Package io.github.kirstenali.deepj.tensor.metal

Class MetalBackend

java.lang.Object

io.github.kirstenali.deepj.tensor.metal.MetalBackend

All Implemented Interfaces:

TensorBackend

public final class MetalBackend
extends Object
implements TensorBackend

Constructor Summary

Constructors

Constructor	Description
MetalBackend()

Method Summary

Modifier and Type	Method	Description
void	adamWUpdate(Tensor w, Tensor g, Tensor mt, Tensor vt, float lr, float beta1, float beta2, float eps, float weightDecay, float bc1, float bc2)	In-place AdamW update.
Tensor	add(Tensor a, Tensor b)
Tensor	addBroadcastCols(Tensor a, Tensor v)
Tensor	addBroadcastRows(Tensor a, Tensor v)
void	addInPlace(Tensor a, Tensor b)
Tensor	addRowVector(Tensor a, Tensor v)
Tensor	addScalar(Tensor a, float scalar)
void	addScalarInPlace(Tensor a, float s)
Tensor	clamp(Tensor a, float min, float max)
Tensor	crossEntropyGradient(Tensor logits, int[] targets)
float	crossEntropyLoss(Tensor logits, int[] targets)
Tensor	divide(Tensor a, Tensor b)
Tensor	divideBroadcastCols(Tensor a, Tensor v)
void	divideInPlace(Tensor a, Tensor b)
Tensor	divideScalar(Tensor a, float scalar)
void	divideScalarInPlace(Tensor a, float s)
Tensor	exp(Tensor a)
void	expInPlace(Tensor a)
Tensor	gelu(Tensor a)
Tensor	geluBackward(Tensor input, Tensor gradOutput)
void	geluInPlace(Tensor a)
static boolean	isAvailable()
Tensor	layerNormBackward(Tensor dXHat, Tensor xHat, Tensor std, int dim)	LayerNorm backward through normalization (given dXHat, xHat, std).
Tensor	log(Tensor a)
void	logInPlace(Tensor a)
void	materializeTensor(Tensor t)	Materialize a tensor: flush any pending GPU computation and download the result to the tensor's CPU data[].
Tensor	matmul(Tensor a, Tensor b)
Tensor	maxAlongRows(Tensor a)
Tensor	meanAlongRows(Tensor a)
Tensor	multiply(Tensor a, Tensor b)
Tensor	multiplyBroadcastCols(Tensor a, Tensor v)
Tensor	multiplyBroadcastRows(Tensor a, Tensor v)
void	multiplyInPlace(Tensor a, Tensor b)
Tensor	multiplyScalar(Tensor a, float scalar)
void	multiplyScalarInPlace(Tensor a, float s)
Tensor	neg(Tensor a)
void	negInPlace(Tensor a)
Tensor	pow(Tensor a, float exponent)
void	releaseResources()	Release backend-owned resources (GPU buffers, native handles, etc.).
Tensor	relu(Tensor a)
Tensor	reluBackward(Tensor input, Tensor gradOutput)
void	reluInPlace(Tensor a)
void	scatterAddRows(Tensor target, int[] indices, Tensor grad)
Tensor	sigmoid(Tensor a)
void	sigmoidInPlace(Tensor a)
Tensor	softmaxBackward(Tensor gradOutput, Tensor softmaxOut)
Tensor	softmaxRows(Tensor logits)
Tensor	sqrt(Tensor a)
void	sqrtInPlace(Tensor a)
Tensor	subtract(Tensor a, Tensor b)
Tensor	subtractBroadcastCols(Tensor a, Tensor v)
void	subtractInPlace(Tensor a, Tensor b)
float	sum(Tensor a)
float	sumAbs(Tensor a)
Tensor	sumAlongCols(Tensor a)
Tensor	sumAlongRows(Tensor a)
Tensor	sumRows(Tensor a)
Tensor	tanh(Tensor a)
void	tanhInPlace(Tensor a)
Tensor	transpose(Tensor a)
Tensor	varianceAlongRows(Tensor a)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

MetalBackend

public MetalBackend()

Method Details

isAvailable

public static boolean isAvailable()

materializeTensor

public void materializeTensor(Tensor t)

Description copied from interface: TensorBackend

Materialize a tensor: flush any pending GPU computation and download the result to the tensor's CPU data[]. Default is a no-op (for CpuBackend).

