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>

18 files changed, 376 insertions, 110 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;
 };
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 2f748de53e..a5e75df8be 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -347,6 +347,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_float", "gemmlowp.cl" },
     { "generate_proposals_compute_all_anchors", "generate_proposals.cl" },
+    { "generate_proposals_compute_all_anchors_quantized", "generate_proposals_quantized.cl" },
     { "harris_score_3x3", "harris_corners.cl" },
     { "harris_score_5x5", "harris_corners.cl" },
     { "harris_score_7x7", "harris_corners.cl" },
@@ -793,6 +794,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/generate_proposals.clembed"
     },
     {
+        "generate_proposals_quantized.cl",
+#include "./cl_kernels/generate_proposals_quantized.clembed"
+    },
+    {
         "harris_corners.cl",
 #include "./cl_kernels/harris_corners.clembed"
     },
diff --git a/src/core/CL/cl_kernels/generate_proposals_quantized.cl b/src/core/CL/cl_kernels/generate_proposals_quantized.cl
new file mode 100644
index 0000000000..690d1cfdf8
--- /dev/null
+++ b/src/core/CL/cl_kernels/generate_proposals_quantized.cl
@@ -0,0 +1,87 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers_asymm.h"
+
+/** Generate all the region of interests based on the image size and the anchors passed in. For each element (x,y) of the
+ * grid, it will generate NUM_ANCHORS rois, given by shifting the grid position to match the anchor.
+ *
+ * @attention The following variables must be passed at compile time:
+ * -# -DDATA_TYPE= Tensor data type. Supported data types: QASYMM8
+ * -# -DHEIGHT= Height of the feature map on which this kernel is applied
+ * -# -DWIDTH= Width of the feature map on which this kernel is applied
+ * -# -DNUM_ANCHORS= Number of anchors to be used to generate the rois per each pixel
+ * -# -DSTRIDE= Stride to be applied at each different pixel position (i.e., x_range = (1:WIDTH)*STRIDE and y_range = (1:HEIGHT)*STRIDE
+ * -# -DNUM_ROI_FIELDS= Number of fields used to represent a roi
+ *
+ * @param[in]  anchors_ptr                           Pointer to the anchors tensor. Supported data types: QASYMM8
+ * @param[in]  anchors_stride_x                      Stride of the anchors tensor in X dimension (in bytes)
+ * @param[in]  anchors_step_x                        anchors_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  anchors_stride_y                      Stride of the anchors tensor in Y dimension (in bytes)
+ * @param[in]  anchors_step_y                        anchors_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  anchors_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  anchors_step_z                        anchors_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  anchors_offset_first_element_in_bytes The offset of the first element in the boxes tensor
+ * @param[out] rois_ptr                              Pointer to the rois. Supported data types: same as @p in_ptr
+ * @param[out] rois_stride_x                         Stride of the rois in X dimension (in bytes)
+ * @param[out] rois_step_x                           pred_boxes_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[out] rois_stride_y                         Stride of the rois in Y dimension (in bytes)
+ * @param[out] rois_step_y                           pred_boxes_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[out] rois_stride_z                         Stride of the rois in Z dimension (in bytes)
+ * @param[out] rois_step_z                           pred_boxes_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[out] rois_offset_first_element_in_bytes    The offset of the first element in the rois
+ */
+#if defined(DATA_TYPE) && defined(WIDTH) && defined(HEIGHT) && defined(NUM_ANCHORS) && defined(STRIDE) && defined(NUM_ROI_FIELDS) && defined(OFFSET) && defined(SCALE)
+__kernel void generate_proposals_compute_all_anchors_quantized(
+    VECTOR_DECLARATION(anchors),
+    VECTOR_DECLARATION(rois))
+{
+    Vector anchors = CONVERT_TO_VECTOR_STRUCT_NO_STEP(anchors);
+    Vector rois    = CONVERT_TO_VECTOR_STRUCT(rois);
+
+    const size_t idx = get_global_id(0);
+    // Find the index of the anchor
+    const size_t anchor_idx = idx % NUM_ANCHORS;
+
+    // Find which shift is this thread using
+    const size_t shift_idx = idx / NUM_ANCHORS;
+
+    // Compute the shift on the X and Y direction (the shift depends exclusively by the index thread id)
+    const float shift_x = (float)(shift_idx % WIDTH) * STRIDE;
+    const float shift_y = (float)(shift_idx / WIDTH) * STRIDE;
+
+    VEC_DATA_TYPE(float, NUM_ROI_FIELDS)
+    shift = (VEC_DATA_TYPE(float, NUM_ROI_FIELDS))(shift_x, shift_y, shift_x, shift_y);
+
+    // Read the given anchor
+    VEC_DATA_TYPE(float, NUM_ROI_FIELDS)
+    anchor = DEQUANTIZE(VLOAD(NUM_ROI_FIELDS)(0, (__global DATA_TYPE *)vector_offset(&anchors, anchor_idx * NUM_ROI_FIELDS)), OFFSET, SCALE, DATA_TYPE, NUM_ROI_FIELDS);
+
+    // Apply the shift to the anchor
+    VEC_DATA_TYPE(float, NUM_ROI_FIELDS)
+    shifted_anchor = anchor + shift;
+
+    VSTORE(NUM_ROI_FIELDS)
+    (QUANTIZE(shifted_anchor, OFFSET, SCALE, DATA_TYPE, NUM_ROI_FIELDS), 0, (__global DATA_TYPE *)rois.ptr);
+}
+#endif //defined(DATA_TYPE) && defined(WIDTH) && defined(HEIGHT) && defined(NUM_ANCHORS) && defined(STRIDE) && defined(NUM_ROI_FIELDS) && defined(OFFSET) && defined(SCALE)
diff --git a/src/core/CL/cl_kernels/helpers_asymm.h b/src/core/CL/cl_kernels/helpers_asymm.h
index ad06451f13..53e6719cd7 100644
--- a/src/core/CL/cl_kernels/helpers_asymm.h
+++ b/src/core/CL/cl_kernels/helpers_asymm.h
@@ -375,9 +375,11 @@ inline float dequantize_qasymm8(uchar input, float offset, float scale)
 
