Fix out-of-bound reads in cl gemm kernels

* Revert "Remove padding in FP Cl Gemm kernels"
  This reverts commit 48717a3d38fef8d316cd4b9fd9a3bc1a43db736b.
* Allow different boundary row handling strategies across native,
reshaped and reshaped_only_rhs kernels by introducing a
ELTWISE_OPERAND_ROW parameter to the macro

Resolves COMPMID-4919

Change-Id: Icefc23c0760a6abb838fef1d0d5bda06b07c79e3
Signed-off-by: SiCongLi <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6569
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

6 files changed, 145 insertions, 123 deletions

diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h
index fc9704f13b..070c47bd7f 100644
--- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h
+++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h
@@ -54,6 +54,9 @@
  * @param[in] N0                     The number of consecutive columns
  * @param[in] BASENAME               The basename of the result variables
  * @param[in] ELTWISE_OPERAND_NAME   The basename of the other operand variables
+ * @param[in] ELTWISE_OPERAND_ROW    The starting row of the other operand variables. Required as different boundary handling strategies are used by different kernels
+ *                                   E.g. reshaped_only_rhs and native kernels shifts rows (by using COMPUTE_M0_START_ROW) to handle boundary rows,
+ *                                   whereas reshaped kernels do not shift rows
  * @param[in] DATA_TYPE              Data type of the result variables
  * @param[in] DATA_TYPE_ACCUMULATR   Higher-precision accumulator data type in case of mixed-precision op
  * @param[in] ZERO                   Zero vector for z offset
@@ -66,26 +69,25 @@
 #if defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH)
 #if P2_ELTWISE_ARG1_HEIGHT == 1
 #if P2_ELTWISE_ARG1_WIDTH == 1 // Case 1: Broadcasting in both X and Y; op2 arg tile shape[YxX] == [1x1]
-#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X)         \
-    __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \
-    VEC_DATA_TYPE(DATA_TYPE, 1)                                                                                                                                                           \
-    ELTWISE_OPERAND_NAME##0 = VLOAD(1)(0, (__global DATA_TYPE *)ELTWISE_OPERAND_NAME##_addr);                                                                                             \
+#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_ROW, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
+    __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z;              \
+    VEC_DATA_TYPE(DATA_TYPE, 1)                                                                                                                                                                        \
+    ELTWISE_OPERAND_NAME##0 = VLOAD(1)(0, (__global DATA_TYPE *)ELTWISE_OPERAND_NAME##_addr);                                                                                                          \
     MIXED_PRECISION_ELTWISE_OP_BLOCK_BROADCAST(OP, M0, 1, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE_ACCUMULATOR, ELTWISE_OPERAND_NAME##_hp);
 #else // P2_ELTWISE_ARG1_WIDTH == 1; Case 2: Broadcasting in only Y; op2 arg tile shape[YxX] == [1xN0]
-#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X)                                                             \
+#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_ROW, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X)                                        \
     __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \
     LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_NAME##_addr, 0, ELTWISE_OPERAND_NAME##_stride_y, ZERO, 1, PARTIAL_LOAD_N0, false, PARTIAL_COND_X);                                                      \
     MIXED_PRECISION_ELTWISE_OP_BLOCK_BROADCAST(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE_ACCUMULATOR, ELTWISE_OPERAND_NAME##_hp);
 #endif // P2_ELTWISE_ARG1_WIDTH == 1
 #else  // P2_ELTWISE_ARG1_HEIGHT == 1; Case 3: No broadcasting; op2 arg tile shape[YxX] == [M0xN0]
-#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X)                                                                                                                               \
-    __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * ELTWISE_OPERAND_NAME##_stride_y) + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \
-    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_NAME##_addr, 0, ELTWISE_OPERAND_NAME##_stride_y, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X);                                                                                                \
+#define POST_OP2_ELTWISE_OP(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_ROW, DATA_TYPE, DATA_TYPE_ACCUMULATOR, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X)                                                                                                  \
+    __global uchar *ELTWISE_OPERAND_NAME##_addr = ELTWISE_OPERAND_NAME##_ptr + ELTWISE_OPERAND_NAME##_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (ELTWISE_OPERAND_ROW * ELTWISE_OPERAND_NAME##_stride_y) + get_global_id(2) * ELTWISE_OPERAND_NAME##_stride_z; \
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, ELTWISE_OPERAND_NAME, ELTWISE_OPERAND_NAME##_addr, 0, ELTWISE_OPERAND_NAME##_stride_y, ZERO, PARTIAL_LOAD_M0, PARTIAL_LOAD_N0, PARTIAL_COND_Y, PARTIAL_COND_X);                                                                                        \
     MIXED_PRECISION_ELTWISE_OP_BLOCK(OP, M0, N0, BASENAME, ELTWISE_OPERAND_NAME, DATA_TYPE_ACCUMULATOR, ELTWISE_OPERAND_NAME##_hp);
 #endif    // P2_ELTWISE_ARG1_HEIGHT == 1
 #endif    // defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH)
 /** @} */ // end of group POST_OP2_ELTWISE_OP
-
 /** Post Op 3: Activation Block (Optional)
  * @name POST_OP3_ACTIVATION_OPTIONAL
  * Toggled by -DP3_ACTIVATION_TYPE
diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl
index e53ce3d1b2..bbe97b2781 100644
--- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl
+++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl
@@ -170,7 +170,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
     // Compute RHS matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);
@@ -187,7 +187,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 
 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -294,17 +294,13 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
         rhs_offset += rhs_stride_y;
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -327,7 +323,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
 