Specified by:

materializeTensor in interface TensorBackend

matmul

public Tensor matmul(Tensor a,
 Tensor b)

Specified by:

matmul in interface TensorBackend

add

public Tensor add(Tensor a,
 Tensor b)

Specified by:

add in interface TensorBackend

subtract

public Tensor subtract(Tensor a,
 Tensor b)

Specified by:

subtract in interface TensorBackend

multiply

public Tensor multiply(Tensor a,
 Tensor b)

Specified by:

multiply in interface TensorBackend

divide

public Tensor divide(Tensor a,
 Tensor b)

Specified by:

divide in interface TensorBackend

addRowVector

public Tensor addRowVector(Tensor a,
 Tensor v)

Specified by:

addRowVector in interface TensorBackend

addBroadcastCols

public Tensor addBroadcastCols(Tensor a,
 Tensor v)

Specified by:

addBroadcastCols in interface TensorBackend

divideBroadcastCols

public Tensor divideBroadcastCols(Tensor a,
 Tensor v)

Specified by:

divideBroadcastCols in interface TensorBackend

subtractBroadcastCols

public Tensor subtractBroadcastCols(Tensor a,
 Tensor v)

Specified by:

subtractBroadcastCols in interface TensorBackend

multiplyBroadcastCols

public Tensor multiplyBroadcastCols(Tensor a,
 Tensor v)

Specified by:

multiplyBroadcastCols in interface TensorBackend

addBroadcastRows

public Tensor addBroadcastRows(Tensor a,
 Tensor v)

Specified by:

addBroadcastRows in interface TensorBackend

multiplyBroadcastRows

public Tensor multiplyBroadcastRows(Tensor a,
 Tensor v)

Specified by:

multiplyBroadcastRows in interface TensorBackend

multiplyScalar

public Tensor multiplyScalar(Tensor a,
 float scalar)

Specified by:

multiplyScalar in interface TensorBackend

addScalar

public Tensor addScalar(Tensor a,
 float scalar)

Specified by:

addScalar in interface TensorBackend

divideScalar

public Tensor divideScalar(Tensor a,
 float scalar)

Specified by:

divideScalar in interface TensorBackend

sumRows

public Tensor sumRows(Tensor a)

Specified by:

sumRows in interface TensorBackend

sumAlongRows

public Tensor sumAlongRows(Tensor a)

Specified by:

sumAlongRows in interface TensorBackend

sumAlongCols

public Tensor sumAlongCols(Tensor a)

Specified by:

sumAlongCols in interface TensorBackend

meanAlongRows

public Tensor meanAlongRows(Tensor a)

Specified by:

meanAlongRows in interface TensorBackend

varianceAlongRows

public Tensor varianceAlongRows(Tensor a)

Specified by:

varianceAlongRows in interface TensorBackend

maxAlongRows

public Tensor maxAlongRows(Tensor a)

Specified by:

maxAlongRows in interface TensorBackend

sum

public float sum(Tensor a)

Specified by:

sum in interface TensorBackend

sumAbs

public float sumAbs(Tensor a)

Specified by:

sumAbs in interface TensorBackend

transpose

public Tensor transpose(Tensor a)

Specified by:

transpose in interface TensorBackend

clamp

public Tensor clamp(Tensor a,
 float min,
 float max)

Specified by:

clamp in interface TensorBackend

pow

public Tensor pow(Tensor a,
 float exponent)

Specified by:

pow in interface TensorBackend

sqrt

public Tensor sqrt(Tensor a)

Specified by:

sqrt in interface TensorBackend

neg

public Tensor neg(Tensor a)

Specified by:

neg in interface TensorBackend

exp

public Tensor exp(Tensor a)

Specified by:

exp in interface TensorBackend

log

public Tensor log(Tensor a)

Specified by:

log in interface TensorBackend

tanh

public Tensor tanh(Tensor a)

Specified by:

tanh in interface TensorBackend

sigmoid

public Tensor sigmoid(Tensor a)