 QUANTIZE_IMPL(uchar, 4)
 QUANTIZE_IMPL(ushort, 4)
+QUANTIZE_IMPL(short, 4)
 
 DEQUANTIZE_IMPL(uchar, 4)
 DEQUANTIZE_IMPL(ushort, 4)
+DEQUANTIZE_IMPL(short, 4)
 
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4)
diff --git a/src/core/CL/cl_kernels/slice_ops.cl b/src/core/CL/cl_kernels/slice_ops.cl
index 97decee6fc..2163c699dd 100644
--- a/src/core/CL/cl_kernels/slice_ops.cl
+++ b/src/core/CL/cl_kernels/slice_ops.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -32,7 +32,7 @@
  * @attention Absolute start coordinates for each dimension should be given as preprocessor -DSTART_index=value e.g. -DSTART_0=2
  * @attention Strides for each dimension should be given as preprocessor -DSTRIDE_index=value e.g. -DSTRIDE_1=1
  *
- * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/QASYMM16/QSYMM16/F16/U32/S32/F32
  * @param[in]  input_stride_x                       Stride of the source tensor in X dimension (in bytes)
  * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
diff --git a/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp b/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp
index 79e364caf7..16d0e86d7d 100644
--- a/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp
+++ b/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp
@@ -44,7 +44,7 @@ Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anc
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(anchors, all_anchors);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(anchors);
     ARM_COMPUTE_RETURN_ERROR_ON(anchors->dimension(0) != info.values_per_roi());
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::QSYMM16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON(anchors->num_dimensions() > 2);
     if(all_anchors->total_size() > 0)
     {
@@ -55,6 +55,11 @@ Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anc
         ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->num_dimensions() > 2);
         ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(0) != info.values_per_roi());
         ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(1) != feature_height * feature_width * num_anchors);
+
+        if(is_data_type_quantized(anchors->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(anchors, all_anchors);
+        }
     }
     return Status{};
 }
@@ -78,12 +83,14 @@ void CLComputeAllAnchorsKernel::configure(const ICLTensor *anchors, ICLTensor *a
 
     // Initialize the output if empty
     const TensorShape output_shape(info.values_per_roi(), width * height * num_anchors);
-    auto_init_if_empty(*all_anchors->info(), output_shape, 1, data_type);
+    auto_init_if_empty(*all_anchors->info(), TensorInfo(output_shape, 1, data_type, anchors->info()->quantization_info()));
 
     // Set instance variables
     _anchors     = anchors;
     _all_anchors = all_anchors;
 
+    const bool is_quantized = is_data_type_quantized(anchors->info()->data_type());
+
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
@@ -93,8 +100,16 @@ void CLComputeAllAnchorsKernel::configure(const ICLTensor *anchors, ICLTensor *a
     build_opts.add_option("-DNUM_ANCHORS=" + support::cpp11::to_string(num_anchors));
     build_opts.add_option("-DNUM_ROI_FIELDS=" + support::cpp11::to_string(info.values_per_roi()));
 
+    if(is_quantized)
+    {
+        const UniformQuantizationInfo qinfo = anchors->info()->quantization_info().uniform();
+        build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(qinfo.scale));
+        build_opts.add_option("-DOFFSET=" + float_to_string_with_full_precision(qinfo.offset));
+    }
+
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("generate_proposals_compute_all_anchors", build_opts.options()));
+    const std::string kernel_name = (is_quantized) ? "generate_proposals_compute_all_anchors_quantized" : "generate_proposals_compute_all_anchors";
+    _kernel                       = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // The tensor all_anchors can be interpreted as an array of structs (each structs has values_per_roi fields).
     // This means we don't need to pad on the X dimension, as we know in advance how many fields
diff --git a/src/core/CL/kernels/CLStridedSliceKernel.cpp b/src/core/CL/kernels/CLStridedSliceKernel.cpp
index 9dd488b678..248a55717d 100644
--- a/src/core/CL/kernels/CLStridedSliceKernel.cpp
+++ b/src/core/CL/kernels/CLStridedSliceKernel.cpp
@@ -48,7 +48,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
                                                          DataType::U8, DataType::S8, DataType::QASYMM8,
-                                                         DataType::U16, DataType::S16, DataType::QSYMM16,
+                                                         DataType::U16, DataType::S16, DataType::QASYMM16, DataType::QSYMM16,
                                                          DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
 
diff --git a/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp b/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
index 62568b4b45..3058a0c977 100644
--- a/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
+++ b/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
@@ -360,6 +360,7 @@ void CPPBoxWithNonMaximaSuppressionLimitKernel::configure(const ITensor *scores_
 
