flutter_ml 3.0.4 
flutter_ml: ^3.0.4 copied to clipboard
nomoproject.devBring native GPU-accelerated machine learning to Flutter. Train models, build transformers, and run local inference without Python or cloud APIs.
example/lib/main.dart
import 'dart:math';
import 'package:flutter/material.dart';
import 'package:flutter_ml/full_library.dart';
/*
void main() {
runApp(MaterialApp(home: SimpleGPUApp()));
}
class SimpleGPUApp extends StatelessWidget {
Target target=Target.cuda;
SimpleGPUApp({super.key});
@override
Widget build(BuildContext context) {
return Scaffold(
appBar: AppBar(title: const Text('WebGPU Test')),
body: Center(
child: ElevatedButton(
onPressed: () async {
try {
// useCuda: false zwingt die Engine, das WebGPU-Backend zu nutzen
//print("✅ GPUEngine erfolgreich initialisiert!");
//await testMath();
//print("_____________________________________________________________________________________________");
//await testLinear();
//print("_____________________________________________________________________________________________");
//await testBackward();
//print("_____________________________________________________________________________________________");
//await runTransformer();
//print("_____________________________________________________________________________________________");
//await testResultTransfer();
//print("_____________________________________________________________________________________________");
//await testSpeeds();
await runSimple();
} catch (e) {
print("❌ Fehler bei der Initialisierung: $e");
}
},
child: const Text('Initialize GPU Engine'),
),
),
);
}
Future<bool> runSimple() async {
await GPUEngine.initialize(target: target);
CommandBuffer tape = CommandBuffer();
GPUTensor<Vector> VecA = GPUTensor([1.1, 2.2, 3.3]);
GPUTensor<Vector> VecB = GPUTensor([4.4, 5.5, 6.6]);
GPUTensor<Vector> VecC = GPUTensor([1.0, 0.1, -1.0]);
GPUTensor<Vector> VecD = addVectorGPU(VecA, VecB, tape);
GPUTensor<Vector> VecE = elementWiseMultiplyGPU(VecC, VecD, tape);
GPUEngine.run(tape.bytes());
VecE.toCpu();
print("Result: ${VecE.value}");
VecE.printGraph();
TapeDecoder(tape.bytes()).decode();
return true;
}
Future<bool> runTransformer() async {
await GPUEngine.initialize(debug: false, target: target);
Random random = Random();
int numHeads = 8;
int runsPerTest = 50;
print(
'====================================================================================================',
);
print(
' GPU TRANSFORMER ENCODER BLOCK INFERENCE BENCHMARK (FORWARD ONLY) ',
);
print(
'====================================================================================================',
);
print(
'Heads: $numHeads | Feed-Forward Ratio: 4x dModel | Inference Runs/Test: $runsPerTest',
);
print(
'----------------------------------------------------------------------------------------------------',
);
print(
' SeqLen | dModel | Compile(ms) | Latency(ms) | Throughput (GB/s) | Compute (TFLOPs)',
);
print(
'----------------------------------------------------------------------------------------------------',
);
List<int> sequences = <int>[32, 64, 128, 512];
List<int> dimensions = <int>[32, 64, 128, 256];
for (int s = 0; s < sequences.length; s = s + 1) {
int seqLength = sequences[s];
for (int d = 0; d < dimensions.length; d = d + 1) {
int dModel = dimensions[d];
int dff = dModel * 4; // Standard Transformer configuration
// 1. Generate Input Matrix (Batch = 1)
List<List<double>> hInput = <List<double>>[];
for (int i = 0; i < seqLength; i = i + 1) {
List<double> row = <double>[];
for (int j = 0; j < dModel; j = j + 1) {
row.add((random.nextDouble() * 2.0) - 1.0);
}
hInput.add(row);
}
GPUTensor<Matrix> input = GPUTensor<Matrix>(hInput);
// 2. Build Transformer Block
TransformerEncoderBlockTapeLayer transformerBlock =
TransformerEncoderBlockTapeLayer(dModel, numHeads, dff);
transformerBlock.build(input);
Stopwatch compileSw = Stopwatch();
compileSw.start();
// ===================================================================
// FORWARD TAPE (INFERENCE ONLY)
// ===================================================================
CommandBuffer fTape = CommandBuffer();
List<GPUTensor> intermediates = <GPUTensor>[];
GPUTensor<Matrix> output =
transformerBlock.forward(input, fTape, intermediates)
as GPUTensor<Matrix>;
Uint8List forwardBytes = fTape.bytes();
compileSw.stop();
// ===================================================================
// EXECUTION LOOP & METRIC CALCULATIONS
// ===================================================================
double seqD = seqLength.toDouble();
double modD = dModel.toDouble();
double ffD = dff.toDouble();
// FLOPs purely for the forward pass
// MHA (Proj + Out): 8 * Seq * D^2
// MHA (Attention): 4 * Seq^2 * D
// FFN (W1 + W2): 4 * Seq * D * FF
double totalFlopsStep =
(8.0 * seqD * modD * modD) +
(4.0 * seqD * seqD * modD) +
(4.0 * seqD * modD * ffD);
// Memory Traffic (Bytes Read/Written to VRAM)
// Weights: ~4D^2 (MHA) + 2*D*FF (FFN) -> * 4 bytes
double weightBytes = (4.0 * modD * modD + 2.0 * modD * ffD) * 4.0;
// Activations: Rough estimate of intermediate reads/writes per step
double actBytes =
((20.0 * seqD * modD) + (5.0 * seqD * ffD) + (4.0 * seqD * seqD)) *
4.0;
double totalBytesStep = weightBytes + actBytes;
Stopwatch runSw = Stopwatch();
// Warmup (Push weights into VRAM caches)
GPUEngine.run(forwardBytes);
// Inference Loop
runSw.start();
for (int run = 1; run <= runsPerTest; run = run + 1) {
GPUEngine.run(forwardBytes);
}
runSw.stop();
double avgRunSec =
(runSw.elapsedMicroseconds / 1000000.0) / runsPerTest;
double avgRunMs = avgRunSec * 1000.0;
double tflops = (totalFlopsStep / avgRunSec) / 1000000000000.0;
double gbps = (totalBytesStep / avgRunSec) / 1000000000.0;
String sSeq = seqLength.toString().padRight(6);
String sDim = dModel.toString().padRight(6);
String sComp = compileSw.elapsedMilliseconds.toString().padRight(11);
String sAvg = avgRunMs.toStringAsFixed(2).padRight(11);
String sGbps = gbps.toStringAsFixed(2).padRight(17);
String sTflops = tflops.toStringAsFixed(4).padRight(16);
print(' $sSeq | $sDim | $sComp | $sAvg | $sGbps | $sTflops');
// Free Memory
transformerBlock.free();
input.free();
output.free();
for (int i = 0; i < intermediates.length; i = i + 1) {
intermediates[i].free();
}
}
}
print(
'----------------------------------------------------------------------------------------------------',
);
print('Benchmark Complete.');
return true;
}
Future<bool> testResultTransfer() async {
await GPUEngine.initialize(debug: false, target: target);
GPUTensor<Vec> a = GPUTensor<Vec>([
1.0,
2.0,
3.0,
4.0,
5.0,
6.0,
7.0,
8.0,
9.0,
10.0,
]);
