COMPMID-2310: CLGenerateProposalsLayer: support for QASYMM8

Change-Id: I48b77e09857cd43f9498d28e8f4bf346e3d7110d
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1969
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 40 insertions, 26 deletions

diff --git a/arm_compute/core/CL/kernels/CLGenerateProposalsLayerKernel.h b/arm_compute/core/CL/kernels/CLGenerateProposalsLayerKernel.h
index 5900d79821..e2b20f667f 100644
--- a/arm_compute/core/CL/kernels/CLGenerateProposalsLayerKernel.h
+++ b/arm_compute/core/CL/kernels/CLGenerateProposalsLayerKernel.h
@@ -48,7 +48,7 @@ public:
 
     /** Set the input and output tensors.
      *
-     * @param[in]  anchors     Source tensor. Original set of anchors of size (4, A), where A is the number of anchors. Data types supported: F16/F32
+     * @param[in]  anchors     Source tensor. Original set of anchors of size (4, A), where A is the number of anchors. Data types supported: QSYMM16/F16/F32
      * @param[out] all_anchors Destination tensor. Destination anchors of size (4, H*W*A) where H and W are the height and width of the feature map and A is the number of anchors. Data types supported: Same as @p input
      * @param[in]  info        Contains Compute Anchors operation information described in @ref ComputeAnchorsInfo
      *
@@ -57,7 +57,7 @@ public:
 
     /** Static function to check if given info will lead to a valid configuration of @ref CLComputeAllAnchorsKernel
      *
-     * @param[in] anchors     Source tensor info. Original set of anchors of size (4, A), where A is the number of anchors. Data types supported: F16/F32
+     * @param[in] anchors     Source tensor info. Original set of anchors of size (4, A), where A is the number of anchors. Data types supported: QSYMM16/F16/F32
      * @param[in] all_anchors Destination tensor info. Destination anchors of size (4, H*W*A) where H and W are the height and width of the feature map and A is the number of anchors. Data types supported: Same as @p input
      * @param[in] info        Contains Compute Anchors operation information described in @ref ComputeAnchorsInfo
      *
diff --git a/arm_compute/core/CL/kernels/CLStridedSliceKernel.h b/arm_compute/core/CL/kernels/CLStridedSliceKernel.h
index 5b69b3fd16..d579d1ceb9 100644
--- a/arm_compute/core/CL/kernels/CLStridedSliceKernel.h
+++ b/arm_compute/core/CL/kernels/CLStridedSliceKernel.h
@@ -54,7 +54,7 @@ public:
      *
      * @note Supported tensor rank: up to 4
      *
-     * @param[in]  input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QSYMM16/U32/S32/F16/F32
+     * @param[in]  input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QASYMM16/QSYMM16/U32/S32/F16/F32
      * @param[out] output           Destination tensor. Data type supported: Same as @p input
      * @param[in]  starts           The starts of the dimensions of the input tensor to be sliced. The length must be of rank(input).
      * @param[in]  ends             The ends of the dimensions of the input tensor to be sliced. The length must be of rank(input).
@@ -72,7 +72,7 @@ public:
      *
      * @note Supported tensor rank: up to 4
      *
-     * @param[in] input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QSYMM16/U32/S32/F16/F32
+     * @param[in] input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QASYMM16/QSYMM16/U32/S32/F16/F32
      * @param[in] output           Destination tensor. Data type supported: Same as @p input
      * @param[in] starts           The starts of the dimensions of the input tensor to be sliced. The length must be of rank(input).
      * @param[in] ends             The ends of the dimensions of the input tensor to be sliced. The length must be of rank(input).
diff --git a/arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h b/arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h
index 8546261fef..827f19d130 100644
--- a/arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h
@@ -24,16 +24,17 @@
 #ifndef __ARM_COMPUTE_CLGENERATEPROPOSALSLAYER_H__
 #define __ARM_COMPUTE_CLGENERATEPROPOSALSLAYER_H__
 #include "arm_compute/core/CL/kernels/CLBoundingBoxTransformKernel.h"
+#include "arm_compute/core/CL/kernels/CLDequantizationLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLGenerateProposalsLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLPadLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLPermuteKernel.h"
+#include "arm_compute/core/CL/kernels/CLQuantizationLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLReshapeLayerKernel.h"
-#include "arm_compute/core/CL/kernels/CLStridedSliceKernel.h"
-#include "arm_compute/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CPP/CPPScheduler.h"
+#include "arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/MemoryGroup.h"
 
