COMPMID-3703 Remove OpenCL padding: CLActivationLayerKernel + create utility macro

Change-Id: I73edadc7299247e7bc51ae37c00d3709023da44a
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4073
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 410 insertions, 354 deletions

diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl
index f846cb2764..499378c87f 100644
--- a/src/core/CL/cl_kernels/activation_layer.cl
+++ b/src/core/CL/cl_kernels/activation_layer.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 Arm Limited.
+ * Copyright (c) 2016-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -61,23 +61,24 @@ __kernel void activation_layer(
 #endif /* not IN_PLACE */
 )
 {
+    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);
+
     // Get pixels pointer
-    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+    __global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
 #ifdef IN_PLACE
-    Tensor3D output = input;
+    __global uchar *output_addr = input_addr;
 #else  /* IN_PLACE */
-    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
 #endif /* IN_PLACE */
 
     // Load data
-    TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
+    TYPE data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr);
 
     // Perform activation
-    data = ACTIVATION(ACT, DATA_TYPE, data, A_VAL, B_VAL);
+    data0 = ACTIVATION(ACT, DATA_TYPE, data0, A_VAL, B_VAL);
 
     // Store result
-    VSTORE(VEC_SIZE)
-    (data, 0, (__global DATA_TYPE *)output.ptr);
+    STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
 
 #endif /* defined(ACT) */
diff --git a/src/core/CL/cl_kernels/activation_layer_quant.cl b/src/core/CL/cl_kernels/activation_layer_quant.cl
index 0481319428..d8f56c093a 100644
--- a/src/core/CL/cl_kernels/activation_layer_quant.cl
+++ b/src/core/CL/cl_kernels/activation_layer_quant.cl
@@ -66,34 +66,35 @@ __kernel void activation_layer_quant_f32(
 #endif /* not IN_PLACE */
 )
 {
+    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);
+
     // Get pixels pointer
-    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+    __global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
 #ifdef IN_PLACE
-    Tensor3D output = input;
+    __global uchar *output_addr = input_addr;
 #else  /* IN_PLACE */
-    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
 #endif /* IN_PLACE */
 
     // Load data
-    TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
+    TYPE data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr);
 
-    VEC_FLOAT data_flt = CONVERT(data, VEC_FLOAT);
+    VEC_FLOAT data_flt = CONVERT(data0, VEC_FLOAT);
 #if defined(O1_VAL)
     data_flt = round(data_flt - (float)O1_VAL) * ((float)S1_VAL);
 #else  // defined(O1_VAL)
-    data_flt        = round(data_flt) * ((float)S1_VAL);
+    data_flt                    = round(data_flt) * ((float)S1_VAL);
 #endif // defined(O1_VAL)
     data_flt = ACTIVATION(ACT, float, data_flt, A_VAL, B_VAL);
 
 #if defined(O2_VAL)
-    data = CONVERT_SAT(round(data_flt / ((float)S2_VAL)) + (float)O2_VAL, TYPE);
+    data0 = CONVERT_SAT(round(data_flt / ((float)S2_VAL)) + (float)O2_VAL, TYPE);
 #else  // defined(O2_VAL)
-    data            = CONVERT_SAT(round(data_flt / ((float)S2_VAL)), TYPE);
+    data0                       = CONVERT_SAT(round(data_flt / ((float)S2_VAL)), TYPE);
 #endif // defined(O2_VAL)
 
     // Store result
-    VSTORE(VEC_SIZE)
-    (data, 0, (__global DATA_TYPE *)output.ptr);
+    STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
 
 #else // defined(FLOAT_DOMAIN)
@@ -137,22 +138,23 @@ __kernel void activation_layer_quant(
 #endif /* not IN_PLACE */
 )
 {
+    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);
+
     // Get pixels pointer
-    Tensor3D input  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    __global uchar *input_addr  = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
 #ifdef IN_PLACE
-    Tensor3D output = input;
+    __global uchar *output_addr = input_addr;
 #else  /* IN_PLACE */
-    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
 #endif /* IN_PLACE */
 
     // Load data
-    TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
+    TYPE data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr);
 
-    data = PERFORM_ACTIVATION_QUANT(ACT, data);
+    data0 = PERFORM_ACTIVATION_QUANT(ACT, data0);
 
