Rewrote CLArgMinMax for axis 0

* Simpler implementation without stages for axis 0

* Removed considerable amount of code.

Resolves COMPMID-6271

Change-Id: Ie8bcb2f0b55f87472f44b38872a23a922619a211
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9849
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 200 insertions, 372 deletions

diff --git a/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h b/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h
index a971163c45..ce5bee8d95 100644
--- a/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h
+++ b/arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -107,13 +107,11 @@ public:
     void run() override;
 
 private:
-    MemoryGroup                                          _memory_group;
-    std::vector<CLTensor>                                _results_vector;
-    CLTensor                                             _not_reshaped_output;
-    std::vector<std::unique_ptr<CLArgMinMaxLayerKernel>> _reduction_kernels_vector;
-    CLReshapeLayer                                       _reshape;
-    unsigned int                                         _num_of_stages;
-    unsigned int                                         _reduction_axis;
+    MemoryGroup                             _memory_group;
+    CLTensor                                _not_reshaped_output;
+    std::unique_ptr<CLArgMinMaxLayerKernel> _arg_min_max_kernel;
+    CLReshapeLayer                          _reshape;
+    unsigned int                            _reduction_axis;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_CLARGMINMAXLAYER_H */
diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp
index 1d53b9a093..77f0d6ac32 100644
--- a/src/core/CL/CLHelpers.cpp
+++ b/src/core/CL/CLHelpers.cpp
@@ -145,7 +145,6 @@ std::string get_cl_select_type_from_data_type(const DataType &dt)
     {
         case DataType::U8:
         case DataType::QASYMM8:
-            return "uchar";
         case DataType::S8:
         case DataType::QASYMM8_SIGNED:
         case DataType::QSYMM8:
diff --git a/src/core/CL/cl_kernels/common/arg_min_max.cl b/src/core/CL/cl_kernels/common/arg_min_max.cl
index 6e57ed0af1..438f46eb24 100644
--- a/src/core/CL/cl_kernels/common/arg_min_max.cl
+++ b/src/core/CL/cl_kernels/common/arg_min_max.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 #include "helpers.h"
+#include "tile_helpers.h"
 
 #if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE_OUTPUT)
 
@@ -52,246 +53,183 @@
 #endif // defined(ARG_MAX)
 