Specified by:

sigmoid in interface TensorBackend

relu

public Tensor relu(Tensor a)

Specified by:

relu in interface TensorBackend

reluBackward

public Tensor reluBackward(Tensor input,
 Tensor gradOutput)

Specified by:

reluBackward in interface TensorBackend

gelu

public Tensor gelu(Tensor a)

Specified by:

gelu in interface TensorBackend

geluBackward

public Tensor geluBackward(Tensor input,
 Tensor gradOutput)

Specified by:

geluBackward in interface TensorBackend

softmaxRows

public Tensor softmaxRows(Tensor logits)

Specified by:

softmaxRows in interface TensorBackend

softmaxBackward

public Tensor softmaxBackward(Tensor gradOutput,
 Tensor softmaxOut)

Specified by:

softmaxBackward in interface TensorBackend

addInPlace

public void addInPlace(Tensor a,
 Tensor b)

Specified by:

addInPlace in interface TensorBackend

subtractInPlace

public void subtractInPlace(Tensor a,
 Tensor b)

Specified by:

subtractInPlace in interface TensorBackend

multiplyInPlace

public void multiplyInPlace(Tensor a,
 Tensor b)

Specified by:

multiplyInPlace in interface TensorBackend

divideInPlace

public void divideInPlace(Tensor a,
 Tensor b)

Specified by:

divideInPlace in interface TensorBackend

multiplyScalarInPlace

public void multiplyScalarInPlace(Tensor a,
 float s)

Specified by:

multiplyScalarInPlace in interface TensorBackend

addScalarInPlace

public void addScalarInPlace(Tensor a,
 float s)

Specified by:

addScalarInPlace in interface TensorBackend

divideScalarInPlace

public void divideScalarInPlace(Tensor a,
 float s)

Specified by:

divideScalarInPlace in interface TensorBackend

sqrtInPlace

public void sqrtInPlace(Tensor a)

Specified by:

sqrtInPlace in interface TensorBackend

negInPlace

public void negInPlace(Tensor a)

Specified by:

negInPlace in interface TensorBackend

expInPlace

public void expInPlace(Tensor a)

Specified by:

expInPlace in interface TensorBackend

logInPlace

public void logInPlace(Tensor a)

Specified by:

logInPlace in interface TensorBackend

reluInPlace

public void reluInPlace(Tensor a)

Specified by:

reluInPlace in interface TensorBackend

geluInPlace

public void geluInPlace(Tensor a)

Specified by:

geluInPlace in interface TensorBackend

tanhInPlace

public void tanhInPlace(Tensor a)

Specified by:

tanhInPlace in interface TensorBackend

sigmoidInPlace

public void sigmoidInPlace(Tensor a)

Specified by:

sigmoidInPlace in interface TensorBackend

crossEntropyGradient

public Tensor crossEntropyGradient(Tensor logits,
 int[] targets)

Specified by:

crossEntropyGradient in interface TensorBackend

crossEntropyLoss

public float crossEntropyLoss(Tensor logits,
 int[] targets)

Specified by:

crossEntropyLoss in interface TensorBackend

adamWUpdate

public void adamWUpdate(Tensor w,
 Tensor g,
 Tensor mt,
 Tensor vt,
 float lr,
 float beta1,
 float beta2,
 float eps,
 float weightDecay,
 float bc1,
 float bc2)

Description copied from interface: TensorBackend

In-place AdamW update. Mutates w, mt, vt.

Specified by:

adamWUpdate in interface TensorBackend

layerNormBackward

public Tensor layerNormBackward(Tensor dXHat,
 Tensor xHat,
 Tensor std,
 int dim)

Description copied from interface: TensorBackend

LayerNorm backward through normalization (given dXHat, xHat, std).

Specified by:

layerNormBackward in interface TensorBackend

scatterAddRows

public void scatterAddRows(Tensor target,
 int[] indices,
 Tensor grad)

Specified by:

scatterAddRows in interface TensorBackend

releaseResources

public void releaseResources()

Description copied from interface: TensorBackend

Release backend-owned resources (GPU buffers, native handles, etc.). Default is a no-op for backends without external resources.

Specified by:

releaseResources in interface TensorBackend