     // Store result
-    VSTORE(VEC_SIZE)
-    (data, 0, (__global DATA_TYPE *)output.ptr);
+    STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
 }
 #endif // defined(ACT)
 #endif // defined(FLOAT_DOMAIN)
diff --git a/src/core/CL/cl_kernels/gemm_helpers.h b/src/core/CL/cl_kernels/gemm_helpers.h
index 6f6edc1bcf..3943a8de78 100644
--- a/src/core/CL/cl_kernels/gemm_helpers.h
+++ b/src/core/CL/cl_kernels/gemm_helpers.h
@@ -799,101 +799,6 @@
     (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @} */ // end of groupd STORE_ROW_n
 
-/** Partially store the 0 to (n-1)th rows of the given variables
- * @name STORE_ROW_PARTIAL_n
- * Within each row, store the lower @p STORE_N0 elements of vectors of width @p N0
- *
- * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
- *
- * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
- * @param[in] STORE_N0  The **lower** size of the vectors to store. Supported: [1-16 and <= @p N0
- * @param[in] DATA_TYPE The data type of the vectors
- * @param[in] BASENAME  The basename of the variables
- * @param[in] PTR       The base pointer
- * @param[in] STRIDE_Y  The stride value in y-axis direction
- * @param[in] Z         The offset in z-axis direction
- * @{
- */
-#define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));
-
-#define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));
-
-#define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));
-
-#define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));
-
-#define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));
-
-#define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));
-
-#define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));
-
-#define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));
-
-#define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
-    (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));
-
-#define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
-    (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));
-
-#define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
-    (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));
-
-#define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
-    (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));
-
-#define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
-    (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));
-
-#define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
-    (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));
-
-#define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
-    (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));
-
-#define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
-    STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
-    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
-    (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
-/** @} */ // end of groupd STORE_ROW_PARTIAL_n
-
 /** Convert and store the 0th to (n-1)th rows of the given variables
  * @name CONVERT_STORE_ROW_n
  *
@@ -1008,127 +913,6 @@
 #define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** @} */ // end of group STORE_BLOCK
 