 #if defined(WIDTH)
-#if defined(ARG_MIN)
-#if defined(PREV_OUTPUT)
-/** Find index minimum value of a vector
- *
- * @param[in] input Pointer to the first value.
- *
- * @return index of the vector.
- */
-inline DATA_TYPE_OUTPUT arg_idx_min_prev_out(__global const DATA_TYPE *input, __global const DATA_TYPE_OUTPUT *prev_res, const int x_idx)
+
+#if defined(ARG_MAX)
+#define VECTOR_PREDICATE_EQ(x, y) ((x) >= (y))
+#define VECTOR_PREDICATE(x, y) ((x) > (y))
+#define SCALAR_SELECT_OP(x, y) ((x) > (y)) ? (x) : (y);
+#elif defined(ARG_MIN)
+#define VECTOR_PREDICATE_EQ(x, y) ((x) <= (y))
+#define VECTOR_PREDICATE(x, y) ((x) < (y))
+#define SCALAR_SELECT_OP(x, y) ((x) < (y)) ? (x) : (y);
+#else // !(defined(ARG_MAX) || defined(ARG_MIN))
+#error "Unsupported reduction operation!"
+#endif // defined(ARG_MAX)
+
+inline DATA_TYPE_OUTPUT vectorized_compute_arg_min_max_2(DATA_TYPE *min_max_val, DATA_TYPE_OUTPUT *min_max_idx, VEC_DATA_TYPE(DATA_TYPE, 2) in, VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 2) res)
 {
-    int end_elem = (x_idx + 1) * 16;
-    if(end_elem > WIDTH)
+    if( VECTOR_PREDICATE_EQ(in.s0,in.s1) )
     {
-        end_elem = WIDTH - x_idx * 16;
+        *min_max_val  = in.s0;
+        *min_max_idx  = res.s0;
     }
-    DATA_TYPE_OUTPUT res = prev_res[0];
-    for(int x_v = 1; x_v < end_elem; ++x_v)
+    else
     {
-        res = select(res, prev_res[x_v], *(input + prev_res[x_v]) < * (input + res));
+        *min_max_val  = in.s1;
+        *min_max_idx  = res.s1;
     }
-    return res;
 }
-#else // !defined(PREV_OUTPUT)
-/** Find index minimum value of a vector
- *
- * @param[in] input Pointer to the first value.
- *
- * @return index of the vector.
- */
-inline DATA_TYPE_OUTPUT arg_idx_min(__global const DATA_TYPE *input, const int x_idx)
+
+inline DATA_TYPE_OUTPUT vectorized_compute_arg_min_max_4(DATA_TYPE *min_max_val, DATA_TYPE_OUTPUT *min_max_idx, VEC_DATA_TYPE(DATA_TYPE, 4) in, VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 4) res)
 {
-#if WIDTH < 16
-    DATA_TYPE_OUTPUT res = 0;
-    for(DATA_TYPE_OUTPUT x_v = res + 1; x_v < WIDTH; ++x_v)
-    {
-        res = select(res, x_v, *(input + x_v) < * (input + res));
-    }
-    return res;
-#else  // WIDTH >= 16
-    int       x_elem   = x_idx * 16;
-    const int x_goback = select(0, 16 - WIDTH % 16, x_elem + 16 > WIDTH);
-    x_elem -= x_goback;
-
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    in = vload16(0, input - x_goback);
-    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)
-    res = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
-
-    SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 8)
-    idx_sel       = (in.s01234567 <= in.s89abcdef);
+    VEC_DATA_TYPE(COND_DATA_TYPE, 2)
+    idx_sel       = VECTOR_PREDICATE_EQ(in.s01, in.s23);
+    in.s01      = select(in.s23, in.s01, idx_sel);
+    res.s01     = select(res.s23, res.s01, CONVERT(idx_sel, int2));
+    idx_sel.s0    = VECTOR_PREDICATE(in.s0, in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), COND_DATA_TYPE));
+    res.s0        = select(res.s1, res.s0, CONVERT(idx_sel.s0, int));
+    *min_max_val  = SCALAR_SELECT_OP(in.s0, in.s1);
+    *min_max_idx  = res.s0;
+}
+
+inline DATA_TYPE_OUTPUT vectorized_compute_arg_min_max_8(DATA_TYPE *min_max_val, DATA_TYPE_OUTPUT *min_max_idx, VEC_DATA_TYPE(DATA_TYPE, 8) in, VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 8) res)
+{
+    VEC_DATA_TYPE(COND_DATA_TYPE, 4)
+    idx_sel       = VECTOR_PREDICATE_EQ(in.s0123, in.s4567);
+    in.s0123      = select(in.s4567, in.s0123, idx_sel);
+    res.s0123     = select(res.s4567, res.s0123, CONVERT(idx_sel, int4));
+    idx_sel.s01   = (VECTOR_PREDICATE(in.s01, in.s23)) || (in.s01 == in.s23 && CONVERT(((res.s01 < res.s23)), VEC_DATA_TYPE(COND_DATA_TYPE, 2)));
+    in.s01        = select(in.s23, in.s01, idx_sel.s01);
+    res.s01       = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2));
+    idx_sel.s0    = VECTOR_PREDICATE(in.s0, in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), COND_DATA_TYPE));
+    res.s0        = select(res.s1, res.s0, CONVERT(idx_sel.s0, int));
+    *min_max_val  = SCALAR_SELECT_OP(in.s0, in.s1);
+    *min_max_idx  = res.s0;
+}
+
+inline DATA_TYPE_OUTPUT vectorized_compute_arg_min_max_16(DATA_TYPE *min_max_val, DATA_TYPE_OUTPUT *min_max_idx, VEC_DATA_TYPE(DATA_TYPE, 16) in, VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16) res)
+{
+    VEC_DATA_TYPE(COND_DATA_TYPE, 8)
+    idx_sel       = VECTOR_PREDICATE_EQ(in.s01234567, in.s89abcdef);
     in.s01234567  = select(in.s89abcdef, in.s01234567, idx_sel);
     res.s01234567 = select(res.s89abcdef, res.s01234567, CONVERT(idx_sel, int8));
-
-    idx_sel.s0123 = (in.s0123 < in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 4)));
+    idx_sel.s0123 = VECTOR_PREDICATE(in.s0123, in.s4567) || (in.s0123 == in.s4567 && CONVERT(((res.s0123 < res.s4567)), VEC_DATA_TYPE(COND_DATA_TYPE, 4)));
     in.s0123      = select(in.s4567, in.s0123, idx_sel.s0123);
     res.s0123     = select(res.s4567, res.s0123, CONVERT(idx_sel.s0123, int4));
+    idx_sel.s01   = (VECTOR_PREDICATE(in.s01, in.s23)) || (in.s01 == in.s23 && CONVERT(((res.s01 < res.s23)), VEC_DATA_TYPE(COND_DATA_TYPE, 2)));
+    in.s01        = select(in.s23, in.s01, idx_sel.s01);
+    res.s01       = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2));
+    idx_sel.s0    = VECTOR_PREDICATE(in.s0, in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), COND_DATA_TYPE));
+    res.s0        = select(res.s1, res.s0, CONVERT(idx_sel.s0, int));
+    *min_max_val  = SCALAR_SELECT_OP(in.s0, in.s1);
+    *min_max_idx  = res.s0;
+}
 