     ARM_COMPUTE_ERROR_ON(scores_out->info()->dimension(0) != boxes_out->info()->dimension(1));
     ARM_COMPUTE_ERROR_ON(boxes_out->info()->dimension(0) != 4);
+    ARM_COMPUTE_ERROR_ON(scores_out->info()->dimension(0) != classes->info()->dimension(0));
     if(keeps != nullptr)
     {
         ARM_COMPUTE_ERROR_ON_MSG(keeps_size == nullptr, "keeps_size cannot be nullptr if keeps has to be provided as output");
diff --git a/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp b/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp
index 94aa5e7198..c9eb8abc29 100644
--- a/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp
+++ b/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp
@@ -30,7 +30,7 @@
 namespace arm_compute
 {
 CLGenerateProposalsLayer::CLGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)),
+    : _memory_group(memory_manager),
       _permute_deltas_kernel(),
       _flatten_deltas_kernel(),
       _permute_scores_kernel(),
@@ -38,17 +38,25 @@ CLGenerateProposalsLayer::CLGenerateProposalsLayer(std::shared_ptr<IMemoryManage
       _compute_anchors_kernel(),
       _bounding_box_kernel(),
       _pad_kernel(),
-      _cpp_nms_kernel(),
+      _dequantize_anchors(),
+      _dequantize_deltas(),
+      _quantize_all_proposals(),
+      _cpp_nms(memory_manager),
       _is_nhwc(false),
+      _is_qasymm8(false),
       _deltas_permuted(),
       _deltas_flattened(),
+      _deltas_flattened_f32(),
       _scores_permuted(),
       _scores_flattened(),
       _all_anchors(),
+      _all_anchors_f32(),
       _all_proposals(),
+      _all_proposals_quantized(),
       _keeps_nms_unused(),
       _classes_nms_unused(),
       _proposals_4_roi_values(),
+      _all_proposals_to_use(nullptr),
       _num_valid_proposals(nullptr),
       _scores_out(nullptr)
 {
@@ -60,63 +68,93 @@ void CLGenerateProposalsLayer::configure(const ICLTensor *scores, const ICLTenso
     ARM_COMPUTE_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
     ARM_COMPUTE_ERROR_THROW_ON(CLGenerateProposalsLayer::validate(scores->info(), deltas->info(), anchors->info(), proposals->info(), scores_out->info(), num_valid_proposals->info(), info));
 
-    _is_nhwc                         = scores->info()->data_layout() == DataLayout::NHWC;
-    const DataType data_type         = deltas->info()->data_type();
-    const int      num_anchors       = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::CHANNEL));
-    const int      feat_width        = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::WIDTH));
-    const int      feat_height       = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::HEIGHT));
-    const int      total_num_anchors = num_anchors * feat_width * feat_height;
-    const int      pre_nms_topN      = info.pre_nms_topN();
-    const int      post_nms_topN     = info.post_nms_topN();
-    const size_t   values_per_roi    = info.values_per_roi();
+    _is_nhwc                        = scores->info()->data_layout() == DataLayout::NHWC;
+    const DataType scores_data_type = scores->info()->data_type();
+    _is_qasymm8                     = scores_data_type == DataType::QASYMM8;
+    const int    num_anchors        = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::CHANNEL));
+    const int    feat_width         = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::WIDTH));
+    const int    feat_height        = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::HEIGHT));
+    const int    total_num_anchors  = num_anchors * feat_width * feat_height;
+    const int    pre_nms_topN       = info.pre_nms_topN();
+    const int    post_nms_topN      = info.post_nms_topN();
+    const size_t values_per_roi     = info.values_per_roi();
+
+    const QuantizationInfo scores_qinfo   = scores->info()->quantization_info();
+    const DataType         rois_data_type = (_is_qasymm8) ? DataType::QASYMM16 : scores_data_type;
+    const QuantizationInfo rois_qinfo     = (_is_qasymm8) ? QuantizationInfo(0.125f, 0) : scores->info()->quantization_info();
 
     // Compute all the anchors
     _memory_group.manage(&_all_anchors);
     _compute_anchors_kernel.configure(anchors, &_all_anchors, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale()));
 
     const TensorShape flatten_shape_deltas(values_per_roi, total_num_anchors);
-    _deltas_flattened.allocator()->init(TensorInfo(flatten_shape_deltas, 1, data_type));
+    _deltas_flattened.allocator()->init(TensorInfo(flatten_shape_deltas, 1, scores_data_type, deltas->info()->quantization_info()));
 
     // Permute and reshape deltas
+    _memory_group.manage(&_deltas_flattened);
     if(!_is_nhwc)
     {
         _memory_group.manage(&_deltas_permuted);
-        _memory_group.manage(&_deltas_flattened);
         _permute_deltas_kernel.configure(deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 });
         _flatten_deltas_kernel.configure(&_deltas_permuted, &_deltas_flattened);
         _deltas_permuted.allocator()->allocate();
     }
     else
     {
-        _memory_group.manage(&_deltas_flattened);
         _flatten_deltas_kernel.configure(deltas, &_deltas_flattened);
     }
 
     const TensorShape flatten_shape_scores(1, total_num_anchors);
-    _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, data_type));
+    _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, scores_data_type, scores_qinfo));
 
