Remove OpenCL padding: CLL2NormalizeLayerKernel

Resolves: COMPMID-3909

Change-Id: I00a1705ed202002e2a6053702272181805fa6869
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5360
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 127 insertions, 102 deletions

diff --git a/src/core/CL/cl_kernels/l2_normalize.cl b/src/core/CL/cl_kernels/l2_normalize.cl
index 14b37e3257..fbe3406239 100644
--- a/src/core/CL/cl_kernels/l2_normalize.cl
+++ b/src/core/CL/cl_kernels/l2_normalize.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,142 +23,167 @@
  */
 #include "helpers.h"
 
+#if defined(VEC_SIZE_X) && defined(VEC_SIZE_LEFTOVER_X)
 /** This kernel performs l2 normalization on x-axis
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
- * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE_X=size. e.g. -DVEC_SIZE_X=16
+ * @note The leftover size in the X dimension shoud be given as preprocessor argument using -DVEC_SIZE_LEFTOVER_X is; x_dimension % VEC_SIZE_X. e.g. -DVEC_SIZE_LEFTOVER_X=1
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: 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 X 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]  sum_ptr                           Pointer to the source tensor. Supported data types: F16/F32
- * @param[in]  sum_stride_x                      Stride of the source tensor in X dimension (in bytes)
- * @param[in]  sum_step_x                        sum_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  sum_stride_y                      Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  sum_step_y                        sum_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  sum_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- * @param[in]  epsilon                           Epsilon value
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: 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 X 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]  sum_ptr                              Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  sum_stride_x                         Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  sum_step_x                           sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_stride_y                         Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  sum_step_y                           sum_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_offset_first_element_in_bytes    The offset of the first element in the source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination 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 destination 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 destination tensor
+ * @param[in]  epsilon                              Epsilon value
  */
 __kernel void l2_normalize_x(
-    IMAGE_DECLARATION(src),
+    IMAGE_DECLARATION(input),
     IMAGE_DECLARATION(sum),
-    IMAGE_DECLARATION(dst),
+    IMAGE_DECLARATION(output),
     DATA_TYPE epsilon)
 {
-    Image src = CONVERT_TO_IMAGE_STRUCT(src);
-    Image sum = CONVERT_TO_IMAGE_STRUCT(sum);
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+    // Offset computation
+    const uint x_offs = max((int)(get_global_id(0) * VEC_SIZE_X - (VEC_SIZE_X - VEC_SIZE_LEFTOVER_X) % VEC_SIZE_X), 0);
 
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    in = vload16(0, (__global DATA_TYPE *)src.ptr);
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    normalize_value = (VEC_DATA_TYPE(DATA_TYPE, 16))rsqrt(fmax(((__global DATA_TYPE *)sum.ptr)[0], epsilon));
+    // Address computation
+    __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 *sum_addr    = sum_ptr + sum_offset_first_element_in_bytes + get_global_id(1) * sum_stride_y;
+    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * output_stride_y;
 
-    vstore16(in * normalize_value, 0, (__global DATA_TYPE *)dst.ptr);
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    in = VLOAD(VEC_SIZE_X)(0, (__global DATA_TYPE *)input_addr);
+
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    normalize_value = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X))rsqrt(fmax(*((__global DATA_TYPE *)sum_addr), epsilon));
+
+    const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    data0 = in * normalize_value;
+
+    STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE_X, VEC_SIZE_LEFTOVER_X, VEC_SIZE_LEFTOVER_X != 0 && get_global_id(0) == 0);
 }
 
 /** This kernel performs l2 normalization on y-axis.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
- * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE_X=size. e.g. -DVEC_SIZE_X=16
+ * @note The leftover size in the X dimension shoud be given as preprocessor argument using -DVEC_SIZE_LEFTOVER_X is; x_dimension % VEC_SIZE_X. e.g. -DVEC_SIZE_LEFTOVER_X=1
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: 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 Y 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 X 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]  sum_ptr                           Pointer to the source tensor. Supported data types: F16/F32
- * @param[in]  sum_stride_x                      Stride of the source tensor in X dimension (in bytes)
- * @param[in]  sum_step_x                        sum_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  sum_stride_y                      Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  sum_step_y                        sum_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  sum_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- * @param[in]  epsilon                           Epsilon value
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: 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 Y 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]  sum_ptr                              Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  sum_stride_x                         Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  sum_step_x                           sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_stride_y                         Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  sum_step_y                           sum_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_offset_first_element_in_bytes    The offset of the first element in the source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination 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 destination 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 destination tensor
+ * @param[in]  epsilon                              Epsilon value
  */
 __kernel void l2_normalize_y(
-    IMAGE_DECLARATION(src),
+    IMAGE_DECLARATION(input),
     IMAGE_DECLARATION(sum),
-    IMAGE_DECLARATION(dst),
+    IMAGE_DECLARATION(output),
     DATA_TYPE epsilon)
 {
-    Image src = CONVERT_TO_IMAGE_STRUCT(src);
-    Image sum = CONVERT_TO_IMAGE_STRUCT(sum);
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+    // Offset computation
+    const uint x_offs = max((int)(get_global_id(0) * VEC_SIZE_X - (VEC_SIZE_X - VEC_SIZE_LEFTOVER_X) % VEC_SIZE_X), 0);
+
+    // Address computation
+    __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 *sum_addr    = sum_ptr + sum_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE);
+    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * output_stride_y;
 