-/** Partially store a block of the given size STORE_M0xSTORE_N0
- * @name STORE_BLOCK_PARTIAL
- *
- * @note The vector width @p N0 is also required for correct partial storing behaviour.
- * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
- *
- * The data to store is expected to have consecutive names for each row.
- * E.g., for STORE_M0=3 and basename=c, the expected names are c0, c1 and c2.
- * The Z offset is expected to have consecutive names.
- * E.g., for STORE_M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
- *
- * @param[in] STORE_M0  The number of rows to store. Supported: 1-16
- * @param[in] STORE_N0  The lower number of elements of vectors to store. Supported: 1-16 and <= @p N0
- * @param[in] N0        The size of each vector. Supported: 1, 2, 3, 4, 8, 16
- * @param[in] DATA_TYPE The data type of the vectors
- * @param[in] BASENAME  The basename of the variables
- * @param[in] PTR       The base pointer
- * @param[in] STRIDE_Y  The stride value in y-axis direction
- * @param[in] Z         The offset in z-axis direction
- * @{
- */
-#define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
-#define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
-/** Store a block that can be partial in both x and y dimensions
- *
- * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
- *
- * The data to store is expected to have consecutive names for each row.
- * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
- * The Z offset is expected to have consecutive names.
- * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
- *
- * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
- * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
- * @param[in] DATA_TYPE        The data type of the vectors
- * @param[in] BASENAME         The basename of the variables
- * @param[in] PTR              The base pointer
- * @param[in] STRIDE_Y         The stride value in y-axis direction
- * @param[in] Z                The offset in z-axis direction
- * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
- * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
- * @param[in] N                Total number of columns. Used to detect if current block is at the boundary in x.
- * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
- * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
- */
-#define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
-    if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y))                                                                                                               \
-    {                                                                                                                                        \
-        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                              \
-    }                                                                                                                                        \
-    else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X))                                                                                                           \
-    {                                                                                                                                        \
-        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                \
-    }                                                                                                                                        \
-    else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X))                                                                                                           \
-    {                                                                                                                                        \
-        STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                \
-    }                                                                                                                                        \
-    else                                                                                                                                     \
-    {                                                                                                                                        \
-        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                  \
-    }
-/** Store a block that can only be partial in x but not y.
- *
- * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
- *
- * The data to store is expected to have consecutive names for each row.
- * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
- * The Z offset is expected to have consecutive names.
- * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
- *
- * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
- * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
- * @param[in] DATA_TYPE        The data type of the vectors
- * @param[in] BASENAME         The basename of the variables
- * @param[in] PTR              The base pointer
- * @param[in] STRIDE_Y         The stride value in y-axis direction
- * @param[in] Z                The offset in z-axis direction
- * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
- * @param[in] N                Total number of columns. Used to detect if current block is at the boundary in x.
- * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
- */
-#define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, N, PARTIAL_COND_X) \
-    if(!(PARTIAL_COND_X))                                                                                            \
-    {                                                                                                        \
-        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                              \
-    }                                                                                                        \
-    else                                                                                                     \
-    {                                                                                                        \
-        STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                \
-    }
-/** Store a block that can only be partial in y but not x.
- *
- * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
- *
- * The data to store is expected to have consecutive names for each row.
- * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
- * The Z offset is expected to have consecutive names.
- * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
- *
- * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
- * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
- * @param[in] DATA_TYPE        The data type of the vectors
- * @param[in] BASENAME         The basename of the variables
- * @param[in] PTR              The base pointer
- * @param[in] STRIDE_Y         The stride value in y-axis direction
- * @param[in] Z                The offset in z-axis direction
- * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
- * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
- */
-#define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \
-    if(!(PARTIAL_COND_Y))                                                                                         \