     // Permute and reshape scores
+    _memory_group.manage(&_scores_flattened);
     if(!_is_nhwc)
     {
         _memory_group.manage(&_scores_permuted);
-        _memory_group.manage(&_scores_flattened);
         _permute_scores_kernel.configure(scores, &_scores_permuted, PermutationVector{ 2, 0, 1 });
         _flatten_scores_kernel.configure(&_scores_permuted, &_scores_flattened);
         _scores_permuted.allocator()->allocate();
     }
     else
     {
-        _memory_group.manage(&_scores_flattened);
         _flatten_scores_kernel.configure(scores, &_scores_flattened);
     }
 
+    CLTensor *anchors_to_use = &_all_anchors;
+    CLTensor *deltas_to_use  = &_deltas_flattened;
+    if(_is_qasymm8)
+    {
+        _all_anchors_f32.allocator()->init(TensorInfo(_all_anchors.info()->tensor_shape(), 1, DataType::F32));
+        _deltas_flattened_f32.allocator()->init(TensorInfo(_deltas_flattened.info()->tensor_shape(), 1, DataType::F32));
+        _memory_group.manage(&_all_anchors_f32);
+        _memory_group.manage(&_deltas_flattened_f32);
+        // Dequantize anchors to float
+        _dequantize_anchors.configure(&_all_anchors, &_all_anchors_f32);
+        _all_anchors.allocator()->allocate();
+        anchors_to_use = &_all_anchors_f32;
+        // Dequantize deltas to float
+        _dequantize_deltas.configure(&_deltas_flattened, &_deltas_flattened_f32);
+        _deltas_flattened.allocator()->allocate();
+        deltas_to_use = &_deltas_flattened_f32;
+    }
     // Bounding box transform
     _memory_group.manage(&_all_proposals);
     BoundingBoxTransformInfo bbox_info(info.im_width(), info.im_height(), 1.f);
-    _bounding_box_kernel.configure(&_all_anchors, &_all_proposals, &_deltas_flattened, bbox_info);
-    _deltas_flattened.allocator()->allocate();
-    _all_anchors.allocator()->allocate();
+    _bounding_box_kernel.configure(anchors_to_use, &_all_proposals, deltas_to_use, bbox_info);
+    deltas_to_use->allocator()->allocate();
+    anchors_to_use->allocator()->allocate();
 
+    _all_proposals_to_use = &_all_proposals;
+    if(_is_qasymm8)
+    {
+        _memory_group.manage(&_all_proposals_quantized);
+        // Requantize all_proposals to QASYMM16 with 0.125 scale and 0 offset
+        _all_proposals_quantized.allocator()->init(TensorInfo(_all_proposals.info()->tensor_shape(), 1, DataType::QASYMM16, QuantizationInfo(0.125f, 0)));
+        _quantize_all_proposals.configure(&_all_proposals, &_all_proposals_quantized);
+        _all_proposals.allocator()->allocate();
+        _all_proposals_to_use = &_all_proposals_quantized;
+    }
     // The original layer implementation first selects the best pre_nms_topN anchors (thus having a lightweight sort)
     // that are then transformed by bbox_transform. The boxes generated are then fed into a non-sorting NMS operation.
     // Since we are reusing the NMS layer and we don't implement any CL/sort, we let NMS do the sorting (of all the input)
@@ -127,12 +165,12 @@ void CLGenerateProposalsLayer::configure(const ICLTensor *scores, const ICLTenso
     _memory_group.manage(&_keeps_nms_unused);
 
     // Note that NMS needs outputs preinitialized.
-    auto_init_if_empty(*scores_out->info(), TensorShape(scores_nms_size), 1, data_type);
-    auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, data_type);
+    auto_init_if_empty(*scores_out->info(), TensorShape(scores_nms_size), 1, scores_data_type, scores_qinfo);
+    auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, rois_data_type, rois_qinfo);
     auto_init_if_empty(*num_valid_proposals->info(), TensorShape(1), 1, DataType::U32);
 
     // Initialize temporaries (unused) outputs
-    _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(1, 1), 1, data_type));
+    _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(scores_nms_size), 1, scores_data_type, scores_qinfo));
     _keeps_nms_unused.allocator()->init(*scores_out->info());
 
     // Save the output (to map and unmap them at run)
@@ -140,11 +178,11 @@ void CLGenerateProposalsLayer::configure(const ICLTensor *scores, const ICLTenso
     _num_valid_proposals = num_valid_proposals;
 
     _memory_group.manage(&_proposals_4_roi_values);
-    _cpp_nms_kernel.configure(&_scores_flattened, &_all_proposals, nullptr, scores_out, &_proposals_4_roi_values, &_classes_nms_unused, nullptr, &_keeps_nms_unused, num_valid_proposals,
-                              BoxNMSLimitInfo(0.0f, info.nms_thres(), scores_nms_size, false, NMSType::LINEAR, 0.5f, 0.001f, true, min_size_scaled, info.im_width(), info.im_height()));
+    _cpp_nms.configure(&_scores_flattened, _all_proposals_to_use, nullptr, scores_out, &_proposals_4_roi_values, &_classes_nms_unused, nullptr, &_keeps_nms_unused, num_valid_proposals,
+                       BoxNMSLimitInfo(0.0f, info.nms_thres(), scores_nms_size, false, NMSType::LINEAR, 0.5f, 0.001f, true, min_size_scaled, info.im_width(), info.im_height()));
     _keeps_nms_unused.allocator()->allocate();
     _classes_nms_unused.allocator()->allocate();
-    _all_proposals.allocator()->allocate();
+    _all_proposals_to_use->allocator()->allocate();
     _scores_flattened.allocator()->allocate();
 
