Sparse BLAS Functionality

In the following table for functionality, sm = sparse matrix, dm = dense matrix, sv = sparse vector, dv = dense vector, sc = scalar.

In the following table for operations, dense vectors = x, y, sparse vectors = w,v, dense matrices = X,Y, sparse matrices = A, B, C, and scalars = alpha, beta, d.

Level 1

Functionality

Operations

CPU

Intel GPU

Sparse Vector - Dense Vector addition (AXPY)

y <- alpha * w + y

No

No

Sparse Vector - Sparse Vector Dot product (SPDOT) (sv.sv -> sc)

d <- dot(w,v)

N/A

N/A

dot(w,v) = sum(wi* vi)

No

No

dot(w,v) = sum(conj(wi) * vi)

No

No

Sparse Vector - Dense Vector Dot product (SPDOT) (sv.dv -> sc)

d <- dot(w,x)

N/A

N/A

dot(w,v) = sum(wi* vi)

No

No

dot(w,v) = sum(conj(wi) * vi)

No

No

Dense Vector - Sparse Vector Conversion (sv <-> dv)

N/A

N/A

x = scatter(w)

No

No

w = gather(x,windx)

No

No

In the following table for functionality, sm = sparse matrix, dm = dense matrix, sv = sparse vector, dv = dense vector, sc = scalar.

In the following table for operations, dense vectors = x, y, sparse vectors = w,v, dense matrices = X,Y, sparse matrices = A, B, C, and scalars = alpha, beta, d.

Level 2

Functionality

Operations

CPU

Intel GPU

General Matrix-Vector multiplication (GEMV) (sm*dv->dv)

y <- beta*y + alpha * op(A)*x

N/A

N/A

op(A) = A

Yes

Yes

op(A) = AT

Yes

Yes

op(A) = AH

No

No

Symmetric Matrix-Vector multiplication (SYMV) (sm*dv->dv)

y <- beta*y + alpha * op(A)*x

N/A

N/A

op(A) = A

Yes

Yes

op(A) = AT

Yes

Yes

op(A) = AH

No

No

Triangular Matrix-Vector multiplication (TRMV) (sm*dv->dv)

y <- beta*y + alpha * op(A)*x

N/A

N/A

op(A) = A

Yes

No

op(A) = AT

Yes

No

op(A) = AH

No

No

General Matrix-Vector mult with dot product (GEMVDOT) (sm*dv -> dv, dv.dv->sc)

y <- beta*y + alpha * op(A)*x, d = dot(x,y)

N/A

N/A

op(A) = A

Yes

Yes

op(A) = AT

Yes

Yes

op(A) = AH

No

No

Triangular Solve (TRSV) (inv(sm)*dv -> dv)

solve for y, op(A)*y = alpha*x

N/A

N/A

op(A) = A

Yes

Yes

op(A) = AT

Yes

Yes

op(A) = AH

No

No

In the following table for functionality, sm = sparse matrix, dm = dense matrix, sv = sparse vector, dv = dense vector, sc = scalar.

In the following table for operations, dense vectors = x, y, sparse vectors = w,v, dense matrices = X,Y, sparse matrices = A, B, C, and scalars = alpha, beta, d.

Level 3

Functionality

Operations

CPU

Intel GPU

General Sparse Matrix - Dense Matrix Multiplication (GEMM) (sm*dm->dm)

Y <- alpha*op(A)*op(X) + beta*Y

N/A

N/A

op(A) = A, op(X) = X

Yes

Yes

op(A) = AT, op(X) = X

Yes

Yes

op(A) = AH, op(X) = X

Yes

Yes

op(A) = A, op(X) = XT

No

No

op(A) = AT, op(X) = XT

No

No

op(A) = A, op(X) = XH

No

No

op(A) = AH

No

No

op(A) = AT, op(X) = XH

No

No

op(A) = AH, op(X) = XH

No

No

General Dense Matrix - Sparse Matrix Multiplication (GEMM) (dm*sm->dm)

Y <- alpha*op(X)*op(A) + beta*Y

N/A

N/A

op(X) = X, op(A)=A

No

No

op(X) = XH, op(A)=A

No

No

op(X) = XH, op(A)=A

No

No

op(X) = X, op(A)=AH

No

No

op(X) = XH, op(A)=AH

No

No

op(X) = XH, op(A)=AH

No

No

op(X) = X, op(A)=AH

No

No

op(X) = XH, op(A)=AH

No

No

op(X) = XH, op(A)=AH

No

No

General Sparse Matrix - Sparse Matrix Multiplication (GEMM) (sm*sm->sm)

C <- alpha*op(A)*op(B) + beta*C

N/A

N/A

op(A)=A, op(B)=B

No

No

op(A)=AT, op(B)=B

No

No

op(A)=AH, op(B)=B

No

No

op(A)=A, op(B)=BT

No

No

op(A)=AT, op(B)=BT

No

No

op(A)=AH, op(B)=BT

No

No

op(A)=A, op(B)=BH

No

No

op(A)=AT, op(B)=BH

No

No

op(A)=AH, op(B)=BH

No

No

General Sparse Matrix - Sparse Matrix Multiplication (GEMM) (sm*sm->dm)

