MLCE-166: Add support for extracting indices in NEPoolingLayer 2x2 NCHW

 * Added initial support for pooling indices
 * Only supported for NCHW Poolsize 2

Change-Id: I92ce767e64fcc01aae89411064b4cb2be272a1e9
Signed-off-by: morgolock <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2927
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 35 insertions, 14 deletions

diff --git a/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h b/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h
index 4b3ee24333..fdd10f3f66 100644
--- a/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h
+++ b/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h
@@ -55,17 +55,19 @@ public:
      * @param[in]  input     Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[out] output    Destination tensor. Data types supported: Same as @p input.
      * @param[in]  pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[out] indices   (optional) The indices of the maximal values. Data type supported: U32.
      */
-    void configure(const ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info);
+    void configure(const ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info, ICLTensor *indices = nullptr);
     /** Static function to check if given info will lead to a valid configuration of @ref CLPoolingLayerKernel
      *
      * @param[in] input     Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] output    Destination tensor info. Data types supported: Same as @p input.
      * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[in] indices   (optional) The indices of the maximal values. Data type supported: U32.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -74,6 +76,7 @@ public:
 public:
     const ICLTensor *_input;
     ICLTensor       *_output;
+    ICLTensor       *_indices;
     PoolingLayerInfo _pool_info;
     DataLayout       _data_layout;
     BorderSize       _border_size;
diff --git a/arm_compute/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.h b/arm_compute/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.h
index 85c051c326..7a2fb84f34 100644
--- a/arm_compute/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.h
+++ b/arm_compute/core/GLES_COMPUTE/kernels/GCPoolingLayerKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -54,18 +54,20 @@ public:
      * @param[in]  input     Source tensor. Data types supported: F16/F32.
      * @param[out] output    Destination tensor. Data types supported: Same as @p input.
      * @param[in]  pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[out] indices   (optional) The indices of the maximal values. Data type supported: U32.
      */
-    void configure(const IGCTensor *input, IGCTensor *output, const PoolingLayerInfo &pool_info);
+    void configure(const IGCTensor *input, IGCTensor *output, const PoolingLayerInfo &pool_info, IGCTensor *indices = nullptr);
 
     /** Static function to check if given info will lead to a valid configuration of @ref GCPoolingLayerKernel
      *
      * @param[in] input     Source tensor info. Data types supported: F16/F32.
      * @param[in] output    Destination tensor info. Data types supported: Same as @p input.
      * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[in] indices   (optional) The indices of the maximal values. Data type supported: U32.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr);
 