-    {                                                                                                     \
-        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                           \
-    }                                                                                                     \
-    else                                                                                                  \
-    {                                                                                                     \
-        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);             \
-    }
-/** @} */ // end of group STORE_BLOCK_PARTIAL
-
 /** Convert and store a block of the given size M0xN0
  * @name CONVERT_STORE_BLOCK
  *
@@ -1732,113 +1516,4 @@
  */
 #define CONVERT_BLOCK_STR(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) CONVERT_ROW_##M(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)
 #define CONVERT_BLOCK(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) CONVERT_BLOCK_STR(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)
-/** @} */ // end of group CONVERT_BLOCK
-
-#if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
-
-/** Boundary-aware GEMM block store
- * @name STORE_BLOCK_BOUNDARY_AWARE
- * This macro assumes the following schemes to achieve boundary-awareness:
- *  - Overlapping load in Y axis from lhs tensor. This implies lhs has no padding along y dim.
- *  - Non-Overlapping(normal) load from rhs tensor. This imples rhs can have paddings.
- *  - Overlapping load in Y axis from bias tensor. This implies rhs has no padding along y dim.
- * The macro then ensures that the dst tensor can be stored without any paddings in both x and y dim.
- *
- * In the y dimension, we place the partial blocks **at the beginning** while in the x dimension, we place the partial
- * blocks **at the end**.
- * Say, the dst tensor is of shape MxN and we have M0 and N0 as the block size, this is how we define "partial blocks"/
- * "boundary block" (we use the 2 terms "partial blocks" and "boundary blocks" interchangeably) and its various parameters:
- *
- *  *--x-->                         x == 0                        x == 1
- *  |                  |<------------------------------N-------------------------->|
- *  y                  |<--------------N0------------->|<----PARTIAL_STORE_N0----->|
- *  |     -------------#############################################################
- *  *     |          | |...............................|...........................|
- * y == 0 | PAR_..._M0 |......Boundary block in y......|.Boundary block in x and y.|
- *        |          | |...............................|...........................|
- *        M          --#############################################################
- *        |          | |                               |...........................|
- * y == 1 |         M0 |      Non-boundary block       |....Boundary block in x....|
- *        |          | |                               |...........................|
- *        |------------#############################################################
- *
- * Then @p PARTIAL_STORE_M0 = M % M0      and @p PARTIAL_STORE_N0 = N % N0
- *
- * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
- *
- * It automatically detects if a giving M,N,M0,N0 combination can yield partial blocks in either X and Y dimension,
- * and select corresponding store methods such that the boundary detection logic is only added when needed.
- *
- * The data to store is expected to have consecutive names for each row.
- * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
- * The Z offset is expected to have consecutive names.
- * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
- *
- * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
- * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
- * @param[in] DATA_TYPE        The data type of the vectors
- * @param[in] BASENAME         The basename of the variables
- * @param[in] PTR              The base pointer
- * @param[in] STRIDE_Y         The stride value in y-axis direction
- * @param[in] Z                The offset in z-axis direction
- * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
- * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported: [0, @p N0)
- * @param[in] N                Total number of columns. Used to detect if current block is at the boundary in x.
- * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
- * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
- * @{
- */
-#if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
-// Case1: No partial blocks in either x or y
-#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
-    STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
-
-#elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0
-// Case2: Partial blocks in y
-#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
-    STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y)
-
-#elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0
-// Case3: Partial blocks in x
-#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
-    STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, N, PARTIAL_COND_X)
-
-#else // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
-// Case4: Partial blocks in both x and y
-#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
-    STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X)
-
-#endif // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
-
-#else // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
-
-#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
-    STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
-
-#endif    // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
-/** @} */ // end of group STORE_BLOCK_BOUNDARY_AWARE
-
-#if defined(PARTIAL_STORE_M0)
-/** Compute the start m0 row (LHS, BIAS and DST) in a boundary-aware way so as to avoid padding
- * @name COMPUTE_M0_START_ROW
- * If there're any partial blocks in y dimension, they are placed at the beginning of the rows.
- * This shift amount is added to all rows such that the partial block (at the beginning) overlaps with the subsequent
- * blocks in the y dimension to avoid any padding.
- * EG: M0=4, PARTIAL_STORE_M0=1:
- *                  | Non-overlapping | +M0_ROW_SHIFT (Overlapping)
- * block 0 (partial)| start row = 0   | start row = 0
- * block 1 (full)   | start row = 4   | start row = 1
- * block 2 (full)   | start row = 8   | start row = 5
- *
- * @param[in] y                Global id of current block in y.
- * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
- * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
- * @{
- */
-#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
-    ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0))))
-#else // defined(PARTIAL_STORE_M0)
-#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
-    ((uint)(y * M0))
-#endif    // defined(PARTIAL_STORE_M0)
-/** @} */ // end of group COMPUTE_M0_START_ROW
+/** @} */ // end of group CONVERT_BLOCK
\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h
index 7b08233029..10f04e9e31 100644
--- a/src/core/CL/cl_kernels/helpers.h
+++ b/src/core/CL/cl_kernels/helpers.h
@@ -24,6 +24,8 @@
 #ifndef ARM_COMPUTE_HELPER_H
 #define ARM_COMPUTE_HELPER_H
 
