Remove redundant implementations of Add/Sub operators

Allows only implementations where inputs/output are of the same data
type and removes legacy Computer Vision ones.

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: Ia2b3d23a04236aab682f0c36a1110a30f7c06d1c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5900
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 29 insertions, 30 deletions

diff --git a/src/core/CL/cl_kernels/elementwise_operation.cl b/src/core/CL/cl_kernels/elementwise_operation.cl
index c8250045dc..99f725645d 100644
--- a/src/core/CL/cl_kernels/elementwise_operation.cl
+++ b/src/core/CL/cl_kernels/elementwise_operation.cl
@@ -23,7 +23,7 @@
  */
 #include "helpers.h"
 
-#if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_OUT)
+#if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(DATA_TYPE)
 
 /** List of all the operations supported by this kernel.
  * @note ADD and SUB operations, when executed on integers, support saturation */
@@ -43,17 +43,17 @@
 #if VEC_SIZE_OUT == 1
 #define PRELU(x, y) (x > 0 ? x : x * y)
 #else // VEC_SIZE_OUT == 1
-#define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE_OUT)0), SELECT_VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))))
+#define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE)0), SELECT_VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))))
 #endif // VEC_SIZE_OUT == 1
 
 #if defined(S32)
-#define DIV(x, y) CONVERT(floor(CONVERT(x, VEC_DATA_TYPE(float, VEC_SIZE_OUT)) / CONVERT(y, VEC_DATA_TYPE(float, VEC_SIZE_OUT))), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT));
+#define DIV(x, y) CONVERT(floor(CONVERT(x, VEC_DATA_TYPE(float, VEC_SIZE_OUT)) / CONVERT(y, VEC_DATA_TYPE(float, VEC_SIZE_OUT))), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT));
 #else /* S32 */
 #define DIV(x, y) (x / y)
 #endif /* S32 */
 
-#define AND(x, y) (CONVERT((x && y), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))1))
-#define OR(x, y) (CONVERT((x || y), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))1))
+#define AND(x, y) (CONVERT((x && y), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))1))
+#define OR(x, y) (CONVERT((x || y), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))1))
 
 #define OP_FUN_NAME_STR(op) elementwise_operation_##op
 #define OP_FUN_NAME(op) OP_FUN_NAME_STR(op)
@@ -66,8 +66,7 @@
  *
  * @note Vector sizes of inputs and output have to be passed at compile time using -DVEC_SIZE_IN1, -DVEC_SIZE_IN2, -DVEC_SIZE_OUT.
  * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_OUT=3. It is defined as the remainder between the input's first dimension and VEC_SIZE_OUT
- * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT:
- * e.g. -DDATA_TYPE_IN1=uchar -DDATA_TYPE_IN2=uchar -DDATA_TYPE_OUT=short
+ * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=uchar
  * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
  * @note The element-wise operation to be executed has to be passed at compile time using -DOP (e.g., -DOP=ADD)
  *
@@ -114,23 +113,23 @@ __kernel void OP_FUN_NAME(OP)(
     uint out_x_offs = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
 
     // Get pixels pointer
-    __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs * sizeof(DATA_TYPE_IN1) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z;
-    __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs * sizeof(DATA_TYPE_IN2) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z;
-    __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs * sizeof(DATA_TYPE_OUT) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z;
+    __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z;
+    __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z;
+    __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z;
 
     // Load values
-    VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
-    in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE_IN1, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE_IN1 *)in1_addr)), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT));
-    VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
-    in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE_IN2, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE_IN2 *)in2_addr)), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT));
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
+    in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE *)in1_addr)), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT));
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
+    in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE *)in2_addr)), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT));
 
     // Calculate and store result
-    VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
     res0 = OP(in_a, in_b);
 #if defined(ACTIVATION_TYPE)
-    res0 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE_OUT, VEC_SIZE_OUT, res0, A_VAL, B_VAL);
+    res0 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE_OUT, res0, A_VAL, B_VAL);
 #endif // defined(ACTIVATION_TYPE)
 
-    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
+    STORE_VECTOR_SELECT(res, DATA_TYPE, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
-#endif /* defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_OUT) */
+#endif /* defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(DATA_TYPE) */
diff --git a/src/core/CL/cl_kernels/elementwise_operation_quantized.cl b/src/core/CL/cl_kernels/elementwise_operation_quantized.cl
index a08c3b2d47..0051babf03 100644
--- a/src/core/CL/cl_kernels/elementwise_operation_quantized.cl
+++ b/src/core/CL/cl_kernels/elementwise_operation_quantized.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -28,7 +28,7 @@
 #define MAX(x, y) max((x), (y))
 #define MIN(x, y) min((x), (y))
 #define SQUARED_DIFF(x, y) (x - y) * (x - y)
-#define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE_OUT)0), SELECT_VEC_DATA_TYPE(float, VEC_SIZE_OUT))))
+#define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE)0), SELECT_VEC_DATA_TYPE(float, VEC_SIZE_OUT))))
 #define DIV(x, y) (x / y)
 
 #define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
@@ -37,11 +37,11 @@
 #define OP_FUN_NAME_STR(op) elementwise_operation_##op##_quantized
 #define OP_FUN_NAME(op) OP_FUN_NAME_STR(op)
 
-#if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT)
+#if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE)
 
 #define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE_OUT)
 #define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE_OUT)
-#define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
+#define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
 
 /** This function executes an element-wise operation among two tensors.
  *
@@ -57,7 +57,7 @@
  * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
  * @note The element-wise operation to be executed has to be passed at compile time using -DOP (e.g., -DOP=ADD)
  * @note For QSYMM16 operations OFFSET_IN1, OFFSET_IN2 and OFFSET_OUT must be set to zero
- * @note The data type must be passed at compile time using -DDATA_TYPE_OUT, i.e. -DDATA_TYPE_OUT=uchar
+ * @note The data type must be passed at compile time using -DDATA_TYPE, i.e. -DDATA_TYPE=uchar
  *
  * @param[in]  in1_ptr                           Pointer to the source tensor. Supported data types: QASYMM8/QSYMM16
  * @param[in]  in1_stride_x                      Stride of the source tensor in X dimension (in bytes)
@@ -102,12 +102,12 @@ __kernel void OP_FUN_NAME(OP)(
     uint out_x_offs = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
 
     // Get pixels pointer
-    __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs * sizeof(DATA_TYPE_OUT) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z;
-    __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs * sizeof(DATA_TYPE_OUT) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z;
-    __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs * sizeof(DATA_TYPE_OUT) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z;
+    __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z;
+    __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z;
+    __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z;
 
-    VEC_INT in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE_OUT *)in1_addr)), VEC_INT);
-    VEC_INT in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE_OUT *)in2_addr)), VEC_INT);
+    VEC_INT in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE *)in1_addr)), VEC_INT);
+    VEC_INT in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE *)in2_addr)), VEC_INT);
 
     in_a = SUB(in_a, (VEC_INT)((int)OFFSET_IN1));
     in_b = SUB(in_b, (VEC_INT)((int)OFFSET_IN2));
@@ -118,6 +118,6 @@ __kernel void OP_FUN_NAME(OP)(
     const VEC_TYPE  res0    = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
 
     // Store result
-    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
+    STORE_VECTOR_SELECT(res, DATA_TYPE, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
-#endif /* defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) */
+#endif /* defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE) */