COMPMID-3704 Remove OpenCL padding: CLBatchConcatenateLayerKernel

COMPMID-3709 Remove OpenCL padding: CLDepthConcatenateLayerKernel

Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Change-Id: Iaea4fafd5d0f081fd5b45b0f6945302dc3365bd9
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4105
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

3 files changed, 14 insertions, 6 deletions

diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl
index 499378c87f..fca4e85255 100644
--- a/src/core/CL/cl_kernels/activation_layer.cl
+++ b/src/core/CL/cl_kernels/activation_layer.cl
@@ -33,6 +33,7 @@
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @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 Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH
  * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
  *
diff --git a/src/core/CL/cl_kernels/activation_layer_quant.cl b/src/core/CL/cl_kernels/activation_layer_quant.cl
index d8f56c093a..dbaefacc13 100644
--- a/src/core/CL/cl_kernels/activation_layer_quant.cl
+++ b/src/core/CL/cl_kernels/activation_layer_quant.cl
@@ -36,6 +36,7 @@
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @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 A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
  * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively.
  * @note Quantization offsets of the input/output tensors are passed in only if asymmetric with -DO1_VAL= and -DO2_VAL= respectively.
@@ -107,6 +108,7 @@ __kernel void activation_layer_quant_f32(
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @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 Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH
  * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
  * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively.
diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl
index 4281e675d7..0b211a6d1f 100644
--- a/src/core/CL/cl_kernels/concatenate.cl
+++ b/src/core/CL/cl_kernels/concatenate.cl
@@ -383,10 +383,13 @@ __kernel void concatenate_height(
 
 #endif /* defined(HEIGHT_OFFSET) && defined(DEPTH) */
 
+#if defined(VEC_SIZE_LEFTOVER)
+
 /** This kernel concatenates the input tensor into the output tensor along the third dimension
  *
  * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
  * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
+ * @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
  *
  * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: All
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
@@ -411,17 +414,19 @@ __kernel void concatenate(
     TENSOR3D_DECLARATION(dst),
     int offset)
 {
-    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
-    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+    uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0);
+
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x_offs + get_global_id(1) * src_stride_y + get_global_id(2) * src_stride_z;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_offs + get_global_id(1) * dst_stride_y + get_global_id(2) * dst_stride_z;
 
     VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
+    source_values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr);
 
 #if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
-    source_values = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
+    source_values0 = requantize(source_values0, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
 #endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
 
-    VSTORE(VEC_SIZE)
-    (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offset));
+    STORE_VECTOR_SELECT(source_values, DATA_TYPE, dst_addr + offset, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
+#endif /* defined(VEC_SIZE_LEFTOVER) */
 #endif /* defined(DATA_TYPE) && defined(VEC_SIZE) */