Y <- alpha*op(A)*op(B) + beta*Y

N/A

N/A

op(A)=A, op(B)=B

No

No

op(A)=AT, op(B)=B

No

No

op(A)=AH, op(B)=B

No

No

op(A)=A, op(B)=BT

No

No

op(A)=AT, op(B)=BT

No

No

op(A)=AH, op(B)=BT

No

No

op(A)=A, op(B)=BH

No

No

op(A)=AT, op(B)=BH

No

No

op(A)=AH, op(B)=BH

No

No

Symmetric Rank-K update (SYRK) (sm*sm->sm)

C <- op(A)*op(A)H

N/A

N/A

op(A)=A

No

No

op(A)=AT

No

No

op(A)=AH

No

No

Symmetric Rank-K update (SYRK) (sm*sm->dm)

Y <- op(A)*op(A)H

N/A

N/A

op(A)=A

No

No

op(A)=AT

No

No

op(A)=AH

No

No

Symmetric Triple Product (SYPR) (op(sm)*sm*sm -> sm)

C <- op(A)*B*op(A)H

N/A

N/A

op(A)=A

No

No

op(A)=AT

No

No

op(A)=AH

No

No

Triangular Solve (TRSM) (inv(sm)*dm -> dm)

solve for Y, op(A)*Y = alpha*X

N/A

N/A

op(A)=A

No

No

op(A)=AT

No

No

op(A)=AH

No

No

In the following table for functionality, sm = sparse matrix, dm = dense matrix, sv = sparse vector, dv = dense vector, sc = scalar.

In the following table for operations, dense vectors = x, y, sparse vectors = w,v, dense matrices = X,Y, sparse matrices = A, B, C, and scalars = alpha, beta, d.

Other

Functionality

Operations

CPU

Intel GPU

Symmetric Gauss-Seidel Preconditioner (SYMGS) (update A*x=b, A=L+D+U)

x0 <- x*alpha; (L+D)*x1=b-U*x0; (U+D)*x=b-L*x1

No

No

Symmetric Gauss-Seidel Preconditioner with Matrix-Vector product (SYMGS_MV) (update A*x=b, A=L+D+U)

x0 <- x*alpha; (L+D)*x1=b-U*x0; (U+D)*x=b-L*x1; y=A*x

No

No

LU Smoother (LU_SMOOTHER) (update A*x=b, A=L+D+U, E~inv(D) )

r=b-A*x; (L+D)*E*(U+D)*dx=r; y=x+dr

No

No

Sparse Matrix Add (ADD)

C <- alpha*op(A) + B

No

No

op(A) = AT

No

No

op(A) = AH

No

No

In the following table for operations, dense vectors = x, y, sparse vectors = w,v, dense matrices = X,Y, sparse matrices = A, B, C, and scalars = alpha, beta, d.

Helper Functions

Functionality

Operations

CPU

Intel GPU

Sort Indices of Matrix (ORDER)

N/A

No

No

Transpose of Sparse Matrix (TRANSPOSE)

A <- op(A) with op=trans or conjtrans

N/A

N/A

transpose CSR/CSC matrix

No

No

transpose BSR matrix

No

No

Sparse Matrix Format Converter (CONVERT)

N/A

No

No

Dense to Sparse Matrix Format Converter (CONVERT)

N/A

No

No

Copy Matrix Handle (COPY)

N/A

No

No

Create CSR Matrix Handle

N/A

Yes

Yes

Create CSC Matrix Handle

N/A

No

No

Create COO Matrix Handle

N/A

No

No

Create BSR Matrix Handle

N/A

No

No

Export CSR Matrix

Allows access to internal data in the CSR Matrix handle

No

No

Export CSC Matrix

Allows access to internal data in the CSC Matrix handle

No

No

Export COO Matrix

Allows access to internal data in the COO Matrix handle

No

No

Export BSR Matrix

Allows access to internal data in the BSR Matrix handle

No

No

Set Value in Matrix

N/A

No

No

In the following table for functionality, sm = sparse matrix, dm = dense matrix, sv = sparse vector, dv = dense vector, sc = scalar.

In the following table for operations, dense vectors = x, y, sparse vectors = w,v, dense matrices = X,Y, sparse matrices = A, B, C, and scalars = alpha, beta, d.

Optimize Stages

Functionality

Operations

CPU

Intel GPU

add MEMORY hint and optimize

Chooses to allow larger memory requiring optimizations or not.

No

No

Add GEMV hint and optimize

N/A

Yes

No

Add SYMV hint and optimize

N/A

Yes

No

Add TRMV hint and optimize

N/A

Yes

No

add TRSV hint and optimize

N/A

Yes

No

add GEMM hint and optimize

N/A

Yes

No

add TRSM hint and optimize

N/A

No

No

add DOTMV hint and optimize

N/A

Yes

No

add SYMGS hint and optimize

N/A

No

No

add SYMGS_MV hint and optimize

N/A

No

No

add LU_SMOOTHER hint and optimize

N/A

No

No