     // Add the first column that represents the batch id. This will be all zeros, as we don't support multiple images
@@ -156,8 +194,10 @@ Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITens
                                           const ITensorInfo *num_valid_proposals, const GenerateProposalsInfo &info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(scores, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(scores, DataLayout::NCHW, DataLayout::NHWC);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(scores, deltas);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores, deltas);
 
     const int num_anchors       = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::CHANNEL));
     const int feat_width        = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::WIDTH));
@@ -166,8 +206,17 @@ Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITens
     const int total_num_anchors = num_anchors * feat_width * feat_height;
     const int values_per_roi    = info.values_per_roi();
 
+    const bool is_qasymm8 = scores->data_type() == DataType::QASYMM8;
+
     ARM_COMPUTE_RETURN_ERROR_ON(num_images > 1);
 
+    if(is_qasymm8)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(anchors, 1, DataType::QSYMM16);
+        const UniformQuantizationInfo anchors_qinfo = anchors->quantization_info().uniform();
+        ARM_COMPUTE_RETURN_ERROR_ON(anchors_qinfo.scale != 0.125f);
+    }
+
     TensorInfo all_anchors_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
     ARM_COMPUTE_RETURN_ON_ERROR(CLComputeAllAnchorsKernel::validate(anchors, &all_anchors_info, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale())));
 
@@ -187,14 +236,36 @@ Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITens
     TensorInfo deltas_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
     ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&deltas_permuted_info, &deltas_flattened_info));
 
-    TensorInfo scores_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true));
+    TensorInfo scores_flattened_info(scores->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true));
     TensorInfo proposals_4_roi_values(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
 
     ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&scores_permuted_info, &scores_flattened_info));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_info, &proposals_4_roi_values, &deltas_flattened_info, BoundingBoxTransformInfo(info.im_width(), info.im_height(),
-                                                                       1.f)));
 
-    ARM_COMPUTE_RETURN_ON_ERROR(CLPadLayerKernel::validate(&proposals_4_roi_values, proposals, PaddingList{ { 1, 0 } }));
+    TensorInfo *proposals_4_roi_values_to_use = &proposals_4_roi_values;
+    TensorInfo  proposals_4_roi_values_quantized(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
+    proposals_4_roi_values_quantized.set_data_type(DataType::QASYMM16).set_quantization_info(QuantizationInfo(0.125f, 0));
+    if(is_qasymm8)
+    {
+        TensorInfo all_anchors_f32_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayerKernel::validate(&all_anchors_info, &all_anchors_f32_info));
+
+        TensorInfo deltas_flattened_f32_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayerKernel::validate(&deltas_flattened_info, &deltas_flattened_f32_info));
+
+        TensorInfo proposals_4_roi_values_f32(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_f32_info, &proposals_4_roi_values_f32, &deltas_flattened_f32_info,
+                                                                           BoundingBoxTransformInfo(info.im_width(), info.im_height(), 1.f)));
+
+        ARM_COMPUTE_RETURN_ON_ERROR(CLQuantizationLayerKernel::validate(&proposals_4_roi_values_f32, &proposals_4_roi_values_quantized));
+        proposals_4_roi_values_to_use = &proposals_4_roi_values_quantized;
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_info, &proposals_4_roi_values, &deltas_flattened_info,
+                                                                           BoundingBoxTransformInfo(info.im_width(), info.im_height(), 1.f)));
+    }
+
+    ARM_COMPUTE_RETURN_ON_ERROR(CLPadLayerKernel::validate(proposals_4_roi_values_to_use, proposals, PaddingList{ { 1, 0 } }));
 
     if(num_valid_proposals->total_size() > 0)
     {
@@ -208,7 +279,17 @@ Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITens
         ARM_COMPUTE_RETURN_ERROR_ON(proposals->num_dimensions() > 2);
         ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(0) != size_t(values_per_roi) + 1);
         ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(1) != size_t(total_num_anchors));
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, deltas);
+        if(is_qasymm8)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(proposals, 1, DataType::QASYMM16);
+            const UniformQuantizationInfo proposals_qinfo = proposals->quantization_info().uniform();
+            ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.scale != 0.125f);
+            ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.offset != 0);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, scores);
+        }
     }
 
     if(scores_out->total_size() > 0)
@@ -225,7 +306,7 @@ void CLGenerateProposalsLayer::run_cpp_nms_kernel()
 {
     // Map inputs
     _scores_flattened.map(true);
-    _all_proposals.map(true);
+    _all_proposals_to_use->map(true);
 
     // Map outputs
     _scores_out->map(CLScheduler::get().queue(), true);
@@ -235,7 +316,7 @@ void CLGenerateProposalsLayer::run_cpp_nms_kernel()
     _classes_nms_unused.map(true);
 
     // Run nms
-    CPPScheduler::get().schedule(&_cpp_nms_kernel, Window::DimX);
+    _cpp_nms.run();
 
     // Unmap outputs
     _keeps_nms_unused.unmap();
@@ -246,7 +327,7 @@ void CLGenerateProposalsLayer::run_cpp_nms_kernel()
 