+#include "load_store_utility.h"
+
 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
@@ -273,21 +275,30 @@
 #define VSTORE_PARTIAL_STR(size, store_size) vstore_partial_##size##_##store_size
 #define VSTORE_PARTIAL(size, store_size) VSTORE_PARTIAL_STR(size, store_size)
 
+#define NO_STORE(data, offs, ptr) \
+    {                             \
+    }
+
 // Size == 1 (scalar)
+#define vstore_partial_1_0 NO_STORE
 #define vstore_partial_1_1 vstore1
 // Size == 2
+#define vstore_partial_2_0 NO_STORE
 #define vstore_partial_2_1 vstore_partial_1
 #define vstore_partial_2_2 vstore_partial_2
 // Size == 3
+#define vstore_partial_3_0 NO_STORE
 #define vstore_partial_3_1 vstore_partial_1
 #define vstore_partial_3_2 vstore_partial_2
 #define vstore_partial_3_3 vstore_partial_3
 // Size == 4
+#define vstore_partial_4_0 NO_STORE
 #define vstore_partial_4_1 vstore_partial_1
 #define vstore_partial_4_2 vstore_partial_2
 #define vstore_partial_4_3 vstore_partial_3
 #define vstore_partial_4_4 vstore_partial_4
 // Size == 8
+#define vstore_partial_8_0 NO_STORE
 #define vstore_partial_8_1 vstore_partial_1
 #define vstore_partial_8_2 vstore_partial_2
 #define vstore_partial_8_3 vstore_partial_3
@@ -297,6 +308,7 @@
 #define vstore_partial_8_7 vstore_partial_7
 #define vstore_partial_8_8 vstore_partial_8
 // Size == 16
+#define vstore_partial_16_0 NO_STORE
 #define vstore_partial_16_1 vstore_partial_1
 #define vstore_partial_16_2 vstore_partial_2
 #define vstore_partial_16_3 vstore_partial_3
@@ -376,15 +388,15 @@
 