GPUTensor<Vec> b = GPUTensor<Vec>([
10.0,
9.0,
8.0,
7.0,
6.0,
5.0,
4.0,
3.0,
2.0,
1.0,
]);
CommandBuffer tape = CommandBuffer();
GPUTensor<Scalar> c = dotProductGPU(a, b, tape);
GPUTensor<Scalar> d = l2NormGPU(a, tape);
GPUTensor<Scalar> e = euclideanDistanceGPU(a, b, tape);
GPUTensor<Scalar> f = cosineSimilarityGPU(a, b, tape);
GPUTensor<Scalar> g = maeLossGPU(a, b, tape);
GPUEngine.run(tape.bytes());
a.toCpu();
b.toCpu();
c.toCpu();
d.toCpu();
e.toCpu();
f.toCpu();
g.toCpu();
print("Original Tensors: ${a.value}, ${b.value}");
print("Dot Product: ${c.value}");
print("L2 Norm: ${d.value}");
print("Euclidean Distance: ${e.value}");
print("Cosine Similarity: ${f.value}");
print("MAE Loss: ${g.value}");
return true;
}
Future<bool> testLearning() async {
print("🚀 Initialisiere GPU Engine...");
// 1. Engine starten (passe das Target an, falls du auf Android/Web bist)
await GPUEngine.initialize(debug: false, target: this.target); //
print("📦 Allokiere Tensoren im VRAM...");
// 2. Tensoren anlegen
// Shape: [1, 2] bedeutet 1 Sample mit 2 Features
var input = GPUTensor<Matrix>.empty([1, 2]); //
var target = GPUTensor<Matrix>.empty([1, 1]); //
// Gewichte & Bias direkt im VRAM randomisiert initialisieren
var weights = GPUTensor<Matrix>.randomUniform([2, 1], 1.0); //[cite: 2]
var bias = GPUTensor<Vector>.randomUniform([1], 1.0); //[cite: 2]
print("📼 Zeichne Computational Graph (Tape) auf...");
// 3. CommandBuffer erstellen
var tape = CommandBuffer(); //[cite: 5]
// --- FORWARD PASS ---
// input * weights
var step1 = matMulGPU(input, weights, tape); //[cite: 3]
// + bias
var output = addBiasToMatMulOutGPU(step1, bias, tape); //[cite: 3]
// Loss berechnen (Mean Squared Error)
var loss = mseMatrixGPU(output, target, tape); //[cite: 3]
// --- BACKWARD PASS ---
// Vorherige Gradienten im Graphen nullen
loss.zeroGraphGrads(tape); //[cite: 2]
// Backpropagation auslösen (füllt den ersten Gradienten mit 1.0)
loss.backward(tape, fillOnes: true); //[cite: 2]
// --- OPTIMIZER STEP ---
// Gewichte und Bias via Stochastic Gradient Descent anpassen
sgdUpdateGPU(weights, 0.01, tape); //[cite: 3]
sgdUpdateGPU(bias, 0.01, tape); //[cite: 3]
// Tape zu Bytes kompilieren
Uint8List compiledTape = tape.bytes(); //[cite: 5]
print("🔥 Starte Training auf der Engine...");
// 4. Test-Trainings-Loop feuern
for (int epoch = 1; epoch <= 50; epoch++) {
// a) Daten von der Dart-CPU in den VRAM pushen
input.pushData([0.5, 0.8]); //[cite: 2]
target.pushData([1.0]); //[cite: 2]
// b) Das komplette kompilierte Tape auf der C++ Engine ausführen
GPUEngine.run(compiledTape); //[cite: 1]
// c) Den berechneten Loss-Wert zurück in den RAM holen, um ihn anzuzeigen
loss.toCpu(); //[cite: 2]
print("Epoche $epoch - Loss: ${loss.value}"); //[cite: 2]
}
print("🧹 Räume VRAM auf...");
// 5. Sauberes Memory Management
input.free(); //[cite: 2]
target.free(); //[cite: 2]
weights.free(); //[cite: 2]
bias.free(); //[cite: 2]
step1.free(); //[cite: 2]
output.free(); //[cite: 2]
loss.free(); //[cite: 2]
GPUEngine.dispose(); //[cite: 1]
print("✅ Test erfolgreich abgeschlossen!");
return true;
}
Future<bool> testMath() async {
print("🚀 Starte massive GPU Math-Validierung...");
await GPUEngine.initialize(
debug: false,
target: this.target,
); //[cite: 1]
// ======================================================================
// 1. BASIC MATH (Element-wise)
// ======================================================================
print("\n--- 1. BASIC MATH (Vectors) ---");
{
var vA = GPUTensor<Vector>.empty([2]); //[cite: 2]
var vB = GPUTensor<Vector>.empty([2]); //[cite: 2]
vA.pushData([4.0, 9.0]); //[cite: 2]
vB.pushData([2.0, 3.0]); //[cite: 2]
var tape = CommandBuffer(); //[cite: 5]
var tAdd = addGPU<Vector>(vA, vB, tape); //[cite: 3]
var tSub = subtractGPU<Vector>(vA, vB, tape); //[cite: 3]
var tMul = multiplyGPU<Vector>(vA, vB, tape); //[cite: 3]
var tDiv = divideGPU<Vector>(vA, vB, tape); //[cite: 3]
var tExp = vectorExpGPU(vA, tape); //[cite: 3]
var tAbs = absGPU<Vector>(
GPUTensor<Vector>([-4.0, -9.0]),
tape,
); //[cite: 2, 3]
var tSqrt = sqrtGPU<Vector>(vA, tape); //[cite: 3]
var tLog = logGPU<Vector>(vA, tape); //[cite: 3]
var tPow = powGPU<Vector>(vA, 2.0, tape); //[cite: 3]
var tClamp = clampGPU<Vector>(vA, 0.0, 5.0, tape); //[cite: 3]
GPUEngine.run(tape.bytes()); //[cite: 1]
tAdd.toCpu();
tSub.toCpu();
tMul.toCpu();
tDiv.toCpu(); //[cite: 2]
tExp.toCpu();
tAbs.toCpu();
tSqrt.toCpu();
tLog.toCpu();
tPow.toCpu();
tClamp.toCpu(); //[cite: 2]
print("Inputs: vA=[4, 9], vB=[2, 3]");
print("ADD: ${tAdd.value}");
print("SUB: ${tSub.value}");
print("MUL: ${tMul.value}");
print("DIV: ${tDiv.value}");
print("EXP(vA):${tExp.value}");
print("ABS([-4, -9]): ${tAbs.value}");
print("SQRT(vA): ${tSqrt.value}");
print("LOG(vA): ${tLog.value}");
print("POW(vA, 2): ${tPow.value}");
print("CLAMP(vA,0,5): ${tClamp.value}");
vA.free();
vB.free();
tAdd.free();
tSub.free();
tMul.free();
tDiv.free(); //[cite: 2]
tExp.free();
tAbs.free();
tSqrt.free();
tLog.free();
tPow.free();
tClamp.free(); //[cite: 2]
}
// ======================================================================
// 2. LINEAR ALGEBRA & MATRIX OPS
// ======================================================================
print("\n--- 2. LINEAR ALGEBRA ---");
{
print("Make GPUTensors");
var mA = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
var mB = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
var vC = GPUTensor<Vector>.empty([2]); //[cite: 2]
print("Push Data to GPU");
mA.pushData([1.0, 2.0, 3.0, 4.0]); // [[1, 2], [3, 4]] //[cite: 2]
mB.pushData([2.0, 0.0, 1.0, 2.0]); // [[2, 0], [1, 2]] //[cite: 2]
vC.pushData([10.0, 20.0]); //[cite: 2]
print("Tape");
var tape = CommandBuffer(); //[cite: 5]
print("tMatMul");
var tMatMul = matMulGPU(mA, mB, tape); //[cite: 3]
print("tMatVec");
var tMatVec = matVecMulGPU(mA, vC, tape); //[cite: 3]
print("tTrans");
var tTrans = transposeGPU(mA, tape); //[cite: 3]
print("tAddBi");
var tAddBi = addBiasToMatMulOutGPU(mA, vC, tape); //[cite: 3]
print("tScale");
var tScale = scaleMatrixGPU(mA, 10.0, tape); //[cite: 3]
print("tAddSca");
var tAddSca = addScalarMatrixGPU(
mA,
GPUTensor<Scalar>(5.0),
tape,
); //[cite: 2, 3]
print("Run tape");
GPUEngine.run(tape.bytes()); //[cite: 1]
print("Push to CPU");
tMatMul.toCpu();
tMatVec.toCpu();
tTrans.toCpu();
tAddBi.toCpu();
tScale.toCpu();
tAddSca.toCpu(); //[cite: 2]
print("mA=[[1,2],[3,4]], mB=[[2,0],[1,2]], vC=[10,20]");
print("MATMUL(mA, mB): ${tMatMul.value}");
print("MATVEC(mA, vC): ${tMatVec.value}");
print("TRANSPOSE(mA): ${tTrans.value}");
print("ADDBIAS(mA, vC): ${tAddBi.value}");
print("SCALE(mA, 10): ${tScale.value}");
