aboutsummaryrefslogtreecommitdiff
path: root/docs/user_guide/operator_list.dox
diff options
context:
space:
mode:
Diffstat (limited to 'docs/user_guide/operator_list.dox')
-rw-r--r--docs/user_guide/operator_list.dox227
1 files changed, 180 insertions, 47 deletions
diff --git a/docs/user_guide/operator_list.dox b/docs/user_guide/operator_list.dox
index 05cc892d40..e7f1823f8b 100644
--- a/docs/user_guide/operator_list.dox
+++ b/docs/user_guide/operator_list.dox
@@ -1,5 +1,5 @@
///
-/// Copyright (c) 2021 Arm Limited.
+/// Copyright (c) 2021-2024 Arm Limited.
///
/// SPDX-License-Identifier: MIT
///
@@ -52,9 +52,10 @@ Compute Library supports the following data layouts (fast changing dimension fro
<ul>
<li>NHWC: The native layout of Compute Library that delivers the best performance where channels are in the fastest changing dimension
<li>NCHW: Legacy layout where width is in the fastest changing dimension
+ <li>NDHWC: New data layout for supporting 3D operators
<li>All: Agnostic to any specific data layout
</ul>
-where N = batches, C = channels, H = height, W = width
+where N = batches, C = channels, H = height, W = width, D = depth
<table>
<caption id="multi_row"></caption>
@@ -108,6 +109,26 @@ where N = batches, C = channels, H = height, W = width
<tr><td>F32<td>F32
</table>
<tr>
+ <td rowspan="1">AddMulAdd
+ <td rowspan="1" style="width:200px;"> Performs a fused Add + Mul + Add [+ Relu-based-Activation] operation.
+ <td rowspan="1">
+ <ul>
+ <li>n/a
+ </ul>
+ <td>NEAddMulAdd
+ <td>
+ <ul>
+ <li>Any
+ </ul>
+ <td>
+ <table>
+ <tr><th>input1<th>input2<th>bn_mul<th>bn_add<th>add_output<th>final_output
+ <tr><td>QASYMM8<td>QASYMM8<td>QASYMM8<td>QASYMM8<td>QASYMM8<td>QASYMM8
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>QASYMM8_SIGNED
+ <tr><td>F16<td>F16<td>F16<td>F16<td>F16<td>F16
+ <tr><td>F32<td>F32<td>F32<td>F32<td>F32<td>F32
+ </table>
+<tr>
<td rowspan="2">ArgMinMaxLayer
<td rowspan="2" style="width:200px;"> Function to calculate the index of the minimum or maximum values in a tensor based on an axis.
<td rowspan="2">
@@ -125,7 +146,7 @@ where N = batches, C = channels, H = height, W = width
<tr><th>src<th>dst
<tr><td>QASYMM8<td>U32, S32
<tr><td>QASYMM8_SIGNED<td>U32, S32
- <tr><td>S32<td>U32, S32
+ <tr><td>S32<td>U32, S32, S64
<tr><td>F16<td>U32, S32
<tr><td>F32<td>U32, S32
</table>
@@ -164,9 +185,6 @@ where N = batches, C = channels, H = height, W = width
<tr><td>QSYMM16<td>QSYMM16<td>QASYMM16
<tr><td>QSYMM16<td>QSYMM16<td>S32
<tr><td>U8<td>U8<td>U8
- <tr><td>U8<td>U8<td>S16
- <tr><td>U8<td>S16<td>S16
- <tr><td>S16<td>U8<td>S16
<tr><td>S16<td>S16<td>S16
<tr><td>S32<td>S32<td>S32
<tr><td>F16<td>F16<td>F16
@@ -192,9 +210,6 @@ where N = batches, C = channels, H = height, W = width
<tr><td>QSYMM16<td>QSYMM16<td>QASYMM16
<tr><td>QSYMM16<td>QSYMM16<td>S32
<tr><td>U8<td>U8<td>U8
- <tr><td>U8<td>U8<td>S16
- <tr><td>U8<td>S16<td>S16
- <tr><td>S16<td>U8<td>S16
<tr><td>S16<td>S16<td>S16
<tr><td>S32<td>S32<td>S32
<tr><td>F16<td>F16<td>F16
@@ -442,12 +457,15 @@ where N = batches, C = channels, H = height, W = width
<table>
<tr><th>src<th>dst
<tr><td>U8<td>S8, U16, S16, U32, S32, F16, F32
+ <tr><td>S8<td>U8, U16, S16, U32, S32, F16, F32
<tr><td>U16<td>U8, S8, S16, U32, S32, F16, F32
<tr><td>S16<td>U8, S8, U16, U32, S32, F16, F32
<tr><td>U32<td>U8, S8, U16, S16, S32, F16, F32
<tr><td>S32<td>U8, S8, U16, S16, U32, F16, F32
- <tr><td>F16<td>U8, S8, U16, S16, U32, F32
- <tr><td>F32<td>U8, S8, U16, S16, U32, F16
+ <tr><td>U64<td>U8, S8, U16, S16, U32, S32, F16, F32
+ <tr><td>S64<td>U8, S8, U16, S16, U32, S32, F16, F32
+ <tr><td>F16<td>U8, S8, U16, S16, S32, U32, F32
+ <tr><td>F32<td>U8, S8, U16, S16, S32, U32, F16
</table>
<tr>
<td rowspan="2">ChannelShuffleLayer
@@ -460,6 +478,7 @@ where N = batches, C = channels, H = height, W = width
<td>
<ul>
<li>NCHW
+ <li>NHWC
</ul>
<td>
<table>
@@ -471,6 +490,7 @@ where N = batches, C = channels, H = height, W = width
<td>
<ul>
<li>NCHW
+ <li>NHWC
</ul>
<td>
<table>
@@ -604,6 +624,40 @@ where N = batches, C = channels, H = height, W = width
<tr><td>QASYMM8_SIGNED<td>QSYMM8_PER_CHANNEL<td>S32<td>QASYMM8_SIGNED
</table>
<tr>
+ <td rowspan="2">Conv3D
+ <td rowspan="2" style="width:200px;"> Function to compute a 3d convolution layer.
