Apple MPS Backend

MPS/Metal inference entry points and helpers. More...

Namespaces
namespace	Darknet
	The namespace for the C++ Darknet API.

Functions
bool	mps_avgpool_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, bool defer_readback, const char **reason)
static bool	mps_avgpool_forward (const Darknet::Layer &l, const float *input, float *output)
bool	mps_connected_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, bool defer_readback, bool *activation_applied, const char **reason)
	Try to execute connected layer forward using MPS (batchnorm/activation).
static bool	mps_connected_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, const char **reason)
bool	mps_convolution_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, bool defer_readback, bool *activation_applied, const char **reason)
	Try to execute convolution forward using MPS.
static bool	mps_convolution_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, const char **reason)
static bool	mps_convolution_forward (const Darknet::Layer &l, const float *input, float *output)
static bool	mps_convolution_forward (const Darknet::Layer &l, const float *input, float *output, const char **reason)
void	mps_flush_deferred_output (const Darknet::Layer *layer)
void	mps_flush_output_if_needed (const Darknet::Layer *layer, float *output)
bool	mps_gemm (int TA, int TB, int M, int N, int K, float ALPHA, float *A, int lda, float *B, int ldb, float BETA, float *C, int ldc)
	Try to execute GEMM using MPS.
bool	mps_global_avgpool_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, bool defer_readback, const char **reason)
	Try to execute global avgpool forward using MPS.
static bool	mps_global_avgpool_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, const char **reason)
bool	mps_is_available ()
	Returns true if MPS is available and initialized.
bool	mps_is_output_deferred (const Darknet::Layer *layer)
bool	mps_layer_can_run (const Darknet::Layer &l, bool train)
bool	mps_maxpool_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, bool defer_readback, const char **reason)
static bool	mps_maxpool_forward (const Darknet::Layer &l, const float *input, float *output)
const Darknet::Layer *	mps_prev_layer (const Darknet::NetworkState &state)
bool	mps_reorg_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, bool defer_readback, const char **reason)
	Try to execute reorg using a Metal kernel on GPU.
static bool	mps_reorg_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, const char **reason)
bool	mps_route_forward (const Darknet::Layer &l, const Darknet::Network &net, float *output, bool defer_readback, const char **reason)
	Try to concatenate route inputs using MPS.
static bool	mps_route_forward (const Darknet::Layer &l, const Darknet::Network &net, float *output, const char **reason)
static bool	mps_shortcut_forward (const Darknet::Layer &l, const Darknet::Layer *from, const float *input, float *output, const char **reason)
bool	mps_shortcut_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const Darknet::Layer *from, const float *input, float *output, bool defer_readback, bool *activation_applied, const char **reason)
	Try to execute shortcut add using MPS.
bool	mps_should_defer_readback (const Darknet::NetworkState &state)
bool	mps_softmax_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, const char **reason)
	Try to execute softmax on GPU.
bool	mps_upsample_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, bool defer_readback, const char **reason)
	Try to upsample using a Metal kernel on GPU.
static bool	mps_upsample_forward (const Darknet::Layer &l, const Darknet::Layer *prev, const float *input, float *output, const char **reason)

Detailed Description

MPS/Metal inference entry points and helpers.

Function Documentation

mps_avgpool_forward() [1/2]

bool mps_avgpool_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		bool	defer_readback,
		const char **	reason
	)

Here is the caller graph for this function:

mps_avgpool_forward() [2/2]

static bool mps_avgpool_forward ( const Darknet::Layer &  l, const float *  input, float *  output  )

inlinestatic

Here is the call graph for this function:

mps_connected_forward() [1/2]

bool mps_connected_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		bool	defer_readback,
		bool *	activation_applied,
		const char **	reason
	)

Try to execute connected layer forward using MPS (batchnorm/activation).

Here is the call graph for this function:

Here is the caller graph for this function:

mps_connected_forward() [2/2]

static bool mps_connected_forward ( const Darknet::Layer &  l, const Darknet::Layer *  prev, const float *  input, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function:

mps_convolution_forward() [1/4]

bool mps_convolution_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		bool	defer_readback,
		bool *	activation_applied,
		const char **	reason
	)

Try to execute convolution forward using MPS.

Parameters

defer_readback	If true, output may remain on GPU for a subsequent MPS layer.
activation_applied	Set to true if MPS applied the activation.
reason	Optional fallback reason string when returning false.

