From cb9466f319c2a72866019cafa6006c272d58c1bc Mon Sep 17 00:00:00 2001 From: Aaron Markham Date: Tue, 26 Feb 2019 14:25:39 -0800 Subject: [PATCH 1/5] fix render issue in docs --- src/operator/tensor/la_op.cc | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-) diff --git a/src/operator/tensor/la_op.cc b/src/operator/tensor/la_op.cc index 0f3c2954a0f6..4159b68a3432 100644 --- a/src/operator/tensor/la_op.cc +++ b/src/operator/tensor/la_op.cc @@ -48,11 +48,12 @@ Here, *alpha* and *beta* are scalar parameters, and *op()* is either the identit matrix transposition (depending on *transpose_a*, *transpose_b*). If *n>2*, *gemm* is performed separately for a batch of matrices. The column indices of the matrices -are given by the last dimensions of the tensors, the row indices by the axis specified with the *axis* +are given by the last dimensions of the tensors, the row indices by the axis specified with the *axis* parameter. By default, the trailing two dimensions will be used for matrix encoding. For a non-default axis parameter, the operation performed is equivalent to a series of swapaxes/gemm/swapaxes -calls. For example let *A*, *B*, *C* be 5 dimensional tensors. Then gemm(*A*, *B*, *C*, axis=1) is equivalent to +calls. For example let *A*, *B*, *C* be 5 dimensional tensors. Then gemm(*A*, *B*, *C*, axis=1) is equivalent +to the following without the overhead of the additional swapaxis operations:: A1 = swapaxes(A, dim1=1, dim2=3) B1 = swapaxes(B, dim1=1, dim2=3) @@ -60,8 +61,6 @@ calls. For example let *A*, *B*, *C* be 5 dimensional tensors. Then gemm(*A*, *B C = gemm(A1, B1, C) C = swapaxis(C, dim1=1, dim2=3) -without the overhead of the additional swapaxis operations. - When the input data is of type float32 and the environment variables MXNET_CUDA_ALLOW_TENSOR_CORE and MXNET_CUDA_TENSOR_OP_MATH_ALLOW_CONVERSION are set to 1, this operator will try to use pseudo-float16 precision (float32 math with float16 I/O) precision in order to use @@ -126,19 +125,18 @@ Here *alpha* is a scalar parameter and *op()* is either the identity or the matr transposition (depending on *transpose_a*, *transpose_b*). If *n>2*, *gemm* is performed separately for a batch of matrices. The column indices of the matrices -are given by the last dimensions of the tensors, the row indices by the axis specified with the *axis* +are given by the last dimensions of the tensors, the row indices by the axis specified with the *axis* parameter. By default, the trailing two dimensions will be used for matrix encoding. For a non-default axis parameter, the operation performed is equivalent to a series of swapaxes/gemm/swapaxes calls. For example let *A*, *B* be 5 dimensional tensors. Then gemm(*A*, *B*, axis=1) is equivalent to +the following without the overhead of the additional swapaxis operations:: A1 = swapaxes(A, dim1=1, dim2=3) B1 = swapaxes(B, dim1=1, dim2=3) C = gemm2(A1, B1) C = swapaxis(C, dim1=1, dim2=3) -without the overhead of the additional swapaxis operations. - When the input data is of type float32 and the environment variables MXNET_CUDA_ALLOW_TENSOR_CORE and MXNET_CUDA_TENSOR_OP_MATH_ALLOW_CONVERSION are set to 1, this operator will try to use pseudo-float16 precision (float32 math with float16 I/O) precision in order to use From e0bbcdd0fa9f136810b2ae77d0344297ffd5f9c9 Mon Sep 17 00:00:00 2001 From: Aaron Markham Date: Wed, 27 Feb 2019 10:56:00 -0800 Subject: [PATCH 2/5] nudge ci --- src/operator/tensor/la_op.cc | 1 - 1 file changed, 1 deletion(-) diff --git a/src/operator/tensor/la_op.cc b/src/operator/tensor/la_op.cc index 4159b68a3432..9981b3867906 100644 --- a/src/operator/tensor/la_op.cc +++ b/src/operator/tensor/la_op.cc @@ -315,7 +315,6 @@ If *n>2*, *trmm* is performed separately on the trailing two dimensions for all .. note:: The operator supports float32 and float64 data types only. - Examples:: // Single triangular matrix multiply From 14b19f3bf782fa0d756d49ee861710b4edf31274 Mon Sep 17 00:00:00 2001 From: Aaron Markham Date: Mon, 4 Mar 2019 15:14:46 -0800 Subject: [PATCH 3/5] update copyright year; mostly for bumping ci --- src/operator/tensor/la_op.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/operator/tensor/la_op.cc b/src/operator/tensor/la_op.cc index 9981b3867906..82b525f9f756 100644 --- a/src/operator/tensor/la_op.cc +++ b/src/operator/tensor/la_op.cc @@ -18,7 +18,7 @@ */ /*! - * Copyright (c) 2017 by Contributors + * Copyright (c) 2019 by Contributors * \file la_op.cc * \brief CPU implementation of Operators for advanced linear algebra. */ From d3afa774b510e0af07bda6c518b22f3a48054e76 Mon Sep 17 00:00:00 2001 From: Aaron Markham Date: Tue, 5 Mar 2019 12:35:02 -0800 Subject: [PATCH 4/5] bumping ci --- src/operator/tensor/la_op.cc | 1 + 1 file changed, 1 insertion(+) diff --git a/src/operator/tensor/la_op.cc b/src/operator/tensor/la_op.cc index 82b525f9f756..9853f1f1d0a2 100644 --- a/src/operator/tensor/la_op.cc +++ b/src/operator/tensor/la_op.cc @@ -22,6 +22,7 @@ * \file la_op.cc * \brief CPU implementation of Operators for advanced linear algebra. */ + #include "./la_op.h" #include "./la_op-inl.h" From f74b02646b64be8e51fb54d824cc1a9e4127c9e7 Mon Sep 17 00:00:00 2001 From: JackieWu Date: Wed, 6 Mar 2019 12:03:28 +0800 Subject: [PATCH 5/5] delete whitespace --- src/operator/tensor/la_op.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/operator/tensor/la_op.cc b/src/operator/tensor/la_op.cc index 9853f1f1d0a2..6246a4bb213b 100644 --- a/src/operator/tensor/la_op.cc +++ b/src/operator/tensor/la_op.cc @@ -22,7 +22,7 @@ * \file la_op.cc * \brief CPU implementation of Operators for advanced linear algebra. */ - + #include "./la_op.h" #include "./la_op-inl.h"