print("ADD_SCALAR(mA,5): ${tAddSca.value}");
mA.free();
mB.free();
vC.free();
tMatMul.free();
tMatVec.free();
tTrans.free();
tAddBi.free();
tScale.free();
tAddSca.free(); //[cite: 2]
}
// ======================================================================
// 3. ACTIVATIONS
// ======================================================================
print("\n--- 3. ACTIVATIONS ---");
{
var vA = GPUTensor<Vector>.empty([3]); //[cite: 2]
vA.pushData([-1.0, 0.0, 1.0]); //[cite: 2]
var mA = GPUTensor<Matrix>.empty([1, 2]); //[cite: 2]
mA.pushData([0.0, 1.0]); //[cite: 2]
var tape = CommandBuffer(); //[cite: 5]
var tRelu = reluGPU(vA, tape); //[cite: 3]
var tSig = sigmoidGPU(vA, tape); //[cite: 3]
var tTanh = vectorTanhGPU(vA, tape); //[cite: 3]
var tGelu = geluGPU(vA, tape); //[cite: 3]
var tSoftmax = softmaxMatrixGPU(mA, tape); //[cite: 3]
GPUEngine.run(tape.bytes()); //[cite: 1]
tRelu.toCpu();
tSig.toCpu();
tTanh.toCpu();
tGelu.toCpu();
tSoftmax.toCpu(); //[cite: 2]
print("Inputs: vA=[-1, 0, 1], mA=[[0, 1]]");
print("RELU(vA): ${tRelu.value}");
print("SIGMOID(vA): ${tSig.value}");
print("TANH(vA): ${tTanh.value}");
print("GELU(vA): ${tGelu.value}");
print("SOFTMAX(mA): ${tSoftmax.value}");
vA.free();
mA.free();
tRelu.free();
tSig.free();
tTanh.free();
tGelu.free();
tSoftmax.free(); //[cite: 2]
}
// ======================================================================
// 4. REDUCTIONS & STATS
// ======================================================================
print("\n--- 4. REDUCTIONS & STATS ---");
{
var vA = GPUTensor<Vector>.empty([3]); //[cite: 2]
var vB = GPUTensor<Vector>.empty([3]); //[cite: 2]
var mA = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
vA.pushData([1.0, 2.0, 3.0]); //[cite: 2]
vB.pushData([4.0, 5.0, 6.0]); //[cite: 2]
mA.pushData([1.0, 2.0, 3.0, 4.0]); // [[1, 2], [3, 4]] //[cite: 2]
var tape = CommandBuffer(); //[cite: 5]
var tSumVec = sumGPU(vA, tape); //[cite: 3]
var tSumMat = sumMatrixGPU(mA, tape); //[cite: 3]
var tSumCol = sumReduceColumnsGPU(mA, tape); //[cite: 3]
var tSumRow = sumReduceRowsGPU(mA, tape); //[cite: 3]
var tDot = dotProductGPU(vA, vB, tape); //[cite: 3]
var tL2 = l2NormGPU(
GPUTensor<Vector>([3.0, 4.0]),
tape,
); // Sollte 5 ergeben //[cite: 2, 3]
var tEuc = euclideanDistanceGPU(vA, vB, tape); //[cite: 3]
var tCos = cosineSimilarityGPU(vA, vB, tape); //[cite: 3]
GPUEngine.run(tape.bytes()); //[cite: 1]
tSumVec.toCpu();
tSumMat.toCpu();
tSumCol.toCpu();
tSumRow.toCpu(); //[cite: 2]
tDot.toCpu();
tL2.toCpu();
tEuc.toCpu();
tCos.toCpu(); //[cite: 2]
print("Inputs: vA=[1,2,3], vB=[4,5,6], mA=[[1,2],[3,4]]");
print("SUM(vA): ${tSumVec.value}");
print("SUM(mA): ${tSumMat.value}");
print("SUM_COLS(mA): ${tSumCol.value}");
print("SUM_ROWS(mA): ${tSumRow.value}");
print("DOT(vA, vB): ${tDot.value}");
print("L2NORM([3,4]): ${tL2.value}");
print("EUC_DIST(vA, vB): ${tEuc.value}");
print("COS_SIM(vA, vB): ${tCos.value}");
vA.free();
vB.free();
mA.free();
tSumVec.free();
tSumMat.free();
tSumCol.free();
tSumRow.free();
tDot.free();
tL2.free();
tEuc.free();
tCos.free(); //[cite: 2]
}
// ======================================================================
// 5. LOSS FUNCTIONS
// ======================================================================
print("\n--- 5. LOSS FUNCTIONS ---");
{
var vPred = GPUTensor<Vector>.empty([2]); //[cite: 2]
var vTarg = GPUTensor<Vector>.empty([2]); //[cite: 2]
vPred.pushData([0.8, 0.2]); //[cite: 2]
vTarg.pushData([1.0, 0.0]); //[cite: 2]
var tape = CommandBuffer(); //[cite: 5]
var tBce = binaryCrossEntropyGPU<Vector>(vPred, vTarg, tape); //[cite: 3]
var tMse = mseGPU(vPred, vTarg, tape); //[cite: 3]
var tMae = maeLossGPU(vPred, vTarg, tape); //[cite: 3]
GPUEngine.run(tape.bytes()); //[cite: 1]
tBce.toCpu();
tMse.toCpu();
tMae.toCpu(); //[cite: 2]
print("Inputs: Pred=[0.8, 0.2], Targ=[1.0, 0.0]");
print("BCE_LOSS: ${tBce.value}");
print("MSE_LOSS: ${tMse.value}");
print("MAE_LOSS: ${tMae.value}");
vPred.free();
vTarg.free();
tBce.free();
tMse.free();
tMae.free(); //[cite: 2]
}
// ======================================================================
// 6. SHAPE / TENSOR MANIPULATION
// ======================================================================
print("\n--- 6. SHAPE & TENSOR MANIPULATION ---");
{
var vA = GPUTensor<Vector>.empty([4]); //[cite: 2]
vA.pushData([1.0, 2.0, 3.0, 4.0]); //[cite: 2]
var mA = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
mA.pushData([1.0, 2.0, 3.0, 4.0]); //[cite: 2]
var mB = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
mB.pushData([5.0, 6.0, 7.0, 8.0]); //[cite: 2]
var tape = CommandBuffer(); //[cite: 5]
var tReshape = reshapeVectorToMatrixGPU(vA, 2, 2, tape); //[cite: 3]
var tSliceCol = sliceColumnGPU(mA, 0, 1, tape); //[cite: 3]
var tSelRow = selectRowGPU(mA, 1, tape); //[cite: 3]
var tConcat = concatenateGPU(
vA,
GPUTensor<Vector>([5.0, 6.0]),
tape,
); //[cite: 2, 3]
var tStack = stackMatricesGPU([mA, mB], tape); //[cite: 3]
var tPad = padMatrixGPU(mA, 1, tape); //[cite: 3]
GPUEngine.run(tape.bytes()); //[cite: 1]
tReshape.toCpu();
tSliceCol.toCpu();
tSelRow.toCpu();
tConcat.toCpu();
tStack.toCpu();
tPad.toCpu(); //[cite: 2]
print("RESHAPE(v[1,2,3,4] -> 2x2): ${tReshape.value}");
print("SLICE_COL(mA, 0, 1): ${tSliceCol.value}");
print("SELECT_ROW(mA, 1): ${tSelRow.value}");
print("CONCAT_VEC(vA, [5,6]): ${tConcat.value}");
print("STACK_MAT(mA, mB): ${tStack.value}");
print("PAD_MAT(mA, 1): ${tPad.value}");
vA.free();
mA.free();
mB.free();
tReshape.free();
tSliceCol.free();
tSelRow.free();
tConcat.free();
tStack.free();
tPad.free(); //[cite: 2]
}
// ======================================================================
// 7. ADVANCED LAYERS (Conv, Pool)
// ======================================================================
print("\n--- 7. ADVANCED LAYERS ---");
{
var mIn = GPUTensor<Matrix>.empty([3, 3]); //[cite: 2]
// 3x3 Matrix voller 1er
mIn.pushData([1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]); //[cite: 2]
var mKernel = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
// 2x2 Kernel voller 1er
mKernel.pushData([1.0, 1.0, 1.0, 1.0]); //[cite: 2]
var tape = CommandBuffer(); //[cite: 5]
// Simple 2D Convolution
var tConv = conv2dSimpleGPU(mIn, mKernel, tape); //[cite: 3]
// MaxPool2D (2x2 pool, stride 1)
var tPool = maxPool2dGPU(mIn, 2, 1, tape); //[cite: 3]
GPUEngine.run(tape.bytes()); //[cite: 1]
tConv.toCpu();
tPool.toCpu(); //[cite: 2]
print("Inputs: 3x3 Matrix (ones), 2x2 Kernel (ones)");
print("CONV2D_SIMPLE: ${tConv.value} (Erwartet 2x2 mit Wert 4.0)");
print("MAXPOOL_2D: ${tPool.value}");
mIn.free();
mKernel.free();
tConv.free();