     // Unmap inputs
     _scores_flattened.unmap();
-    _all_proposals.unmap();
+    _all_proposals_to_use->unmap();
 }
 
 void CLGenerateProposalsLayer::run()
@@ -266,8 +347,20 @@ void CLGenerateProposalsLayer::run()
     CLScheduler::get().enqueue(_flatten_deltas_kernel, false);
     CLScheduler::get().enqueue(_flatten_scores_kernel, false);
 
+    if(_is_qasymm8)
+    {
+        CLScheduler::get().enqueue(_dequantize_anchors, false);
+        CLScheduler::get().enqueue(_dequantize_deltas, false);
+    }
+
     // Build the boxes
     CLScheduler::get().enqueue(_bounding_box_kernel, false);
+
+    if(_is_qasymm8)
+    {
+        CLScheduler::get().enqueue(_quantize_all_proposals, false);
+    }
+
     // Non maxima suppression
     run_cpp_nms_kernel();
     // Add dummy batch indexes
diff --git a/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp b/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp
index 158f45a320..782771bc50 100644
--- a/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp
+++ b/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp
@@ -30,9 +30,10 @@ namespace arm_compute
 {
 namespace
 {
-void dequantize_tensor(const ITensor *input, ITensor *output, DataType data_type)
+void dequantize_tensor(const ITensor *input, ITensor *output)
 {
-    const UniformQuantizationInfo qinfo = input->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qinfo     = input->info()->quantization_info().uniform();
+    const DataType                data_type = input->info()->data_type();
 
     Window window;
     window.use_tensor_dimensions(input->info()->tensor_shape());
@@ -60,9 +61,10 @@ void dequantize_tensor(const ITensor *input, ITensor *output, DataType data_type
     }
 }
 
-void quantize_tensor(const ITensor *input, ITensor *output, DataType data_type)
+void quantize_tensor(const ITensor *input, ITensor *output)
 {
-    const UniformQuantizationInfo qinfo = input->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qinfo     = output->info()->quantization_info().uniform();
+    const DataType                data_type = output->info()->data_type();
 
     Window window;
     window.use_tensor_dimensions(input->info()->tensor_shape());
@@ -102,7 +104,6 @@ CPPBoxWithNonMaximaSuppressionLimit::CPPBoxWithNonMaximaSuppressionLimit(std::sh
       _classes(),
       _batch_splits_out(),
       _keeps(),
-      _keeps_size(),
       _scores_in_f32(),
       _boxes_in_f32(),
       _batch_splits_in_f32(),
@@ -111,7 +112,6 @@ CPPBoxWithNonMaximaSuppressionLimit::CPPBoxWithNonMaximaSuppressionLimit(std::sh
       _classes_f32(),
       _batch_splits_out_f32(),
       _keeps_f32(),
-      _keeps_size_f32(),
       _is_qasymm8(false)
 {
 }
@@ -119,7 +119,7 @@ CPPBoxWithNonMaximaSuppressionLimit::CPPBoxWithNonMaximaSuppressionLimit(std::sh
 void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, const ITensor *boxes_in, const ITensor *batch_splits_in, ITensor *scores_out, ITensor *boxes_out, ITensor *classes,
                                                     ITensor *batch_splits_out, ITensor *keeps, ITensor *keeps_size, const BoxNMSLimitInfo info)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(scores_in, boxes_in, scores_out, boxes_out, classes);
 
     _is_qasymm8 = scores_in->info()->data_type() == DataType::QASYMM8;
 
@@ -131,20 +131,22 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co
     _classes          = classes;
     _batch_splits_out = batch_splits_out;
     _keeps            = keeps;
-    _keeps_size       = keeps_size;
 
     if(_is_qasymm8)
     {
         // Manage intermediate buffers
         _memory_group.manage(&_scores_in_f32);
         _memory_group.manage(&_boxes_in_f32);
-        _memory_group.manage(&_batch_splits_in_f32);
         _memory_group.manage(&_scores_out_f32);
         _memory_group.manage(&_boxes_out_f32);
         _memory_group.manage(&_classes_f32);
         _scores_in_f32.allocator()->init(scores_in->info()->clone()->set_data_type(DataType::F32));
         _boxes_in_f32.allocator()->init(boxes_in->info()->clone()->set_data_type(DataType::F32));
-        _batch_splits_in_f32.allocator()->init(batch_splits_in->info()->clone()->set_data_type(DataType::F32));
+        if(batch_splits_in != nullptr)
+        {
+            _memory_group.manage(&_batch_splits_in_f32);
+            _batch_splits_in_f32.allocator()->init(batch_splits_in->info()->clone()->set_data_type(DataType::F32));
+        }
         _scores_out_f32.allocator()->init(scores_out->info()->clone()->set_data_type(DataType::F32));
         _boxes_out_f32.allocator()->init(boxes_out->info()->clone()->set_data_type(DataType::F32));
         _classes_f32.allocator()->init(classes->info()->clone()->set_data_type(DataType::F32));
@@ -158,15 +160,11 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co
             _memory_group.manage(&_keeps_f32);
             _keeps_f32.allocator()->init(keeps->info()->clone()->set_data_type(DataType::F32));
         }
-        if(keeps_size != nullptr)
-        {
-            _memory_group.manage(&_keeps_size_f32);
-            _keeps_size_f32.allocator()->init(keeps_size->info()->clone()->set_data_type(DataType::F32));
-        }
 