-    idx_sel.s01 = (in.s01 < in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 2)));
-    in.s01      = select(in.s23, in.s01, idx_sel.s01);
-    res.s01     = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2));
 
-    idx_sel.s0 = (in.s0 < in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), SIGNED_INT_DATA_TYPE(DATA_TYPE)));
-    res.s0     = select(res.s1, res.s0, CONVERT(idx_sel.s0, int));
 
-    return res.s0 + x_elem;
-#endif // WIDTH < 16
-}
-#endif // defined(PREV_OUTPUT)
-#endif // defined(ARG_MIN)
-#if defined(ARG_MAX)
-#if defined(PREV_OUTPUT)
-/** Find index maximum value of a vector
- *
- * @param[in] input Pointer to the first value.
- *
- * @return index of the vector.
- */
-inline DATA_TYPE_OUTPUT arg_idx_max_prev_out(__global const DATA_TYPE *input, __global const DATA_TYPE_OUTPUT *prev_res, const int x_idx)
+inline void scalar_compute_global_min_max(DATA_TYPE in_val, int idx, DATA_TYPE *out_min_max_val, DATA_TYPE_OUTPUT *out_idx)
 {
-    int end_elem = (x_idx + 1) * 16;
-    if(end_elem > WIDTH)
-    {
-        end_elem = WIDTH - x_idx * 16;
-    }
-    DATA_TYPE_OUTPUT res = prev_res[0];
-    for(int x_v = 1; x_v < end_elem; ++x_v)
+#if defined(ARG_MAX)
+    if(in_val > *out_min_max_val)
+#else  // defined(ARG_MAX)
+    if(in_val < *out_min_max_val)
+#endif // defined(ARG_MAX)
     {
-        res = select(res, prev_res[x_v], *(input + prev_res[x_v]) > *(input + res));
+        *out_min_max_val = in_val;
+        *out_idx         = idx;
     }
-    return res;
 }
-#else // !defined(PREV_OUTPUT)
-/** Find index maximum value of a vector
- *
- * @param[in] input Pointer to the first value.
- *
- * @return index of the vector.
- */
-inline DATA_TYPE_OUTPUT arg_idx_max(__global const DATA_TYPE *input, const int x_idx)
-{
-#if WIDTH < 16
-    DATA_TYPE_OUTPUT res = 0;
-    for(DATA_TYPE_OUTPUT x_v = res + 1; x_v < WIDTH; ++x_v)
-    {
-        res = select(res, x_v, *(input + x_v) > *(input + res));
-    }
-    return res;
-#else  // WIDTH >= 16
-    int       x_elem   = x_idx * 16;
-    const int x_goback = select(0, 16 - WIDTH % 16, x_elem + 16 > WIDTH);
-    x_elem -= x_goback;
-
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    in = vload16(0, input - x_goback);
-    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, 16)
-    res = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
-
-    SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 8)
-    idx_sel       = (in.s01234567 >= in.s89abcdef);
-    in.s01234567  = select(in.s89abcdef, in.s01234567, idx_sel);
-    res.s01234567 = select(res.s89abcdef, res.s01234567, CONVERT(idx_sel, int8));
-
-    idx_sel.s0123 = (in.s0123 > in.s4567) || (in.s0123 == in.s4567 && CONVERT((res.s0123 < res.s4567), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 4)));
-    in.s0123      = select(in.s4567, in.s0123, idx_sel.s0123);
-    res.s0123     = select(res.s4567, res.s0123, CONVERT(idx_sel.s0123, int4));
-
-    idx_sel.s01 = (in.s01 > in.s23) || (in.s01 == in.s23 && CONVERT((res.s01 < res.s23), SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, 2)));
-    in.s01      = select(in.s23, in.s01, idx_sel.s01);
-    res.s01     = select(res.s23, res.s01, CONVERT(idx_sel.s01, int2));
-
-    idx_sel.s0 = (in.s0 > in.s1) || (in.s0 == in.s1 && CONVERT((res.s0 < res.s1), SIGNED_INT_DATA_TYPE(DATA_TYPE)));
-    res.s0     = select(res.s1, res.s0, CONVERT(idx_sel.s0, int));
 