-    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+    LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
@@ -337,10 +333,9 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
-                                    2) * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
-    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
@@ -352,10 +347,13 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl
index 758fd327fe..9e9a73ccf6 100644
--- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl
+++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl
@@ -409,7 +409,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_post_act_eltwise_op_act(IMAGE_DECLAR
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
@@ -626,7 +626,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
@@ -1069,7 +1069,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLAR
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
@@ -1384,7 +1384,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture_post_act_eltwise_op_act(IMAG
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, get_global_id(1) * (uint)M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
index 508ee96d2d..fe2d103de5 100644
--- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
+++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
@@ -214,6 +214,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -222,7 +225,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
     // Compute RHS reshaped matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
@@ -239,7 +242,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT
 
 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -338,18 +341,14 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT
         rhs_offset += sizeof(DATA_TYPE);
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
 
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -382,7 +381,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -399,7 +398,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
@@ -480,6 +479,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -488,7 +490,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
@@ -506,7 +508,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_
 
 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -654,18 +656,14 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_
 
 #endif // LEFTOVER_K != 0
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
 
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -698,7 +696,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -715,7 +713,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
@@ -870,6 +868,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -878,7 +879,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
     // Compute RHS reshaped matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
@@ -896,7 +897,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA
 #if defined(REINTERPRET_INPUT_AS_3D)
 
     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -1020,17 +1021,13 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA
         rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE);
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -1063,7 +1060,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -1080,7 +1077,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
@@ -1157,6 +1154,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -1165,7 +1165,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
@@ -1184,7 +1184,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE
 #if defined(REINTERPRET_INPUT_AS_3D)
 
     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -1299,17 +1299,13 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE
         x_rhs += RHS_STEP_X;
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -1342,7 +1338,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -1359,7 +1355,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
 
diff --git a/src/core/CL/cl_kernels/common/gemm.cl b/src/core/CL/cl_kernels/common/gemm.cl
index 6502dd496a..a76ad458a6 100644
--- a/src/core/CL/cl_kernels/common/gemm.cl
+++ b/src/core/CL/cl_kernels/common/gemm.cl
@@ -1097,6 +1097,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -1105,7 +1108,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
     // Compute RHS reshaped matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
@@ -1122,7 +1125,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
 
 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -1221,18 +1224,14 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
         rhs_offset += sizeof(DATA_TYPE);
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
 
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -1265,7 +1264,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -1395,6 +1394,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -1403,7 +1405,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
@@ -1421,7 +1423,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
 
 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -1569,18 +1571,14 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
 