tPool.free(); //[cite: 2]
}
print("\n🧹 Aufräumen...");
GPUEngine.dispose(); //[cite: 1]
print("✅ Alle Funktionen gefeuert und validiert!");
return true;
}
Future<bool> testLinear() async {
print("🚀 Initialisiere chirurgischen Linear-Algebra-Debugger...");
await GPUEngine.initialize(
debug: false,
target: target,
); //[cite: 1]
// Basis-Tensoren anlegen
var mA = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
var mB = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
var vC = GPUTensor<Vector>.empty([2]); //[cite: 2]
var sD = GPUTensor<Scalar>(
5.0,
); // Scalar für addScalarMatrixGPU //[cite: 2]
mA.pushData([1.0, 2.0, 3.0, 4.0]); //[cite: 2]
mB.pushData([2.0, 0.0, 1.0, 2.0]); //[cite: 2]
vC.pushData([10.0, 20.0]); //[cite: 2]
print("\n--- TEST 1: MATMUL ---");
var tape1 = CommandBuffer(); //[cite: 5]
var tMatMul = matMulGPU(mA, mB, tape1); //[cite: 3]
print("Feuere MatMul...");
GPUEngine.run(tape1.bytes()); //[cite: 1]
tMatMul.toCpu(); //[cite: 2]
print("✅ MATMUL OK: ${tMatMul.value}");
print("\n--- TEST 2: MATVEC ---");
var tape2 = CommandBuffer(); //[cite: 5]
var tMatVec = matVecMulGPU(mA, vC, tape2); //[cite: 3]
print("Feuere MatVec...");
GPUEngine.run(tape2.bytes()); //[cite: 1]
tMatVec.toCpu(); //[cite: 2]
print("✅ MATVEC OK: ${tMatVec.value}");
print("\n--- TEST 3: TRANSPOSE ---");
var tape3 = CommandBuffer(); //[cite: 5]
var tTrans = transposeGPU(mA, tape3); //[cite: 3]
print("Feuere Transpose...");
GPUEngine.run(tape3.bytes()); //[cite: 1]
tTrans.toCpu(); //[cite: 2]
print("✅ TRANSPOSE OK: ${tTrans.value}");
print("\n--- TEST 4: SCALE MATRIX ---");
var tape4 = CommandBuffer(); //[cite: 5]
var tScale = scaleMatrixGPU(mA, 10.0, tape4); //[cite: 3]
print("Feuere Scale Matrix...");
GPUEngine.run(tape4.bytes()); //[cite: 1]
tScale.toCpu(); //[cite: 2]
print("✅ SCALE MATRIX OK: ${tScale.value}");
print("\n--- TEST 5: ADD BIAS TO MATMUL (Nutzt OP_BROADCAST_ADD) ---");
var tape5 = CommandBuffer(); //[cite: 5]
var tAddBi = addBiasToMatMulOutGPU(mA, vC, tape5); //[cite: 3]
print("Feuere Add Bias...");
GPUEngine.run(tape5.bytes()); //[cite: 1]
tAddBi.toCpu(); //[cite: 2]
print("✅ ADD BIAS OK: ${tAddBi.value}");
print("\n--- TEST 6: ADD SCALAR MATRIX (Nutzt OP_BROADCAST_ADD) ---");
var tape6 = CommandBuffer(); //[cite: 5]
var tAddSca = addScalarMatrixGPU(mA, sD, tape6); //[cite: 3]
print("Feuere Add Scalar Matrix...");
GPUEngine.run(tape6.bytes()); //[cite: 1]
tAddSca.toCpu(); //[cite: 2]
print("✅ ADD SCALAR MATRIX OK: ${tAddSca.value}");
print("\n🧹 Räume auf...");
mA.free();
mB.free();
vC.free();
sD.free(); //[cite: 2]
tMatMul.free();
tMatVec.free();
tTrans.free();
tScale.free();
tAddBi.free();
tAddSca.free(); //[cite: 2]
GPUEngine.dispose(); //[cite: 1]
print("🎉 Alle Tests überlebt!");
return true;
}
Future<bool> testBackward() async{
print("🚀 Starte massive BACKWARD (Gradient) Validierung...");
await GPUEngine.initialize(debug: false, target: target); //[cite: 1]
// =========================================================
// 1. BASIC MATH BACKWARDS
// =========================================================
print("\n--- 1. BASIC MATH (ADD, SUB, MUL, DIV) ---");
{
var tape = CommandBuffer(); //[cite: 5]
// ADD: Ableitung von (A + B) nach A ist 1, nach B ist 1
var aAdd = GPUTensor<Vector>.empty([2]); aAdd.pushData([2.0, 3.0]); //[cite: 2]
var bAdd = GPUTensor<Vector>.empty([2]); bAdd.pushData([4.0, 5.0]); //[cite: 2]
var outAdd = addGPU<Vector>(aAdd, bAdd, tape); //[cite: 3]
outAdd.backward(tape, fillOnes: true); //[cite: 2]
// SUB: Ableitung von (A - B) nach A ist 1, nach B ist -1
var aSub = GPUTensor<Vector>.empty([2]); aSub.pushData([2.0, 3.0]); //[cite: 2]
var bSub = GPUTensor<Vector>.empty([2]); bSub.pushData([4.0, 5.0]); //[cite: 2]
var outSub = subtractGPU<Vector>(aSub, bSub, tape); //[cite: 3]
outSub.backward(tape, fillOnes: true); //[cite: 2]
// MUL: Ableitung von (A * B) nach A ist B, nach B ist A
var aMul = GPUTensor<Vector>.empty([2]); aMul.pushData([2.0, 3.0]); //[cite: 2]
var bMul = GPUTensor<Vector>.empty([2]); bMul.pushData([4.0, 5.0]); //[cite: 2]
var outMul = multiplyGPU<Vector>(aMul, bMul, tape); //[cite: 3]
outMul.backward(tape, fillOnes: true); //[cite: 2]
// DIV: Ableitung von (A / B) nach A ist 1/B, nach B ist -A/B^2
var aDiv = GPUTensor<Vector>.empty([2]); aDiv.pushData([2.0, 3.0]); //[cite: 2]
var bDiv = GPUTensor<Vector>.empty([2]); bDiv.pushData([4.0, 5.0]); //[cite: 2]
var outDiv = divideGPU<Vector>(aDiv, bDiv, tape); //[cite: 3]
outDiv.backward(tape, fillOnes: true); //[cite: 2]
GPUEngine.run(tape.bytes()); //[cite: 1]
// Gradienten (.grad) zusammen mit den Daten in den RAM holen
aAdd.toCpu(); bAdd.toCpu(); //[cite: 2]
aSub.toCpu(); bSub.toCpu(); //[cite: 2]
aMul.toCpu(); bMul.toCpu(); //[cite: 2]
aDiv.toCpu(); bDiv.toCpu(); //[cite: 2]
print("Inputs: A=[2,3], B=[4,5]");
print("ADD Grad A: ${aAdd.gradValue} (Erwartet: [1.0, 1.0])"); //[cite: 2]
print("ADD Grad B: ${bAdd.gradValue} (Erwartet: [1.0, 1.0])"); //[cite: 2]
print("SUB Grad A: ${aSub.gradValue} (Erwartet: [1.0, 1.0])"); //[cite: 2]
print("SUB Grad B: ${bSub.gradValue} (Erwartet: [-1.0, -1.0])"); //[cite: 2]
print("MUL Grad A: ${aMul.gradValue} (Erwartet: [4.0, 5.0])"); //[cite: 2]
print("MUL Grad B: ${bMul.gradValue} (Erwartet: [2.0, 3.0])"); //[cite: 2]
print("DIV Grad A: ${aDiv.gradValue} (Erwartet: [0.25, 0.2])"); //[cite: 2]
print("DIV Grad B: ${bDiv.gradValue} (Erwartet: [-0.125, -0.12])"); //[cite: 2]
aAdd.free(); bAdd.free(); outAdd.free(); //[cite: 2]
aSub.free(); bSub.free(); outSub.free(); //[cite: 2]
aMul.free(); bMul.free(); outMul.free(); //[cite: 2]
aDiv.free(); bDiv.free(); outDiv.free(); //[cite: 2]
}
// =========================================================
// 2. UNARY & ACTIVATIONS
// =========================================================
print("\n--- 2. UNARY & ACTIVATIONS ---");
{
var tape = CommandBuffer(); //[cite: 5]
// POW: Ableitung von x^2 ist 2*x
var aPow = GPUTensor<Vector>.empty([2]); aPow.pushData([2.0, 4.0]); //[cite: 2]
var outPow = powGPU<Vector>(aPow, 2.0, tape); //[cite: 3]
outPow.backward(tape, fillOnes: true); //[cite: 2]
// LOG: Ableitung von ln(x) ist 1/x
var aLog = GPUTensor<Vector>.empty([2]); aLog.pushData([2.0, 4.0]); //[cite: 2]
var outLog = logGPU<Vector>(aLog, tape); //[cite: 3]
outLog.backward(tape, fillOnes: true); //[cite: 2]
// RELU: Ableitung von max(0, x) ist 1 wenn x > 0, sonst 0
var aRelu = GPUTensor<Vector>.empty([2]); aRelu.pushData([2.0, -1.0]); //[cite: 2]
var outRelu = reluGPU(aRelu, tape); //[cite: 3]
outRelu.backward(tape, fillOnes: true); //[cite: 2]
GPUEngine.run(tape.bytes()); //[cite: 1]
aPow.toCpu(); aLog.toCpu(); aRelu.toCpu(); //[cite: 2]
