Merge "FL7: Refine MIRROR_PAD specification" am: f914da59a6 am: 7dc880a386
Original change: https://android-review.googlesource.com/c/platform/hardware/interfaces/+/1940673
Change-Id: Ia333ba1f744aa73ff51ddf8745d69384ecc41d3b
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl
index 0ad254d..5f7810b 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl
@@ -5331,6 +5331,18 @@
/**
* Pads a tensor with mirrored values.
*
+ * This operator specifies one of two padding modes: REFLECT or SYMMETRIC.
+ * In the case of REFLECT mode, the mirroring excludes the border element
+ * on the padding side.
+ * In the case of SYMMETRIC mode, the mirroring includes the border element
+ * on the padding side.
+ *
+ * For example, if the input is the 1-D tensor `[1, 2, 3]` and the padding
+ * is `[0, 2]` (i.e., pad no elements before the first (and only) dimension,
+ * and two elements after the first (and only) dimension), then:
+ * - REFLECT mode produces the output `[1, 2, 3, 2, 1]`
+ * - SYMMETRIC mode produces the output `[1, 2, 3, 3, 2]`
+ *
* Supported tensor {@link OperandType}:
* * {@link OperandType::TENSOR_FLOAT16}
* * {@link OperandType::TENSOR_FLOAT32}
@@ -5349,6 +5361,11 @@
* front of dimension i.
* padding[i, 1] specifies the number of elements to be padded after the
* end of dimension i.
+ * Each padding value must be nonnegative.
+ * In the case of REFLECT mode, each padding value must be less than the
+ * corresponding dimension.
+ * In the case of SYMMETRIC mode, each padding value must be less than or
+ * equal to the corresponding dimension.
* * 2: An {@link OperandType::INT32} scalar, specifying the mode.
* Options are 0:REFLECT and 1:SYMMETRIC.
*