     // Inherited methods overridden:
     void run(const Window &window) override;
@@ -74,6 +76,7 @@ public:
 private:
     const IGCTensor *_input;
     IGCTensor       *_output;
+    IGCTensor       *_indices;
     PoolingLayerInfo _pool_info;
     BorderSize       _border_size;
     unsigned int     _num_elems_processed_per_iteration;
diff --git a/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h b/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h
index 654dfad701..6519ac72fe 100644
--- a/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h
@@ -57,8 +57,9 @@ public:
      * @param[in]  input     Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[out] output    Destination tensor. Data types supported: Same as @p input.
      * @param[in]  pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[out] indices   (optional) The indices of the maximal values. Data type supported: U32.
      */
-    void configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info);
+    void configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices = nullptr);
     /** Static function to check if given info will lead to a valid configuration of @ref NEPoolingLayerKernel
      *
      * @note F16 are supported for pool sizes 2 and 3 only
@@ -66,10 +67,11 @@ public:
      * @param[in] input     Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] output    Destination tensor. Data types supported: Same as @p input.
      * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[in] indices   (optional) The indices of the maximal values. Data type supported: U32.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
@@ -84,6 +86,12 @@ private:
      * @param[in] exclude_padding Flag to specify exclusion of padding from the operation.
      */
     void pooling2_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false);
+    /** Function to perform 2x2 pooling and compute the pooling indices. The indices can be used for max unpool.
+     *
+     * @param[in] window_input Input region on which to execute the kernel.
+     * @param[in] window       Output region on which to execute the kernel.
+     */
+    void pooling2_f32_nchw_maxpool_indices(const Window &window_input, const Window &window);
     /** Function to perform MxN pooling for 32-bit floating point values.
      *
      * @param[in] window_input    Input region on which to execute the kernel.
@@ -197,6 +205,7 @@ private:
     PoolingFunction  _func;
     const ITensor   *_input;
     ITensor         *_output;
+    ITensor         *_indices;
     PoolingLayerInfo _pool_info;
     DataLayout       _data_layout;
     unsigned int     _num_elems_processed_per_iteration;
diff --git a/arm_compute/runtime/CL/functions/CLPoolingLayer.h b/arm_compute/runtime/CL/functions/CLPoolingLayer.h
index c78b558ac8..05b35dcee8 100644
--- a/arm_compute/runtime/CL/functions/CLPoolingLayer.h
+++ b/arm_compute/runtime/CL/functions/CLPoolingLayer.h
@@ -46,17 +46,19 @@ public:
      * @param[in,out] input     Source tensor. (Written to only when padding != 0) Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[out]    output    Destination tensor. Data types supported: Same as @p input.
      * @param[in]     pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[out]    indices   (optional) The indices of the maximal values. Data type supported: U32.
      */
-    void configure(ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info);
+    void configure(ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info, ICLTensor *indices = nullptr);
     /** Static function to check if given info will lead to a valid configuration of @ref CLPoolingLayer
      *
      * @param[in] input     Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] output    Destination tensor info. Data types supported: Same as @p input.
      * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[in] indices   (optional) The indices of the maximal values. Data type supported: U32.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr);
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_CLPOOLINGLAYER_H */
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCPoolingLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCPoolingLayer.h
index 13b7ad363f..b29f808f99 100644
--- a/arm_compute/runtime/GLES_COMPUTE/functions/GCPoolingLayer.h
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCPoolingLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -51,17 +51,19 @@ public:
      * @param[in,out] input     Source tensor. (Written to only when padding != 0) Data types supported: F16/F32.
      * @param[out]    output    Destination tensor. Data types supported: Same as @p input.
      * @param[in]     pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[out]    indices   (optional) The indices of the maximal values. Data type supported: U32.
      */
-    void configure(IGCTensor *input, IGCTensor *output, const PoolingLayerInfo &pool_info);
+    void configure(IGCTensor *input, IGCTensor *output, const PoolingLayerInfo &pool_info, IGCTensor *indices = nullptr);
     /** Static function to check if given info will lead to a valid configuration of @ref GCPoolingLayer
      *
      * @param[in] input     Source tensor info. Data types supported: F16/F32.
      * @param[in] output    Destination tensor info. Data types supported: Same as @p input.
      * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[in] indices   (optional) The indices of the maximal values. Data type supported: U32.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr);
 
     void run() override final;
 
diff --git a/arm_compute/runtime/NEON/functions/NEPoolingLayer.h b/arm_compute/runtime/NEON/functions/NEPoolingLayer.h
index eb840b52f2..e43741c95b 100644
--- a/arm_compute/runtime/NEON/functions/NEPoolingLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEPoolingLayer.h
@@ -51,8 +51,9 @@ public:
      * @param[in, out] input     Source tensor. (Written to only when padding != 0) Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[out]     output    Destination tensor. Data types supported: Same as @p input.
      * @param[in]      pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[out]     indices   (optional) The indices of the maximal values. Data type supported: U32.
      */
-    void configure(ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info);
+    void configure(ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices = nullptr);
     /** Static function to check if given info will lead to a valid configuration of @ref NEPoolingLayer
      *
      * @note F16 is supported for pool sizes 2 and 3 only
@@ -60,10 +61,11 @@ public:
      * @param[in] input     Source tensor. (Written to only when padding != 0) Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] output    Destination tensor. Data types supported: Same as @p input.
      * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo.
+     * @param[in] indices   (optional) The indices of the maximal values. Data type supported: U32.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr);
 
     // Inherited methods overridden:
     void run() override;