-        _box_with_nms_limit_kernel.configure(&_scores_in_f32, &_boxes_in_f32, &_batch_splits_in_f32, &_scores_out_f32, &_boxes_out_f32, &_classes_f32,
+        _box_with_nms_limit_kernel.configure(&_scores_in_f32, &_boxes_in_f32, (batch_splits_in != nullptr) ? &_batch_splits_in_f32 : nullptr,
+                                             &_scores_out_f32, &_boxes_out_f32, &_classes_f32,
                                              (batch_splits_out != nullptr) ? &_batch_splits_out_f32 : nullptr, (keeps != nullptr) ? &_keeps_f32 : nullptr,
-                                             (keeps_size != nullptr) ? &_keeps_size_f32 : nullptr, info);
+                                             keeps_size, info);
     }
     else
     {
@@ -177,7 +175,10 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co
     {
         _scores_in_f32.allocator()->allocate();
         _boxes_in_f32.allocator()->allocate();
-        _batch_splits_in_f32.allocator()->allocate();
+        if(_batch_splits_in != nullptr)
+        {
+            _batch_splits_in_f32.allocator()->allocate();
+        }
         _scores_out_f32.allocator()->allocate();
         _boxes_out_f32.allocator()->allocate();
         _classes_f32.allocator()->allocate();
@@ -189,17 +190,13 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co
         {
             _keeps_f32.allocator()->allocate();
         }
-        if(keeps_size != nullptr)
-        {
-            _keeps_size_f32.allocator()->allocate();
-        }
     }
 }
 
 Status validate(const ITensorInfo *scores_in, const ITensorInfo *boxes_in, const ITensorInfo *batch_splits_in, const ITensorInfo *scores_out, const ITensorInfo *boxes_out, const ITensorInfo *classes,
                 const ITensorInfo *batch_splits_out, const ITensorInfo *keeps, const ITensorInfo *keeps_size, const BoxNMSLimitInfo info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores_in, boxes_in, scores_out, boxes_out, classes);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(scores_in, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
 
     const bool is_qasymm8 = scores_in->data_type() == DataType::QASYMM8;
@@ -218,31 +215,33 @@ Status validate(const ITensorInfo *scores_in, const ITensorInfo *boxes_in, const
 
 void CPPBoxWithNonMaximaSuppressionLimit::run()
 {
+    // Acquire all the temporaries
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
     if(_is_qasymm8)
     {
-        dequantize_tensor(_scores_in, &_scores_in_f32, _scores_in->info()->data_type());
-        dequantize_tensor(_boxes_in, &_boxes_in_f32, _boxes_in->info()->data_type());
-        dequantize_tensor(_batch_splits_in, &_batch_splits_in_f32, _batch_splits_in->info()->data_type());
+        dequantize_tensor(_scores_in, &_scores_in_f32);
+        dequantize_tensor(_boxes_in, &_boxes_in_f32);
+        if(_batch_splits_in != nullptr)
+        {
+            dequantize_tensor(_batch_splits_in, &_batch_splits_in_f32);
+        }
     }
 
     Scheduler::get().schedule(&_box_with_nms_limit_kernel, Window::DimY);
 
     if(_is_qasymm8)
     {
-        quantize_tensor(&_scores_out_f32, _scores_out, _scores_out->info()->data_type());
-        quantize_tensor(&_boxes_out_f32, _boxes_out, _boxes_out->info()->data_type());
-        quantize_tensor(&_classes_f32, _classes, _classes->info()->data_type());
+        quantize_tensor(&_scores_out_f32, _scores_out);
+        quantize_tensor(&_boxes_out_f32, _boxes_out);
+        quantize_tensor(&_classes_f32, _classes);
         if(_batch_splits_out != nullptr)
         {
-            quantize_tensor(&_batch_splits_out_f32, _batch_splits_out, _batch_splits_out->info()->data_type());
+            quantize_tensor(&_batch_splits_out_f32, _batch_splits_out);
         }
         if(_keeps != nullptr)
         {
-            quantize_tensor(&_keeps_f32, _keeps, _keeps->info()->data_type());
-        }
-        if(_keeps_size != nullptr)
-        {
-            quantize_tensor(&_keeps_size_f32, _keeps_size, _keeps_size->info()->data_type());
+            quantize_tensor(&_keeps_f32, _keeps);
         }
     }
 }
diff --git a/tests/validation/CL/GenerateProposalsLayer.cpp b/tests/validation/CL/GenerateProposalsLayer.cpp
index 4ebffd7e79..bfad8e8381 100644
--- a/tests/validation/CL/GenerateProposalsLayer.cpp
+++ b/tests/validation/CL/GenerateProposalsLayer.cpp
@@ -82,6 +82,8 @@ const auto ComputeAllInfoDataset = framework::dataset::make("ComputeAllInfo",
     ComputeAnchorsInfo(100U, 100U, 1. / 4.f),
 
 });
+
+constexpr AbsoluteTolerance<int16_t> tolerance_qsymm16(1);
 } // namespace
 
 TEST_SUITE(CL)
@@ -364,7 +366,7 @@ DATA_TEST_CASE(IntegrationTestCaseGenerateProposals, framework::DatasetMode::ALL
     proposals_final.allocator()->allocate();
     select_proposals.run();
 
-    // Select the first N entries of the proposals
+    // Select the first N entries of the scores
     CLTensor scores_final;
     CLSlice  select_scores;
     select_scores.configure(&scores_out, &scores_final, Coordinates(0), Coordinates(N));
@@ -395,6 +397,22 @@ FIXTURE_DATA_TEST_CASE(ComputeAllAnchors, CLComputeAllAnchorsFixture<half>, fram
 TEST_SUITE_END() // FP16
 TEST_SUITE_END() // Float
 