+ <td rowspan="2">
+ <ul>
+ <li>ANEURALNETWORKS_CONV_3D
+ </ul>
+ <td>NEConv3D
+ <td>
+ <ul>
+ <li>NDHWC
+ </ul>
+ <td>
+ <table>
+ <tr><th>src0<th>src1<th>src2<th>dst
+ <tr><td>F16<td>F16<td>F16<td>F16
+ <tr><td>F32<td>F32<td>F32<td>F32
+ <tr><td>QASYMM8<td>QASYMM8<td>S32<td>QASYMM8
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>S32<td>QASYMM8_SIGNED
+ </table>
+<tr>
+ <td>CLConv3D
+ <td>
+ <ul>
+ <li>NDHWC
+ </ul>
+ <td>
+ <table>
+ <tr><th>src0<th>src1<th>src2<th>dst
+ <tr><td>F16<td>F16<td>F16<td>F16
+ <tr><td>F32<td>F32<td>F32<td>F32
+ <tr><td>QASYMM8<td>QASYMM8<td>S32<td>QASYMM8
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>S32<td>QASYMM8_SIGNED
+ </table>
+<tr>
<td rowspan="2">Copy
<td rowspan="2" style="width:200px;"> Function to copy a tensor.
<td rowspan="2">
@@ -1408,9 +1462,9 @@ where N = batches, C = channels, H = height, W = width
<tr><td>All<td>All
</table>
<tr>
- <td rowspan="2">FillBorder
- <td rowspan="2" style="width:200px;"> Function to fill the borders within the XY-planes.
- <td rowspan="2">
+ <td rowspan="1">FillBorder
+ <td rowspan="1" style="width:200px;"> Function to fill the borders within the XY-planes.
+ <td rowspan="1">
<ul>
<li>n/a
</ul>
@@ -1425,17 +1479,6 @@ where N = batches, C = channels, H = height, W = width
<tr><td>All<td>All
</table>
<tr>
- <td>CLFillBorder
- <td>
- <ul>
- <li>All
- </ul>
- <td>
- <table>
- <tr><th>src<th>dst
- <tr><td>All<td>All
- </table>
-<tr>
<td rowspan="2">FlattenLayer
<td rowspan="2" style="width:200px;"> Reshape a tensor to be 1D
<td rowspan="2">
@@ -1730,6 +1773,7 @@ where N = batches, C = channels, H = height, W = width
<tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>S32<td>S32
<tr><td>QASYMM8_SIGNED<td>QSYMM8_PER_CHANNEL<td>S32<td>S32
<tr><td>QASYMM8_SIGNED<td>QSYMM8<td>S32<td>S32
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>F32<td>F32
</table>
<tr>
<td>CLGEMMLowpMatrixMultiplyCore
@@ -2032,6 +2076,41 @@ where N = batches, C = channels, H = height, W = width
<tr><td>QASYMM8<td>S32<td>QSYMM16<td>QASYMM8<td>QSYMM16<td>QASYMM8
</table>
<tr>
+ <td rowspan="2">MatMul
+ <td rowspan="2" style="width:200px;"> Computes a matrix multiplication in batches.
+ <td rowspan="2">
+ <ul>
+ <li>ANEURALNETWORKS_BATCH_MATMUL
+ </ul>
+ <td>NEMatMul
+ <td>
+ <ul>
+ <li>Any
+ </ul>
+ <td>
+ <table>
+ <tr><th>lhs<th>rhs<th>dst
+ <tr><td>F32<td>F32<td>F32
+ <tr><td>F16<td>F16<td>F16
+ <tr><td>BFLOAT16<td>BFLOAT16<td>BFLOAT16
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>QASYMM8_SIGNED
+ <tr><td>QASYMM8<td>QASYMM8<td>QASYMM8
+ </table>
+<tr>
+ <td>CLMatMul
+ <td>
+ <ul>
+ <li>All
+ </ul>
+ <td>
+ <table>
+ <tr><th>lhs<th>rhs<th>dst
+ <tr><td>F32<td>F32<td>F32
+ <tr><td>F16<td>F16<td>F16
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>QASYMM8_SIGNED
+ <tr><td>QASYMM8<td>QASYMM8<td>QASYMM8
+ </table>
+<tr>
<td rowspan="2">MaxUnpoolingLayer
<td rowspan="2" style="width:200px;"> Function to perform MaxUnpooling.
