweep that justifies it  nandithebull 40## How many isolates?
42The slider goes to 32, and past the core count is the right place to be. Bands
43cost very different amounts — rows crossing the set's interior run the full
44iteration cap, rows in open space escape almost immediately — and a frame is
45not done until its slowest band is. Over-decomposing lets a free core start the
46next small band instead of idling.
48Measured here (800x600 / 500 iterations, 8 logical cores):
50| tiles | best ms | speedup |
51|---|---|---|
52| serial | 153 | 1.00x |
53| 4 | 64 | 2.39x |
54| 8 | 43 | 3.56x |
55| 16 | 34 | 4.50x |
56| 32 | 29 | 5.28x |
57| 64 | 27 | 5.67x |
59Even at 32 bands the slowest is still 2.7x the mean, which is the remaining
60inefficiency — not isolate overhead. Spawning 24 isolates costs ~0.8 ms and
61moving the whole 1.83 MB frame back across them ~1.4 ms, against a ~29 ms
62frame. Run `just sweep` to get these numbers for your own machine.
Add a Mandelbrot example app, and a V kernel worth demonstrating  nandithebull 64
Add a justfile  nandithebull 66
67just mandelbrot
70or, driving Flutter yourself:
Add a Mandelbrot example app, and a V kernel worth demonstrating  nandithebull 72
73
74flutter run -d linux
77Only the Linux runner is generated. For the other platforms, generate their
78runner directories once first:
81flutter create --platforms=android,ios,macos,windows .
84`flutter create` leaves `lib/`, `test/` and `pubspec.yaml` alone. The native
85library is built by the plugin itself: CMake invokes `v` on Linux, Windows and
86Android; iOS and macOS compile the pre-generated C in `ios/Classes/`.
90`flutter test` runs on the host VM rather than in the app bundle, so it needs
91the native library on the loader path:
94./tool/build.sh
95cd example && LD_LIBRARY_PATH=../build flutter test
98The suite covers the zoom maths as a property (the point under the cursor stays
99put), the buffer contract, the band split matching a single-shot render, and
100the V and Dart kernels agreeing exactly.
102## Headless scripts
104No Flutter SDK needed — only `dart` and the host library. Run from the
105repository root, which resolves without Flutter:
108./tool/build.sh
109LD_LIBRARY_PATH=build dart example/lib/fractal_bench.dart   # V vs Dart timings
110LD_LIBRARY_PATH=build dart example/lib/main_test.dart       # string round-trips
111LD_LIBRARY_PATH=build dart example/lib/memory_check.dart    # RSS regression guard
114For a like-for-like comparison with a release app, compile the benchmark AOT
115first — `dart run` uses the JIT:
118dart compile exe example/lib/fractal_bench.dart -o /tmp/bench
119LD_LIBRARY_PATH=build /tmp/bench

Add a Mandelbrot example app, and a V kernel worth demonstrating 81540bb nandithebull 11h ago	1	# vflutter_ffi example
	2
	3	A Mandelbrot explorer whose pixels are computed in V. Scroll to zoom smoothly,
	4	tap to zoom in, right-click to zoom out — all anchored at the cursor. Sliders
	5	control the iteration cap and how many isolates the frame is split across;
	6	"Reset view" restores the framing without disturbing either.
	7
	8	It exists to exercise the parts of the bridge the string demo cannot:
	9
	10	* Bulk data across the boundary. Each frame is a full RGBA buffer, not a
	11	string. Dart allocates it and `vf_mandelbrot` fills it in place, so no V
	12	memory crosses the boundary and there is nothing to `vf_free`.
	13	* Band parallelism. The image splits into horizontal strips rendered on
	14	separate isolates. Bands are independent by construction, and `-gc none`
	15	means no stop-the-world phase to serialise them.
	16	* An honest benchmark. The same algorithm is written twice — in V
	17	(`src/vflutter.v`) and in Dart (`lib/mandelbrot_dart.dart`) — and the
	18	Benchmark button times both. The tests assert the two agree pixel-for-pixel,
	19	which is what makes the comparison meaningful.