print("POW(A=[2,4], 2) Grad: ${aPow.gradValue} (Erwartet: [4.0, 8.0])"); //[cite: 2]
print("LOG(A=[2,4]) Grad: ${aLog.gradValue} (Erwartet: [0.5, 0.25])"); //[cite: 2]
print("RELU(A=[2,-1]) Grad: ${aRelu.gradValue} (Erwartet: [1.0, 0.0])"); //[cite: 2]
aPow.free(); outPow.free(); //[cite: 2]
aLog.free(); outLog.free(); //[cite: 2]
aRelu.free(); outRelu.free(); //[cite: 2]
}
// =========================================================
// 3. MATMUL & LINEAR ALGEBRA
// =========================================================
print("\n--- 3. MATMUL & LINEAR ALGEBRA ---");
{
var tape = CommandBuffer(); //[cite: 5]
// C = A * B. Grad(A) = Grad(C) * B^T. Grad(B) = A^T * Grad(C)
var mA = GPUTensor<Matrix>.empty([1, 2]); mA.pushData([1.0, 2.0]); //[cite: 2]
var mB = GPUTensor<Matrix>.empty([2, 1]); mB.pushData([3.0, 4.0]); //[cite: 2]
var outMatMul = matMulGPU(mA, mB, tape); //[cite: 3]
outMatMul.backward(tape, fillOnes: true); //[cite: 2]
GPUEngine.run(tape.bytes()); //[cite: 1]
mA.toCpu(); mB.toCpu(); //[cite: 2]
print("MATMUL A=[[1,2]], B=[[3],[4]]");
print("Grad A: ${mA.gradValue} (Erwartet: [[3.0, 4.0]])"); //[cite: 2]
print("Grad B: ${mB.gradValue} (Erwartet: [[1.0], [2.0]])"); //[cite: 2]
mA.free(); mB.free(); outMatMul.free(); //[cite: 2]
}
// =========================================================
// 4. LOSSES & REDUCTIONS
// =========================================================
print("\n--- 4. LOSSES & REDUCTIONS ---");
{
var tape = CommandBuffer(); //[cite: 5]
// MSE Loss: (Pred - Target) * 2 / N
var vPred = GPUTensor<Vector>.empty([2]); vPred.pushData([0.8, 0.2]); //[cite: 2]
var vTarg = GPUTensor<Vector>.empty([2]); vTarg.pushData([1.0, 0.0]); //[cite: 2]
var outMse = mseGPU(vPred, vTarg, tape); //[cite: 3]
outMse.backward(tape, fillOnes: true); //[cite: 2]
// Sum Reduction: Gradient verteilt sich gleichmäßig als 1.0
var vSum = GPUTensor<Vector>.empty([3]); vSum.pushData([1.0, 2.0, 3.0]); //[cite: 2]
var outSum = sumGPU(vSum, tape); //[cite: 3]
outSum.backward(tape, fillOnes: true); //[cite: 2]
GPUEngine.run(tape.bytes()); //[cite: 1]
vPred.toCpu(); vSum.toCpu(); //[cite: 2]
print("MSE Pred=[0.8, 0.2], Targ=[1.0, 0.0]");
print("Grad Pred: ${vPred.gradValue} (Erwartet: [-0.2, 0.2])"); //[cite: 2]
print("SUM V=[1,2,3]");
print("Grad V: ${vSum.gradValue} (Erwartet: [1.0, 1.0, 1.0])"); //[cite: 2]
vPred.free(); vTarg.free(); outMse.free(); //[cite: 2]
vSum.free(); outSum.free(); //[cite: 2]
}
// =========================================================
// 5. CONVOLUTION 2D
// =========================================================
print("\n--- 5. CONVOLUTION 2D ---");
{
var tape = CommandBuffer(); //[cite: 5]
// Faltung eines 3x3 Inputs mit einem 2x2 Kernel.
// Der Gradient des Kernels ist die Summe der überlappenden Fenster.
var mIn = GPUTensor<Matrix>.empty([3, 3]); //[cite: 2]
mIn.pushData([1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0]); //[cite: 2]
var mKer = GPUTensor<Matrix>.empty([2, 2]); //[cite: 2]
mKer.pushData([1.0, 1.0, 1.0, 1.0]); //[cite: 2]
var outConv = conv2dSimpleGPU(mIn, mKer, tape); //[cite: 3]
outConv.backward(tape, fillOnes: true); //[cite: 2]
GPUEngine.run(tape.bytes()); //[cite: 1]
mKer.toCpu(); //[cite: 2]
print("CONV2D In=3x3(1..9), Ker=2x2(1s)");
print("Grad Kernel: ${mKer.gradValue} (Erwartet: [[12.0, 16.0], [24.0, 28.0]])"); //[cite: 2]
mIn.free(); mKer.free(); outConv.free(); //[cite: 2]
}
print("\n🧹 Aufräumen...");
GPUEngine.dispose(); //[cite: 1]
print("✅ Backward Validierung abgeschlossen!");
return true;
}
Future<bool> testSpeeds()async{
print("🚀 Starte TFLOP & Overhead Benchmark...");
// Target anpassen (z.B. Target.cuda oder Target.android_x86_64)
await GPUEngine.initialize(debug: false, target: target); //
// ====================================================================
// TEST 1: FFI & DISPATCH OVERHEAD
// ====================================================================
print("\n=======================================================");
print("⏱️ TEST 1: ENGINE DISPATCH OVERHEAD");
print("=======================================================");
// Ein winziger Tensor (1 Element), um die GPU-Rechenzeit gegen 0 zu drücken
var tinyVec = GPUTensor<Vector>.empty([1]); //[cite: 2]
var tapeOverhead = CommandBuffer(); //[cite: 5]
// Nur ein simpler Fill-Befehl
tapeOverhead.putInt(OP_FILL); //[cite: 2, 5]
tapeOverhead.putString(tinyVec.id); //[cite: 2, 5]
tapeOverhead.putFloat(1.0); //[cite: 2, 5]
Uint8List compiledOverhead = tapeOverhead.bytes(); //[cite: 5]
// Warmup (Weckt die Engine und füllt Caches)
for (int i = 0; i < 100; i++) {
GPUEngine.run(compiledOverhead); //[cite: 1]
}
int overheadIterations = 10000;
Stopwatch swOverhead = Stopwatch()..start();
for (int i = 0; i < overheadIterations; i++) {
GPUEngine.run(compiledOverhead); //[cite: 1]
}
swOverhead.stop();
double totalMs = swOverhead.elapsedMicroseconds / 1000.0;
double msPerDispatch = totalMs / overheadIterations;
print("Befehle gesendet: $overheadIterations");
print("Gesamtzeit: ${totalMs.toStringAsFixed(2)} ms");
print("Overhead pro Run: ${(msPerDispatch * 1000).toStringAsFixed(2)} Mikrosekunden (µs) pro Dispatch");
tinyVec.free(); //[cite: 2]
// ====================================================================
// TEST 2: MAXIMAL TFLOPS (Matrix Multiplication)
// ====================================================================
print("\n=======================================================");
print("🔥 TEST 2: MAXIMAL TFLOPS (MATRIX MULTIPLICATION)");
print("=======================================================");
// Matrix-Dimension: 4096 x 4096
// Belegt ca. 3x 67 MB VRAM (A, B, und Out) = ~201 MB total
int M = 512*2;
print("Matrix Größe: ${M}x$M Elemente");
var matA = GPUTensor<Matrix>.empty([M, M]); //[cite: 2]
var matB = GPUTensor<Matrix>.empty([M, M]); //[cite: 2]
// Wir füllen sie auf der GPU, damit wir keine Zeit mit CPU->GPU Transfers verlieren
var setupTape = CommandBuffer(); //[cite: 5]
setupTape.putInt(OP_FILL); setupTape.putString(matA.id); setupTape.putFloat(1.0); //[cite: 2, 5]
setupTape.putInt(OP_FILL); setupTape.putString(matB.id); setupTape.putFloat(1.0); //[cite: 2, 5]
GPUEngine.run(setupTape.bytes()); //[cite: 1]
var tflopTape = CommandBuffer(); //[cite: 5]
// Das ist unsere schwere Last
var matOut = matMulGPU(matA, matB, tflopTape); //[cite: 3]
Uint8List compiledTflop = tflopTape.bytes(); //[cite: 5]
// Formel für MatMul Flops: 2 * M * K * N
// Da M = K = N = 4096:
double flopsPerRun = 2.0 * M * M * M;
double tflopsPerRun = flopsPerRun / 1e12; // In TeraFLOPs umrechnen
// Warmup
GPUEngine.run(compiledTflop); //[cite: 1]
