xtrakerns

package
v0.0.0-...-c9f06ed Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: May 13, 2020 License: MIT Imports: 5 Imported by: 0

Documentation

Index

Constants

View Source 
const Defines = `` /* 445-byte string literal not displayed */

Defines are used for all kernels

View Source 
const Headers = `
#include <cuda.h>
#include <stdbool.h>
#include <cuda_fp16.h>
`

Headers are used for all kernels

Variables

This section is empty.

Functions

func CreateModule 

func CreateModule(k Kernel, dev cuda.Device) (*cuda.Module, error)

Types

type Kernel 

type Kernel struct {
	Name string
	Code string
}

Kernel is used to build kernels

func AdaDelta 

func AdaDelta() Kernel

AdaDelta is a weight trainer

func AdaDeltaFP16 

func AdaDeltaFP16() Kernel

AdaDeltaFP16 is a weight trainer

func AdaGrad 

func AdaGrad() Kernel

AdaGrad is the AdaGrad training weight updater

func AdaGradFP16 

func AdaGradFP16() Kernel

AdaGradFP16 is a weight updater

func Adam 

func Adam() Kernel

Adam is a weight trainer

func AdamFP16 

func AdamFP16() Kernel

AdamFP16 is a weight updater probably needs fixed

func AllKerns 

func AllKerns() (all []Kernel)

func ConcatBackwardNCHW 

func ConcatBackwardNCHW() Kernel

ConcatBackwardNCHW is a concat for NCHW that hasn't been tested.

func ConcatBackwardNCHWFP16 

func ConcatBackwardNCHWFP16() Kernel

ConcatBackwardNCHWhalf for concat half

func ConcatForwardNCHW 

func ConcatForwardNCHW() Kernel
func ConcatForwardNHWC()Kernel{
	return Kernel{
		Name:`ConcatForwardHWC`,
		Code:`extern "C" __global__ void ConcatForwardNHWC(const int XThreads,
			const int YThreads,
			const int Batch,
			const int nsrcs,
			const float **srcs
			const int **srcchansize,
			const int destchansize,
			float* Dest){

				CUDA_GRID_LOOP_AXIS(i, YThreads,y){
				CUDA_GRID_LOOP_AXIS(j, XThreads,x){

				}
				}`
	}
}

ConcatForwardNCHW is a concat for NCHW that hasn't been tested.

func ConcatForwardNCHWFP16 

func ConcatForwardNCHWFP16() Kernel

ConcatForwardNCHWhalf is concat func in halfs

func L1L2 

func L1L2() Kernel

L1L2 are the l1l2 functions for weight normalization

func L1L2FP16 

func L1L2FP16() Kernel

L1L2FP16 is the L1L2 normalization functions

func LeakyBackward 

func LeakyBackward() Kernel

LeakyBackward --backwards function for forward

func LeakyBackwardAlpha 

func LeakyBackwardAlpha() Kernel

LeakyBackwardAlpha --backwards function for forward

func LeakyBackwardAlphaBeta 

func LeakyBackwardAlphaBeta() Kernel

LeakyBackwardAlphaBeta --backwards function for forward

func LeakyBackwardAlphaBetaFP16 

func LeakyBackwardAlphaBetaFP16() Kernel

LeakyBackwardAlphaBetaFP16 --backwards function for forward

func LeakyBackwardAlphaFP16 

func LeakyBackwardAlphaFP16() Kernel

LeakyBackwardAlphaFP16 --backwards function for forward

func LeakyBackwardFP16 

func LeakyBackwardFP16() Kernel

LeakyBackwardFP16 --backwards function for forward

func LeakyForward 

func LeakyForward() Kernel

LeakyForward is a leaky function

func LeakyForwardAlpha 

func LeakyForwardAlpha() Kernel

LeakyForwardAlpha is a leaky function

func LeakyForwardAlphaBeta 

func LeakyForwardAlphaBeta() Kernel

LeakyForwardAlphaBeta - is the leaky activation

func LeakyForwardAlphaBetaFP16 

func LeakyForwardAlphaBetaFP16() Kernel

LeakyForwardAlphaBetaFP16 is the leaky activation

func LeakyForwardAlphaFP16 

func LeakyForwardAlphaFP16() Kernel

LeakyForwardAlphaFP16 is a leaky function

func LeakyForwardFP16 

func LeakyForwardFP16() Kernel

LeakyForwardFP16 is a leaky function

func MSELoss 

func MSELoss() Kernel

MSELoss performs the mean squared error loss function

func MSELossFP16 

func MSELossFP16() Kernel

MSELossFP16 performs the mean squared error loss function

func MSELossbyBatches 

func MSELossbyBatches() Kernel

MSELossbyBatches performs the mean squared error loss function by batches Good for gans

func MSELossbyBatchesFP16 

func MSELossbyBatchesFP16() Kernel

MSELossbyBatchesFP16 performs the mean squared error loss function by batches Good for gans