@@ -47,10 +48,11 @@ class ICLTensor;
  * -# @ref CLComputeAllAnchors
  * -# @ref CLPermute x 2
  * -# @ref CLReshapeLayer x 2
- * -# @ref CLStridedSlice x 3
  * -# @ref CLBoundingBoxTransform
  * -# @ref CLPadLayerKernel
- * And the following CPP kernels:
+ * -# @ref CLDequantizationLayerKernel
+ * -# @ref CLQuantizationLayerKernel
+ * And the following CPP functions:
  * -# @ref CPPBoxWithNonMaximaSuppressionLimit
  */
 class CLGenerateProposalsLayer : public IFunction
@@ -72,11 +74,13 @@ public:
 
     /** Set the input and output tensors.
      *
-     * @param[in]  scores              Scores from convolution layer of size (W, H, A), where H and W are the height and width of the feature map, and A is the number of anchors. Data types supported: F16/F32
+     * @param[in]  scores              Scores from convolution layer of size (W, H, A), where H and W are the height and width of the feature map, and A is the number of anchors.
+     *                                 Data types supported: QASYMM8/F16/F32
      * @param[in]  deltas              Bounding box deltas from convolution layer of size (W, H, 4*A). Data types supported: Same as @p scores
-     * @param[in]  anchors             Anchors tensor of size (4, A). Data types supported: Same as @p input
-     * @param[out] proposals           Box proposals output tensor of size (5, W*H*A). Data types supported: Same as @p input
-     * @param[out] scores_out          Box scores output tensor of size (W*H*A). Data types supported: Same as @p input
+     * @param[in]  anchors             Anchors tensor of size (4, A). Data types supported: QSYMM16 with scale of 0.125 if @p scores is QASYMM8, otherwise same as @p scores
+     * @param[out] proposals           Box proposals output tensor of size (5, W*H*A).
+     *                                 Data types supported: QASYMM16 with scale of 0.125 and 0 offset if @p scores is QASYMM8, otherwise same as @p scores
+     * @param[out] scores_out          Box scores output tensor of size (W*H*A). Data types supported: Same as @p scores
      * @param[out] num_valid_proposals Scalar output tensor which says which of the first proposals are valid. Data types supported: U32
      * @param[in]  info                Contains GenerateProposals operation information described in @ref GenerateProposalsInfo
      *
@@ -88,12 +92,14 @@ public:
 
     /** Static function to check if given info will lead to a valid configuration of @ref CLGenerateProposalsLayer
      *
-     * @param[in] scores              Scores info from convolution layer of size (W, H, A), where H and W are the height and width of the feature map, and A is the number of anchors. Data types supported: F16/F32
+     * @param[in] scores              Scores info from convolution layer of size (W, H, A), where H and W are the height and width of the feature map, and A is the number of anchors.
+     *                                Data types supported: QASYMM8/F16/F32
      * @param[in] deltas              Bounding box deltas info from convolution layer of size (W, H, 4*A). Data types supported: Same as @p scores
-     * @param[in] anchors             Anchors tensor info of size (4, A). Data types supported: Same as @p input
-     * @param[in] proposals           Box proposals info  output tensor of size (5, W*H*A). Data types supported: Data types supported: U32
-     * @param[in] scores_out          Box scores output tensor info of size (W*H*A). Data types supported: Same as @p input
-     * @param[in] num_valid_proposals Scalar output tensor info which says which of the first proposals are valid. Data types supported: Same as @p input
+     * @param[in] anchors             Anchors tensor of size (4, A). Data types supported: QSYMM16 with scale of 0.125 if @p scores is QASYMM8, otherwise same as @p scores
+     * @param[in] proposals           Box proposals info  output tensor of size (5, W*H*A).
+     *                                Data types supported: QASYMM16 with scale of 0.125 and 0 offset if @p scores is QASYMM8, otherwise same as @p scores
+     * @param[in] scores_out          Box scores output tensor info of size (W*H*A). Data types supported: Same as @p scores
+     * @param[in] num_valid_proposals Scalar output tensor info which says which of the first proposals are valid. Data types supported: U32
      * @param[in] info                Contains GenerateProposals operation information described in @ref GenerateProposalsInfo
      *
      * @return a Status
@@ -117,23 +123,33 @@ private:
     CLComputeAllAnchorsKernel    _compute_anchors_kernel;
     CLBoundingBoxTransformKernel _bounding_box_kernel;
     CLPadLayerKernel             _pad_kernel;
+    CLDequantizationLayerKernel  _dequantize_anchors;
+    CLDequantizationLayerKernel  _dequantize_deltas;
+    CLQuantizationLayerKernel    _quantize_all_proposals;
 
-    // CPP kernels
-    CPPBoxWithNonMaximaSuppressionLimitKernel _cpp_nms_kernel;
+    // CPP functions
+    CPPBoxWithNonMaximaSuppressionLimit _cpp_nms;
 
     bool _is_nhwc;
+    bool _is_qasymm8;
 
