COMPMID-2257: Add support for QASYMM8 in NEGenerateProposals

Change-Id: I7d9aa21ecac97847fce209f97dff0dea6e62790a
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2020
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 36 insertions, 20 deletions

diff --git a/arm_compute/core/NEON/kernels/NEGenerateProposalsLayerKernel.h b/arm_compute/core/NEON/kernels/NEGenerateProposalsLayerKernel.h
index a7b2603648..9ee9d5dd08 100644
--- a/arm_compute/core/NEON/kernels/NEGenerateProposalsLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGenerateProposalsLayerKernel.h
@@ -53,7 +53,7 @@ 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
      *
@@ -62,7 +62,7 @@ public:
 
     /** Static function to check if given info will lead to a valid configuration of @ref NEComputeAllAnchorsKernel
      *
-     * @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
      *
@@ -74,6 +74,9 @@ public:
     void run(const Window &window, const ThreadInfo &info) override;
 
 private:
+    template <typename T>
+    void internal_run(const Window &window, const ThreadInfo &info);
+
     const ITensor     *_anchors;
     ITensor           *_all_anchors;
     ComputeAnchorsInfo _anchors_info;
diff --git a/arm_compute/core/NEON/kernels/NEStridedSliceKernel.h b/arm_compute/core/NEON/kernels/NEStridedSliceKernel.h
index a272a8118b..12075207b1 100644
--- a/arm_compute/core/NEON/kernels/NEStridedSliceKernel.h
+++ b/arm_compute/core/NEON/kernels/NEStridedSliceKernel.h
@@ -58,7 +58,7 @@ public:
      *
      * @note Supported tensor rank: up to 4
      *
-     * @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).
@@ -76,7 +76,7 @@ public:
      *
      * @note Supported tensor rank: up to 4
      *
-     * @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/CLGenerateProposalsLayer.h b/arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h
index 827f19d130..e14e195ec6 100644
--- a/arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h
@@ -50,7 +50,7 @@ class ICLTensor;
  * -# @ref CLReshapeLayer x 2
  * -# @ref CLBoundingBoxTransform
  * -# @ref CLPadLayerKernel
- * -# @ref CLDequantizationLayerKernel
+ * -# @ref CLDequantizationLayerKernel x 2
  * -# @ref CLQuantizationLayerKernel
  * And the following CPP functions:
  * -# @ref CPPBoxWithNonMaximaSuppressionLimit
diff --git a/arm_compute/runtime/NEON/functions/NEGenerateProposalsLayer.h b/arm_compute/runtime/NEON/functions/NEGenerateProposalsLayer.h
index c6d3628e37..cd370a03dd 100644
--- a/arm_compute/runtime/NEON/functions/NEGenerateProposalsLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGenerateProposalsLayer.h
@@ -23,15 +23,16 @@
  */
 #ifndef __ARM_COMPUTE_NEGENERATEPROPOSALSLAYER_H__
 #define __ARM_COMPUTE_NEGENERATEPROPOSALSLAYER_H__
-#include "arm_compute/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.h"
 #include "arm_compute/core/NEON/kernels/NEBoundingBoxTransformKernel.h"
+#include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGenerateProposalsLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEPadLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEPermuteKernel.h"
+#include "arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h"
-#include "arm_compute/core/NEON/kernels/NEStridedSliceKernel.h"
 #include "arm_compute/core/Types.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"
 #include "arm_compute/runtime/Tensor.h"
@@ -46,9 +47,10 @@ class ITensor;
  * -# @ref NEComputeAllAnchors
  * -# @ref NEPermute x 2
  * -# @ref NEReshapeLayer x 2
- * -# @ref NEStridedSlice x 3
  * -# @ref NEBoundingBoxTransform
  * -# @ref NEPadLayerKernel
+ * -# @ref NEDequantizationLayerKernel x 2
+ * -# @ref NEQuantizationLayerKernel
  * And the following CPP kernels:
  * -# @ref CPPBoxWithNonMaximaSuppressionLimit
  */
@@ -71,10 +73,12 @@ 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[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 input
      * @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
@@ -87,12 +91,14 @@ public:
 
     /** Static function to check if given info will lead to a valid configuration of @ref NEGenerateProposalsLayer
      *
-     * @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] anchors             Anchors tensor info 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 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] 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
@@ -116,29 +122,36 @@ private:
     NEComputeAllAnchorsKernel    _compute_anchors_kernel;
     NEBoundingBoxTransformKernel _bounding_box_kernel;
     NEPadLayerKernel             _pad_kernel;
+    NEDequantizationLayerKernel  _dequantize_anchors;
+    NEDequantizationLayerKernel  _dequantize_deltas;
+    NEQuantizationLayerKernel    _quantize_all_proposals;
 
-    // CPP kernels
-    CPPBoxWithNonMaximaSuppressionLimitKernel _cpp_nms_kernel;
+    // CPP functions
+    CPPBoxWithNonMaximaSuppressionLimit _cpp_nms;
 
     bool _is_nhwc;
+    bool _is_qasymm8;
 
     // Temporary tensors
     Tensor _deltas_permuted;
     Tensor _deltas_flattened;
+    Tensor _deltas_flattened_f32;
     Tensor _scores_permuted;
     Tensor _scores_flattened;
     Tensor _all_anchors;
+    Tensor _all_anchors_f32;
     Tensor _all_proposals;
+    Tensor _all_proposals_quantized;
     Tensor _keeps_nms_unused;
     Tensor _classes_nms_unused;
     Tensor _proposals_4_roi_values;
 
+    // Temporary tensor pointers
+    Tensor *_all_proposals_to_use;
+
     // Output tensor pointers
     ITensor *_num_valid_proposals;
     ITensor *_scores_out;
-
-    /** Internal function to run the CPP BoxWithNMS kernel */
-    void run_cpp_nms_kernel();
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_NEGENERATEPROPOSALSLAYER_H__ */