<td rowspan="2">
@@ -2132,6 +2211,27 @@ where N = batches, C = channels, H = height, W = width
<tr><td>F16<td>F16
</table>
<tr>
+ <td rowspan="1">NormalizePlanarYUVLayer
+ <td rowspan="1" style="width:200px;"> Function to compute normalization planar YUV layer.
+ <td rowspan="1">
+ <ul>
+ <li>n/a
+ </ul>
+ <td>CLNormalizePlanarYUVLayer
+ <td>
+ <ul>
+ <li>NHWC
+ <li>NCHW
+ </ul>
+ <td>
+ <table>
+ <tr><th>src<th>dst
+ <tr><td>F32<td>F32
+ <tr><td>F16<td>F16
+ <tr><td>QASYMM8<td>QASYMM8
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED
+ </table>
+<tr>
<td rowspan="2">PadLayer
<td rowspan="2" style="width:200px;"> Function to pad a tensor.
<td rowspan="2">
@@ -2280,6 +2380,40 @@ where N = batches, C = channels, H = height, W = width
<tr><td>F32<td>F32
</table>
<tr>
+ <td rowspan="2">Pooling3dLayer
+ <td rowspan="2" style="width:200px;"> Function to perform pooling 3D with the specified pooling operation.
+ <td rowspan="2">
+ <ul>
+ <li>N/A
+ </ul>
+ <td>NEPooling3dLayer
+ <td>
+ <ul>
+ <li>NDHWC
+ </ul>
+ <td>
+ <table>
+ <tr><th>src<th>dst
+ <tr><td>F16<td>F16
+ <tr><td>F32<td>F32
+ <tr><td>QASYMM8<td>QASYMM8
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED
+ </table>
+<tr>
+ <td>CLPooling3dLayer
+ <td>
+ <ul>
+ <li>NDHWC
+ </ul>
+ <td>
+ <table>
+ <tr><th>src<th>dst
+ <tr><td>F16<td>F16
+ <tr><td>F32<td>F32
+ <tr><td>QASYMM8<td>QASYMM8
+ <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED
+ </table>
+<tr>
<td rowspan="2">PReluLayer
<td rowspan="2" style="width:200px;"> Function to compute the activation layer with the PRELU activation function.
<td rowspan="2">
@@ -2525,6 +2659,23 @@ where N = batches, C = channels, H = height, W = width
<tr><td>S32<td>S32
</table>
<tr>
+ <td rowspan="1">ReorderLayer
+ <td rowspan="1" style="width:200px;"> Reorders a tensor to a different weights format.
+ <td rowspan="1">
+ <ul>
+ <li>n/a
+ </ul>
+ <td>NEReorderLayer
+ <td>
+ <ul>
+ <li>NCHW
+ </ul>
+ <td>
+ <table>
+ <tr><th>src<th>dst
+ <tr><td>F32<td>F32
+ </table>
+<tr>
<td rowspan="2">ReorgLayer
<td rowspan="2" style="width:200px;"> Performs a reorganization layer of input tensor to the output tensor.
<td rowspan="2">
@@ -2598,7 +2749,7 @@ where N = batches, C = channels, H = height, W = width
<td>
<table>
<tr><th>src0<th>src1<th>dst
- <tr><td>All<td>U32<td>All
+ <tr><td>All<td>U32, S32<td>All
</table>
<tr>
<td>CLReverse
@@ -2609,7 +2760,7 @@ where N = batches, C = channels, H = height, W = width
<td>
<table>
<tr><th>src0<th>src1<th>dst
- <tr><td>All<td>U32<td>All
+ <tr><td>All<td>U32, S32<td>All
</table>
<tr>
<td rowspan="2">RNNLayer
@@ -2730,6 +2881,7 @@ where N = batches, C = channels, H = height, W = width
<tr><td>F16<td>F16
<tr><td>F32<td>F32
<tr><td>U8<td>U8
+ <tr><td>S8<td>S8
<tr><td>S16<td>S16
</table>
<tr>
@@ -3100,26 +3252,7 @@ where N = batches, C = channels, H = height, W = width
<tr><td>F16<td>F16<td>F16<td>F16
<tr><td>F32<td>F32<td>F32<td>F32
</table>
-<tr>
- <td rowspan="1">WinogradInputTransform
- <td rowspan="1" style="width:200px;"> Function to perform a Winograd transform on the input tensor.
- <td rowspan="1">
- <ul>
- <li>n/a
- </ul>
- <td>CLWinogradInputTransform
- <td>
- <ul>
- <li>NHWC
- <li>NCHW
- </ul>
- <td>
- <table>
- <tr><th>src<th>dst
- <tr><td>F16<td>F16
- <tr><td>F32<td>F32
- </table>
</table>
*/
-} // namespace \ No newline at end of file
+} // namespace
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-03 18:54:25 +0000