-    return res.s0 + x_elem;
-#endif // WIDTH < 16
+#if VEC_SIZE > 1
+#if VEC_SIZE == 16
+    #define VECTORIZED_OP(min_max_val,min_max_idx,in,res) vectorized_compute_arg_min_max_16(min_max_val,min_max_idx,in,res)
+#elif VEC_SIZE == 8 // #if VEC_SIZE == 16
+    #define VECTORIZED_OP(min_max_val,min_max_idx,in,res) vectorized_compute_arg_min_max_8(min_max_val,min_max_idx,in,res)
+#elif VEC_SIZE == 4 // # elif VEC_SIZE == 8
+    #define VECTORIZED_OP(min_max_val,min_max_idx,in,res) vectorized_compute_arg_min_max_4(min_max_val,min_max_idx,in,res)
+#elif VEC_SIZE == 2 // elif VEC_SIZE == 4
+    #define VECTORIZED_OP(min_max_val,min_max_idx,in,res) vectorized_compute_arg_min_max_2(min_max_val,min_max_idx,in,res)
+#else // elif VEC_SIZE == 2
+    #error "Not supported"
+#endif // #if VEC_SIZE == 16
+
+inline VEC_DATA_TYPE(DATA_TYPE_OUTPUT, VEC_SIZE) init_idx_vector()
+{
+#if VEC_SIZE == 16
+    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, VEC_SIZE)
+    vidx = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
+#elif VEC_SIZE == 8 // #if VEC_SIZE == 16
+    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, VEC_SIZE)
+    vidx = { 0, 1, 2, 3, 4, 5, 6, 7 };
+#elif VEC_SIZE == 4 // elif VEC_SIZE == 8
+    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, VEC_SIZE)
+    vidx = { 0, 1, 2, 3 };
+#elif VEC_SIZE == 2 // elif VEC_SIZE == 4
+    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, VEC_SIZE)
+    vidx = { 0, 1 };
+#else  // elif VEC_SIZE == 2
+#error "Not supported"
+#endif // #if VEC_SIZE == 16
+    return vidx;
 }
-#endif // defined(PREV_OUTPUT)
-#endif // defined(ARG_MAX)
+#endif // VEC_SIZE > 1
 
