COMPMID-1816: Use parallel reduction on 0 axis in CL ARG_MIN/ARG_MAX

Introducing new CLArgMinMax kernel

Change-Id: I0b8254207cc3859d19ceef9b6429cf5c1c586db0
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2202
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

7 files changed, 148 insertions, 33 deletions

diff --git a/arm_compute/core/CL/CLHelpers.h b/arm_compute/core/CL/CLHelpers.h
index cd65eafc9c..7e549be989 100644
--- a/arm_compute/core/CL/CLHelpers.h
+++ b/arm_compute/core/CL/CLHelpers.h
@@ -190,5 +190,16 @@ bool preferred_dummy_work_items_support(const cl::Device &device);
  * @return An opencl kernel
  */
 cl::Kernel create_opencl_kernel(CLCoreRuntimeContext *ctx, const std::string &kernel_name, const CLBuildOptions &build_opts);
+
+/** Creates a suitable LWS hint object for parallel implementations. Sets the number of WG based on the input size.
+ *  If input width is smaller than 128 we can use fewer threads than 8.
+ *
+ * @param[in] input_dimension number of elements along the dimension to apply the parallellization
+ * @param[in] vector_size     size of the vector in OpenCL
+ *
+ * @return An LWS hint object
+ */
+cl::NDRange create_lws_hint_parallel_implementations(unsigned int input_dimension, unsigned int vector_size);
+
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_CLHELPERS_H__ */
diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index c3c485db7c..78437beffb 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -21,13 +21,14 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_CLKERNELS_H__
-#define __ARM_COMPUTE_CLKERNELS_H__
+#ifndef ARM_COMPUTE_CLKERNELS_H
+#define ARM_COMPUTE_CLKERNELS_H
 
 /* Header regrouping all the CL kernels */
 #include "arm_compute/core/CL/kernels/CLAbsoluteDifferenceKernel.h"
 #include "arm_compute/core/CL/kernels/CLAccumulateKernel.h"
 #include "arm_compute/core/CL/kernels/CLActivationLayerKernel.h"
+#include "arm_compute/core/CL/kernels/CLArgMinMaxLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLBatchConcatenateLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLBatchNormalizationLayerKernel.h"
 #include "arm_compute/core/CL/kernels/CLBatchToSpaceLayerKernel.h"
@@ -160,4 +161,4 @@
 #include "arm_compute/core/CL/kernels/CLYOLOLayerKernel.h"
 #include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
 