     // Temporary tensors
     CLTensor _deltas_permuted;
     CLTensor _deltas_flattened;
+    CLTensor _deltas_flattened_f32;
     CLTensor _scores_permuted;
     CLTensor _scores_flattened;
     CLTensor _all_anchors;
+    CLTensor _all_anchors_f32;
     CLTensor _all_proposals;
+    CLTensor _all_proposals_quantized;
     CLTensor _keeps_nms_unused;
     CLTensor _classes_nms_unused;
     CLTensor _proposals_4_roi_values;
 
+    // Temporary tensor pointers
+    CLTensor *_all_proposals_to_use;
+
     // Output tensor pointers
     ICLTensor *_num_valid_proposals;
     ICLTensor *_scores_out;
diff --git a/arm_compute/runtime/CL/functions/CLSlice.h b/arm_compute/runtime/CL/functions/CLSlice.h
index acd4f0d3ad..5e8d0199c2 100644
--- a/arm_compute/runtime/CL/functions/CLSlice.h
+++ b/arm_compute/runtime/CL/functions/CLSlice.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,7 +42,7 @@ public:
      * @note End coordinates can be negative, which represents the number of elements before the end of that dimension.
      * @note End indices are not inclusive unless negative.
      *
-     * @param[in]  input  Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32
+     * @param[in]  input  Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QASYMM16/QSYMM16/U32/S32/F16/F32
      * @param[out] output Destination tensor. Data type supported: Same as @p input
      * @param[in]  starts The starts of the dimensions of the input tensor to be sliced. The length must be of rank(input).
      * @param[in]  ends   The ends of the dimensions of the input tensor to be sliced. The length must be of rank(input).
@@ -56,7 +56,7 @@ public:
      * @note End coordinates can be negative, which represents the number of elements before the end of that dimension.
      * @note End indices are not inclusive unless negative.
      *
-     * @param[in] input  Source tensor info. Data type supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32
+     * @param[in] input  Source tensor info. Data type supported: U8/S8/QASYMM8/U16/S16/QASYMM16/QSYMM16/U32/S32/F16/F32
      * @param[in] output Destination tensor info. Data type supported: Same as @p input
      * @param[in] starts The starts of the dimensions of the input tensor to be sliced. The length must be of rank(input).
      * @param[in] ends   The ends of the dimensions of the input tensor to be sliced. The length must be of rank(input).
diff --git a/arm_compute/runtime/CL/functions/CLStridedSlice.h b/arm_compute/runtime/CL/functions/CLStridedSlice.h
index bb97b17fea..885751788c 100644
--- a/arm_compute/runtime/CL/functions/CLStridedSlice.h
+++ b/arm_compute/runtime/CL/functions/CLStridedSlice.h
@@ -39,7 +39,7 @@ public:
      *
      * @note Supported tensor rank: up to 4
      *
-     * @param[in]  input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QSYMM16/U32/S32/F16/F32
+     * @param[in]  input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QASYMM16/QSYMM16/U32/S32/F16/F32
      * @param[out] output           Destination tensor. Data type supported: Same as @p input
      * @param[in]  starts           The starts of the dimensions of the input tensor to be sliced. The length must be of rank(input).
      * @param[in]  ends             The ends of the dimensions of the input tensor to be sliced. The length must be of rank(input).
@@ -57,7 +57,7 @@ public:
      *
      * @note Supported tensor rank: up to 4
      *
-     * @param[in] input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QSYMM16/U32/S32/F16/F32
+     * @param[in] input            Source tensor. Data type supported: U8/S8/QASYMM8/U16/S16/QASYMM16/QSYMM16/U32/S32/F16/F32
      * @param[in] output           Destination tensor. Data type supported: Same as @p input
      * @param[in] starts           The starts of the dimensions of the input tensor to be sliced. The length must be of rank(input).
      * @param[in] ends             The ends of the dimensions of the input tensor to be sliced. The length must be of rank(input).
diff --git a/arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h b/arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h
index 4857f74f93..dc23d42126 100644
--- a/arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h
+++ b/arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h
@@ -100,7 +100,6 @@ private:
     ITensor       *_classes;
     ITensor       *_batch_splits_out;
     ITensor       *_keeps;
-    ITensor       *_keeps_size;
 
     Tensor _scores_in_f32;
     Tensor _boxes_in_f32;
@@ -110,7 +109,6 @@ private:
     Tensor _classes_f32;
     Tensor _batch_splits_out_f32;
     Tensor _keeps_f32;
-    Tensor _keeps_size_f32;
 
     bool _is_qasymm8;
 };