-/** This kernel performs parallel reduction given an operation on x-axis.
+/** This kernel performs reduction on x-axis.
  *
- * @note In case the results of previous stages are passed the flag PREV_OUTPUT has to be passed using -DPREV_OUTPUT
- * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The input data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
  * @note The data type of the output must be passed at compile time using -DDATA_TYPE_OUTPUT: e.g. -DDATA_TYPE_OUTPUT=uint
- * @note The arg_max flag must be passed at compile time using -DARG_MAX if we want to compute the ArgMax
- * @note The arg_min flag must be passed at compile time using -DARG_MIN if we want to compute the ArgMin
+ * @note The data type used for the comparing indexe must be passed at compile type using -DCOND_DATA_TYPE: e.g -DCOND_DATA_TYPE=uint
+ * @note The height size must be passed at compile time using -DHEIGHT e.g. -DHEIGHT=128
  *
- * @param[in] src_ptr                                   Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/S32/F16/F32
- * @param[in] src_stride_x                              Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x                                src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y                              Stride of the source tensor in Y dimension (in bytes)
- * @param[in] src_step_y                                src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes         The offset of the first element in the source tensor
- * @param[in] prev_res_ptr                              (Optional) Pointer to previous results tensor. Supported data types: U32/S32
- * @param[in] prev_res_stride_x                         (Optional) Stride of the output tensor in X dimension (in bytes)
- * @param[in] prev_res_step_x                           (Optional) prev_res_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] prev_res_stride_y                         (Optional) Stride of the output tensor in Y dimension (in bytes)
- * @param[in] prev_res_step_y                           (Optional) prev_res_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] prev_res_offset_first_element_in_bytes    (Optional) The offset of the first element in the previous results tensor
- * @param[in] partial_res_ptr                           The local buffer to hold partial result values. Supported data types: U32/S32
- * @param[in] partial_res_stride_x                      Stride of the output tensor in X dimension (in bytes)
- * @param[in] partial_res_step_x                        partial_res_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] partial_res_stride_y                      Stride of the output tensor in Y dimension (in bytes)
- * @param[in] partial_res_step_y                        partial_res_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] partial_res_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[in] local_results                             Local buffer for storing the partial result
+ * @param[in] input_ptr                            Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/S32/F16/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)
+ * @param[in] input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[in] output_ptr                           The local buffer to hold sumed values. Supported data types: U32/S32
+ * @param[in] output_stride_x                      Stride of the output tensor in X dimension (in bytes)
+ * @param[in] output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] output_stride_y                      Stride of the output tensor in Y dimension (in bytes)
+ * @param[in] output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor
  */
 __kernel void arg_min_max_x(
-    IMAGE_DECLARATION(src),
-#if defined(PREV_OUTPUT)
-    IMAGE_DECLARATION(prev_res),
-#endif // defined(PREV_OUTPUT)
-    IMAGE_DECLARATION(partial_res),
-    __local DATA_TYPE_OUTPUT *local_results)
+    IMAGE_DECLARATION(input),
+    IMAGE_DECLARATION(output))
 {
-#if defined(PREV_OUTPUT)
-    Image src      = CONVERT_TO_IMAGE_STRUCT_NO_STEP(src);
-    Image prev_res = CONVERT_TO_IMAGE_STRUCT(prev_res);
-#else  // !defined(PREV_OUTPUT)
-    Image src                      = CONVERT_TO_IMAGE_STRUCT(src);
-#endif // defined(PREV_OUTPUT)
-    Image partial_res = CONVERT_TO_IMAGE_STRUCT(partial_res);
-
-    unsigned int lsize = get_local_size(0);
-    unsigned int lid   = get_local_id(0);
-
-    const uint     x_idx                 = get_global_id(0);
-    const uint     y_idx                 = get_global_id(1);
-    const __global DATA_TYPE *src_in_row = (const __global DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + y_idx * src_step_y);
-
-    for(unsigned int y = 0; y < get_local_size(1); ++y)
+    __global DATA_TYPE *input_addr         = (__global DATA_TYPE *)(input_ptr + input_offset_first_element_in_bytes + get_global_id(1) * input_stride_y);
+    __global DATA_TYPE_OUTPUT *output_addr = (__global DATA_TYPE_OUTPUT *)(output_ptr + output_offset_first_element_in_bytes + get_global_id(1) * output_stride_y);
+
+    DATA_TYPE        final_value = input_addr[0];
+    DATA_TYPE_OUTPUT final_idx   = 0;
+
+#if VEC_SIZE > 1
+    VEC_DATA_TYPE(DATA_TYPE_OUTPUT, VEC_SIZE)
+    vidx = init_idx_vector();
+
+    int x = 0;
+    for(; x <= (WIDTH - VEC_SIZE); x += VEC_SIZE)
     {
-#if defined(ARG_MAX)
-#if defined(PREV_OUTPUT)
-        local_results[lid] = arg_idx_max_prev_out(src_in_row, (__global DATA_TYPE_OUTPUT *)offset(&prev_res, 0, y), x_idx);
-#else  // !defined(PREV_OUTPUT)
-        local_results[lid] = arg_idx_max((__global DATA_TYPE *)offset(&src, 0, y), x_idx);
-#endif // defined(PREV_OUTPUT)
-#else  // defined(ARG_MIN)
-#if defined(PREV_OUTPUT)
-        local_results[lid]         = arg_idx_min_prev_out(src_in_row, (__global DATA_TYPE_OUTPUT *)offset(&prev_res, 0, y), x_idx);
-#else  // !defined(PREV_OUTPUT)
-        local_results[lid] = arg_idx_min((__global DATA_TYPE *)offset(&src, 0, y), x_idx);
-#endif // defined(PREV_OUTPUT)
-#endif // defined(ARG_MAX) || defined(ARG_MIN)
-
-        barrier(CLK_LOCAL_MEM_FENCE);
-
-        // Looking for the next highest power of 2 (maximum value of lsize is 8)
-        unsigned int middle = lsize - 1;
-        middle |= middle >> 1;
-        middle |= middle >> 2;
-        middle += 1;
-        // Perform parallel reduction
-        for(unsigned int i = middle; i > 0; i >>= 1)
-        {
-            if(lid < i && lid + i < lsize)
-            {
-                DATA_TYPE tmp0 = *(src_in_row + local_results[lid]);
-                DATA_TYPE tmp1 = *(src_in_row + local_results[lid + i]);
-#if defined(ARG_MAX)
-                local_results[lid] = select(
-                                         local_results[lid],
-                                         local_results[lid + i],
-                                         ((tmp0 == tmp1) && (local_results[lid + i] < local_results[lid])) || (tmp0 < tmp1));
-#else  // defined(ARG_MIN)
-                local_results[lid] = select(
-                                         local_results[lid],
-                                         local_results[lid + i],
-                                         ((tmp0 == tmp1) && (local_results[lid + i] < local_results[lid])) || (tmp0 > tmp1));
-#endif // defined(ARG_MAX) || defined(ARG_MIN)
-            }
-            barrier(CLK_LOCAL_MEM_FENCE);
-        }
-
-        if(lid == 0)
-        {
-            ((__global DATA_TYPE_OUTPUT *)offset(&partial_res, get_group_id(0), y))[0] = local_results[0];
-        }
+        VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+        vals = VLOAD(VEC_SIZE)(0, (input_addr + x));
+        DATA_TYPE        local_min_max_value;
+        DATA_TYPE_OUTPUT local_min_max_idx;
+
+        VECTORIZED_OP(&local_min_max_value, &local_min_max_idx, vals, vidx);
+        local_min_max_idx += x;
+        scalar_compute_global_min_max(local_min_max_value, local_min_max_idx, &final_value, &final_idx);
     }
+#endif // VEC_SIZE > 1
+
+#if(WIDTH % VEC_SIZE)
+    LOOP_UNROLLING(int, j, 0, 1, WIDTH % VEC_SIZE,
+    {
+        scalar_compute_global_min_max(*(input_addr + j + x), j + x, &final_value, &final_idx);
+    })
+#endif // (WIDTH % VEC_SIZE)
+
+    output_addr[0] = final_idx;
 }
 #endif // defined(WIDTH)
 