 #define vstore_partial_13(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
-    vstore_partial_5(DATA.s89abc, OFFSET, PTR + 8);
+    vstore_partial_5(DATA.s89abcdef, OFFSET, PTR + 8);
 
 #define vstore_partial_14(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
-    vstore_partial_6(DATA.s89abcd, OFFSET, PTR + 8);
+    vstore_partial_6(DATA.s89abcdef, OFFSET, PTR + 8);
 
 #define vstore_partial_15(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
-    vstore_partial_7(DATA.s89abcde, OFFSET, PTR + 8);
+    vstore_partial_7(DATA.s89abcdef, OFFSET, PTR + 8);
 
 #define vstore_partial_16(DATA, OFFSET, PTR) \
     vstore16(DATA, OFFSET, PTR);
diff --git a/src/core/CL/cl_kernels/load_store_utility.h b/src/core/CL/cl_kernels/load_store_utility.h
new file mode 100644
index 0000000000..cb833d016b
--- /dev/null
+++ b/src/core/CL/cl_kernels/load_store_utility.h
@@ -0,0 +1,366 @@
+/*
+ * Copyright (c) 2020 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+/** Partially store the 0 to (n-1)th rows of the given variables
+ * @name STORE_ROW_PARTIAL_n
+ * Within each row, store the lower @p STORE_N0 elements of vectors of width @p N0
+ *
+ * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
+ *
+ * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
+ * @param[in] STORE_N0  The **lower** size of the vectors to store. Supported: [1-16 and <= @p N0
+ * @param[in] DATA_TYPE The data type of the vectors
+ * @param[in] BASENAME  The basename of the variables
+ * @param[in] PTR       The base pointer
+ * @param[in] STRIDE_Y  The stride value in y-axis direction
+ * @param[in] Z         The offset in z-axis direction
+ * @{
+ */
+#define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));
+
+#define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));
+
+#define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));
+
+#define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));
+
+#define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));
+
+#define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));
+
+#define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));
+
+#define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));
+
+#define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                 \
+    (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));
+
+#define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
+    (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));
+
+#define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
+    (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));
+
+#define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
+    (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));
+
+#define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
+    (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));
+
+#define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
+    (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));
+
+#define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
+    (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));
+
+#define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
+    STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
+    VSTORE_PARTIAL(N0, STORE_N0)                                                  \
+    (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
+/** @} */ // end of groupd STORE_ROW_PARTIAL_n
+
+/** Partially store a block of the given size STORE_M0xSTORE_N0
+ * @name STORE_BLOCK_PARTIAL
+ *
+ * @note The vector width @p N0 is also required for correct partial storing behaviour.
+ * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
+ *
+ * The data to store is expected to have consecutive names for each row.
+ * E.g., for STORE_M0=3 and basename=c, the expected names are c0, c1 and c2.
+ * The Z offset is expected to have consecutive names.
+ * E.g., for STORE_M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
+ *
+ * @param[in] STORE_M0  The number of rows to store. Supported: 1-16
+ * @param[in] STORE_N0  The lower number of elements of vectors to store. Supported: 1-16 and <= @p N0
+ * @param[in] N0        The size of each vector. Supported: 1, 2, 3, 4, 8, 16
+ * @param[in] DATA_TYPE The data type of the vectors
+ * @param[in] BASENAME  The basename of the variables
+ * @param[in] PTR       The base pointer
+ * @param[in] STRIDE_Y  The stride value in y-axis direction
+ * @param[in] Z         The offset in z-axis direction
+ * @{
+ */
+#define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
+#define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
+/** Store a block that can be partial in both x and y dimensions
+ *
+ * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
+ *
+ * The data to store is expected to have consecutive names for each row.
+ * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
+ * The Z offset is expected to have consecutive names.
+ * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
+ *
+ * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
+ * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
+ * @param[in] DATA_TYPE        The data type of the vectors
+ * @param[in] BASENAME         The basename of the variables
+ * @param[in] PTR              The base pointer
+ * @param[in] STRIDE_Y         The stride value in y-axis direction
+ * @param[in] Z                The offset in z-axis direction
+ * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
+ * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
+ * @param[in] N                Total number of columns. Used to detect if current block is at the boundary in x.
+ * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
+ * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
+ */
+#define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
+    if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y))                                                                                                               \
+    {                                                                                                                                                        \
+        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                                              \
+    }                                                                                                                                                        \
+    else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X))                                                                                                           \
+    {                                                                                                                                                        \
+        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                                \
+    }                                                                                                                                                        \
+    else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X))                                                                                                           \
+    {                                                                                                                                                        \
+        STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                                \
+    }                                                                                                                                                        \
+    else                                                                                                                                                     \
+    {                                                                                                                                                        \
+        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                  \
+    }
+/** Store a block that can only be partial in x but not y.
+ *
+ * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
+ *
+ * The data to store is expected to have consecutive names for each row.
+ * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
+ * The Z offset is expected to have consecutive names.
+ * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
+ *
+ * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
+ * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
+ * @param[in] DATA_TYPE        The data type of the vectors
+ * @param[in] BASENAME         The basename of the variables
+ * @param[in] PTR              The base pointer
+ * @param[in] STRIDE_Y         The stride value in y-axis direction
+ * @param[in] Z                The offset in z-axis direction
+ * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
+ * @param[in] N                Total number of columns. Used to detect if current block is at the boundary in x.
+ * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
+ */
+#define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, N, PARTIAL_COND_X) \
+    if(!(PARTIAL_COND_X))                                                                                            \
