Remove padding in FP Cl Gemm kernels

* Remove rhs and bias padding in ClGemmMatrixMultiplyNativeKernel
* Rework ClGemmMatrixMultiplyReshapedOnlyRHSKernel to use the same
  padding boundary condition as the other kernels

Partially resolves COMPMID-4435

Change-Id: I1c17af9cca0b5cb3be087ce160948b7b0e62d297
Signed-off-by: SiCongLi <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6549
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 60 insertions, 83 deletions

diff --git a/src/core/CL/cl_kernels/common/gemm.cl b/src/core/CL/cl_kernels/common/gemm.cl
index dd03147ad6..9732588b31 100644
--- a/src/core/CL/cl_kernels/common/gemm.cl
+++ b/src/core/CL/cl_kernels/common/gemm.cl
@@ -1096,9 +1096,6 @@ __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))
     {
@@ -1107,7 +1104,7 @@ __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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * 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;
@@ -1124,7 +1121,7 @@ __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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * 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
@@ -1223,14 +1220,18 @@ __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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_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);
 
     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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, y * 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
@@ -1263,7 +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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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)) + (y * 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);
 
@@ -1392,9 +1393,6 @@ __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 +1401,7 @@ __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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * 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 +1419,7 @@ __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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * 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
@@ -1569,14 +1567,18 @@ __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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_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);
 
     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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, y * 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
@@ -1609,7 +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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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)) + (y * 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);
 
@@ -1813,9 +1815,6 @@ __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))
     {
@@ -1824,7 +1823,7 @@ __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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * 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;
@@ -1842,7 +1841,7 @@ __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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zin, y * 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
@@ -1966,13 +1965,17 @@ __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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_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);
 
     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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, y * 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
@@ -2005,7 +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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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)) + (y * 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);
 
@@ -2130,9 +2133,6 @@ __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))
     {
@@ -2141,7 +2141,7 @@ __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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * 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
@@ -2160,7 +2160,7 @@ __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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zin, y * 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
@@ -2275,13 +2275,17 @@ __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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_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);
 
     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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, y * 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
@@ -2314,7 +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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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)) + (y * 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);
 
@@ -2706,6 +2710,7 @@ __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);
 
@@ -2977,6 +2982,7 @@ __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);
 
@@ -3287,6 +3293,7 @@ __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);
 
@@ -3842,6 +3849,7 @@ __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);
 
@@ -4111,7 +4119,7 @@ __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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
 
     // Compute RHS matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);
@@ -4128,7 +4136,7 @@ __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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * 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
@@ -4235,13 +4243,17 @@ __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)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_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);
 
     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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, y * 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
@@ -4264,7 +4276,7 @@ __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(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
@@ -4274,9 +4286,10 @@ __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 + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+    __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;
 
-    LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
+    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);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
@@ -4292,9 +4305,6 @@ __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 6c872fd48c..e389ce5b0c 100644
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
+++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
@@ -119,15 +119,12 @@ 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)
@@ -135,9 +132,6 @@ 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)
@@ -153,44 +147,14 @@ 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;
 
-    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
-    }
+    Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
 
     // 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);
-    collapsed                                = win.collapse(win, dimension_to_collapse);
+    Window             collapsed             = win.collapse(win, dimension_to_collapse);
 
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, collapsed);
+    return std::make_pair(Status{}, collapsed);
 }
 } // namespace
 
@@ -208,7 +172,10 @@ 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, dst });
+    // 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)));
+
+    auto padding_info         = get_padding_info({ src0, src1, src2, 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());