-#endif /* __ARM_COMPUTE_CLKERNELS_H__ */
+#endif /* ARM_COMPUTE_CLKERNELS_H */
diff --git a/arm_compute/core/CL/kernels/CLArgMinMaxLayerKernel.h b/arm_compute/core/CL/kernels/CLArgMinMaxLayerKernel.h
new file mode 100644
index 0000000000..7f4cfe3edc
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLArgMinMaxLayerKernel.h
@@ -0,0 +1,94 @@
+/*
+ * 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.
+ */
+#ifndef ARM_COMPUTE_CLARGMINMAXLAYERKERNEL_H
+#define ARM_COMPUTE_CLARGMINMAXLAYERKERNEL_H
+
+#include "arm_compute/core/CL/ICLKernel.h"
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the reduction operation kernel
+ *
+ * @note The default data type for an uninitialized output tensor is
+ *       signed 32-bit integer (S32). It is the user's responsibility to check
+ *       that the results do not overflow because the indices are computed
+ *       in unsigned 32-bit (U32).
+ */
+class CLArgMinMaxLayerKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLArgMinMaxLayerKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLArgMinMaxLayerKernel(const CLArgMinMaxLayerKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLArgMinMaxLayerKernel &operator=(const CLArgMinMaxLayerKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLArgMinMaxLayerKernel(CLArgMinMaxLayerKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLArgMinMaxLayerKernel &operator=(CLArgMinMaxLayerKernel &&) = default;
+    /** Default destructor */
+    ~CLArgMinMaxLayerKernel() = default;
+
+    /** Set the input and output tensors.
+     *
+     * @param[in]  input       Source tensor. Data types supported: S32/F16/F32.
+     * @param[in]  prev_output Destination tensor of the previous iterations of @ref CLArgMinMaxLayerKernel. Data types supported: U32/S32
+     *                         Has to be nullptr for the first iteration
+     * @param[out] output      Destination tensor. Data types supported: U32/S32
+     *                         Output will have the same number of dimensions as input.
+     * @param[in]  axis        Axis along which to reduce. Supported reduction axis : 0,1,2,3
+     * @param[in]  op          Reduction operation to perform. Only ArgMin and ArgMax are supported.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *prev_output, ICLTensor *output, unsigned int axis, ReductionOperation op);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CLArgMinMaxLayerKernel.
+     *
+     * @param[in] input       Source tensor info. Data types supported: S32/F16/F32.
+     * @param[in] prev_output Destination tensor info of the previous iterations. Data types supported: U32/S32
+     *                        Has to be nullptr for the first iteration
+     * @param[in] output      Destination tensor info. Data types supported: U32/S32
+     *                        Output will have the same number of dimensions as input.
+     * @param[in] axis        Axis along which to reduce. Supported reduction axis : 0,1,2,3
+     * @param[in] op          Reduction operation to perform.  Only ArgMin and ArgMax are supported.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *prev_output, const ITensorInfo *output, unsigned int axis, ReductionOperation op);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor   *_input;
+    const ICLTensor   *_prev_output;
+    ICLTensor         *_output;
+    unsigned int       _reduction_axis;
+    ReductionOperation _op;
+};
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CLARGMINMAXLAYERKERNEL_H */
diff --git a/arm_compute/core/CL/kernels/CLReductionOperationKernel.h b/arm_compute/core/CL/kernels/CLReductionOperationKernel.h
index 172ed8985a..1ed7e6e5aa 100644
--- a/arm_compute/core/CL/kernels/CLReductionOperationKernel.h
+++ b/arm_compute/core/CL/kernels/CLReductionOperationKernel.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_CLREDUCTIONOPERATIONKERNEL_H__
-#define __ARM_COMPUTE_CLREDUCTIONOPERATIONKERNEL_H__
+#ifndef ARM_COMPUTE_CLREDUCTIONOPERATIONKERNEL_H
+#define ARM_COMPUTE_CLREDUCTIONOPERATIONKERNEL_H
 
 #include "arm_compute/core/CL/ICLKernel.h"
 #include "arm_compute/core/Types.h"
@@ -32,11 +32,6 @@ namespace arm_compute
 class ICLTensor;
 