Here is the caller graph for this function:

mps_convolution_forward() [2/4]

static bool mps_convolution_forward ( const Darknet::Layer &  l, const Darknet::Layer *  prev, const float *  input, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function:

mps_convolution_forward() [3/4]

static bool mps_convolution_forward ( const Darknet::Layer &  l, const float *  input, float *  output  )

inlinestatic

Here is the call graph for this function:

mps_convolution_forward() [4/4]

static bool mps_convolution_forward ( const Darknet::Layer &  l, const float *  input, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function:

mps_flush_deferred_output()

void mps_flush_deferred_output

(

const Darknet::Layer *

layer

)

Here is the call graph for this function:

Here is the caller graph for this function:

mps_flush_output_if_needed()

void mps_flush_output_if_needed	(	const Darknet::Layer *	layer,
		float *	output
	)

Here is the caller graph for this function:

mps_gemm()

bool mps_gemm	(	int	TA,
		int	TB,
		int	M,
		int	N,
		int	K,
		float	ALPHA,
		float *	A,
		int	lda,
		float *	B,
		int	ldb,
		float	BETA,
		float *	C,
		int	ldc
	)

Try to execute GEMM using MPS.

Returns true if MPS handled the op.

mps_global_avgpool_forward() [1/2]

bool mps_global_avgpool_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		bool	defer_readback,
		const char **	reason
	)

Try to execute global avgpool forward using MPS.

Here is the caller graph for this function:

mps_global_avgpool_forward() [2/2]

static bool mps_global_avgpool_forward ( const Darknet::Layer &  l, const Darknet::Layer *  prev, const float *  input, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function:

mps_is_available()

bool mps_is_available

(

)

Returns true if MPS is available and initialized.

mps_is_output_deferred()

bool mps_is_output_deferred

(

const Darknet::Layer *

layer

)

Here is the caller graph for this function:

mps_layer_can_run()

bool mps_layer_can_run	(	const Darknet::Layer &	l,
		bool	train
	)

Here is the caller graph for this function:

mps_maxpool_forward() [1/2]

bool mps_maxpool_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		bool	defer_readback,
		const char **	reason
	)

Here is the caller graph for this function:

mps_maxpool_forward() [2/2]

static bool mps_maxpool_forward ( const Darknet::Layer &  l, const float *  input, float *  output  )

inlinestatic

Here is the call graph for this function:

mps_prev_layer()

const Darknet::Layer * mps_prev_layer

(

const Darknet::NetworkState &

state

)

Here is the caller graph for this function:

mps_reorg_forward() [1/2]

bool mps_reorg_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		bool	defer_readback,
		const char **	reason
	)

Try to execute reorg using a Metal kernel on GPU.

Here is the call graph for this function:

Here is the caller graph for this function:

mps_reorg_forward() [2/2]

static bool mps_reorg_forward ( const Darknet::Layer &  l, const Darknet::Layer *  prev, const float *  input, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function:

mps_route_forward() [1/2]

bool mps_route_forward	(	const Darknet::Layer &	l,
		const Darknet::Network &	net,
		float *	output,
		bool	defer_readback,
		const char **	reason
	)

Try to concatenate route inputs using MPS.

Here is the call graph for this function:

Here is the caller graph for this function:

mps_route_forward() [2/2]

static bool mps_route_forward ( const Darknet::Layer &  l, const Darknet::Network &  net, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function:

mps_shortcut_forward() [1/2]

static bool mps_shortcut_forward ( const Darknet::Layer &  l, const Darknet::Layer *  from, const float *  input, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function:

mps_shortcut_forward() [2/2]

bool mps_shortcut_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const Darknet::Layer *	from,
		const float *	input,
		float *	output,
		bool	defer_readback,
		bool *	activation_applied,
		const char **	reason
	)

Try to execute shortcut add using MPS.

Here is the call graph for this function:

Here is the caller graph for this function:

mps_should_defer_readback()

bool mps_should_defer_readback

(

const Darknet::NetworkState &

state

)

Here is the call graph for this function:

Here is the caller graph for this function:

mps_softmax_forward()

bool mps_softmax_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		const char **	reason
	)

Try to execute softmax on GPU.

Here is the call graph for this function:

Here is the caller graph for this function:

mps_upsample_forward() [1/2]

bool mps_upsample_forward	(	const Darknet::Layer &	l,
		const Darknet::Layer *	prev,
		const float *	input,
		float *	output,
		bool	defer_readback,
		const char **	reason
	)

Try to upsample using a Metal kernel on GPU.

Here is the caller graph for this function:

mps_upsample_forward() [2/2]

static bool mps_upsample_forward ( const Darknet::Layer &  l, const Darknet::Layer *  prev, const float *  input, float *  output, const char **  reason  )

inlinestatic

Here is the call graph for this function: