summaryrefslogtreecommitdiff
path: root/scripts
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2018-05-06 11:35:34 +0200
committerCedric Nugteren <web@cedricnugteren.nl>2018-05-06 11:35:34 +0200
commit2d1f6ba7fe842ba938490fc599b6ebd209b6560b (patch)
treef1a284e5dc0163b7fed938a3efeb39432b9d3788 /scripts
parent2776d761768295b01a8be7c333dbb337805d7f77 (diff)
Added convgemm skeleton, test infrastructure, and first reference implementation
Diffstat (limited to 'scripts')
-rwxr-xr-xscripts/generator/generator.py2
1 files changed, 1 insertions, 1 deletions
diff --git a/scripts/generator/generator.py b/scripts/generator/generator.py
index e2837dd5..f04d9f3d 100755
--- a/scripts/generator/generator.py
+++ b/scripts/generator/generator.py
@@ -181,7 +181,7 @@ ROUTINES = [
Routine(True, True, 0, False, "x", "had", T, [S,D,C,Z,H], ["n"], [], ["x","y"], ["z"], [xn,yn,zn], ["alpha","beta"], "", "Element-wise vector product (Hadamard)", "Performs the Hadamard element-wise product _z = alpha * x * y + beta * z_, in which _x_, _y_, and _z_ are vectors and _alpha_ and _beta_ are scalar constants.", []),
Routine(True, True, 0, False, "x", "omatcopy", T, [S,D,C,Z,H], ["m","n"], ["layout","a_transpose"], ["a"], ["b"], [amn,bnma], ["alpha"], "", "Scaling and out-place transpose/copy (non-BLAS function)", "Performs scaling and out-of-place transposition/copying of matrices according to _B = alpha*op(A)_, in which _A_ is an input matrix (_m_ rows by _n_ columns), _B_ an output matrix, and _alpha_ a scalar value. The operation _op_ can be a normal matrix copy, a transposition or a conjugate transposition.", [ald_m, bld_n]),
Routine(True, True, 0, False, "x", "im2col", T, [S,D,C,Z,H], im2col_constants, [], ["im"], ["col"], [im,col], [""], "", "Im2col function (non-BLAS function)", "Performs the im2col algorithm, in which _im_ is the input matrix and _col_ is the output matrix.", []),
- Routine(False, True, 0, False, "x", "convgemm", T, [S,D,C,Z,H], convgemm_constants, [], ["im","kernel"], ["result"], [imb,kernel,result],[""], "", "Batched convolution as GEMM (non-BLAS function)", "Integrates im2col and GEMM for batched convolution, in which _im_ is the 4D input tensor, _kernel_ the 3D kernel weights tensor, and _result_ the 4D output tensor.", []),
+ Routine(True, True, 0, False, "x", "convgemm", T, [S,D,C,Z,H], convgemm_constants, [], ["im","kernel"], ["result"], [imb,kernel,result],[""], "", "Batched convolution as GEMM (non-BLAS function)", "Integrates im2col and GEMM for batched 3D convolution, in which _im_ is the 4D input tensor, _kernel_ the 4D kernel weights tensor, and _result_ the 4D output tensor.", []),
# Batched routines:
Routine(True, True, 1, False, "x", "axpy", T, [S,D,C,Z,H], ["n"], [], ["x"], ["y"], [xn,yn], ["alpha"], "", "Batched version of AXPY", "As AXPY, but multiple operations are batched together for better performance.", []),
Routine(True, True, 1, False, "x", "gemm", T, [S,D,C,Z,H], ["m","n","k"], ["layout","a_transpose","b_transpose"], ["a","b"], ["c"], [amk,bkn,cmn], ["alpha","beta"], "", "Batched version of GEMM", "As GEMM, but multiple operations are batched together for better performance.", [ald_transa_m_k, bld_transb_k_n, cld_m]),