 /** Interface for the reduction operation kernel
- *
- * @note For ARG_MIN/ARG_MAX reduction, the default data type for an uninitialized
- *       output tensor is signed 32-bit integer (S32). It is the user's responsibility
- *       to check that the results do not overflow because the indices are computed
- *       in unsigned 32-bit (U32).
  */
 class CLReductionOperationKernel : public ICLKernel
 {
@@ -57,10 +52,10 @@ public:
     /** Set the input and output tensors.
      *
      * @param[in]  input  Source tensor. Data types supported: QASYMM8/S32/F16/F32.
-     * @param[out] output Destination tensor. Data types and data layouts supported: Same as @p input, U32/S32 for ARG_MIX/ARG_MAX.
+     * @param[out] output Destination tensor. Data types and data layouts supported: Same as @p input.
      *                    Output will have the same number of dimensions as input.
      * @param[in]  axis   Axis along which to reduce. Supported reduction axis : 0,1,2,3
-     * @param[in]  op     Reduction operation to perform.
+     * @param[in]  op     Reduction operation to perform. Operations supported: MEAN_SUM, PROD, SUM_SQUARE, SUM, MIN, MAX
      * @param[in]  width  (Optional)  In case of x-axis we also need to provide the width of the input image.
      */
     void configure(const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, unsigned int width = 0);
@@ -68,10 +63,10 @@ public:
     /** Static function to check if given info will lead to a valid configuration of @ref CLReductionOperationKernel.
      *
      * @param[in] input  Source tensor info. Data types supported: QASYMM8/S32/F16/F32.
-     * @param[in] output Destination tensor info. Data types and data layouts supported: Same as @p input, U32/S32 for ARG_MIX/ARG_MAX.
+     * @param[in] output Destination tensor info. Data types and data layouts supported: Same as @p input.
      *                   Output will have the same number of dimensions as input.
      * @param[in] axis   Axis along which to reduce. Supported reduction axis : 0,1,2,3
-     * @param[in] op     Reduction operation to perform.
+     * @param[in] op     Reduction operation to perform. Operations supported: MEAN_SUM, PROD, SUM_SQUARE, SUM, MIN, MAX
      * @param[in] width  (Optional)  In case of x-axis we also need to provide the width of the input image.
      *
      * @return a status
@@ -90,4 +85,4 @@ private:
     BorderSize         _border_size;
 };
 } // namespace arm_compute
-#endif /*__ARM_COMPUTE_CLREDUCTIONOPERATIONKERNEL_H__ */
+#endif /*ARM_COMPUTE_CLREDUCTIONOPERATIONKERNEL_H */
diff --git a/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h b/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h
index 1b465a4866..21cded0417 100644
--- a/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h
+++ b/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h
@@ -21,10 +21,13 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_CLARGMINMAXLAYER_H__
-#define __ARM_COMPUTE_CLARGMINMAXLAYER_H__
+#ifndef ARM_COMPUTE_CLARGMINMAXLAYER_H
+#define ARM_COMPUTE_CLARGMINMAXLAYER_H
 
+#include "arm_compute/core/CL/kernels/CLArgMinMaxLayerKernel.h"
+#include "arm_compute/core/CL/kernels/CLReshapeLayerKernel.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
@@ -33,7 +36,6 @@ namespace arm_compute
 {
 class ITensorInfo;
 class ICLTensor;
-class CLReductionOperation;
 
 /** Function to calculate the index of the minimum or maximum values in a
  *  tensor based on an axis.
@@ -53,19 +55,18 @@ public:
     CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Set the input and output tensors.
      *
-     * @param[in]  input  Input source tensor, this could be written if @ref CLReductionOperation
-     *                    manipulates its border for better performance. Data types supported: F16/F32.
+     * @param[in]  input  Input source tensor. Data types supported: F16/F32.
      * @param[in]  axis   Axis to find max/min index.
      * @param[out] output Output source tensor. Data types supported: U32/S32.
-     * @param[in]  op     Operation to perform: min or max
+     * @param[in]  op     Reduction operation to perform. Operations supported: ARG_IDX_MAX, ARG_IDX_MIN
      */
-    void configure(ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op);
+    void configure(const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op);
     /** Static function to check if given info will lead to a valid configuration of @ref CLArgMinMaxLayer
      *
      * @param[in] input  Input source tensor info. Data types supported: F16/F32.
      * @param[in] axis   Axis to find max/min index.
      * @param[in] output Output source tensor info. Data types supported: U32/S32.
-     * @param[in] op     Operation to perform: min or max
+     * @param[in] op     Reduction operation to perform. Operations supported: ARG_IDX_MAX, ARG_IDX_MIN
      *
      * @return a status
      */
@@ -75,7 +76,13 @@ public:
     void run() override;
 