@@ -320,8 +258,7 @@ __kernel void arg_min_max_y(
     IMAGE_DECLARATION(input),
     IMAGE_DECLARATION(output))
 {
-    const int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
-
+    const int x_offs            = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
     __global uchar *input_addr  = input_ptr + input_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * input_stride_y;
     __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE_OUTPUT) + get_global_id(1) * output_stride_y;
 
@@ -448,4 +385,4 @@ __kernel void arg_min_max_w(
     STORE_VECTOR_SELECT(indx, DATA_TYPE_OUTPUT, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
 }
 #endif /* defined(BATCH) && defined(DEPTH) */
-#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE_OUTPUT)
\ No newline at end of file
+#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE_OUTPUT)
diff --git a/src/core/CL/kernels/CLArgMinMaxLayerKernel.cpp b/src/core/CL/kernels/CLArgMinMaxLayerKernel.cpp
index 7af2fa1e64..8438739764 100644
--- a/src/core/CL/kernels/CLArgMinMaxLayerKernel.cpp
+++ b/src/core/CL/kernels/CLArgMinMaxLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -33,14 +33,13 @@
 #include "src/core/CL/CLValidate.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
-
 #include "support/StringSupport.h"
 
 namespace arm_compute
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *prev_output, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
@@ -53,31 +52,23 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *prev_outp
     {
         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32, DataType::S32);
     }
-    if(prev_output != nullptr && prev_output->total_size() != 0)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(prev_output, 1, DataType::U32, DataType::S32);
-        if(output->total_size() != 0)
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(prev_output, output);
-        }
-    }
 
     return Status{};
 }
 } // namespace
 
 CLArgMinMaxLayerKernel::CLArgMinMaxLayerKernel()
-    : _input(nullptr), _prev_output(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::ARG_IDX_MAX)
+    : _input(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::ARG_IDX_MAX)
 {
     _type = CLKernelType::ELEMENTWISE;
 }
 
-void CLArgMinMaxLayerKernel::configure(const ICLTensor *input, const ICLTensor *prev_output, ICLTensor *output, unsigned int axis, ReductionOperation op)
+void CLArgMinMaxLayerKernel::configure(const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op)
 {
-    configure(CLKernelLibrary::get().get_compile_context(), input, prev_output, output, axis, op);
+    configure(CLKernelLibrary::get().get_compile_context(), input, output, axis, op);
 }
 
-void CLArgMinMaxLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *prev_output, ICLTensor *output, unsigned int axis, ReductionOperation op)
+void CLArgMinMaxLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
@@ -85,42 +76,35 @@ void CLArgMinMaxLayerKernel::configure(const CLCompileContext &compile_context,
     output_shape.set(axis, 1);
     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_type(DataType::S32).reset_padding().set_is_resizable(true));
 
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (prev_output != nullptr) ? prev_output->info() : nullptr, output->info(), axis, op));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis, op));
 
-    auto padding_info = get_padding_info({ input, prev_output, output });
+    auto padding_info = get_padding_info({ input, output });
 
     _input          = input;
-    _prev_output    = prev_output;
     _output         = output;
     _reduction_axis = axis;
     _op             = op;
 
     // Set build options
-    const auto vector_size = (axis == 0) ? 16U : adjust_vec_size(16U, input->info()->dimension(0));
-
+    const auto     vector_size = adjust_vec_size(16U, input->info()->dimension(0));
     CLBuildOptions build_opts;
-    build_opts.add_option_if(_prev_output != nullptr, "-DPREV_OUTPUT");
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->info()->dimension(0) % vector_size));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vector_size));
     build_opts.add_option_if(is_data_type_float(input->info()->data_type()), "-DFLOAT_DATA_TYPE");
     build_opts.add_option_if_else(op == ReductionOperation::ARG_IDX_MAX, "-DARG_MAX", "-DARG_MIN");
     build_opts.add_option("-DDATA_TYPE_OUTPUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+    build_opts.add_option("-DCOND_DATA_TYPE=" + get_cl_select_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DUNROLL_WITH_PRAGMA=1");
 
