COMPMID-3715: Remove OpenCL padding - CLElementwiseOperationKernel

Change-Id: I8a685d4ac5de747a0f775bd10be9c411cf394953
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4140
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 38 insertions, 25 deletions

diff --git a/src/core/CL/cl_kernels/elementwise_operation_quantized.cl b/src/core/CL/cl_kernels/elementwise_operation_quantized.cl
index eb57da828d..9edfb4f9a2 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) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,25 +37,27 @@
 #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) && 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_OUT)
 
-#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
-#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
-#define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
+#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)
 
 /** This function executes an element-wise operation among two tensors.
  *
- * @attention The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
- * @attention The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10
- * @attention The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10
- * @attention The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10
- * @attention The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10
- * @attention The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10
- * @attention To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
- * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
- * @attention The element-wise operation to be executed has to be passed at compile time using -DOP (e.g., -DOP=ADD)
- * @attention For QSYMM16 operations OFFSET_IN1, OFFSET_IN2 and OFFSET_OUT must be set to zero
- * @attention The data type must be passed at compile time using -DDATA_TYPE_OUT, i.e. -DDATA_TYPE_OUT=uchar
+ * @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=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
+ * @note In case of broadcasting along the X dimension the proper preprocessor argument should be passed depending on the input (e.g. -DIS_IN1_X_BROADCASTING, -DIS_IN2_X_BROADCASTING)
+ * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
+ * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10
+ * @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10
+ * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10
+ * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10
+ * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10
+ * @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
  *
  * @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)
@@ -87,13 +89,25 @@ __kernel void OP_FUN_NAME(OP)(
     TENSOR3D_DECLARATION(in2),
     TENSOR3D_DECLARATION(out))
 {
+#if VEC_SIZE_IN1 == 1
+    uint in1_x_offs = 0;
+#else  // VEC_SIZE_IN1 == 1
+    uint in1_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN1 - (VEC_SIZE_IN1 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN1), 0);
+#endif // VEC_SIZE_IN1 == 1
+#if VEC_SIZE_IN2 == 1
+    uint in2_x_offs = 0;
+#else  // VEC_SIZE_IN2 == 1
+    uint in2_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN2 - (VEC_SIZE_IN2 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN2), 0);
+#endif // VEC_SIZE_IN2 == 1
+    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
-    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
-    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
-    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+    __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;
 
-    VEC_INT in_a = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_OUT *)in1.ptr), VEC_INT);
-    VEC_INT in_b = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_OUT *)in2.ptr), VEC_INT);
+    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_IN1)(0, (__global DATA_TYPE_OUT *)in2_addr)), VEC_INT);
 
     in_a = SUB(in_a, (VEC_INT)((int)OFFSET_IN1));
     in_b = SUB(in_b, (VEC_INT)((int)OFFSET_IN2));
@@ -101,10 +115,9 @@ __kernel void OP_FUN_NAME(OP)(
     const VEC_FLOAT in1f32  = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1);
     const VEC_FLOAT in2f32  = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2);
     const VEC_FLOAT qresf32 = OP(in1f32, in2f32) / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFFSET_OUT));
-    const VEC_TYPE  res     = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
+    const VEC_TYPE  res0    = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
 
     // Store result
-    VSTORE(VEC_SIZE)
-    (res, 0, (__global DATA_TYPE_OUT *)out.ptr);
+    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
-#endif /* 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_OUT) */