int tflopIterations = 20;
print("Starte $tflopIterations Durchläufe...");
Stopwatch swTflop = Stopwatch()..start();
for (int i = 0; i < tflopIterations; i++) {
GPUEngine.run(compiledTflop); //[cite: 1]
}
swTflop.stop();
double seconds = swTflop.elapsedMicroseconds / 1000000.0;
double totalTflops = tflopsPerRun * tflopIterations;
double tflopsPerSecond = totalTflops / seconds;
double gflopsPerSecond = tflopsPerSecond * 1000.0;
print("Gesamtzeit: ${seconds.toStringAsFixed(3)} s");
print("Berechnete TFLOPs: ${totalTflops.toStringAsFixed(3)} TFLOPs total");
print("-------------------------------------------------------");
print("Leistung (GFLOP/s): ${gflopsPerSecond.toStringAsFixed(2)} GFLOP/s");
print("Leistung (TFLOP/s): ${tflopsPerSecond.toStringAsFixed(4)} TFLOP/s");
print("=======================================================\n");
matA.free(); matB.free(); matOut.free(); //[cite: 2]
GPUEngine.dispose(); //[cite: 1]
return true;
}
Future<bool> testLearning2() async {
print("🚀 Initialisiere GPU Engine für Forward-Pass-Test...");
await GPUEngine.initialize(debug: false, target:target);
SNetworkGPU net = SNetworkGPU();
net.add(DenseTL(8));
net.add(ReLULayerMatrixTapeLayer());
net.add(DenseTL(1));
net.add(SigmoidMatrixTL());
// Kompilieren (allokiert VRAM und bereitet das Forward-Tape vor)
print("⚙️ Kompiliere Graphen...");
net.compile([4, 2], [4, 1], 0.001, useAdam: true);
// Unser Test-Batch
List<double> xBatch = [
0.0, 0.0,
0.0, 1.0,
1.0, 0.0,
1.0, 1.0
];
print("🔮 Führe Forward-Pass (predict) aus...");
// predict() schiebt xBatch in den VRAM, triggert fTape und holt das Ergebnis in den RAM
var prediction = net.predict(xBatch);
print("\n📊 Rohe Vorhersagen (Untrainiert, basierend auf Zufallsgewichten):");
print(prediction);
// Optional: Wir können auch direkt prüfen, ob die Gewichte ausgelesen werden können
print("\n⚖️ Lade initiale Gewichte aus Layer 1 (DenseTL) herunter...");
var initialWeights = net.layers[0].getWeights();
print(initialWeights['weights']);
print("\n🧹 Aufräumen...");
net.free();
GPUEngine.dispose();
return true;
}
Future<bool> testBackwardPass() async {
print("🚀 Initialisiere GPU Engine für Backward-Pass-Test...");
await GPUEngine.initialize(debug: false, target: target);
SNetworkGPU net = SNetworkGPU();
net.add(DenseTL(8));
net.add(ReLULayerMatrixTapeLayer());
net.add(DenseTL(1));
net.add(SigmoidMatrixTL());
// Wir nutzen eine sehr hohe Lernrate (0.5), damit wir den Sprung der Gewichte deutlich sehen
net.compile([4, 2], [4, 1], 0.5, useAdam: true);
List<double> xBatch = [
0.0, 0.0,
0.0, 1.0,
1.0, 0.0,
1.0, 1.0
];
List<double> yBatch = [0.0, 1.0, 1.0, 0.0];
// 1. Gewichte VOR dem Gradienten-Update lesen
print("\n⚖️ Initiale Gewichte in Layer 1 (Zeile 1):");
var wBefore = net.layers[0].getWeights()['weights'] as List;
print(wBefore[0]);
// 2. Ein einziger Durchlauf (Forward -> Backward -> Adam Update)
print("\n🔮 Führe Backpropagation aus...");
double loss = net.trainStep(xBatch, yBatch);
print("📊 Initialer Loss: $loss");
// 3. Gewichte NACH dem Gradienten-Update lesen
print("\n📉 Gewichte NACH dem Update (Zeile 1):");
var wAfter = net.layers[0].getWeights()['weights'] as List;
print(wAfter[0]);
// 4. Prüfen, ob sich etwas bewegt hat
bool didLearn = false;
for (int i = 0; i < (wBefore[0] as List).length; i++) {
if (wBefore[0][i] != wAfter[0][i]) {
didLearn = true;
break;
}
}
print(didLearn ? "✅ Gradienten fließen! Gewichte wurden erfolgreich geupdatet." : "❌ FEHLER: Gewichte sind eingefroren (Gradient ist 0.0).");
net.free();
GPUEngine.dispose();
return true;
}
Future<bool> testShortTraining() async {
print("🚀 Initialisiere GPU Engine für Kurz-Training...");
await GPUEngine.initialize(debug: false, target: target);
SNetworkGPU net = SNetworkGPU();
net.add(DenseTL(16)); // 16 Neuronen, um das 0.69-Minimum sicher zu knacken
net.add(ReLULayerMatrixTapeLayer());
net.add(DenseTL(1));
net.add(SigmoidMatrixTL());
// Kompilieren mit Adam und einer guten Lernrate
net.compile([4, 2], [4, 1], 0.05, useAdam: true);
List<double> xBatch = [
0.0, 0.0,
0.0, 1.0,
1.0, 0.0,
1.0, 1.0
];
List<double> yBatch = [0.0, 1.0, 1.0, 0.0];
print("\n🔮 Starte Training für 300 Epochen...");
for (int epoch = 1; epoch <= 300; epoch++) {
double loss = net.trainStep(xBatch, yBatch);
if (epoch % 50 == 0 || epoch == 1) {
print("Epoch ${epoch.toString().padLeft(3)} | Loss: ${loss.toStringAsFixed(6)}");
}
}
print("\n📊 Vorhersagen NACH dem Training (Ziel: 0, 1, 1, 0):");
var prediction = net.predict(xBatch);
// Etwas schöner formatiert ausgeben
List<List<double>> preds = prediction as List<List<double>>;
print("Input [0,0] -> ${preds[0][0].toStringAsFixed(4)} (Ziel: 0.0)");
print("Input [0,1] -> ${preds[1][0].toStringAsFixed(4)} (Ziel: 1.0)");
print("Input [1,0] -> ${preds[2][0].toStringAsFixed(4)} (Ziel: 1.0)");
print("Input [1,1] -> ${preds[3][0].toStringAsFixed(4)} (Ziel: 0.0)");
print("\n🧹 Aufräumen...");
net.free();
GPUEngine.dispose();
return true;
}
Future<bool> testConvergence() async {
print("🚀 Initialisiere GPU Engine für Konvergenz-Test...");
await GPUEngine.initialize(debug: false, target: target);
SNetworkGPU net = SNetworkGPU();
net.add(DenseTL(16));
net.add(ReLULayerMatrixTapeLayer());
net.add(DenseTL(1));
net.add(SigmoidMatrixTL());
net.compile([4, 2], [4, 1], 0.05, useAdam: true);
List<double> xBatch = [
0.0, 0.0,
0.0, 1.0,
1.0, 0.0,
1.0, 1.0
];
List<double> yBatch = [0.0, 1.0, 1.0, 0.0];
print("\n🔍 Beobachte Vorhersagen und Loss über 50 Epochen...");
for (int epoch = 1; epoch <= 50; epoch++) {
// 1. Trainieren (Forward, Backward, Optimize)
double loss = net.trainStep(xBatch, yBatch);
// 2. Aktuellen Stand der Vorhersagen abholen
var prediction = net.predict(xBatch);
List<List<double>> preds = prediction as List<List<double>>;
// 3. Ein einzelnes Gewicht auslesen, um die Bewegung zu tracken
var weights = net.layers[0].getWeights()['weights'] as List;
double sampleWeight = weights[0][0];
// Alle 10 Epochen (und beim ersten Mal) loggen
if (epoch % 10 == 0 || epoch == 1) {
print("\n--- Epoche ${epoch.toString().padLeft(2)} ---");
print("Loss: ${loss.toStringAsFixed(6)}");
print("Gewicht 0: ${sampleWeight.toStringAsFixed(6)}");
print("Preds: [${preds[0][0].toStringAsFixed(4)}, ${preds[1][0].toStringAsFixed(4)}, ${preds[2][0].toStringAsFixed(4)}, ${preds[3][0].toStringAsFixed(4)}]");
}
}
print("\n🧹 Aufräumen...");
net.free();
GPUEngine.dispose();
return true;
}
Future<bool> testTrainingAndPrintWeights() async {
print("🚀 Initialisiere GPU Engine...");
await GPUEngine.initialize(debug: false, target: target);
SNetworkGPU net = SNetworkGPU();
net.add(DenseTL(16));
// ✅ DER LEBENSRETTER: Sigmoid stirbt niemals!
