cuda

package
v3.9.3+incompatible Latest Latest
Warning

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

 Go to latest Published: Sep 30, 2016 License: CC-BY-3.0, Freetype, GPL-3.0-or-later Imports: 14 Imported by: 109

Documentation

Overview

Package cuda provides GPU interaction

Index

Constants

View Source 
const (
	BlockSize    = 512
	TileX, TileY = 32, 32
	MaxGridSize  = 65535
)

CUDA Launch parameters. there might be better choices for recent hardware, but it barely makes a difference in the end.

View Source 
const (
	X = 0
	Y = 1
	Z = 2
)
View Source 
const CONV_TOLERANCE = 1e-6

Maximum tolerable error on demag convolution self-test.

View Source 
const FFT_IMAG_TOLERANCE = 1e-6

Maximum tolerable imaginary/real part for demag kernel in Fourier space. Assures kernel has correct symmetry.

View Source 
const REDUCE_BLOCKSIZE = C.REDUCE_BLOCKSIZE

Block size for reduce kernels.

Variables

View Source 
var (
	Version     int    // cuda version
	DevName     string // GPU name
	TotalMem    int64  // total GPU memory
	GPUInfo     string // Human-readable GPU description
	Synchronous bool   // for debug: synchronize stream0 at every kernel launch

)
View Source 
var UseCC = 0

Functions

func AddCubicAnisotropy 

func AddCubicAnisotropy(Beff, m *data.Slice, k1_red, k2_red, k3_red LUTPtr, c1, c2 LUTPtrs, regions *Bytes)

Adds cubic anisotropy field to Beff. see cubicanisotropy.cu

func AddDMI 

func AddDMI(Beff *data.Slice, m *data.Slice, Aex_red, Dex_red SymmLUT, regions *Bytes, mesh *data.Mesh)

Add effective field of Dzyaloshinskii-Moriya interaction to Beff (Tesla). According to Bagdanov and Röβler, PRL 87, 3, 2001. eq.8 (out-of-plane symmetry breaking). See dmi.cu

func AddDMIBulk 

func AddDMIBulk(Beff *data.Slice, m *data.Slice, Aex_red, D_red SymmLUT, regions *Bytes, mesh *data.Mesh)

Add effective field due to bulk Dzyaloshinskii-Moriya interaction to Beff. See dmibulk.cu

func AddDotProduct 

func AddDotProduct(dst *data.Slice, prefactor float32, a, b *data.Slice)

dst += prefactor * dot(a, b), as used for energy density

func AddExchange 

func AddExchange(B, m *data.Slice, Aex_red SymmLUT, regions *Bytes, mesh *data.Mesh)

Add exchange field to Beff. 

m: normalized magnetization
B: effective field in Tesla
Aex_red: Aex / (Msat * 1e18 m2)

see exchange.cu

func AddSlonczewskiTorque 

func AddSlonczewskiTorque(torque, m, J, fixedP *data.Slice, Msat, alpha, pol, λ, ε_prime LUTPtr, regions *Bytes, mesh *data.Mesh)

Add Slonczewski ST torque to torque (Tesla). see slonczewski.cu

func AddUniaxialAnisotropy 

func AddUniaxialAnisotropy(Beff, m *data.Slice, k1_red, k2_red LUTPtr, u LUTPtrs, regions *Bytes)

Add uniaxial magnetocrystalline anisotropy field to Beff. see uniaxialanisotropy.cu

func AddZhangLiTorque 

func AddZhangLiTorque(torque, m, J *data.Slice, bsat, alpha, xi, pol LUTPtr, regions *Bytes, mesh *data.Mesh)

Add Zhang-Li ST torque (Tesla) to torque. see zhangli.cu

func Buffer 

func Buffer(nComp int, size [3]int) *data.Slice

Returns a GPU slice for temporary use. To be returned to the pool with Recycle

func Crop 

func Crop(dst, src *data.Slice, offX, offY, offZ int)

Crop stores in dst a rectangle cropped from src at given offset position. dst size may be smaller than src.

func Dot 

func Dot(a, b *data.Slice) float32

Dot product.

func ExchangeDecode 

func ExchangeDecode(dst *data.Slice, Aex_red SymmLUT, regions *Bytes, mesh *data.Mesh)

Finds the average exchange strength around each cell, for debugging.

func FreeBuffers 

func FreeBuffers()

Frees all buffers. Called after mesh resize.

func GPUCopy 

func GPUCopy(in *data.Slice) *data.Slice

Returns a copy of in, allocated on GPU.