 private:
-    std::unique_ptr<CLReductionOperation> _reduction_function;
+    MemoryGroup                         _memory_group;
+    std::vector<CLTensor>               _results_vector;
+    CLTensor                            _not_reshaped_output;
+    std::vector<CLArgMinMaxLayerKernel> _reduction_kernels_vector;
+    CLReshapeLayerKernel                _reshape_kernel;
+    unsigned int                        _num_of_stages;
+    unsigned int                        _reduction_axis;
 };
 } // namespace arm_compute
-#endif /* __ARM_COMPUTE_CLARGMINMAXLAYER_H__ */
+#endif /* ARM_COMPUTE_CLARGMINMAXLAYER_H */
diff --git a/arm_compute/runtime/CL/functions/CLReductionOperation.h b/arm_compute/runtime/CL/functions/CLReductionOperation.h
index 405e1177fd..9e0bf03ffe 100644
--- a/arm_compute/runtime/CL/functions/CLReductionOperation.h
+++ b/arm_compute/runtime/CL/functions/CLReductionOperation.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_CLREDUCTIONOPERATION_H__
-#define __ARM_COMPUTE_CLREDUCTIONOPERATION_H__
+#ifndef ARM_COMPUTE_CLREDUCTIONOPERATION_H
+#define ARM_COMPUTE_CLREDUCTIONOPERATION_H
 
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLReductionOperationKernel.h"
@@ -57,7 +57,7 @@ public:
      * @param[in]  input     Source tensor. Data types supported: QASYMM8/F16/F32.
      * @param[out] output    Destination tensor. Data types and data layouts supported: Same as @p input.
      * @param[in]  axis      Axis along which to reduce. Supported reduction axis : 0, 1, 2, 3
-     * @param[in]  op        Reduction operation to perform.
+     * @param[in]  op        Reduction operation to perform. Operations supported: MEAN_SUM, PROD, SUM_SQUARE, SUM, MIN, MAX
      * @param[in]  keep_dims (Optional) Whether to keep the reduced dimension after the operation. Defaults to true.
      */
     void configure(ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, bool keep_dims = true);
@@ -67,7 +67,7 @@ public:
      * @param[in] input     Source tensor info. Data types supported: QASYMM8/F16/F32.
      * @param[in] output    Destination tensor info. Data types and data layouts supported: Same as @p input.
      * @param[in] axis      Axis along which to reduce. Supported reduction axis : 0, 1, 2, 3
-     * @param[in] op        Reduction operation to perform.
+     * @param[in] op        Reduction operation to perform. Operations supported: MEAN_SUM, PROD, SUM_SQUARE, SUM, MIN, MAX
      * @param[in] keep_dims (Optional) Whether to keep the reduced dimension after the operation. Defaults to true.
      *
      * @return a status
@@ -92,4 +92,4 @@ private:
     bool                                    _is_reshape_required;
 };
 } // namespace arm_compute
-#endif /*__ARM_COMPUTE_CLREDUCTIONOPERATION_H__ */
+#endif /* ARM_COMPUTE_CLREDUCTIONOPERATION_H */
\ No newline at end of file
diff --git a/arm_compute/runtime/Utils.h b/arm_compute/runtime/Utils.h
index 15c0042a33..9a5b20eb26 100644
--- a/arm_compute/runtime/Utils.h
+++ b/arm_compute/runtime/Utils.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef __ARM_COMPUTE_RUNTIME_UTILS_H__
-#define __ARM_COMPUTE_RUNTIME_UTILS_H__
+#ifndef ARM_COMPUTE_RUNTIME_UTILS_H
+#define ARM_COMPUTE_RUNTIME_UTILS_H
 
 #include "arm_compute/runtime/IRuntimeContext.h"
 #include "arm_compute/runtime/Scheduler.h"
@@ -46,5 +46,12 @@ const std::string &string_from_scheduler_type(Scheduler::Type t);
  * @param[in] hints  Hints to use.
  */
 void schedule_kernel_on_ctx(IRuntimeContext *ctx, ICPPKernel *kernel, const IScheduler::Hints &hints);
+
+/** Calculate number of stages for parallel implementations
+ *
+ * @param[in] input_x_dimension input tensor x dimension
+ * @param[in] axis              axis to be used
+ */
+unsigned int calculate_number_of_stages_only_x_axis(size_t input_x_dimension, unsigned int axis);
 } // namespace arm_compute
-#endif /* __ARM_COMPUTE_RUNTIME_UTILS_H__ */
+#endif /* ARM_COMPUTE_RUNTIME_UTILS_H */