// Es liefert auch bei negativen Werten immer einen kleinen Gradienten zurück,
// sodass Adam weiterarbeiten kann und das Netzwerk nicht einfriert.
net.add(SigmoidMatrixTL());
net.add(DenseTL(1));
net.add(SigmoidMatrixTL());
// Wir können die Lernrate jetzt mutig auf 0.05 setzen, da Sigmoid robuster ist
net.compile([4, 2], [4, 1], 0.05, useAdam: true);
List<double> xBatch = [
0.0, 0.0,
0.0, 1.0,
1.0, 0.0,
1.0, 1.0
];
List<double> yBatch = [0.0, 1.0, 1.0, 0.0];
print("\n🔥 Starte Training für 2000 Epochen...");
for (int epoch = 1; epoch <= 2000; epoch++) {
double loss = net.trainStep(xBatch, yBatch);
if (epoch % 200 == 0 || epoch == 1) {
print("Epoch ${epoch.toString().padLeft(4)} | Loss: ${loss.toStringAsFixed(6)}");
}
}
print("\n🧠 Vorhersagen nach 2000 Epochen (Ziel: 0, 1, 1, 0):");
var prediction = net.predict(xBatch);
List<List<double>> preds = prediction as List<List<double>>;
print("Input [0,0] -> ${preds[0][0].toStringAsFixed(4)}");
print("Input [0,1] -> ${preds[1][0].toStringAsFixed(4)}");
print("Input [1,0] -> ${preds[2][0].toStringAsFixed(4)}");
print("Input [1,1] -> ${preds[3][0].toStringAsFixed(4)}");
print("\n🧹 Aufräumen...");
net.free();
GPUEngine.dispose();
return true;
}
Future<bool> debugAllGradients() async {
print("🚀 Initialisiere GPU Engine für GRADIENTEN-RÖNTGEN...");
await GPUEngine.initialize(debug: false, target: target);
SNetworkGPU net = SNetworkGPU();
net.add(DenseTL(4)); // 4 Neuronen = Übersichtlicher Output!
net.add(ReLULayerMatrixTapeLayer());
net.add(DenseTL(1));
net.add(SigmoidMatrixTL());
// 🛑 WICHTIG: SGD mit Lernrate 1.0 !
// Dadurch gilt: Gradient = Gewicht_Vorher - Gewicht_Nachher
net.compile([4, 2], [4, 1], 1.0, useAdam: false);
List<double> xBatch = [
0.0, 0.0,
0.0, 1.0,
1.0, 0.0,
1.0, 1.0
];
List<double> yBatch = [0.0, 1.0, 1.0, 0.0];
// 1. Snapshot: Gewichte VOR dem Forward-Pass aus dem VRAM holen
var l1Before = net.layers[0].getWeights()['weights'] as List;
var l2Before = net.layers[2].getWeights()['weights'] as List;
// 2. Aktion: EIN einziger Durchlauf (Forward + Backward + Optimize)
print("\n🔥 Führe exakt 1 Backpropagation-Schritt aus...");
double loss = net.trainStep(xBatch, yBatch);
print("📊 Berechneter Loss: $loss");
// 3. Snapshot: Gewichte NACH dem Update aus dem VRAM holen
var l1After = net.layers[0].getWeights()['weights'] as List;
var l2After = net.layers[2].getWeights()['weights'] as List;
// 4. Analyse: Berechnen und Printen!
print("\n" + "="*60);
print("🔍 GRADIENTEN-ANALYSE (LAYER 1 - Hidden Layer)");
print("="*60);
for (int i = 0; i < l1Before.length; i++) {
List<double> rowBefore = (l1Before[i] as List).cast<double>();
List<double> rowAfter = (l1After[i] as List).cast<double>();
print("Neuron $i:");
for (int j = 0; j < rowBefore.length; j++) {
double delta = rowAfter[j] - rowBefore[j];
double grad = -delta; // W_alt - W_neu
String marker = (grad == 0.0) ? " 💀 DEAD" : " ✅ ALIVE";
print(" Weight $j | Vorher: ${rowBefore[j].toStringAsFixed(6).padLeft(9)} | Nachher: ${rowAfter[j].toStringAsFixed(6).padLeft(9)} | Gradient: ${grad.toStringAsFixed(8)}$marker");
}
}
print("\n" + "="*60);
print("🔍 GRADIENTEN-ANALYSE (LAYER 2 - Output Layer)");
print("="*60);
for (int i = 0; i < l2Before.length; i++) {
List<double> rowBefore = (l2Before[i] as List).cast<double>();
List<double> rowAfter = (l2After[i] as List).cast<double>();
print("Output-Verbindung Zeile $i:");
for (int j = 0; j < rowBefore.length; j++) {
double delta = rowAfter[j] - rowBefore[j];
double grad = -delta; // W_alt - W_neu
String marker = (grad == 0.0) ? " 💀 DEAD" : " ✅ ALIVE";
print(" Weight $j | Vorher: ${rowBefore[j].toStringAsFixed(6).padLeft(9)} | Nachher: ${rowAfter[j].toStringAsFixed(6).padLeft(9)} | Gradient: ${grad.toStringAsFixed(8)}$marker");
}
}
print("\n🧹 Aufräumen...");
net.free();
GPUEngine.dispose();
return true;
}
Future<bool> extractRawGradients() async {
print("🚀 Initialisiere GPU Engine für ROH-GRADIENTEN-EXTRAKTION...");
await GPUEngine.initialize(debug: false, target: target);
SNetworkGPU net = SNetworkGPU();
net.add(DenseTL(4));
net.add(SigmoidMatrixTL());
net.add(DenseTL(1));
net.add(SigmoidMatrixTL());
net.compile([4, 2], [4, 1], 1.0, useAdam: false);
List<double> xBatch = [
0.0, 0.0,
0.0, 1.0,
1.0, 0.0,
1.0, 1.0
];
List<double> yBatch = [0.0, 1.0, 1.0, 0.0];
net.inputRef.pushData(xBatch);
net.targetRef.pushData(yBatch);
print("\n🔥 Führe Forward und Backward Pass aus (OHNE Optimizer)...");
GPUEngine.run(net.zTape);
GPUEngine.run(net.fTape);
GPUEngine.run(net.bTape);
print("\n" + "="*60);
print("🔍 ROH-GRADIENTEN DIREKT AUS DEM C++ VRAM (via toCpu())");
print("="*60);
for (int i = 0; i < net.allParams.length; i++) {
GPUTensor param = net.allParams[i];
// ✅ KORREKTUR: Wir nutzen sauber die integrierte toCpu() Methode der Tensor-Klasse[cite: 3]
param.toCpu();
// Sobald toCpu() aufgerufen wurde, ist die .grad Liste des Tensors direkt befüllt[cite: 3]
var gradData = param.grad;
print("\n--- Parameter $i (Shape: ${param.shape}, ID: ${param.id}) ---");
print("Rohe Gradienten im Speicher:");
print(gradData);
}
print("\n🧹 Aufräumen...");
net.free();
GPUEngine.dispose();
return true;
}
}
*/
void main() async {
WidgetsFlutterBinding.ensureInitialized();
/// Adjust for your architecture
await GPUEngine.initialize(target: Target.cuda);
runApp(MaterialApp(home: BenchmarkScreen()));
}
class BenchmarkScreen extends StatefulWidget {
@override
State<BenchmarkScreen> createState() => _BenchmarkScreenState();
}
class _BenchmarkScreenState extends State<BenchmarkScreen> {
String status = 'Ready to benchmark';
bool isRunning = false;
void runBenchmarkSuite() async {
setState(() {
isRunning = true;
status = 'Running benchmarks...';
});
await Future<void>.delayed(Duration(milliseconds: 100));
// 1. Training and Static Prediction
print('--- Test 1: Training and Static Prediction ---');
GPUTensor<Matrix> trainInput1 = GPUTensor<Matrix>(<List<double>>[
<double>[1.0, 1.0],
<double>[2.0, 2.0],
]);
GPUTensor<Matrix> trainTarget1 = GPUTensor<Matrix>(<List<double>>[
<double>[2.0, 2.0],
<double>[4.0, 4.0],
]);
SeqModel<Matrix, Matrix> model1 = SeqModel<Matrix, Matrix>(
<TapeLayer>[DenseTL(2)],
trainInput1,
target: trainTarget1,
lossFunction: mseMatrixGPU,
optimizerBuilder: (List<GPUTensor> params) {
return SGDGPU(params, 0.01);
},
);
model1.compile();
int epochs1 = 1000;
for (int i = 0; i < epochs1; i = i + 1) {
model1.runTraining();
}
model1.loss?.toCpu();
print('Final Training Loss: ${model1.loss?.value}');
GPUTensor<Matrix> inferInput1 = GPUTensor<Matrix>(<List<double>>[
<double>[3.0, 3.0],
]);
GPUTensor<Matrix> inferResult1 = model1.predict(inferInput1);
model1.runForward();
inferResult1.toCpu();
print('Prediction for [3.0, 3.0]: ${inferResult1.value}');
List<double> newInferData = <double>[4.0, 4.0];