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    in = vload16(0, (__global DATA_TYPE *)src.ptr);
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    sums = vload16(0, (__global DATA_TYPE *)sum.ptr);
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    in = VLOAD(VEC_SIZE_X)(0, (__global DATA_TYPE *)input_addr);
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    sums = VLOAD(VEC_SIZE_X)(0, (__global DATA_TYPE *)sum_addr);
 
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    normalize_value = (VEC_DATA_TYPE(DATA_TYPE, 16))rsqrt(fmax(sums, epsilon));
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    normalize_value = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X))rsqrt(fmax(sums, epsilon));
 
-    vstore16(in * normalize_value, 0, (__global DATA_TYPE *)dst.ptr);
+    const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    data0 = in * normalize_value;
+
+    STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE_X, VEC_SIZE_LEFTOVER_X, VEC_SIZE_LEFTOVER_X != 0 && get_global_id(0) == 0);
 }
+
 /** This kernel performs l2 normalization on z-axis.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
- * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE_X=size. e.g. -DVEC_SIZE_X=16
+ * @note The leftover size in the X dimension shoud be given as preprocessor argument using -DVEC_SIZE_LEFTOVER_X is; x_dimension % VEC_SIZE_X. e.g. -DVEC_SIZE_LEFTOVER_X=1
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: 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 Y 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 X processed per workitem(in bytes)
- * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  src_step_z                        src_stride_z * number of elements along Z 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]  sum_ptr                           Pointer to the source tensor. Supported data types: F16/F32
- * @param[in]  sum_stride_x                      Stride of the source tensor in X dimension (in bytes)
- * @param[in]  sum_step_x                        sum_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  sum_stride_y                      Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  sum_step_y                        sum_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  sum_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  sum_step_z                        sum_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  sum_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  dst_step_z                        dst_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- * @param[in]  epsilon                           Epsilon value
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: 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 Y 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_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z 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]  sum_ptr                              Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  sum_stride_x                         Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  sum_step_x                           sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_stride_y                         Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  sum_step_y                           sum_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_stride_z                         Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  sum_step_z                           sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  sum_offset_first_element_in_bytes    The offset of the first element in the source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination 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 destination 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_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in]  epsilon                              Epsilon value
  */
 __kernel void l2_normalize_z(
-    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(input),
     TENSOR3D_DECLARATION(sum),
-    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(output),
     DATA_TYPE epsilon)
 {
-    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
-    Tensor3D sum = CONVERT_TO_TENSOR3D_STRUCT(sum);
-    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+    // Offset computation
+    const uint x_offs = max((int)(get_global_id(0) * VEC_SIZE_X - (VEC_SIZE_X - VEC_SIZE_LEFTOVER_X) % VEC_SIZE_X), 0);
 
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    in = vload16(0, (__global DATA_TYPE *)src.ptr);
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    sums = vload16(0, (__global DATA_TYPE *)sum.ptr);
+    // Address computation
+    __global uchar *input_addr  = input_ptr + input_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
+    __global uchar *sum_addr    = sum_ptr + sum_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * sum_stride_y;
+    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
 
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    normalize_value = (VEC_DATA_TYPE(DATA_TYPE, 16))rsqrt(fmax(sums, epsilon));
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    in = VLOAD(VEC_SIZE_X)(0, (__global DATA_TYPE *)input_addr);
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    sums = VLOAD(VEC_SIZE_X)(0, (__global DATA_TYPE *)sum_addr);
 
-    vstore16(in * normalize_value, 0, (__global DATA_TYPE *)dst.ptr);
-}
\ No newline at end of file
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X)
+    data0 = in * ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_X))(rsqrt(fmax(sums, epsilon))));
+
+    STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE_X, VEC_SIZE_LEFTOVER_X, VEC_SIZE_LEFTOVER_X != 0 && get_global_id(0) == 0);
+}
+#endif // defined(VEC_SIZE_X) && defined(VEC_SIZE_LEFTOVER_X)
\ No newline at end of file