func MakePlanarImageBatchesUint8 

func MakePlanarImageBatchesUint8() Kernel

MakePlanarImageBatchesUint8 - for this to work all the each batch should have the same amount of channels and all the channels need to be the same size

func NearestNeighborNCHW 

func NearestNeighborNCHW() Kernel

NearestNeighborNCHW is a nearest neighbor resize function

func NearestNeighborNCHWBack 

func NearestNeighborNCHWBack() Kernel

NearestNeighborNCHWBack is a nearest neighbor resize function

func NearestNeighborNCHWBackFP16 

func NearestNeighborNCHWBackFP16() Kernel

NearestNeighborNCHWBackFP16 is a nearest neighbor resize function

func NearestNeighborNCHWFP16 

func NearestNeighborNCHWFP16() Kernel

NearestNeighborNCHWFP16 is a nearest neighbor resize function

func NearestNeighborNHWC 

func NearestNeighborNHWC() Kernel

NearestNeighborNHWC is a nearest neighbor resize function

func NearestNeighborNHWCBack 

func NearestNeighborNHWCBack() Kernel

NearestNeighborNHWCBack is a nearest neighbor resize function

func NearestNeighborNHWCBackFP16 

func NearestNeighborNHWCBackFP16() Kernel

NearestNeighborNHWCBackFP16 is a nearest neighbor resize function

func NearestNeighborNHWCFP16 

func NearestNeighborNHWCFP16() Kernel

NearestNeighborNHWCFP16 is a nearest neighbor resize function

func PreluBackward 

func PreluBackward() Kernel

PreluBackward --backwards function for forward

func PreluBackwardFP16 

func PreluBackwardFP16() Kernel

PreluBackwardFP16 --backwards function for forward

func PreluForward 

func PreluForward() Kernel

PreluForward does the forward Prelu

func PreluForwardFP16 

func PreluForwardFP16() Kernel

PreluForwardFP16 does the forward Prelu

func Segment1stDim 

func Segment1stDim() Kernel

Segment1stDim -- is paired with the host -- it segments the first dim of a tensor

func Segment1stDimFP16 

func Segment1stDimFP16() Kernel

Segment1stDimFP16 -- is paired with the host -- it segments the first dim of a tensor

func ShapeToBatch4DNHWC 

func ShapeToBatch4DNHWC() Kernel

ShapeToBatch4DNHWC Does a stride shape to batch. Make sure values on receiving end are set to zero when s2b is 0

func ShapeToBatch4DNHWCFP16 

func ShapeToBatch4DNHWCFP16() Kernel

ShapeToBatch4DNHWCFP16 Does a stride shape to batch. Make sure values on receiving end are set to zero when s2b is 0

func ShapetoBatch4DNCHW 

func ShapetoBatch4DNCHW() Kernel

ShapetoBatch4DNCHW Does a stride shape to batch. Make sure values on receiving end are set to zero when s2b is 0

func ShapetoBatch4DNCHWFP16 

func ShapetoBatch4DNCHWFP16() Kernel

ShapetoBatch4DNCHWFP16 is like ShapetoBatch4DNCHW

func SwapEveryOther 

func SwapEveryOther() Kernel

SwapEveryOther will swap the batches between 2 tensors. It will be either the even or the odd. Both tensors have to be equal in size and dims. if even is >0 then it will do the even batches. Make sure labels are swapped on host end.

func SwapEveryOtherFP16 

func SwapEveryOtherFP16() Kernel

SwapEveryOtherFP16 will swap the batches between 2 tensors. It will be either the even or the odd. Both tensors have to be equal in size and dims. if even is >0 then it will do the even batches. Make sure labels are swapped on host end.

func SwapUpperLower 

func SwapUpperLower() Kernel

SwapUpperLower will swap either the upper or lower batches Right Now inverse doesn't do anything

func SwapUpperLowerFP16 

func SwapUpperLowerFP16() Kernel

SwapUpperLowerFP16 is like the FP32 version

func ThreshBackward 

func ThreshBackward() Kernel

ThreshBackward --backwards function for forward

func ThreshBackwardFP16 

func ThreshBackwardFP16() Kernel

ThreshBackwardFP16 --backwards function for forward

func ThreshForward 

func ThreshForward() Kernel

ThreshForward is kind of memory expensive, mostly because it is experimental. To test start the positive at random uniform numbers between .9 and 1.1 and do the negcoefs between .01 and .2 or something along those lines. maybe the threshold should be between -.3 and .3 uniform number

func ThreshForwardFP16 

func ThreshForwardFP16() Kernel

ThreshForwardFP16 is kind of memory expensive, mostly because it is experimental. To test start the positive at random uniform numbers between .9 and 1.1 and do the negcoefs between .01 and .2 or something along those lines. maybe the threshold should be between -.3 and .3 uniform number

func Transpose 

func Transpose() Kernel

Transpose is the kernel for transpose

func TransposeFP16 

func TransposeFP16() Kernel

TransposeFP16 is the kernel for transpose

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.