model1.runForward(inputData: newInferData);
inferResult1.toCpu();
print('Prediction for [4.0, 4.0]: ${inferResult1.value}');
model1.free();
// 2. GPU Engine Benchmark - Deep Network Static Unrolling
print('--- Test 2: Deep Network Static Unrolling ---');
int batchSize2 = 1024;
int inputFeatures2 = 64 * 64;
int outputFeatures2 = 16;
Random rnd2 = Random(42);
List<List<double>> rawInput2 = <List<double>>[];
List<List<double>> rawTarget2 = <List<double>>[];
for (int i = 0; i < batchSize2; i = i + 1) {
List<double> inRow = <double>[];
for (int j = 0; j < inputFeatures2; j = j + 1) {
inRow.add(rnd2.nextDouble());
}
rawInput2.add(inRow);
List<double> outRow = <double>[];
for (int j = 0; j < outputFeatures2; j = j + 1) {
outRow.add(rnd2.nextDouble());
}
rawTarget2.add(outRow);
}
GPUTensor<Matrix> trainInput2 = GPUTensor<Matrix>(rawInput2);
GPUTensor<Matrix> trainTarget2 = GPUTensor<Matrix>(rawTarget2);
SeqModel<Matrix, Matrix> model2 = SeqModel<Matrix, Matrix>(
<TapeLayer>[
DenseTL(128),
GeluLayerMatrixTL(),
DenseTL(64),
GeluLayerMatrixTL(),
DenseTL(outputFeatures2),
],
trainInput2,
target: trainTarget2,
lossFunction: mseMatrixGPU,
optimizerBuilder: (List<GPUTensor> params) {
return AdamGPU(params, 0.001);
},
);
Stopwatch timer2 = Stopwatch();
print('Compiling static tapes...');
timer2.start();
model2.compile();
timer2.stop();
print('Compilation took: ${timer2.elapsedMilliseconds} ms');
print(
'Starting high-speed training loop (20 seconds max) for Batch Size $batchSize2...',
);
timer2.reset();
timer2.start();
int actualEpochs2 = 0;
while (timer2.elapsedMilliseconds < 20000) {
model2.runTraining();
actualEpochs2 = actualEpochs2 + 1;
}
timer2.stop();
model2.loss?.toCpu();
double seconds2 = timer2.elapsedMilliseconds / 1000.0;
double epochsPerSec2 = actualEpochs2 / seconds2;
double dense1Flops2 = batchSize2 * 128 * inputFeatures2 * 2.0;
double dense2Flops2 = batchSize2 * 64 * 128 * 2.0;
double dense3Flops2 = batchSize2 * outputFeatures2 * 64 * 2.0;
double forwardFlops2 = dense1Flops2 + dense2Flops2 + dense3Flops2;
double totalFlopsPerEpoch2 = forwardFlops2 * 3.0;
double totalFlopsAchieved2 = totalFlopsPerEpoch2 * actualEpochs2;
double flopsPerSecond2 = totalFlopsAchieved2 / seconds2;
double gflops2 = flopsPerSecond2 / 1000000000.0;
double tflops2 = gflops2 / 1000.0;
print('Final Training Loss: ${model2.loss?.value}');
print(
'Completed $actualEpochs2 epochs in ${seconds2.toStringAsFixed(2)} seconds',
);
print('Engine Performance: ${epochsPerSec2.toStringAsFixed(2)} Epochs/sec');
print(
'Compute Performance: ${gflops2.toStringAsFixed(2)} GFLOPS (${tflops2.toStringAsFixed(4)} TFLOPS)',
);
model2.free();
// 3. GPU Engine Benchmark - Pure MatMul Isolated
print('--- Test 3: Pure MatMul Isolated ---');
int m3 = 2048;
int k3 = 2048;
int n3 = 2048;
Random rnd3 = Random(42);
List<List<double>> rawA3 = <List<double>>[];
for (int i = 0; i < m3; i = i + 1) {
List<double> row = <double>[];
for (int j = 0; j < k3; j = j + 1) {
row.add(rnd3.nextDouble());
}
rawA3.add(row);
}
List<List<double>> rawB3 = <List<double>>[];
for (int i = 0; i < k3; i = i + 1) {
List<double> row = <double>[];
for (int j = 0; j < n3; j = j + 1) {
row.add(rnd3.nextDouble());
}
rawB3.add(row);
}
GPUTensor<Matrix> tensorA3 = GPUTensor<Matrix>(rawA3);
GPUTensor<Matrix> tensorB3 = GPUTensor<Matrix>(rawB3);
GPUTensor<Matrix> tensorC3 = GPUTensor<Matrix>.empty(<int>[m3, n3]);
CommandBuffer tape3 = CommandBuffer();
matMulGPU(tensorA3, tensorB3, tape3, outTensor: tensorC3);
Stopwatch timer3 = Stopwatch();
print('Starting high-speed pure MatMul loop (20 seconds max)...');
timer3.start();
int actualSteps3 = 0;
while (timer3.elapsedMilliseconds < 20000) {
GPUEngine.run(tape3.bytes());
actualSteps3 = actualSteps3 + 1;
}
timer3.stop();
double seconds3 = timer3.elapsedMilliseconds / 1000.0;
double flopsPerStep3 = 2.0 * m3 * k3 * n3;
double totalFlopsAchieved3 = flopsPerStep3 * actualSteps3;
double flopsPerSecond3 = totalFlopsAchieved3 / seconds3;
double gflops3 = flopsPerSecond3 / 1000000000.0;
double tflops3 = gflops3 / 1000.0;
print(
'Completed $actualSteps3 MatMul steps in ${seconds3.toStringAsFixed(2)} seconds',
);
print(
'Compute Performance: ${gflops3.toStringAsFixed(2)} GFLOPS (${tflops3.toStringAsFixed(4)} TFLOPS)',
);
tensorA3.free();
tensorB3.free();
tensorC3.free();
// 4. GPU Engine Benchmark - Absolute VRAM Saturation MatMul
print('--- Test 4: Scaled VRAM Saturation MatMul ---');
int m4 = 2048;
int k4 = 2048;
int n4 = 2048;
GPUTensor<Matrix> tensorA4 = GPUTensor<Matrix>.empty(<int>[m4, k4]);
GPUTensor<Matrix> tensorB4 = GPUTensor<Matrix>.empty(<int>[k4, n4]);
GPUTensor<Matrix> tensorC4 = GPUTensor<Matrix>.empty(<int>[m4, n4]);
CommandBuffer tape4 = CommandBuffer();
matMulGPU(tensorA4, tensorB4, tape4, outTensor: tensorC4);
int steps4 = 5;
print('Starting VRAM saturation test for $steps4 steps...');
Stopwatch timer4 = Stopwatch();
timer4.start();
for (int i = 0; i < steps4; i = i + 1) {
GPUEngine.run(tape4.bytes());
}
timer4.stop();
double seconds4 = timer4.elapsedMilliseconds / 1000.0;
double flopsPerStep4 = 2.0 * m4 * k4 * n4;
double totalFlopsAchieved4 = flopsPerStep4 * steps4;
double flopsPerSecond4 = totalFlopsAchieved4 / seconds4;
double gflops4 = flopsPerSecond4 / 1000000000.0;
double tflops4 = gflops4 / 1000.0;
print(
'Completed $steps4 massive MatMul steps in ${seconds4.toStringAsFixed(2)} seconds',
);
print(
'Compute Performance: ${gflops4.toStringAsFixed(2)} GFLOPS (${tflops4.toStringAsFixed(4)} TFLOPS)',
);
tensorA4.free();
tensorB4.free();
tensorC4.free();
setState(() {
isRunning = false;
status = 'All benchmarks completed. Check logcat/console.';
});
}
@override
Widget build(BuildContext context) {
return Scaffold(
appBar: AppBar(title: Text('GPU Engine Benchmark')),
body: Center(
child: Padding(
padding: EdgeInsets.all(24.0),
child: Column(
mainAxisAlignment: MainAxisAlignment.center,
children: <Widget>[
Text(
status,
textAlign: TextAlign.center,
style: TextStyle(fontSize: 18.0),
),
SizedBox(height: 32.0),
ElevatedButton(
onPressed: isRunning ? null : runBenchmarkSuite,
child: Padding(
padding: EdgeInsets.symmetric(
horizontal: 24.0,
vertical: 12.0,
),
child: Text(
isRunning ? 'Running...' : 'Run Benchmarks',
style: TextStyle(fontSize: 16.0),
),
),
),
],
),
),
),
);
}
}Metadata
4
likes
0
points
589
downloads
Publisher
Weekly Downloads
Metadata
Bring native GPU-accelerated machine learning to Flutter. Train models, build transformers, and run local inference without Python or cloud APIs.
Repository (GitHub)
View/report issues
License
unknown (license)
Dependencies
ffi, flutter, flutter_web_plugins, plugin_platform_interface, web
More
Packages that depend on flutter_ml
Packages that implement flutter_ml