	20
	21	## What the numbers actually say
	22
	23	Measured here (800x600, maxIter 500, release build, 8-core Linux):
	24
	25	\| \| time \|
	26	\|---\|---\|
	27	\| V, one isolate \| ~152 ms \|
	28	\| V, 4 isolates \| ~65 ms \|
	29	\| Dart AOT, one isolate \| ~165 ms \|
	30
	31	V is not meaningfully faster than Dart at this. For a scalar
	32	floating-point loop, V-through-C and Dart AOT land within ~10% of each other.
	33	The 2.5x comes from the parallel split, which Dart could do on its own without
	34	any FFI. Treat this app as a demonstration of the bridge — bulk buffers,
	35	caller-owned memory, safe multi-isolate calls — not as a speed claim for V.
	36
	37	Reach for V here because you want to write the logic in V, or already have it
	38	in V, not because C-via-V is expected to outrun Dart AOT on arithmetic.
	39
Raise the isolate ceiling to 32, and add the sweep that justifies it 59a731a nandithebull 11h ago	40	## How many isolates?
	41
	42	The slider goes to 32, and past the core count is the right place to be. Bands
	43	cost very different amounts — rows crossing the set's interior run the full
	44	iteration cap, rows in open space escape almost immediately — and a frame is
	45	not done until its slowest band is. Over-decomposing lets a free core start the
	46	next small band instead of idling.
	47
	48	Measured here (800x600 / 500 iterations, 8 logical cores):
	49
	50	\| tiles \| best ms \| speedup \|
	51	\|---\|---\|---\|
	52	\| serial \| 153 \| 1.00x \|
	53	\| 4 \| 64 \| 2.39x \|
	54	\| 8 \| 43 \| 3.56x \|
	55	\| 16 \| 34 \| 4.50x \|
	56	\| 32 \| 29 \| 5.28x \|
	57	\| 64 \| 27 \| 5.67x \|
	58
	59	Even at 32 bands the slowest is still 2.7x the mean, which is the remaining
	60	inefficiency — not isolate overhead. Spawning 24 isolates costs ~0.8 ms and
	61	moving the whole 1.83 MB frame back across them ~1.4 ms, against a ~29 ms
	62	frame. Run `just sweep` to get these numbers for your own machine.
	63
Add a Mandelbrot example app, and a V kernel worth demonstrating 81540bb nandithebull 11h ago	64	## Run it
	65
Add a justfile f7c3825 nandithebull 11h ago	66	```bash
	67	just mandelbrot
	68	```
	69
	70	or, driving Flutter yourself:
	71
Add a Mandelbrot example app, and a V kernel worth demonstrating 81540bb nandithebull 11h ago	72	```bash
	73	cd example
	74	flutter run -d linux
	75	```
	76
	77	Only the Linux runner is generated. For the other platforms, generate their
	78	runner directories once first:
	79
	80	```bash
	81	flutter create --platforms=android,ios,macos,windows .
	82	```
	83
	84	`flutter create` leaves `lib/`, `test/` and `pubspec.yaml` alone. The native
	85	library is built by the plugin itself: CMake invokes `v` on Linux, Windows and
	86	Android; iOS and macOS compile the pre-generated C in `ios/Classes/`.
	87
	88	## Tests
	89
	90	`flutter test` runs on the host VM rather than in the app bundle, so it needs
	91	the native library on the loader path:
	92
	93	```bash
	94	./tool/build.sh
	95	cd example && LD_LIBRARY_PATH=../build flutter test
	96	```
	97
	98	The suite covers the zoom maths as a property (the point under the cursor stays
	99	put), the buffer contract, the band split matching a single-shot render, and
	100	the V and Dart kernels agreeing exactly.
	101
	102	## Headless scripts
	103
	104	No Flutter SDK needed — only `dart` and the host library. Run from the
	105	repository root, which resolves without Flutter:
	106
	107	```bash
	108	./tool/build.sh
	109	LD_LIBRARY_PATH=build dart example/lib/fractal_bench.dart # V vs Dart timings
	110	LD_LIBRARY_PATH=build dart example/lib/main_test.dart # string round-trips
	111	LD_LIBRARY_PATH=build dart example/lib/memory_check.dart # RSS regression guard
	112	```
	113
	114	For a like-for-like comparison with a release app, compile the benchmark AOT
	115	first — `dart run` uses the JIT:
	116
	117	```bash
	118	dart compile exe example/lib/fractal_bench.dart -o /tmp/bench
	119	LD_LIBRARY_PATH=build /tmp/bench
	120	```