     // Create kernel
-    cl::NDRange lws_hint = CLKernelLibrary::get().default_ndrange();
     std::string kernel_axis_name;
     switch(axis)
     {
         case 0:
-        {
-            const ICLTensor *input_for_width = prev_output != nullptr ? _prev_output : _input;
-            build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input_for_width->info()->dimension(0)));
-
+            build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
             kernel_axis_name = "x";
-            lws_hint         = create_lws_hint_parallel_implementations(input_for_width->info()->dimension(0), vector_size);
-        }
-        break;
+            break;
         case 1:
             build_opts.add_option("-DHEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
             kernel_axis_name = "y";
@@ -140,15 +124,15 @@ void CLArgMinMaxLayerKernel::configure(const CLCompileContext &compile_context,
     _kernel = create_kernel(compile_context, "arg_min_max_" + kernel_axis_name, build_opts.options());
 
     // Configure kernel window
-    Window win = calculate_max_window((prev_output != nullptr) ? (*prev_output->info()) : (*input->info()), Steps(vector_size));
-    ICLKernel::configure_internal(win, lws_hint);
+    Window win = calculate_max_window(*input->info(), Steps(vector_size));
+    ICLKernel::configure_internal(win);
 
     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
 
-Status CLArgMinMaxLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *prev_output, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
+Status CLArgMinMaxLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, prev_output, output, axis, op));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis, op));
     return Status{};
 }
 
@@ -163,30 +147,22 @@ void CLArgMinMaxLayerKernel::run(const Window &window, cl::CommandQueue &queue)
         {
             // Set out window
             Window out_window(window);
+            Window in_window(window);
             out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
+            in_window.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));
+            in_window.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), 1u));
 
             // Get first input and output slices
-            Window in_slice  = window.first_slice_window_2D();
+            Window in_slice  = in_window.first_slice_window_2D();
             Window out_slice = out_window.first_slice_window_2D();
-
-            // Reshape window
-            const unsigned int num_tensors = _prev_output != nullptr ? 3 : 2;
-
-            // Set local sums buffer
-            unsigned int local_res_size = lws_hint()[0] * _output->info()->element_size();
-            _kernel.setArg(num_arguments_per_2D_tensor() * num_tensors, local_res_size, nullptr);
             do
             {
                 unsigned int idx = 0;
                 add_2D_tensor_argument(idx, _input, in_slice);
-                if(_prev_output != nullptr)
-                {
-                    add_2D_tensor_argument(idx, _prev_output, in_slice);
-                }
                 add_2D_tensor_argument(idx, _output, out_slice);
                 enqueue(queue, *this, in_slice, lws_hint());
             }
-            while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
+            while(in_window.slide_window_slice_2D(in_slice) && out_window.slide_window_slice_2D(out_slice));
         }
         break;
         case 1:
diff --git a/src/core/CL/kernels/CLArgMinMaxLayerKernel.h b/src/core/CL/kernels/CLArgMinMaxLayerKernel.h
index 929677f905..5f36bdf113 100644
--- a/src/core/CL/kernels/CLArgMinMaxLayerKernel.h
+++ b/src/core/CL/kernels/CLArgMinMaxLayerKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2020, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -56,48 +56,41 @@ public:
 
     /** Set the input and output tensors.
      *
-     * @param[in]  input       Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/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.
+     * @param[in]  input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/S32/F16/F32.
+     * @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);
+    void configure(const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op);
     /** Set the input and output tensors.
      *
      * @param[in]  compile_context The compile context to be used.
      * @param[in]  input           Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/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 CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *prev_output, ICLTensor *output, unsigned int axis, ReductionOperation op);
+    void configure(const CLCompileContext &compile_context, const ICLTensor *input, 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: QASYMM8/QASYMM8_SIGNED/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.
+     * @param[in] input  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/S32/F16/F32.
+     * @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);
+    static Status validate(const ITensorInfo *input, 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;
diff --git a/src/runtime/CL/functions/CLArgMinMaxLayer.cpp b/src/runtime/CL/functions/CLArgMinMaxLayer.cpp
index 1b0a86a864..ea6311afdb 100644
--- a/src/runtime/CL/functions/CLArgMinMaxLayer.cpp
+++ b/src/runtime/CL/functions/CLArgMinMaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,7 +39,7 @@
 namespace arm_compute
 {
 CLArgMinMaxLayer::CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(), _reduction_kernels_vector(), _reshape(), _num_of_stages(), _reduction_axis()
+    : _memory_group(std::move(memory_manager)), _not_reshaped_output(), _arg_min_max_kernel(), _reshape(), _reduction_axis()
 {
 }
 
@@ -53,7 +53,6 @@ Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITen
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX && op != ReductionOperation::ARG_IDX_MIN, "Invalid reduction operation");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= static_cast<int>(TensorShape::num_max_dimensions), "Reduction axis greater than max number of dimensions");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
-    const unsigned int num_of_stages = utils::calculate_number_of_stages_only_x_axis(input->dimension(0), axis);
 