+template <typename T>
+using CLComputeAllAnchorsQuantizedFixture = ComputeAllAnchorsQuantizedFixture<CLTensor, CLAccessor, CLComputeAllAnchors, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(ComputeAllAnchors, CLComputeAllAnchorsQuantizedFixture<int16_t>, framework::DatasetMode::ALL,
+                       combine(combine(combine(framework::dataset::make("NumAnchors", { 2, 4, 8 }), ComputeAllInfoDataset),
+                                       framework::dataset::make("DataType", { DataType::QSYMM16 })),
+                               framework::dataset::make("QuantInfo", { QuantizationInfo(0.125f, 0) })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_qsymm16);
+}
+TEST_SUITE_END() // QASYMM8
+TEST_SUITE_END() // Quantized
+
 TEST_SUITE_END() // GenerateProposals
 TEST_SUITE_END() // CL
 
diff --git a/tests/validation/fixtures/ComputeAllAnchorsFixture.h b/tests/validation/fixtures/ComputeAllAnchorsFixture.h
index 6f2db3e623..e837bd4838 100644
--- a/tests/validation/fixtures/ComputeAllAnchorsFixture.h
+++ b/tests/validation/fixtures/ComputeAllAnchorsFixture.h
@@ -41,14 +41,14 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ComputeAllAnchorsFixture : public framework::Fixture
+class ComputeAllAnchorsGenericFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(size_t num_anchors, const ComputeAnchorsInfo &info, DataType data_type)
+    void setup(size_t num_anchors, const ComputeAnchorsInfo &info, DataType data_type, QuantizationInfo qinfo)
     {
-        _target    = compute_target(num_anchors, data_type, info);
-        _reference = compute_reference(num_anchors, data_type, info);
+        _target    = compute_target(num_anchors, data_type, info, qinfo);
+        _reference = compute_reference(num_anchors, data_type, info, qinfo);
     }
 
 protected:
@@ -58,11 +58,11 @@ protected:
         library->fill_tensor_uniform(tensor, 0, T(0), T(100));
     }
 
-    TensorType compute_target(size_t num_anchors, DataType data_type, const ComputeAnchorsInfo &info)
+    TensorType compute_target(size_t num_anchors, DataType data_type, const ComputeAnchorsInfo &info, QuantizationInfo qinfo)
     {
         // Create tensors
         TensorShape anchors_shape(4, num_anchors);
-        TensorType  anchors = create_tensor<TensorType>(anchors_shape, data_type);
+        TensorType  anchors = create_tensor<TensorType>(anchors_shape, data_type, 1, qinfo);
         TensorType  all_anchors;
 
         // Create and configure function
@@ -88,10 +88,11 @@ protected:
 
     SimpleTensor<T> compute_reference(size_t                    num_anchors,
                                       DataType                  data_type,
-                                      const ComputeAnchorsInfo &info)
+                                      const ComputeAnchorsInfo &info,
+                                      QuantizationInfo          qinfo)
     {
         // Create reference tensor
-        SimpleTensor<T> anchors(TensorShape(4, num_anchors), data_type);
+        SimpleTensor<T> anchors(TensorShape(4, num_anchors), data_type, 1, qinfo);
 
         // Fill reference tensor
         fill(anchors);
@@ -101,6 +102,28 @@ protected:
     TensorType      _target{};
     SimpleTensor<T> _reference{};
 };
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ComputeAllAnchorsFixture : public ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(size_t num_anchors, const ComputeAnchorsInfo &info, DataType data_type)
+    {
+        ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(num_anchors, info, data_type, QuantizationInfo());
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ComputeAllAnchorsQuantizedFixture : public ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(size_t num_anchors, const ComputeAnchorsInfo &info, DataType data_type, QuantizationInfo qinfo)
+    {
+        ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(num_anchors, info, data_type, qinfo);
+    }
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/reference/ComputeAllAnchors.cpp b/tests/validation/reference/ComputeAllAnchors.cpp
index 3f0498015a..60be7ef8a8 100644
--- a/tests/validation/reference/ComputeAllAnchors.cpp
+++ b/tests/validation/reference/ComputeAllAnchors.cpp
@@ -73,6 +73,15 @@ SimpleTensor<T> compute_all_anchors(const SimpleTensor<T> &anchors, const Comput
 }
 template SimpleTensor<float> compute_all_anchors(const SimpleTensor<float> &anchors, const ComputeAnchorsInfo &info);
 template SimpleTensor<half> compute_all_anchors(const SimpleTensor<half> &anchors, const ComputeAnchorsInfo &info);
+
+template <>
+SimpleTensor<int16_t> compute_all_anchors(const SimpleTensor<int16_t> &anchors, const ComputeAnchorsInfo &info)
+{
+    SimpleTensor<float>   anchors_tmp     = convert_from_symmetric(anchors);
+    SimpleTensor<float>   all_anchors_tmp = compute_all_anchors(anchors_tmp, info);
+    SimpleTensor<int16_t> all_anchors     = convert_to_symmetric<int16_t>(all_anchors_tmp, anchors.quantization_info());
+    return all_anchors;
+}
 } // namespace reference
 } // namespace validation
 } // namespace test