func GetCell 

func GetCell(s *data.Slice, comp, ix, iy, iz int) float32

func GetElem 

func GetElem(s *data.Slice, comp int, index int) float32

func Init 

func Init(gpu int)

Locks to an OS thread and initializes CUDA for that thread.

func LLNoPrecess 

func LLNoPrecess(torque, m, B *data.Slice)

Landau-Lifshitz torque with precession disabled. Used by engine.Relax().

func LLTorque 

func LLTorque(torque, m, B *data.Slice, alpha LUTPtr, regions *Bytes)

Landau-Lifshitz torque divided by gamma0:

  • 1/(1+α²) [ m x B + α m x (m x B) ] torque in Tesla m normalized B in Tesla

see lltorque.cu

func Madd2 

func Madd2(dst, src1, src2 *data.Slice, factor1, factor2 float32)

multiply-add: dst[i] = src1[i] * factor1 + src2[i] * factor2

func Madd3 

func Madd3(dst, src1, src2, src3 *data.Slice, factor1, factor2, factor3 float32)

multiply-add: dst[i] = src1[i] * factor1 + src2[i] * factor2 + src3 * factor3

func MaxAbs 

func MaxAbs(in *data.Slice) float32

Maximum of absolute values of all elements.

func MaxVecDiff 

func MaxVecDiff(x, y *data.Slice) float64

Maximum of the norms of the difference between all vectors (x1,y1,z1) and (x2,y2,z2) 

(dx, dy, dz) = (x1, y1, z1) - (x2, y2, z2)
max_i sqrt( dx[i]*dx[i] + dy[i]*dy[i] + dz[i]*dz[i] )

func MaxVecNorm 

func MaxVecNorm(v *data.Slice) float64

Maximum of the norms of all vectors (x[i], y[i], z[i]). 

max_i sqrt( x[i]*x[i] + y[i]*y[i] + z[i]*z[i] )

func MemAlloc 

func MemAlloc(bytes int64) unsafe.Pointer

Wrapper for cu.MemAlloc, fatal exit on out of memory.

func MemCpy 

func MemCpy(dst, src unsafe.Pointer, bytes int64)

func MemCpyDtoH 

func MemCpyDtoH(dst, src unsafe.Pointer, bytes int64)

func MemCpyHtoD 

func MemCpyHtoD(dst, src unsafe.Pointer, bytes int64)

func Memset 

func Memset(s *data.Slice, val ...float32)

Memset sets the Slice's components to the specified values. To be carefully used on unified slice (need sync)

func Minimize 

func Minimize(m, m0, torque *data.Slice, dt float32)

m = 1 / (4 + τ²(m x H)²) [{4 - τ²(m x H)²} m - 4τ(m x m x H)] note: torque from LLNoPrecess has negative sign

func Mul 

func Mul(dst, a, b *data.Slice)

multiply: dst[i] = a[i] * b[i]

func NewSlice 

func NewSlice(nComp int, size [3]int) *data.Slice

Make a GPU Slice with nComp components each of size length.

func Normalize 

func Normalize(vec, vol *data.Slice)

Normalize vec to unit length, unless length or vol are zero.

func Recycle 

func Recycle(s *data.Slice)

Returns a buffer obtained from GetBuffer to the pool.

func RegionAddV 

func RegionAddV(dst *data.Slice, lut LUTPtrs, regions *Bytes)

dst += LUT[region], for vectors. Used to add terms to excitation.

func RegionDecode 

func RegionDecode(dst *data.Slice, lut LUTPtr, regions *Bytes)

decode the regions+LUT pair into an uncompressed array

func RegionSelect 

func RegionSelect(dst, src *data.Slice, regions *Bytes, region byte)

select the part of src within the specified region, set 0's everywhere else.

func Resize 

func Resize(dst, src *data.Slice, layer int)

Select and resize one layer for interactive output

func SetCell 

func SetCell(s *data.Slice, comp int, ix, iy, iz int, value float32)

func SetElem 

func SetElem(s *data.Slice, comp int, index int, value float32)

func SetMaxAngle 

func SetMaxAngle(dst, m *data.Slice, Aex_red SymmLUT, regions *Bytes, mesh *data.Mesh)

SetMaxAngle sets dst to the maximum angle of each cells magnetization with all of its neighbors, provided the exchange stiffness with that neighbor is nonzero.

func SetTemperature 

func SetTemperature(Bth, noise *data.Slice, temp_red LUTPtr, k2mu0_VgammaDt float64, regions *Bytes)