     DataType   output_data_type = DataType::S32;
     TensorInfo not_reshaped_output;
@@ -76,39 +75,7 @@ Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITen
 
     initialize_tensorinfo(not_reshaped_output, shape_before_reshape, output_data_type, input_num_channles, input_qinfo);
 
-    if(num_of_stages == 1)
-    {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, &not_reshaped_output, axis, op));
-    }
-    else
-    {
-        // Create temporary tensor infos
-        std::vector<TensorInfo> sums_vector(num_of_stages - 1);
-
-        // Create intermediate tensor info
-        TensorShape shape{ input->tensor_shape() };
-
-        for(unsigned int i = 0; i < num_of_stages - 1; i++)
-        {
-            shape.set(0, ceil(shape.x() / 128.f));
-            sums_vector[i].set_data_type(input->data_type());
-            sums_vector[i].set_tensor_shape(shape);
-            sums_vector[i].set_num_channels(input->num_channels());
-        }
-
-        // Validate ReductionOperation only on first kernel
-        ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, &sums_vector[0], axis, op));
-
-        // Validate ReductionOperation on intermediate stages
-        for(unsigned int i = 1; i < num_of_stages - 1; ++i)
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[i - 1], &sums_vector[i], axis, op));
-        }
-
-        // Validate ReductionOperation on the last stage
-        const unsigned int last_stage = num_of_stages - 1;
-        ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[last_stage - 1], &not_reshaped_output, axis, op));
-    }
+    ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &not_reshaped_output, axis, op));
     ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(&not_reshaped_output, output));
     return Status{};
 }
@@ -123,55 +90,16 @@ void CLArgMinMaxLayer::configure(const CLCompileContext &compile_context, const
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_LOG_PARAMS(input, axis, output, op);
 
-    _num_of_stages  = utils::calculate_number_of_stages_only_x_axis(input->info()->dimension(0), axis);
     _reduction_axis = axis;
 
     const TensorShape output_shape     = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false);
     DataType          output_data_type = (output->info()->data_type() == DataType::UNKNOWN) ? DataType::S32 : output->info()->data_type();
     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true));
 
-    // Configure reduction operation kernels
-    _reduction_kernels_vector.reserve(_num_of_stages);
-
-    auto add_reduction_kernel = [this, &compile_context, axis, op](const ICLTensor * input, const ICLTensor * prev_output, ICLTensor * output)
-    {
-        _reduction_kernels_vector.emplace_back(std::make_unique<CLArgMinMaxLayerKernel>());
-        _reduction_kernels_vector.back()->configure(compile_context, input, prev_output, output, axis, op);
-    };
+    _arg_min_max_kernel = std::make_unique<CLArgMinMaxLayerKernel>();
+    _arg_min_max_kernel->configure(compile_context, input, &_not_reshaped_output, axis, op);
 
     _memory_group.manage(&_not_reshaped_output);
-    // Create temporary tensors
-    if(_num_of_stages == 1)
-    {
-        add_reduction_kernel(input, nullptr, &_not_reshaped_output);
-    }
-    else
-    {
-        _results_vector.resize(_num_of_stages - 1);
-        TensorShape shape{ input->info()->tensor_shape() };
-        for(unsigned int i = 0; i < _num_of_stages - 1; i++)
-        {
-            shape.set(0, ceil(shape.x() / 128.f));
-            _results_vector[i].allocator()->init(input->info()->clone()->set_tensor_shape(shape).set_data_type(output_data_type));
-        }
-
-        // Apply ReductionOperation only on first kernel
-        _memory_group.manage(&_results_vector[0]);
-        add_reduction_kernel(input, nullptr, &_results_vector[0]);
-
-        // Apply ReductionOperation on intermediate stages
-        for(unsigned int i = 1; i < _num_of_stages - 1; ++i)
-        {
-            _memory_group.manage(&_results_vector[i]);
-            add_reduction_kernel(input, &_results_vector[i - 1], &_results_vector[i]);
-            _results_vector[i - 1].allocator()->allocate();
-        }
-
-        // Apply ReductionOperation on the last stage
-        const unsigned int last_stage = _num_of_stages - 1;
-        add_reduction_kernel(input, &_results_vector[last_stage - 1], &_not_reshaped_output);
-        _results_vector[last_stage - 1].allocator()->allocate();
-    }
     _reshape.configure(compile_context, &_not_reshaped_output, output);
     _not_reshaped_output.allocator()->allocate();
 }
@@ -180,10 +108,7 @@ void CLArgMinMaxLayer::run()
 {
     MemoryGroupResourceScope scope_mg(_memory_group);
 
-    for(unsigned int i = 0; i < _num_of_stages; ++i)
-    {
-        CLScheduler::get().enqueue(*_reduction_kernels_vector[i], false);
-    }
+    CLScheduler::get().enqueue(*_arg_min_max_kernel, false);
     _reshape.run();
 }
 } // namespace arm_compute