 #endif // LEFTOVER_K != 0
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
 
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -1613,7 +1611,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -1818,6 +1816,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -1826,7 +1827,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
     // Compute RHS reshaped matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
@@ -1844,7 +1845,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
 #if defined(REINTERPRET_INPUT_AS_3D)
 
     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -1968,17 +1969,13 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
         rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE);
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -2011,7 +2008,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -2137,6 +2134,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
     uint y = get_global_id(1);
     uint z = get_global_id(2);
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
@@ -2145,7 +2145,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
@@ -2164,7 +2164,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
 #if defined(REINTERPRET_INPUT_AS_3D)
 
     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -2279,17 +2279,13 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
         x_rhs += RHS_STEP_X;
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -2322,7 +2318,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
     LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 
@@ -2714,7 +2710,6 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
 
@@ -2986,7 +2981,6 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs),
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
 
@@ -3297,7 +3291,6 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs),
     const uint y = get_global_id(1);
     const uint z = get_global_id(2);
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
 
@@ -3853,7 +3846,6 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs),
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
 
@@ -4124,7 +4116,7 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
     // Compute RHS matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);
@@ -4141,7 +4133,7 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
 
 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -4248,17 +4240,13 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
         rhs_offset += rhs_stride_y;
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -4281,7 +4269,7 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
 
-    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+    LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
@@ -4291,10 +4279,9 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
-                                    2) * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
-    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
@@ -4310,6 +4297,9 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(ACTIVATION_TYPE)
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
     // Store output block
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 }
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
index 7ad3d55fe0..c3efc24fa9 100644
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
+++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
@@ -134,12 +134,15 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITens
                                                         const GEMMRHSMatrixInfo &rhs_info,
                                                         const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
 {
-    ARM_COMPUTE_UNUSED(src0, src1, src2);
     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
     bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d;
     bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;
 
+    Window win{};
+    Window win_out{};
+    bool   window_changed = false;
+
     // In case both input and dst have to be reinterpreted as 3D tensors,
     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
     if(reinterpret_input_as_3d == reinterpret_output_as_3d)
@@ -147,6 +150,9 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITens
         reinterpret_output_as_3d = false;
     }
 
+    // dst tensor auto initialization if not yet initialized
+    auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
+
     TensorInfo tmp_info(*dst);
 
     if(reinterpret_output_as_3d)
@@ -162,14 +168,44 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITens
     num_elems_processed_per_iteration_x = rhs_info.n0;
     num_elems_processed_per_iteration_y = lhs_info.m0;
 
-    Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    win_out = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+    AccessWindowStatic src0_access(src0, 0, 0,
+                                   src0->dimension(0),
+                                   src0->dimension(1));
+    AccessWindowStatic src1_access(src1, 0, 0,
+                                   ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x),
+                                   src1->dimension(1));
+    AccessWindowStatic dst_access(dst, 0, 0,
+                                  dst->dimension(0),
+                                  dst->dimension(1));
+
+    if(src2 != nullptr)
+    {
+        const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
+
+        AccessWindowStatic src2_access(src2, 0, 0,
+                                       ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x),
+                                       src2->dimension(1));
+
+        window_changed = update_window_and_padding(win, src0_access, src1_access, src2_access) || // window used by the execute_window_loop
+                         update_window_and_padding(win_out, dst_access);                          // window used to update the padding requirements of dst tensor
+    }
+    else
+    {
+        window_changed = update_window_and_padding(win, src0_access, src1_access) || // window used by the execute_window_loop
+                         update_window_and_padding(win_out, dst_access);             // window used to update the padding requirements of dst tensor
+    }
 
     // Collapse along the Z direction
     // This collapse needs to be here in order to tune the Z dimension of LWS
+    Window             collapsed             = win;
     const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
-    Window             collapsed             = win.collapse(win, dimension_to_collapse);
+    collapsed                                = win.collapse(win, dimension_to_collapse);
 
-    return std::make_pair(Status{}, collapsed);
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, collapsed);
 }
 } // namespace
 
@@ -190,7 +226,7 @@ void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile
 
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
 
-    auto padding_info         = get_padding_info({ src0, src1, src2, dst });
+    auto padding_info         = get_padding_info({ src0, dst });
     _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d;
     _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());