Set Bth to thermal noise (Brown). see temperature.cu

func SetTopologicalCharge 

func SetTopologicalCharge(s *data.Slice, m *data.Slice, mesh *data.Mesh)

Set s to the toplogogical charge density s = m · (m/∂x ❌ ∂m/∂y) See topologicalcharge.cu

func ShiftBytes 

func ShiftBytes(dst, src *Bytes, m *data.Mesh, shiftX int, clamp byte)

Like Shift, but for bytes

func ShiftX 

func ShiftX(dst, src *data.Slice, shiftX int, clampL, clampR float32)

shift dst by shx cells (positive or negative) along X-axis. new edge value is clampL at left edge or clampR at right edge.

func Sum 

func Sum(in *data.Slice) float32

Sum of all elements.

func Sync 

func Sync()

Synchronize the global stream This is called before and after all memcopy operations between host and device.

func Zero 

func Zero(s *data.Slice)

Set all elements of all components to zero.

func ZeroMask 

func ZeroMask(dst *data.Slice, mask LUTPtr, regions *Bytes)

Sets vector dst to zero where mask != 0.

Types

type Bytes 

type Bytes struct {
	Ptr unsafe.Pointer
	Len int
}

3D byte slice, used for region lookup.

func NewBytes 

func NewBytes(Len int) *Bytes

Construct new byte slice with given length, initialised to zeros.

func (*Bytes) Copy 

func (dst *Bytes) Copy(src *Bytes)

Copy on device: dst = src.

func (*Bytes) Download 

func (src *Bytes) Download(dst []byte)

Copy to host: dst = src.

func (*Bytes) Free 

func (b *Bytes) Free()

Frees the GPU memory and disables the slice.

func (*Bytes) Get 

func (src *Bytes) Get(index int) byte

Get one element. data.Index can be used to find the index for x,y,z.

func (*Bytes) Set 

func (dst *Bytes) Set(index int, value byte)

Set one element to value. data.Index can be used to find the index for x,y,z.

func (*Bytes) Upload 

func (dst *Bytes) Upload(src []byte)

Upload src (host) to dst (gpu).

type DemagConvolution 

type DemagConvolution struct {
	// contains filtered or unexported fields
}

Stores the necessary state to perform FFT-accelerated convolution with magnetostatic kernel (or other kernel of same symmetry).

func NewDemag 

func NewDemag(inputSize, PBC [3]int, kernel [3][3]*data.Slice, test bool) *DemagConvolution

Initializes a convolution to evaluate the demag field for the given mesh geometry. Sanity-checked if test == true (slow-ish for large meshes).

func (*DemagConvolution) Exec 

func (c *DemagConvolution) Exec(B, m, vol *data.Slice, Bsat LUTPtr, regions *Bytes)

Calculate the demag field of m * vol * Bsat, store result in B. 

m:    magnetization normalized to unit length
vol:  unitless mask used to scale m's length, may be nil
Bsat: saturation magnetization in Tesla
B:    resulting demag field, in Tesla

func (*DemagConvolution) Free 

func (c *DemagConvolution) Free()

type LUTPtr 

type LUTPtr unsafe.Pointer // points to 256 float32's

type LUTPtrs 

type LUTPtrs []unsafe.Pointer // elements point to 256 float32's

type MFMConvolution 

type MFMConvolution struct {
	// contains filtered or unexported fields
}

Stores the necessary state to perform FFT-accelerated convolution

func NewMFM 

func NewMFM(mesh *data.Mesh, lift, tipsize float64) *MFMConvolution

Initializes a convolution to evaluate the demag field for the given mesh geometry.

func (*MFMConvolution) Exec 

func (c *MFMConvolution) Exec(outp, inp, vol *data.Slice, Bsat LUTPtr, regions *Bytes)

store MFM image in output, based on magnetization in inp.

func (*MFMConvolution) Free 

func (c *MFMConvolution) Free()

func (*MFMConvolution) Reinit 

func (c *MFMConvolution) Reinit(lift, tipsize float64)

type SymmLUT 

type SymmLUT unsafe.Pointer // points to 256x256 symmetric matrix, only lower half stored. See exchange.cu

Source Files

View all Source files

Directories

Path Synopsis
Go bindings for the CUDA driver API.
 Go bindings for the CUDA driver API.
Go bindings for the CUDA CUFFT API.
 Go bindings for the CUDA CUFFT API.

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.