+    {                                                                                                                \
+        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                      \
+    }                                                                                                                \
+    else                                                                                                             \
+    {                                                                                                                \
+        STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                        \
+    }
+/** Store a block that can only be partial in y but not x.
+ *
+ * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
+ *
+ * The data to store is expected to have consecutive names for each row.
+ * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
+ * The Z offset is expected to have consecutive names.
+ * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
+ *
+ * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
+ * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
+ * @param[in] DATA_TYPE        The data type of the vectors
+ * @param[in] BASENAME         The basename of the variables
+ * @param[in] PTR              The base pointer
+ * @param[in] STRIDE_Y         The stride value in y-axis direction
+ * @param[in] Z                The offset in z-axis direction
+ * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
+ * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
+ */
+#define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \
+    if(!(PARTIAL_COND_Y))                                                                                         \
+    {                                                                                                             \
+        STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
+    }                                                                                                             \
+    else                                                                                                          \
+    {                                                                                                             \
+        STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
+    }
+/** @} */ // end of group STORE_BLOCK_PARTIAL
+
+#if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
+
+/** Boundary-aware GEMM block store
+ * @name STORE_BLOCK_BOUNDARY_AWARE
+ * This macro assumes the following schemes to achieve boundary-awareness:
+ *  - Overlapping load in Y axis from lhs tensor. This implies lhs has no padding along y dim.
+ *  - Non-Overlapping(normal) load from rhs tensor. This imples rhs can have paddings.
+ *  - Overlapping load in Y axis from bias tensor. This implies rhs has no padding along y dim.
+ * The macro then ensures that the dst tensor can be stored without any paddings in both x and y dim.
+ *
+ * In the y dimension, we place the partial blocks **at the beginning** while in the x dimension, we place the partial
+ * blocks **at the end**.
+ * Say, the dst tensor is of shape MxN and we have M0 and N0 as the block size, this is how we define "partial blocks"/
+ * "boundary block" (we use the 2 terms "partial blocks" and "boundary blocks" interchangeably) and its various parameters:
+ *
+ *  *--x-->                         x == 0                        x == 1
+ *  |                  |<------------------------------N-------------------------->|
+ *  y                  |<--------------N0------------->|<----PARTIAL_STORE_N0----->|
+ *  |     -------------#############################################################
+ *  *     |          | |...............................|...........................|
+ * y == 0 | PAR_..._M0 |......Boundary block in y......|.Boundary block in x and y.|
+ *        |          | |...............................|...........................|
+ *        M          --#############################################################
+ *        |          | |                               |...........................|
+ * y == 1 |         M0 |      Non-boundary block       |....Boundary block in x....|
+ *        |          | |                               |...........................|
+ *        |------------#############################################################
+ *
+ * Then @p PARTIAL_STORE_M0 = M % M0      and @p PARTIAL_STORE_N0 = N % N0
+ *
+ * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
+ *
+ * It automatically detects if a giving M,N,M0,N0 combination can yield partial blocks in either X and Y dimension,
+ * and select corresponding store methods such that the boundary detection logic is only added when needed.
+ *
+ * The data to store is expected to have consecutive names for each row.
+ * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
+ * The Z offset is expected to have consecutive names.
+ * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
+ *
+ * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
+ * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
+ * @param[in] DATA_TYPE        The data type of the vectors
+ * @param[in] BASENAME         The basename of the variables
+ * @param[in] PTR              The base pointer
+ * @param[in] STRIDE_Y         The stride value in y-axis direction
+ * @param[in] Z                The offset in z-axis direction
+ * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
+ * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported: [0, @p N0)
+ * @param[in] N                Total number of columns. Used to detect if current block is at the boundary in x.
+ * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
+ * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
+ * @{
+ */
+#if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
+// Case1: No partial blocks in either x or y
+#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
+    STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
+
+#elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0
+// Case2: Partial blocks in y
+#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
+    STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y)
+
+#elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0
+// Case3: Partial blocks in x
+#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
+    STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, N, PARTIAL_COND_X)
+
+#else // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
+// Case4: Partial blocks in both x and y
+#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
+    STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X)
+
+#endif // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
+
+#else // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
+
+#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, PARTIAL_COND_Y, PARTIAL_COND_X) \
+    STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
+
+#endif    // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
+/** @} */ // end of group STORE_BLOCK_BOUNDARY_AWARE
+
+#if defined(PARTIAL_STORE_M0)
+/** Compute the start m0 row (LHS, BIAS and DST) in a boundary-aware way so as to avoid padding
+ * @name COMPUTE_M0_START_ROW
+ * If there're any partial blocks in y dimension, they are placed at the beginning of the rows.
+ * This shift amount is added to all rows such that the partial block (at the beginning) overlaps with the subsequent
+ * blocks in the y dimension to avoid any padding.
+ * EG: M0=4, PARTIAL_STORE_M0=1:
+ *                  | Non-overlapping | +M0_ROW_SHIFT (Overlapping)
+ * block 0 (partial)| start row = 0   | start row = 0
+ * block 1 (full)   | start row = 4   | start row = 1
+ * block 2 (full)   | start row = 8   | start row = 5
+ *
+ * @param[in] y                Global id of current block in y.
+ * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
+ * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
+ * @{
+ */
+#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
+    ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0))))
+#else // defined(PARTIAL_STORE_M0)
+#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
+    ((uint)(y * M0))
+#endif    // defined(PARTIAL_STORE_M0)
+/** @} */ // end of group COMPUTE_M0_START_ROW
+
+/** Store a vector that can only be partial in x.
+ *
+ * @note in case @p vec_size or @p leftover != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
+ *
+ * The data to store is expected to end in a 0.
+ * E.g., for basename=c, the expected name is c0.
+ *
+ * @param[in] basename  The name of the variable without trailing 0
+ * @param[in] data_type The data type of the vector
+ * @param[in] ptr       The base pointer
+ * @param[in] vec_size  The vector size if cond = false. Supported: 1, 2, 3, 4, 8, 16
+ * @param[in] leftover  The vector size if cond = true. Supported range: [1, @p vec_size0)
+ * @param[in] cond      Condition to select either vec_size0 or vec_size1
+ * @{
+ */
+#define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond) \
+    STORE_BLOCK_PARTIAL_IN_X(1, vec_size, data_type, basename, ptr, 0, 0, leftover, 0, cond)
+/** @} */ // end of group STORE_VECTOR_SELECT
\ No newline at end of file